Deep Reinforcement Learning-Based Approach for Proportional Resonance Power System Stabilizer to Prevent Ultra-Low-Frequency Oscillations
Guozhou Zhang, Weihao Hu, Di Cao, Qi Huang, Jianbo Yi, Zhe Chen, Frede Blaabjerg
Abstract
Recent studies have shown that due to the hammer effect of the governor, hydropower units are easily creating negative damping torque at the common mode frequency (below 0.1 Hz). Therefore, there is a risk of ultra low frequency oscillations (ULFO) in hydropower-dominated systems. ULFO is a small-signal frequency oscillation problem, which is quite different from low frequency oscillations (LFO). A conventional power system stabilizer (CPSS) has less effect on suppressing ULFO. To solve this problem, this paper proposes a high-order polynomial structure to replace the CPSS, and combine it with a proportional resonance controller to form a novel PR-PSS. In order to ensure the robustness of PR-PSS, based on the characteristic analysis results of the PR-PSS, a deep reinforcement learning (DRL) algorithm asynchronous advantage actor-critic (A3C) is introduced to train an agent. After training, the proposed agent can provide optimal parameter settings for PR-PSS under various operating conditions. Simulation results verify the effectiveness of the proposed method.