Fault detection,classification in multiterminal HVDC transmission system with MC-SVM

Abstract

Due to the recent advancement in power electronics devices in past few decades, HVDC system became mature but still has some protection issues, like tripping of the circuit breaker for a temporary fault as to load changes. Therefore, in this paper, a scheme of complete protection for fast, and accurate classification and detection of a fault in HVDC transmission line using support vector machine (SVM) is presented. In the proposed scheme, ac and dc side voltage and current at each converter station are measured and treated as the input of SVM binary classifier. For classification of fault, SVM module with multi-classification feature is used. For the normalization purposes of the signals, the standard deviation is used over half cycle before and after the occurrence of the fault. Features have been extracted through wavelet transform of predefined function for detection and classification of a fault. The proposed scheme is easy to use as it requires only one end data and a standard deviation over one cycle data.

Keywords

Support vector machines (SVM)hyper plane multiclass SVM multiterminal HVDC regression confusion matrix receiver operating characteristic (ROC)

1 Introduction

HVDC system has multiple advantages like improved stability, flexible control, and greater power transfer capability, reduced losses and lesser cost in comparison with HVAC system. HVDC is economical in the contest with HVAC above a distance called break-even distance is brought down to almost 500 km due to recent advancement in the power electronics area in past few years. In recent years efficiency of ac-dc or dc-ac converters is improved which further improves the efficiency of the entire HVDC transmission system.

In electrical power system generated electrical power is carried to distribution system by transmission lines so the performance of transmission line improves the performance of whole power system. The protection scheme of transmission system must be very efficient because it improves the reliability and efficiency of the entire transmission system. In DC line series inductance is less in comparison with AC line. Therefore, damage caused by fault can be reduced by detecting and isolating the faulty section as soon as possible. This requires a detector and locator algorithm which should be fast andaccurate.

In the Field of HVAC transmission system lot of research work and publication has been done but in case of HVDC transmission system, much more research work is needed. The methods implemented in HVAC system are not equally applicative for HVDC transmission system. So some new strategies are implemented for HVDC system some of them are briefly described in below paragraph.

Methods based on traveling wave are employed in HVDC transmission system in the present-day primary protection scheme. Whereas minimum dc voltage and dc line differential protection methods are employed for back up protection. The traveling wave algorithms are based on the estimation time consumed by the fault current wave to pass around the transmission line [1 –3]. For synchronization of measurements, the global positioning system is generally employed in the methods which require data of both terminals. These methods are fast and accurate but less reliable [4, 5]. The reliability of these methods depend upon the identification accuracy of surge arrival time. In [6] describe traveling wave based technique which works with un-synchronized two end data but it using very high sampling frequency of 1 MHz. Fault detection technique for multi-terminal HVDC cable system with transient sheath voltage only but it does not provide classification of different types fault in wind farm connectedsystem [7].

In artificial neural network (ANN) based paper used two end or one end data for training and testing purpose it provide good result in classification, few paper use it for location of fault also. Most of paper use radial biased function (RBF) of ANN to solve the problems [8]. The method based on support vector machine requires single end data and less sampling frequency [9]. SVM binary classifier needs data for half cycle before fault and half cycle data of after fault [10]. The wavelet transforms based methods are also accurate and fast methods which use handshaking method [11]. The fault current of dc side is transmitted to AC side in handshaking methods and then ac CB is operated according to this AC current. Wavelet transform observes that how the signal change over time and the signal are decomposed into each frequency band over a range of frequency at the same time. Decomposition of the signal is done by multi-resolution signal decomposition(MSD).

This paper proposed a method that requires very less number of data for identification and classification of the different type of DC faults. It provides a much accurate result with the variation of sampling frequency because its feature vector removes stander deviation of data in time and frequency domain. Some of the important highlights of this purposed algorithms as,

The method aims to classify all types of faults.

The first step of the method is obtaining fault section.

Feature selection is achieved with the help of Principle Component analysis (PCA) method that provides easy selection of dominant feature.

Multiclass SVM algorithm is utilized for precise multiclass fault classification.

The rest of paper organized as follows. Section 2, describe basic of Support Vector Machine technique and its modification. Section 3, provide overview of test system used for validation of proposed algorithms and analysis of the experimental results. Finally, Section 4, provide a brief conclusion.

2 Support vector machine

In this work, SVM is used to the classification of four different types of fault insisted in HVDC transmission line. It is much accurate and robust and easy to apply. It is binary classifier in nature but modification is possible for higher order classification. It’s actually estimated an optimal hyperplane that suitable for the dataset to separate in their respective classes. It is basically linear classifier but also classifies non-linear data with kernel trick. It is very efficient and powerful classifier. Basic equations of this classifier given in next sections.

2.1 SVM as binary classifier

Our training data consist of N pairs (x_i, y_i) i = 1, 2, 3…. N with x ∈ Rⁿ and, define a hyperplaneby $f (x) = ω^{T} x_{i} + d = 0$ (1)

Where d is a unit vector ∥d ∥ = 1, A classification rule induced by f (x)is $G (x) = sign (ω^{T} x_{i} + d)$ (2)

Where f(x) in Equation (1), denotes signed distance between point x and hyperplane f (x) = ω^Tx_i + d = 0. Here, classes are idiosyncratic, a function could be found as f (x) = w^Tx_i + d = 0 with y_if (x_i) > 0 ∀ i. Hence, hyperplane could be obtained that has a high margin between training points of class 1 & – 1 (as shown in Fig. 1) [12, 13].

Fig.1

SVM Classification.

$M a x M f o r y_{i} (ω^{T} x_{i} + d) \geq M i = 1, ....., N$ (3)

The bandwidth in the figure is M units away from the optimal hyperplane on either side. Therefore total bandwidth is 2 M units wider that is define as margin of hyperplane. It shows that, this problem can be more appropriately rephrased as, $min | d | for y_{i} (ω^{T} x_{i} + d) \geq 1$ (4)

2.2 Multiclass SVM

Basically, SVMs are two class classifiers but with appropriate modification, it can be used in the multiclass application. The technique commonly used is to build |C| a one-versus-rest classifier. It is also referred as one-versus-all (OVA). This technique chooses the class which the data item with the largest margin. Another scheme which can be used is that choose that class from a set of already built OVA which is selected by most no. of classifiers. While this involves building |C| (|C| - 1)/2 classifiers which reduces the time for training classifiers because of the use of much smaller training dataset [14].

However, the accuracy of the above-mentioned scheme is quite poor for problem-solving in multiclass SVM. Overcome for this problem is the construction of multiclass SVMs.

2.3 SVM Regression

SVM regression is first proposed in 1995 by Drucker [15]. It has to find a function f (x) which has the largest margin of ε from the hyperplane y_i for all the given training data. Our training data consist of N pairs (x_i, y_i) i = 1, 2, 3…. N with x ∈ Rⁿ and y_i∈ { + 1, - 1 } is optimal hyperplane, then standard form of SVM regression is given, $min \frac{1}{2} ω^{T} ω + c \sum_{i = 1}^{n} δ_{i} + c \sum_{i = 1}^{n} δ_{i}^{*}$ (5) $[ω^{T} φ (x_{i}) + d - y_{i}] \leq ε + δ_{i}$ (6) $[y_{i} - ω^{T} ψ (x_{i}) - d] \leq ε + δ_{i}^{*}$ (7) $δ_{i}, δ_{i}^{*} \geq 0, i = 1, . . . . ., N$ (8)

The performance of SVM regression controlled by the selecting of parameters i.e. c, ε and appropriate kernel function. With the fact that depends on all the three parameters mentioned above therefore selection of optimal parameter is a complex problem itself. SVM parameters are treated as user-defined inputs by the SVM regression implementation software. Selection of kernel function usually depends upon the knowledge of application and selection of kernel function also reflect the distribution of input training data. Some Commonly used kernel function is as follows:

(a) Polynomial $k (x, y) = {(〈 x, y 〉 + θ)}^{p}$ (9)

(b) Gaussian radial basis function (RBF) $k (x, y) = exp (- \frac{{∥ x - y ∥}^{2}}{γ})$ (10)

Among the above given three functions, Gaussian RBF two variable can be varied i.e. c and γ [16]. Gamma parameter decides closeness of a single training data. Misclassification of training data is carried by parameter c against the simplicity of decision surface. The low value of parameter c makes smooth decision surface and leads to poor accuracy while a high value of c provide classifier to selecting more no of samples as support vectors for good accuracy [17].

3 Test system for simulation

For algorithm verification, a Multi-terminal HVDC transmission test system has been used, as shown Fig. 2. The VSC Transmission model available with PSCAD/EMTDC module was originally modeled for two-terminal and is modified into a multiterminal transmission line. This transmission line is now connected to a cable having the length of 900 km and 600 km. These cables have been used to create fault at every 25 km. With this condition, the dataset is prepared for training and testing the supervised SVM classification algorithm. The cable structure along with their configuration is given in Fig. 3. The detailed information of HVDC system parameter is given in Table 1.

Fig.2

Schematic diagram of Multiterminal HVDC model.

Fig.3

Configuration of HVDC cable.

Table 1

System specification

Particulars	Description
Voltage source-1 at Terminal-1	420 kV,100MVA, 60 Hz
Voltage source-2 at Terminal-2	420 kV,100MVA, 50 Hz
Voltage source-3 at Terminal-3	5000 kV,100MVA, 50 Hz
Rectifier section	230 kV, IGBT, on resistance = 0.005 Ω, off resistance = 100 M Ω, breakdown voltage = 100 MV, reverse withstand voltage = 100 MV, PWM control
Inverter section	230 kV, IGBT, on resistance = 0.005 Ω, off resistance = 100 M Ω, breakdown voltage = 100 MV, reverse withstand voltage = 100 MV, PWM control
Transformer	420 kV/230 kV, Y- Y, solid grounded- resistive grounded with 1 M Ω, 1500MVA
High Pass Filter parameter	Resistance = 1 Ω, Inductance = 0.1 H, Capacitance = 5μF
Smoothing Reactor in series	0.0724 H
Transmission cable connect Terminal 1-2	±500 kv, 900 km
Transmission cable connect Terminal 1-2	±500 kv, 600 km

6-pulse IGBT based on VSC control technique considering AC side transmission characteristics gives up to 300 MW power to the DC side (500 kV cable) of the transmission line. Rectifier/Inverter controls both real & imaginary power flows. In this, PWM control technique is used with carrier frequency up to 33 times of the fundamental frequency. This control technique uses AC voltage phase angle difference measure at sending end and receiving end. Generator exciter is used to control the sending end voltages.

3.1 Result and discussion

Purposed algorithms are based on the support vector machine technique using principal component analysis for feature selection [18]. The simulation of the tested system is carried out in PSCAD/EMTP transient program at a sampling rate of 4.80 kHz (i.e. 81 sample/cycle at 60 Hz). The PSCAD/EMTP generated fault case data loaded into the Matlab software. From this dataset calculate different feature and after that use feature selection algorithms that reduce this dataset on the basis of dominating feature. The step of this purposed algorithms are as:

Design multilevel HVDC system in PSCAD transient program.

Measure voltage, current and power data at a different node of AC and DC transmission section.

Calculate different features on the basis of feature function (i.e. Mean, Mode, Median, Weighted mean, standard deviation, skewness, correlation coefficient, Variance, Covariance, Signal energy, etc.) in both time and frequency domain.

Select dominant feature on the basic of PCA criteria and reduce dataset.

Use this reduced dataset for training and testing of multiclass supervised support vector machine classifier.

For each fault case, initiate fault at every 25 km length in both the transmission line. In first transmission line i.e. cable, terminal1-2 is 600 km long and second transmission line i.e. cable terminal1-3 is 900 km in length. So, it is initiated 23 events for each fault and collected 23 sample data for each fault (i.e. LG and LL) from the first transmission line and similarly get 35 samples for each fault in the second transmission line. In present work, it is supposed that a short circuit fault has occurred the system. Each fault initiated at 1 s and exists for 0.1 s (approx. 5 cycles). The collected data for every sample is having the sampling frequency of 4.8 kHz. Once the fault occurs to the system, there would be immediate fall in line voltage. In rectifier side, there would be rise in DC current and there would be fall in inverter side DC current. Therefore, in order to maintain these currents at some predefined level, a control circuit is required. This predefined value is selected in accordance with the DC voltage. During the fault event, this DC voltage will tend to reduce to a lower value than its threshold value. Therefore, this will bring down the line current to a lower value. It is advisable to limit this current to +ve non-zero value. If it goes beyond this value the reactive power demand on ac part of the system reduced thereby it will prevent the system torestore [19].

LL-Fault is initiated in DC line at connecting terminal1-3 at 525 km far from terminal-1. Figure 4 shows the generated side voltage and current that is not affected by the fault. Figure 5 shows the AC voltage and current of each terminal near converter end voltage and current variation in terminal2 and terminal-3 are more comparable to terminal-1. It is also observed that current and voltage has harmonic contains due to fault present. Due to fault present in system DC line current is much more disturbed but DC terminal voltage, not a much-disturbed line -2 and line-3 voltage somehow fall. The current of the line connected terminal1-2 rise due to inverter action and line connected terminal1-3 fall due to rectifier action is shown in Fig. 6. Due to VSC technique, some voltage of DC side maintained constant.

Fig.4

Zoomed Signal of High Voltage Side. (a) Voltage. (b) Current.

Fig.5

Zoomed signal of Converter Terminal End. (a) Rms Voltage and Current at Terminal-1. (b) Rms Voltage and Current at Terminal-2. (c) Rms Voltage and Current at Terminal-3.

Fig.6

Zoomed signal of DC Side. (a) Voltage. (b) Current.

For classification algorithms, used 10-fold cross-validation, cubic kernel function, 70 data used for training purpose and 30 for testing purpose. Data contain 116 instant or sample of four different class, 646 feature used after selection with PCA algorithms for training and testing purpose. Obsessed result obtains with this algorithm shown in Figs. 7 and 8. In Fig.7 diagonal element show, the correct classify number or percentage and upper and lower triangular element shows false classified element orpercentage.

Fig.7

(a) Confusion Matrix with no of sample. (b) Confusion Matrix with TPR/NPR. (c) Confusion Matrix with PPD/FDR for SVM classifier.

A much better way to assess classifiers is by accounting for their performance over the full operating range by accounting for the confidence of predictions. This is done, for example, using the (area under) receiver operating characteristic (ROC) or precision-recall (PR) curves. Observed result for all class are shown in Fig. 8. Overall fault classification accuracy is 99.23 achived shown in Figs. 7 and 8. From Fig. 7 (a) seen that only one sample of LG fault of cable connected terminal1-3 is misclassified with LG fault of cable connected terminal1-2 sample and that class has accuracy 97 except this all other class of fault classifies with 100 accuracy with this proposed algorithms. The Same result is verified with receiver operating characteristic (ROC) curves. Confusion matrix description is given in Table 2. Basic equation behind confusion matrix is shown in Equations (12), (13)and (14).

Fig.8

(a) ROC of LG fault in cable connected terminal 1-2. (b) ROC of LG fault in cable connected terminal 1–3. (c) ROC of LL fault in cable connected terminal 1-2. (d) ROC of LL fault in cable connected terminal 1–3.

Where,

$True Positive Rate = \frac{True Classification}{\sum (True Classification + False Classification)}$ (12)

$True Negative Rate = \frac{False Classification}{\sum (True Classification + False Classification)}$ (13)

$Accuracy = \frac{True Classification}{Total no . of Cases} \times 100$ (14)

Table 2

Percentage classification accuracy

Classes	True Positive Rate (Recall) (%)	True Negative Rate (%)	Accuracy (%)
LG fault at cable T1-2	100	0	99.285
LG fault at cable T1-3	97.143	2.857
LG fault at cable T1-2	100	0
LG fault at cable T1-3	100	0

4 Conclusion

This paper proposed a scheme for detection and classification of a fault and tested on VSC based HVDC transmission system. The proposed scheme uses SVM classifier that classifies the different type of faults and provides a much accurate result with less number of observed data. The accuracy and applicability of the scheme for classification of different faults in DC system have been clearly indicated by the results obtained. The accuracy of the scheme is 99.285 for a system under investigation with four different class of faults. The proposed scheme is also suitable for all type of fault detection, and classification in DC and AC transmission as well as distribution system.

References

Suonan

, Gao

, Song

, Jiao

and Kang

, A novel fault-location method for HVDC transmission lines, IEEE Transaction power delivery25(2) (2010), 1203–1209.

Nanayakkara

, Rajapakse

and Wachal

, Traveling-wave-based line fault location in star-connected multiterminal HVDC systems, 2013 IEEE Power Energy Society General Meeting, 2013, p. 1.

Dewe

M.B.

, Sankar

and Arrillaga

, The application of satellite time references to HVDC fault location, IEEE Transactions on Power Delivery8(3) (1993), 1295–1302.

Shang

, Herold

, Jaeger

, Krebs

and Kumar

, High-speed fault identification and protection for HVDC line using wavelet technique,, 2001 IEEE Porto Power Tech Proceedings (Cat. No.01EX502)3 (2001), 5.

Chen

, Xu

and Li

, A traveling wave based fault locating system for HVDC transmission lines, 2006 International Conference on Power System Technology, 2006, pp. 1–4.

Yuangsheng

, Gang

and Haifeng

, Time-domain fault-location method on HVDC transmission lines under unsynchronized twoend measurement and uncertain line parameters, IEEE Transaction on Power Delivery30 (2015), 1031–1038.

Niaki

S.H.A.

, Hosseini

S.M.

and Abdoos

A.K.

, Fault detection of HVDC cable in multi-terminal offshore wind farms using transient sheath voltage, IET Renewable Power Generation11(13) (2017), 1707–1713.

Cui

, Tu

, et al.HVDC transmission line fault localization base on RBF neural network with wavelet packet decomposition, 12th International Conference on Service Systems and Service Management (ICSSSM), 2015, pp. 1–4.

Johnson

J.M.

and Yadav

, Complete protection scheme for fault detection, classification and location estimation in HVDC transmission lines using support vector machines, IET Science, Measurement Technology11(3) (2017), 279–287.

10.

Yeap

Y.M.

and Ukil

, Wavelet based fault analysis in HVDC system, IECON – Conference of the IEEE Industrial Electronics Society, (2014), 2472–2478.

11.

Cristianini

, and Shawe-Taylor , An introduction to support vector machines: And other kernel-based learning methods, Cambridge University Press, New York, U.S.A., 2000.

12.

Hykin

, A comprehensive foundation,, Neural Network 2004, (2004), 340–348.

13.

Manning

C.D.

, Raghavan

and Schutze

, Introduction to information retrieval, Cambridge University Press, 2008, 2008, pp. 327–330.

14.

Drucker

, Burges

C.J.

and Kaufman

, Support vector regression machines, Advances in Neural Network Information Processing System9 (1997), 155–161.

15.

Cherkassky

and Ma

, Practical selection Of SVM parameter and noise estimation For SVM Regression, Advance Neural Network, 20042004, pp. 113–126.

16.

Pedregosa

, Varoquaux

and Gramfort

, RBF SVM parameters, [online]. Available at http://scikit-learn.org/stable/auto_examples/svm/plot_rbf_parameters.html.

17.

Ilin

and Raiko

, Practical approaches to principal component analysis in the presence of missing values,, Proc of IEEE Int Conf on Power System Technology11 (2010), 1957–2000.

18.

Naidoo

and Ijumba

N.M.

, HVDC line protection for the proposed future HVDC system,, Proc of IEEE Int Conf on Power System Technology PowerCon, 20042 (2004), 1327–1332.

19.

Farhangi

, The path of the smart grid, IEEE Power Energy Mag8(1) (2010), 18–28.