The tunneling working face is an area prone to frequent gas accidents during coal mining. In view of the strong nonlinear and spatiotemporal dynamic correlation characteristics of gas concentration variations, a prediction model for gas concentration at the tunneling working face, named KAN-iTransformer, was proposed, which integrates the Kolmogorov-Arnold Networks (KAN) and an improved iTransformer (Inverted Transformer). Firstly, a relevant feature dataset for gas concentration prediction was constructed through data preprocessing and Spearman correlation coefficient analysis. Secondly, based on the inverted Transformer structure, the complete variable sequence was regarded as a token, and a graph attention mechanism was introduced to explicitly model the spatial topological relationships of the sensor network, enabling effective modeling of the dynamic dependencies among sensors. Thirdly, the learnable activation functions of KAN were used to replace the traditional linear weights, enhancing the model's ability to characterize complex nonlinear relationships among multi-source monitoring parameters. Finally, a hierarchical response mechanism combining point prediction and interval estimation was proposed to enable advanced dynamic hierarchical response to gas overrun risk. The field test results show that the MSE, MAE and MAPE of the improved KAN-iTransformer model in single-step prediction are 0.000 307, 0.012 921 and 2.321 373%, respectively, significantly outperforming both the iTransformer and Transformer models. In multi-step prediction, the MSE of the improved KAN-iTransformer model is 0.000 913, which is 63.7% and 80.9% lower than those of the comparison models, respectively, demonstrating better long-term prediction performance and effectively realizing accurate prediction and dynamic early warning of gas concentration at the tunneling working face.