Transient Stability Preventive Control Based on Graph Convolution Neural Network and Transfer Deep Reinforcement Learning
Tianjing Wang, Yong Tang
Abstract
This study proposes a new transient-stability preventive control (TSPC) method based on graph convolutional neural networks (GCNN) and transfer deep reinforcement learning (DRL) to address non-convergence problems of traditional optimization algorithms and slow training speed of artificial intelligence algorithms for TSPC. First, a transient stability assessor (TSA) with GCNN is developed to assess current-power-flow state. Sensitivities of the transient-stability index relative to the generators are approximately calculated using TSA; generators with significant influence able to narrow action space are identified. Subsequently, the Markov decision-making process of TSPC is derived by introducing the process of TSPC. A DRL for TSPC is constructed by adding entropy to twin delayed deep deterministic policy gradient (TD3). Knowledge learned by TSA is transferred to DRL based on transfer learning, which improves learning efficiency. Finally, case studies based on the IEEE 39-bus system and an actual power grid prove the effectiveness of the proposed method. Comparisons performed with reference algorithms in literature demonstrate the proposed method has better performance in both control effect and speed.