Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Transfer Learning for Electricity Demand Forecasting in Nordic Cities : A Case Study of Copenhagen and Reykjavik

Javidsharifi, Mahshid ; Arabani, Hamoun Pourroshanfekr LU ; Vasquez, Juan C. and Guerrero, Josep M. (2025) 32nd International Conference on Systems, Signals and Image Processing, IWSSIP 2025 In International Conference on Systems, Signals, and Image Processing
Abstract

In Nordic countries, demand response (DR) programs are critical for enhancing energy flexibility and reducing peak electricity consumption in buildings and industry. Accurate short-term demand forecasting is essential to support these programs, especially in the context of cities with varying levels of data availability. This paper compares three different forecasting approaches for hourly electricity demand prediction in the Nordic region: a baseline Long Short-Term Memory (LSTM) model, a transfer learning (TL)-based LSTM model, and a tree-based Extreme Gradient Boosting (XGBoost) model. We use Copenhagen, Denmark as a data-rich source city to pretrain an LSTM and fine-tune it on limited data from Reykjavik, Iceland simulating a... (More)

In Nordic countries, demand response (DR) programs are critical for enhancing energy flexibility and reducing peak electricity consumption in buildings and industry. Accurate short-term demand forecasting is essential to support these programs, especially in the context of cities with varying levels of data availability. This paper compares three different forecasting approaches for hourly electricity demand prediction in the Nordic region: a baseline Long Short-Term Memory (LSTM) model, a transfer learning (TL)-based LSTM model, and a tree-based Extreme Gradient Boosting (XGBoost) model. We use Copenhagen, Denmark as a data-rich source city to pretrain an LSTM and fine-tune it on limited data from Reykjavik, Iceland simulating a low-data target environment. The results show that while XGBoost performs best with well-engineered features, TL offers a robust and scalable alternative in data-scarce scenarios, especially where traditional features may be limited or incomplete.

(Less)
Please use this url to cite or link to this publication:
author
; ; and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
electricity demand forecasting, extreme gradient boosting, long short-term memory, transfer learning
host publication
32nd International Conference on Systems, Signals and Image Processing, IWSSIP 2025 - Proceedings
series title
International Conference on Systems, Signals, and Image Processing
publisher
IEEE Computer Society
conference name
32nd International Conference on Systems, Signals and Image Processing, IWSSIP 2025
conference location
Skopje, Macedonia, The Former Yugoslav Republic of
conference dates
2025-06-24 - 2025-06-26
external identifiers
  • scopus:105017433059
ISSN
2157-8702
2157-8672
ISBN
9798350392890
DOI
10.1109/IWSSIP66997.2025.11151865
language
English
LU publication?
yes
additional info
Publisher Copyright: © 2025 IEEE.
id
4bba1d76-8d7b-45bb-b1a6-adbb7875f328
date added to LUP
2025-12-08 10:50:40
date last changed
2025-12-08 10:51:45
@inproceedings{4bba1d76-8d7b-45bb-b1a6-adbb7875f328,
  abstract     = {{<p>In Nordic countries, demand response (DR) programs are critical for enhancing energy flexibility and reducing peak electricity consumption in buildings and industry. Accurate short-term demand forecasting is essential to support these programs, especially in the context of cities with varying levels of data availability. This paper compares three different forecasting approaches for hourly electricity demand prediction in the Nordic region: a baseline Long Short-Term Memory (LSTM) model, a transfer learning (TL)-based LSTM model, and a tree-based Extreme Gradient Boosting (XGBoost) model. We use Copenhagen, Denmark as a data-rich source city to pretrain an LSTM and fine-tune it on limited data from Reykjavik, Iceland simulating a low-data target environment. The results show that while XGBoost performs best with well-engineered features, TL offers a robust and scalable alternative in data-scarce scenarios, especially where traditional features may be limited or incomplete.</p>}},
  author       = {{Javidsharifi, Mahshid and Arabani, Hamoun Pourroshanfekr and Vasquez, Juan C. and Guerrero, Josep M.}},
  booktitle    = {{32nd International Conference on Systems, Signals and Image Processing, IWSSIP 2025 - Proceedings}},
  isbn         = {{9798350392890}},
  issn         = {{2157-8702}},
  keywords     = {{electricity demand forecasting; extreme gradient boosting; long short-term memory; transfer learning}},
  language     = {{eng}},
  publisher    = {{IEEE Computer Society}},
  series       = {{International Conference on Systems, Signals, and Image Processing}},
  title        = {{Transfer Learning for Electricity Demand Forecasting in Nordic Cities : A Case Study of Copenhagen and Reykjavik}},
  url          = {{http://dx.doi.org/10.1109/IWSSIP66997.2025.11151865}},
  doi          = {{10.1109/IWSSIP66997.2025.11151865}},
  year         = {{2025}},
}