Using Hyperparameter Tuning via Deep Reinforcement Learning to Optimize Energy Storage Systems in an Integrated Energy System

The integration of nuclear power plants (NPPs) with intermittent renewable energy sources, energy storage systems (ESS), and hydrogen-producing electrolyzers within an integrated energy system (IES) has the potential to enhance both its flexibility and profitability. Optimizing the ESS capacity can further improve the flexibility and the practicability of installing such storage solutions in a nuclear renewable-IES (NR-IES). This paper demonstrates the use of hyperparameter tuning in optimizing ESS capacities within such an envisioned NR-IES. This optimization results in decreased maintenance and installation costs for the ESS as they are operated at optimal capacities, where both the cumulative revenue and overall NPP flexibility in the NR-IES are high. The proposed ESS optimization solution involves a comparative analysis of various search optimization algorithms to determine the most effective hyperparameter tuning algorithm. This paper presents the NR-IES optimization using deep reinforcement learning (DRL) based on a prior study, where an agent is trained via proximal policy optimization (PPO). It demonstrates enhanced results in comparison to the earlier work. Compared to a non-hyperparameter-tuned framework, the ESS capacities tuned by Bayesian optimization (BO) would result in lower expenses and a significant revenue increase of 24.41% over a 120-day period.