AI-Driven Digital Transformation in Petroleum Product Distribution: A Multi-Agent Reinforcement Learning and Generative Modeling Approach
Abstract
The distribution of petroleum products is a high-stakes, multi-scale optimization problem: decisions at the terminal, fleet, and financial layers interact under deep uncertainty in prices, demand, and geopolitics. This work introduces a three-layer AI platform. Layer 1 embeds a generative scenario engine built on time-series diffusion models that produce realistic price and demand paths. Layer 2 coordinates logistics with a multi-agent deep reinforcement learning (MADRL) framework that balances service reliability, inventory cost, and utilization. Layer 3 couples the scenario distribution to a financial risk module performing dynamic hedging under VaR and CVaR constraints. Across a simulated US network of 10 terminals, 20 heterogeneous carriers and 5 pipelines, the platform reduces delivered cost by 17.3%, improves service reliability by 23.8%, cuts value-at-risk by 31.2%, and delivers a five-year ROI of 1,332% versus the static baseline.
DOI: 10.1016/j.apenergy.2025.125000
Layer 1 · Generative Scenario Engine
Time-series diffusion models sample joint price–demand futures. The engine stress-tests every downstream decision against the full ensemble instead of point forecasts.
Layer 2 · MADRL Logistics Coordinator
Multi-agent deep RL — one agent per terminal, carrier and corridor — learns dispatch, routing and inventory policies that optimize service reliability against cost.
Layer 3 · Financial Risk Module
Scenario distributions drive dynamic hedge rebalancing under VaR and CVaR constraints across futures, options and swaps.
Keywords
Code & Reproducibility
The full training pipelines, scenario engine, MADRL environments and hedging backtests are open-sourced. The simulated US network (10 terminals, 20 carriers, 5 pipelines) used across this demo mirrors the experimental setup in the paper.