Bridge deflection was regarded as a critical indicator that reflected structural deformation and bearing performance, and its accurate prediction was considered essential for structural safety assessment and maintenance decision-making. Traditional inspection methods, which rely heavily on manual inspection, were insufficient for real-time health assessment. The integration of modern sensing technologies and intelligent algorithms has become a prominent research focus in structural health monitoring. A hybrid deflection prediction model was proposed based on real-time data collected from bridge structural health monitoring sensors, including deflection, temperature, and other key parameters. Outliers were identified and removed using the isolation forest algorithm. The time convolutional network (TCN) was employed for its time-series modeling capability, the long short-term memory (LSTM) network was used for capturing long-term dependencies, and the Attention mechanism was integrated for adaptive feature selection. A TCN-LSTM-Attention hybrid model was proposed to predict bridge deflection. To optimize the model’s performance, the whale optimization algorithm was used for hyperparameter tuning. The results demonstrate that the TCN-LSTM-Attention model performs excellently across various evaluation metrics, with an R2 of 0.97 and an RMSE of 0.206 6. Compared to the TCN and LSTM models, it improves R2 by 5.24% and 3.78%, and reduces RMSE by 5.47% and 4.21%, respectively. Notably, the model exhibits strong robustness and accuracy, especially in long-term predictions scenarios. This work provides an effective and intelligent approach to bridge deflection prediction and offers valuable support for condition-based maintenance planning.