AI agents envisioning the future: Forecast-based operation of renewable energy storage systems using hydrogen with Deep Reinforcement Learning

Abstract

Hydrogen-based energy storage has the potential to compensate for the volatility of renewable power generation in energy systems with a high renewable penetration. The operation of these storage facilities can be optimized using automated energy management systems. This work presents a Reinforcement Learning-based energy management approach in the context of CO2-neutral hydrogen production and storage for an industrial combined heat and power application. The economic performance of the presented approach is compared to a rule-based energy management strategy as a lower benchmark and a Dynamic Programming-based unit commitment as an upper benchmark. The comparative analysis highlights both the potential benefits and drawbacks of the implemented Reinforcement Learning approach. The simulation results indicate a promising potential of Reinforcement Learning-based algorithms for hydrogen production planning, outperforming the lower benchmark. Furthermore, a novel approach in the scientific literature demonstrates that including energy and price forecasts in the Reinforcement Learning observation space significantly improves optimization results and allows the algorithm to take variable prices into account. An unresolved challenge, however, is balancing multiple conflicting objectives in a setting with few degrees of freedom. As a result, no parameterization of the reward function could be found that fully satisfied all predefined targets, highlighting one of the major challenges for Reinforcement Learning -based energy management algorithms to overcome.