Assessing the Transferability of Machine-Learning Models for EV-Charging Demand Potential: Cross-City External Validation and an Isfahan, Iran Data-Readiness Case Study

Assessing the Transferability of Machine-Learning Models for EV-Charging Demand Potential: Cross-City External Validation and an Isfahan, Iran Data-Readiness Case Study

Parisa GhahremanShahraki1

1) Computer Engineer

IOI Code : AAHF-ACZBZH
Publication : National Conference on Architecture and Art Technologies, Smart City, Soft and Creative Industries and Technological Culture(icact.ir)
Abstract :
    The rapid expansion of electric vehicles (EVs) has increased the need for public charging infrastructure that is both adequately utilised and spatially accessible. However, models developed to estimate charging demand in one city are often applied elsewhere without a sufficiently strict assessment of whether their predictive relationships transfer across urban, infrastructural, and behavioural contexts. This study develops and tests a transferable modelling workflow for identifying areas with high EV-charging demand potential. The workflow was trained using the CHARGED dataset from five cities—Amsterdam, Johannesburg, Los Angeles, Melbourne, and São Paulo—and evaluated through leave-one-city-out validation. The target was a within-city classification of the top 20% of spatial cells according to long-term daily charging-duration value per charger. The final feature set comprised transferable spatial indicators derived from points of interest (POIs), population, and built-up surface. HistGradientBoosting was selected solely using the internal cross-city evaluation, achieving a mean ROC-AUC of 0.5671, a mean PR-AUC of 0.2905, and a mean precision of 0.3056 among the highest-ranked 20% of cells, compared with a random-prevalence benchmark of 0.2134. The selected model was then evaluated once, without retraining or tuning, on the independent UrbanEV dataset for Shenzhen, China. External performance deteriorated to ROC-AUC = 0.4773, PR-AUC = 0.2100, and top-20% precision = 0.1786; the latter was below the Shenzhen high-potential prevalence of 0.2036. Feature-shift analysis indicated substantial distributional differences between the training cities and Shenzhen, particularly in charger capacity, station density, POI intensity, population, and built form. An exploratory data-readiness workflow was subsequently implemented for Isfahan, Iran, including a 500-m analytical grid, OpenStreetMap-derived road and POI layers, and Global Human Settlement Layer indicators. These outputs are presented as preparation for future local modelling and validation, not as construction recommendations. The results show that internally promising cross-city performance does not establish external transferability. Local charging-behaviour data, infrastructure attributes, travel patterns, and power-network information are needed before a model can support operational charging-station siting in a new city.

Keywords : electric vehicle charging charging demand potential transfer learning external validation spatial machine learning urban mobility isfahan