Predictive Modeling of Nigeria's Crude Oil and Condensate Production Using Multi-Algorithm Machine Learning and Ensemble Approaches

Abstract Nigeria's crude oil and condensate production is subject to persistent volatility driven by infrastructure deterioration, security disruptions, and OPEC quota compliance requirements, making accurate production forecasting a critical challenge for energy planning and policy. This study develops a multi-algorithm predictive modeling framework applied to 60 months of disaggregated terminal-level production data (January 2020 - December 2024) sourced from the Nigerian Upstream Petroleum Regulatory Commission (NUPRC), covering 35 terminals and streams. Five forecasting models were trained on 48 months of data and evaluated on a 12-month held-out test set: a Seasonal ARIMA (SARIMA) baseline, Random Forest and XGBoost regressors trained on a 20-feature lag-based representation, and Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning networks trained on 12-step sequential windows. A sixth ensemble model was constructed by combining all five forecasts using inverse-RMSE weighting. On the 2024 test set, the GRU achieved the best individual performance (RMSE = 2.974 MMbbl, MAPE = 5.26%), followed by the LSTM (MAPE = 5.46%), while the ensemble delivered the overall best RMSE (3.065 MMbbl) at a MAPE of 5.41%, which is well within the 10% threshold accepted for energy production forecasting. All models were subsequently used to generate recursive 10-year monthly forecasts (2025-2034), with the ensemble projecting gradual production growth from approximately 530 MMbbl/year to 557 MMbbl by 2034. The findings demonstrate that inverse-RMSE weighted ensemble modeling effectively reduces individual model error and provides a robust framework for production forecasting in complex, multi-source petroleum systems.