Abstract: In the underground coal gasification (UCG) process, different operating parameters and coal properties can cause the UCG system to exhibit different gasification processes, and the calorific value of the product gas will also show a decreasing trend after a certain time of stable output. Two coaxial coal underground gasification models are constructed in an artificial coal seam to study the effects of coaxial gasification channels on the gas composition and calorific value of the generated gasification products, and the effects of pressure, gasifier flow rate, and gas components on the calorific value are comparatively analyzed. The Informer calorific value prediction model is proposed, using E_RMS, E_MA and R² as evaluation indexes, and the results are compared with four machine learning models (LSTM, MLP, RNN, and ARIMA) to comparatively analyze the actual and predicted values and errors of the calorific value of the product gas under different prediction time lengths. The results show that different coaxial gasification channels can affect the gasification effect, and changing the extended length of gasification channels can change the proportion of combustible gas components in the product gas. In coaxial model 1, the volume fraction of effective gas components (CO, H₂ and CH₄) in the product gas was 33.45%, and the average calorific value was 4.68 MJ/Nm³; in coaxial model 2, the volume fraction of effective gas components in the product gas was 35.09%, and the average calorific value was 4.75 MJ/Nm³. Compared with the coaxial model 1, the volume fraction of H₂ in the coaxial model 2 increased from 6.17% to 8.10%, while the volume fraction of CH₄ decreases from 1.19% to 0.61%. Adjusting the gasifier flow rate changes the gasifier equilibrium, which transiently increases changes the calorific value of the product gas. Compared with other reference models, the E_MA decreased by 22.42%-42.78%, and the E_RMS decreased by 15.38%-30.49%. The Informer calorific value prediction model performed outstandingly with high prediction accuracy. The model has high prediction accuracy across different datasets, prediction lengths, and sampling frequencies. The average error for the coaxial model 1 dataset is in the range of 8%-18%, and the average error for the coaxial model 2 dataset is in the range of 5%-12%. The Informer model is able to effectively predict the different trends of thermal value changes when predicting thermal value curves in different time periods, and it can also predict the performance of the system after parameter adjustment.