TL;DR
This paper introduces a novel two-stage approach for analyzing historical wind turbine SCADA data to identify hidden operational anomalies. By first accurately modeling the normal power curve and then applying a robust, distribution-free control chart, the method effectively detects subtle deviations that traditional methods miss, leading to improved wind turbine monitoring and maintenance.
What Problem Does This Paper Address?
Wind turbines are a crucial part of the renewable energy landscape, but their operation and maintenance (OM) costs can be significant. Effective condition monitoring (CM) is key to reducing these costs and improving reliability. SCADA data, collected at regular intervals, is a rich source of information for CM. However, this data is often noisy, suffers from low sampling frequencies, and is influenced by complex turbine dynamics. Existing CM methods often struggle with these challenges, particularly in identifying subtle, "hidden" operating anomalies that don't trigger immediate alarms. These anomalies, such as gradual degradation or minor control issues, can lead to reduced efficiency and eventual failures if not detected early. This paper focuses on "Phase I analysis," which involves retrospectively analyzing historical data to characterize the "in-control" (normal) state and filter out variabilities, a crucial prerequisite for reliable online monitoring.
Methods
The paper proposes a two-part methodology:
- Power Curve Estimation with MARS and IFGLS:
- Multivariate Adaptive Regression Splines (MARS): This is a flexible, non-parametric technique used to model the complex relationship between various weather conditions (wind speed, direction, temperature, etc.) and the turbine's power output. MARS is chosen for its ability to handle high-dimensional, interactive inputs and its interpretable basis functions.
- Iterative Feasible Generalized Least Squares (IFGLS): SCADA data residuals often exhibit serial correlation (autocorrelation). IFGLS is applied iteratively to reduce this autocorrelation in the residuals after the MARS model is fitted. This step is crucial for improving the accuracy of the power curve estimation and the subsequent anomaly detection.
- Distribution-Free Control Chart (RS/P):
- Recursive Segmentation and Permutation (RS/P): After the power curve residuals are adjusted for autocorrelation, they are assumed to be independent and identically distributed (i.i.d.) but with an unknown distribution. RS/P is a powerful, distribution-free control chart that can detect a wide range of "out-of-control" (OC) patterns, including level changes (shifts in the mean). It achieves this by recursively segmenting the data and using permutation tests to determine significance, thus avoiding assumptions about the underlying data distribution. The paper specifically focuses on detecting level changes in the residual process.
Datasets
The study utilizes SCADA data from an open data wind farm called "La Haute Borne" in Meuse, France, provided by ENGIE Renewable Energy. Two datasets were used for experiments, differing in their time span and number of data points after initial filtering:
- Dataset 1: 2012-12-31 23:00 to 2013-04-01 22:00 (10014 points after filtering)
- Dataset 2: 2016-01-01 23:00 to 2016-12-31 22:50 (41374 points after filtering)
The paper notes that these datasets lack explicit control/alarm records, making the proposed data-driven approach particularly relevant.
Key Insights and Ideas
- The Power Curve as a Diagnostic Tool: The paper highlights the power curve not just as a performance indicator but as a critical benchmark for detecting anomalies in wind turbine operation.
- Autocorrelation is a Major Hurdle: Ignoring serial correlation in SCADA data residuals can lead to less effective monitoring and higher false alarm rates. The IFGLS step is essential for mitigating this.
- Distribution-Free is Key: Assuming a specific distribution for SCADA residuals is risky. RS/P's distribution-free nature makes it robust and applicable to real-world data where distributions are often unknown or complex.
- Subtle Anomalies Matter: The RS/P chart is capable of detecting patterns and deviations that might be missed by simpler clustering or threshold-based methods, offering a more nuanced view of turbine health.
- Iterative Anomaly Removal: The method iteratively removes detected out-of-control segments and re-analyzes, allowing for the identification of multiple anomalies and a clearer picture of the normal operating state.
Key Results
- Reduced Autocorrelation: Table II demonstrates that MARS with IFGLS significantly reduces the Root Mean Squared Error (RMSE) of the fitted power curve compared to MARS alone, indicating a more accurate model and effective reduction of autocorrelation. For Dataset 1, RMSE dropped from 39.18 to 30.08, and for Dataset 2, from 42.86 to 35.84.
- Detection of Informative OC Patterns: Figure 3 visually showcases the detected out-of-control segments, illustrating various deviations from the normal power curve. These include periods of bad sensor readings, inefficient data transmission, effects of unmeasured variables like air density, down-rating periods, and segments preceding emergency shutdowns.
- Improved Significance: Excluding detected OC segments led to an increase in the significance level (p-value) of the aggregated control statistic, indicating a better characterization of the in-control state.
Limitations and Future Work
- Focus on Level Changes: The current RS/P implementation primarily focuses on detecting level changes (shifts in the mean) of the residual process. The authors acknowledge that detecting "scale shifts" (changes in variance) is also important and requires further development.
- Unmeasured Variables: The analysis of OC patterns revealed the impact of unmeasured variables (e.g., air density) on power output. Future work could explore incorporating more data sources or advanced modeling techniques to account for these.
- Coefficient Monitoring: The paper did not monitor the coefficients of the power curve model itself for anomalies. This is identified as a non-trivial area for future research.
- Expert Diagnosis: While the method identifies anomalies, providing more "considerate diagnoses" requires further integration with expert knowledge and potentially more detailed operational data.
Summary updated