Physics-Data Hybrid Productivity Prediction Considering Realistic Fracture Geometry for Tight Sandstone Hydraulic Fracturing

Multistage hydraulic fracturing in horizontal wells induces significant spatial heterogeneity in fracture networks because of the complex interactions between reservoir geological characteristics and fracturing operations. This heterogeneity is difficult to capture using conventional productivity models, even when realistic fracture geometry is considered, resulting in large prediction deviations. To address this issue, this study proposes a physics-data hybrid productivity prediction model optimized by a production-informed loss function, with a tight sandstone gas reservoir in the Ordos Basin used as the case study. By integrating field data with asymmetric-fracture numerical simulations, eight controlling parameters, including formation pressure and reservoir thickness, were identified through weighted sensitivity analysis. On this basis, a refined productivity forecasting model was established by combining the Goose Optimization Algorithm (GOOSE) with long short-term memory (LSTM). Four physics-data hybrid models were developed based on the GOOSE-LSTM framework. Among them, the loss-function-optimized model exhibited the best performance, achieving an R2 of 0.953 and a prediction error below 0.05. The proposed methodology provides a refined decision-making tool for fracturing scheme optimization and development plan formulation, thereby improving productivity forecasting accuracy and supporting the efficient development of target horizontal-well intervals.