Urban electric vehicle fast charging load forecasting: an LGRL approach

Abstract

To address the challenges that traditional single models have in balancing high-dimensional spatio-temporal correlation, long-sequence dynamic dependence, and a lack of a multi-station coordination mechanism in electric vehicle (EV) charging load forecasting, a hybrid forecasting method based on long short-term memory-graph reinforcement learning (LGRL) is proposed in this paper. First, a spatio-temporal feature decoupling architecture is constructed. The graph attention network (GAT) is used to dynamically aggregate high-order spatial features of neighboring nodes via adaptive attention coefficients, thereby characterizing the topological correlations of the charging station road network. Meanwhile, residual connections and a gated noise suppression mechanism are introduced into the long short-term memory network (LSTM) to adaptively filter irregular load disturbances, which enables accurate capture and noise-resistant extraction of non-linear temporal evolution laws. Second, a multi-agent coordinated forecasting framework based on value decomposition (QMIX) is established. Each charging station is modeled as an independent agent, and the global forecasting objective is decomposed into a decentralized local decision-making process in a lossless manner by introducing monotonicity constraints. A two-layer network structure comprising local and global mixing layers is designed. A parameterized hypernetwork is employed to fit the non-linear mapping between global states and local action value functions, thus achieving multi-station coordinated forecasting while ensuring the optimal performance of the global system. Third, a robust optimization strategy integrated with prioritized experience replay (PER) is designed. Key load samples are reweighted and trained according to temporal difference errors, which solves the problems of difficult algorithm convergence and weak generalization ability to extreme samples in dynamic heterogeneous scenarios. Finally, a case study is conducted based on the actual power grid topology of Fengxian District, Shanghai. Test results show that the proposed method outperforms traditional algorithms in forecasting accuracy, convergence speed, and robustness under extreme conditions such as data missing, and it can effectively quantify the impact of load fluctuations on the voltage of distribution networks.