Accurate fuel consumption prediction is recognized as crucial support for promoting energy conservation, emission reduction, and the adoption of eco-driving. To achieve precise fuel consumption prediction for urban expressway vehicles across different time periods. Time-series data including fuel consumption, speed, and operating status were collected through field experiments conducted on Hetan Expressway in Urumqi. A multi-dimensional analysis of fuel consumption influencing factors was performed using mutual information(MI), Pearson correlation coefficient, and Spearman correlation coefficient methods, followed by feature reconstruction. A hybrid BO-Transformer-LSTM model was developed by integrating Bayesian optimization algorithms, Transformer architecture, and long short-term memory(LSTM) neural networks. Comparative assessments were performed against six benchmark models. Experimental results demonstrate that the BO-Transformer-LSTM model has a coefficient of determination of 98.3% and 98.36% for fuel consumption prediction during peak and peak hours on weekdays, respectively, and the mean absolute error(MAE) is 0.030 9 g/s and 0.042 9 g/s. In non-working day scenarios, prediction accuracy remains sTable at 97.43% for off-peak hours and 97.89% for peak hours, indicating superior generalization capabilities. This confirms that the proposed model effectively enhances the robustness of fuel consumption prediction across various urban expressway periods. This approach provides technical support for accurate vehicle energy consumption prediction and facilitates the construction of time-specific emission inventories.