Edge-cloud collaboration-driven predictive planning of electric vehicle charging load for microgrids

The rapid development of smart grids has driven the intelligent transformation of charging-load forecasting for electric vehicle (EV), opening new avenues for real-time microgrid dispatch and load management. However, deep-learning models typically demand substantial computational resources, posing a challenge for deployment on resource-constrained edge devices. To address this, an Edge-Cloud Collaboration (ECC) architecture is proposed that offloads heavy training tasks to the cloud while enabling real-time inference at the edge. At its core lies a hybrid Multi-Head Attention–Long Short-Term Memory (MHA-LSTM) model: The Multi-Head Attention module computes attention weights across all input features to capture complex, nonlinear inter-feature relationships; these weighted representations are then fed sequentially into the LSTM, which learns both short-term fluctuations and long-term temporal dependencies. The proposed approach is validated on a one-year, real-world charging load dataset from a microgrid station in Baoji. Its performance is benchmarked against two traditional machine learning methods, including GBDT and MLP, and several deep learning models such as LSTM, Attention-LSTM, TPA-LSTM and Transformer. Compared with the next-best Transformer, MHA-LSTM reduces MAE by 13.3 % and RMSE by 10.6 %, achieves an R2 of 0.7979 and a peak-period MAE of just 10.7915, and attains the CPU inference time of 0.2367 s, demonstrating its feasibility for edge deployment. Furthermore, SHAP analysis reveals that Price_Band and Hour are the primary drivers of load variation, while fine-grained features such as Minute and Work_Rest improve accuracy through nonlinear interactions. The ECC-driven MHA-LSTM architecture thus delivers both high precision and low latency, providing a practical tool for energy optimization and real-time EV charging management in intelligent microgrids.