Research on Synergistic Prediction Method of Gas Emission from Working Face Based on Feature Selection and BO-GBDT

To rapidly and accurately predict the gas emission volume in mine working faces, a Gradient Boosting Decision Tree (GBDT) model for gas emission prediction was established. Given the numerous influencing factors of gas emission in working faces, data set redundancy, and the need to reduce subsequent data collection intensity, five feature selection algorithms were applied to filter features from the data set. By analyzing the fitness, computation time, and prediction outcomes of each feature combination within the GBDT model, the wrapper method was identified as the optimal feature selection algorithm. The study revealed that the number of features does not necessarily correlate positively with prediction accuracy and generalization; instead, the presence of redundant or irrelevant features can detract from the model's predictive accuracy.To further enhance model precision, five hyperparameter optimization algorithms were utilized to tune the GBDT model. Comparative analysis of each hyperparameter combination's prediction outcomes indicated that while the optimization algorithm itself had a limited impact on the GBDT model's accuracy and generalization, the optimal hyperparameter combination derived from Tree-structured Parzen Estimator-based Bayesian Optimization (BO) achieved the highest accuracy with a relatively short optimization time, demonstrating the best optimization performance. Consequently, a BO-GBDT model was established.After dividing the feature-selected data set into training and testing sets, the BO-GBDT model was employed to predict gas emission in working faces. Comparisons with Random Forest, Support Vector Machine, and Neural Network models revealed that the GBDT model exhibited superior accuracy and generalization, with an average relative error of 2.7%, representing reductions of 39.73%, 40.13%, and 22.41% respectively compared to the 4.48%, 4.51%, and 3.48% relative errors of the Random Forest, Support Vector Machine, and Neural Network models. This demonstrates that the GBDT model can effectively meet on-site engineering requirements and provide theoretical guidance for safe mine production.