Abstract: It is essential to predict multistage fractured horizontal well (MFHW) productivity of tight sand gas reservoirs for evaluation of hydraulic fracturing performance and optimization of hydraulic fracturing design. However, most of the current predictive methods introduce multiple assumptions and simplifications. Therefore, these methods cannot fully account for multi-scale fluid flow mechanisms in the tight formations of well productivity. A machine learning approach for predicting MFHW productivity is proposed. A well productivity model is built by machine learning algorithms to uncover hidden patterns in a data set including geological, fractured well productivity, drilling and completion data. In addition, to solve the "black box" issue of conventional machine learning modelling, the SHAP (SHapley Additive exPlanations) method is used to explain the built ML model globally and locally. The efficiency and practicality of the proposed method is demonstrated by the application to the Eastern Sulige Gas Field. Compared with petroleum reservoir simulation, the method not only improves the prediction performance of the well productivity, but also reduces modelling cycle and improve computational speed.

Key words: multistage fractured horizontal well, machine learning, productivity prediction, interpretablity, data-driven, SHAP method