LSTM-based PSS Design for Modern Power Systems


  • Khaled Aleikish
  • Thomas Øyvang



Machine learning, Artificial neural networks, Power system modeling, Low-frequency oscillations, Power system stability


With the ever-increasing incorporation of wind and solar power in the electric power system, enhanced performance of classical bulk hydropower plants for robust operation of the power system is required. This current energy transition may cause a rapid increase in undesirable low-frequency oscillations (LFOs) in modern power system operations. A power system stabilizer (PSS) located at hydropower plants is one solution to damp such oscillations. This paper presents a new method based on Long Short-Term Memory (LSTM) neural networks for sine-wave phase shifting to possibly enhance PSS damping. The proposed controller considers the PSS input and the rotor speed deviation as a damped sinusoidal signal, simplifying PSS control and real-time optimization of PSSs parameters. Results show that the proposed LSTM architecture is able to learn multiple damped sine waves with different frequencies and decay rates. Therefore, the proposed controller can operate on the entire range of LFOs, unlike simple feedforward neural network (FNN) controllers, which can only learn and function on a single LFO frequency.