Tight Gas Production Prediction in the Southern Montney Play Using Machine Learning Approaches

Abstract

Abstract Recently, the machine learning approach has been used to forecast tight gas production from unconventional resources. However, the performance of machine learning-based predictive models has not been successful with respect to actual field production. The poor performance has been ascribed to several factors, including the relatively few field data and few input data from geological, geomechanical, and operational information. This study uses big data analytics to develop a prediction model for Southern Montney Play in the province of British Columbia, Canada. First, a complete dataset is built, including ten geological, geomechanical, and operational characteristics for 3146 horizontal wells in Southern Montney Play. Then, the relationships between the first-year production and input parameters are evaluated, and controlling factors are identified. Finally, a comparative study of prediction models with distinctive training algorithms is conducted to find the best algorithm for predicting first-year production. The results reveal that the top features that contribute most to tight gas productivity are total injection volume, porosity, and formation pressure. Features with secondary effects are net thickness, fracturing depth, and number of stages. The other features, including permeability, gas saturation, horizontal length, and cumulative proppant injection, are the least related. The Random Forest algorithm with the highest correlation coefficient (R2=0.82) and lowest mean absolute error (MSE=0.15) is picked. The Random Forest-based production prediction matches the observed field production, indicating that the northeastern portion of the study area has the highest estimated tight gas productivity. This procedure can be applied to additional scenarios involving tight gas production and used to guide the future site selection and fracturing job size, thereby achieving effective tight gas development.