Non-linear soft sensing method for temperature of coal spontaneous combustion

Abstract

The accurate measurement of the temperature of coal spontaneous combustion is a key technology for coal spontaneous combustion and fire prevention and is important to realising safe coal mining and reducing energy losses. A soft sensor model based on a hybrid-kernel-function support vector machine with an improved genetic algorithm is proposed for the nonlinear relationship between the concentration of a characteristic gas of coal spontaneous combustion and the temperature of the coal body. The natural oxidation process of the coal body is simulated using a coal-spontaneous-combustion oxidation warming experiment bench. The gas concentration and maximum temperature of the coal body at different moments in an oxidation warming experiment are taken as the model input data. Various soft sensor models are automatically optimised by a genetic algorithm in the training process, and the performances of the models are compared and evaluated in the field. The results show that the evaluation indexes of the hybrid-kernel-function model are better than those of the single-kernel function. In the training process, the evaluation indexes of the model with hyperparameters automatically optimised by the genetic algorithm are better than those other models, further verifying the nonlinear relationship between the gas concentration and coal temperature in the process of coal spontaneous combustion and oxidation. The use of the improved support vector machine model basically results in no deviation between the soft sensor results and actual temperature data in the field test, indicating that the soft sensor model has high accuracy, generalisation ability and applicability. The improved support vector machine model has a mean absolute percentage error of ≤ 1.616%, mean absolute error of ≤ 0.533 ℃, and root-mean-square error of ≤ 0.608 ℃. These results indicate that the soft sensor problem of coal spontaneous combustion can be considered a typical small-sample regression method. The proposed method can be further applied to similar problems in the energy industry.