<p>Accurate prediction of outlet temperature and outlet pressure of heat exchangers is critical for production safety in industries such as chemical industry. However, traditional physical modeling requires domain experts to formulate equations and perform time-consuming parameter calibration, while single deep learning models (e.g., standalone CNN or RNN) exhibit limitations in simultaneously extracting both temporal and spatial features. To address these limitations, this study proposes a hybrid neural network model based on convolutional (CNN) + bi-directional gated recurrent unit fused with time-domain features and frequency-domain features, which combines the advantages of each of the Fourier transform, channel attention, and global attention mechanisms and improves the accuracy of the prediction of the data. The performance of this model is verified to be superior to other models by comparing the existing prediction methods with this paper’s method through publicly available datasets. Finally, the proposed model is evaluated by MAE and RMSE evaluation metrics using the operational data of a heat exchanger in a coal coking company as a case study. The results show that the proposed model achieves good accuracy in predicting the assessment metrics of heat exchanger outlet variables, and the reduction of root mean square error and mean absolute error is remarkable, which meets the demand for real-time and accurate safety production prediction in the chemical industry.</p>