Large Language Model-Derived Spatial Embedding for Revealing Electric Vehicle Energy Consumption: An Urban–Rural Case
TL;DR
Conditional, location-level associations are described for the sampled fleet and season of private EVs operating in Shanghai from October to December 2020, integrating about 29 million high-resolution GPS records with multi-source built-environment data.
Abstract
Understanding spatial variation in electric vehicle energy consumption (EC) can support more targeted urban transport analysis, but observational data do not identify the causal effects of urban design. This study analyzes approximately 1700 private EVs operating in Shanghai from October to December 2020, integrating about 29 million high-resolution GPS records with multi-source built-environment data. Segment-level EC was aggregated to 500 m grids and communities. Community functional zones were derived by encoding POI sequences with Qwen2.5-1.5B-Instruct and clustering the resulting embeddings, while Random Forest models and SHAP were used to characterize model-attributed associations with spatial EC variation. At the grid level, mean speed and speed variability received approximately 84% of the native Random Forest importance, whereas the 5D-inspired indicators received the remaining 16%, led by destination accessibility, road density, and bus-stop density. At the community level, the Urban core zone had approximately 9% higher fleet-weighted EC than the Rural zone, while within-zone importance rankings differed between dense and sparse zones. These results describe conditional, location-level associations for the sampled fleet and season.