Gerholm Lindstrom EV Fast-Charging Peak-Load Forecasting (2026)

“Risk-aware peak-load profile forecasting for new electric vehicle fast-charging stations without historical data” — Markus Emanuel Gerholm, Erik Lindström (Lund University, Sweden). Energy and AI 25 (2026) 100866. Received 18 June 2026, accepted 5 August 2026, open access (CC BY 4.0). Data collection facilitated by Öresundskraft AB.

Summary

A machine-learning framework for predicting 24-hour peak-load profiles of a planned DC fast-charging station — one with no historical load observations of its own — from static pre-construction metadata alone (capacity, number of charging points, geographic location, distance to nearest highway, count of nearby stations). This is a genuine cold-start prediction problem: grid planners need to know whether existing capacity can accommodate a new station’s expected peak load, or whether reinforcement is needed first, before the station has generated a single hour of data.

Method: splits the problem into two parts — (1) predicting the shape of the daily load curve via K-means clustering of existing stations’ load patterns plus random-forest classification to assign a new station to a cluster, and (2) predicting the magnitude of the peak via quantile random forest regression. Combining shape and magnitude gives a full risk-adjustable peak-load profile: planners can choose a low-risk (conservative, high-quantile) or a tighter (median) prediction depending on how much overload risk they’re willing to accept versus how much spare capacity they’re willing to allocate — explicitly framed as a newsvendor-style capacity-planning trade-off.

Data: hourly average load from 33 DC fast-charging stations (32 after removing one outlier station) in the distribution grid of “an energy company in northwestern Skåne” — Öresundskraft’s service area. Observation period from 2025-01-01 (or each station’s install date if later) through 2026-03-30.

Results: the conservative (99th-percentile) quantile prediction covered the majority of observed hourly peaks for 93.4% of stations, while staying below theoretical maximum capacity — i.e., a usably tight conservative bound, not just “assume worst case.” The median prediction achieved the lowest average error (mean absolute error ≈ 108 kW). Smaller stations (fewer charging points) are harder to predict — they more easily approach full simultaneous-charging capacity, producing higher and more variable normalized peak loads, while larger stations’ peaks are damped by the low probability of many vehicles charging simultaneously. Two stations broke the pattern by being unusually close to a highway, suggesting traffic exposure — not just charging-point count — matters for prediction difficulty.

Relevance to wiki topics