A Closed-Loop Framework Integrating Diffusion Models with Multi-Agent Reinforcement Learning for Risk-Aware Petroleum Distribution Networks

The distribution of petroleum products is a multi-layered optimization with a multitude of scales, involving gigantic decisions in the terminal, the fleet and the finances, all over a backdrop of high prices uncertainty, high demand uncertainty and geopolitical factors. In this paper, an artificial intelligence (AI) system of three layers is presented. Layer 1 is a generative scenario engine based on time-series diffusion models, providing realistic price and demand paths and is regime-aware. The optimization of service reliability and inventory cost/asset utilization is done in Layer 2 with a multi-agent deep reinforcement learning (MADRL) architecture. The scenario distribution is added to Layer 3 and a financial risk layer is added for dynamic hedging in the scenario space with the constraint of VaR (Value-at-Risk) and CVaR (Conditional Value-at-Risk). The platform achieves a 17.3% cost savings on delivered cost for a simulated US network of 10 terminals, 20 different network carriers and 5 pipelines, a 23.8% improvement in service reliability, and a 31.2% savings in value-at-risk when compared to a static baseline, to achieve a five-year ROI of 1332% against the static baseline. In this talk we discuss the quality measures for the scenarios, the measures of convergence of logistics and the measures of effectiveness of the hedges and we discuss the deployment aspects which makes it difficult to converge the logistics and which aligns with ISO 31000:2018 risk-management guideline and ISO 20815:2018 production-assurance practice.

DOI: 10.1016/j.apenergy.2025.125000

Authors

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.