Comprehensive flexible framework for using multi-machine learning methods to optimal dynamic transient stability prediction by considering prediction accuracy and time

• Comprehensive flexible framework for combining machine learning methods to transient stability prediction. • Independent from number and type of stability prediction algorithms. • Dynamic time-dependent prediction of stability condition and the remaining time until instability. • Prediction the stability status and time along with the corresponding accuracy level. • Using only local data of the generator as the input of model and Robust against miss and noisy data. Fig. 1 shows the graphical concept of the proposed approach. The local data are measured from the power system: voltage (V), current (I), mechanical power (Pm), electrical power (Pe), and speed (W) to be hired with MLs for TSP. The ML results are combined by using two novel hard voting and soft voting systems to extract stability status and the remaining time until instability, respectively. The final decision is updated by a time frame and represents a dynamic response. Transient stability, a crucial aspect of power system research, is the subject of this paper. It determines the system's stability under severe disturbances. In recent years, Machine/Deep Learning (ML/DL) techniques have been widely applied to predict transient stability conditions. This paper presents a flexible framework for using the desired number of ML algorithms and combines the results of them to extract the final optimal transient stability perdition (TSP). This prediction includes stability status (stable/unstable) and remaining time until instability is labeled with related accuracy. To show the effectiveness of the proposed framework, for instance, four different ML approaches are used: Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), and K-nearest neighbor (KNN). The introduced framework combines the output of ML methods in two stages considering the time-dependent perdition accuracy; the first stage predicts stability status by using a hard voting system, and the second one estimates the remaining time until instability with a soft voting system. The final optimal outputs of the proposed approach are dynamic time-dependent curves for the prediction of stability status and time. The prediction accuracy changes by data size and can reach to 100 % and the remaining time until instability can predict with 0.03 s error, averagely. Supplementary studies examine how noise, missing data, and important inputs affect the projections. The stability dataset is collected from the DIgSILENT Power Factory and tested on the IEEE 39-bus system. Also, the proposed framework is coded with PYTHON software.