A Novel Multidimensional Hybrid Machine Learning Model for CO₂ Injection to Separate Coalbed Methane: Comprehensive Prediction of Methane Diffusion Rate, Production Volume, and CO₂ Sequestration

In the process of replacing coal seam CH4 with CO2, the CH4 diffusion rate, production, and CO2 storage determine the effectiveness of the replacement. Predicting the changes in CH4 diffusion rates, production, and CO2 sequestration during the CO2 injection process in coal beds, based on historical data, presents a highly nonlinear challenge influenced by various gas injection parameters. In this study, we developed a model based on time window segmentation, key feature extraction using support vector machines-recursive feature elimination (SVM-RFE), and optimization through the Lizard Tail Split Algorithm (LTSA) for an Enhanced Learning Machine (ELM), named SVM-RFE-LTSA-ELM. For the same characteristic variables, our proposed SVM-RFE-LTSA-ELM model accurately predicted the CH4 diffusion rate, yield, and CO2 sequestration, outperforming five baseline machine learning models: linear regression, random forest, support vector machine, decision tree, and gradient boosting. Furthermore, the model can reliably extrapolate to data beyond the experimental range using historical data. The proposed model outperforms LTSA-ELM, ELM, and other models lacking engineering optimization, demonstrating exceptional generalization and predictive capabilities. After 10-fold cross-validation, the model consistently exhibits excellent stability and reliability. Our innovative Lizard Tail Splitting heuristic algorithm dynamically adjusts the number of hidden neurons in the extreme learning machine, thereby enhancing the model’s generalization and predictive accuracy. The model’s extrapolation capability can improve energy efficiency and reduce CO2 emissions, contributing to energy savings and emission reductions.