Reservoir Oil Rate Forecasting Using Machine Learning Regression Algorithms

Abstract

Accurate prediction of oil rates is critical for optimizing production and reservoir management. This dissertation investigates the effectiveness of machine learning (ML) regression algorithms techniques as an alternative to traditional numerical simulation methods in oil rate prediction. The work focuses on pre-processing and cleaning well data, including outlier removal, handling missing values, and label encoding for well names. Feature selection utilizes correlation analysis to identify relevant features impacting oil rates, such as Gas-Oil Ratio (GOR), water cut, choke size, and wellhead pressure. Five regression algorithms – Linear Regression, Random Forest Regressor, Gradient Boosting Regressor, Support Vector Regressor (SVR), and XGBoost – are evaluated for their performance in predicting oil rate by using different Python libraries. XGBoost emerges as the best-performing model with the lowest Mean Squared Error (MSE). Random Forest Regressor and Gradient Boosting Regressor also demonstrate promising results. Linear Regression and SVR exhibit significantly higher MSE, indicating lower accuracy. The dissertation showcases the effectiveness of XGBoost in predicting oil rates with high accuracy (over 90% for sample wells). This approach offers a valuable alternative to traditional numerical simulation methods, potentially leading to improved production forecasting and reservoir management strategies