Heterogeneous graph neural networks for short-term state forecasting in power systems across domains and time scales: a hydroelectric power plant case study
Machine Learning: Engineering · 01.06.2026 · 2 (1) · 015009
DOI: 10.1088/3049-4761/ae565c ↗Kurz gefasst
Fachartikel in der Zeitschrift Machine Learning: Engineering, online erschienen am 13.04.2026. Verfasst von Raffael Theiler und 1 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Swiss Federal Office of Transport FOT.
Abstract (Originalsprache)
Accurate short-term state forecasting is crucial for the efficient and stable operation of modern power systems, especially as renewable and distributed energy resources introduce greater variability in power generation and load dynamics. Power systems today span multiple physical domains such as electrical, hydraulic, mechanical, and thermal, making modeling and prediction particularly challenging due to complex multi-rate dynamics and heterogeneous subsystem interactions. While graph neural networks (GNNs) have shown promise for system state estimation by leveraging network topology to capture inter-sensor relationships, most existing approaches assume homogeneous sensor types and are typically restricted to a single physical domain, which limits their real-world applicability. To address these challenges, we propose a heterogeneous graph attention network (HGAT) framework that integrates both homogeneous intra-domain and heterogeneous inter-domain relationships, focusing on sensor data from hydraulic and electrical subsystems, each exhibiting distinct temporal dynamics. HGAT constructs a unified system-wide heterogeneous graph that fuses data across domains, enabling the model to learn cross-domain interactions often neglected in traditional GNNs. The heterogeneous processing introduces signal grouping as an inductive bias, which improves forecasting performance compared to strong homogeneous and monolithic baseline models. We evaluate our approach on two real-world hydroelectric power plant datasets with sensor measurements at one-minute and one-second resolutions. Experimental results show that our method outperforms state-of-the-art baselines by an average of 18.55% in terms of normalized root mean square error, confirming its effectiveness for multi-domain, multi-rate state forecasting in practical energy system settings.
Förderung
- Swiss Federal Office of Transport FOT
Zitieren
APA
Theiler, R., & Fink, O. (2026). Heterogeneous graph neural networks for short-term state forecasting in power systems across domains and time scales: a hydroelectric power plant case study. Machine Learning: Engineering, 2(1), 015009. https://doi.org/10.1088/3049-4761/ae565c
Vancouver
Theiler R, Fink O. Heterogeneous graph neural networks for short-term state forecasting in power systems across domains and time scales: a hydroelectric power plant case study. Machine Learning: Engineering. 2026;2(1):015009. doi:10.1088/3049-4761/ae565c
BibTeX
@article{Theiler2026,
author = {Theiler, Raffael and Fink, Olga},
title = {Heterogeneous graph neural networks for short-term state forecasting in power systems across domains and time scales: a hydroelectric power plant case study},
journal = {Machine Learning: Engineering},
year = {2026},
volume = {2},
number = {1},
pages = {015009},
doi = {10.1088/3049-4761/ae565c}
}
EUVIMED