Qualified Two-Hybrid Techniques by DWT Output to Predict Fault Location

Azriyenni Azhari Zakri, Syukri Darmawan, Sandy Ahmad, Mohd Wazir Mustafa, Jafaru Usman Department of Electrical Engineering, Faculty of Engineering, Universitas Riau, Pekanbaru, Indonesia School of Electrical Engineering, Faculty of Engineering, Universiti Teknologi Malaysia, Johor Bahru, Malaysia Department of Electrical & Electronic Engineering, Faculty of Engineering, University of Maiduguri Borno State, Nigeria


INTRODUCTION
The transmission system is the electrical power energy element to transfer electrical power from remote generation location to distribution systems. A fault that occurs in the transmission system may disturb the power source in the delivery scheme which may shut down the customer stations [1]. The fault of a shortcircuit might happen either balanced or unbalanced [2]. Unbalanced fault involves single-phase fault to ground, dual-phase, and dual phases to ground, while a balanced fault is a three-phase fault. Some categories of shortcircuit fault have different appearances in current and voltage. Accordingly, faults that occur on the transmission system need to be identified and categorized correctly to make it easier to resolve. Implementation of Artificial Neural Networks (ANN) has been performed for the fault category as well as the fault location in the transmission line. According to Hessine's research result, using modular ANN can shorten the duration of the training and improve the accuracy of ANN for the types of faults as well as estimating the fault location [3]. It has also proposed the application of DWT and support vector machine (SVM) using a sample frequency of 50 kHz. Summing the absolute numbers of the detailed derivative at levels 8 and 9 is used with SVM. The results show that this technique is applicable in parallel transmission line [4,5]. The demonstration of a mixture performance for the grouping of short circuit faults is available from this research. The work used a combination of Discrete Wavelet Transformation (DWT) and Support Vector Machine (SVM) techniques. Wavelet Transforms (WT) compensate the Fourier Transforms (FT), which has incomplete data on the frequency domain. Signal gained from wavelet transforms is devoted to frequency and time domains [6]. However, WT is generally indecisive for the category of fault in the transmission link [4]. Essentially, DWT is more extensive than Continuous Wavelet Transformation (CWT) to transform voltage and current signal to frequency domain [8,9]. Previous research applied WT for the grouping of fault in equivalent transmission system. They operated DWT to take wavelet level 8 at a sample frequency of 12.5 kHz. The researchers successfully classified faults in an equivalent transmission system using the ANN method [10].

RESEARCH METHOD 2.1. DWT
DWT can analyse various categories of fault with data gained from frequency and time domain. The DWT is beneficial in identifying numerous failures as it is sensitive to signal irregularities [13]. The wavelet transformation can be separated into two kinds: DWT and Continuous Wavelet Transforms (CWT). The CWT is the amount of indicators overtime increased by the scale and locus of wavelet utility, as in equation (1) and (2)  CWT transforms a shifted and scaled copy of signal in the basic wavelet. As a redundant transform, CWT has wavelets overlapped and require larger computation resources to compute and store the coefficients than DWT. DWT is used to analyse wave indications correctly. The DWT determination employs the Mallat algorithm. DWT is separated in dual wave indicators, i.e., filter techniques and down-sample processes. The high-pass filter technique produces different signs through high frequency. The low-pass filters provide  ISSN: 2089-3272 IJEEI, Vol.8, No. 4, December 2020: 806 -817 808 dissimilar signs through low frequency. Subsequently, the determination of the number is through downsampling operation. Thus, it only proceeds half of each data gotten previously [13]. Figure 1 shows the diagram of DWT operation. High frequency indicators remain particulars signal, and signs using low frequencies termed rough calculation. A procedure is decayed through recapitulation with a view of the information gained earlier. It will be decayed to generate separate estimates and element information. Figure 2 demonstrate the decompose repetition procedure to go for an initial wave indication. The situation will be expected to complete by summing the approximation information and element information.

Figure 2. Repetition of DWT decomposition
In the additional decomposition procedure, the indicator can be divided on numerous low-resolution mechanisms. Therefore, the minimal frequency element filter procedure will be pursued continuously. Table 1 designates the decay of wavelet stage 9 for a cycle when the fault is sampled with frequency of 50 kHz [12].

SVM
SVM is a structure that uses a theoretical linear plane in a high-dimensional interplanetary, also proficient through procedures established on bias knowledge optimisation concept. The principal determination of this procedure creates optimum separating hyperplane, which purpose optimal parting that can continue the arrangement in best way. Information on the borderline field is known as a support-vector. Figure 3 shows a couple of lines separating two record groups. Also, it demonstrates the delimiter ground termed support-vector. Two categories are distributed as a result of the equivalent jumping planes. The arrangement of the midpoint is synchronised toward the Euclidean distance. A primary delimiter field bounds the first class after the next demarcating zone over the next period, the formulation articulated in equation (3) [16]: Where w is a standard plane, and b in another field of a usual plane. The bordering value between the bounding planes is following the formulation of space to the midpoint [17]: The value of this boundary will make the best use of two parameters; by increasing b and w, its determination significantly increased. A constraint in Equation (3) is a scaling constraint by re-scaling b and w. Hence, in order to exploit m equals to minimise ||w|| 2 . Two boundary planes on Equation (3) can be articulated using the following equation [17]: The greatest partition of the plane is the most substantial boundary which is expressed into constraint optimisation [18]: The SVM of formula was used to classify linear data using a variable called soft margin hyper-plane. Then, the best formula of separator become [17]: with, (x i w + ) ≥ 1 − , ≥ 0 Parameter C is used to determine inaccuracies in data grouping, and its value is assigned by user. The role of C is to minimize training inaccuracies and reduce complexity. The parameter C of SVM is customarily termed box constraint [17]. The proportion of accuracy is defined as [9]:

RESULTS AND DISCUSSION
The data model is run between 0.00002 s in separated model. When the fault happens, a single round current surge is recycled as DWT input through the number of models 1 / (10-5 * 50) = 1000 example numbers. The constraints used for training and test data are noted in Table 6. The fault resistance is 10 Ω. The quantity of models accomplished to gain the training numbers is 6930 data (10 faults * 11 distance of faults * 7 resistance of faults * 9 FIA). The number of models accomplished to acquire the test data is 4900 data (10 faults * 10 distance of faults * 7 resistance of faults * 7 FIA). The current measured on the Bus KP becomes reference for the phase to ground (AG) of a short circuit. The fault distances through different points through 10% increments, resistance of fault 35 Ω, then FIA 50° that can be located is shown in Figure 6.  Figure 6 The fault current simulation at bus KP for one phase to ground

Processing Data
The fault current occured in one cycle (post-fault) was processed using mother wavelet for Daubechies level 4 (DB 4) at level 9 (D9). DB 4 is beneficial for the investigation of momentary indicators [6,7,10]. Figure  7 shows outcomes of signal processing using DWT procured at level 8 elements (D8) as well as element level 9 (D9). The value on D8 and D9 of instantaneous currents for a cycle after the fault happens were calculated by measuring the Root Mean Square (RMS). This RMS value was then used as input for the SVM method. Moreover, the ground current is calculated from the sum of currents of each phase divided by three as following [2,19]:

Estimation of Short Circuit Fault
The ANFIS structures has been trained to estimate the A-G fault location. Table 7 demonstrates that the ANFIS structure test results have the smallest RMSE and MSE values using the Gaussian membership function (FLAG1) with values of 0.027835 and 0.022561. The ANFIS structure that has the largest RMSE and MSE values in the ANFIS structure occurs in the generalised bell (FLAG2) membership function with a value of 0.060488 and 0.043865.  Table 8 is the result of error calculation for the estimated location of the fault for each type of fault that has been replicated. The fault that has been computer-generated includes; A-G, B-G, C-G, A-B, A-C, B-C, A-B-G, A-C-G, B-C-G, then A-B-C. From the test, ANFIS results for estimation fault location with the smallest average error testing is generated by the ANFIS is 6.05*10-4 % (A-G), while the most significant average error is produced by the ANFIS estimated fault location is 2.9*10-2 % (A-C). The RMS element of D8 is also D9 for separate phase then ground current is recycled as contribution SVM to acquire a hyper-plane utility. Constraint SVM is involved in the limitation box constraint C significance is 1, and the kernel measure significance is 0.35. Hyper-plane established from SVM exercise expending the contribution value of exercise numbers. The SVM for approximating the fault location is given six input, namely; the RMS values of the D8 also D9 coefficients of the current signal post fault in each phase. Data of the SVM estimated fault location can be sketched as following: a) two RMS coefficient details of the phase current signal A b) two RMS coefficient details of the phase current signal B c) two RMS coefficient details of the phase current signal C The SVM used is a Gaussian kernel type, the data for estimating fault locations are ,  and box constraints (C). Table 9 is the output on the RMSE and MAE values, where the value is error limit with a fixed input value of 10 -3 . The gamma parameter and box constraint values are varied. It aims to get a small RMSE result and a short training duration. The constraint  has values 10 -2 and 10 -1 , while the parameter C with values 10 0 , 10 1 , 10 2 , 10 3 , and 10 4 . The training simulations were conducted ten times.  The best SVM models for training and testing are selected with parameters  = 0.01 and C = 1000 while the preferred SVM model for the estimated fault location is with parameters of  = 0.01 and C = 10000. Figure  8 shows the result of the SVM testing of the estimated fault location by calculating the error value of the SVM model for determining the fault location.
The test result of fault classification and fault location shows the error percentage value on each ANFIS. ANFIS fault type has determined the fault in each phase and ground correctly according to the nature of fault that occurs. Accuracy of ANFIS fault classification is 100% without any errors. The error value of ANFIS results for estimating each fault's location has been calculated, as shown in Table 9. The variation of SVM parameters for fault location estimation is presented in Table 10. It shows the average number of iterations needed to achieve the epsilon () 10 -3 target and the acquisition of RMSE and MSE values during training and testing on each SVM model when an A-G fault occurs.  Figure 9 shows the simulation results that the minimum and maximum error range of hybrid technique one is from 2*10 -6 % to 0.107867%, while for hybrid technique two is from 8*10 -6 % to 0.194868%. The comparison results of the minimum and the maximum percentage of errors and the average percentage of errors for hybrid techniques one and hybrid techniques two are to prove that between these two hybrid techniques does not have significant differences in error results.

CONCLUSION
In this study, the estimation of the fault location in the electric power transmission system has been carried out by simulating the classification of short circuit faults. The DWT has been applied to analyze the type of interference obtained from the frequency domain and time domain, therefore, DWT is very useful in detecting and processing various interference data. The results of the DWT were carried out using two techniques, namely ANFIS and SVM with training and test variables; resistance disturbance and FIA. Furthermore, the significance of the RMS value for levels D8 and D9 for the training and test data has a very small error value.
In the SVM technique there is an alternative hyperplane dividing line between two classes to find the maximum point, the closest pattern as a support vector. Then, ANFIS simulations have been trained to estimate the location of the same disturbances as was done in the SVM (A-G fault) technique. The ANFIS structure test has the smallest RMSE and MSE values using the Gaussian membership function (FLAG1) with values of 2.7835 * 10 -2 and 2.2561 * 10 -2 . The ANFIS structure has the largest RMSE and MSE values in the ANFIS structure found in the generalized bell membership function (FLAG2), with a magnitude of 6.0488 * 10 -2 and at 4.3865 * 10 -2 , but the values are still within tolerance limits. The design of the FL16 SVM model as an SVM approximation model was selected for each type of disturbance with a value of 0.01 and a value of 10000 resulting in the final average test error, specifically for the type of disturbance ABC 2.281 * 10 -3 % RMSE is 2.3 * 10 -3 , and MAE is 1.5 * 10 -3 . The comparison of the accuracy of the simulation results is shown in the form of percentage errors. It is found that hybrid technique one and two have a difference but insignificant in results, within the standard of fault tolerance. Finally, the simulation results of these two hybrid techniques shows that both hybrid techniques can be applied to predict the location of the disturbance with a satisfactory level of accuracy.