AI-driven hypergraph neural network for predicting gasoline price trends

Abstract

Accurate gasoline price trend forecasting is crucial for policymaking, risk management, and maintaining international trade balances. Despite advances, existing artificial intelligence (AI) deep learning models often fail to fully address the complex multifactor interactions behind gasoline price fluctuation and lack interpretability. This study introduces a hypergraph neural network (TADHGCN) model with temporal attention embedding to examine these higher-order interactions and extract time-dependent information effectively. We also integrate SHAP (SHapley Additive exPlanations) for interpretability to enable the exploration of how various factors influence gasoline price predictions. We validate the proposed model using gasoline price data from the United States. Energy Information Administration, from January 1995 to February 2011. Empirical results demonstrate that the proposed model improves accuracy by 6.35% compared with the state-of-the-art benchmark, and the model’s robustness is verified by ablation study. Gasoline sales volume and finished gasoline inventories are identified as key drivers of price movement wherein sales volume has a positive impact, while inventories exert a negative effect. The TADHGCN model with time-series attention and SHAP analysis provide accurate price trend forecasts and offer valuable insights into the contributing factors. This study provides both theoretical and practical guidance for anticipating gasoline price fluctuations more accurately.