Oil Production Decline Prediction: A Comparative Study of Conventional Machine Learning, Ensemble Residual Learning, and Sarima Models

Oil production forecasting is a key phase in the life of an oilfield. Unfortunately, the conventional empirical and statistical predictive models exhibit several lapses. Also, traditional machine learning model training fails to capture the high non-linearity in the time series dataset. In this current work, the performance of ensemble residual machine learning models is compared to conventional machine learning and SARIMA models which have been largely used for time series forecasting. A monthly production data composed of 72 characteristics and 3 observations was used in this study. Hybrid models developed from the ensemble residual learning, especially the ridge plus the XGBoost/Decision Tree models showed enhanced performance. Their metrics were better than the best-performing conventional model with 11.30% RMSE, 14.13% MAPE, and the SARIMA model with 21.11% RMSE, 23.00% MAPE.

Keywords: oil production forecasting, time series forecasting, machine learning, residual modeling, ARIMA, SARIMA

Suggested Citation: Suggested Citation

Adjei, Stephen and Ojuu, John Oleka and Laryea-Adjei, Nii Sowah and Sokama- Neuyam, Yen Adams and Mensah, Isaac Adjei and Quaye, Jonathan Atuquaye and Duodu, Turkson Kwesi, Oil Production Decline Prediction: A Comparative Study of Conventional Machine Learning, Ensemble Residual Learning, and Sarima Models. Available at SSRN: https://ssrn.com/abstract=5124652 or http://dx.doi.org/10.2139/ssrn.5124652