Tar yield prediction of tar-rich coal based on geophysical logging data: Comparison between semi-supervised and supervised learning

Abstract

Accurate and economical prediction of tar yield is essential for precise evaluation of tar-rich coal geological distribution. Through the correlation analysis of conventional logging parameters and tar yield, this identified the highly correlated logging parameters including high-definition deep investigate resistivity log (HLLD), density (DEN), caliper (CAL), acoustic (AC), natural gamma ray (GR), and spontaneous potential (SP). Based on them, the predictive models for tar yield have been built by semi-supervised learning and supervised learning methods, and a comparison was made. First, applying the self-training algorithm based on a semi-supervised learning framework, this research has built a Semi-Supervised Recurrent Neural Network (SSRNN) model for the prediction of tar yield. Second, based on supervised learning, this research has built tar yield prediction models, such as backpropagation Neural Network (BPNN), Support Vector Regression (SVR), and Random Forest (RF). Semi-supervised learning can effectively utilize unlabeled data to enhance model performance and address the problem of expensive access to labeled data. Supervised learning, based on mapping inputs directly to outputs, offers a clear and intuitive training process, making it ideal for tasks with well-defined input-output relationships. This paper builds a tar yield prediction model for multiple drilling wells in the Santanghu Basin, utilizing 121 labeled and 1952 unlabeled data sets, of which the labeled data were used for supervised learning, and the unlabeled data were employed for semi-supervised learning. In addition, the generalization abilities of different prediction models were evaluated by the use of 48 labeled data from the GM2 well. The results indicated that the SSRNN model, which demonstrates better generalization capability, is superior to the RNN, BPNN, SVR, and RF models in performance.