A machine learning-based load shedding method for improving resilience of power system

Modern power systems require fast and effective load shedding to maintain frequency stability and enhance resilience, especially during disturbances. Traditional load shedding methods suffer from several shortcomings, including slow computational time and negligence of load shedding priorities. To address these challenges, this paper introduces a novel load shedding approach based on frequency prediction and deep reinforcement learning. First, a dynamic frequency response model of the power system is constructed. This model imitates the dynamic frequency response of the power system through the transfer function, and based on this, the dynamic frequency response indices including the steady-state frequency difference and maximum frequency difference can be calculated. Then, a deep Q network (DQN) based load shedding method is presented through designing DQN parameters including DQN state, action, reward function, and training method. Finally, the empirical analysis indicates that the proposed method can achieve a lower frequency nadir and smaller maximum frequency difference than the method based on real-time frequency measurement. Moreover, relative to the model-based method, the proposed method provides faster decision-making speed, contributing positively to system frequency stability and enhancing the resilience of power systems against disturbances.