An optimal scheduling method of hydrogen production systems using curtailed electricity through artificial intelligence and deep reinforcement learning

Hydrogen production using curtailed wind and solar energy is an efficient way of energy consumption, but the power generation model using curtailed wind and solar energy suffers the same shortcoming as the conventional wind and solar power generation models-unstable transfer of energy, which makes it imperative to perform energy scheduling in hydrogen production systems that use curtailed wind and solar energy to avoid secondary waste of energy. Deep reinforcement learning (DRL), due to its outstanding performance in finding the optimal solution in time series problems, has gained popularity in energy scheduling tasks. However, when applied to the unstable hydrogen production systems using curtailed wind and solar energy, the reinforcement learning (RL) decision-making processes are not intelligent enough and the decisions are likely to violate the safety limits. In the present work, the energy scheduling process decided by DRL agent was converted into textual information, which was then assessed and scored by a large language model (LLM); the scores were then fed back to the RL reward function to allow the DRL agent adjust the decisions. The research revealed that the coupling between $L L M$ and DRL could improve the security of the energy storage module of the hydrogen production system and make the energy scheduling more intelligent without affecting the RL agent’s performance in finding the optimal solution.