[Objective] Operational back pressure significantly impacts the coal consumption and generation efficiency of direct air-cooled units. Traditional data-driven prediction methods primarily rely on scalar time-series of historical process parameters

largely ignoring the spatial thermodynamic distribution information contained in infrared images of finned tube bundles. This omission limits the model's ability to capture complex spatial-temporal correlations under variable working conditions. This study aims to bridge this gap by developing a high-precision prediction model that effectively fuses machine vision data with process parameters. [Methods] A multimodal deep learning model integrating convolutional neural networks(CNN)

bidirectional long short-term memory(BiLSTM)

and an Attention mechanism is proposed. Data from a 660 MW unit

comprising numerical operation parameters and infrared thermal images

were utilized. To evaluate the model's adaptability to different temporal dynamics

datasets with four distinct time granularities(1 s

5 s

30 s

and 60 s) were constructed via data resampling. Data engineering techniques

including ANOVA and Spearman correlation analysis

screened input features to select critical variables like unit load and ambient temperature while removing redundancy. The model architecture processes these inputs: The CNN extracts spatial thermal features from infrared images

while the BiLSTM models temporal dependencies in the operating parameters. An attention mechanism emphasizes informative time steps

and the two modalities are subsequently fused for back-pressure prediction. Bayesian optimization is employed to tune hyperparameters for optimal performance. [Results] Extensive comparative experiments across 1 s

5 s

30 s

and 60 s granularities demonstrate that the proposed BiLSTM+CNN+Attention model consistently outperforms traditional machine learning(e.g.

SVR

GBM) and deep learning models(RNN

GRU

standard BiLSTM). Ablation studies on the 1 s dataset show that the introduction of the Attention mechanism reduces the Mean Absolute Percentage Error(MAPE) from 1.528 8%(standard BiLSTM) to 0.824 0%

and integrating CNN further reduces the MAPE to 0.500 4%. Specifically

the model achieves a Mean Absolute Error(MAE) of 0.055 2 kPa and a root mean square error(RMSE) of 0.076 0 kPa on the 1 s dataset. Crucially

despite the increased cumulative uncertainty at larger time scales

the model maintains a low MAPE of approximately 2.604 9% on the 60 s dataset. This indicates that the model is robust and can effectively capture both high-frequency transient fluctuations and low-frequency thermodynamic trends. [Conclusion] The fusion of infrared thermal imaging and process data effectively captures both the spatial thermodynamic state and temporal evolution of the air-cooled system. The proposed model exhibits superior generalization ability and robustness across multiple time scales. This research provides a reliable online modeling solution for the model predictive control(MPC) and multi-objective operational optimization of aircooled units

contributing to the intelligent and economic operation of thermal power plants.