Prediction model of coal spontaneous combustion temperature based on data filling
-
Graphical Abstract
-
Abstract
Most of the existing coal spontaneous combustion temperature prediction models are based on relatively complete index gas sample data. However, the index gas data are affected by instruments or human factors. There are often data missing phenomena, resulting in low accuracy and over-fitting of coal spontaneous combustion temperature prediction. In order to solve the above problems, the paper proposes to apply filling algorithms such as K-nearest neighbor algorithm (KNN), random forest algorithm (RF), decision tree algorithm (DT) and support vector regression algorithm based on particle swarm optimization (PSO-SVR) to fill in the missing values. The missing data and the filled data are trained by RF, SVR and extreme gradient boosting (XGBoost) algorithm respectively. The parameters are optimized by the PSO algorithm. The RF, XGBoost and SVR coal spontaneous combustion temperature prediction models based on data filling are constructed. CO, CO2, CH4, C2H6 and O2 are selected as index gas in coal spontaneous combustion experiment, and six kinds of random data missing are designed. The overall missing rates are designed as 10%, 20% and 30%. The missing rates of CO and CO2 are designed as 40%, 50% and 60%. The average absolute error percentage (MAPE) is used as the filling effect evaluation index. The MAPE, the judgment coefficient R2 and the root mean square error (RMSE) are used as the model performance evaluation index. Four filling algorithms and three prediction models are compared. The results of the comparative analysis show the following points. The DT filling algorithm has better filling effect than the other three algorithms in six kinds of missing data cases. When there are more missing values of CO and CO2, the MAPE between the filling value and the actual values of the RF algorithm is large. The XGBoost model works extremely well in the training set without adjusting the parameters, but it is very prone to overfitting. The prediction effect of SVR model is very poor and the model cannot meet the prediction requirements. In the case of six kinds of data missing, the MAPE of PSO-SVR, RF and PSO-RF coal spontaneous combustion temperature prediction models based on the DT filling algorithm are about 4%. The RF model based on the DT filling algorithm can predict the coal spontaneous combustion temperature without optimization and has good stability.
-
-