Prediction of Total Organic Carbon Content in Deep Marine Shale Reservoirs Based on a Super Hybrid Machine Learning Model

The total organic carbon (TOC) content is crucial for assessing the gas-bearing potential of shale reservoirs. Thus, quantitative characterization and intelligent prediction of TOC content play important roles in determining geological sweet spots and the development of shale reservoirs. Unfortunately, directly obtaining TOC content data in deep marine shale reservoirs is difficult, and the accuracy of indirect prediction remains insufficient. To efficiently and accurately predict TOC content, we propose a super hybrid prediction model, CVMD-CNN-BiLSTM-AT, which integrates correlation variational modal decomposition (CVMD), a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism (AT). The model employs CVMD to remove noise signals from the original TOC sequence, decomposes the denoised sequence into stable subsequence components, and a CNN-BiLSTM prediction model is constructed for each one. In addition, we incorporate AT to assign the hidden layer probability weights of BiLSTM, which makes the model focus on high-importance features and assigns weights accordingly. Finally, the predicted subsequences are combined and reconstructed according to the decomposition law to obtain the final TOC content. Herein, 1007 core samples and their related well logging data were collected from 13 typical wells, among which data from 705 samples were utilized to train the model and the remaining data were utilized to validate and test the model. The study results indicate that the CVMD-CNN-BiLSTM-AT model has excellent and reliable predictive ability, with an R2 of 0.967 and can accurately predict TOC content. This achievement can provide adequate technical support and insight for deep marine shale gas exploration and development.