Abstract

Shale gas production is governed by complex geological and engineering factors, and its production dynamics are often highly variable. Conventional methods, which can incorporate only a limited number of production-related variables, often struggle to provide accurate forecasts under fluctuating operating conditions. Focusing on the natural flowing stage of shale gas wells, this study proposes a probabilistic forecasting framework that integrates physical decline characteristics with dynamic production data. A dual-branch TCN–LSTM network constrained by decline features is constructed, and Student’s t-distribution is introduced to quantify the uncertainty caused by short-term production fluctuations. The results show that embedding physical decline constraints into the deep learning architecture helps bridge the gap between conventional models with limited parameter representation and purely data-driven models with insufficient interpretability. The proposed method improves forecasting accuracy while preserving the physical meaning of the predictions, and it can generate noise-robust confidence intervals with stable coverage. This method provides decision support for short-term production tracking and production-regime adjustment in shale gas wells.

IPC Classification

Keywords