A Time-Adaptive Method to Transient Stability Assessment Based on Reinforcement Learning

The timely evaluation of post-contingency transient stability is crucial for power system control and decision-making. However, a major challenge lies in how to evaluate the stability within the shortest possible time. Most existing methods require a fixed length of time window for training an evaluation model, which may not be suitable for all cases. The fixed time window can delay the identification of events that could be predicted in the early stage until the specified time is reached. To address this issue, a novel time-adaptive method using Recurrent Replay Distributed DQN (R2D2) is proposed, which adapts the prediction time for each sample without human intervention. The transient stability assessment (TSA) is formulated as a decision-making process in the reinforcement learning framework. To achieve high overall accuracy at an early stage while reducing the misjudgment of unstable samples, a specific reward function has been designed to encourage the agent to explore policies. Our method is tested on the IEEE-39 bus system and the 197-bus system. The experiment results demonstrate that the agent trained by the proposed method can efficiently evaluate stability while balancing accuracy and response time. In comparison to other methods, the proposed approach achieves superior accuracy with minimal data length.