Gulaydin, Oguzhan  ORCID: https://orcid.org/0000-0002-1795-7939 and Mourshed, Monjur ORCID: https://orcid.org/0000-0001-8347-1366
2025.
Machine learning for subnational residential electricity demand forecasting to 2050 under shared socioeconomic pathways: Comparing tree-based, neural and kernel methods.
Energy
336
, 138195.
10.1016/j.energy.2025.138195
|
Preview |
PDF
- Published Version
Available under License Creative Commons Attribution. Download (2MB) | Preview |
Abstract
A scenario-based machine learning framework is presented for long-term, subnational electricity demand forecasting, integrating Shared Socioeconomic Pathways (SSPs) with spatially downscaled demographic, economic, and climatic variables. Using Turkey as a case study, the framework projects residential electricity demand to 2050 across all 81 provinces. The subnational approach enables the use of data-intensive machine learning algorithms by expanding the training dataset through the multiplicative effect of combining spatial and temporal dimensions. Six machine learning models: tree-based (Random Forest, XGBoost), neural networks (Feed-forward Neural Network, Long Short-Term Memory), and kernel-based methods (Support Vector Regression, Gaussian Process Regression), are systematically compared against a traditional linear regression benchmark. Random Forest achieves the highest accuracy (�2 = 0.9359, MAE= 0.04 TWh), outperforming neural and kernel-based models and substantially improving on the linear baseline. Socioeconomic variables, especially family households, population, and GDP, have a greater influence on electricity demand than climatic indicators such as heating and cooling degree days. Turkey’s residential electricity demand is projected to increase by 78% from 65.5 TWh in 2023 to 116.7±2.9 TWh by 2050, with substantial variation across provinces. The spatial variation in demand forecasts highlights the value of subnational modelling for energy planning and the limitations of national-level projections. The use of SSPs enables a consistent and policy-relevant exploration of plausible long-term demand trajectories. By combining subnational resolution, scenario-based inputs, and a structured comparison of algorithm families, the study offers a transferable framework for electricity demand forecasting in regionally diverse or data-scarce contexts, supporting infrastructure planning and decarbonisation strategies.
| Item Type: | Article |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Engineering |
| Subjects: | T Technology > TA Engineering (General). Civil engineering (General) |
| Uncontrolled Keywords: | Long-term energy projections Machine learning Shared Socioeconomic Pathways Sub-national energy demand Residential electricity forecasting Random Forest algorithm Turkey (Türkiye) energy planning |
| Publisher: | Elsevier |
| ISSN: | 1873-6785 |
| Date of First Compliant Deposit: | 18 September 2025 |
| Date of Acceptance: | 25 August 2025 |
| Last Modified: | 22 Sep 2025 11:01 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/181154 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions