ThévNN: Data-driven Online Thévenin Equivalent Estimation Using BiLSTM

An on-line estimation method was developed to estimate the Thévenin equivalent reactance of a power system at regular intervals or on demand, using measured voltage and current values. The method is applied on-line, to track the Thévenin equivalent in a fine time scale. Using the optimally estimated Thévenin values as a multivariate time-series target, a Long Short Term Memory Recurrent Neural Network (LSTM-RNN) is trained toward performing automatic Thévenin equivalent estimation from measured synchrophasor data in real-time. To enable transferability of the trained model and limit the parameter space, the back-calculation of the bus voltage from the estimated Thévenin pair is incorporated as an additional constraint penalty during training. The proposed method has been validated using two weeks of real-world phasor measurement unit (PMU) ambient data from a high-voltage substation in the US Eastern interconnection.