【摘要】 N-1状态下潮流不收敛问题对N-1校验和电网的安全运行造成极大困扰，但当前的潮流收敛研究主要集中于静态潮流，且调整方法不仅动作有效性低，也难以兼顾快速性和成功率。因此提出一种基于BNN-DS的DDPG改进算法，通过深度强化学习对N-1潮流不收敛网络进行智能调整。首先，根据N-1方案校验元件类型及潮流重载量等指标确定了方案的调整措施，通过广度优先算法确定调整元件组以保证动作的有效性，根据CRITIC权重法计算了多重奖励之和，据此，设计了N-1潮流收敛调整MDP模型。其次对MDP模型中所用DDPG算法进行改进，搭建了轻量BNN网络以降低计算复杂度、提高计算速度，设计了高奖励经验池以及存量判定机制以优化模型的收敛性。最后，在某分部2 179节点网络和某分部12 732节点网络上对改进算法进行测试验证，结果表明基于BNN-DS的DDPG改进算法比传统方法的成功率提高36.535%,平均用时减少95.01%。更多还原

【Abstract】 The problem of non-convergence in N-1 power flow poses great challenges to N-1 verification and the safe operation of power grid. However, current research on power flow convergence mainly focuses on static power flow, and existing adjustment methods suffer from low effectiveness, making it difficult to balance speed and success rate. Therefore, we propose an improved DDPG algorithm based on BNN-DS, which intelligently adjusts the N-1 power flow non-convergence scenarios through deep reinforcement learning. Firstly, we determine the adjustment measures for the N-1 scenario based on indicators such as the type of verified components and the overload of the power flow. Then, we determine the adjustment component group to ensure the effectiveness of the action by the breadth first algorithm, and calculate the sum of multiple rewards based on the CRITIC weight method. Based on this, we design the N-1 power flow convergence adjustment MDP model. Secondly, we improve the DDPG algorithm used in the MDP model, build a lightweight BNN network to reduce computational complexity and improve computational speed, and design a high reward experience pool and stock determination mechanism to optimize the convergence of the model. Finally, we test and validate the improved algorithm on a 2 179 node network and a 12 732 node network of a certain branch. The results show that the improved DDPG algorithm based on BNN-DS achieves a 36.535% increase in success rate and a 95.01% reduction in average time compared to traditional methods.更多还原