It is very crucial to be able to accurately predict the energy consumption of electric heavy-duty vehicles. This is especially true as that market is expanding rapidly. This thesis aims to create vector embeddings through the use of multi-task learning to help improve the prediction of energy consumption of auxiliary systems in electric heavy-duty vehicles, thus allowing for better route planning and preservation of energy. Previous studies have shown that multi-task learning has improved the predictive performance of various downstream tasks, such as time series forecasting. Therefore, we propose a Batched Multi-task Learning method (BaMLE). To perform the time series forecasting, we use a forecasting transformer model, with its state-of-the-art multi-head attention mechanism, that other forecasting models lack.

We evaluate our proposed method by comparing them to other stateof the art task agnostic embedding methods through downstream tasks. We also compare the choice of transformer model to one othertransformer model, as well as some other models that do not use the transformer architecture. Our research illustrates the potential benefits that multi-task learning has, and it contributes to future work for improving multi-task performance. To summarise, this thesis proposes a novel approach of using state-of-the-art methods and models to improve the downstream task performance of auxiliary systems energy consumption in electric heavy-duty vehicles.