Discovering EV Charging Site Archetypes Through Few Shot Forecasting: The First U.S.-Wide Study

TL;DR

This work tackles the challenge of forecasting EV charging demand across diverse charging sites by addressing data scarcity and limited transferability. It introduces a framework that combines k-means clustering with few-shot forecasting, training cluster-specific expert models (based on the Temporal Fusion Transformer) on a large nationwide dataset to uncover site archetypes. The approach identifies an optimal number of archetypes (around 12) and demonstrates that archetype-specific models outperform a global baseline in forecasting demand at unseen sites, with semantic labeling enabling geospatial interpretability and operational insights. The findings have practical implications for siting, pricing, and grid resilience, supporting scalable, grid-aware management of EV charging infrastructure. The forecasting objective is defined as $\hat{\mathbf y}_{t+1:t+H \mid t} = f_\theta(\mathbf y_{t-L+1:t})$ with context length $L=28$ and horizon $H=7$ across $k=1..20$ clusters, selecting $k$ by predictive performance on unseen sites.

Abstract

The decarbonization of transportation relies on the widespread adoption of electric vehicles (EVs), which requires an accurate understanding of charging behavior to ensure cost-effective, grid-resilient infrastructure. Existing work is constrained by small-scale datasets, simple proximity-based modeling of temporal dependencies, and weak generalization to sites with limited operational history. To overcome these limitations, this work proposes a framework that integrates clustering with few-shot forecasting to uncover site archetypes using a novel large-scale dataset of charging demand. The results demonstrate that archetype-specific expert models outperform global baselines in forecasting demand at unseen sites. By establishing forecast performance as a basis for infrastructure segmentation, we generate actionable insights that enable operators to lower costs, optimize energy and pricing strategies, and support grid resilience critical to climate goals.

Figures (6)

• Figure 1: The framework leverages historical demand data to generate features for clustering, with expert models trained on resulting site clusters to forecast demand. This is repeated for different $k$.
• Figure 2: (a) A8: Recurring weekday patterns across downtown cores; (b) A1: Stable daytime demand in metro areas adjacent to retail chains; (c) A3: Utilization concentrated along major highways; (d) A11: Elevated weekend demand near leisure hubs such as Lake Tahoe and Hampton Bays. See Appendix \ref{['subsec:cluster_spread']} for complete geospatial distribution and demand profiles.
• Figure 3: Global $(k=1)$ vs expert $(k>1)$ performance.
• Figure 4: Geospatial distribution of charging site archetypes across the United States. Sites are aggregated into hexagonal cells, with each hexagon colored by the dominant utilization cluster.
• Figure 5: Representative demand profiles for the identified archetypes. Each profile shows the barycenter of clustered sites, illustrating characteristic temporal patterns ranging from stable weekday cycles to highly volatile weekend or tourism-driven surges.