SHAC framework cuts industrial microgrid costs 91% with multi-agent RL

A team of researchers from Tsinghua University has introduced SHAC (Sequential Heterogeneous-Agent Coordination), a multi-agent reinforcement learning framework designed to optimize tie-line power shaping in industrial microgrids. The framework treats process loads, hydrogen storage, and battery storage as distinct agents with role-specific rewards, critics, and asynchronous decision intervals. By embedding process-knowledge-based action masking and feasibility projection into policy execution, SHAC ensures safe real-time operation while capturing the heterogeneous temporal effects of each resource on the grid connection.

Tested on a steel industrial microgrid with real-world renewable generation and electricity market data, SHAC achieved dramatic improvements over baseline operations. The framework reduced total grid purchase costs by 91.27%, eliminated 98.64% of contract-demand exceedance time, and cut cumulative ramp excess by 96.91%. Critically, it maintained zero production failures while operating with an average computational time of just 0.4 milliseconds per decision step—enabling fully adaptive 1-minute online control without needing a predefined reference trajectory.