Energy consumption prediction of electric vehicles for data-scarce scenarios using pre-trained model

Abstract

Accurate energy consumption prediction is essential for mitigating range anxiety in electric vehicles (EVs). However, data-driven approaches require substantial user-specific data, limiting their effectiveness during the early stages of vehicle ownership or for newly registered EVs with limited historical records. This study introduces the Energy Consumption Transformer (ECT), a large-scale pre-trained model designed for data-scarce scenarios. ECT leverages transferable prior knowledge learned from over 310 thousand real-world trips. This pre-trained knowledge is then integrated with proprietary knowledge from limited EV trips, enabling accurate prediction even with minimal data. The pre-training is conducted using multi-city, multi-model datasets, ensuring the generalization of the learned knowledge. Results show that ECT outperforms existing methods in few-shot scenarios and bridges the multi-month prediction service gap during data collection. Furthermore, ECT provides efficient inference, with a response time ranging from 0.14 to 0.28 ms/trip. This study facilitates immediately available prediction without requiring extensive user-specific data collection.