Optimal dispatch of hydrogen production systems: Integrating electrolyzer start-up optimization and deep reinforcement learning

Severe fluctuations of photovoltaic power in off-grid operation can lead to frequent start-ups of the electrolyzer, thereby reducing hydrogen production efficiency, yet this issue is often overlooked in existing energy management system. To address this problem, this work proposes an energy management strategy that integrates a dynamic model with a deep reinforcement learning algorithm. The dynamic model is employed to quantitatively characterize the impact of frequent start-up conditions on the internal dynamics of the electrolyzer, such as temperature fluctuations and voltage instability. The resulting multi-objective optimization problem is solved using an energy management strategy based on the deep deterministic policy gradient algorithm. Minute-level optimal scheduling is performed under typical operating conditions in different seasons to reduce the number of start-up events and thereby improve hydrogen production efficiency. The analysis shows that the proposed strategy can effectively reduce start–up frequency, enhance hydrogen production efficiency, and maintain stable system operation.