Abstract
The wind turbine power curve model is critical to a wind turbine’s power prediction and performance analysis. However, abnormal data in the training set decrease the prediction accuracy of trained models. This paper proposes a sample average approach-based method to construct an interval model of a wind turbine, which increases robustness against abnormal data and further improves the model accuracy. We compare our proposed methods with the traditional neural network-based and Bayesian neural network-based models in experimental data-based validations. Our model shows better performance in both accuracy and computational time.
Introduction: The Challenge of "Dirty" Data
Wind turbine power curves—the relationship between wind speed and electrical power output—are essential "fingerprints" for assessing turbine health and predicting energy yield. Engineers rely on these curves to detect mechanical faults and optimize performance. However, data collected via Supervisory Control and Data Acquisition (SCADA) systems is frequently "dirty." Factors such as communication failures, sensor icing, and scheduled maintenance introduce significant outliers that do not reflect actual turbine performance.
Standard regression models are highly sensitive to these outliers. As illustrated in Figure 1, abnormal data points can significantly pull the estimated curve away from the real performance line, leading to biased predictions.
Methodology: The SSELM Approach
To address this, researchers developed the SSELM (Sample average approach-based extreme learning machine). This method integrates two distinct frameworks:
- Extreme Learning Machine (ELM): A single-layer feedforward neural network characterized by randomly assigned hidden layer weights. By only training the output weights, ELM offers exceptional computational speed compared to traditional deep learning models.
- Sample Average Approach (SAA): A technique from chance-constrained optimization. Rather than fitting every data point, SAA identifies a "confidence region" or interval that encompasses a specific majority of the data (e.g., 91%).
By modeling the power curve as an interval rather than a single line, the system creates a "corridor" of normal operation. Data points falling outside this corridor are automatically identified as abnormal, allowing for simultaneous data cleaning and regression.
Experimental Validation
The model was tested using real-world SCADA data from a wind farm in Hubei, China. To establish a ground truth, human specialists manually labeled data points as "normal" or "abnormal."
The SSELM was benchmarked against several models:
- TrueELM: Trained on manually cleaned data.
- Standard ELM: Trained on raw, noisy data.
- Bayesian ELM (BELM) and Scenario-based Neural Network (SNN).
Key Results
- Superior Outlier Detection: SSELM achieved 91.3% accuracy in identifying abnormal data, outperforming BELM (78.9%) and SNN (63.5%).
- High Prediction Precision: In terms of Mean Absolute Error (MAE), SSELM (21.39) nearly matched the performance of TrueELM (21.34), despite the latter having the advantage of pre-cleaned data. The standard ELM trained on raw data lagged significantly with an MAE of 23.35.
- Efficiency: The ELM-based architecture ensures the model remains computationally light, making it ideal for real-time SCADA monitoring.
Conclusion and Future Directions
The SSELM provides a robust, one-step solution for wind turbine power curve modeling. By shifting from point-based regression to interval-based estimation, it naturally filters sensor noise and maintenance-related outliers.
Future research will focus on integrating temporal dependencies (time-series analysis) and developing automated methods for setting the probability threshold based on specific farm noise profiles. Expanding this approach across diverse turbine fleets will be the next step in validating its scalability.
Summary updated