Residual RL cuts pumped storage degradation by 56% in new two-layer control

A new paper on arXiv (2607.06911) introduces a degradation-aware control system for variable-speed pumped storage hydropower (VS-PSH), a critical technology for grid-scale energy storage. The authors—Kyung-bin Kwon, SangWoo Park, and Dam Kim—tackle the fundamental trade-off between accurate power dispatch and component wear. Traditional single-controller approaches force a conflict: every improvement in tracking accuracy increases mechanical and hydraulic degradation. The proposed solution uses a two-layer architecture: a deterministic feedforward-PI gate controller handles the guaranteed 5-minute block dispatch commitments (auditable and certifiable for grid operations), while a residual reinforcement learning policy adjusts only the rotor speed within a narrow, safe band. This design ensures the worst-case command is bounded by construction, preventing any learning-based action from violating operational limits.

The RL speed policy is trained on a novel operation-degradation index that combines off-best-efficiency hydraulic loss with power and actuation variation into a single, interpretable signal. It learns to follow a demand-dependent best-efficiency-point reference. In tests across normal and stressed dispatch scenarios, the new architecture achieves a 96% reduction in best-efficiency-point tracking error compared to a fixed-speed baseline. More importantly, it cuts total operational degradation by up to 56% under the most demanding dispatch conditions. The system matches or slightly exceeds the efficiency of a full-information model-based optimizer while maintaining tighter block tracking. This work demonstrates how residual reinforcement learning can safely and significantly improve both performance and longevity in critical energy infrastructure.