[Purposes] Vehicle load is one of the critical factors causing deflection variation of bridges during the service stage, characterized by strong spatial-temporal randomness and great challenges in prediction. To improve the prediction accuracy of deflection induced by vehicle loads, a CNN-LSTM-Attention deep learning model integrated with feature engineering was established, and the effects of feature engineering and model structure on prediction performance were systematically analyzed. [Methods] Firstly, based on data from the bridge deck monitoring and weighing systems, a lane-level vehicle load time series was constructed. Feature engineering methods including lag features, difference features, rolling standard deviation, and moving average were introduced to enhance the representation capability of the time series. The CNN-LSTM-Attention prediction model was thus established and compared with the long short-term memory (LSTM) model and the model without feature engineering. Finally, the model performance was quantitatively evaluated using indicators such as mean absolute error rMAE, mean square error rMSE, coefficient of determination R2, and explained variance score SEVS. [Findings] The prediction accuracy of the model is significantly improved after introducing feature engineering. Compared with the model without feature engineering, the rMAE is reduced by 87.2%, and R2 is increased by 27.7%. Compared with the traditional LSTM model, the rMAE is decreased by 87.3%, and R2 is improved by 34.4%. [Conclusions] The CNN-LSTM-Attention model integrated with feature engineering can effectively capture the temporal characteristics and long-term dependencies of vehicle load sequences, enabling high-precision prediction of bridge deflection.