Improving Shale Gas Production Forecasting through the Integration of Bidirectional Gated Recurrent Unit and Attention Mechanism

Recurrent neural networks (RNN) are widely used for predicting oil and gas production.However, in the analysis of shale gas production, these models often suffer from the loss of training information.This study aimed to enhance the learning performance of a bidirectional gated recurrent unit (BiGRU) and improve the accuracy of shale gas production forecasting.To achieve this, we evaluated the limitations and applicability of BiGRU by comparing its performance with that of long short-term memory (LSTM).To enhance the performance of BiGRU, the decline characteristics were considered using the attention mechanism.This resulted in the prediction errors clustering tightly at lower levels, reducing the average mean absolute percentage error (MAPE) and normalized root mean squared error (NRMSE) from 35.7% and 40.8% to 16.2% and 18.2%, respectively, compared with those of the LSTM-based model.This study demonstrated that reliable production forecasting can be achieved by effectively reflecting the characteristics of various types of shale gas production data.