Enhancing shale gas production forecasting performance using bidirectional gated recurrent unit with integration of attention mechanism and Fourier analysis networks

Abstract

Recurrent Neural Networks (RNNs) often lose initial training information and treat all time-series data as continuously periodic, which limits their ability to account for production changes in shale gas caused by sporadic non-periodic features. 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 series-based time-series representations, while selectively incorporating key features. To achieve this, Bidirectional Gated Recurrent Unit (BiGRU) combined with the attention mechanism (AM) was applied to forecast production up to 60 months using data from Barnett Shale, including 24 months of production history, shut-in periods, monthly operation days, and decline characteristics. Furthermore, by comparing with Time2Vec, this study demonstrated the advantage of Fourier Analysis Networks (FAN) in capturing production decline from both periodic and non-periodic perspectives. Based on this capability, a novel approach, termed FAN-BiGAD, was proposed to address the limitations of RNNs-based models. Compared with the BiGRU-based model, FAN-BiGAD reduced the mean average absolute percentage error (MAPE) from 29.8% to 14.5% and the normalized root mean squared error (NRMSE) from 35% to 16.1%. In addition, an independent validation on the Marcellus Shale dataset yielded average MAPE and NRMSE values of 11.1% and 17.4%, respectively. Moreover, the cumulative production forecasts generated by FAN-BiGAD exhibited a high degree of linear agreement with the observed cumulative production in both development areas, resulting in R 2 values of approximately 0.99. The proposed approach may facilitate optimizing operational planning for shale gas wells, thereby preventing supply imbalance.