Abstract:

In order to predict the spontaneous combustion period of coal quickly and accurately, firstly, based on the large-scale coal spontaneous combustion low-temperature oxidation test and literature data, the data was composed, and considering the many influencing factors of coal spontaneous combustion and the complex nonlinear relationship with the combustion period, the data set including the spontaneous combustion period of coal, ambient temperature, coal calorific value, moisture and other parameters was established. Secondly, machine learning models such as MLP and RF are used to establish a coal spontaneous combustion period prediction model to characterize the influence of internal and external factors on the combustion period. At the same time, in order to enhance the fitting ability and generalization ability of the model, feature engineering is used to study the correlation of feature variable to filter the input features of the model. Then, the gridsearch method is used to optimize the model hyperparameters to improve the prediction ability of the model. Finally, the learning curve method is used to evaluate the model state to prevent overfitting. The results show that both RF and MLP models can predict cola spontaneous combustion period, and the generalization capability of the RF model is higher. The mean absolute error (MAE) predicted by the RF and MLP models were 9.34 days and 12.10 days, respectively, indicating that the machine learning model can consider the complex effects of multiple internal and external factors.

Key words: coal spontaneous combustion period, machine learning, prediction model, random forest (RF), multilayer perceptron (MLP)