As nations race toward carbon neutrality, the intermittency of wind and solar power poses a major challenge to grid reliability. While hydrogen energy storage systems (HESS) offer a promising buffer for days or even seasons, intelligently coordinating these diverse energy sources in real time remains daunting for traditional methods. To tackle this, a team led by He, L. from Northwestern Polytechnical University, China, developed a distributed deep reinforcement learning dispatch framework.
The framework first condenses year-long electricity demand patterns using PCA-enhanced K-means clustering, preserving over 95% of original information. To capture renewable generation uncertainty, the team employed Dynamic Time Warping (DTW) with DBSCAN to extract representative seasonal scenarios that account for nonlinear timing shifts conventional averaging misses.
At its core, a distributed Deep Deterministic Policy Gradient (DDPG) algorithm deploys multiple parallel “actors” exploring different data segments, with a central learner synchronizing their insights—achieving a 5.5-fold speedup (from over 72 hours to 11.5 hours). The system dispatches thermal power, grid purchases, and hydrogen storage while minimizing coal, carbon, and electricity purchase costs. In simulations, the HESS-integrated framework cut total operational costs by 6% (from $56.96 million to $53.6 million) and proved highly robust under meteorological noise, with costs rising by only 0.35%. This work establishes a scalable blueprint for hydrogen storage as an active participant in future low-carbon grids. The work entitled “An energy-efficient scheduling approach for wind-solar-hydrogen systems based on distributed reinforcement learning” was published on AI Agent (published on May 29, 2026).
DOI：10.20517/aiagent.2026.01