Real-time transient stability assessment based on deep recurrent neural network

Abstract

Real-time transient stability assessment (TSA) is conducted in a data driven framework, by considering the temporal relations of the predictors using recurrent neural network (RNN) with long short-term memory (LSTM) units. Specifically, the TSA problem is converted to approximate the stability boundary based on the fact that the disturbed system is stable if and only if the initial post-fault state is in the region of attraction in state space. Since the state variables are difficult to monitor in real world applications, multiple time-step algebraic variables (power flows and voltages of the transmission network, i. e. P, Q, V, θ) are utilized to approximate the state space by using a LSTM-based model to capture the long-term dependencies along the time steps. The proposed scheme is illustrated on the IEEE 39-bus test system and gets remarkably better testing results compared with a SVM benchmark. The proposed method has shown that involving power system stability domain knowledge is significant to data driven TSA analysis.