Machine Learning Application for Hydraulic Fracturing Optimization in a China Tight Gas Field

Abstract

Abstract The normal approach to fracturing optimization in any tight formation involves comparing the production from multiple wells after one or two controlled changes to the design parameters. This approach might not adequately consider the importance of each factor nor disclose hidden relationships between them. The new approach presented in this paper involves applying machine learning to optimize fracturing designs by understanding factors affecting production and by predicting production using a data set of historic wells’ geological, fracturing and production parameters. This approach proposed here, when applied to our development asset, increased the prior frac optimization parameters by multiple factors, including geological properties, fracturing parameters, and production data, which will also be applied to predict gas production. The data from 110 wells containing Shan2 and Benxi reservoirs were prepared, trained, and analyzed. Then, 14 trial algorithms were ranked by R2 scores and the best one was used to perform a sensitivity analysis between the factors and production. This result can be used to optimize design parameters and achieve the most economic design. Randomly chosen training and testing data sets were used to compare algorithms. Based on R2 scores, a gradient-boosted tree algorithm applicable for both reservoirs was determined to be the best. This algorithm showed that effective thickness and pumping rate have had the greatest impact to gas production from the Shan2 reservoir production, but gas saturation and proppant concentration are the most important for wells producing from the Benxi reservoir. These factors then become the main adjustable parameters in the fracturing design. These results were then applied to guide hydraulic fracturing designs of six new wells. The final designs were based on optimizing both production and operating costs. Well testing showed that optimized designs achieved an overall 27% increase in absolute open flow compared with standard designs. Optimized designs incur marginally higher costs, this is offset by a 33% increase in 2 years of cumulative production, which represents an overall economic improvement in this development scenario. This principal benefit of the machine learning approach is to create a robust decision making for adjusting hydraulic fracturing designs based on production and economics. In addition, prediction accuracy improves over time with the addition of new wells and longer production history. Finally, this novel approach is proven to maximize gas well productivity and to optimize materials and logistics.