| To improve the accuracy and robustness of shale gas production capacity prediction models, a CNN-BiLSTM-Attention prediction model integrating Bayesian parameter optimization is proposed. This method combines the local feature extraction of convolutional neural networks, the temporal dependency capture of bidirectional long short-term memory networks, and the key feature weighting ability of the attention mechanism, achieving efficient extraction and representation of spatio-temporal coupling features of multi-source time series data from shale gas production. The Bayesian optimization algorithm is used to achieve adaptive global optimization of hyperparameters, further enhancing the stability and training efficiency of the model. The proposed method is compared and verified for its prediction performance using the Longning shale gas field dataset. The model achieves an MSE of 0.220, RMSE of 0.465, MAE of 0.339, R2 of 0.897, and MAPE of 5.71% on the test set, demonstrating significantly superior overall performance compared to single comparison models such as LSTM and Bi-LSTM; compared with the CNN-BiLSTM-Attention model, the MSE is reduced by 14.4%, the fluctuation range of prediction indicators is reduced by 15.2%, and the prediction deviation in the capacity decline stage can be controlled within 0.3×10?m3. The proposed model has good robustness to noise in shale gas production time series data, outstanding generalization ability, and provides stable and reliable technical support and feasible ideas for optimizing fracturing construction parameters, regulating production and recovery processes, and formulating development strategies in shale gas development. |
分享按钮