Prediction of Productivity of Fractured Horizontal Wells Based on Tree Regression Method in Shale Reservoirs

Productivity prediction and fracturing impact evaluation are problematic by traditional methods because of the huge differential productivity and the number of horizontal fracturing sections in shale oil wells. The path to effective shale reservoir development is to build a reliable and efficient intelligent productivity prediction method using machine learning. The geological data, fracturing data, and production database of 91 production wells in the Jimsar shale reservoir were used as data sources in this study. The vector-based feature recursive elimination method is used to screen and reduce the dimension and cross-validation of the data, and determine 5 optimal main control factors covering geological factors and construction factors from 15 characteristic parameters. Root mean square error was used to assess the model performance. The learning decision trees, random forests, and gradient boosting models (GBDT) of three tree regression methods—were utilized to predict productivity. The results show that the single storage coefficient, cluster spacing, reconstruction volume, sand volume, and fracturing section length are the primary governing elements impacting the productivity of fractured horizontal wells. In the three regression machine learning models, the random forest algorithm using self-sampling solves the over-fitting problem of other tree models, which is superior to the decision tree model and GBDT model in the tree model. It has the best prediction effect, with a test set root mean square error of 0.934.