CatBoost-based fuel consumption modeling and explainable analysis for heavy-duty diesel trucks: Impact of engine, driving behavior, and vehicle weight

The limited understanding of complex causal relationships and interactions among influencing factors and fuel consumption hinders the development of high-precision fuel consumption models and restricts their practical application. This study leverages high-resolution operational data from 7311 in-use heavy-duty diesel trucks to model the impacts of engine parameters, driving behavior, and vehicle weight on fuel consumption using machine learning. Specifically, six feature availability scenarios were designed, and CatBoost-based prediction models were developed to evaluate model robustness under missing feature conditions. SHapley Additive exPlanations (SHAP) were further applied for post-hoc analysis to reveal the global, individual, and interaction-level effects of key features on fuel consumption. Results show that models incorporating engine features yielded the highest accuracy, achieving a minimum MAPE of 5.62 %, and exhibited strong robustness to the exclusion of other feature categories. In contrast, without engine features, the best MAPE rose to 11.60 %, and vehicle weight became the most critical predictor; omitting it led to a further 7.96 % rise in prediction error. Vehicle load rate has a clear positive effect on fuel consumption, and the effect of driving behavior varies with loading conditions. Under high-speed conditions (>80 km/h), maintaining engine speed within the economical range effectively reduces fuel consumption. Additionally, high engine load significantly increases the sensitivity of fuel consumption to friction torque.