PROBABILISTIC PHYSICS-AWARE MACHINE LEARNING PREDICTIONS OF ELECTRIC TRUCK ENERGY CONSUMPTION WITH FIELD DATA

Research Question and Approach

Accurate energy consumption prediction is vital for the operational efficiency of electric truck fleets, yet traditional physics-based models often struggle with the complexity of real-world variables, while purely data-driven machine learning models may lack interpretability or fail to quantify uncertainty. This paper investigates a hybrid approach, developing physics-aware machine learning models that incorporate fundamental vehicle dynamics—such as acceleration, road inclination, and aerodynamic drag—directly into the learning process. The researchers evaluated three classes of models: Bayesian linear regression, neural networks, and gradient boosted regression trees. By training these models on high-frequency (1Hz) field data from four electric trucks, the study aimed to improve both point-estimate accuracy and the quantification of predictive uncertainty.

The study demonstrates that embedding physical principles into machine learning architectures consistently outperforms standard, non-physics-aware versions of the same models. Bayesian linear regression provided more reliable estimates than standard linear regression, while more complex models like neural networks and gradient boosted regression trees achieved even higher accuracy. A key contribution is the development of a framework to estimate predictive uncertainty (standard deviation) alongside energy consumption. The authors successfully calibrated these uncertainty estimates to account for dependencies between road segments, ensuring that the models provide trustworthy ranges for energy needs rather than just single-point predictions.

For long-haul freight and logistics, underestimating energy consumption can lead to delivery failures and range anxiety, while overestimation results in inefficient charging infrastructure use and unnecessary costs. By bridging the gap between theoretical vehicle dynamics and data-driven predictive modeling, this research provides a practical, robust method for fleet operators to make informed, risk-aware decisions. The ability to quantify uncertainty allows for more reliable route planning and battery management without requiring the massive datasets often needed for purely empirical approaches.