Deep reinforcement learning-based optimization of an island energy-water microgrid system

Islanded microgrids often struggle with limited resources and heavy reliance on fossil fuels. This study optimizes an island energy-water microgrid using reinforcement learning (RL) to schedule the water system as a virtual battery. Using the Shoals Marine Laboratory microgrid as a testbed, a dynamic model simulating water and energy interactions was integrated with an RL algorithm to improve profit, sustainability, and reliability. The RL scenario outperformed the status quo, leading to a 7.04 % overall improvement in the equally weighted total score. It also exceeded a single-objective, heuristic management approach in profit and reliability, though slightly reducing sustainability. The RL model strategically reserves battery storage for peak renewable energy generation and extends water system operation to circumvent pumping constraints. Increasing desalination rates further improved performance, while larger water tank capacity offered minimal advantages. Future research may expand RL decision space to incorporate energy system actions for enhanced benefits.