Initial production prediction for horizontal wells in tight sandstone gas reservoirs based on data-driven methods.

Accurate prediction of the initial production in horizontal wells targeting tight sandstone gas reservoirs (IPHTSG) is critical for assessing the exploitation potential of well locations and identifying reservoir sweet spots. Traditional methods for estimating horizontal well productivity exhibit limited applicability due to reservoir heterogeneity and unfavourable petrophysical properties; therefore, this study proposes the use of machine learning for IPHTSG forecasting by systematically analysing the engineering parameters and production metrics. First, an IPHTSG database is established by categorizing and compiling the collected engineering and production parameters in addition to the classified initial production data. Second, on the basis of the IPHTSG database, prediction models for the IPHTSG are developed by employing various machine learning algorithms. The dimensionality of the input data is reduced via correlation analysis of the feature parameters, and the parameters of each prediction model are optimized using a grid search and 10-fold cross-validation. Finally, the models are applied to make predictions on a test set to validate their reliability, forming a set of methods and procedures for IPHTSG prediction. Then, this work describes a case study that was conducted on the tight gas reservoir of the H8 Member in the Sulige Southeast Field (Ordos Basin). The effective reservoir length, vertical thickness, open-flow capacity, bottom hole pressure, and amount of sand inclusion from 155 horizontal wells were selected as feature parameters, with data from 140 wells used as the training set and data from 15 wells used as the test set. Six machine learning algorithms were utilized to establish models, and the relevant calculation indicators of different models are compared. Ultimately, the XGBoost prediction model, which exhibits superior performance, is selected. This model achieves a training accuracy of 95% and a testing accuracy of 93.33%, with precision, recall, and F1-score values of 95%, 94.12%, and 93.14%, respectively, and it also has a relatively short training time. The method proposed in this paper successfully realizes IPHTSG prediction, providing a decision-making basis for formulating reasonable development plans and optimizing production parameters. This interdisciplinary methodology provides a replicable template for data-intelligent decision-making in tight gas reservoir management.