In recent years, creating a robust strategy to allocate crude oil has been proven difficult due to the sudden and harsh changes in product prices, energy markets, and trading volumes. Traditional optimization methods, whether deterministic or heuristic, fail to optimize the allocation of crude oil, mainly because of various economic goals and the nonlinear dynamics of refinery and petrochemical markets. To address these difficulties, this study aims to develop a policy-driven framework that combines Reinforcement Learning (RL) with the Analytic Hierarchy Process (AHP). We will use AHP to assess expert preferences through pairwise comparisons. This process will generate structured weights for multiple economic and operational criteria. These weights are then embedded directly into the RL reward function, which will enable the agent to learn allocation strategies both adaptive to market transitions and aligned with managerial priorities.

The proposed RL–AHP decision framework is trained and evaluated using live transaction, price, and volume data from the Energy Exchange. Simulation results demonstrate that the agent reliably converges to a stable policy, with high agreement between expert optimal and learned decisions (Accuracy: 94.44%). We will gain substantial economic growth by implementing the learned framework. This growth leads to improved total profitability by 10.23%, increased revenue by 1.32%, and reduced operational costs by 8.91%. As a result, the general profit increases from 4.12 billion USD to 4.54 billion USD.