Machine Learning Framework for Oil Production Forecasting

Abstract

The prediction of oil well performance and decline rates is important for field management, because these factors can be directly used to model the planning of production as well as operational decision. Classical prediction tools suffer from several difficulties, particularly in terms of the complexity and variable nature of the data. This work proposes an Artificial Intelligence (AI) approach based on machine learning (ML) for predicting oil well production. A case study was based on data from the Volve oil field in Norway. Pre- processing was performed to ensure effective and robust modelling predictions. This included searching for outliers using the IQR method and replacing them with NaNs (missing values). A machine learning based method, Random Forest that learned the entire data to fill in these missing entries was adopted. each feature having missing values were temporarily considered as a target and predicted from the rest of features. Feature relevance was evaluated using the correlation coefficient of each dependent variable with the target (BORE_OIL_VOL). Scaling of all the numerical attributes was carried out and data set has been divided into train set (80%) and test sets (20%). The prediction performance of the forecasting method was improved using an ensemble model of Support Vector Regression (SVR) and Random Forest (RF). The experimental results indicated the new model improved the prediction ability of oil well production, since we achieved to RMSE = 37.7969, R2 = 0.9954 and MAE =17.5485. These metrics demonstrate the reliability of our method. The study shows that the ensemble learning models provide good prediction results of oil well productivity and can act as a promising decision-making approach for oil field management.