Enhancing Production Forecasting in Fractured Shale Gas Reservoirs Using a Hybrid ARIMA-EMD-LSTM Approach

Abstract

Abstract This study presents a robust machine-learning protocol for predicting production from fractured shale gas reservoirs. The field data of an active shale gas reservoir in Sichuan Basin was utilized to construct and validate the proposed method. The main objective is to establish a framework to rapidly and accurately forecast the production, which aides in assessments of economic viability for shale gas reservoir developments. This study utilized field production data from a representative shale gas reservoir. Statistical analysis and sensitivity analysis were conducted to characterize the input-output data structure. To train the prediction model, a novel hybrid approach was proposed, combining the Long Short-Term Memory network (LSTM) with the Autoregressive Integrated Moving Average model (ARIMA) and Empirical Mode Decomposition (EMD).Shale gas production is influenced by numerous factors, leading to historical production sequences that exhibit significant nonlinearity and abrupt fluctuations. Consequently, precise prediction using conventional approaches becomes challenging. By applying the proposed workflow, the ARIMA model was applied to filter both linear and nonlinear features from the original production data. ARIMA was utilized to fit and predict the linear components, while the nonlinear components were passed through EMD. EMD was employed to extract Intrinsic Mode Functions (IMFs), which represent local characteristic signals at different time scales within the nonlinear components. Multiple Long Short-Term Memory (LSTM) models were then trained to predict the various IMFs. The hyperparameters of the LSTM models were optimized to enhance performance. The LSTM predictions were combined with the ARIMA results, resulting in a comprehensive production forecasting model. To validate the proposed hybrid method, a comparison was made with an existing approach LSTM. The superiority and feasibility of the ARIMA-EMD-LSTM model were confirmed, as it demonstrated strong generalization ability and increased prediction accuracy by effectively capturing the input-output relationship. It was proved that this framework can enable rapid and accurate forecasting of shale gas production. This study introduces a novel hybrid ARIMA-EMD-LSTM model, which represents the first application of this approach to fractured shale gas production. The ARIMA effectively separates the linear and nonlinear components of the original data, while the EMD decomposes the nonlinear parts into IMFs. The proposed workflow could enhance the prediction accuracy and intensively extend its feasibility to gas production forecasting in fractured shale reservoirs.