This study introduces the Spatio-Temporal Attention Enhanced Encoder-Decoder Damage Prediction Network (STAE-EDDPNet), an innovative deep learning model designed to enhance the predictive capabilities of coal-rock damage infrared temperature fields, which is crucial for the safe production in rock engineering and mining engineering. STAE-EDDPNet integrates a spatio-temporal attention mechanism, significantly improving the capture of complex nonlinear spatio-temporal information in rock infrared radiation. Compared with baseline models such as 3DCNN, ConvLSTM, and EDDPNet, STAE-EDDPNet demonstrated superior performance in both single-step and multi-step forecasting tasks. Test set results show that its predictive accuracy is 25.56% higher than 3DCNN, 5.69% higher than ConvLSTM, and 0.19% higher than EDDPNet. The study also found that the characteristics of brittle failure rock data significantly affect model training and predictive performance, providing a direction for future data collection and experimental design improvements. The introduction of STAE-EDDPNet not only promotes the application of infrared monitoring technology in the field of safety monitoring but also provides valuable reference for rock damage early warning.