Publication

Abstract : The rapid increase in the use of electric vehicles (EVs) leads to a critical need to address range anxiety in the drivers by accurately predicting the energy consumption of the vehicle. This research presents an innovative approach to enhancing the prediction of EV energy consumption by integrating modern advanced deep learning models, including TabTransformer and tabular neural network (TabNet), with traditional machine learning techniques, including light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), and light gradient boosting machine (CatBoost). The study introduces bidirectional encoder representations from transformers (BERT) embeddings for text feature extraction in EV energy consumption prediction, showing their usefulness in enhancing model performance. Utilizing the Volkswagen e-Golf dataset with 19 attributes, TabTransformer emerged as the most effective model, achieving an R2 of 0.9812, a significant improvement over traditional models such as LightGBM. TabNet offered a competitive R2 of 0.9758 while maintaining reduced computational complexity. In addition to that, BERT-enhanced XGBoost exhibited a solid performance, particularly in terms of R2 and MAE. Among given traditional approaches, Bayesian-optimized LightGBM showed outstanding efficiency, with a training time of only 0.476 s. These results highlight how well TabTransformer can make predictions and its potential to optimize energy consumption predictions and efficiency of Bayesian optimization for LightGBM. The findings contribute valuable insights into model selection and optimization, fostering advances in EV range estimation and promoting the broader adoption of sustainable transportation solutions.