A Sequential Multi-Stage Framework for Electric Vehicle Charging Management Integrating Forecasting, Optimisation, and LLM-Based Explanation

Abstract

As electric vehicle (EV) adoption grows, improving EV charging management depends on accurate forecasts of charging demand and station occupancy, along with effective route guidance and dynamic charging-selection policies. Yet many existing approaches address only one component and do not provide an end-to-end pipeline that handles multiple goals. To overcome this limitation, this study introduces a sequential, multi-stage framework that predicts EV charging volume and occupancy under dynamic conditions while simultaneously optimising objectives such as routing, congestion, and cost. For forecasting, the framework uses an LSTM–Transformer model to learn both local and global dependencies and to deliver multi-horizon predictions. For optimisation, it applies reinforcement learning with proximal policy optimisation (RL-PPO), improving decisions by utilising a predicted feature map. In addition, the forecasting signals and agent outputs are leveraged to produce interpretable, LLM-based explanations, helping users to better understand the model’s predictions. The framework is evaluated on diverse datasets, including urban networks from two public benchmarks (UrbanEV and Multi-Faceted) and custom data from the Al Jouf region in Saudi Arabia. The results demonstrate the model’s effectiveness, achieving minimum RMSE values of 0.3244, 0.9339, and 0.15, with predictions closely matching the observed signals. In the optimisation stage, the proposed RL-PPO method achieves the best performance across all three datasets in terms of distance (km), congestion, and waiting time, resulting in higher customer satisfaction and a stronger trade-off than competing methods. Overall, the study delivers accurate forecasting, optimised decision support, and human-readable explanations, enabling an intelligent EV charging system.