Hybrid Neural Networks for Production Rate Forecast in Unconventional Reservoirs

Abstract Unconventional reservoirs, such as shale formations and tight sands, play a critical role in global oil and gas production. However, their complex structures, characterized by intricate fracture networks and a wide range of physical events, challenge conventional modeling methods. Empirical and analytical approaches often fail to achieve high accuracy due to their reliance on oversimplified assumptions. This study introduces a hybrid workflow that integrates the Fast-Marching Method (FMM) with a Long Short-Term Memory (LSTM) model to forecast production rates in large, heterogeneous unconventional reservoirs with parameters such as well length, fracture half-length and fracture spacing. By leveraging the resulting FMM-based drainage pore volume derivative as an input feature, the LSTM model effectively captures the non-linear relationships governing production performance. Validation using a synthetic heterogeneous reservoir demonstrates strong prediction accuracy, with average training and testing R2 scores of 0.94 and 0.92, respectively, indicating a robust match between reservoir simulation results and model forecasts. This paper also demonstrates that initial production data from appraisal increases test R2 scores by +0.04. Additionally, the proposed workflow is significantly more computationally efficient than commercial simulators, reducing forecast time by up to 50 times. This novel approach enhances predictive accuracy and efficiency, providing a cost-effective tool for field development planning and aiding in optimal decision-making for well placement and production strategy design.