Temporal feature augmentation for transient fuel consumption prediction of heavy-duty trucks: Toward sustainable road freight and cleaner operations

Transient operations substantially increase fuel consumption in heavy-duty trucks, while conventional static models fail to capture these short-term dynamics. Existing data-driven models either treat the engine-vehicle system as memoryless or rely on black-box sequence models that sacrifice interpretability. This study introduces a physics-informed, model-agnostic Temporal Feature Augmentation (TFA) framework that encodes unobserved internal states, such as turbocharger spool state, manifold filling, and thermal inertia, into compact and interpretable derivative-based proxy features computed causally from standard onboard measurements without additional hardware. TFA enriches the feature space by incorporating instantaneous temporal derivatives of measurable engine and air-path states, including the excess air coefficient λ, intake pressure, and pre-turbine pressure and temperature, which serve as compact proxies for unobserved transient dynamics. Validated on standardized CHTC-TT (1800 s) and extensive real-world cycles (5350 s) using five algorithms (RF, XGBoost, SVM, RNN, LSTM), TFA consistently enhances prediction accuracy and cross-cycle robustness: the best model achieves R 2 = 0.996 and reduces RMSE by 23.4% relative to static-feature baselines. Feature-importance analysis confirms that temporal-derivative features effectively proxy load transients. The proposed framework supports eco-driving assistance, fleet fuel monitoring, and data-driven quantification of fuel-saving opportunities, contributing to low-carbon, cleaner production operations in commercial road freight.