the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Spatio-Temporal Graph Neural Networks for Power Prediction in Offshore Wind Farms Using SCADA Data
Abstract. This paper introduces a novel model for predicting wind turbine power output within a wind farm at a high temporal resolution of 30 seconds. 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 added by feeding a timeseries of input features into the graph and utilizing a hybrid GNN-LSTM model architecture. This approach sequentially extracts temporal features with the LSTM component and spatial features with the GNN component. 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 Spatio-Temporal GNN models outperform traditional data-driven power curve methods, with the Spatio-Temporal GNN demonstrating superior performance due to its ability to capture both spatial and temporal dynamics. Additionally, 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.
- Preprint
(1169 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on wes-2024-113', Anonymous Referee #1, 25 Oct 2024
This paper presents a spatial graph neural network (GNN) and spatio-temporal GNN for modeling the power of individual turbines in a wind farm using wind speed, wind direction, and turbulence intensity measurements from all turbines. The spatio-temporal GNN also uses lagged values of these measurements to capture temporal dependencies within the wind farm. The GNN model's accuracy relative to a simple power curve model for predicting power output is assessed. Further, the ability of the GNN model to identify underperformance events, by comparing the predicted power during normal operation to the actual power, is analyzed, showing that the GNN model can detect underperformance events with high accuracy.
The paper is overall well-written, but could use more details about the GNN models, especially for readers less familiar with neural networks. For example, clarifying terms and adding a figure illustrating the full GNN model would be helpful. Additionally, although the benefits of using the GNN compared to the simple power curve are quantified in terms of power prediction errors, to better show the value of using the GNN models, the GNN performance in predicting abnormal events should be compared to the performance using the simple power curve model. The power curve model is much simpler to implement, so if the abnormal event detection accuracy is similar when using the power curve model, it might not be worth implementing the much more complex GNN model. Finally, since many of the conclusions of the paper depend on comparison with the simple power curve approach, more details on how the power curve model was developed would be useful.
Specific comments are provided below.
- Pg. 2, ln. 53: Please define LSTM.
- Pg. 3, ln. 69: "it cannot provide a precise forecast of wind power as it fails to accurately model the nonlinear relationship between wind speed and power output": A power curve does not necessarily have to be a linear function, and can capture the nonlinear relationship between wind speed and power if modeled correctly. Can you explain this statement in more detail?
- Pg. 5, ln. 127: "both computed over a 10-minute interval centered around the 1-second data point.": Does this mean that the method for predicting power outputs will have at least a 5-minute lag, because the computed TI value is delayed by 5 minutes?
- Pg. 5, ln. 137: "uses properties of the power curve based on the IEC standard to annotate steady-state control conditions.": Please mention which IEC standard is used and provide a reference.
- Pg. 5, ln. 138: "Using this approach, different control regimes can be identified, and outliers and abnormal behavior can be removed from the dataset.": Do you remove entire timestamps (for all turbines) when any individual turbine has an outlier/abnormal operation identified? Or do you only remove the data for the specific turbine with abnormal operation?
- Pg. 5, ln. 140: "lagged values of the input features are incorporated into the model": How many samples in the past are used? Do you just use the current and previous timestamps, or several timestamps in the past?
- Pg. 5, ln. 142: "to ensure that the lagged values utilized by the model are consecutive and there are no discontinuities due to missing data.": Can you explain the criteria for determining continuous data? Are you identifying timestamps where there is no missing data for all turbines for the current and lagged timestamps? Further, for a large wind farm, it could be rare that all turbines are operating normally at any given time (due to repairs, derating, etc.). How do you handle the case when there are always some turbines with missing data?
- Pg. 5, ln. 143: "partitioned into distinct subsets for model training, validation, and testing": A brief explanation of these three subsets would be helpful for readers not as familiar with AI/ML training and validation.
- Algorithm 1: Please mention the value of "tolerance" that is used.
- Pg. 7, ln. 154: "containing a set of geometric features, such as the length and direction of each edge": Can you state exactly which features are used as edge attributes in this work? This sentence makes it sounds like length and direction are just two examples of what could be used.
- Pg. 7, lns. 175-183: These two paragraphs contain a lot of terms that require further explanation for readers less familiar with neural networks. For example, "generalized message aggregators", "pre-activation residual connectors", "message normalization layers", "feature encoder", "sigmoid activation", "multi-layer perceptrons", "ReLU", "DeepGCNLayer", "preactivation residual connections". I would suggest describing these terms in more detail and what their purpose is in the GNN model. Or references could be provided for some of the terms. Also, in addition to Figure 1, a block diagram outlining the entire GNN model from the inputs to the estimated power outputs would greatly clarify the model and put some of these terms in context.
- Pg. 8, lns. 184-186: Similarly, please explain "LSTM" and "node feature encoder" and how they are used in the model. A figure showing the full model architecture for the Spatio-Temporal GNN would be helpful as well to clarify how it differs from the spatial GNN.
- Figure 1: Please label the subfigures in the figure (a, b, c, d).
- Figure 1: Please describe what specific functions are used for the "UPDATE" and "AGGREGATE" functions. Earlier in Section 3.3 you mention that they are arbitrary differentiable function (i.e., neural networks), but is isn’t clear what is actually used in this work. Further, the equations in Fig. 1c do not match the form of the equation in line 169. Is one of these incorrect? If so, please make sure they match or clarify which form is actually implemented in your work.
- Pg. 8, ln. 194: "based on the objective function.": What specific objective function is used? Is it the MSE of the turbine power for all turbines?
- Pg. 8, ln. 196: "Hyperparameter combinations that resulted in the lowest MSE for the validation dataset were retained…": What were the specific combinations that were ultimately selected for the final model?
- Table 3: Please explain the hyperparameters and how they are used in the GNN models.
- Section 4.1: Many of the conclusions of this work are based on a comparison of the GNN performance with the simple power curve approach. Please discuss in more detail what kind of power curve model is used. This will help put the results in context.
- Pg. 12, ln. 237: "computed for each wind direction bin": How do you estimate the wind direction?
- Pg. 12, ln. 241: "One turbine is positioned in the free flow relative to the dominant direction": Is this the turbine in the numerator or denominator of Eq. 2?
- Pg. 12, ln. 245: "Our spatio-temporal GNN model demonstrates remarkable agreement with the energy ratios from the SCADA data": Can you include the power curve-based energy ratio estimates in Fig. 3? This would help show how much value there is in using the GNN model instead of a simple power curve estimator.
- Pg. 12, ln. 255: "shows the Spatio-Temporal GNN model's effectiveness in capturing complex dynamics of wake interactions within the wind farm.": Much of the complex wake behavior would be captured by the turbines' own wind speed measurements, which are used as features in the GNN model and would also be the inputs to the simple power curve model. So it seems likely that the simple power curve model would already capture these wake dynamics.
- Sections 4.2 and 4.3: For the results in these sections, please mention whether you are using the spatial GNN or the spatio-temporal GNN.
- Pg. 12, ln. 264: "the standard deviation of the produced power for a given wind speed": Should this be "for a given wind speed bin"? And if so, what bin width is used?
- Fig. 4: Comparing these results to the confusion matrix results when using the simple power curve model would show how much value there is in using the GNN model. If the results are similar, it seems like it would be preferable to just use the simple power curve model.
- Figs. 5 and 6: Can you also compare these results to the estimated reduced performance periods based on the simple power curve model?
- Fig. 6: The "number of turbines with reduced performance" axis label should include "normalized" since the values are between 0 and 1.
Citation: https://doi.org/10.5194/wes-2024-113-RC1 -
RC2: 'Comment on wes-2024-113', Anonymous Referee #2, 06 Nov 2024
This paper is about spatial and spatio-temporal GNN models for power prediction and fault detection. It is a ver interesting and important topic, and the paper is generally well-written and clear. However, there are a number of improvements required to both the structure and the content, as well as some small edits, which can all be found in the annotations to the paper in the attached file.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
212 | 80 | 5 | 297 | 3 | 6 |
- HTML: 212
- PDF: 80
- XML: 5
- Total: 297
- BibTeX: 3
- EndNote: 6
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1