A new research framework combining reinforcement learning (RL) and model decomposition is gaining attention as a transformative tool for refinery scheduling and planning. The method splits complex refinery decision problems into sub models, each governed by RL agents that coordinate pricing and production across the plant.

Tested across single- and multi-period scenarios, the approach delivered major gains in computational efficiency while preserving profitability, enabling quicker adaptation to margin shifts or feedstock variance. The RL agents dynamically communicate via pricing signals, while the decomposition layer ensures the complex interdependencies in refinery networks are respected.

This is especially relevant as refiners pursue flexibility: with volatile crude markets and shifting product demand, planning that reacts in near real time becomes a differentiator. Instead of static monthly schedules, RL-enabled systems can continually adjust flows, blending, and constraints to maximize throughput and margins. For digital refinery leaders, this strategy could define the next frontier of planning intelligence.