Digital Twin-Based Remaining Driving Range Prediction for Connected Electric Vehicles

• Published in: SAE International journal of electrified vehicles (Print) 2024-01, Vol.13 (1), p.23-36, Article 14-13-01-0004
• Main Authors: Zhuo, Shilong, Li, Heng, Bin Kaleem, Muaaz, Peng, Hui, Wu, Yue
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Batteries; Datasets; Digital twin; Digital twins; Driving range prediction; Electric charge; Electric potential; Electric vehicle charging; Electric vehicles; Energy consumption; Feature extraction; Prediction models; Predictions; Random forest; Road conditions; Sliding; Sliding window; Voltage
• ISSN: 2691-3747; 2691-3755
• DOI: 10.4271/14-13-01-0004

Electric vehicles (EVs) suffer from long charging time and inconvenient charging due to limited charging stations, which are the main causes of drivers’ range anxiety. Real-time and accurate driving range prediction can help drivers plan journeys, alleviate range anxiety, and promote EV development. However, predicting the EV driving range is challenging due to different weather, road conditions, driver habits, and limited available data. To address this issue, this article proposes a novel digital twin-based driving range prediction method. First, a one-year real-world EV dataset in Beijing is utilized. Detailed feature selection is conducted for the dataset, and six key features are extracted: battery SOC, consumed battery SOC, battery total voltage, battery maximum cell voltage, battery minimum cell voltage, and mileage already driven. Then, a random forest method is used to train the EV driving range prediction model using the features described earlier. Four prediction models with different adopted features are trained, respectively. Finally, the sliding window algorithm is proposed for the input of random forest to investigate its impact on prediction accuracy in the four prediction models, and different window sizes are evaluated. Results show that the sliding window algorithm can significantly improve the prediction model with only SOC as input, while it can deteriorate other models with more features. The most accurate model taking all six features as inputs provides 89.8% data that has an accuracy of over 80%, while data proportion of the prediction model without past energy consumption is only 31.8%.