Abstract: During high-speed flight, severe aerodynamic heating is encountered by the vehicle, and the accurate identification and prediction of surface heat flux are regarded as essential requirements for the design and optimization of thermal protection systems. Real time feedback on the spatiotemporal evolution of active thermal protection systems such as transpiration cooling is difficult to obtain by conventional approaches, and complex internal processes including microporous permeation and gas liquid phase change introduce significant thermal lag effects. These effects further increase the difficulty of real time estimation and prediction and reduce the responsiveness of coolant dynamics in active thermal protection systems. In this study, a heat flux identification and rapid prediction method for transpiration cooling systems is proposed based on a Channel Attention Temporal U-Net model, in which spatial features are extracted by convolutional networks, temporal dependencies are modeled by convolutional long-Short-Term-Memory units, and channel features are adaptively calibrated through squeeze and excitation modules. Through this integrated structure, the accuracy of spatiotemporal heat flux modeling is enhanced and both the current heat flux and its future evolution can be predicted with high fidelity. It is shown that under typical transpiration cooling conditions a root mean square error of 0.0189 MW/m² is achieved by the proposed model, corresponding to an 84.7% reduction relative to the baseline U-Net model, and the prediction error within a 10-30 s forecast window is maintained within 2.5%. The influence of thermal lag on system performance is thereby alleviated, and improvements in accuracy, robustness, and computational efficiency are demonstrated, providing theoretical support and technical foundations for real time dynamic thermal management in high-speed vehicle active thermal protection systems.