Multi-source data-driven short-term remaining driving range prediction for electric vehicles: A hybrid CNN-transformer framework

The prediction of the remaining driving range (RDR) of electric vehicles is crucial for optimizing vehicle energy management. However, traditional remaining driving range predictions tend to focus on long-term estimates, which lead to issues such as accumulated long-term errors and insufficient sensitivity to short-term driving conditions. This makes them inadequate for the condition of short driving distance in urban road network. To address this issue, a CNN-Transformer (CT) hybrid model is designed to conduct short-term remaining driving range (STRDR) prediction. The CT model integrates the local dynamic feature extraction capabilities of the CNN model with the Transformer's ability to model long-term dependencies, significantly improving prediction performance. This study constructs different features from various aspect through the fusion of multi-source data, including vehicle driving data, battery status data, and weather data. At the same time, driving behavior characteristics are taken into account. Random forest is used to identify key features related to STRDR predictions for different drivers. Verification using real vehicle data from six vehicles across three cities in Hunan Province, China, demonstrated that the CT model achieved an average R2 of 0.95 and a SMAPE of 18.01 %, which was 6.81 %–9.69 % lower than candidate models such as Transformer and LightGBM. The CT model also exhibited strong robustness across different driving styles. Error analysis confirms that predicted and true values are equally distributed across low and high STRDR intervals, and the slope error approaches zero, confirming the model's stability. This study is conducive to the promotion of EVs.