Research on early warning model of coal spontaneous combustion based on interpretability.
Journal:
Scientific reports
Published Date:
May 29, 2025
Abstract
Predicting the temperature of the coal spontaneous combustion (CSC) is essential for preventing and managing coal mine fires. In this paper, a Rough Set-Stacking-SHapley Additive Explanations (RS-Stacking-SHAP) prediction model of CSC based on grid search optimized is proposed. Compared with the traditional machine learning model, the model has better prediction accuracy and generalization ability. Based on the data collected from experimental coal samples in Lijiahao Coal Mine, rough set algorithm was used for attribute approximation to identify O, CO, CH, CH, CH, CH, CO/CH, CH/CH as the model indexes, thereby establishing the system of warning indexes for spontaneous combustion of coal. XGBoost, SVR, RF, LightGBM and BP models were selected as base models to establish an early warning model for CSC based on the stacking integration architecture. The grid search algorithm was utilized to optimize the model parameters, ensuring the selection of the most suitable parameter configurations. The dataset was then divided into the training and test sets in a 7:3 ratio, and the extracted indicators of each gas were used as inputs to the model and the temperature was used as outputs. The mean absolute error (MAE), root mean square error (RMSE), r-square (R), mean absolute percentage error (MAPE), weighted mean absolute percentage error (WMAPE) and variance account for (VAF) were chosen to evaluate the results. The predictive performance of the model was compared with that of the individual base models, and the results displayed that the R value of the RS-Stacking model was 0.991, representing improvements of 12.7%, 14.1%, 0.6%, 3.5% and 17.7% over the XGBoost, SVR, RF, LightGBM, and BP models, respectively. GS-RS-Stacking was considered to be the best model, where MAPE = 5.14%, WMAPE = 3.76%, VAF = 99.08%, MAE = 5.081, RMSE = 6.461, close to the ideal value. Finally, we used SHAP to provide global feature interaction interpretation and local interpretation for the model, analyzing the contributions of CH, CH, CH, and CO to the model's predictive outcomes. The results show that the model proposed in this paper has better prediction effect and robustness for temperature of CSC.
Authors
Keywords
No keywords available for this article.
