Hierarchical Personalized Federated Learning for Privacy-Preserving Oil Production Forecasting Under Data Heterogeneity

Summary Accurate oil production forecasting is critical for optimizing reservoir management, but data-driven approaches face severe constraints: Raw data sharing is restricted by privacy concerns, and high data acquisition costs lead to limited training data, seriously undermining model performance. Federated learning (FL) offers a solution by enabling collaborative modeling without centralizing raw data, yet its effectiveness in oil forecasting is further hampered by substantial interwell data heterogeneity—variations in reservoir geology, wellbore conditions, and operations cause conventional FL to fail at capturing well-specific patterns, resulting in poor forecasts. To address these dual challenges, we propose a hierarchical personalized FL (HierPFL) framework with three key innovations: (1) Wells are clustered using encrypted production data statistical features, preserving privacy while capturing intrinsic production behavior; (2) a two-level structure uses a global model for general patterns and cluster proxy models for personalized characteristics; (3) a regularized training objective aligns each client’s update with its cluster model, ensuring personalization and training stability. Experiments on a 172-well field dataset show that HierPFL converges more efficiently and achieves significantly lower prediction errors than conventional FL, underscoring its potential to enable robust, privacy-preserving oil production forecasting in petroleum engineering.