现有采空区遗煤温度预测研究多侧重于温度与气体浓度之间的关系，较少考虑采空区内遗煤温度与距工作面距离及漏风风速之间的复杂非线性关系。针对该问题，提出了一种基于蚁群优化算法优化核极限学习机（ACO−KELM）的采空区遗煤温度预测模型。在葫芦素煤矿21404工作面采空区布置束管及分布式光纤，对21404工作面采空区内O₂，CO，CO₂浓度和温度数据进行采集，同时结合采空区内漏风强度和距工作面水平距离构建KELM模型，通过ACO对KELM模型中的正则系数和核参数进行寻优，获得最优超参数组合，进而得到性能最优的KELM模型。与基于极限学习机（ELM）和基于随机森林（RF）算法的预测模型相比，ACO−KELM模型在测试集上的平均绝对误差为0.070 1℃，均方根误差为0.074 8℃，较基于ELM模型分别降低了65%和195%，较基于RF模型分别降低了53%和156%；ACO−KELM模型在测试集上的判定系数为0.963 5，与训练集的判定系数仅相差0.01，说明该模型未陷入过拟合且拟合程度较高。

Existing studies on the temperature prediction of residual coal in goafs have mainly focused on the relationship between temperature and gas concentration, with limited attention given to the complex nonlinear relationships between the residual coal temperature in the goaf, the distance from the working face, and the air leakage velocity. To address this gap, a prediction model based on ant colony optimization (ACO) and kernel extreme learning machine (KELM) (ACO-KELM) was proposed. In the 21404 working face goaf of Hulususu Coal Mine, beam tubes and distributed fiber optics were arranged to collect data on O₂ concentration, CO concentration, CO₂ concentration, and temperature within the goaf. Simultaneously, the air leakage intensity and horizontal distances from the working face were incorporated to construct the KELM model. ACO was employed to optimize the regularization coefficients and kernel parameters in the KELM model, thereby obtaining the best-performing hyperparameter combination and generating the optimal KELM model. Compared to the prediction models based on extreme learning machine (ELM) and random forest (RF) algorithms, the ACO-KELM model achieved an average absolute error of 0.0701 ℃ and a root mean square error (RMSE) of 0.0748 ℃ on the test set, reducing these errors by 65% and 195%, respectively, compared to the ELM-based model, and by 53% and 156%, respectively, compared to the RF-based model. The coefficient of determination (R²) for the ACO-KELM model on the test set was 0.9635, which was only 0.01 lower than that of the training set, indicating that the model was not overfitted and demonstrated a high degree of accuracy.