Two-stage model for heavy-duty truck fuel consumption prediction based on incomplete data

To address uncertainty and accuracy loss in heavy-duty truck fuel consumption prediction under missing-data conditions, this study proposed an end-to-end two-stage deep learning model for incomplete data scenarios. First, a continual adversarial imputation method combined Transformer-based feature extraction with a generative adversarial network, so that the generated data distribution closely matched the real distribution. Next, a fuel consumption prediction model integrated regional attention and pure graph attention mechanisms. Joint local and global modeling improved representation of complex dynamic features, enabling accurate fuel consumption imputation and prediction. Experiments on a real-world heavy-duty truck fuel consumption dataset with missing rates from 0 to 40% validated the proposed model. Research results show that, across different missing rates, the imputation mean squared error (MSE) ranges from 0.873 to 0.906, and overall performance surpasses seven baseline imputation methods. Impact analysis reveals that local dynamic variables, such as vehicle speed and acceleration, significantly affect fuel consumption imputation and prediction, whereas global factors, such as altitude and mileage, have a weaker effect. This indicates that local features contribute more to performance improvement. Comparative experiments demonstrate that the proposed two-stage imputation-prediction framework achieves the highest predictive accuracy, with an MSE of 0.602 0. Ablation experiments further confirm the effectiveness of each functional module. This study provides technical support for heavy-duty truck fuel management, carbon emission monitoring, and intelligent fleet scheduling, especially under complex operating conditions with incomplete data.