Identification and Detection of Ferroresonance Phenomenon in Active Distribution Networks Using Long Short-Term Memory Neural Networks Enhanced by Genetic Algorithm

Given the importance of detecting the ferroresonance phenomenon and distinguishing it from other transient events in active distribution networks, this paper proposes the use of Long Short-Term Memory (LSTM) recurrent neural networks to achieve this goal. In the proposed LSTM network, the optimization of learning period parameters, hidden layers, and corresponding weight coefficients is performed using a Genetic Algorithm (GA) to enhance the detection accuracy and speed. To create a suitable time-sequence database for LSTM training, an active distribution network comprising wind turbines, photovoltaic systems, and synchronous diesel generators is modeled in PSCAD software. By simulating various transient events, such as phase disconnections, capacitor bank switching, and load switching, the calculated detail and approximation coefficients up to six levels of the three-phase voltage waveforms of feeders are used as input features during the LSTM network training phase. Subsequently, to validate and confirm the effectiveness of the proposed LSTM-GA model, numerical studies are conducted on this active distribution network for the identification and distinction of the ferroresonance phenomenon. Simulation results confirm that optimizing and tuning the LSTM network parameters with GA significantly improves performance metrics, including precision, accuracy (P), recall (R), and others.