Decline Curve Analysis Enhancement with Machine Learning

Abstract Production forecasting is crucial as it directly influences financial planning, operation adjustment, and resource management. Despite its significance, the application of machine learning technique with production data is rarely reported. A common approach in the academic and industry still relies upon the traditional Arps’ decline curve analysis (DCA) models, which assume constant reservoir properties and no changes in underlying field settings. In this paper, the author analyzes the effectiveness of the Arp’s methods and demonstrates an improved approach to apply machine learning model on top of the fitted DCA modelling results. Data from hundreds of historical production wells over a 5-year period has been collected. The data is divided into train/test/validation sets. Three primary DCA methods (exponential, hyperbolic, and harmonic decline curves) are fitted, and their residuals are compared. The residual results contain useful information and the traditional DCA method is an effective way to de-trend production data. Residuals are subsequently processed using machine learning models. The models used in this study include gradient boosted tree-based models, linear models, statistical time series models, and ensembling techniques. The best fit of three DCA curves shows a fascinating result. The data is best fit by a hyperbolic curve in 51% of cases, a harmonic curve in 35% of cases, and an exponential curve in the remaining 14% of cases. The effectiveness of the classical feature engineering technique, specifically the application of logarithmic transformation, is analyzed and appears to be ineffective when used with general production data as it is only able to detrend the exponential DCA curve only. The integration of modern machine learning techniques significantly outperforms the traditional DCA methods, as demonstrated by a notable table 30 – 40% reduction in mean absolute error (MAE) scores in both train, test and hold-out sets. While some applied feature engineering techniques are beneficial, others are not. Useful techniques include applying changepoint detection algorithm prior to fitting DCA models. The ineffective methods include lagging transformation, feature extraction from date variables, and applying logarithmic transformation for non-exponential decline rate. With the trade-off of more processing time, stacking and ensembling approach of the residuals can further improve the results. This study demonstrates the value of combining machine learning techniques with traditional DCA models to enhance production data forecasting accuracy. The model has been proven to have strong generalization capabilities and should be seriously assessed alongside the traditional DCA technique. This study is beneficial for oil and gas professionals, students, and academics involved in production data who may profit from enhancing the accuracy of production forecasting.