Spatio-temporal graph neural networks for power prediction in offshore wind farms using SCADA data

Simon Daenens, Timothy Verstraeten, Pieter-Jan Daems, Ann Nowé, Jan Helsen

TL;DR

Researchers have developed a spatio-temporal Graph Neural Network (GNN) that predicts offshore wind turbine power output with 30.3% less error than traditional methods. By combining LSTMs for time-series data with GNNs for spatial relationships, the model effectively captures complex "wake effects" and reduces undetected power loss events from 12.6% to a staggering 0.02%.

What Problem Does This Paper Address?

Predicting how much power a wind turbine will produce isn't as simple as looking at a manufacturer's power curve. In a real wind farm, turbines interact. Upstream turbines create "wakes"—regions of slower, more turbulent air—that significantly reduce the efficiency of turbines downstream.

Traditional models often treat turbines in isolation or use simplistic "binning" methods that fail to account for these dynamic, multi-directional spatial interactions. Furthermore, most models operate on 10-minute averages, which are too coarse to capture the rapid fluctuations (30-second scale) needed for real-time farm control and precise anomaly detection. The authors sought to bridge this gap by creating a model that "understands" both the layout of the farm and the flow of wind over time.

Methods

The authors introduced a Spatio-Temporal Graph Neural Network (ST-GNN). Here is the breakdown of their novel architecture:

  1. Graph Representation: The wind farm is modeled as a graph where each turbine is a node. Edges (connections) are drawn between turbines that have a clear "line-of-sight" to one another, allowing the model to learn spatial dependencies based on the physical layout.
  2. Temporal Processing (LSTM): Before looking at the neighbors, the model uses a Long Short-Term Memory (LSTM) network to process a 5-minute window of historical SCADA data for each individual turbine. This captures the "momentum" and trends in wind behavior.
  3. Spatial Aggregation (GENConv): The processed temporal features are then passed through a Generalized Graph Convolution (GENConv) layer. This is the "spatial" engine; it aggregates information from neighboring turbines (like wind speed and direction) to understand how one turbine's state affects another.
  4. Normal Behavior Modeling (NBM): The model is trained only on "normal" operating data. By comparing the model’s "ideal" prediction against actual performance, operators can identify anomalies (like a turbine being forced to slow down or a sensor failing) that aren't always caught by standard status logs.

Datasets

The study utilized a massive, high-resolution dataset:

  • Source: A real-world offshore wind farm in the Dutch–Belgian zone.
  • Scale: Over 40 turbines, each rated at >8 MW.
  • Duration: Two years of data (2021–2022).
  • Resolution: Original data was 1 Hz (one sample per second), resampled to 30-second averages for the model.
  • Features: SCADA signals including wind speed, direction (sine/cosine), turbulence intensity, and active power.

Key Insights and Ideas

  • Spatial Awareness is Non-Negotiable: The "aha moment" is that a turbine's performance is inextricably linked to its neighbors. The GNN outperformed standard Multi-Layer Perceptrons (MLPs) because it could explicitly model the "neighborhood" effect of wakes.
  • The Power of 30 Seconds: Moving from 10-minute averages to 30-second resolutions allows the model to capture the "propagation" of wind across the farm.
  • Adaptive Error Thresholds: Instead of a flat error margin for anomaly detection, the authors used a threshold based on the standard deviation of power at specific wind speeds. This "context-aware" detection prevents false alarms during high-turbulence periods where power output is naturally more volatile.
  • Capturing the "Unlogged": The model identified significant power loss events (like "derations" or subtle sensor issues) that were completely absent from the official operator logs, proving its value as a diagnostic tool.

Key Results

The ST-GNN showed dominant performance across the board compared to the traditional "Power Curve" (PC) baseline and standard MLPs:

  • Accuracy: Achieved a 30.3% reduction in Mean Absolute Error (MAE) and a 30.5% reduction in Mean Absolute Percentage Error (MAPE) compared to the power curve method.
  • Wake Modeling: In "waked" conditions (where turbines are behind others), the ST-GNN improved the energy ratio error by 57.4% over traditional binning.
  • Anomaly Detection: The most striking result was in capturing abnormal events. The traditional power curve method missed 12.6% of power loss events (False Negatives). The ST-GNN reduced this failure rate to just 0.02%.

Limitations and Future Work

While the results are impressive, the authors note a few areas for growth:

  • Transferability: The current model was trained and tested on the same farm. Future work needs to determine if a model trained on "Farm A" can be applied to "Farm B" without extensive retraining.
  • Environmental Variables: The model currently relies on wind and power data. Incorporating air density, wave height, or tide levels could further refine predictions in harsh offshore environments.
  • Data Standardization: To scale this across the industry, there needs to be a push for standardized SCADA data formats (like RDS-PP) to make integration easier across different turbine manufacturers.

Summary updated

Abstract

This paper introduces a novel model for predicting wind turbine power output in a wind farm at a high temporal resolution of 30 s. The wind farm is represented as a graph, with graph neural networks (GNNs) used to aggregate selected input features from neighboring turbines. A temporal component is introduced by feeding a time series of input features into the graph, processed through a long short-term memory (LSTM) network before being passed to the GNN. Our model is integrated into a normal behavior model (NBM) framework for analyzing power loss events in wind farms. The results show that both the spatial and the spatio-temporal GNN models outperform traditional data-driven power curve methods, achieving reductions in the mean absolute error (MAE) of approximately 22.6 % and 30.3 %, respectively, and in the mean absolute percentage error (MAPE) of around 20.7 % and 30.5 %, respectively. Notably, the spatio-temporal GNN demonstrates superior performance, attributed to its ability to effectively capture both spatial and temporal dynamics. Additionally, the model achieves remarkable agreement with SCADA-derived energy ratios across the full range of wind directions, with a weighted average error of 0.0373, an improvement of approximately 57.4 % compared to the power curve binning method. This advantage is especially pronounced under waked conditions, where traditional methods such as the power curve and multilayer perceptron (MLP) models exhibit significantly higher error rates. Beyond power prediction, we illustrate the model's effectiveness in detecting and analyzing instances of reduced performance and its ability to identify various types of abnormal events beyond what is recorded in standard status logs. Compared to the power curve method, the spatio-temporal GNN reduces the rate of undetected power loss events from 12.6 % to just 0.02 %, demonstrating a substantial improvement in capturing abnormal events.
BibTeX
@inproceedings{daenens2025spatio,
  author = {Simon Daenens and Timothy Verstraeten and Pieter-Jan Daems and Ann Nowé and Jan Helsen},
  title = {{Spatio-temporal graph neural networks for power prediction in offshore wind farms using SCADA data}},
  booktitle = {Wind},
  year = {2025},
  doi = {10.5194/wes-10-1137-2025},
}
Figure