Microgrids constitute a major lever of the Algerian energy transition. We propose a controller based on deep reinforcement learning capable of arbitrating in real time between photovoltaic production, storage, and grid withdrawal.
Method
The agent is trained on real load profiles measured on the campus. The reward function penalizes both energy cost and constraint violations.
Results
Simulations show a reduction of 18% of the energy bill compared to a heuristic regulation, while respecting the quality of service constraints.