Forecasting-Centric Multi-Agent Reinforcement Learning for Real-Time and Scalable Virtual Power Plant Optimization under Renewable and EV Uncertainty

Accurate energy forecasting has become a fundamental enabler for the reliable, economic, and sustainable operation of Virtual Power Plants (VPPs), particularly under high renewable penetration and large-scale electric vehicle (EV) integration. The inherent intermittency of photovoltaic and wind generation, combined with stochastic EV charging and discharging behavior, introduces significant operational uncertainty across day-ahead, real-time, and flexibility markets. To address these challenges, this paper proposes a forecasting-centric artificial intelligence framework that tightly integrates hybrid probabilistic deep learning with multi-agent reinforcement learning (MARL) for real-time, multi-objective VPP optimization. A hybrid Transformer–BiLSTM architecture with attention mechanisms and probabilistic output layers is developed to generate calibrated forecasts of renewable generation, load demand, EV availability, market prices, and flexibility requirements. Compared with advanced LSTM-based and AOLSTM approaches, the proposed forecasting model reduces RMSE by 14% and 30%, respectively, while improving probabilistic calibration (CRPS) by over 20%, demonstrating superior uncertainty representation and prediction reliability. The resulting predictive distributions and scenario sets are embedded directly into a scenario-aware Multi-Agent Proximal Policy Optimization (MAPPO) framework, enabling coordinated yet decentralized decision-making among VPP operator, EV, ESS, and microgrid agents under centralized training and decentralized execution. Extensive simulations on a modified IEEE 33-bus system show that the proposed framework achieves a 13.6% reduction in operating cost, a 16.9% increase in net profit, and an 11.2% reduction in carbon emissions compared with deterministic scheduling. Relative to single-agent reinforcement learning, additional improvements of 3–5% are observed in cost efficiency and voltage stability, while flexibility adequacy increases from 88.2% to 97.6%. The framework maintains stable convergence under 1000 EV penetration with sub-second decision latency, confirming scalability and real-time feasibility. These results demonstrate that embedding probabilistic forecasting within decentralized MARL substantially enhances economic performance, reliability, and sustainability compared with state-of-the-art deterministic, robust, and single-agent approaches.