Long short-term memory suggests a model for predicting shale gas production. (15th September 2022)

• Record Type: Journal Article
• Authors: Yang, Run; Liu, Xiangui; Yu, Rongze; Hu, Zhiming; Duan, Xianggang
• Is Part Of: Applied energy. Volume 322(2022)
• Publication Date: 2022-09-15
• Subjects: Shale gas -- Exponential smoothing method -- LSTM model -- Production prediction -- Deep learning; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.119415
• Languages: English
• ISSNs: 0306-2619

Abstract

Abstract: Predicting the production behaviors of shale gas wells is of great importance for further developing future unconventional hydrocarbon strategies. An accurate prediction production, as well as reliable shale gas production models, are required to fully understand the shale gas exploitation budget. However, a major problem with classical analytic methods is the insufficient accuracy of the existing models, the time-consuming collection of historical production data, and the costly computational expense. To minimize this problem, a combination of the exponential smoothing method, autoregressive integrated moving average (ARIMA) model, and long short-term memory (LSTM) model was proposed to provide robust support for the production behaviors of shale gas. In this paper, we employed shale gas well production data to establish a database for model training and optimized the predicted model. Hereby, we sought to evaluate the production data predicted by conventional analytical methods, the exponential smoothing method, the ARIMA model, and the LSTM model. Shortly afterward, we objectively compared the predicted results obtained by the novel LSTM model and traditional analytical methods, such as Arps, stretched exponential decline (SEPD), and the Duong model. Herein, we compared the computational cost between the LSTM model and traditional numerical simulation. The combined interpretation of the proposed model demonstrates that the LSTM model achieved scientific accuracy and outstanding results in both short-term and long-term predictions, and realized production prediction of the adjacent well, with excellent agreement with the real shale gas production and a low error, making it an effective tool in forecasting shale gas production. This assay could be used as a potential approach for evaluating deep learning in the petroleum industry and for predicting the future production of unconventional hydrocarbons. Graphical abstract: Highlights: A framework for shale gas production potential estimation was developed. A deep-learning-based time series analysis model was built to predict the short-term and longterm shale gas production, based on thousands of in-situ production data samples from three shale gas wells. Computational cost of the shale gas production prediction was significantly reduced with proposed long short-term memory model. Objectively compare the advantage and drawback forecasting production among conventional analytical methods, exponential smoothing algorithm, ARIMA, and LSTM model. The study shed lights on the few-shot learning applications in shale gas production prediction.