Cased-Hole Formation Density Prediction Based on CNN-Transformer Model

Formation density measurement of cased holes is a vital reference parameter for formation evaluation. The traditional calculation method of formation density is greatly affected by the thickness of cement sheath, and the calculation accuracy is relatively poor. There is a nonlinear relationship between the formation density and the detector count rate of different energy windows. Considering that convolution neural network (CNN) is one of the most often used techniques to solve the nonlinear regression problem, a system structure combining CNN and Transformer is proposed to achieve the purpose of automatic formation density prediction. Besides, the new S-ReLU function is added to CNN-Transformer model. The count rates of four energy windows for different detector are adopted as the inputs of CNN-Transformer model. In comparison with BP, GBDT, and CNN models, the prediction accuracy is evaluated. The findings demonstrate that CNN-Transformer model is more robust than conventional network models at predicting formation density, and the predicted formation density differs less from the actual formation density. In test set, the error value is only 0.041, and the predicted numerical correlation reaches 0.941, which achieves an excellent fitting effect in the prediction.