A Reliability Assessment Method for Distribution Networks Based on Conditional Generative Adversarial Network and Multi-agent Reinforcement Learning

To enhance computational efficiency and accuracy in reliability assessment of distribution networks with large-scale distributed photovoltaic integration, a novel assessment method is proposed based on conditional generative adversarial network and multi-agent reinforcement learning. Firstly, the sequential state sequences of the system are generated using Sequential Monte Carlo simulation, and a conditional generative adversarial network (CGAN) is combined with multi-resolution meteorological factors to characterize the multivariate characteristics of source-load scenarios, including temporal dependency, volatility, randomness, and source-load correlation. Secondly, a multi-agent reinforcement learning (MARL) model is established, and a training algorithm integrating imitation learning and exploratory learning is proposed, enabling the agents to acquire optimal policies through interactive learning with an expert experience model. Finally，the simulationg is verified based on the IEEE RBTS BUS-2 system. Simulation results demonstrate that the proposed method outperforms traditional methods in terms of learning curve and stability, significantly improving both the accuracy and computational efficiency in distribution network reliability assessment, possessing superior practical values.