Abstract: In view of the problems of existing tunneling face temperature prediction methods, such as weak generalization ability, poor robustness, and limited predictive capacity for nonlinear multidimensional data, a tunneling face temperature prediction model based on ensemble learning-enhanced Back Propagation Neural Network (BPNN), namely t-SNE-BPNN-AdaBoost, was proposed. First, the t-Distributed Stochastic Neighbor Embedding (t-SNE) nonlinear dimensionality reduction technique was adopted to reduce seven high-dimensional features, including air volume, temperature, and relative humidity in front of the ventilator, to three dimensions, retaining the local structure of data and removing noise. Then, the reduced-dimensional data were input into BPNN as the base classifier, and the preliminary model was obtained through iterative training. Finally, ensemble learning was carried out by Adaptive Boosting (AdaBoost), in which multiple weak BPNN classifiers were iteratively trained and combined into a strong classifier by weighted integration, thereby enhancing the generalization ability of the model. Sixty sets of measured tunneling face data were divided into training and testing sets at a ratio of 8:2, and 5-fold cross-validation was conducted to determine that the optimal number of AdaBoost weak learners was 30. The experimental results showed that: ① the prediction curve of t-SNE-BPNN-AdaBoost fit the true values best, with the smallest overall error, strong adaptability in sudden temperature change intervals, and stability far superior to Support Vector Machine (SVM), BPNN, and t-SNE-BPNN. ② The relative prediction error of t-SNE-BPNN-AdaBoost was the smallest, almost all within 5%, demonstrating the best prediction accuracy. ③ On the test set, the coefficient of determination of t-SNE-BPNN-AdaBoost was 0.9784, which was improved by 60.3%, 17.2%, and 8.1% compared with SVM, BPNN, and t-SNE-BPNN, respectively. The Mean Absolute Error (MAE) was 0.1676, the Mean Squared Error (MSE) was 0.0567, and the Mean Absolute Percentage Error (MAPE) was 0.9640. All metrics were significantly better than those of SVM, BPNN, and t-SNE-BPNN, and the adaptability in sudden temperature change intervals was stronger.