A Machine Learning-Based Model For Performance Prediction Of Hydraulic Fracturing In Shale Oil Reservoirs

With the continuous improvement of oil and gas extraction technology,fracturing has been widely used as an important production-enhancing technology in unconventional oil and gas development such as shale oil.Accurate prediction of the fracturing effect is an important basis for decision-making and optimization and has significant theoretical and practical value.However,current research still has deficiencies such as incomplete data processing,low prediction accuracy,and poor model generalization ability.This article selects 14 parameters that reflect reservoir properties and fracturing construction conditions as influential factors based on the comprehensive logging quantitative interpretation model for fracturing “sweet spot” selection technology and the basic principles of fracturing construction process technology.The dynamic and static fracturing effect prediction ideas are represented based on multiple linear regression and Transformer neural network architecture.The multiple linear regression model uses cumulative production in the first 90 days as the fracturing effect evaluation standard.Using principal component regression analysis,it also takes into account the interaction of influential factors and establishes a multiple linear regression equation between 14 influential factors and cumulative oil production in 90 days,explaining the influential factors and their degree of influence on the initial production capacity of shale oil reservoirs after fracturing.The Transformer model is suitable for dynamic prediction on a longer time scale,with the first year’s daily production curve as the fracturing effect evaluation standard.With the geological properties,fracturing construction data,and production data of 20 wells as training sets,the model adaptively learns,trains,encodes,and decodes the features,and can dynamically predict the daily production curve within one year after production for single wells based on the production data of the first 90 days.Experiments have shown that both the dynamic and static models established in this article have high prediction accuracy and generalization ability.The multiple linear regression model establishes a numerical relationship between logging data,fracturing construction data,and the initial production capacity of shale oil reservoirs,providing data support for fracturing construction decision-making and optimization and reference for fracturing construction effect evaluation.The Transformer model can establish a prediction model for future production trends,greatly improving the prediction accuracy of the fracturing effect.In summary,the dynamic and static models established in this article can improve the fracturing effect’s prediction accuracy and generalization ability by selecting,processing,and optimizing data features,providing effective support and a decision-making basis for fracturing technology optimization.In the future,by analyzing and modeling more data factors,the accuracy and effectiveness of fracturing effect prediction can be further improved.