Power System Transient Stability Assessment Based on Voltage Phasor and Convolution Neural Network

Recent Developments in Machine Learning for Energy Systems Reliability Management

“…The database generation is then usually done offline, given the extensive simulation cost to build it, while the application of the resulting model trained on the dataset can be done offline or online, depending on the application and the context. [15], [16], [19], [22], [27]- [29], [32], [34], [35], [37], [38], [41], [43], [45], [47], [49]- [52], [54]- [56], [63], [64], [67]- [73], [75], [76], [81]- [83], [85]- [88], [90], [92], [93], [98], [102], [105], [107], [113], [115]- [117] Voltage stability [26], [30], [39], [40], [42], [44], [46], [48], [53],…”

“…Table II provides an overview of the main data pre-processing methods used for DSA and DSC discussed below. [34], [35], [37] Genetic algorithms [38], [39] Tree-based algorithms [40]- [43] Feature extraction PCA and variants [44]- [46] Fisher's linear discriminant [47] Shapelets for time series [48] Deep learning auto-encoders [49]- [52] Class imbalance Oversampling [24], [53], [54] Cost-sensitive learning [53], [55] Ensemble methods [41], [45], [56], [57] 1) Feature engineering: Given the large number of features necessary to fully describe the state of a power system and the need for fast algorithms, feature selection techniques are proposed in many papers. Too many features can lead to excessive training time and, if many features are not relevant, could decrease the performance of the learnt model.…”

“…The problem is particularly important, given that unstable events must be detected to guarantee the reliability of the system. To overcome this issue, several solutions have been proposed in the literature, such as oversampling the minority class [24], [53], [54], undersampling the majority class, exploiting a cost-sensitive learning [53], [55] and using ensemble methods [41], [45], [56], [57].…”

“…Costsensitive learning consists in using different costs for different classes during learning, in order to, for instance, avoid as much as possible misclassifying unstable samples. Using another approach, suitable because of their data representation as images, Hou et al artificially increase the number of samples of the minority class by using a multi-window sliding recognition method [54]. The database generation method proposed by Venzke et al in [24] allows to generate balanced dataset between secure and insecure labels.…”

“…Convolutional Neural Networks (CNNs) have also been used, in particular for transient stability assessment [49], [54], [55], [70]- [73]. Reference [74] proposes to represent the power system state as a (3-channel) image to take advantage of the convolutional neural network algorithm for small-signal stability.…”