Graph Neural Networks for Electric and Hydraulic Data Fusion to Enhance Short-term Forecasting of Pumped-storage Hydroelectricity

R. Theiler, Olga Fink

Abstract Pumped-storage hydropower plants (PSH) actively participate in grid power-frequency control and therefore often operate under dynamic conditions, which results in rapidly varying system states. Predicting these dynamically changing states is essential for comprehending the underlying sensor and machine conditions. This understanding aids in detecting anomalies and faults, ensuring the reliable operation of the connected power grid, and in identifying faulty and miscalibrated sensors. PSH are complex, highly interconnected systems encompassing electrical and hydraulic subsystems, each characterized by their respective underlying networks that can individually be represented as graphs. To take advantage of this relational inductive bias, graph neural networks (GNNs) have been separately applied to state forecasting tasks in the individual subsystems, but without considering their interdependencies. In PSH, however, these subsystems depend on the same control input, making their operations highly interdependent and interconnected. Consequently, hydraulic and electrical sensor data should be fused across PSH subsystems to improve state forecasting accuracy. This approach has not been explored in GNN literature yet because many available PSH graphs are limited to their respective subsystem boundaries, which makes the method unsuitable to be applied directly. In this work, we introduce the application of spectral-temporal graph neural networks, which leverage self-attention mechanisms to concurrently capture and learn meaningful subsystem interdependencies and the dynamic patterns observed in electric and hydraulic sensors. Our method effectively fuses data from the PSH's subsystems by operating on a unified, system-wide graph, learned directly from the data. This approach leads to demonstrably improved state forecasting performance and enhanced generalizability.

Introduction

Pumped-storage hydropower (PSH) plants are the "giant batteries" of the power grid, essential for balancing supply and demand. Because they respond dynamically to grid fluctuations, their internal states change rapidly. Accurate short-term forecasting is critical for detecting anomalies and preventing equipment failure.

However, PSH plants are complex systems consisting of two distinct physical domains: the hydraulic system (water flow, pressure, lake levels) and the electrical system (currents, voltages). Traditionally, these are modeled in "silos" using separate engineering principles. While Graph Neural Networks (GNNs) are effective at modeling relational data, previous attempts in this field usually applied them to each subsystem separately, ignoring the mechanical links and shared control inputs. Furthermore, building these graphs manually from technical schematics is labor-intensive and often results in graphs that aren't optimized for machine learning.

Methodology

The researchers developed a Spectral-Temporal Graph Neural Network (STGNN) designed to bridge the gap between the hydraulic and electrical domains. The architecture consists of two main stages:

Architecture of the Spectral-Temporal Graph Neural Network (STGNN) for PSH state forecasting, featuring attention-based graph learning and spectral filtering.
  1. Attention-based Graph Learning: Instead of using a fixed graph based on a blueprint, the model learns the relationships between sensors. It uses a Gated Recurrent Unit (GRU) to embed time-series data and a self-attention mechanism to calculate "attention scores" between sensor nodes. These scores are converted into a symmetrical adjacency matrix, effectively discovering which sensors influence one another across domains.
  2. Spectral Filtering: Once the graph is established, the model applies filters in the spectral domain. It uses Chebyshev polynomials to approximate graph convolutions (capturing spatial relationships) and Discrete Fourier Transforms (DFT) combined with Gated Linear Units (GLU) to capture temporal patterns.

Experimental Setup

The study used a real-world dataset collected in collaboration with the Swiss Federal Railways (SBB), covering four months of data from a PSH plant in Switzerland. The dataset included 58 signals, encompassing hydraulic readings (pressures, flow rates, lake levels) and electrical readings (phasor currents). Data was averaged to a 1-minute resolution, with the primary goal of forecasting phasor currents for the electrical subsystem.

Key Results and Discussion

The STGNN was compared against several baselines using Normalized Mean Squared Error (NMSE). The proposed model outperformed all baselines, including LSTMs by 28% and larger Fully Connected Networks by 14%.

Forecasting performance comparison: The fused model (EL+HYD) shows closer alignment with ground truth and lower NMSE spikes compared to the electric-only (EL) model.

The Benefit of Cross-Domain Fusion

Incorporating hydraulic data into the electrical forecast provided a measurable 6.5% boost in accuracy. The fused model achieved an NMSE of 5.34e-3, compared to 5.71e-3 for the electric-only model. This improvement was robust, with the forecast improving for 19 out of 21 electrical sensors when hydraulic data was added.

Learned vs. Schematic Graphs

A striking finding was that a GNN using a graph derived from a technical schematic actually performed worse than a simple LSTM. Schematic-based graphs often have properties like high diameters and low node degrees that hinder message passing. Learning the graph from data bypasses this bottleneck and allows the model to focus on physical causal chains, such as the relationship between penstock flow rates and electrical currents.

Conclusion

This work demonstrates that fusing electrical and hydraulic sensor data through a learned, system-wide graph significantly improves short-term forecasting in PSH plants. By leveraging spectral-temporal graph neural networks, the model captures complex interdependencies that traditional siloed approaches miss. Future work will explore the integration of physics-informed loss functions and the combination of prior engineering knowledge with data-driven attention mechanisms to further enhance generalization.

Summary updated

Abstract

Pumped-storage hydropower plants (PSH) actively participate in grid power-frequency control and therefore often operate under dynamic conditions, which results in rapidly varying system states. Predicting these dynamically changing states is essential for comprehending the underlying sensor and machine conditions. This understanding aids in detecting anomalies and faults, ensuring the reliable operation of the connected power grid, and in identifying faulty and miscalibrated sensors. PSH are complex, highly interconnected systems encompassing electrical and hydraulic subsystems, each characterized by their respective underlying networks that can individually be represented as graphs. To take advantage of this relational inductive bias, graph neural networks (GNNs) have been separately applied to state forecasting tasks in the individual subsystems, but without considering their interdependencies. In PSH, however, these subsystems depend on the same control input, making their operations highly interdependent and interconnected. Consequently, hydraulic and electrical sensor data should be fused across PSH subsystems to improve state forecasting accuracy. This approach has not been explored in GNN literature yet because many available PSH graphs are limited to their respective subsystem boundaries, which makes the method unsuitable to be applied directly. In this work, we introduce the application of spectral-temporal graph neural networks, which leverage self-attention mechanisms to concurrently capture and learn meaningful subsystem interdependencies and the dynamic patterns observed in electric and hydraulic sensors. Our method effectively fuses data from the PSH's subsystems by operating on a unified, system-wide graph, learned directly from the data, This approach leads to demonstrably improved state forecasting performance and enhanced generalizability.
BibTeX
@inproceedings{theiler2024graph,
  author = {R. Theiler and Olga Fink},
  title = {{Graph Neural Networks for Electric and Hydraulic Data Fusion to Enhance Short-term Forecasting of Pumped-storage Hydroelectricity}},
  booktitle = {PHM},
  year = {2024},
  doi = {10.36001/phme.2024.v8i1.4129},
}
Figure