The shift towards battery electric trucks (BETs) represents a transformativestep for the transportation industry, motivated by globalsustainability goals and the logistics sector’s emphasis on long-termprofitability. While physics-based models have successfully predictedpropulsion energy consumption in heavy-duty vehicles (HDVs), auxiliarysystems such as compressors, heaters, and converters presentunique challenges due to their dynamic nature and dependence onenvironmental factors.

Existing methods, which rely on fleet-widestatistical averages, often fail to capture the variations in operationalconditions, leading to less effective energy management and inaccurateforecasting.This thesis proposes a data-driven framework to address these challengesand improve predictions for auxiliary energy consumption inHDVs. At the heart of this work is a hybrid model that combinesthe ability of Long Short-Term Memory (LSTM) networks to capturetemporal dependencies with the stability of a mean predictor. Thishybrid approach mitigates error accumulation over longer forecastinghorizons, a common issue in time-series forecasting. Additionally,an attention mechanism is integrated into the model, enabling it tofocus on the most relevant parts of the input data, further enhancingprediction accuracy.A notable contribution of this thesis is the introduction of a fleetbasedregression approach. Recognizing that auxiliary systems behavedifferently across vehicle types and fleets, this approach developsspecialized models tailored to each system and fleet. For instance,systems like heaters and compressors, which are highly sensitive toexternal factors such as temperature and vehicle load, are modeledindividually to capture their unique behaviors.

A key aspect of thiswork is the selection of contextual features informed by expert domainknowledge. Recognizing the diverse behaviors of auxiliary systemsfeatures such as temperature, vehicle load, and operational conditionswere carefully chosen to ensure the model captures systemspecificpatterns effectively. This fleet-specific strategy ensures predictionsare precise and context-aware, addressing the diversity inHDV operations.To further enhance the model’s applicability, domain adaptationtechniques are employed to overcome challenges posed by domainshifts between different fleets or environments. By learning domaininvariantfeatures, the model can generalize effectively across diversedatasets, reducing the reliance on large amounts of labeled data fromthe target domain. This approach is particularly valuable for handling variations across different HDV configurations and operating conditions.

The thesis contributes in three key areas: (1) developing a hybridLSTM-mean predictor model with attention for improved long-horizonforecasting, (2) creating a fleet-based regression approach tailored toindividual auxiliary systems, and (3) implementing domain adaptationtechniques to improve model generalization across differentfleets and environments. Experimental results highlight the superiorityof this framework over traditional methods, providing accurate,context-aware predictions for auxiliary energy consumption. This workpaves the way for optimizing energy management in HDVs, enhancingrange estimation, and supporting sustainable electric trucking solutions.