Combining machine learning and game theory for forecasting production capacity and influence factors analysis

Forecasting production capacity efficiently and obtaining dominant influence factors are important and necessary for oil and gas reservoirs development. Therefore, an innovative framework (CNN-SHAP) is proposed combining machine learning model and game theory. In a real case, the proposed framework presents a high precision and low errors with R2&0.87 and MSE is 0.35, and has great convergence on predict production capacity. In addition, the CNN-SHAP framework has accurately quantified the individual/joint/interaction effects of influence factors on forecasting tasks and identified whether the contribution of each input feature is positive or negative. The results demonstrate that the top five dominant factors are equivalent permeability, formation thickness, crude oil viscosity, matrix permeability and cluster number. Furthermore, formation thickness, injection volume and saturation are found strongly coupled to other parameters. Therefore, this framework sheds important insights into providing cost-effective production development strategies on tight reservoirs. Meanwhile, this framework has captured the complex non-linear mapping relationships among input variables and production capacity, and presents them in an explainable manner, thus improving the interpretability of machine learning models. Overall, the CNN-SHAP framework can contribute to a robust and high-precision prediction for production capacity and contribute to a more advanced development strategy on tight reservoirs.