Beyond Day-Ahead: Multi-Day Electricity Price Forecasting for Smarter Electric Vehicle Charging
(2026) In Master's Theses in Mathematical Sciences FMSM01 20261Mathematical Statistics
- Abstract
- Smart charging of electric vehicles can reduce cost by scheduling against time-varying electricity spot prices, but commercial systems generally rely on day-ahead prices and capture only within-day variation. This thesis investigates whether forecasting spot prices beyond the day-ahead horizon adds value for vehicle-integrated smart charging, and how that value varies across charging contexts. A three-stage pipeline is built for the Swedish SE3 bidding zone. Probabilistic hourly forecasts of wind production, solar production and electricity consumption for up to one week ahead are produced from weather forecasts together with standard calendar and historical features. These upstream forecasts, along with day-ahead price history and... (More)
- Smart charging of electric vehicles can reduce cost by scheduling against time-varying electricity spot prices, but commercial systems generally rely on day-ahead prices and capture only within-day variation. This thesis investigates whether forecasting spot prices beyond the day-ahead horizon adds value for vehicle-integrated smart charging, and how that value varies across charging contexts. A three-stage pipeline is built for the Swedish SE3 bidding zone. Probabilistic hourly forecasts of wind production, solar production and electricity consumption for up to one week ahead are produced from weather forecasts together with standard calendar and historical features. These upstream forecasts, along with day-ahead price history and supply-side market data, then feed an hourly spot price model running over the same horizon. The forecasts are evaluated in an EV charging simulation under two specific frameworks: continuous deferral and a weekly deadline. This evaluation is conducted across two use cases representative of vehicle-manufacturer- controlled charging, specifically home overnight and workplace daytime.
Forecast accuracy is dominated by tree-based model families (Random Forest and XGBoost), which outperform neural sequence families (Long Short-Term Memory and Gated Recurrent Unit) by 15–18 percentage points in normalised mean absolute error at short leads, narrowing to roughly 3 points beyond five days. Accuracy declines markedly across the horizon, driven mostly by the loss of weather forecast skill at multi-day leads. Specifically, the coefficient of determination drops from 0.83 day-ahead to between 0.27 and 0.33 at one week. Despite this decline, the broader price pattern remains clear over the first two to four days, with electricity consumption being the most reliably forecast upstream input. The marginal annual saving from acting on multi-day forecasts, above and beyond within-day optimisation, falls in the range 44–56 EUR per vehicle per year across all four framework × use-case combinations at a reference 50 km/day driver on a 7.4 kW home wallbox, scaling from roughly 20–28 EUR/year at 25 km/day to 72–78 EUR/year at 100 km/day. The total annual savings against a do-nothing baseline are larger (103–168 EUR/year), but the variation between configurations comes almost entirely from within-day scheduling, which requires no forecast at all. The residual gap to a perfect- information benchmark stays within 3.6 % of the do-nothing cost, indicating that further gains would have to come from a richer decision problem rather than better forecasts. The value is real but dependent on a customer relationship to charging that does not require a fully charged battery every morning. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9244798
- author
- Kjellergren, Karl LU and Lindqvist Sabel, Johan
- supervisor
- organization
- course
- FMSM01 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Electricity price forecasting, Day-ahead market, Multi-day forecasting, Probabilistic forecasting, SE3, Random Forest, XGBoost, Long Short-Term Memory, Gated Recurrent Unit, Electric Vehicle Charging, Smart Charging, Charging Optimization, Wind Power Forecasting, Solar Power Forecasting, Electricity Consumption Forecasting, Numerical Weather Prediction, Nord Pool, Spot Price
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMS-3572-2026
- ISSN
- 1404-6342
- other publication id
- 2026:E99
- language
- English
- id
- 9244798
- date added to LUP
- 2026-06-29 13:06:54
- date last changed
- 2026-06-29 13:06:54
@misc{9244798,
abstract = {{Smart charging of electric vehicles can reduce cost by scheduling against time-varying electricity spot prices, but commercial systems generally rely on day-ahead prices and capture only within-day variation. This thesis investigates whether forecasting spot prices beyond the day-ahead horizon adds value for vehicle-integrated smart charging, and how that value varies across charging contexts. A three-stage pipeline is built for the Swedish SE3 bidding zone. Probabilistic hourly forecasts of wind production, solar production and electricity consumption for up to one week ahead are produced from weather forecasts together with standard calendar and historical features. These upstream forecasts, along with day-ahead price history and supply-side market data, then feed an hourly spot price model running over the same horizon. The forecasts are evaluated in an EV charging simulation under two specific frameworks: continuous deferral and a weekly deadline. This evaluation is conducted across two use cases representative of vehicle-manufacturer- controlled charging, specifically home overnight and workplace daytime.
Forecast accuracy is dominated by tree-based model families (Random Forest and XGBoost), which outperform neural sequence families (Long Short-Term Memory and Gated Recurrent Unit) by 15–18 percentage points in normalised mean absolute error at short leads, narrowing to roughly 3 points beyond five days. Accuracy declines markedly across the horizon, driven mostly by the loss of weather forecast skill at multi-day leads. Specifically, the coefficient of determination drops from 0.83 day-ahead to between 0.27 and 0.33 at one week. Despite this decline, the broader price pattern remains clear over the first two to four days, with electricity consumption being the most reliably forecast upstream input. The marginal annual saving from acting on multi-day forecasts, above and beyond within-day optimisation, falls in the range 44–56 EUR per vehicle per year across all four framework × use-case combinations at a reference 50 km/day driver on a 7.4 kW home wallbox, scaling from roughly 20–28 EUR/year at 25 km/day to 72–78 EUR/year at 100 km/day. The total annual savings against a do-nothing baseline are larger (103–168 EUR/year), but the variation between configurations comes almost entirely from within-day scheduling, which requires no forecast at all. The residual gap to a perfect- information benchmark stays within 3.6 % of the do-nothing cost, indicating that further gains would have to come from a richer decision problem rather than better forecasts. The value is real but dependent on a customer relationship to charging that does not require a fully charged battery every morning.}},
author = {{Kjellergren, Karl and Lindqvist Sabel, Johan}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Beyond Day-Ahead: Multi-Day Electricity Price Forecasting for Smarter Electric Vehicle Charging}},
year = {{2026}},
}