Reinforcement-learning-driven Prescriptive Operation and Maintenance of distributed energy storage for cost-effective grid asset protection

Growing integration of stochastic loads and distributed energy resources strains ageing distribution grids. Hardware upgrades, such as cable replacements, are costly, requiring predictive software solutions for grid reliability. This paper presents a Prescriptive Operation and Maintenance (POM) model with forecasting modules using a novel hybrid Graph Neural Network and Transformer to predict energy consumption and production. Combined with load flow simulations, these forecasts guide a reinforcement learning (RL)-agent to optimise battery operation, minimising electricity costs and grid alarms. Validated with real-world data, the POM model significantly reduces alarms and costs. The POM model's performance with forecasted energy nearly matches that achieved under perfect forecasting, due to accurate predictions and a robust RL-agent. Three RL algorithms, Deep Deterministic Policy Gradient (DDPG), Soft Actor–Critic (SAC), and Proximal Policy Optimisation (PPO), were tested. SAC outperformed the others across various efficiencies, consumption scales, and battery sizes, while PPO was more adaptable to parameter changes. Future work could explore higher temporal resolution, along with tests and demonstrations on larger grids operating near their capacity limit.