Real-Time AI-Driven Energy and Flexibility Management for a Technical Virtual Power Plant

This paper presents an AI-powered framework for real-time energy and flexibility self-scheduling of a technical virtual power plant (TVPP). The TVPP integrates electric vehicle charging stations (EVCSs), energy storage systems (ESSs), photovoltaic (PV) panels, wind turbines (WTs), and distributed generators (DGs) to participate in wholesale energy and flexibility markets while adhering to distribution network constraints. The framework employs a two-stage approach to maximize profitability. In the first stage, a deep learning model forecasts short-term flexibility requirements. In the second stage, optimization is divided into two interdependent sub-problems. The first sub-problem coordinates EVCS and ESS operations through multi-agent deep reinforcement learning (MADRL), using multi-agent proximal policy optimization (MAPPO) with centralized training and decentralized execution (CTDE). The second, formulated as a linear programming (LP) problem, addresses the remaining system-level operations and constraints. These sub-problems interact iteratively through a feedback mechanism to align local decisions with global objectives. To enhance training efficiency, single-agent policies are pre-trained using proximal policy optimization (PPO) in simplified environments and later fine-tuned via transfer learning in the full MADRL framework. The proposed method is evaluated on the IEEE-33 bus test system, demonstrating superior operational efficiency and profitability over existing approaches.