Deep-Learning Based LSTM for Production Data Analysis of Hydraulically Fractured Wells

During the production and operations of hydraulically fractured wells, large amounts of data are collected through numerous sensors or flowmeters, which can provide valuable understanding on the formation and hydraulic fractures. Although much studies try to use physical-justification based approaches to analyze these well history data, the analysis accuracy is significantly limited due to many assumptions made in physical models. This paper developed a deep-learning based Long Short-term Memory (LSTM) approach for production data analysis in shale reservoir and proposed a workflow to quantitatively evaluate fracture parameters. The proxy model is based on deep-learning algorithm of LSTM and is combined with a semianalytical (base) model for multiphase water and hydrocarbon (oil or gas) flow in the hydraulically fractured reservoirs. To rigorously consider the multiphase flow mechanism in the semi-analytical model, LSTM and attention mechanism are introduced to forecast the key relationship of average saturation and pressure for semi-analytical model by training and predicting the time-dependent pressure and saturation series. We generated thousands of numerical simulation cases of wells in hydraulically fractured rservoirs, which provide production data and static reservoir data to train the deep-learning based proxy model. Model verification and comparison show that the proxy model can effectively predict pressure-dependent average saturation relationship with high accuracy. The numerical validation confirms the superiority of the proposed deep-learning based model than the semi-analytical model in accuracy with the error of estimated reservoir and fracture parameters less than 10% and in calculation efficiency with the speed two orders of magnitude faster.