Machine Learning-Driven Early Productivity Forecasting for Post-Fracturing Multilayered Wells

Hydraulic fracturing technology significantly enhances reservoir conductivity by creating artificial fractures, serving as a crucial means for the economically viable development of low-permeability reservoirs. Accurate prediction of post-fracturing productivity is essential for optimizing fracturing parameter design and establishing scientific production strategies. However, current limitations in understanding post-fracturing production dynamics and the lack of efficient prediction methods severely constrain the evaluation of fracturing effectiveness and the adjustment of development plans. This study proposes a machine learning-based method for predicting post-fracturing productivity in multi-layer commingled production wells and validates its effectiveness using a key block from the PetroChina North China Huabei Oilfield Company. During the data preprocessing stage, the three-sigma rule, median absolute deviation, and density-based spatial clustering of applications with noise were employed to detect outliers, while missing values were imputed using the K-nearest neighbors method. Feature selection was performed using Pearson correlation coefficient and variance inflation factor, resulting in the identification of twelve key parameters as input features. The coefficient of determination served as the evaluation metric, and model hyperparameters were optimized using grid search combined with cross-validation. To address the multi-layer commingled production challenge, seven distinct datasets incorporating production parameters were constructed based on four geological parameter partitioning methods: thickness ratio, porosity–thickness product ratio, permeability–thickness product ratio, and porosity–permeability–thickness product ratio. Twelve machine learning models were then applied for training. Through comparative analysis, the most suitable productivity prediction model for the block was selected, and the block’s productivity patterns were revealed. The results show that after training with block-partitioned data, the accuracy of all models has improved; further stratigraphic subdivision based on block partitioning has led the models to reach peak performance. However, data volume is a critical limiting factor—for blocks with insufficient data, stratigraphic subdivision instead results in a decline in prediction performance.