Recurrent Neural Networks (RNNs) have been widely adopted as effective alternatives to conventional decline curve analysis for forecasting oil and gas production. However, these models often lose early training information and treat all time-series da...
Recurrent Neural Networks (RNNs) have been widely adopted as effective alternatives to conventional decline curve analysis for forecasting oil and gas production. However, these models often lose early training information and treat all time-series data as continuously periodic, which limits their ability to account for abrupt production changes in shale gas caused by sporadic non-periodic features. Previous studies have mitigated the uncertainty in RNNs-based shale gas production forecasting and evaluated their applicability across diverse development sites by incorporating operational parameters of production wells or integrating machine learning algorithms that complement data interpretation. However, few studies have evaluated optimal modeling strategies by introducing diverse machine learning techniques based on their inherent data-processing principles, or explored long-term shale gas production forecasting by learning the influence of factors associated with early production decline through the linkage between time-series feature extraction and information selection techniques.
This study aimed to overcome the limitations of RNNs-based shale gas production forecasting by improving the learning architecture to transform early production decline trends into a series-based time-series representations, while selectively incorporating key features. To achieve this, 24 months of production history, shut-in period and monthly well operation days were used to forecast production up to 60 months, enabling the applicability of Bidirectional Gated Recurrent Unit (BiGRU) to be evaluated against Long Short-Term Memory (LSTM). In addition, the incorporation of the attention mechanism (AM) and decline characteristics resulted in prediction errors gradually converging to lower levels. Furthermore, by integrating Fourier Analysis Networks (FAN), which required validation in specific domains, into the data processing stage, the proposed model reduced the deviations from the measured production data by approximately 3~7 MMcf and demonstrated advantages over Time2Vec, thereby enabling the effective application of FAN to shale gas production forecasting. The FAN-integrated forecasting model demonstrated excellent applicability to data with large fluctuations of approximately 5~10 MMcf, reducing the mean absolute percentage error (MAPE) on average from 35.7% to 14.5% and normalized root mean squared error (NRMSE) from 40.8% to 16.1%, compared with those of the conventional LSTM-based model. The proposed approach can serve as a reliable tool for the development and operational planning of shale gas wells.