基于CNN-BiLSTM-Attention的TOC测井预测方法OA

To address the problem of limited generalization ability in traditional models for logging prediction of total organic carbon(TOC)content in the Longmaxi Formation shale reservoirs of the Sichuan Basin,this study systematically evaluates the performance of various deep learning models in both single-well and cross-well pre-diction tasks and identifies practical models suitable for different scenarios.Based on the acoustic time difference,density,and natural gamma-ray logging data from three wells(well1,well2,and well3)in the Longmaxi Forma-tion,multiple models are constructed,including multiple linear regression(MLR),support vector regression(SVR),convolutional neural network(CNN),bidirectional long short-term memory network(BiLSTM),and hy-brid CNN-BiLSTM and CNN-BiLSTM-Attention models.In the single-well prediction experiment,a random split strategy is applied to well1 for modeling and validation.The results show that the CNN model achieves the best per-formance,with the coefficient of determination(R2)reaching 0.9519 on the prediction set,demonstrating excel-lent local feature extraction capability and strong resistance to overfitting.To further evaluate model generalization ability,a leave-one-well-out(LOWO)cross-validation strategy is designed for cross-well prediction.The results indicate that the CNN-BiLSTM-Attention model exhibits the strongest generalization performance,achieving the highest R2 of 0.9653 on the prediction set,with mean absolute error(MAE)and root mean square error(RMSE)as low as 0.131%and 0.170%,respectively,which significantly outperforms other models.The attention mecha-nism effectively integrates the local features extracted by CNN with the long-term sequential dependencies cap-tured by BiLSTM,enhancing the model's ability to focus on key information and adapt to inter-well variations.This study verifies the effectiveness and robustness of deep learning models integrated with an attention mechanism for TOC prediction under complex geological conditions,emphasizes the importance of cross-well validation in practical applications,and provides a reliable methodological foundation for shale gas sweet-spot prediction.