Reinforcement learning-based framework for integrated green methanol supply–demand management with demand forecasting under renewable energy uncertainty

The transition toward a sustainable methanol economy requires developing integrated green methanol systems powered by renewable energy sources (RES). To this end, a novel reinforcement learning (RL)-based framework for green methanol supply–demand management is proposed in this study. This framework integrates production (power-to-methanol), transportation, and utilization (methanol-to-power) into a unified operational system. To address challenges posed by the temporal variability of RES and methanol demand, the proposed framework incorporates time-series forecasting models for embedding future demand information into the observation space of an agent, thereby enabling proactive and adaptive decision making under uncertainty. Both single-agent and multi-agent RL architectures are implemented and compared to evaluate their performances in managing energy flows, optimizing power planning, and minimizing grid dependency. The proposed system is validated using real-world renewable power and demand profiles, which demonstrate a trade-off between profit and renewable energy penetration (REP). Multi-agent RL achieves higher profitability, whereas single-agent RL achieves superior REP performance. In addition, the effect of forecasting accuracy on RL performance is analyzed, which indicates that higher forecasting errors led to a 305.33% increase in profit, reaching maximum profits of 1,173,422$/month while reducing REP by 79%. This analysis highlights the synergy between predictive models and intelligent control strategies. This study presents new insights into the integration of RL-based energy management techniques for future green methanol infrastructures.