Hybrid prediction model of electric vehicle charging load based on STL decomposition and multi-branch neural network

In the context of the rapid development of the global electric vehicle (EV) industry, accurately predicting charging loads is of vital importance for ensuring grid stability and resource optimization. Addressing the challenges posed by existing models in handling the spatiotemporal discreteness, high-frequency fluctuations, and randomness of load data, this paper proposes a prediction model named AP-STNet based on a multi-branch structure. This model utilizes STL decomposition to divide the load sequence into trend, periodic, and residual components, which are modeled by the TCENet, APGCNet, and MLRDNet respectively, and is optimized collaboratively within a unified framework. Empirical results based on 18,061 charging stations in Shenzhen show that AP-STNet outperforms multiple mainstream models in terms of MAE, RMSE, and R${}^2$ by 36.8%, 44.6%, and 23.7% respectively. This study constructs a clear-structured and highly interpretable deep hybrid prediction framework, filling the research gap in large-scale charging load modeling, and has significant engineering and academic value.