Integrated Deep Learning Framework for History Matching and Production Forecasting in Shale Reservoirs

Abstract

Shale reservoirs are vital energy resources, but their complex fracture networks, low porosity, and permeability make production forecasting highly uncertain. Traditional models like Arps decline and reservoir simulations struggle with shale's nonlinear, heterogeneous behavior, often leading to inaccurate forecasts. To address this, data-driven deep learning (DL) methods were applied using datasets from the Bakken (training/testing) and Eagle Ford (validation). Hybrid DL models (GRU, LSTM, TCN) were developed, with the TCN-GRU showing superior performance. Advanced hybrid frameworks integrating decomposition techniques (EMD, EEMD, CEEMDAN) with DL further improved prediction, though at a high computational cost. A novel CEEMDAN-SE-LSTM model reduced this cost while maintaining strong accuracy across shale and conventional datasets. In addition, a physics-informed neural network (PINN) was proposed to capture micro- and macro-scale fluid transport mechanisms (e.g., Knudsen diffusion, desorption, capillary effects) and predict dynamic gas composition shifts. Validated with field data, the framework enhances production forecasting, well performance evaluation, and fracture optimization, offering a robust tool for cost-effective shale reservoir development.