HI-LSAC framework uses human guidance to boost grid voltage safety

Researchers Mahmuda Akter and Hamidreza Nazaripouya introduced HI-LSAC (Human-Interactive Lagrangian Soft Actor-Critic), a framework that marries deep reinforcement learning with human oversight for voltage regulation in power distribution networks. As distributed energy resources (like solar and batteries) increase grid variability, traditional DRL methods suffer from unsafe exploration and slow convergence. HI-LSAC addresses this by adding an adaptive Lagrange constraint to enforce voltage limits and a human-guidance module that provides corrective actions based on sensitivity analysis, feeding into a regularized actor loss. This enables the policy to learn safe, interpretable control behaviors.

Simulated on the IEEE 13-node feeder using PowerGym-OpenDSS, HI-LSAC significantly outperformed baseline DRL methods in reducing voltage-violation severity and overall power losses. The framework demonstrates how human expertise can be integrated into autonomous control to improve reliability and safety in critical infrastructure. This work points toward practical, collaborative AI systems for managing increasingly complex power grids.