Hybrid Deep-Learning Model to Forecast the Shale Gas Production Based on the Decomposition-Reconstruction Principle

Abstract

Abstract. A multi-stage hydraulic fracturing technique and horizontal well drilling are necessary for effective gas production because of the ultralow porosity and permeability. A reasonable development plan must have an accurate estimate of gas output. Due to the intricate nature of the hydraulic fracture network and gas flow mechanism, the physics-based neural network model for gas production prediction is currently being developed. The data-driven model offers an alternative to solving the production forecast problem. In this study, we first evaluated the data using advanced decline curve analysis and individual deep-learning models to forecast the production performance. The low accuracy of these models is mainly due to the nonlinearity and nonstationary nature of the datasets. Hybrid models have been developed to improve accuracy by integrating deep-learning models (long short-term memory (LSTM) and gated recurrent units) with decomposition techniques like empirical mode decomposition, ensemble empirical mode decomposition, and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN). The performance of these models is observed to be very good, but the computational cost is also high. To minimize the computational cost, sample entropy (SE) has been used in the novel framework (CEEMDAN-SE-LSTM), which effectively forecasts the dynamic production data. Its prediction performances are assessed using the evaluation metrics after a detailed analysis of the appropriate window size and hyperparameters. The proposed model is validated with two different shale gas production datasets, and reliability is evaluated on a large dataset from a conventional reservoir. The superiority and applicability of the proposed hybrid model on dynamic production datasets with the capability to capture operational disruptions, such as well shut-in and well interventions, will help to give valuable information for production enhancement. This research work represents the first application of this approach to analyze the nonlinear and nonstationary dynamics inherent in shale production datasets. The proposed framework is specifically designed to successfully tackle the highly nonstationary and nonlinear nature of shale gas production data.