Curriculum-based deep evolutionary learning for large-scale grid look-ahead transient stability preventive dispatch

Abstract

This paper focuses on the look-ahead transient stability preventive dispatch (LA-TSPD) problem in large-scale power systems. The main objective is to derive look-ahead dispatch strategies in real-time to achieve safe and economical operation of the power grid under credible contingencies. Deep reinforcement learning (DRL) methods have been developed for the same or similar scenarios, but they still suffer from several challenges such as computational inefficiency and poor exploration ability. To overcome these issues, a novel curriculum-based deep evolutionary learning (DEL) method is developed for large-scale LA-TSPD problem. Unlike regular DRL methods, DEL methods introduce perturbations directly in neural network parameter space rather than the action space to facilitate exploration, which makes it particularly well-suited for the highly complex LA-TSPD problem. Besides, drawing on the physics knowledge from LA-TSPD, a novel curriculum-based learning framework is further developed to alleviate the problem complexity in large-scale grids. Numerical simulations on the IEEE 39-bus system, a real 58-bus system, and a large-scale 500-bus system demonstrate that compared with the state-of-the-art (SOTA) DRL methods, the proposed method shows better solution optimality, training robustness, parallel scalability, as well as adaptability to large-scale power grids.