Multi-Agent and Forecast-Conscious DRL-MPC Framework of Safe and Efficient EV Fast-Charging With Digital Twins

Abstract

The growing adoption of electric vehicles (EVs) puts a lot of pressure on fast-charging and distribution grids. The current methods of charging control, which are rule based, stand-alone Model Predictive Control (MPC), and deep reinforcement learning (DRL), have several important limitations: poor scalability, absence of coordination across chargers, the safety of exploration and reactive behavior because of the no-forecasting of demand. Such deficiencies lead to grid constraint breach, ineffective charging as well as poor system reliability in peak demand states. To solve these dilemmas, this paper suggests a Digital Twin-based, forecast-aware Multi- Agent Deep Reinforcement Learning (MARL) architecture with MPC-based safety oversight in coordinated EV fast charging. The electro-thermal battery dynamics, charger constraints and grid limits of a high-fidelity digital twin is electronegative and allows safe learning and realistic evaluation in closed-loop. The agent observations are equipped with short-term load forecasting, which facilitates grid-aware decision-making, that is, proactive charging. MARL will regulate individual chargers, and an MPC supervisor will implement strict operational and safety limitations. Extensive experimental analyses reveal that the proposed framework has a significant reduction of charging completion time, charging efficiency, smooths aggregate grid load, and an overall result that eliminates grid and thermal constraint violations using MPC only, DRL only, and single-agent DRL MPC baselines. Scalability test also proves its strong performance when scaling the number of charging agents, which proves the appropriateness of the framework to large-scale rollout of EV fast-charging infrastructure.