Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Discrimination of Internal Faults and Other Transients in an Interconnected System with Power Transformers and Phase Angle Regulators

Discrimination of Internal Faults and Other Transients in an Interconnected System with Power... IEEE SYSTEMS JOURNAL 1 Discrimination of Internal Faults and Other Transients in an Interconnected System with Power Transformers and Phase Angle Regulators Pallav Kumar Bera, Can Isik, Senior Member, IEEE, and Vajendra Kumar Abstract—This study solves the problem of accurate detection to magnetizing and sympathetic inrush, and CT saturation of internal faults and classification of transients in a 5-bus during external faults are the major problems associated with interconnected system for Phase Angle Regulators (PARs) and differential protection. Second-harmonic restraint method is Power Transformers (PTs). The analysis prevents mal-operation extensively used to distinguish internal faults from magnetiz- of differential relays in case of transients other than faults which ing inrush since more second-harmonic component exists in include magnetizing inrush, sympathetic inrush, external faults with Current Transformer (CT) saturation, capacitor switch- inrush currents than in internal faults [1]. However, higher ing, non-linear load switching, and ferroresonance. A gradient second-harmonics are generated during internal faults with boosting classifier (GBC) is used to distinguish the internal CT saturation, presence of shunt capacitance, or because of faults from the transient disturbances based on 1.5 cycles of the distributed capacitance of EHV lines [2]. In addition, 3-phase differential currents registered by a change detector. the second-harmonic content in inrush currents has reduced After the detection of an internal fault, GBCs are used to locate the faulty unit (PT, PAR series, or exciting unit) and in modern transformers with soft core material [3]. Hence, identify the type of fault. In case a transient disturbance is several cases of mal-operation of conventional relays in dis- detected, another GBC classifies them into the six disturbances. tinguishing faults and inrush have been reported [4]. CT Five most relevant frequency and time domain features obtained saturation during external faults may also cause false trips due using Information Gain are used to train and test the classifiers. to the inefficient setting of commonly used dual-slope biased The proposed algorithm distinguishes the internal faults from the other transients with a balanced accuracy () of 99.95%. differential relays [5]. The faulty transformer unit is located with  of 99.5% and Phase Angle Regulators or Phase Shifters or Phase Shift the different transient disturbances are identified with  of Transformers are a special class of transformers used to 99.3%. Moreover, the reliability of the scheme is verified for control real power flow in parallel transmission lines. They different rating and connection of the transformers involved, CT ensure the system reliability and allow easier integration of saturation, and noise level in the signals. These GBC classifiers can work together with a conventional differential relay and new generations with the grid. By regulating the phase angle offer a supervisory control over its operation. PSCAD/EMTDC between the sending and receiving ends they prevent over- software is used for simulation of the transients and to develop loading of a line and re-routes power via another line. PARs the two and three-winding transformer models for creating the can be categorized on the basis of the number of cores and internal faults including inter-turn and inter-winding faults. magnitude of sending end voltage with respect to the receiving Index Terms—Phase Angle Regulators, Power Transformers, end. Indirect Symmetrical Phase Angle Regulators (ISPAR) Fault detection, Transients, Gradient Boosting Classifier having the same sending and receiving end voltages with two transformer units, namely, series and exciting (Fig.1(b)), has I. INTRODUCTION been chosen as one of the subjects (other being the PTs) in this study because of their popularity and security against higher OWER Transformers are an integral part of an electrical voltage levels as the load tap changer (LTC) is not exposed grid and their protection is vital for reliable and stable to system disturbances. The exciting unit is responsible for operation of the power system. An important requirement creating the required phase difference to regulate the power of the protection system is the faithful discrimination of which can be controlled by the LTC connections and an faults from other transients. Differential protection has been advance-retard-switch located on its secondary winding [6]. the primary protection scheme in transformers because of The modified real power flow in a transmission line with a its inherent selectivity and sensitivity. Mal-operations due PAR is given by Manuscript received February 20, 2020; revised May 23, 2020; accepted V  V July 10, 2020.(Corresponding author: Pallav Kumar Bera.) S L P =  sin( + ) (1) Pallav Kumar Bera & Can Isik are with the Department of Electrical X + X line PAR Engineering and Computer Science, Syracuse University, NY, USA (e-mail: pkbera@syr.edu; cisik@syr.edu). where, V is source voltage, V is load voltage;  is the phase S L Vajendra Kumar was with Indian Institute of Technology, Roorkee, India angle difference between V and V ; X ; X are the S L line PAR (email: kumarvajendra@gmail.com). 2020 IEEE. Personal use of this material is permitted. Permission from transmission line and PAR reactance respectively; and is the IEEE must be obtained for all other uses, in any current or future media, new constraint added which is responsible for controlling the including reprinting/republishing this material for advertising or promotional power flow. The PARs similar to PTs require a fast, sensitive, purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. secure, and dependable protection system. Discriminating ex- arXiv:2004.06003v2 [eess.SP] 2 Aug 2020 IEEE SYSTEMS JOURNAL 2 ternal faults with CT saturation, magnetizing inrush, and other evaluates the effect of noise, CT saturation, and change in transient disturbances from internal faults is a challenge for the transformer rating and connection on the proposed scheme. protection systems of PARs as well. Moreover, methods used Section V concludes the paper. to compensate the phase for differential relays in PTs with a fixed phase shift are not applicable in PARs with variable phase shift [7]. Authors have used different intelligent methods to distin- guish internal faults and magnetizing inrush in PTs in the past decade. A combination of Artificial Neural Network (ANN) and spectral energies of wavelet components was used to discriminate internal faults and inrush in [8]. Support Vector Machines (SVM) and Decision Tree (DT) based trans- former protection were proposed in [9], [10] and [11]–[13] respectively. Probabilistic Neural Network (PNN) has been used to detect different conditions in PT operation in [14]. Random Forest Classifier (RFC) was proposed to discrimi- nate internal faults and inrush in [15]. Works of literature also suggest extensive use of S-Transform, Hyperbolic S- Transform, Wavelet Transform (WT) to detect Power Quality Fig. 1: (a) 5-bus interconnected system with ISPARs, PTs, T-lines, and AC (PQ) transient disturbances and then classify them using DT, sources, (b) Series and exciting transformers in ISPAR SVM, ANN, PNN [16]–[21]. These transient disturbances are caused by variations in load, capacitor switching, charging of transformers, starting of induction machines, use of non-linear II. MODELING AND S IMULATION loads, etc. Contrarily, literature investigating internal faults and inrush in an ISPAR is limited. However, attempts were made in The power network chosen for the simulation of the internal faults and the transient disturbances is based on a proposed [22] where internal faults are distinguished from magnetizing Pumped-storage (efficient form of renewable storage designed inrush using WT and then the internal faults are classified to meet energy needs and reduce emissions by utilizing the using ANN and in [23] where the internal faults in series and energy stored in an upper water body pumped from a lower exciting transformers of the ISPAR are classified using RFC. water body) project in California, USA [25]. Further, the authors have predominately used an isolated and PSCAD/EMTDC is used for the modeling and simulation simple network having a PT [8]–[14] or a PAR [22], [23] to of the transients in the ISPAR and PT in the chosen intercon- support their proposed protection scheme. Also, the transient nected power system. Fig.1(a) shows the single-line diagram disturbances have not been studied rigorously in these works. This paper studies the use of Decision Tree based algo- of the 5-bus interconnected model consisting of the AC source, rithms to discriminate the internal faults and other transient transmission lines, ISPARs, PTs, and 3-phase loads working at disturbances including magnetizing inrush and CT saturation 60Hz. The ISPARs have a rating of 500 MVA, 230kV/230kV, during external faults in a 5-bus interconnected system with with phase angle variations of 25 and the PTs are rated at Phase Angle Regulators and Power Transformers which has 500 MVA, 500kV/230kV. The AC source consists of 9 units not been attempted before. Customized two-winding and three- of 120 MVA, 13.8kV hydro-generators. Two transformers are winding transformers are developed to simulate the internal used in cascade to step up the voltage from 13.8kV to 500kV. faults. A change detector has been used to detect and register 3 ISPARs (ISPAR , ISPAR , and ISPAR ) are connected 1 2 3 the differential currents. Five most relevant time and frequency between bus4 and bus5 through transformers T , T , and T . 1 2 3 domain features have been used to train SVM, RFC, DT, and Only the internal faults in ISPAR and T are studied. 1 1 GBC classifiers to detect, locate and identify the internal faults The three-winding transformer required for the series units and classify six transient disturbances. The proposed scheme of ISPAR and the two-winding transformer required for the is tested on 101,088 transients cases simulated on Power exciting units of ISPAR and PTs for the simulation of various System Computer Aided Design (PSCAD)/ Electromagnetic internal faults including turn-to-turn and primary-to-secondary Transients including DC (EMTDC) by varying various system winding faults are developed in PSCAD/EMTDC with Fortran. parameters. The entire dataset of internal faults and other The self-inductance terms (Li) and the mutual inductance disturbances is made available on IEEE Dataport [24]. terms (Mij) of the 4 4 L matrix (Eq.2) of the single-phase The rest of the paper is organized as follows. Section II two-winding transformer and 6  6 L matrix of the single- illustrates the modeling and simulation of the internal faults phase three-winding transformer are evaluated from primary and other transient disturbances in the power network con- and secondary voltages, the magnetizing component of the taining PTs and ISPARs. Section III comprises the detection no-load excitation current (I ), and the short-circuit tests. of internal faults, feature extraction and selection, and the The modeled components have the provision to change the classifiers used for the detection and identification of tran- saturation characteristics, % of winding shorted and other sients. Section IV includes the results for detection of internal parameters. The Fortran script of the two-winding transformer faults, identification of faults and transient disturbances, and is shown in the Appendix. IEEE SYSTEMS JOURNAL 3 exciting unit. Fig.2(b) shows the differential currents for LTC= 2 3 Lx Mxy Mxz Mxw full, fault inception time=15.0124s, backward phase shift, 6Myx Ly Myz Myw7 L = (2) 4 5 Mzx Mzy Lz Mzw fault resistance=0.01 and % turns shorted=40 in primary of Mwx Mwy Mwz Lw series unit. Fig.2(c) shows the differential currents for LTC= full, fault inception time=15.01518s, forward phase shift, fault The study covers various internal faults in the ISPAR resistance=0.01 and % turns shorted=60 in primary of the PT. and PT, capacitor switching, switching of non-linear loads, magnetizing inrush, sympathetic inrush, external faults with CT saturation, and ferroresonance. In the following para- graphs, these conditions are considered one after the other. The simulation run-time, fault/disturbance inception time, and fault duration time are 15.2s, 15.0s, and 0.05s (3 cycles) respectively in all cases. The multi-run component is used to change the parameter values wherever possible to get the different simulation cases and snapshots of the first simulation runs are used to start the simulation from initialized conditions to reduce the simulation time. A. Internal Faults The internal faults are created in the PT, ISPAR series, and ISPAR exciting unit. 88,128 internal fault cases which include basic internal faults, turn-to-turn, and winding-to- winding faults are simulated by varying the fault resistance, % of winding shorted, fault inception time, forward or backward shift, and the LTC in the exciting unit. Fig. 2: 3-phase differential currents for turn-to-turn faults in (a) primary of 1) Internal phase & ground faults (ph & g): Phase winding exciting unit, (b) primary of series unit, and (c) primary of PT to ground (w -g, w -g, w -g), phase winding to phase winding a b c to ground (w -w -g, w -w -g, w -w -g), phase winding to a b a c b c TABLE II: Parameters for winding-to-winding & turn-to-turn faults in the phase winding (w -w , w -w , w -w ), 3-phase winding (w - a b a c b c a ISPAR and PT w -w ), and 3-phase winding to ground (w -w -w -g) faults b c a b c Variables Values are simulated in the primary (P) and secondary (S) sides of Fault resistance 0.01, 0.5 & 10 (3) the PT and on the primary and secondary sides of exciting and % of winding shorted 20%, 40%, 60%, 80% (4) Fault inception time 15s to 15.0153s in steps of 1.38ms (12) series transformer units in the ISPAR. Table I shows the values Transformer phase a,b,c (P & S) (6) of different system and fault parameters in T and ISPAR 1 1 Fault location ISPAR Exciting phase a,b,c (P & S) (6) (Fig.1(a)) which are varied to get the training and testing cases & ISPAR Series phase a,b,c (P & S) (6) Phase shift Forward and backward (2) for the internal phase & ground faults. LTC 0.2,0.4,0.6,0.8,1 [1 & 0.5 in ISPAR exciting] TABLE I: Parameters for ph & g faults in the ISPAR and PT Transformer or ISPAR series(T-T) faults = 3412625 = 8640 ISPAR exciting(T-T) faults = 3412622 = 3456 Variables Values Transformer or ISPAR series(W-W) faults = 3412325= 4320 Fault resistance 0.01, 0.5 & 10 (3) ISPAR exciting(W-W) faults = 3412322 = 1728 % of winding shorted 20%, 50%, 80% (3) Fault type w-g, w-w-g, w-w, w-w-w & w-w-w-g (11) Fault inception time 15s to 15.0153s in steps of 1.38ms (12) 3) Winding-to-winding (W-W) faults: The electrical, ther- Transformer (P & S) (2) mal and mechanical stress due to short circuits and transformer Fault location ISPAR Exciting unit (P & S) (2) aging reduces the mechanical and dielectric strength of the & ISPAR Series unit (P & S) (2) Phase shift Forward and backward (2) winding and results in degradation of the insulation between LTC 0.2,0.4,0.6,0.8,1[1 & 0.5 in ISPAR exciting] LV and HV winding and may damage the winding eventually Transformer or ISPAR series faults = 331112225 = 23,760 [26]. Table II shows the values of different parameters of the ISPAR exciting faults = 331112222 = 9504 PT and the series and exciting unit of ISPAR used to simulate 10,368 winding-to-winding faults. 2) Turn-to-turn (T-T) faults: About 70-80% of faults in transformers are due to turn-to-turn insulation failures. Ther- B. Magnetizing inrush mal, mechanical and electrical stress degrades the insulation Transients caused by the energization of transformers are and causes turn-to-turn faults which may lead to more serious common and discrimination of inrush from fault currents has faults and inter-winding faults if not detected quickly [26]. been studied since the 19th century. The harmonic restraint Table II shows the values of different parameters of the PT and relays fail to detect inrush currents in transformers with the series and exciting unit of ISPAR used to simulate 20,736 modern core materials. The flux in a transformer core just turn-to-turn faults. Fig.2(a) shows the differential currents for after switching can be expressed as LTC= half, fault inception time=15s, backward phase shift, 0 0 fault resistance=0.01 and % turns shorted=20 in primary of  =  +  cos!t  cos!(t + t ) (3) R m m IEEE SYSTEMS JOURNAL 4 where,  represents residual flux,  represents maximum during switching of T which drives T to saturation can be R m 2 1 flux, and t is switching time. The transformer draws a high expressed as peaky non-sinusoidal current to meet the high flux demand 2+t = [(R + R )i + R i ] (4) sys T 1 sys 2 where R is the system resistance , and R is the resistance sys T of transformer T , i and i are magnetizing currents of T and 1 1 2 1 T . This interaction between the incoming and the in-service transformers leads to mal-operation of differential relays of the in-service transformer due to failure of harmonic restraint relays and may cause prolonged harmonic over-voltages [27]. The use of superconducting winding, soft magnetic material in the core, and CT local transient saturation are some factors responsible for these mal-operations [3] [28]. Sympathetic inrush is influenced by the residual flux ( ) of the incoming transformer, switching time (t ), and system resistance [29] and Fig. 3: B-H curve of transformer core takes place with the incoming transformer energized in either series or parallel. The magnitude and direction of  , and when switched on. Since this current flows only on one side t are altered and the incoming transformer T1 is connected of the transformer the differential scheme mal-operates. T in parallel to simulate the scenarios. Table III shows the (Fig.1(a)) is the incoming 3-phase transformer and DC sources values of the different parameters used to get the training and are used to get the desired  in the single-phase transformers. testing cases for sympathetic inrush. Fig.4(b) shows the 3- The values for the DC currents in phase-a, b, and c are phase differential currents for LTC = 0.2, switching time = obtained from the B-H curve of the transformer core material 15s, forward phase shift, and -80% residual flux. as shown in Fig.3. Table III shows the values of different parameters including  and t used to get the data for training D. External faults with CT saturation and testing for magnetizing inrush and Fig.4(a) shows the 3- phase differential currents for LTC = full, switching time = The differential currents become non-zero due to CT sat- 15s, forward phase shift, and -80% residual flux. uration in case of heavy through faults and may lead to a false trip. While raising the bias threshold ensures the security (i.e. no mal-operation), the dependability for in-zone resistive faults gets reduced. The external faults with CT saturation are simulated on the 500kV and 230kV buses (bus4 & bus5). The values for the different parameters are given in Table IV. Fig.5(a) shows the 3-phase differential currents for an external line-to-ground (lg) fault when LTC = 0.2, phase shift = forward, fault inception time = 15s, and fault resistance = 0.01 on the 230kV bus. Fig. 4: 3-phase differential currents for (a) Magnetizing inrush, and (b) Sympathetic inrush TABLE III: Parameters for Magnetizing and Sympathetic inrush Variables Values Residual flux 80%;40%; 0% in 3 phases; 5 3 = (15) Switching time 15s to 15.0153s in steps of 1.38ms (12) LTC 0.2 to full tap in steps of 0.2 (5) Phase shift Forward and backward (2) Total=15 12 5 2=1800 C. Sympathetic Inrush Sympathetic inrush occurs in the in-service transformer (T ) Fig. 5: 3-phase differential currents for (a) External fault with CT saturation, when the incoming transformer (T ) is energized in a resistive (b) Capacitor Switching, and (c) Ferroresonance network at no-load. The asymmetrical flux change per cycle IEEE SYSTEMS JOURNAL 5 TABLE IV: Parameters for External faults on 230kV & 500kV bus high transient inrush currents from capacitor-bank switching to avoid malfunctioning of instantaneous and time overcurrent Variables Values relays (50/51). A capacitor bank having 3 Legs of 500 MVAr Fault resistance 0.01, 0.5 & 10 (3) Fault type lg, llg, ll, lll & lllg (11) each is connected to the 230kV bus. Capacitor bank reactors Fault inception time 15s to 15.0153s in steps of 1.38ms (12) and resistors are used in each Leg to reduce the effect of LTC 0.2 to full tap in steps of 0.2 (5) transients in voltages. Table VI shows the different parameters Phase shift Forward and backward (2) Fault location 230kV & 500kV bus (2) and their values used to get the data for training and testing Total=3 11 12 5 2 2=7920 for capacitor switching. Fig.5(b) shows the 3-phase differential currents for LTC = full, switching time = 15.00138s, and switching of 3 Legs of the capacitor bank. E. Non-linear Load Switching TABLE VI: Parameters for Capacitor Switching With the advancement in semiconductor technology and the use of non-linear loads with power converters, harmonic Variables Values contents in the line currents have increased. The differential re- Capacitor bank rating 500,1000,1500 MVAr (3) lays may mal-operate when non-linear loads e.g steel furnaces Switching time 15s to 15.0153s in steps of 1.38ms (12) Phase shift Forward and backward (2) are switched in a network containing transformers because LTC 0.2 to full tap in steps of 0.2 (5) of mutual enhancement effects between the transformer core Total = 3 12 2 5 = 360 and the load causing extreme saturation of the transformer core for several cycles [30]. The harmonic information has been used to discriminate faults from other disturbances and G. Ferroresonance locate the faults in the transmission line using SVM and ANN Initiated by faults and switching operations, ferroresonance [31]. A thyristor-based 6-pulse bridge rectifier with a wye- causes harmonics and overvoltages and may lead to mal- delta transformer as the non-linear load is connected to the operation of protective relays and damage of power equipment 230 kV bus to obtain the training and testing cases for load [33]. Mal-operation of the differential relay occurs because of switching. The values for the different parameters are given the higher magnitude of current in the HV side than the LV in table V and Fig.6 shows the phase-a differential current side [34]. Besides, the low loss, amorphous core transformer for LTC = full, switching time = 15s and firing angle of 0 . increases the intensity and occurrence of ferroresonance [35]. Fig.6(a) shows the transient and Fig.6(b) shows the steady- Several configurations may lead to ferroresonance in electrical state differential current after the switching. systems. In this paper, one such arrangement has been modeled when one of the phases of a 3-phase transformer is switched off. The parameters and their values for ferroresonance con- ditions are presented in Table VII. Fig.5(c) shows the 3-phase differential currents for switching time = 15s and grading capacitance = 0.2F simulated between bus2 and bus4. TABLE VII: Parameters for Ferroresonance Variables Values Grading capacitance 0.02F to 0.2F in steps of 0.02F (10) Location a,b,c phases (3) Switching time 15s to 15.016s in steps of 0.69ms (24) Total=10 3 24=720 Fig. 6: Non-linear Load Switching (a) Transient, and (b) Steady-state differ- ential currents III. DETECTION, DISCRIMINATION & CLASSIFICATION ALGORITHM TABLE V: Parameters for Non-linear Load Switching A. Change detection filter (CDF) for transient detection Variables Values The change in the differential currents in case of transients Firing angle 0 , 10 , 20 , 30 , 40 , 50 (6) is detected by a change detection filter (CDF) which calculates Switching time 15s to 15.0153s in steps of 1.38ms (12) the difference between the cumulative sum of modulus of two LTC 0.2 to full tap in steps of 0.2 (5) Total = 6 12 5 = 360 consecutive cycles. 2n +t 2n +t c c X X CDF (t) = jId(x)j jId(x n )j (5) F. Capacitor Switching x=n +t x=n +t c c Capacitor banks are used to improve voltage profile, re- where x equals sample number which begins at the second nn duce losses, and enhance power factor. Mal-functioning of cycle, n equals number of samples in a cycle, ftg , n t=1 customer equipment due to voltage magnification coinciding equals total number of samples, and Id represents a, b, and c with capacitor switching is common. [32] used WT to detect phase differential currents. IEEE SYSTEMS JOURNAL 6 The change detection filter starts logging the data from the F5, autoregressive coefficients are the least-square estimates instant CDF(t) is greater than a threshold, th in any one of of ' which are obtained by minimizing Eq.8 with respect i s the 3-phases. In normal conditions when there is no transient, to ' ; ' :::; ' and lag P. 0 1 P the values of CDF(t) are nearer to zero [36]. [Id ' '  Id ::: '  Id ] (8) t 0 1 t1 P tP B. Feature Extraction & Selection t=p+1 Time series analysis of the differential currents helps in More than one feature can be extracted from the above time the classification and characterization of power system events. and frequency domain functions by varying their parameters. Features extracted from these time series are used as input to e.g (qh; ql) = (0.8,0.4) & (0.8,0.2) yields 2 features from the machine learning algorithms. Informative and distinctive change quantile and window length = 5, 10, and 15 would features that help to classify the events may range from simple return 3 features of linear trend. statistical functions to complex ones. Researchers have used time-frequency representations like Wavelet Transform [8]– [10], [13], [17], [18], [21] and Stockwell Transform [16], [19]– C. Classifiers [21] to extract features from the non-stationary transients to Tree-based learning algorithms like decision trees, random discriminate inrush and internal faults and for classification of forest, and gradient boosting are considered among the best PQ disturbances. In this paper, to differentiate the faults from and predominantly used supervised learning methods in prob- the other transient disturbances, three time-domain features lems related to data science. These estimators have higher and two frequency-domain features have been used. accuracy, stability and are easy to interpret. They can also A comprehensive number of features (794) from different handle non-linear relationships quite well. DT, RFC, GBC, domains are extracted from the 3-phase differential currents and SVM has been used to detect and classify the transients. obtained from the current transformers, CT1 and CT2 located 1) Decision Tree: Decision trees are distribution-free white near bus4 and bus5. The complete list of the features extracted box Machine Learning models that learn simple decision can be found in [37]. Out of these 794 features, Random rules inferred from the feature values. In 1984 Breiman et Forest is used to rank and select the features with maximum al. introduced Classification and Regression Trees (CART) Information Gain to distinguish between the different classes. [39]. Here, the CART algorithm implemented in scikit-learn The most relevant and common features for each of the is used which constructs binary trees by splitting the training classification tasks obtained after performing feature ranking set recursively till it reaches the maximum depth or a splitting belong to the set F =fF1, F2, F3, F4, F5g where, F1 is average doesn’t reduce the impurity measure. The candidate parent change quantile, F2 is Fourier transform (FT) coefficients, F3 data D is split into D and D at each node using a feature p l r is aggregate linear trend, F4 is spectral welch density, and F5 is (f ) and threshold that yields the largest Information Gain. autoregressive coefficients. Only those features of set F which The objective function IG which is optimized at each split are present in each of the 3-phase differential currents are is defined as used for training the classifiers to detect the faults, localize the N N faulty units, identify the fault type, and identify the disturbance l r IG(D ; f ) = I (D )  I (D )  I (D ) (9) p p l r type (Table XVI). The feature set F is detailed in what follows. N N p p F1, average change quantile calculates the average of ab- where, I is impurity measure, N ; N and N are the number of p l r solute values of consecutive changes of the time series inside samples at the parent and child nodes [40]. Gini, and entropy two constant values qh and ql as impurity measures are used in the hyperparameter search. n 1 1 2) Random Forest: RFC belongs to the family of ensemble avg: change quantile =  jId Id j (6) t+1 t trees which builds numerous base estimators and averages their t=1 predictions which produces a better estimator with reduced where, n equals number of sample points in the differential variance. Each tree constitutes a random sample (drawn with current between qh and ql, Id is a, b, and c phase differential replacement) of the training set and the best split is found currents with n sample points. at each node by considering a subset of input features. The F2, FT coefficients, (Xjk) returns the fourier coefficients of individual trees tend to overfit but averaging the predictions of 1-D discrete Fourier Transform for real input using fast FT as all trees reduces the variance [41]. The main hyperparameters n1 in RFC are no of estimators (number of trees in the forest), j2kt (Xjk) = Id  e( ); k 2 Z (7) t max depth (tree depth), and max features (feature size to t=0 consider when splitting a node). The no of jobs parameter F3, aggregate linear trend calculates the linear least-squares was also used to parallelize the construction of tree and regression for values of the time series over windows and computation of predictions by using more processing units. returns aggregated values of either intercept or standard error. Random Forest has also been used during feature selection F4, spectral welch density uses Welchs method to compute and ranking (III-B) to get the relative importance of the an estimate of the power spectral density by partitioning the features which is measured by the fraction of samples a feature time series into segments and then averaging the periodgrams contributes to and the mean decrease in impurity from splitting of the discrete Fourier transform of each segment [38]. the samples [42]. IEEE SYSTEMS JOURNAL 7 3) Gradient Boosting Classifier: GBC belongs to the class used for localization of faults and classification of transient of ensemble trees which builds the base estimators from weak disturbances. The scheme consists of a four-level classifier learners (w (x)) sequentially in a greedy manner which results design. The level-1 classifier (GBC-1) consists of the fault in a strong estimator [43] [44]. The newly added w tries to detector, which can apply supervisory control over the oper- minimize the loss function given f , step length ( ), and ation of the conventional differential relay. GBC-1 identifies p1 p input (x ; y ) . an internal fault with “0” and other transient disturbances with i i i=1 “1”. Hence, it governs the working of the trip/restrain function f (x) = f (x) +  w (x) p p1 p p by blocking all other power system transients but an internal (10) fault. The level-2 classifier (GBC-2) does further analysis of w = arg min L(y ; f (x ) + w(x )) p i p1 i i the power system events in case the output of GBC-1 is “1”. i=1 The GBC-2 can identify the transient disturbance responsible The minimization problem is solved by taking the negative for the mal-operation of the conventional differential relay gradient of the negative multinomial log-likelihood loss func- (GBC-1 is “1” & Operate relay is “1”). The level-3 classifier tion, L for mutually exclusive classes. (GBC-3) locates the faulty transformer unit (PT, ISPAR series, and ISPAR exciting) if the output of GBC-1 is “0”. The level-4 f (x) = f (x)  r L(y ; f (x )) (11) p p1 p f i p1 i classifiers (GBC-4, GBC-5, and GBC-6) further identifies the i=1 internal faults in the ISPAR exciting, the ISPAR series and the PT. GBC uses shrinkage which scales the contribution of the weak learners by the learning rate and sub-sampling of the training data (stochastic gradient boosting) for regularization. The im- IV. R ESULTS AND DISCUSSION portant hyperparameters of the different GBC classifiers used are the results of grid search on no of estimators = [5000, 1.5 cycles of 3-phase differential currents are used for detec- 7000, 10000, 12000, 15000], max depth = [3,5,7,10,15], and tion, and 3 cycles are used for localization and identification of learning rate = [0.01, 0.05, 0.07, 0.1]. transients from the time of their inception. Thus, at a sampling rate of 10 kHz, 167 data samples per cycle are analyzed. Several factors influence the classification accuracy of an D. Proposed scheme algorithm. Cross-validation and grid search helps in using the data effectively and training the classifier with the best combination of hyperparameters. The data is split randomly into training and test set in a 4:1 ratio. To avoid the problem of overfitting and underfitting of the estimator on the test set, cross-validation is applied on the training data and the hyper- parameters are optimized using grid search over a parameter grid. Grid search comprehensively searches for the parameters over the subset of the hyperparameter space of the estimator. The performance of the selected hyperparameters is then tested on the unseen test data that is not used during the training process. Ten-fold stratified cross-validation (rearrangement of the training data in ten folds such that each fold represents every class well) is used to select the model as it is better both in terms of bias and variance [45]. TABLE VIII: Internal fault detection with CDF & GBC-1 Fault/Disturbances Total TP FN FP Internal Faults 2107 2105 2 0 Disturbances 1852 1852 0 2 Fig. 7: Proposed transient detection and classification algorithm TABLE IX: Comparison of performances with and without CDF (b) Internal fault detection The block diagram description of the CDF and GBC based (a) Internal fault detection with CDF without CDF proposed internal fault detection, fault localization, and tran- sient disturbance classification algorithm is shown in Fig.7. Classifier   Classifier The change detector discovers the change in the 3-phase GBC-1 99.95 % GBC 98.5% differential currents (I -I ) if the CDF index in any phase P S DT 99.5% DT 95.3% is greater than the threshold, th = 0.05. 1/2 cycle pre-transient SVM 99.7% SVM 89.2% RFC 99.9% RFC 94.6% and 1 cycle post-transient differential current samples are used to detect an internal fault and 3 post-transient cycles are IEEE SYSTEMS JOURNAL 8 TABLE X: Localization of faulty transformer unit A. Internal fault detection (b) Comparison of perfor- (a) Localization with GBC-3 mances Transformer Total TP FN FP Classifier The detection of internal faults is performed using GBC in ISPAR Exciting 2937 2899 38 8 GBC-3 99.5% ISPAR Series 7402 7383 19 17 two ways, one with the CDF and the other without it. Most DT 98.6% PT 7287 7287 0 32 authors haven’t considered using some technique to detect SVM 88.9% RFC 98.7% the change in differential currents in case a transient occurs. Rather they fixed the time of occurrence of the transient events and used this specified inception time to store the B. Identification of faulty unit & internal fault type disturbance and fault data. However, faults and disturbances are highly unpredictable in time. In this paper, both methods, Once it is confirmed that an internal fault has been detected, one considering a specified time (without the use of CDF) the locations of those internal faults are determined. 3 cycles and the other with CDF are used to register the data after of post-fault differential current samples are used to locate the inception of transients. The CDF detects the change and the faulty transformer unit (PT or ISPAR Exciting or ISPAR registers 1/2 cycle of pre-transient and 1 cycle of post-transient Series) and determine the type of fault. GBC, SVM, DT, and samples. This 1.5 cycle (250 samples) is used to extract the RFC are used to identify the faulty unit and further locate relevant features which are then fed to GBC, SVM, DT, and and pinpoint the type of fault in the PT and ISPAR units. (TP +TN) RFC classifiers. Accuracy is used as the typical metrics to and accuracy computed as  = , are used as (TP +FN+TN+FP ) measure the performance of the classifiers. But, it is biased to the metrics to measure the performance of the estimators for data imbalance. Since, the classes are imbalanced, balanced localization of faulty unit and identification of internal fault accuracy which is defined as mean of the accuracies obtained type, respectively. 1 TP TN on all classes and computed as   =  [ + ] 2 (TP +FN) (TN+FP ) TABLE XI: Comparison of identification performances of internal fault type for binary classes is used to compute the performance mea- (a) Exciting unit (b) Series unit (c) PT sure where, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false Classifier  Classifier  Classifier negative [46]. GBC-4 99.2% GBC-5 98.0% GBC-6 99.2% DT 98.6% DT 94.7% DT 98.9% The performance of the fault detection scheme composed of SVM 94.8% SVM 90.7% SVM 94.0% the GBC-1 and CDF is shown in Table VIII.   of 99.95% is RFC 98.9% RFC 96.9% RFC 97.8% obtained on a training data of 15,835, testing data of 3959, 1) Localization of faulty unit: To locate the faulty trans- and hyperparameters: learning rate = 0.1, max depth = 5, former unit 70,502 fault cases are trained and 17,626 cases and no of estimators = 7000. The performance of the four classifiers with CDF is shown in Table IXa. One cycle of post- are tested. 18 features are used to train the classifiers (Table fault data is used for training the classifiers for fault detection XVI). GBC-3 with hyperparameters: no of estimators = 5000, without the CDF.   of 98.52% is obtained with GBC for learning rate = 0.07, and max depth = 10 gives   of 99.48%. max depth = 7, no of estimators = 5000, and learning rate Table Xa shows the localization results using GBC-3 and Table = 0.07. The balanced scores of the four classifiers trained on Xb compares the   of the four different classifiers. 80,870 cases and tested on 20,218 cases are shown in Table 2) Identification of internal fault type: The internal faults IXb. 18 features from the 3-phase differential currents (Table in the ISPAR series, ISPAR exciting and the PT are further XVI) are used as the input to the classifiers for training the classified into w -g, w -g, w -g, w -w -g, w -w -g, w -w - a b c a b a c b c fault detection models with and without CDF. GBC with CDF g, w -w , w -w , w -w , turn-to-turn, winding-to-winding, a b a c b c performed better than without CDF (Table IX) as the CDF and very rare w -w -w and w -w -w -g faults. 21 features a b c a b c filtered out the cases where there is no appreciable change in from 3 cycles of the 3-phase differential currents are used differential currents although a transient event occurred. It is as the input to the estimators (Table XVI). Tables XIa, XIb, noticed that the CDF could detect the change in differential and XIc compare the performances of GBC, RFC, DT, and currents in all internal fault cases except turn-to-turn faults SVM classifiers for ISPAR exciting, ISPAR series, and the PT with Rf = 10 , LTC = 0.2, and percentage of winding respectively. shorted = 20%. Also, it detected the change for all transient To identify the internal faults in ISPAR exciting 14,688 disturbances except sympathetic inrush cases for switching fault cases are used to train and test the four classifiers. angles from 120 to 330 . On exploring the data it is observed GBC-4 trained with hyperparameters of max depth = 5, that there is almost no change in the differential currents for no of estimators = 7000, and learning rate = 0.1 achieved these instances. The w-g faults for LTC = 0.2, and percentage the best accuracy of 99.18%. For the identification of in- of winding shorted = 20% which needs higher sensitivity were ternal faults in ISPAR series 36,720 cases are used to train detected. It proves the dependability of the scheme for ground and test the classifiers. GBC-5 trained with learning rate = faults near neutral of wye grounded transformers (PT and 0.05, max depth = 7, and no of estimators = 5000 gives an ISPAR exciting) which is again a challenge for conventional accuracy of 98.0%. Similarly, for PT the classifiers are trained differential relays [7]. & tested on 36,720 fault cases. GBC-6 achieved the best ac- IEEE SYSTEMS JOURNAL 9 curacy of 99.2% obtained by training the hyperparameters on learning rate = 0.05, no of estimators = 5000, and max depth = 5. The identification accuracy obtained in the ISPAR series is lower than in PT and ISPAR exciting because the secondary side of ISPAR series is delta connected. Hence, one type of fault on the primary side confuses with another type on the secondary side. C. Identification of disturbance type The various disturbances: magnetizing inrush, sympathetic inrush, ferroresonance, external faults with CT saturation, capacitor switching, and non-linear load switching are also classified using 3 cycles of post-transient samples after they are differentiated as no-fault by the fault detection scheme. Fig. 8: Kernel Density Estimate plots showing the probability distribution of 15 features are used as input to the classifiers in this case the 5 selected features for the 6 transient disturbances in (a) phase a, (b) phase (Table XVI). It’s always useful to know the probabilities of the b, and (c) phase c. input features taking on various real values. ParzenRosenblatt TABLE XII: Low and high-order statistics of the 5 selected features window method is used to estimate the underlying probability density of the 5 features for the six different disturbances in (a) Magnetizing inrush phases a, b, and c. Fig.8 shows the kernel density estima- autoregressive FT linear change change tion plots for the chosen features. Gaussian is used as the coefficient coefficient trend quantile 1 quantile 2 2 2 2 2 2 kernel function to approximate the univariate features with   ~    ~    ~    ~    ~ 3 3 3 3 3 ph a -4.6 5.9 .74 -.35 1e3 2e6 1.3 .74 2.6 14 1.9 3.9 .02 9e-4 1.9 2.9 .04 .008 2.5 5.1 a bandwidth of 0.2 for autoregressive coefficient, and FT ph b -4.7 6.2 .65 -.13 502 3e5 1.0 -.06 2.6 13 1.8 3.2 .01 8e-5 1.5 1.06 .04 .01 2.5 5.7 ph c 3.3 .72 -1.1 .33 33 1e3 1.1 .25 .19 .05 1.0 -.35 .02 5e-4 .90 -.45 .08 .009 .88 -.84 coefficient and a bandwidth of 0.01 for aggregate linear trend, and avg. change quantile 1 and 2. It is observed that probability (b) CT saturation during external faults density functions of autoregressive and FT coefficients are a autoregressive FT linear change change mixture of multiple normal distributions with varying standard coefficient coefficient trend quantile 1 quantile 2 deviation and mean whereas linear trend, and change quantiles 2 2 2 2 2 ~    ~    ~    ~    ~ 3 3 3 3 3 are unimodal with means near zero and smaller standard ph a -2.3 3.1 -1.6 1.97 190 4e4 1.6 2.5 .47 .23 1.64 2.9 .04 6e-4 1 .98 .003 9e-6 1.8 3.6 ph b -2.2 3 -1.6 1.95 261 2e4 .81 .35 .47 .24 1.55 2.3 7e-4 9e-7 3.5 19 .04 5e-4 .6 -.42 deviations. Table XIIa and XIIb shows the values of mean (), ph c 2.5 .4 1.1 .27 16 150 1.7 2.9 .02 5e-4 1.5 2.4 .003 8e-6 1.7 4.7 .04 4e-4 .67 -.14 variance ( ), skewness ( ~ ), and kurtosis () of the 5 features for magnetizing inrush and CT saturation during external faults respectively in phases a, b and c. Because of space limitations, to verify how effectively the GBC discriminates the internal the feature statistics of only these two transients are shown. faults in the PT and the ISPAR. The table XV shows the Furthermore, to visualize the 15-dimensional input data in classification errors with learning rate of 0.05, max depth = a 2-dimensional plane, the T-distributed Stochastic Neighbor 9, and no of estimators = 5000. The balanced accuracy of Embedding dimensionality reduction technique has been used 99.9% shows that the GBC is capable of distinguishing these which preserves much of the significant structure in the high- faults even in an interconnected network. dimensional data while mapping in the 2-dimension [47]. Fig.9 shows the clusters of similar transients (300 instances each) E. Performances on balanced and imbalanced data and also the relationships between different group of transients Machine learning algorithms are more reliable when they as a scatter plot. operate on a balanced dataset. To adjust the data distribution of The Table XIII shows the classification results using GBC- classes and remove class imbalance, under-sampling of major- 2. Table XIV compares the results of GBC with RFC, DT, and SVM. The classifiers are trained on 10,368 cases and tested on 2592 cases.   of 99.28% is obtained with GBC-2 having hyperparameters: no of estimators = 5000, learning rate = 0.7, and max depth = 3. D. Discriminate faults in PT & ISPAR The PAR controls the power flow through a line and when connected with a PT, it reduces the magnitude of the differential currents and their harmonic contents and alters the wave shapes due to the additional phase shift for the PT in case of external faults. With internal faults, such changes in the differential currents are lesser. 70,502 internal fault Fig. 9: 2D scatterplot of the input features for transient disturbances cases are trained and 17,626 cases are tested on 18 features IEEE SYSTEMS JOURNAL 10 TABLE XIII: Identification of transient disturbances with GBC-2 TABLE XVII: Performance for 400-MVA & Y connection Disturbances Total TP FN FP Fault/ Faults/ Total TP FN Disturbances Abnormalities (%) Magnetizing inrush 365 357 8 0 Sympathetic inrush 336 336 0 8 (a) ph & g, T-T, W-W faults (PT) 1200 1200 0 100 Internal Capacitor switching 73 72 1 1 (b) ph & g, T-T, W-W faults (Series) 1200 1200 0 100 faults Ferroresonance 133 132 1 0 (c) ph & g, T-T, W-W faults (Exciting) 672 672 0 100 (3072) Load switching 69 69 0 0 (d) Total = (a)+ (b)+ (c) 3072 3072 0 100 External faults 1616 1615 1 2 (e) Capacitive switching 60 51 9 85 TABLE XIV: Performance (f) External faults with CT saturation 528 525 3 99.4 Other comparison of identification (g) Ferroresonance 24 24 0 100 TABLE XV: Misclassification disturbances of transient disturbances (h) Magnetizing inrush 60 60 0 100 between ISPAR & PT (876) (i) Load switching 144 144 0 100 (j) Sympathetic inrush 60 60 0 100 Classifier Faults Total TP FN FP (k) Total = (e)+ (f)+ (g)+ (h)+ (i)+ (j) 876 864 12 98.6 GBC-2 99.28% PT 7262 7262 0 13 DT 98.09% ISPAR 10364 10351 13 0 Total (3948) Total faults and disturbances = (d)+(k) 3948 3936 12 99.3 SVM 98.23% RFC 98.89% TABLE XVIII: Effect of Noise Fault/ Number Predicted class Accuracy SNR (dB) ity classes and over-sampling of minority classes is performed. Disturbances of cases (%) Faults Disturbances Synthetic Minority Over-Sampling Technique (SMOTE) [48] 1 1010 1001 9 99.2 is used to create minority synthetic data considering k-nearest Internal 30 1035 1005 30 97.1 neighbours and NearMiss algorithm is used for under-sampling faults 20 1008 934 74 92.7 the majority classes avoiding information loss. The table 10 984 891 93 90.4 XVI shows the balanced accuracy/ accuracy for detecting 1 990 3 987 99.7 the internal faults, identifying the faulty units and type of Other 30 965 24 941 97.6 disturbances 20 992 26 966 97.4 faults in those units, and identifying the disturbances with and 10 1016 28 988 97.2 without using SMOTE and NearMiss. It is observed that the accuracies obtained with SMOTE and NearMiss algorithm for the different classification tasks are similar to those obtained is compatible with the accuracy obtained when trained and by training the GBCs without them. Table XVI also gives tested at 500 MVA and YY connection. the information about the time and frequency domain features (fF g ) that has been used to train the different GBC i=1 classifiers for the different classification tasks. G. Effect of Signal Noise & CT saturation TABLE XVI: Input features and performance of different GBC classifiers with In order to analyse the effect of noise in the differential cur- and without SMOTE analysis rents on the proposed fault detection scheme white Gaussian noise of different levels measured in terms of Signal-to-Noise- n   using Classification task F1 F2 F3 F4 F5 Fi3 (%) SMOTE ratio (SNR) are added to the training and testing cases for fault detection [17], [19]–[21]. Table XVIII shows the accuracy detect faults 2 1 1 1 1 18 99.9 99.9 locate faulty units 2 2 2 - - 18 99.5 99.6 of the GBC for different levels of noise on 5000 cases of identify faults (series) 3 1 2 1 - 21 98.0 98.2 internal faults and other disturbances each. It is observed that identify faults(exciting) 3 2 2 - - 21 99.2 99.1 as the level of noise increases the  of the classifier dips, identify faults (PT) 3 2 2 - - 21 99.2 99.1 90:4+97:2 identify transients 2 1 1 - 1 15 99.3 99.4 but still always above 93.8% ( ). The  changes from 99.4% to 93.8% as the SNR is varied from 1 to 10dB. It is also observed from the table that the misclassification of internal faults increases as the SNR is decreased whereas the F. Effect of different rating and transformer connections misclassifications are nearly the same for other disturbances as It is not necessary to train the fault detection scheme for SNR is decreased from 30dB to 10dB. Moreover, to examine different rating and connection of the PTs, rating of ISPAR, the effect of CT saturation the secondary side impedance and variation in other parameters. In order to validate the (burden and CT secondary impedance) which has the major effectiveness of the proposed scheme with variation in differ- influence over the level of saturation is changed.  of 99.5% ent system parameters, new internal faults and other transient is obtained with GBC on 5000 cases of internal faults and cases are simulated again with 400 MVA, Y connected PTs other disturbances each. Fig.10 shows the 3-phase differential and 400 MVA ISPARs. The fault resistance, LTC, fault type, currents with CT saturation for faults in T1 and ISPAR1. and fault inception time are altered to generate the internal fault cases and switching time, firing angle, LTC, etc. are H. Execution Time altered to generate the transient cases to test the same GBC- 1 model trained using 500 MVA, YY connected PTs and The execution time-averaged over 10 runs for the feature 500 MVA ISPARs. It is observed from Table XVII that the extraction, training, and testing of the GBC classifiers for proposed scheme gives a balanced accuracy of 99.3% which detection of internal faults, identifying the faulty unit and type IEEE SYSTEMS JOURNAL 11 sympathetic inrush, ferroresonance, external faults with CT saturation, capacitor switching, and non-linear load switching transients. A change detector is used to detect the change in the 3-phase differential currents in case a transient event occurs and registers the current samples for detection and classification purposes. Five most relevant time and frequency Fig. 10: 3-phase differential currents with CT saturation for (a) w -w -g fault domain features, selected from the differential currents on the in transformer primary and (b) w -g fault in series primary basis of Information Gain are used to train the DT, RFC, GBC, and SVM classifiers. The fault detection scheme comprising of the CDF and GBC gives an accuracy of 99.95% on 19,794 of fault, and identifying the transient using one CPU core is transient cases obtained by varying different parameters for the reported using the in-built library in python (Table XIX). The internal faults and other transient disturbances confirming its fault/no-fault decision includes the time to compute the feature dependability for internal faults and security against transient and testing a single instance with GBC-1 which adds to 8.7ms disturbances. Once an internal fault is detected and a trip signal with the CDF. Thus, the proposed scheme has a processing is issued using 1.5 cycles, the faulty transformer unit (PT, IS- time of 25.37ms (16.67+1.7+7) or  1 cycle to detect a fault. PAR series, or ISPAR exciting unit) and type of internal faults Considering that these computations can be further optimized in those units are also identified in 3 cycles. Furthermore, the for example by converting Python and MATLAB code to a type of transient disturbance is determined in case the fault compiled low-level language such as C, the fault detection and detection scheme detects a transient other than internal faults. localization, and transient identification schemes are suitable The validity of the scheme is also established for different for future real-time implementation. The DT, SVM, RFC, rating and connection of the transformers, CT saturation, and and GBC classifiers are built in Python 3.7 using Scikit-learn SNR ratio of 30dB to 10dB in the differential currents. The framework [49] while the CDF is implemented in MATLAB proposed fault detection strategy can work together with a 2017. The pre-processing of the data is done in Python and conventional differential relay offering supervisory control MATLAB. All PSCAD simulations are carried out on Intel over its operation and thus avoid false tripping. The transient Core i7-6560U CPU @ 2.20 GHz and 8 GB RAM and the detection and identification accuracies obtained are among the classifiers are run on Intel Core i7-8700 CPU @ 3.20 GHz best even when compared with results from works on isolated and 64 GB RAM. and simple networks. TABLE XIX: Execution time of the GBC classifiers (seconds) APPENDIX Training Testing Training Testing time Feature Classification task Fortran script for two-winding transformer instances instances time extraction One All time 1. NW = 4 18. L2l = Lk1/2*fb detect faults 80870 20218 1506 0.0024 1.1 0.007 2. I 2 = I 1 = I 19. L3l = Lk2/2*fc m m m detect faults with CDF 15835 3959 19 0.0017 .061 0.007 3. fa = fault1 0:01 20. L4l = Lk2/2*fd locate faulty units 70502 17626 1232 0.0049 1.2 0.0088 4. fb = 1:0 fa 21. L1m = (v1/(w*I 1*i1))*fa fa identify faults (series) 29376 7344 1341 0.016 2.1 0.0111 5. fc = fault2 0:01 22. L2m = (v1/(w*I 1*i1))*fb fb identify faults(exciting) 11750 2938 357 0.013 0.48 0.0108 6. fd = 1:0 fc 23. L3m = (v2/(w*I 2*i2))*fc fc identify faults (PT) 29376 7344 2712 0.026 5.1 0.0108 7. i1 = MVA/v1 24. L4m = (v2/(w*I 2*i2))*fd fd identify transients 10368 2592 72.3 0.004 0.09 0.0065 8. i2 = MVA/v2 25. Lx = L1l + L1m 9. z1 = v1/i1 26. Ly = L2l + L2m 10. z2 = v2/i2 27. Lz = L3l + L3m This work distinguishes faults from the six transients (  = 11. w = 2*pi*f 28. Lw = L4l + L4m 12. l1 = v1/(w*I 1*i1) 29. Mxy = sqrt(L1m*L2m) 99.95%), locates the faulty unit (  = 99.5%), identifies the m 13. l2 = v2/(w*I 2*i2) 30. Mxz = sqrt(L1m*L3m) fault type (  99%) and six other transients (  = 99.3%) 14. Lk1 = Xl*z1/w 31. Mxw = sqrt(L1m*L4m) for the ISPAR and PT in an interconnected system, whereas 15. Lk2 = Xl*z2/w 32. Myz = sqrt(L2m*L3m) the publications [22] [23] in the literature focused only on 16. tr = v1/v2 33. Myw = sqrt(L2m*L4m) 17. L1l = Lk1/2*fa 34. Mzw = sqrt(L3m*L4m) the ISPAR. In [22], only the internal faults in ISPAR were identified and in [23], the internal faults were differentiated from magnetizing inrush using WT and then the internal faults REFERENCES were identified. In addition to its broader functionality, the [1] R. Hamilton, “Analysis of transformer inrush current and comparison of current work improves the accuracy from an average of 98.76 harmonic restraint methods in transformer protection,” IEEE Transac- [22] and 97.7% [23] to 99.2%. tions on Industry Applications, vol. 49, no. 4, July 2013. [2] Pei Liu, O. P. Malik, Deshu Chen, G. S. Hope, and Yong Guo, “Improved operation of differential protection of power transformers for internal V. CONCLUSION faults,” IEEE Transactions on Power Delivery, vol. 7, no. 4, Oct 1992. [3] T. S. Sidhu, M. S. Sachdev, H. C. Wood, and M. Nagpal, “Design, In this paper, the task of discrimination of internal faults and implementation and testing of a microprocessor-based high-speed relay other transient disturbances in a 5-bus interconnected power for detecting transformer winding faults,” IEEE Transactions on Power Delivery, vol. 7, no. 1, pp. 108–117, Jan 1992. system for PTs and PARs is presented. The internal faults [4] D. Bi, Y. Sun, D. Li, G. Yu, X. Wang, and W. Wang, “Analysis on including turn-to-turn and winding-to-winding faults in the mal-operation of differential protection caused by sympathetic inrush,” ISPAR and the PT are distinguished from magnetizing inrush, Autom. Elect. Power Syst., vol. 31, pp. 36–40, Nov 2007. IEEE SYSTEMS JOURNAL 12 [5] M. Stanbury and Z. Djekic, “The impact of current-transformer satura- [29] G. B. Kumbhar and S. V. Kulkarni, “Analysis of sympathetic inrush tion on transformer differential protection,” IEEE Transactions on Power phenomena in transformers using coupled field-circuit approach,” in Delivery, vol. 30, no. 3, pp. 1278–1287, June 2015. 2007 IEEE Power Engineering Society General Meeting, June 2007. [6] M. A. Ibrahim and F. P. Stacom, “Phase angle regulating transformer [30] H. Weng and X. Lin, “Studies on the unusual maloperation of trans- protection,” IEEE Transactions on Power Delivery, vol. 9, no. 1, pp. former differential protection during the nonlinear load switch-in,” IEEE 394–404, Jan 1994. Transactions on Power Delivery, vol. 24, no. 4, Oct 2009. [7] “IEEE Guide for the Application of Protective Relaying for Phase- [31] E. Koley, S. K. Shukla, S. Ghosh, and D. K. Mohanta, “Protection Shifting Transformers,” IEEE Std C37.245-2018, pp. 1–71, May 2019. scheme for power transmission lines based on svm and ann consid- [8] P. Mao and R. Aggarwal, “A novel approach to the classification ering the presence of non-linear loads,” IET Generation, Transmission of the transient phenomena in power transformers using combined Distribution, vol. 11, no. 9, pp. 2333–2341, 2017. wavelet transform and neural network,” Power Engineering Review, [32] T. Patcharoen and A. Ngaopitakkul, “Transient inrush current detection IEEE, vol. 21, pp. 70–70, 08 2001. and classification in 230 kv shunt capacitor bank switching under various [9] S. Jazebi, B. Vahidi, and M. Jannati, “A novel application of wavelet transient-mitigation methods based on discrete wavelet transform,” IET based svm to transient phenomena identification of power transformers,” Generation, Transmission Distribution, vol. 12, no. 15, 2018. Energy Conversion and Management, vol. 52, no. 2, 2011. [33] M. K. Siahpoosh, D. Dorrell, and L. Li, “Ferroresonance assessment in [10] A. M. Shah and B. R. Bhalja, “Discrimination between internal faults a case study wind farm with 8 units of 2 mva dfig wind turbines,” in and other disturbances in transformer using the support vector machine- 2017 20th International Conference on Electrical Machines and Systems based protection scheme,” IEEE Transactions on Power Delivery, (ICEMS), Aug 2017, pp. 1–5. vol. 28, no. 3, pp. 1508–1515, July 2013. [34] S. Rezaei, “Impact of ferroresonance on protective relays in manitoba [11] O. Ozgonenel and S. Karagol, “Power transformer protection based on hydro 230 kv electrical network,” in 2015 IEEE 15th International decision tree approach,” IET Electric Power Applications, vol. 8, no. 7, Conference on Environment and Electrical Engineering, June 2015. pp. 251–256, August 2014. [35] M. Hajizadeh, I. Safinejad, and N. Amirshekari, “Study and comparison [12] S. Samantaray and P. Dash, “Decision tree based discrimination between of the effect of conventional, low losses and amorphous transformers on inrush currents and internal faults in power transformer,” International the ferroresonance occurrence in electric distribution networks,” CIRED Journal of Electrical Power & Energy Systems, vol. 33, no. 4, 2011. - Open Access Proceedings Journal, vol. 2017, no. 1, pp. 865–869, 2017. [13] Yong Sheng and S. M. Rovnyak, “Decision trees and wavelet analysis for [36] O. Dharmapandit, R. K. Patnaik, and P. K. Dash, “A fast time-frequency power transformer protection,” IEEE Transactions on Power Delivery, response based differential spectral energy protection of ac microgrids vol. 17, no. 2, pp. 429–433, April 2002. including fault location,” Protection and Control of Modern Power [14] M. Tripathy, R. P. Maheshwari, and H. K. Verma, “Probabilistic neural- Systems, vol. 2, no. 1, p. 30, Aug 2017. network-based protection of power transformer,” IET Electric Power [37] M. Christ, N. Braun, J. Neuffer, and A. W. Kempa-Liehr, “Time series Applications, vol. 1, no. 5, pp. 793–798, Sep. 2007. feature extraction on basis of scalable hypothesis tests (tsfresh a python [15] A. M. Shah and B. R. Bhalja, “Fault discrimination scheme for power package),” Neurocomputing, vol. 307, pp. 72 – 77, 2018. transformer using random forest technique,” IET Generation, Transmis- [38] P. Welch, “The use of fast fourier transform for the estimation of power sion Distribution, vol. 10, no. 6, pp. 1431–1439, 2016. spectra: A method based on time averaging over short, modified peri- [16] R. Kumar, B. Singh, D. T. Shahani, A. Chandra, and K. Al-Haddad, odograms,” IEEE Transactions on Audio and Electroacoustics, vol. 15, “Recognition of power-quality disturbances using s-transform-based ann no. 2, pp. 70–73, June 1967. classifier and rule-based decision tree,” IEEE Transactions on Industry [39] L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, Classification Applications, vol. 51, no. 2, pp. 1249–1258, March 2015. and regression trees. CRC Press, 1984. [17] N. Perera and A. D. Rajapakse, “Recognition of fault transients using [40] J. R. Quinlan, C4.5: Programs for Machine Learning. San Francisco, a probabilistic neural-network classifier,” IEEE Transactions on Power CA, USA: Morgan Kaufmann Publishers Inc., 1993. Delivery, vol. 26, no. 1, pp. 410–419, Jan 2011. [41] L. Breiman, “Random forests,” Machine Learning, vol. 45, pp. 5–32, [18] S. Santoso, E. J. Powers, W. M. Grady, and A. C. Parsons, “Power Oct 2001. quality disturbance waveform recognition using wavelet-based neural [42] G. Louppe, “Understanding random forests: From theory to practice,” classifier. i. theoretical foundation,” IEEE Transactions on Power Deliv- Ph.D. dissertation, University of Lige, Belgium, Oct 2014. ery, vol. 15, no. 1, pp. 222–228, Jan 2000. [43] J. H. Friedman, “Greedy function approximation: A gradient boosting [19] T. Zhong, S. Zhang, G. Cai, Y. Li, B. Yang, and Y. Chen, “Power machine.” Ann. Statist., vol. 29, no. 5, pp. 1189–1232, 10 2001. quality disturbance recognition based on multiresolution s-transform and [44] L. Mason, J. Baxter, P. Bartlett, and M. Frean, “Boosting algorithms as decision tree,” IEEE Access, vol. 7, pp. 88 380–88 392, 2019. gradient descent,” in Proceedings of the 12th International Conference [20] U. Singh and S. N. Singh, “Optimal feature selection via nsga-ii on Neural Information Processing Systems, ser. NIPS’99. Cambridge, for power quality disturbances classification,” IEEE Transactions on MA, USA: MIT Press, 1999, pp. 512–518. Industrial Informatics, vol. 14, no. 7, pp. 2994–3002, July 2018. [45] R. Kohavi, “A study of cross-validation and bootstrap for accuracy [21] P. K. Ray, S. R. Mohanty, N. Kishor, and J. P. S. Catalo, “Optimal feature estimation and model selection,” in Proc. 14th Int. Joint Conf. Artif. and decision tree-based classification of power quality disturbances Intell., 1995, p. 11371143. in distributed generation systems,” IEEE Transactions on Sustainable [46] K. H. Brodersen, C. S. Ong, K. E. Stephan, and J. M. Buhmann, Energy, vol. 5, no. 1, pp. 200–208, Jan 2014. “The balanced accuracy and its posterior distribution,” in 2010 20th [22] S. K. Bhasker, P. K. Bera, V. Kumar, and M. Tripathy, “Differential International Conference on Pattern Recognition, Aug 2010. protection of indirect symmetrical phase shift transformer and internal [47] L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” faults classification using wavelet and ann,” in TENCON 2015 - 2015 Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008. IEEE Region 10 Conference, Nov 2015, pp. 1–6. [48] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: [23] P. K. Bera, R. Kumar, and C. Isik, “Identification of internal faults in Synthetic minority over-sampling technique,” J. Artif. Int. Res., vol. 16, indirect symmetrical phase shift transformers using ensemble learning,” no. 1, p. 321357, Jun. 2002. in 2018 IEEE International Symposium on Signal Processing and [49] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal Information Technology (ISSPIT), Dec 2018, pp. 1–6. of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [24] P. K. Bera, C. Isik, and V. Kumar, “Transients and faults in power transformers and phase angle regulators (dataset),” 2020. [Online]. Available: http://dx.doi.org/10.21227/1d1w-q940 [25] J. McIver, “Phase shifting transformers principles, design aspects and operation,” Siemens Energy, Inc, Mar 2015. [26] S. Kulkarni and S. Khaparde, Transformer Engineering. Boca Raton: CRC Press, 2013. [27] H. Bronzeado and R. Yacamini, “Phenomenon of sympathetic interaction between transformers caused by inrush transients,” IEE Proceedings - Science, Measurement and Technology, vol. 142, no. 4, July 1995. [28] T. Zheng, J. Gu, S. F. Huang, F. Guo, and V. Terzija, “A new algorithm to avoid maloperation of transformer differential protection in substations with an inner bridge connection,” IEEE Transactions on Power Delivery, vol. 27, no. 3, pp. 1178–1185, July 2012. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Electrical Engineering and Systems Science arXiv (Cornell University)

Discrimination of Internal Faults and Other Transients in an Interconnected System with Power Transformers and Phase Angle Regulators

Loading next page...
 
/lp/arxiv-cornell-university/discrimination-of-internal-faults-and-other-transients-in-an-g8DzxsFfsr
ISSN
1932-8184
eISSN
ARCH-3348
DOI
10.1109/JSYST.2020.3009203
Publisher site
See Article on Publisher Site

Abstract

IEEE SYSTEMS JOURNAL 1 Discrimination of Internal Faults and Other Transients in an Interconnected System with Power Transformers and Phase Angle Regulators Pallav Kumar Bera, Can Isik, Senior Member, IEEE, and Vajendra Kumar Abstract—This study solves the problem of accurate detection to magnetizing and sympathetic inrush, and CT saturation of internal faults and classification of transients in a 5-bus during external faults are the major problems associated with interconnected system for Phase Angle Regulators (PARs) and differential protection. Second-harmonic restraint method is Power Transformers (PTs). The analysis prevents mal-operation extensively used to distinguish internal faults from magnetiz- of differential relays in case of transients other than faults which ing inrush since more second-harmonic component exists in include magnetizing inrush, sympathetic inrush, external faults with Current Transformer (CT) saturation, capacitor switch- inrush currents than in internal faults [1]. However, higher ing, non-linear load switching, and ferroresonance. A gradient second-harmonics are generated during internal faults with boosting classifier (GBC) is used to distinguish the internal CT saturation, presence of shunt capacitance, or because of faults from the transient disturbances based on 1.5 cycles of the distributed capacitance of EHV lines [2]. In addition, 3-phase differential currents registered by a change detector. the second-harmonic content in inrush currents has reduced After the detection of an internal fault, GBCs are used to locate the faulty unit (PT, PAR series, or exciting unit) and in modern transformers with soft core material [3]. Hence, identify the type of fault. In case a transient disturbance is several cases of mal-operation of conventional relays in dis- detected, another GBC classifies them into the six disturbances. tinguishing faults and inrush have been reported [4]. CT Five most relevant frequency and time domain features obtained saturation during external faults may also cause false trips due using Information Gain are used to train and test the classifiers. to the inefficient setting of commonly used dual-slope biased The proposed algorithm distinguishes the internal faults from the other transients with a balanced accuracy () of 99.95%. differential relays [5]. The faulty transformer unit is located with  of 99.5% and Phase Angle Regulators or Phase Shifters or Phase Shift the different transient disturbances are identified with  of Transformers are a special class of transformers used to 99.3%. Moreover, the reliability of the scheme is verified for control real power flow in parallel transmission lines. They different rating and connection of the transformers involved, CT ensure the system reliability and allow easier integration of saturation, and noise level in the signals. These GBC classifiers can work together with a conventional differential relay and new generations with the grid. By regulating the phase angle offer a supervisory control over its operation. PSCAD/EMTDC between the sending and receiving ends they prevent over- software is used for simulation of the transients and to develop loading of a line and re-routes power via another line. PARs the two and three-winding transformer models for creating the can be categorized on the basis of the number of cores and internal faults including inter-turn and inter-winding faults. magnitude of sending end voltage with respect to the receiving Index Terms—Phase Angle Regulators, Power Transformers, end. Indirect Symmetrical Phase Angle Regulators (ISPAR) Fault detection, Transients, Gradient Boosting Classifier having the same sending and receiving end voltages with two transformer units, namely, series and exciting (Fig.1(b)), has I. INTRODUCTION been chosen as one of the subjects (other being the PTs) in this study because of their popularity and security against higher OWER Transformers are an integral part of an electrical voltage levels as the load tap changer (LTC) is not exposed grid and their protection is vital for reliable and stable to system disturbances. The exciting unit is responsible for operation of the power system. An important requirement creating the required phase difference to regulate the power of the protection system is the faithful discrimination of which can be controlled by the LTC connections and an faults from other transients. Differential protection has been advance-retard-switch located on its secondary winding [6]. the primary protection scheme in transformers because of The modified real power flow in a transmission line with a its inherent selectivity and sensitivity. Mal-operations due PAR is given by Manuscript received February 20, 2020; revised May 23, 2020; accepted V  V July 10, 2020.(Corresponding author: Pallav Kumar Bera.) S L P =  sin( + ) (1) Pallav Kumar Bera & Can Isik are with the Department of Electrical X + X line PAR Engineering and Computer Science, Syracuse University, NY, USA (e-mail: pkbera@syr.edu; cisik@syr.edu). where, V is source voltage, V is load voltage;  is the phase S L Vajendra Kumar was with Indian Institute of Technology, Roorkee, India angle difference between V and V ; X ; X are the S L line PAR (email: kumarvajendra@gmail.com). 2020 IEEE. Personal use of this material is permitted. Permission from transmission line and PAR reactance respectively; and is the IEEE must be obtained for all other uses, in any current or future media, new constraint added which is responsible for controlling the including reprinting/republishing this material for advertising or promotional power flow. The PARs similar to PTs require a fast, sensitive, purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. secure, and dependable protection system. Discriminating ex- arXiv:2004.06003v2 [eess.SP] 2 Aug 2020 IEEE SYSTEMS JOURNAL 2 ternal faults with CT saturation, magnetizing inrush, and other evaluates the effect of noise, CT saturation, and change in transient disturbances from internal faults is a challenge for the transformer rating and connection on the proposed scheme. protection systems of PARs as well. Moreover, methods used Section V concludes the paper. to compensate the phase for differential relays in PTs with a fixed phase shift are not applicable in PARs with variable phase shift [7]. Authors have used different intelligent methods to distin- guish internal faults and magnetizing inrush in PTs in the past decade. A combination of Artificial Neural Network (ANN) and spectral energies of wavelet components was used to discriminate internal faults and inrush in [8]. Support Vector Machines (SVM) and Decision Tree (DT) based trans- former protection were proposed in [9], [10] and [11]–[13] respectively. Probabilistic Neural Network (PNN) has been used to detect different conditions in PT operation in [14]. Random Forest Classifier (RFC) was proposed to discrimi- nate internal faults and inrush in [15]. Works of literature also suggest extensive use of S-Transform, Hyperbolic S- Transform, Wavelet Transform (WT) to detect Power Quality Fig. 1: (a) 5-bus interconnected system with ISPARs, PTs, T-lines, and AC (PQ) transient disturbances and then classify them using DT, sources, (b) Series and exciting transformers in ISPAR SVM, ANN, PNN [16]–[21]. These transient disturbances are caused by variations in load, capacitor switching, charging of transformers, starting of induction machines, use of non-linear II. MODELING AND S IMULATION loads, etc. Contrarily, literature investigating internal faults and inrush in an ISPAR is limited. However, attempts were made in The power network chosen for the simulation of the internal faults and the transient disturbances is based on a proposed [22] where internal faults are distinguished from magnetizing Pumped-storage (efficient form of renewable storage designed inrush using WT and then the internal faults are classified to meet energy needs and reduce emissions by utilizing the using ANN and in [23] where the internal faults in series and energy stored in an upper water body pumped from a lower exciting transformers of the ISPAR are classified using RFC. water body) project in California, USA [25]. Further, the authors have predominately used an isolated and PSCAD/EMTDC is used for the modeling and simulation simple network having a PT [8]–[14] or a PAR [22], [23] to of the transients in the ISPAR and PT in the chosen intercon- support their proposed protection scheme. Also, the transient nected power system. Fig.1(a) shows the single-line diagram disturbances have not been studied rigorously in these works. This paper studies the use of Decision Tree based algo- of the 5-bus interconnected model consisting of the AC source, rithms to discriminate the internal faults and other transient transmission lines, ISPARs, PTs, and 3-phase loads working at disturbances including magnetizing inrush and CT saturation 60Hz. The ISPARs have a rating of 500 MVA, 230kV/230kV, during external faults in a 5-bus interconnected system with with phase angle variations of 25 and the PTs are rated at Phase Angle Regulators and Power Transformers which has 500 MVA, 500kV/230kV. The AC source consists of 9 units not been attempted before. Customized two-winding and three- of 120 MVA, 13.8kV hydro-generators. Two transformers are winding transformers are developed to simulate the internal used in cascade to step up the voltage from 13.8kV to 500kV. faults. A change detector has been used to detect and register 3 ISPARs (ISPAR , ISPAR , and ISPAR ) are connected 1 2 3 the differential currents. Five most relevant time and frequency between bus4 and bus5 through transformers T , T , and T . 1 2 3 domain features have been used to train SVM, RFC, DT, and Only the internal faults in ISPAR and T are studied. 1 1 GBC classifiers to detect, locate and identify the internal faults The three-winding transformer required for the series units and classify six transient disturbances. The proposed scheme of ISPAR and the two-winding transformer required for the is tested on 101,088 transients cases simulated on Power exciting units of ISPAR and PTs for the simulation of various System Computer Aided Design (PSCAD)/ Electromagnetic internal faults including turn-to-turn and primary-to-secondary Transients including DC (EMTDC) by varying various system winding faults are developed in PSCAD/EMTDC with Fortran. parameters. The entire dataset of internal faults and other The self-inductance terms (Li) and the mutual inductance disturbances is made available on IEEE Dataport [24]. terms (Mij) of the 4 4 L matrix (Eq.2) of the single-phase The rest of the paper is organized as follows. Section II two-winding transformer and 6  6 L matrix of the single- illustrates the modeling and simulation of the internal faults phase three-winding transformer are evaluated from primary and other transient disturbances in the power network con- and secondary voltages, the magnetizing component of the taining PTs and ISPARs. Section III comprises the detection no-load excitation current (I ), and the short-circuit tests. of internal faults, feature extraction and selection, and the The modeled components have the provision to change the classifiers used for the detection and identification of tran- saturation characteristics, % of winding shorted and other sients. Section IV includes the results for detection of internal parameters. The Fortran script of the two-winding transformer faults, identification of faults and transient disturbances, and is shown in the Appendix. IEEE SYSTEMS JOURNAL 3 exciting unit. Fig.2(b) shows the differential currents for LTC= 2 3 Lx Mxy Mxz Mxw full, fault inception time=15.0124s, backward phase shift, 6Myx Ly Myz Myw7 L = (2) 4 5 Mzx Mzy Lz Mzw fault resistance=0.01 and % turns shorted=40 in primary of Mwx Mwy Mwz Lw series unit. Fig.2(c) shows the differential currents for LTC= full, fault inception time=15.01518s, forward phase shift, fault The study covers various internal faults in the ISPAR resistance=0.01 and % turns shorted=60 in primary of the PT. and PT, capacitor switching, switching of non-linear loads, magnetizing inrush, sympathetic inrush, external faults with CT saturation, and ferroresonance. In the following para- graphs, these conditions are considered one after the other. The simulation run-time, fault/disturbance inception time, and fault duration time are 15.2s, 15.0s, and 0.05s (3 cycles) respectively in all cases. The multi-run component is used to change the parameter values wherever possible to get the different simulation cases and snapshots of the first simulation runs are used to start the simulation from initialized conditions to reduce the simulation time. A. Internal Faults The internal faults are created in the PT, ISPAR series, and ISPAR exciting unit. 88,128 internal fault cases which include basic internal faults, turn-to-turn, and winding-to- winding faults are simulated by varying the fault resistance, % of winding shorted, fault inception time, forward or backward shift, and the LTC in the exciting unit. Fig. 2: 3-phase differential currents for turn-to-turn faults in (a) primary of 1) Internal phase & ground faults (ph & g): Phase winding exciting unit, (b) primary of series unit, and (c) primary of PT to ground (w -g, w -g, w -g), phase winding to phase winding a b c to ground (w -w -g, w -w -g, w -w -g), phase winding to a b a c b c TABLE II: Parameters for winding-to-winding & turn-to-turn faults in the phase winding (w -w , w -w , w -w ), 3-phase winding (w - a b a c b c a ISPAR and PT w -w ), and 3-phase winding to ground (w -w -w -g) faults b c a b c Variables Values are simulated in the primary (P) and secondary (S) sides of Fault resistance 0.01, 0.5 & 10 (3) the PT and on the primary and secondary sides of exciting and % of winding shorted 20%, 40%, 60%, 80% (4) Fault inception time 15s to 15.0153s in steps of 1.38ms (12) series transformer units in the ISPAR. Table I shows the values Transformer phase a,b,c (P & S) (6) of different system and fault parameters in T and ISPAR 1 1 Fault location ISPAR Exciting phase a,b,c (P & S) (6) (Fig.1(a)) which are varied to get the training and testing cases & ISPAR Series phase a,b,c (P & S) (6) Phase shift Forward and backward (2) for the internal phase & ground faults. LTC 0.2,0.4,0.6,0.8,1 [1 & 0.5 in ISPAR exciting] TABLE I: Parameters for ph & g faults in the ISPAR and PT Transformer or ISPAR series(T-T) faults = 3412625 = 8640 ISPAR exciting(T-T) faults = 3412622 = 3456 Variables Values Transformer or ISPAR series(W-W) faults = 3412325= 4320 Fault resistance 0.01, 0.5 & 10 (3) ISPAR exciting(W-W) faults = 3412322 = 1728 % of winding shorted 20%, 50%, 80% (3) Fault type w-g, w-w-g, w-w, w-w-w & w-w-w-g (11) Fault inception time 15s to 15.0153s in steps of 1.38ms (12) 3) Winding-to-winding (W-W) faults: The electrical, ther- Transformer (P & S) (2) mal and mechanical stress due to short circuits and transformer Fault location ISPAR Exciting unit (P & S) (2) aging reduces the mechanical and dielectric strength of the & ISPAR Series unit (P & S) (2) Phase shift Forward and backward (2) winding and results in degradation of the insulation between LTC 0.2,0.4,0.6,0.8,1[1 & 0.5 in ISPAR exciting] LV and HV winding and may damage the winding eventually Transformer or ISPAR series faults = 331112225 = 23,760 [26]. Table II shows the values of different parameters of the ISPAR exciting faults = 331112222 = 9504 PT and the series and exciting unit of ISPAR used to simulate 10,368 winding-to-winding faults. 2) Turn-to-turn (T-T) faults: About 70-80% of faults in transformers are due to turn-to-turn insulation failures. Ther- B. Magnetizing inrush mal, mechanical and electrical stress degrades the insulation Transients caused by the energization of transformers are and causes turn-to-turn faults which may lead to more serious common and discrimination of inrush from fault currents has faults and inter-winding faults if not detected quickly [26]. been studied since the 19th century. The harmonic restraint Table II shows the values of different parameters of the PT and relays fail to detect inrush currents in transformers with the series and exciting unit of ISPAR used to simulate 20,736 modern core materials. The flux in a transformer core just turn-to-turn faults. Fig.2(a) shows the differential currents for after switching can be expressed as LTC= half, fault inception time=15s, backward phase shift, 0 0 fault resistance=0.01 and % turns shorted=20 in primary of  =  +  cos!t  cos!(t + t ) (3) R m m IEEE SYSTEMS JOURNAL 4 where,  represents residual flux,  represents maximum during switching of T which drives T to saturation can be R m 2 1 flux, and t is switching time. The transformer draws a high expressed as peaky non-sinusoidal current to meet the high flux demand 2+t = [(R + R )i + R i ] (4) sys T 1 sys 2 where R is the system resistance , and R is the resistance sys T of transformer T , i and i are magnetizing currents of T and 1 1 2 1 T . This interaction between the incoming and the in-service transformers leads to mal-operation of differential relays of the in-service transformer due to failure of harmonic restraint relays and may cause prolonged harmonic over-voltages [27]. The use of superconducting winding, soft magnetic material in the core, and CT local transient saturation are some factors responsible for these mal-operations [3] [28]. Sympathetic inrush is influenced by the residual flux ( ) of the incoming transformer, switching time (t ), and system resistance [29] and Fig. 3: B-H curve of transformer core takes place with the incoming transformer energized in either series or parallel. The magnitude and direction of  , and when switched on. Since this current flows only on one side t are altered and the incoming transformer T1 is connected of the transformer the differential scheme mal-operates. T in parallel to simulate the scenarios. Table III shows the (Fig.1(a)) is the incoming 3-phase transformer and DC sources values of the different parameters used to get the training and are used to get the desired  in the single-phase transformers. testing cases for sympathetic inrush. Fig.4(b) shows the 3- The values for the DC currents in phase-a, b, and c are phase differential currents for LTC = 0.2, switching time = obtained from the B-H curve of the transformer core material 15s, forward phase shift, and -80% residual flux. as shown in Fig.3. Table III shows the values of different parameters including  and t used to get the data for training D. External faults with CT saturation and testing for magnetizing inrush and Fig.4(a) shows the 3- phase differential currents for LTC = full, switching time = The differential currents become non-zero due to CT sat- 15s, forward phase shift, and -80% residual flux. uration in case of heavy through faults and may lead to a false trip. While raising the bias threshold ensures the security (i.e. no mal-operation), the dependability for in-zone resistive faults gets reduced. The external faults with CT saturation are simulated on the 500kV and 230kV buses (bus4 & bus5). The values for the different parameters are given in Table IV. Fig.5(a) shows the 3-phase differential currents for an external line-to-ground (lg) fault when LTC = 0.2, phase shift = forward, fault inception time = 15s, and fault resistance = 0.01 on the 230kV bus. Fig. 4: 3-phase differential currents for (a) Magnetizing inrush, and (b) Sympathetic inrush TABLE III: Parameters for Magnetizing and Sympathetic inrush Variables Values Residual flux 80%;40%; 0% in 3 phases; 5 3 = (15) Switching time 15s to 15.0153s in steps of 1.38ms (12) LTC 0.2 to full tap in steps of 0.2 (5) Phase shift Forward and backward (2) Total=15 12 5 2=1800 C. Sympathetic Inrush Sympathetic inrush occurs in the in-service transformer (T ) Fig. 5: 3-phase differential currents for (a) External fault with CT saturation, when the incoming transformer (T ) is energized in a resistive (b) Capacitor Switching, and (c) Ferroresonance network at no-load. The asymmetrical flux change per cycle IEEE SYSTEMS JOURNAL 5 TABLE IV: Parameters for External faults on 230kV & 500kV bus high transient inrush currents from capacitor-bank switching to avoid malfunctioning of instantaneous and time overcurrent Variables Values relays (50/51). A capacitor bank having 3 Legs of 500 MVAr Fault resistance 0.01, 0.5 & 10 (3) Fault type lg, llg, ll, lll & lllg (11) each is connected to the 230kV bus. Capacitor bank reactors Fault inception time 15s to 15.0153s in steps of 1.38ms (12) and resistors are used in each Leg to reduce the effect of LTC 0.2 to full tap in steps of 0.2 (5) transients in voltages. Table VI shows the different parameters Phase shift Forward and backward (2) Fault location 230kV & 500kV bus (2) and their values used to get the data for training and testing Total=3 11 12 5 2 2=7920 for capacitor switching. Fig.5(b) shows the 3-phase differential currents for LTC = full, switching time = 15.00138s, and switching of 3 Legs of the capacitor bank. E. Non-linear Load Switching TABLE VI: Parameters for Capacitor Switching With the advancement in semiconductor technology and the use of non-linear loads with power converters, harmonic Variables Values contents in the line currents have increased. The differential re- Capacitor bank rating 500,1000,1500 MVAr (3) lays may mal-operate when non-linear loads e.g steel furnaces Switching time 15s to 15.0153s in steps of 1.38ms (12) Phase shift Forward and backward (2) are switched in a network containing transformers because LTC 0.2 to full tap in steps of 0.2 (5) of mutual enhancement effects between the transformer core Total = 3 12 2 5 = 360 and the load causing extreme saturation of the transformer core for several cycles [30]. The harmonic information has been used to discriminate faults from other disturbances and G. Ferroresonance locate the faults in the transmission line using SVM and ANN Initiated by faults and switching operations, ferroresonance [31]. A thyristor-based 6-pulse bridge rectifier with a wye- causes harmonics and overvoltages and may lead to mal- delta transformer as the non-linear load is connected to the operation of protective relays and damage of power equipment 230 kV bus to obtain the training and testing cases for load [33]. Mal-operation of the differential relay occurs because of switching. The values for the different parameters are given the higher magnitude of current in the HV side than the LV in table V and Fig.6 shows the phase-a differential current side [34]. Besides, the low loss, amorphous core transformer for LTC = full, switching time = 15s and firing angle of 0 . increases the intensity and occurrence of ferroresonance [35]. Fig.6(a) shows the transient and Fig.6(b) shows the steady- Several configurations may lead to ferroresonance in electrical state differential current after the switching. systems. In this paper, one such arrangement has been modeled when one of the phases of a 3-phase transformer is switched off. The parameters and their values for ferroresonance con- ditions are presented in Table VII. Fig.5(c) shows the 3-phase differential currents for switching time = 15s and grading capacitance = 0.2F simulated between bus2 and bus4. TABLE VII: Parameters for Ferroresonance Variables Values Grading capacitance 0.02F to 0.2F in steps of 0.02F (10) Location a,b,c phases (3) Switching time 15s to 15.016s in steps of 0.69ms (24) Total=10 3 24=720 Fig. 6: Non-linear Load Switching (a) Transient, and (b) Steady-state differ- ential currents III. DETECTION, DISCRIMINATION & CLASSIFICATION ALGORITHM TABLE V: Parameters for Non-linear Load Switching A. Change detection filter (CDF) for transient detection Variables Values The change in the differential currents in case of transients Firing angle 0 , 10 , 20 , 30 , 40 , 50 (6) is detected by a change detection filter (CDF) which calculates Switching time 15s to 15.0153s in steps of 1.38ms (12) the difference between the cumulative sum of modulus of two LTC 0.2 to full tap in steps of 0.2 (5) Total = 6 12 5 = 360 consecutive cycles. 2n +t 2n +t c c X X CDF (t) = jId(x)j jId(x n )j (5) F. Capacitor Switching x=n +t x=n +t c c Capacitor banks are used to improve voltage profile, re- where x equals sample number which begins at the second nn duce losses, and enhance power factor. Mal-functioning of cycle, n equals number of samples in a cycle, ftg , n t=1 customer equipment due to voltage magnification coinciding equals total number of samples, and Id represents a, b, and c with capacitor switching is common. [32] used WT to detect phase differential currents. IEEE SYSTEMS JOURNAL 6 The change detection filter starts logging the data from the F5, autoregressive coefficients are the least-square estimates instant CDF(t) is greater than a threshold, th in any one of of ' which are obtained by minimizing Eq.8 with respect i s the 3-phases. In normal conditions when there is no transient, to ' ; ' :::; ' and lag P. 0 1 P the values of CDF(t) are nearer to zero [36]. [Id ' '  Id ::: '  Id ] (8) t 0 1 t1 P tP B. Feature Extraction & Selection t=p+1 Time series analysis of the differential currents helps in More than one feature can be extracted from the above time the classification and characterization of power system events. and frequency domain functions by varying their parameters. Features extracted from these time series are used as input to e.g (qh; ql) = (0.8,0.4) & (0.8,0.2) yields 2 features from the machine learning algorithms. Informative and distinctive change quantile and window length = 5, 10, and 15 would features that help to classify the events may range from simple return 3 features of linear trend. statistical functions to complex ones. Researchers have used time-frequency representations like Wavelet Transform [8]– [10], [13], [17], [18], [21] and Stockwell Transform [16], [19]– C. Classifiers [21] to extract features from the non-stationary transients to Tree-based learning algorithms like decision trees, random discriminate inrush and internal faults and for classification of forest, and gradient boosting are considered among the best PQ disturbances. In this paper, to differentiate the faults from and predominantly used supervised learning methods in prob- the other transient disturbances, three time-domain features lems related to data science. These estimators have higher and two frequency-domain features have been used. accuracy, stability and are easy to interpret. They can also A comprehensive number of features (794) from different handle non-linear relationships quite well. DT, RFC, GBC, domains are extracted from the 3-phase differential currents and SVM has been used to detect and classify the transients. obtained from the current transformers, CT1 and CT2 located 1) Decision Tree: Decision trees are distribution-free white near bus4 and bus5. The complete list of the features extracted box Machine Learning models that learn simple decision can be found in [37]. Out of these 794 features, Random rules inferred from the feature values. In 1984 Breiman et Forest is used to rank and select the features with maximum al. introduced Classification and Regression Trees (CART) Information Gain to distinguish between the different classes. [39]. Here, the CART algorithm implemented in scikit-learn The most relevant and common features for each of the is used which constructs binary trees by splitting the training classification tasks obtained after performing feature ranking set recursively till it reaches the maximum depth or a splitting belong to the set F =fF1, F2, F3, F4, F5g where, F1 is average doesn’t reduce the impurity measure. The candidate parent change quantile, F2 is Fourier transform (FT) coefficients, F3 data D is split into D and D at each node using a feature p l r is aggregate linear trend, F4 is spectral welch density, and F5 is (f ) and threshold that yields the largest Information Gain. autoregressive coefficients. Only those features of set F which The objective function IG which is optimized at each split are present in each of the 3-phase differential currents are is defined as used for training the classifiers to detect the faults, localize the N N faulty units, identify the fault type, and identify the disturbance l r IG(D ; f ) = I (D )  I (D )  I (D ) (9) p p l r type (Table XVI). The feature set F is detailed in what follows. N N p p F1, average change quantile calculates the average of ab- where, I is impurity measure, N ; N and N are the number of p l r solute values of consecutive changes of the time series inside samples at the parent and child nodes [40]. Gini, and entropy two constant values qh and ql as impurity measures are used in the hyperparameter search. n 1 1 2) Random Forest: RFC belongs to the family of ensemble avg: change quantile =  jId Id j (6) t+1 t trees which builds numerous base estimators and averages their t=1 predictions which produces a better estimator with reduced where, n equals number of sample points in the differential variance. Each tree constitutes a random sample (drawn with current between qh and ql, Id is a, b, and c phase differential replacement) of the training set and the best split is found currents with n sample points. at each node by considering a subset of input features. The F2, FT coefficients, (Xjk) returns the fourier coefficients of individual trees tend to overfit but averaging the predictions of 1-D discrete Fourier Transform for real input using fast FT as all trees reduces the variance [41]. The main hyperparameters n1 in RFC are no of estimators (number of trees in the forest), j2kt (Xjk) = Id  e( ); k 2 Z (7) t max depth (tree depth), and max features (feature size to t=0 consider when splitting a node). The no of jobs parameter F3, aggregate linear trend calculates the linear least-squares was also used to parallelize the construction of tree and regression for values of the time series over windows and computation of predictions by using more processing units. returns aggregated values of either intercept or standard error. Random Forest has also been used during feature selection F4, spectral welch density uses Welchs method to compute and ranking (III-B) to get the relative importance of the an estimate of the power spectral density by partitioning the features which is measured by the fraction of samples a feature time series into segments and then averaging the periodgrams contributes to and the mean decrease in impurity from splitting of the discrete Fourier transform of each segment [38]. the samples [42]. IEEE SYSTEMS JOURNAL 7 3) Gradient Boosting Classifier: GBC belongs to the class used for localization of faults and classification of transient of ensemble trees which builds the base estimators from weak disturbances. The scheme consists of a four-level classifier learners (w (x)) sequentially in a greedy manner which results design. The level-1 classifier (GBC-1) consists of the fault in a strong estimator [43] [44]. The newly added w tries to detector, which can apply supervisory control over the oper- minimize the loss function given f , step length ( ), and ation of the conventional differential relay. GBC-1 identifies p1 p input (x ; y ) . an internal fault with “0” and other transient disturbances with i i i=1 “1”. Hence, it governs the working of the trip/restrain function f (x) = f (x) +  w (x) p p1 p p by blocking all other power system transients but an internal (10) fault. The level-2 classifier (GBC-2) does further analysis of w = arg min L(y ; f (x ) + w(x )) p i p1 i i the power system events in case the output of GBC-1 is “1”. i=1 The GBC-2 can identify the transient disturbance responsible The minimization problem is solved by taking the negative for the mal-operation of the conventional differential relay gradient of the negative multinomial log-likelihood loss func- (GBC-1 is “1” & Operate relay is “1”). The level-3 classifier tion, L for mutually exclusive classes. (GBC-3) locates the faulty transformer unit (PT, ISPAR series, and ISPAR exciting) if the output of GBC-1 is “0”. The level-4 f (x) = f (x)  r L(y ; f (x )) (11) p p1 p f i p1 i classifiers (GBC-4, GBC-5, and GBC-6) further identifies the i=1 internal faults in the ISPAR exciting, the ISPAR series and the PT. GBC uses shrinkage which scales the contribution of the weak learners by the learning rate and sub-sampling of the training data (stochastic gradient boosting) for regularization. The im- IV. R ESULTS AND DISCUSSION portant hyperparameters of the different GBC classifiers used are the results of grid search on no of estimators = [5000, 1.5 cycles of 3-phase differential currents are used for detec- 7000, 10000, 12000, 15000], max depth = [3,5,7,10,15], and tion, and 3 cycles are used for localization and identification of learning rate = [0.01, 0.05, 0.07, 0.1]. transients from the time of their inception. Thus, at a sampling rate of 10 kHz, 167 data samples per cycle are analyzed. Several factors influence the classification accuracy of an D. Proposed scheme algorithm. Cross-validation and grid search helps in using the data effectively and training the classifier with the best combination of hyperparameters. The data is split randomly into training and test set in a 4:1 ratio. To avoid the problem of overfitting and underfitting of the estimator on the test set, cross-validation is applied on the training data and the hyper- parameters are optimized using grid search over a parameter grid. Grid search comprehensively searches for the parameters over the subset of the hyperparameter space of the estimator. The performance of the selected hyperparameters is then tested on the unseen test data that is not used during the training process. Ten-fold stratified cross-validation (rearrangement of the training data in ten folds such that each fold represents every class well) is used to select the model as it is better both in terms of bias and variance [45]. TABLE VIII: Internal fault detection with CDF & GBC-1 Fault/Disturbances Total TP FN FP Internal Faults 2107 2105 2 0 Disturbances 1852 1852 0 2 Fig. 7: Proposed transient detection and classification algorithm TABLE IX: Comparison of performances with and without CDF (b) Internal fault detection The block diagram description of the CDF and GBC based (a) Internal fault detection with CDF without CDF proposed internal fault detection, fault localization, and tran- sient disturbance classification algorithm is shown in Fig.7. Classifier   Classifier The change detector discovers the change in the 3-phase GBC-1 99.95 % GBC 98.5% differential currents (I -I ) if the CDF index in any phase P S DT 99.5% DT 95.3% is greater than the threshold, th = 0.05. 1/2 cycle pre-transient SVM 99.7% SVM 89.2% RFC 99.9% RFC 94.6% and 1 cycle post-transient differential current samples are used to detect an internal fault and 3 post-transient cycles are IEEE SYSTEMS JOURNAL 8 TABLE X: Localization of faulty transformer unit A. Internal fault detection (b) Comparison of perfor- (a) Localization with GBC-3 mances Transformer Total TP FN FP Classifier The detection of internal faults is performed using GBC in ISPAR Exciting 2937 2899 38 8 GBC-3 99.5% ISPAR Series 7402 7383 19 17 two ways, one with the CDF and the other without it. Most DT 98.6% PT 7287 7287 0 32 authors haven’t considered using some technique to detect SVM 88.9% RFC 98.7% the change in differential currents in case a transient occurs. Rather they fixed the time of occurrence of the transient events and used this specified inception time to store the B. Identification of faulty unit & internal fault type disturbance and fault data. However, faults and disturbances are highly unpredictable in time. In this paper, both methods, Once it is confirmed that an internal fault has been detected, one considering a specified time (without the use of CDF) the locations of those internal faults are determined. 3 cycles and the other with CDF are used to register the data after of post-fault differential current samples are used to locate the inception of transients. The CDF detects the change and the faulty transformer unit (PT or ISPAR Exciting or ISPAR registers 1/2 cycle of pre-transient and 1 cycle of post-transient Series) and determine the type of fault. GBC, SVM, DT, and samples. This 1.5 cycle (250 samples) is used to extract the RFC are used to identify the faulty unit and further locate relevant features which are then fed to GBC, SVM, DT, and and pinpoint the type of fault in the PT and ISPAR units. (TP +TN) RFC classifiers. Accuracy is used as the typical metrics to and accuracy computed as  = , are used as (TP +FN+TN+FP ) measure the performance of the classifiers. But, it is biased to the metrics to measure the performance of the estimators for data imbalance. Since, the classes are imbalanced, balanced localization of faulty unit and identification of internal fault accuracy which is defined as mean of the accuracies obtained type, respectively. 1 TP TN on all classes and computed as   =  [ + ] 2 (TP +FN) (TN+FP ) TABLE XI: Comparison of identification performances of internal fault type for binary classes is used to compute the performance mea- (a) Exciting unit (b) Series unit (c) PT sure where, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false Classifier  Classifier  Classifier negative [46]. GBC-4 99.2% GBC-5 98.0% GBC-6 99.2% DT 98.6% DT 94.7% DT 98.9% The performance of the fault detection scheme composed of SVM 94.8% SVM 90.7% SVM 94.0% the GBC-1 and CDF is shown in Table VIII.   of 99.95% is RFC 98.9% RFC 96.9% RFC 97.8% obtained on a training data of 15,835, testing data of 3959, 1) Localization of faulty unit: To locate the faulty trans- and hyperparameters: learning rate = 0.1, max depth = 5, former unit 70,502 fault cases are trained and 17,626 cases and no of estimators = 7000. The performance of the four classifiers with CDF is shown in Table IXa. One cycle of post- are tested. 18 features are used to train the classifiers (Table fault data is used for training the classifiers for fault detection XVI). GBC-3 with hyperparameters: no of estimators = 5000, without the CDF.   of 98.52% is obtained with GBC for learning rate = 0.07, and max depth = 10 gives   of 99.48%. max depth = 7, no of estimators = 5000, and learning rate Table Xa shows the localization results using GBC-3 and Table = 0.07. The balanced scores of the four classifiers trained on Xb compares the   of the four different classifiers. 80,870 cases and tested on 20,218 cases are shown in Table 2) Identification of internal fault type: The internal faults IXb. 18 features from the 3-phase differential currents (Table in the ISPAR series, ISPAR exciting and the PT are further XVI) are used as the input to the classifiers for training the classified into w -g, w -g, w -g, w -w -g, w -w -g, w -w - a b c a b a c b c fault detection models with and without CDF. GBC with CDF g, w -w , w -w , w -w , turn-to-turn, winding-to-winding, a b a c b c performed better than without CDF (Table IX) as the CDF and very rare w -w -w and w -w -w -g faults. 21 features a b c a b c filtered out the cases where there is no appreciable change in from 3 cycles of the 3-phase differential currents are used differential currents although a transient event occurred. It is as the input to the estimators (Table XVI). Tables XIa, XIb, noticed that the CDF could detect the change in differential and XIc compare the performances of GBC, RFC, DT, and currents in all internal fault cases except turn-to-turn faults SVM classifiers for ISPAR exciting, ISPAR series, and the PT with Rf = 10 , LTC = 0.2, and percentage of winding respectively. shorted = 20%. Also, it detected the change for all transient To identify the internal faults in ISPAR exciting 14,688 disturbances except sympathetic inrush cases for switching fault cases are used to train and test the four classifiers. angles from 120 to 330 . On exploring the data it is observed GBC-4 trained with hyperparameters of max depth = 5, that there is almost no change in the differential currents for no of estimators = 7000, and learning rate = 0.1 achieved these instances. The w-g faults for LTC = 0.2, and percentage the best accuracy of 99.18%. For the identification of in- of winding shorted = 20% which needs higher sensitivity were ternal faults in ISPAR series 36,720 cases are used to train detected. It proves the dependability of the scheme for ground and test the classifiers. GBC-5 trained with learning rate = faults near neutral of wye grounded transformers (PT and 0.05, max depth = 7, and no of estimators = 5000 gives an ISPAR exciting) which is again a challenge for conventional accuracy of 98.0%. Similarly, for PT the classifiers are trained differential relays [7]. & tested on 36,720 fault cases. GBC-6 achieved the best ac- IEEE SYSTEMS JOURNAL 9 curacy of 99.2% obtained by training the hyperparameters on learning rate = 0.05, no of estimators = 5000, and max depth = 5. The identification accuracy obtained in the ISPAR series is lower than in PT and ISPAR exciting because the secondary side of ISPAR series is delta connected. Hence, one type of fault on the primary side confuses with another type on the secondary side. C. Identification of disturbance type The various disturbances: magnetizing inrush, sympathetic inrush, ferroresonance, external faults with CT saturation, capacitor switching, and non-linear load switching are also classified using 3 cycles of post-transient samples after they are differentiated as no-fault by the fault detection scheme. Fig. 8: Kernel Density Estimate plots showing the probability distribution of 15 features are used as input to the classifiers in this case the 5 selected features for the 6 transient disturbances in (a) phase a, (b) phase (Table XVI). It’s always useful to know the probabilities of the b, and (c) phase c. input features taking on various real values. ParzenRosenblatt TABLE XII: Low and high-order statistics of the 5 selected features window method is used to estimate the underlying probability density of the 5 features for the six different disturbances in (a) Magnetizing inrush phases a, b, and c. Fig.8 shows the kernel density estima- autoregressive FT linear change change tion plots for the chosen features. Gaussian is used as the coefficient coefficient trend quantile 1 quantile 2 2 2 2 2 2 kernel function to approximate the univariate features with   ~    ~    ~    ~    ~ 3 3 3 3 3 ph a -4.6 5.9 .74 -.35 1e3 2e6 1.3 .74 2.6 14 1.9 3.9 .02 9e-4 1.9 2.9 .04 .008 2.5 5.1 a bandwidth of 0.2 for autoregressive coefficient, and FT ph b -4.7 6.2 .65 -.13 502 3e5 1.0 -.06 2.6 13 1.8 3.2 .01 8e-5 1.5 1.06 .04 .01 2.5 5.7 ph c 3.3 .72 -1.1 .33 33 1e3 1.1 .25 .19 .05 1.0 -.35 .02 5e-4 .90 -.45 .08 .009 .88 -.84 coefficient and a bandwidth of 0.01 for aggregate linear trend, and avg. change quantile 1 and 2. It is observed that probability (b) CT saturation during external faults density functions of autoregressive and FT coefficients are a autoregressive FT linear change change mixture of multiple normal distributions with varying standard coefficient coefficient trend quantile 1 quantile 2 deviation and mean whereas linear trend, and change quantiles 2 2 2 2 2 ~    ~    ~    ~    ~ 3 3 3 3 3 are unimodal with means near zero and smaller standard ph a -2.3 3.1 -1.6 1.97 190 4e4 1.6 2.5 .47 .23 1.64 2.9 .04 6e-4 1 .98 .003 9e-6 1.8 3.6 ph b -2.2 3 -1.6 1.95 261 2e4 .81 .35 .47 .24 1.55 2.3 7e-4 9e-7 3.5 19 .04 5e-4 .6 -.42 deviations. Table XIIa and XIIb shows the values of mean (), ph c 2.5 .4 1.1 .27 16 150 1.7 2.9 .02 5e-4 1.5 2.4 .003 8e-6 1.7 4.7 .04 4e-4 .67 -.14 variance ( ), skewness ( ~ ), and kurtosis () of the 5 features for magnetizing inrush and CT saturation during external faults respectively in phases a, b and c. Because of space limitations, to verify how effectively the GBC discriminates the internal the feature statistics of only these two transients are shown. faults in the PT and the ISPAR. The table XV shows the Furthermore, to visualize the 15-dimensional input data in classification errors with learning rate of 0.05, max depth = a 2-dimensional plane, the T-distributed Stochastic Neighbor 9, and no of estimators = 5000. The balanced accuracy of Embedding dimensionality reduction technique has been used 99.9% shows that the GBC is capable of distinguishing these which preserves much of the significant structure in the high- faults even in an interconnected network. dimensional data while mapping in the 2-dimension [47]. Fig.9 shows the clusters of similar transients (300 instances each) E. Performances on balanced and imbalanced data and also the relationships between different group of transients Machine learning algorithms are more reliable when they as a scatter plot. operate on a balanced dataset. To adjust the data distribution of The Table XIII shows the classification results using GBC- classes and remove class imbalance, under-sampling of major- 2. Table XIV compares the results of GBC with RFC, DT, and SVM. The classifiers are trained on 10,368 cases and tested on 2592 cases.   of 99.28% is obtained with GBC-2 having hyperparameters: no of estimators = 5000, learning rate = 0.7, and max depth = 3. D. Discriminate faults in PT & ISPAR The PAR controls the power flow through a line and when connected with a PT, it reduces the magnitude of the differential currents and their harmonic contents and alters the wave shapes due to the additional phase shift for the PT in case of external faults. With internal faults, such changes in the differential currents are lesser. 70,502 internal fault Fig. 9: 2D scatterplot of the input features for transient disturbances cases are trained and 17,626 cases are tested on 18 features IEEE SYSTEMS JOURNAL 10 TABLE XIII: Identification of transient disturbances with GBC-2 TABLE XVII: Performance for 400-MVA & Y connection Disturbances Total TP FN FP Fault/ Faults/ Total TP FN Disturbances Abnormalities (%) Magnetizing inrush 365 357 8 0 Sympathetic inrush 336 336 0 8 (a) ph & g, T-T, W-W faults (PT) 1200 1200 0 100 Internal Capacitor switching 73 72 1 1 (b) ph & g, T-T, W-W faults (Series) 1200 1200 0 100 faults Ferroresonance 133 132 1 0 (c) ph & g, T-T, W-W faults (Exciting) 672 672 0 100 (3072) Load switching 69 69 0 0 (d) Total = (a)+ (b)+ (c) 3072 3072 0 100 External faults 1616 1615 1 2 (e) Capacitive switching 60 51 9 85 TABLE XIV: Performance (f) External faults with CT saturation 528 525 3 99.4 Other comparison of identification (g) Ferroresonance 24 24 0 100 TABLE XV: Misclassification disturbances of transient disturbances (h) Magnetizing inrush 60 60 0 100 between ISPAR & PT (876) (i) Load switching 144 144 0 100 (j) Sympathetic inrush 60 60 0 100 Classifier Faults Total TP FN FP (k) Total = (e)+ (f)+ (g)+ (h)+ (i)+ (j) 876 864 12 98.6 GBC-2 99.28% PT 7262 7262 0 13 DT 98.09% ISPAR 10364 10351 13 0 Total (3948) Total faults and disturbances = (d)+(k) 3948 3936 12 99.3 SVM 98.23% RFC 98.89% TABLE XVIII: Effect of Noise Fault/ Number Predicted class Accuracy SNR (dB) ity classes and over-sampling of minority classes is performed. Disturbances of cases (%) Faults Disturbances Synthetic Minority Over-Sampling Technique (SMOTE) [48] 1 1010 1001 9 99.2 is used to create minority synthetic data considering k-nearest Internal 30 1035 1005 30 97.1 neighbours and NearMiss algorithm is used for under-sampling faults 20 1008 934 74 92.7 the majority classes avoiding information loss. The table 10 984 891 93 90.4 XVI shows the balanced accuracy/ accuracy for detecting 1 990 3 987 99.7 the internal faults, identifying the faulty units and type of Other 30 965 24 941 97.6 disturbances 20 992 26 966 97.4 faults in those units, and identifying the disturbances with and 10 1016 28 988 97.2 without using SMOTE and NearMiss. It is observed that the accuracies obtained with SMOTE and NearMiss algorithm for the different classification tasks are similar to those obtained is compatible with the accuracy obtained when trained and by training the GBCs without them. Table XVI also gives tested at 500 MVA and YY connection. the information about the time and frequency domain features (fF g ) that has been used to train the different GBC i=1 classifiers for the different classification tasks. G. Effect of Signal Noise & CT saturation TABLE XVI: Input features and performance of different GBC classifiers with In order to analyse the effect of noise in the differential cur- and without SMOTE analysis rents on the proposed fault detection scheme white Gaussian noise of different levels measured in terms of Signal-to-Noise- n   using Classification task F1 F2 F3 F4 F5 Fi3 (%) SMOTE ratio (SNR) are added to the training and testing cases for fault detection [17], [19]–[21]. Table XVIII shows the accuracy detect faults 2 1 1 1 1 18 99.9 99.9 locate faulty units 2 2 2 - - 18 99.5 99.6 of the GBC for different levels of noise on 5000 cases of identify faults (series) 3 1 2 1 - 21 98.0 98.2 internal faults and other disturbances each. It is observed that identify faults(exciting) 3 2 2 - - 21 99.2 99.1 as the level of noise increases the  of the classifier dips, identify faults (PT) 3 2 2 - - 21 99.2 99.1 90:4+97:2 identify transients 2 1 1 - 1 15 99.3 99.4 but still always above 93.8% ( ). The  changes from 99.4% to 93.8% as the SNR is varied from 1 to 10dB. It is also observed from the table that the misclassification of internal faults increases as the SNR is decreased whereas the F. Effect of different rating and transformer connections misclassifications are nearly the same for other disturbances as It is not necessary to train the fault detection scheme for SNR is decreased from 30dB to 10dB. Moreover, to examine different rating and connection of the PTs, rating of ISPAR, the effect of CT saturation the secondary side impedance and variation in other parameters. In order to validate the (burden and CT secondary impedance) which has the major effectiveness of the proposed scheme with variation in differ- influence over the level of saturation is changed.  of 99.5% ent system parameters, new internal faults and other transient is obtained with GBC on 5000 cases of internal faults and cases are simulated again with 400 MVA, Y connected PTs other disturbances each. Fig.10 shows the 3-phase differential and 400 MVA ISPARs. The fault resistance, LTC, fault type, currents with CT saturation for faults in T1 and ISPAR1. and fault inception time are altered to generate the internal fault cases and switching time, firing angle, LTC, etc. are H. Execution Time altered to generate the transient cases to test the same GBC- 1 model trained using 500 MVA, YY connected PTs and The execution time-averaged over 10 runs for the feature 500 MVA ISPARs. It is observed from Table XVII that the extraction, training, and testing of the GBC classifiers for proposed scheme gives a balanced accuracy of 99.3% which detection of internal faults, identifying the faulty unit and type IEEE SYSTEMS JOURNAL 11 sympathetic inrush, ferroresonance, external faults with CT saturation, capacitor switching, and non-linear load switching transients. A change detector is used to detect the change in the 3-phase differential currents in case a transient event occurs and registers the current samples for detection and classification purposes. Five most relevant time and frequency Fig. 10: 3-phase differential currents with CT saturation for (a) w -w -g fault domain features, selected from the differential currents on the in transformer primary and (b) w -g fault in series primary basis of Information Gain are used to train the DT, RFC, GBC, and SVM classifiers. The fault detection scheme comprising of the CDF and GBC gives an accuracy of 99.95% on 19,794 of fault, and identifying the transient using one CPU core is transient cases obtained by varying different parameters for the reported using the in-built library in python (Table XIX). The internal faults and other transient disturbances confirming its fault/no-fault decision includes the time to compute the feature dependability for internal faults and security against transient and testing a single instance with GBC-1 which adds to 8.7ms disturbances. Once an internal fault is detected and a trip signal with the CDF. Thus, the proposed scheme has a processing is issued using 1.5 cycles, the faulty transformer unit (PT, IS- time of 25.37ms (16.67+1.7+7) or  1 cycle to detect a fault. PAR series, or ISPAR exciting unit) and type of internal faults Considering that these computations can be further optimized in those units are also identified in 3 cycles. Furthermore, the for example by converting Python and MATLAB code to a type of transient disturbance is determined in case the fault compiled low-level language such as C, the fault detection and detection scheme detects a transient other than internal faults. localization, and transient identification schemes are suitable The validity of the scheme is also established for different for future real-time implementation. The DT, SVM, RFC, rating and connection of the transformers, CT saturation, and and GBC classifiers are built in Python 3.7 using Scikit-learn SNR ratio of 30dB to 10dB in the differential currents. The framework [49] while the CDF is implemented in MATLAB proposed fault detection strategy can work together with a 2017. The pre-processing of the data is done in Python and conventional differential relay offering supervisory control MATLAB. All PSCAD simulations are carried out on Intel over its operation and thus avoid false tripping. The transient Core i7-6560U CPU @ 2.20 GHz and 8 GB RAM and the detection and identification accuracies obtained are among the classifiers are run on Intel Core i7-8700 CPU @ 3.20 GHz best even when compared with results from works on isolated and 64 GB RAM. and simple networks. TABLE XIX: Execution time of the GBC classifiers (seconds) APPENDIX Training Testing Training Testing time Feature Classification task Fortran script for two-winding transformer instances instances time extraction One All time 1. NW = 4 18. L2l = Lk1/2*fb detect faults 80870 20218 1506 0.0024 1.1 0.007 2. I 2 = I 1 = I 19. L3l = Lk2/2*fc m m m detect faults with CDF 15835 3959 19 0.0017 .061 0.007 3. fa = fault1 0:01 20. L4l = Lk2/2*fd locate faulty units 70502 17626 1232 0.0049 1.2 0.0088 4. fb = 1:0 fa 21. L1m = (v1/(w*I 1*i1))*fa fa identify faults (series) 29376 7344 1341 0.016 2.1 0.0111 5. fc = fault2 0:01 22. L2m = (v1/(w*I 1*i1))*fb fb identify faults(exciting) 11750 2938 357 0.013 0.48 0.0108 6. fd = 1:0 fc 23. L3m = (v2/(w*I 2*i2))*fc fc identify faults (PT) 29376 7344 2712 0.026 5.1 0.0108 7. i1 = MVA/v1 24. L4m = (v2/(w*I 2*i2))*fd fd identify transients 10368 2592 72.3 0.004 0.09 0.0065 8. i2 = MVA/v2 25. Lx = L1l + L1m 9. z1 = v1/i1 26. Ly = L2l + L2m 10. z2 = v2/i2 27. Lz = L3l + L3m This work distinguishes faults from the six transients (  = 11. w = 2*pi*f 28. Lw = L4l + L4m 12. l1 = v1/(w*I 1*i1) 29. Mxy = sqrt(L1m*L2m) 99.95%), locates the faulty unit (  = 99.5%), identifies the m 13. l2 = v2/(w*I 2*i2) 30. Mxz = sqrt(L1m*L3m) fault type (  99%) and six other transients (  = 99.3%) 14. Lk1 = Xl*z1/w 31. Mxw = sqrt(L1m*L4m) for the ISPAR and PT in an interconnected system, whereas 15. Lk2 = Xl*z2/w 32. Myz = sqrt(L2m*L3m) the publications [22] [23] in the literature focused only on 16. tr = v1/v2 33. Myw = sqrt(L2m*L4m) 17. L1l = Lk1/2*fa 34. Mzw = sqrt(L3m*L4m) the ISPAR. In [22], only the internal faults in ISPAR were identified and in [23], the internal faults were differentiated from magnetizing inrush using WT and then the internal faults REFERENCES were identified. In addition to its broader functionality, the [1] R. Hamilton, “Analysis of transformer inrush current and comparison of current work improves the accuracy from an average of 98.76 harmonic restraint methods in transformer protection,” IEEE Transac- [22] and 97.7% [23] to 99.2%. tions on Industry Applications, vol. 49, no. 4, July 2013. [2] Pei Liu, O. P. Malik, Deshu Chen, G. S. Hope, and Yong Guo, “Improved operation of differential protection of power transformers for internal V. CONCLUSION faults,” IEEE Transactions on Power Delivery, vol. 7, no. 4, Oct 1992. [3] T. S. Sidhu, M. S. Sachdev, H. C. Wood, and M. Nagpal, “Design, In this paper, the task of discrimination of internal faults and implementation and testing of a microprocessor-based high-speed relay other transient disturbances in a 5-bus interconnected power for detecting transformer winding faults,” IEEE Transactions on Power Delivery, vol. 7, no. 1, pp. 108–117, Jan 1992. system for PTs and PARs is presented. The internal faults [4] D. Bi, Y. Sun, D. Li, G. Yu, X. Wang, and W. Wang, “Analysis on including turn-to-turn and winding-to-winding faults in the mal-operation of differential protection caused by sympathetic inrush,” ISPAR and the PT are distinguished from magnetizing inrush, Autom. Elect. Power Syst., vol. 31, pp. 36–40, Nov 2007. IEEE SYSTEMS JOURNAL 12 [5] M. Stanbury and Z. Djekic, “The impact of current-transformer satura- [29] G. B. Kumbhar and S. V. Kulkarni, “Analysis of sympathetic inrush tion on transformer differential protection,” IEEE Transactions on Power phenomena in transformers using coupled field-circuit approach,” in Delivery, vol. 30, no. 3, pp. 1278–1287, June 2015. 2007 IEEE Power Engineering Society General Meeting, June 2007. [6] M. A. Ibrahim and F. P. Stacom, “Phase angle regulating transformer [30] H. Weng and X. Lin, “Studies on the unusual maloperation of trans- protection,” IEEE Transactions on Power Delivery, vol. 9, no. 1, pp. former differential protection during the nonlinear load switch-in,” IEEE 394–404, Jan 1994. Transactions on Power Delivery, vol. 24, no. 4, Oct 2009. [7] “IEEE Guide for the Application of Protective Relaying for Phase- [31] E. Koley, S. K. Shukla, S. Ghosh, and D. K. Mohanta, “Protection Shifting Transformers,” IEEE Std C37.245-2018, pp. 1–71, May 2019. scheme for power transmission lines based on svm and ann consid- [8] P. Mao and R. Aggarwal, “A novel approach to the classification ering the presence of non-linear loads,” IET Generation, Transmission of the transient phenomena in power transformers using combined Distribution, vol. 11, no. 9, pp. 2333–2341, 2017. wavelet transform and neural network,” Power Engineering Review, [32] T. Patcharoen and A. Ngaopitakkul, “Transient inrush current detection IEEE, vol. 21, pp. 70–70, 08 2001. and classification in 230 kv shunt capacitor bank switching under various [9] S. Jazebi, B. Vahidi, and M. Jannati, “A novel application of wavelet transient-mitigation methods based on discrete wavelet transform,” IET based svm to transient phenomena identification of power transformers,” Generation, Transmission Distribution, vol. 12, no. 15, 2018. Energy Conversion and Management, vol. 52, no. 2, 2011. [33] M. K. Siahpoosh, D. Dorrell, and L. Li, “Ferroresonance assessment in [10] A. M. Shah and B. R. Bhalja, “Discrimination between internal faults a case study wind farm with 8 units of 2 mva dfig wind turbines,” in and other disturbances in transformer using the support vector machine- 2017 20th International Conference on Electrical Machines and Systems based protection scheme,” IEEE Transactions on Power Delivery, (ICEMS), Aug 2017, pp. 1–5. vol. 28, no. 3, pp. 1508–1515, July 2013. [34] S. Rezaei, “Impact of ferroresonance on protective relays in manitoba [11] O. Ozgonenel and S. Karagol, “Power transformer protection based on hydro 230 kv electrical network,” in 2015 IEEE 15th International decision tree approach,” IET Electric Power Applications, vol. 8, no. 7, Conference on Environment and Electrical Engineering, June 2015. pp. 251–256, August 2014. [35] M. Hajizadeh, I. Safinejad, and N. Amirshekari, “Study and comparison [12] S. Samantaray and P. Dash, “Decision tree based discrimination between of the effect of conventional, low losses and amorphous transformers on inrush currents and internal faults in power transformer,” International the ferroresonance occurrence in electric distribution networks,” CIRED Journal of Electrical Power & Energy Systems, vol. 33, no. 4, 2011. - Open Access Proceedings Journal, vol. 2017, no. 1, pp. 865–869, 2017. [13] Yong Sheng and S. M. Rovnyak, “Decision trees and wavelet analysis for [36] O. Dharmapandit, R. K. Patnaik, and P. K. Dash, “A fast time-frequency power transformer protection,” IEEE Transactions on Power Delivery, response based differential spectral energy protection of ac microgrids vol. 17, no. 2, pp. 429–433, April 2002. including fault location,” Protection and Control of Modern Power [14] M. Tripathy, R. P. Maheshwari, and H. K. Verma, “Probabilistic neural- Systems, vol. 2, no. 1, p. 30, Aug 2017. network-based protection of power transformer,” IET Electric Power [37] M. Christ, N. Braun, J. Neuffer, and A. W. Kempa-Liehr, “Time series Applications, vol. 1, no. 5, pp. 793–798, Sep. 2007. feature extraction on basis of scalable hypothesis tests (tsfresh a python [15] A. M. Shah and B. R. Bhalja, “Fault discrimination scheme for power package),” Neurocomputing, vol. 307, pp. 72 – 77, 2018. transformer using random forest technique,” IET Generation, Transmis- [38] P. Welch, “The use of fast fourier transform for the estimation of power sion Distribution, vol. 10, no. 6, pp. 1431–1439, 2016. spectra: A method based on time averaging over short, modified peri- [16] R. Kumar, B. Singh, D. T. Shahani, A. Chandra, and K. Al-Haddad, odograms,” IEEE Transactions on Audio and Electroacoustics, vol. 15, “Recognition of power-quality disturbances using s-transform-based ann no. 2, pp. 70–73, June 1967. classifier and rule-based decision tree,” IEEE Transactions on Industry [39] L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, Classification Applications, vol. 51, no. 2, pp. 1249–1258, March 2015. and regression trees. CRC Press, 1984. [17] N. Perera and A. D. Rajapakse, “Recognition of fault transients using [40] J. R. Quinlan, C4.5: Programs for Machine Learning. San Francisco, a probabilistic neural-network classifier,” IEEE Transactions on Power CA, USA: Morgan Kaufmann Publishers Inc., 1993. Delivery, vol. 26, no. 1, pp. 410–419, Jan 2011. [41] L. Breiman, “Random forests,” Machine Learning, vol. 45, pp. 5–32, [18] S. Santoso, E. J. Powers, W. M. Grady, and A. C. Parsons, “Power Oct 2001. quality disturbance waveform recognition using wavelet-based neural [42] G. Louppe, “Understanding random forests: From theory to practice,” classifier. i. theoretical foundation,” IEEE Transactions on Power Deliv- Ph.D. dissertation, University of Lige, Belgium, Oct 2014. ery, vol. 15, no. 1, pp. 222–228, Jan 2000. [43] J. H. Friedman, “Greedy function approximation: A gradient boosting [19] T. Zhong, S. Zhang, G. Cai, Y. Li, B. Yang, and Y. Chen, “Power machine.” Ann. Statist., vol. 29, no. 5, pp. 1189–1232, 10 2001. quality disturbance recognition based on multiresolution s-transform and [44] L. Mason, J. Baxter, P. Bartlett, and M. Frean, “Boosting algorithms as decision tree,” IEEE Access, vol. 7, pp. 88 380–88 392, 2019. gradient descent,” in Proceedings of the 12th International Conference [20] U. Singh and S. N. Singh, “Optimal feature selection via nsga-ii on Neural Information Processing Systems, ser. NIPS’99. Cambridge, for power quality disturbances classification,” IEEE Transactions on MA, USA: MIT Press, 1999, pp. 512–518. Industrial Informatics, vol. 14, no. 7, pp. 2994–3002, July 2018. [45] R. Kohavi, “A study of cross-validation and bootstrap for accuracy [21] P. K. Ray, S. R. Mohanty, N. Kishor, and J. P. S. Catalo, “Optimal feature estimation and model selection,” in Proc. 14th Int. Joint Conf. Artif. and decision tree-based classification of power quality disturbances Intell., 1995, p. 11371143. in distributed generation systems,” IEEE Transactions on Sustainable [46] K. H. Brodersen, C. S. Ong, K. E. Stephan, and J. M. Buhmann, Energy, vol. 5, no. 1, pp. 200–208, Jan 2014. “The balanced accuracy and its posterior distribution,” in 2010 20th [22] S. K. Bhasker, P. K. Bera, V. Kumar, and M. Tripathy, “Differential International Conference on Pattern Recognition, Aug 2010. protection of indirect symmetrical phase shift transformer and internal [47] L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” faults classification using wavelet and ann,” in TENCON 2015 - 2015 Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008. IEEE Region 10 Conference, Nov 2015, pp. 1–6. [48] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: [23] P. K. Bera, R. Kumar, and C. Isik, “Identification of internal faults in Synthetic minority over-sampling technique,” J. Artif. Int. Res., vol. 16, indirect symmetrical phase shift transformers using ensemble learning,” no. 1, p. 321357, Jun. 2002. in 2018 IEEE International Symposium on Signal Processing and [49] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal Information Technology (ISSPIT), Dec 2018, pp. 1–6. of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [24] P. K. Bera, C. Isik, and V. Kumar, “Transients and faults in power transformers and phase angle regulators (dataset),” 2020. [Online]. Available: http://dx.doi.org/10.21227/1d1w-q940 [25] J. McIver, “Phase shifting transformers principles, design aspects and operation,” Siemens Energy, Inc, Mar 2015. [26] S. Kulkarni and S. Khaparde, Transformer Engineering. Boca Raton: CRC Press, 2013. [27] H. Bronzeado and R. Yacamini, “Phenomenon of sympathetic interaction between transformers caused by inrush transients,” IEE Proceedings - Science, Measurement and Technology, vol. 142, no. 4, July 1995. [28] T. Zheng, J. Gu, S. F. Huang, F. Guo, and V. Terzija, “A new algorithm to avoid maloperation of transformer differential protection in substations with an inner bridge connection,” IEEE Transactions on Power Delivery, vol. 27, no. 3, pp. 1178–1185, July 2012.

Journal

Electrical Engineering and Systems SciencearXiv (Cornell University)

Published: Apr 13, 2020

There are no references for this article.