Improved voting model based fault detection and classification in distributed generation power systems

As a result of the increase in energy demand and limited fossil fuels in the current electrical power system, as well as the liberalization of the energy market, and the effects of economic and environmental concerns, the idea of distributed generation, which includes renewable and non-renewable energy sources, has gained popularity. With the inclusion of Distributed Energy Generation Resources (DEGR) in the power system, transition from passive state to active state has been occurred. With the transition to the active state, it naturally changed the direction of current in the power system from one-direction to bi-direction. The integration of renewable energy sources such as solar panels and wind turbines into the distribution grid has resulted in bi-directional power flow and a change in the amplitude of the fault current. The fault current can be affected from the capacity, type, number, and location of DEGRs. On the other hand, renewable DEGRs have an intermittent power generation characteristic, thus leading to potential load demand uncertainty on the consumer side. In such a power distribution system with DEGRs, complexity in load sharing, protection, and controlling is possible. The variation in the level of the fault current, together with the bi-directional in the power flow, led to the loss of coordination between the protection devices and, as a result, the reliability to the power system. Thus, it is inevitable that problems will arise in fault detection and classification. Many studies have been conducted on the use of machine learning models for fault detection and classification. However, one of the most challenging problems in classification is the lack of a general classifier that can correctly learn the data set for each problem. In addition, it is not possible for the classifiers to learn the entire data set. Different classifiers may show different performance or error proneness in different subspaces of a dataset. Therefore, the composing of a near-optimal classifier model for any problem can be achieved with the ensemble learning model. Regarding the voting models, which is a method of the ensemble learning model in the literature, the results of the probabilistic contributions and performances of the classes are not taken into account in the weighted voting model. On the other hand, in the unweighted voting models, especially in the probabilistic based voting model, the performance of the classifiers regarding the training phase is not taken into account. According to the studies in the literature, to the best of our knowledge, there has been no study using the probabilistic weighted voting model consisting of multiple classifiers for the fault detection and classification. Encouraged by the above-explaination, a fault detection and classification scheme (FDCS) has been proposed in this thesis concept. FDCS is based a probabilistic weighted voting model (PWVM) includes multiple diverse learning models. The contribution of the proposed OAOM model compared to other voting models is that the sum of probabilities based voting method, which is one of the probabilistic-based voting methods of unweighted voting models, has turned into the sum of weighted probabilities and the weights are formed by choosing from the weight matrix. The working principle of proposed model relies on training performance-based class contributions and per-class possibilities with test set. The validation metric (mean absolute error) and F-score obtained in the training stage can be improved the model's prediction power and imbalanced class distribution. A class label of any sample in test set is achieved with the training stages, and probabilistic contributions achieved from classes of classifiers. To detect sudden changes in the signal, samples need to be interpreted in detail structurally. Therefore, feature extraction is required to extract more understandable and simplified information representing the core of the signal. The signal processing method used for feature extraction is discrete wavelet transform in this thesis study. Wavelet transform ensures definitive properties of pure signal in the multiresolution time-frequency domain. The used mother wavelet function and wavelet levels are db10 ve level 9-10. The frequency components of the 9th wavelet decomposition level include the 11th and 13th harmonic components, while the frequency components of the 10th wavelet decomposition level include the 5th, 7th, and 9th harmonic components. In this sense, 9 and 10 have been chosen as wavelet decomposition levels. Different types of faults (single-phase ground faults; AG, BG, CG, two-phase faults; AB, BC, CA, two-phase ground faults; ABG, BCG, CAG, three-phase ground fault ABCG) were applied to the model of power distribution system with DEGRs created within the scope of this thesis. The descriptive feature functions are applied to detail coefficients pertaining to wavelet decomposition levels related to these fault states. In this thesis, basically 4 different features were used. However, since there are 2 different wavelet levels, there are 8 features and 3 different phases, a total of 24 features are extracted. These features are absolute maximum of detail coefficients, absolute average of detail coefficients, variance, and entropy. A comprehensive comparative performance analysis was conducted to evaluate the performance of the proposed probabilistic weighted voting model (PWVM) and the basic classifiers, which include support vector machine (SVM), extra trees classifier (ETC), random forest (RF), and multi-layer perceptron (MLP), for fault detection and classification. Firstly, the proposed probabilistic weighted voting model (PWVM) and the base learners/classifiers that formed this model have been evaluated their performance in fault detection and classification. For this purpose, a comprehensive comparative performance analysis was carried out. To analyze and compare performances of the proposed PWVM and machine learning methods on different datasets, the various evaluation metrics including the accuracy, sensitivity, precision, F-score, and area under the curve (AUC) have been utilized. It has been seen that the proposed PWVM structure in FDCS has higher performance than classical learning models on different datasets (normal and noisy data). To confirm generalization capability of the classifiers, additive white Gaussian noise is contributed over dataset with noise-free. These noise levels are SNR 20 dB, SNR 30 dB, and SNR 40 dB. In the second case study, the efficiency of the proposed PWVM was evaluated by using the voting machine learning models in the literature in the relevant fault detection and classification. There are 6 types of voting models: sum of probabilities, product of probabilities, maximum of probabilities, minimum of probabilities, majority, and weighted. The performance of these models on both normal and noisy datasets was investigated. In the third case study, to verify the predictive power or generalization ability of the proposed PWVM, a comprehensive analysis was performed on the training-test datasets. The related fault detection and classification problem into different training-test sets with ratios such as 70-30, 60-40, and 50-50. With this evaluation, it was seen that PWVM had the highest generalization ability compared to other voting models on test sets with different ratios. In the fourth case study, it was performed a performance analysis related to fault detection and classification of PWVM of each faulted (10 different fault types) and no-faulted conditions (11 class labels in total) for normal and noisy datasets. In the fifth case study, the performance efficiency of the proposed PWVM structures formed from different numbers of base classifiers in terms of accuracy criterion was investigated. It turned out that the PWVM structure consisting of 4 classifiers is more successful than the PWVM structure consisting of 3 classifiers. In addition, the importance of the concept of classifier diversity for the ensemble learning model was emphasized. In the sixth case study, a new feature set is created separately for the 9th wavelet decomposition level (11th, and 13th harmonic components) and the 10th wavelet decomposition level (5th, 7th, and 9th harmonic components), and the performance of the proposed PWVM was analyzed for the new fault detection and classification problem. The performance analysis was examined for the case of classifying the PWVM structure only with the feature set belonging to 9th wavelet decomposition level and for classifying only with the feature set belonging to the 10th wavelet decomposition level. Here, the PWVM structure of normal (noiseless) dataset, and noisy datasets with SNR 20 dB of ratio is tested. In this analysis, it was seen that the feature set of 9th decomposition level had a better success in classifying the feature set at SNR 20 dB than the feature set of 10th decomposition level. In the seventh case study, statistical analysis was carried out depending on the performance results obtained from the voting models used for fault detection and classification. The superiority of PWVM over other voting models was investigated by Friedman and Finner post hoc tests. From the results, it was seen that among the voting models, the model that gave meaningful results and was found to be acceptable was PWVM. Finally, apart from the fault detection and classification dataset used in thesis study, performance evaluation was made by comparing the proposed PWVM and different voting methods with the evaluation metrics (accuracy, kappa, and Matthews Correlation Coefficient (MCC)), using datasets with different characteristics accepted in international platforms. It was observed from the obtained results that the efficiency of the PWVM structure was validated against other voting models.