Prediction method of gas emission in working face based on feature selection and BO-GBDT

Gas emission in the working face is influenced by a variety of factors. Dimensionality reduction methods, such as Principal Component Analysis, can reduce computational resources but may alter the original feature structure, leading to a loss of some detailed information in the dataset. To address this issue, a gradient boosting decision tree (GBDT) model for gas emission prediction was developed. Five feature selection algorithms were applied to filter the dataset, and the model fit, computational time, and prediction accuracy of each feature combination in the GBDT model were analyzed. The wrapping method was identified as the most effective feature selection algorithm. Based on field conditions, 8 optimal features were selected for prediction. The results indicated that the number of features did not necessarily correlate with the prediction's accuracy or generalization capability. In fact, redundant or irrelevant features reduced the model's prediction accuracy. To further improve performance, five hyperparameter optimization algorithms were applied to the GBDT model. A comparative analysis of prediction performance for each hyperparameter combination was conducted. The results showed that the optimization algorithm itself had minimal impact on the accuracy and generalization of the GBDT model. However, the optimal hyperparameter combination, obtained through the tree-structured Parzen estimator (TPE) based Bayesian optimization (BO) algorithm, provided the highest accuracy and relatively short optimization time, yielding the best optimization performance. Thus, the BO-GBDT model was established. After feature selection, the dataset was divided into training and testing sets, and the BO-GBDT model was used to predict gas emission in the working face. Comparison with random forest, support vector machine, and neural network models showed that the BO-GBDT model achieved the highest accuracy and generalization, with an average relative error of 2.61%. This was 35.56%, 37.41%, and 32.03% lower than the random forest, support vector machine, and neural network models, respectively. The BO-GBDT model meets the field engineering application requirements and provides theoretical guidance for ensuring safe mining production.