• English
  • Deutsch
  • Log In
    Password Login
    Research Outputs
    Fundings & Projects
    Researchers
    Institutes
    Statistics
Repository logo
Fraunhofer-Gesellschaft
  1. Home
  2. Fraunhofer-Gesellschaft
  3. Artikel
  4. Revealing the Effects of Data Heterogeneity in Federated Learning Regression Models for Short-Term Solar Power Forecasting
 
  • Details
  • Full
Options
2024
Journal Article
Title

Revealing the Effects of Data Heterogeneity in Federated Learning Regression Models for Short-Term Solar Power Forecasting

Abstract
Accurate short-term power forecasting is crucial for the successful commercialization of solar energy, helping to prevent financial losses in energy markets. Federated learning (FL) offers a promising approach for power forecasting with small databases, enabling marketers to collaboratively develop models without exchanging sensitive data. An initial simulation of this FL use case reveals a clear discrepancy between the solar power forecasting performance of the FL model and central learning (CL) models trained on each marketer’s data. While some benefit from participating in FL, others achieve far better results with individually trained forecasting models. This paper demonstrates that the discrepancy is due to heterogeneity between the data of each marketer. As this is a poorly researched area in FL regression, a taxonomy of heterogeneity characteristics (skews) is first proposed. Next the influence of the skews are analyzed conducting a simulation study. The study has shown that differences in autocorrelation and number of data points among time series from FL clients have the greatest impact on the performance of the FL regression model. A first heterogeneity quantification approach based on ARMA-GARCH models is proposed to address this problem in the development of federated solar power forecasting systems and other federated regression problems. In addition, the paper highlights the need for further investigation of heterogeneity in federated regression problems.
Author(s)
Nachtigall, Robin
Haller, Marc Leon
Fraunhofer-Institut für Optronik, Systemtechnik und Bildauswertung IOSB  
Wagner, Andreas
Fraunhofer-Institut für Techno- und Wirtschaftsmathematik ITWM  
Walter, Viktor
Journal
IEEE access  
Open Access
File(s)
Download (8.13 MB)
Rights
CC BY-NC-ND 4.0: Creative Commons Attribution-NonCommercial-NoDerivatives
DOI
10.1109/access.2024.3479737
10.24406/h-477949
Language
English
Fraunhofer-Institut für Optronik, Systemtechnik und Bildauswertung IOSB  
Fraunhofer-Institut für Techno- und Wirtschaftsmathematik ITWM  
Keyword(s)
  • Federated machine learning

  • heterogeneity

  • regression

  • time series

  • solar power forecasting

  • Cookie settings
  • Imprint
  • Privacy policy
  • Api
  • Contact
© 2024