Abstract
<title>Abstract</title> <p>In the petroleum industry, effective reservoir management, operational planning, and economic sustainability are all based on accurate oil output predictions. However, because reservoir systems have nonlinear, time-varying, and unpredictable dynamics, predicting is still very difficult. While the majority of machine learning (ML) models, whether standalone or ensemble-based, function with static architectures that are unable to adjust to changing reservoir behaviors or sensor uncertainties, traditional analytical and numerical approaches frequently fail to capture evolving production patterns due to their fixed physical assumptions. This study proposed the Dynamic Adaptive Fusion Model (DAFM), an intelligent, dynamic, and adaptive forecasting system that can continuously self-adjust in real time, combines Decision Trees, Random Forests, Extreme Gradient Boosting (XGBoost), and Bidirectional Long Short-Term Memory (BiLSTM) networks using neural gating mechanism enabled by real-time adaptive weighting. In contrast to static fusion models, DAFM uses momentum-stabilized weight transitions, sliding-window validation, and a context-aware gating system to continuously recalibrate model contributions based on real-time performance feedback, guaranteeing stability and adaptability. In contrast to static fusion models, DAFM uses momentum-stabilized weight transitions, sliding-window validation, and a context-aware gating system to continuously recalibrate model contributions based on real-time performance feedback, guaranteeing stability and adaptability. In order to provide balanced synergy among diverse learners, a competitive error normalization method further promotes models exhibiting greater predictive accuracy. Tests on an offshore oilfield dataset show that DAFM always does better than static ensemble and standalone models, with a good predictive accuracy (R² = 0.97; MAE = 8.40 bbl/day). A 24-month sliding window provides the optimal trade-off between accuracy and flexibility, according to sensitivity analysis. DAFM provides near real-time deployment with an inference speed of about 49 ms per sample, according to runtime evaluation. All things considered, DAFM bridges the gap between traditional reservoir modeling and next-generation autonomous reservoir management by offering a scalable and reliable framework for intelligent oil production forecasting.</p>