Abstract
Accurate short-term state forecasting is essential for efficient and stable operation of modern power systems, especially in the context of increasing variability introduced by renewable and distributed energy resources. As these systems evolve rapidly, it becomes increasingly important to reliably predict their states in the short term to ensure operational stability, support control decisions, and enable interpretable monitoring of sensor and machine behavior. Modern power systems often span multiple physical domains - including electrical, mechanical, hydraulic, and thermal - posing significant challenges for modeling and prediction. Graph Neural Networks (GNNs) have emerged as a promising data-driven framework for system state estimation and state forecasting in such settings. By leveraging the topological structure of sensor networks, GNNs can implicitly learn inter-sensor relationships and propagate information across the network. However, most existing GNN-based methods are designed under the assumption of homogeneous sensor relationships and are typically constrained to a single physical domain. This limitation restricts their ability to integrate and reason over heterogeneous sensor data commonly encountered in real-world energy systems, such as those used in energy conversion infrastructure. In this work, we propose the use of Heterogeneous Graph Attention Networks to address these limitations. Our approach models both homogeneous intra-domain and heterogeneous inter-domain relationships among sensor data from two distinct physical domains - hydraulic and electrical - which exhibit fundamentally different temporal dynamics. Experimental results demonstrate that our method significantly outperforms conventional baselines on average by 35.5% in terms of normalized root mean square error, confirming its effectiveness in multi-domain, multi-rate power system state forecasting.
Introduction
Modern power systems are becoming increasingly complex, integrating renewable energy sources and high-speed traction networks that introduce rapid, volatile fluctuations. To maintain grid stability, operators require accurate short-term state forecasting—the ability to predict sensor readings seconds or minutes into the future.
A major challenge is that energy conversion systems, such as pumped-storage hydropower plants (PSH), are multi-domain. They involve slow-moving hydraulic dynamics (e.g., water levels, pressures) alongside high-frequency electrical dynamics (e.g., current, voltage). Most existing Graph Neural Networks (GNNs) are homogeneous, treating all sensors and connections identically. This ignores the fundamental physical differences and varying time scales inherent in multi-domain systems.
Methodology
We propose a modular framework based on Heterogeneous Graph ATtention Networks (HGATs). The approach follows a "time-then-graph" paradigm designed to handle multi-rate data and cross-domain dependencies.
The framework consists of four key stages:
- First-order Differentiation: Instead of predicting raw values, the model forecasts the change in state ($\dot{x}$), allowing it to better capture dynamic behavior across different sampling rates.
- Node Embeddings: Local temporal history for each sensor is encoded using Gated Recurrent Units (GRUs). Crucially, distinct GRU parameters are used for electrical and hydraulic nodes to respect their unique physics.
- Heterogeneous Message Passing: The core utilizes GATv2-based layers that allow nodes to "attend" to their neighbors. By defining specific edge types (e.g., electrical-to-electrical, hydraulic-to-electrical), the model learns specialized interaction functions for cross-domain dependencies.
- Euler Integration: The predicted derivatives are integrated using the first-order Euler method to reconstruct the final forecasted signal trajectories.
Experimental Evaluation
The study utilized real-world datasets from a Swiss pumped-storage hydropower plant in collaboration with the Swiss Federal Railways (SBB). Two case studies were conducted:
- 1Min Case Study: Four months of data at one-minute intervals with 84 sensors.
- 1Sec Case Study: High-frequency 1Hz data focusing on rapid load fluctuations (up to 250 MW) typical of railway grids.
The HGAT model consistently outperformed baselines including LSTMs, 1D-CNNs, and standard GNNs, achieving an average improvement of 35.5% in Normalized Root Mean Square Error (NRMSE). It was particularly effective at capturing "change points"—sudden jumps in current and power—and demonstrated superior stability across training runs.
Discussion
A key insight is that explicitly modeling heterogeneity improves accuracy. Ablation studies showed that removing hydraulic data degraded electrical forecasts, proving that cross-domain mechanical couplings provide vital information. Furthermore, unlike models that attempt to infer relationships solely from data (such as StemGNN), this HGAT uses actual plant schematics as a "structural prior." This guide makes the learning process more efficient and the results more interpretable for engineers.
Conclusion
While the HGAT framework is a significant step forward, it currently relies on a static graph. Future work will incorporate dynamic topologies to account for control actions like breaker operations. Additionally, integrating physics-informed loss functions (e.g., mass balance or Kirchhoff’s laws) could further improve the physical consistency of the forecasts and prevent the prediction of impossible system states.
Summary updated