• Burgess, D. W., , Crum T. D. , , and Vogt R. J. , 2008: Impacts of wind farms on WSR-88D radars. Preprints, 24th Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, New Orleans, LA, Amer. Meteor. Soc., 6B.3. [Available online at http://ams.confex.com/ams/pdfpapers/128810.pdf].

    • Search Google Scholar
    • Export Citation
  • Cheong, B. L., , Palmer R. D. , , and Xue M. , 2008: A time-series weather radar simulator based on high-resolution atmospheric model simulations. J. Atmos. Oceanic Technol., 25 , 230243.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cornwall, J. M., 2007: Mitigation of wind turbine effects on FAA and military radars. Adapted from Jason Rep. JSR-07-126, MITRE Corporation, 8 pp.

    • Search Google Scholar
    • Export Citation
  • Department of Defense, 2006: The effect of windmill farms on military readiness. Office of the Director of Defense Research and Engineering, 62 pp. [Available online at http://www1.eere.energy.gov/windandhydro/federalwindsiting/pdfs/dod_windfarms.pdf].

    • Search Google Scholar
    • Export Citation
  • Department of Energy, 2008: 20% wind energy by 2030. Energy Efficiency and Renewable Energy Rep. DOE/GO-102008-2567, 229 pp.

  • Gallardo, B., , Perez F. , , and Aguado F. , 2008: Characterization approach of wind turbine clutter in the Spanish Weather Radar Network. Proc. Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, European Meteorological Society, 9.8.

    • Search Google Scholar
    • Export Citation
  • Green, D. M., , and Swets J. A. , 1966: Signal Detection Theory and Psychophysics. 1st ed. Robert E. Krieger Publishing, 455 pp.

  • Greving, G., , and Malkomes M. , 2006: On the concept of the radar cross section RCS of distorting objects like wind turbines for the weather radar. Proc. Fourth European Conference on Radar in Meteorology and Hydrology, Barcelona, Spain, European Meteorological Society, P5.8.

    • Search Google Scholar
    • Export Citation
  • Greving, G., , and Malkomes M. , 2008: Weather radar and wind turbines—Theoretical and numerical analysis of the shadowing effects and mitigation concepts. Proc. Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, European Meteorological Society, P2.12.

    • Search Google Scholar
    • Export Citation
  • Hubbert, J. C., , Dixon M. , , and Ellis S. M. , 2009: Weather radar ground clutter. Part II: Real-time identification and filtering. J. Atmos. Oceanic Technol., 26 , 11811197.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Isom, B. M., 2007: Characterization and mitigation of wind turbine clutter for the WSR-88D network. M.S. thesis, School of Electrical and Computer Engineering, University of Oklahoma, 93 pp.

  • Isom, B. M., and Coauthors, 2009: Detailed observations of wind turbine clutter with scanning weather radars. J. Atmos. Oceanic Technol., 26 , 894910.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Liu, H., , and Chandrasekar V. , 2000: Classification of hydrometeors based on polarimetric radar measurements: Development of fuzzy logic and neuro-fuzzy systems, and in situ verification. J. Atmos. Oceanic Technol., 17 , 140164.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Meymaris, G., , and Williams J. K. , 2007: Spectrum width estimators for the NEXRAD ORDA: Evaluation and recommendation. Preprints, 23rd Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, San Antonio, TX, Amer. Meteor. Soc., 5B.5. [Available online at http://ams.confex.com/ams/pdfpapers/119628.pdf].

    • Search Google Scholar
    • Export Citation
  • National Oceanic and Atmospheric Administration, 2006: WSR-88D Products and Algorithms. Part C, Doppler Radar Meteorological Observations, Federal Meteorological Handbook 11, FCM-H11C-2006, Office of the Federal Coordinator for Meteorological Services and Supporting Research, 390 pp.

    • Search Google Scholar
    • Export Citation
  • Papoulis, A., , and Pillai S. U. , 2002: Probability, Random Variables, and Stochastic Processes. McGraw-Hill, 852 pp.

  • Perry, J., , and Biss A. , 2007: Wind farm clutter mitigation in air surveillance radar. IEEE Aerosp. Electron. Syst. Mag., 22 , 3540.

  • Siggia, A. D., , and Passarelli R. E. , 2005: Gaussian model adaptive processing (GMAP) for improved ground clutter cancellation and moment calculation. M S-NR:ERAD3-P-00117, Sigmet, Inc., 8 pp.

    • Search Google Scholar
    • Export Citation
  • Skolnik, M. I., 1962: Introduction to Radar Systems. 3rd ed. McGraw-Hill, 648 pp.

  • Vogt, R. J., , Reed J. , , Crum T. , , Snow J. T. , , Palmer R. D. , , Isom B. M. , , and Burgess D. W. , 2007: Impacts of wind farms on WSR-88D operations and policy considerations. Preprints, 23rd Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, San Antonio, TX, Amer. Meteor. Soc., 5B.7. [Available online at http://ams.confex.com/ams/pdfpapers/120352.pdf].

    • Search Google Scholar
    • Export Citation
  • Vogt, R. J., and Coauthors, 2009: Continued progress in assessing and mitigating wind farm impacts on WSR-88Ds. Preprints, 25th Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Phoenix, AZ, Amer. Meteor. Soc., 11B.6. [Available online at http://ams.confex.com/ams/pdfpapers/150646.pdf].

    • Search Google Scholar
    • Export Citation
  • Yu, T-Y., , Wang Y. , , Shapiro A. , , Yeary M. B. , , Zrnić D. S. , , and Doviak R. J. , 2007: Characterization of tornado spectral signatures using higher-order spectra. J. Atmos. Oceanic Technol., 24 , 19972013.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zrnić, D. S., 1975: Simulation of weatherlike Doppler spectra and signals. J. Appl. Meteor., 14 , 619620.

  • Zrnić, D. S., , Ryzhkov A. V. , , Straka J. , , Liu Y. , , and Vivekanandan J. , 2001: Testing a procedure for automatic classification of hydrometeor types. J. Atmos. Oceanic Technol., 18 , 892913.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • View in gallery

    Temporal evolution of WTC spectra. The large power return at 0 Doppler velocity is attributed to the tower of the wind turbine. The flashes, which appear periodically, are from the rotating blades of the turbine when in the vertical position, and the oscillation around 0 velocity is attributed to the slowly moving structures near the rotor hub of the wind turbine.

  • View in gallery

    A hub-contaminated weather spectrum. The maximum value of the spectrum is used to compute the power of the “weather” signal. In this case, the power of the actual weather signal was computed, and when compared to the power of the hub contamination, it was determined that the weather signal was contaminated.

  • View in gallery

    Plot of the delta bias. The bias due to clutter filtering can be seen near 0 Doppler velocity. The increase in the delta bias near 9 m s−1 indicates hub contamination.

  • View in gallery

    Plots of (left) the temporal evolution of WTC spectra with (middle) corresponding delta biases for radial velocity and (right) truth data. Classifications for each CSR and WTC spectra were selected such that 99% of the delta-bias values were lower than the selected threshold of 1 m s−1 (shown in white). The truth data (right) illustrate what range of CSR values are considered contaminated for each WTC spectra.

  • View in gallery

    Histograms for (top) (left) CPA and (right) σS−1; (bottom) (left) μ4 and (right) HWR. Each feature has a range of overlapping values for contaminated (solid) and noncontaminated (dashed) data, making identification of WTC difficult using a single feature. However, the four features can be combined in an FLS to effectively detect WTC contamination.

  • View in gallery

    Diagram of the fuzzy logic system for WTC detection. Inputs to the FLS are the detection features. Each feature is fuzzified using a membership function. The fuzzy feature values are then weighted and aggregated before being compared against a threshold.

  • View in gallery

    Original (dashed) and optimized (solid) membership functions for the FLS: (top) (left) CPA and (right) σS−1; (bottom) (left) μ4 and (right) HWR. Membership functions for CPA and σS−1 were more aggressive in the detection algorithm after the optimization while the membership functions for μ4 and HWR were too aggressive before optimization.

  • View in gallery

    ROC plot before (dashed) and after (solid) optimization. FLS threshold values before optimization range from 0.0 to 1.0 in increments of 0.05 and correspond to the points beginning at the upper-right corner and follow the dashed line to the bottom-left corner. Threshold values after optimization range from 0.15 to 0.5 in increments of 0.05 and correspond to the points on the solid line from right to left. The optimal thresholds before and after optimization are denoted by a circle and cross, respectively. It can be seen that the probability of detection for the optimized algorithm increases while the probability of false alarm is reduced.

  • View in gallery

    Plot of radar images of WTC and (top) weather and (bottom) corresponding detections (yellow) from the fuzzy logic algorithm. The detection algorithm is able to effectively identify WTC contamination within the wind farm, outlined in white, with very few false alarms outside of it. As the power of the weather signal is increased, there are fewer resolution volumes labeled as contaminated, indicating a diminished effect of WTC contamination for weather.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 240 240 19
PDF Downloads 216 216 13

Automatic Detection of Wind Turbine Clutter for Weather Radars

View More View Less
  • 1 School of Electrical and Computer Engineering and Atmospheric Radar Research Center, University of Oklahoma, Norman, Oklahoma
  • | 2 Atmospheric Radar Research Center and Cooperative Institute for Mesoscale Meteorological Studies, University of Oklahoma, and NOAA/OAR/National Severe Storms Laboratory, Norman, Oklahoma
  • | 3 Atmospheric Radar Research Center and School of Meteorology, University of Oklahoma, Norman, Oklahoma
© Get Permissions
Full access

Abstract

Wind turbines cause contamination of weather radar signals that is often detrimental and difficult to distinguish from cloud returns. Because the turbines are always at the same location, it would seem simple to identify where wind turbine clutter (WTC) contaminates the weather radar data. However, under certain atmospheric conditions, anomalous propagation of the radar beam can occur such that WTC corrupts weather data on constantly evolving locations, or WTC can be relatively weak such that contamination on predetermined locations does not occur. Because of the deficiency of using turbine locations as a proxy for WTC, an effective detection algorithm is proposed to perform automatic flagging of contaminated weather radar data, which can then be censored or filtered. Thus, harmful effects can be reduced that may propagate to automatic algorithms or may hamper the forecaster’s ability to issue timely warnings. In this work, temporal and spectral features related to WTC signatures are combined in a fuzzy logic algorithm to classify the radar return as being contaminated by WTC or not. The performance of the algorithm is quantified using simulations and the algorithm is applied to a real data case from the radar facility in Dodge City, Kansas (KDDC). The results illustrate that WTC contamination can be detected automatically, thereby improving the quality control of weather radar data.

Corresponding author address: Sebastián Torres, National Weather Center, 120 David L. Boren Blvd., Norman, OK 73072. Email: sebastian.torres@noaa.gov

Abstract

Wind turbines cause contamination of weather radar signals that is often detrimental and difficult to distinguish from cloud returns. Because the turbines are always at the same location, it would seem simple to identify where wind turbine clutter (WTC) contaminates the weather radar data. However, under certain atmospheric conditions, anomalous propagation of the radar beam can occur such that WTC corrupts weather data on constantly evolving locations, or WTC can be relatively weak such that contamination on predetermined locations does not occur. Because of the deficiency of using turbine locations as a proxy for WTC, an effective detection algorithm is proposed to perform automatic flagging of contaminated weather radar data, which can then be censored or filtered. Thus, harmful effects can be reduced that may propagate to automatic algorithms or may hamper the forecaster’s ability to issue timely warnings. In this work, temporal and spectral features related to WTC signatures are combined in a fuzzy logic algorithm to classify the radar return as being contaminated by WTC or not. The performance of the algorithm is quantified using simulations and the algorithm is applied to a real data case from the radar facility in Dodge City, Kansas (KDDC). The results illustrate that WTC contamination can be detected automatically, thereby improving the quality control of weather radar data.

Corresponding author address: Sebastián Torres, National Weather Center, 120 David L. Boren Blvd., Norman, OK 73072. Email: sebastian.torres@noaa.gov

1. Introduction

There is a concerted effort to expand wind energy such that 20% of the nation’s energy will be generated through wind power by 2030 (Department of Energy 2008). While there are many positive outcomes from this effort, the negative impacts of this expansion cannot be ignored. One such impact of the explosive development of the wind power industry is the interference due to wind turbines on radar systems used to monitor airspace and meteorological phenomena.

Wind turbines are large structures, typically consisting of four main components: the tower, the nacelle, the rotor, and the blades. The large size of these structures, combined with the rotation of the blade cause interference that appears similar to typical radar signatures seen by weather radar systems. (Isom et al. 2009). This interference is commonly known as wind turbine clutter (WTC), and the resulting effects are not specific to any particular radar system (Department of Defense 2006). Returns caused by wind turbines look similar to weather signals and are sometimes difficult to distinguish on a typical plan position indicator (PPI) plot. Further, this contamination biases the estimates of the weather signal parameters commonly used to observe weather phenomena. In turn, biased weather data carry through to other algorithms such as quantitative precipitation estimation (QPE), causing further problems to forecasters (Vogt et al. 2009). False tornado or mesocyclone detections have also been reported, making forecasting even more difficult (Burgess et al. 2008).

Wind turbines are similar to other ground clutter targets as they do not move from one location to another. However, ground clutter typically has near-zero Doppler velocity and narrow spectrum width, so an algorithm can be designed to mitigate this contamination. Conventional clutter filters such as the Gaussian Model Adaptive Processing (GMAP; Siggia and Passarelli 2005) scheme are an effective method in removing these clutter targets but are ineffective for WTC. The clutter filter removes the returns due to the stationary components of the wind turbine, such as the tower, but the moving blades contaminate the data because conventional ground clutter filters are not designed to remove spectral components away from zero Doppler velocity (Vogt et al. 2007).

Previous studies of the characteristics of WTC have been carried out to better understand their signatures and mitigate their effects. Greving and Malkomes (2006, 2008) analyzed the effects of wind turbines on weather radars based on their radar cross section (RCS) and recommended the use of signal-processing-based mitigation schemes. Other studies characterized the WTC Doppler spectral signature over time. Whereas Isom et al. (2009) showed distinct flashes, Gallardo et al. (2008) showed a more complex Doppler spectrum evolution as the blades swept through space. For air traffic control (ATC) radar, mitigation techniques have been proposed or are in development to compensate for WTC. Techniques in development include producing wind turbine blades with low RCS, developing track-inhibition algorithms (Perry and Biss 2007), improving the range resolution of the radar to provide more uncontaminated resolution volumes between wind turbines, and adding telemetry to wind turbines to get real-time information about turbine angles relative to the radar or blade phase (Cornwall 2007). For weather radar, a few mitigation techniques have been proposed. For example, one technique involves nonlinear filtering methods when in “spotlight” mode (Isom 2007) while another technique involves the interpolation of spectral moment data over the wind farm using surrounding noncontaminated data (Isom et al. 2009). This multiquadric interpolation scheme could be a temporary solution for removing the WTC contamination until a more sophisticated signal processing algorithm is developed. However, another problem exists even with an effective interpolation. WTC contamination does not always appear in every radar scan. In one scan, the volumes corresponding to the wind farm may be contaminated, but in a subsequent scan, a fraction of those volumes may not be contaminated. Data corresponding to the clean volumes will be flagged in the interpolation as they are usually contaminated and valid data could be lost. The algorithm also requires knowledge of which volumes are contaminated, and there is currently no automatic detection scheme available for identifying these volumes. As with many artifact mitigation techniques, a detection algorithm is a necessary first step for the effective removal of WTC.

A common statement made in reference to mitigating WTC for weather radar is the notion that the wind turbines are in a fixed location and a flag can be set easily to identify those volumes. Flagging all data as contaminated because of the existence of a wind turbine in a radar volume is too aggressive, as clean data can be recovered when the turbine blades are not moving, for example. Moreover, under certain weather conditions, it has been observed that the weather signal overpowers the WTC signal such that the weather data are not contaminated (S. B. Cocks 2009, personal communication). Furthermore, anomalous propagation (AP) is probably the most important reason why an automatic detection algorithm is necessary. Under specific atmospheric conditions, superrefraction of the radar beam can occur, causing a wind farm that normally is not in the line of sight (LOS) of the radar to come into view.

In this paper, the temporal and spectral features of a canonical WTC signal are proposed to distinguish between contaminated and noncontaminated weather signals. These features based on raw radar data are combined in a fuzzy-logic-based algorithm to identify WTC contamination. To assess the performance of this algorithm, a new method of determining the ground truth to classify data correctly is introduced. Simulations are used to perform a statistical analysis of the distinguishing features; these are the foundation for the membership functions used in the fuzzy-logic detection algorithm. However, it is shown that the performance of the algorithm can be further improved through an optimization procedure based on an ad hoc cost function. Finally, an example application of the algorithm to real data is given with a brief discussion of the results.

2. Distinguishing features of wind turbine clutter

Observation of the spectral evolution by Isom et al. (2009) reveals several characteristics of WTC that can contaminate a weather signal. The tower of the wind turbine is a stationary ground clutter signal and has a strong power return at zero Doppler velocity, as seen in Fig. 1. This return has been well studied and can be mitigated using standard ground clutter filters (GCFs). Two other features are apparent in the Doppler spectral evolution. A periodic flash in the spectrum occurs and contaminates most of the spectrum (herein this is referred to as flash contamination). This flash across the spectrum was hypothesized by Isom et al. (2009) to be the reflected return from the rotating wind turbine blades when in the vertical position (up or down). The elevation in power contaminates the weather signal, and because the flash signal has nonzero velocities, it cannot be effectively filtered out using standard GCFs. In addition to the flash, there is an oscillation of elevated power near zero velocity in the Doppler spectrum. This so-called hub contamination is believed to be caused by the slowly moving turbine structures near the hub, which could have much higher RCSs due to their size and/or construction materials. The hub contamination is wide enough such that it cannot be completely removed with typical GCF notch widths.

Typical WTC contamination on a radar PPI are characterized by high reflectivity, random radial velocities, and very wide or very narrow spectrum widths depending on the GCF control. Hence, in order to detect WTC on a gate-by-gate basis, the spectral moments cannot be used because there is no unique signature that correlates well with WTC contamination. Instead, it is more effective to implement a detection algorithm at the signal processing level using the temporal and spectral features of the time series data. Many temporal and spectral features were considered for the detection algorithm but only the four useful features are discussed next.

Knowledge of the existence of a stationary tower can aid in detecting WTC, as a wind turbine cannot exist without a tower. The phase of the signal returned by a stationary ground clutter target should be close to constant whereas a varying signal, such as that corresponding to a weather return, will have a variable phase depending on the radial velocity and dispersion of velocities of the hydrometeors in the resolution volume. Clutter phase alignment (CPA) is a temporal feature used in the clutter mitigation decision (CMD) algorithm to detect ground clutter contamination (Hubbert et al. 2009) and is a measure of the variability of phase in the complex samples of signals from a particular radar volume.

Spectral flatness (σS−1) is a feature used to measure the flatness of a spectrum and has been used to enhance tornado detection (Yu et al. 2007). It is defined as
i1520-0426-27-11-1868-e1
where Ŝ′( f ) is the set of spectral coefficients with the weakest 5% of the components removed and std is the standard deviation. Flash contamination from WTC can cause aliasing, making the spectrum to appear flat, and this feature is used to identify when this contamination occurs.

The fourth central spectral moment (μ4) is a higher-order moment used to measure the dispersion of velocities in the Doppler spectrum with respect to the mean and is related to the second moment for a Gaussian spectrum (Papoulis and Pillai 2002). For a non-Gaussian spectrum, which occurs when weather signals are contaminated by a flash, μ4 has a high value when compared to a Gaussian spectrum. Hence, this feature can also be used to identify flash contamination.

The last feature, the hub-to-weather ratio (HWR), is used to detect hub contamination. HWR was developed specifically to detect this type of contamination after clutter filtering. HWR is computed by first determining the power of the “hub” contamination residue as
i1520-0426-27-11-1868-e2
where Ŝ is the ground-clutter-filtered Doppler spectrum and is the index of the maximum of the two spectral points closest to the edge of the GCF notch. The “weather” power is estimated as
i1520-0426-27-11-1868-e3
where is the index of the maximum spectral component value. The HWR is computed by taking the ratio of the two estimates and converting to logarithmic units as
i1520-0426-27-11-1868-e4

If the hub power is greater than the weather signal power, the maximum of the spectrum, Ŝ(), is equal to Ŝ(), resulting in an HWR value of 0 dB; for low clutter-to-signal ratios (CSRs), the ratio will be less than 0 dB. A Doppler spectrum with a mean velocity of 18 m s−1 that has hub contamination is shown in Fig. 2. The hub power residue is computed using five points centered at the peak at approximately 2 m s−1, while the weather power is computed using five points centered at the maximum of the filtered spectrum, Ŝ(). For this example spectrum, the weather power is an estimation of the actual weather signal. This feature aids in detecting hub contamination but it has limitations. One such limitation is that the hub power residue is not a true estimate of the power residue from the hub. Two of the spectral components in the calculation of the power are artificial, as they are rebuilt by the spectrum-based clutter-filtering algorithm after removing the near-0 Doppler spectral components. The second limitation of this feature occurs if the weather power is significantly greater than the hub power residue and the mean velocity of the weather signal is near 0 m s−1. If the weather signal has a mean velocity near 0 m s−1, the clutter filter removes a portion of the weather signal. When this occurs, Ŝ() ≈ Ŝ() and a HWR value near 0 dB will be computed incorrectly, signifying hub contamination.

3. Quantifying wind turbine clutter contamination: The delta bias

Performance quantification of any detection algorithm usually requires some knowledge of the truth. To establish the contamination truth for this case, the bias of each spectral moment estimate was computed as
i1520-0426-27-11-1868-e5
where θ̂ is the estimate of a parameter and θ is the true value. Several factors can contribute to the bias of a weather signal estimate. Some estimators are inherently biased and their bias, although small, cannot be ignored. Ground clutter returns also introduce biases into the moment estimates. A GCF such as GMAP minimizes the spectral moment biases through an iterative interpolation technique but, even then, it cannot remove ground clutter without introducing some small bias into the spectral moment estimates. Finally, WTC contamination due to a flash or hub is another source of bias. These biases are especially important in the absence of a mitigation scheme.
Some knowledge of the bias due to only WTC is necessary to independently determine which signals are contaminated by WTC. When estimating the weather signal parameters contaminated by WTC, the estimates are biased because of the estimators, the effects of the GCF, and WTC contamination. To extract the bias due to only WTC, the concept of “delta bias” was created, which is defined as
i1520-0426-27-11-1868-e6
where BiasTotal is the bias due to WTC and signal processing (e.g., moment estimation and GCF) and BiasSP is the bias due to signal processing. Delta bias is a measure of the bias due to WTC only and will be used to objectively distinguish between contaminated data and noncontaminated data. An example illustration of the delta bias for power for a range of radial velocities is given in Fig. 3. The bias due to the GCF can be seen in the Fig. 3 as a dashed–dotted line. In the WTC power bias plot, the bias due to WTC contamination and clutter filtering can be seen as a dashed line. The solid line represents the delta bias, which quantifies the contamination of the weather signal from WTC. Using a threshold on the delta bias, a determination of the truth can be made.

4. Statistical analysis of distinguishing features

In this section, simulations are used to assess the performance of the detection algorithm. To simulate WTC, a simple model was established for the wind turbine using the previously discussed characteristics observed in the spectral evolution, as seen in Fig. 1. The wind turbine blade signal was simulated using a time series weather radar simulator developed by Cheong et al. (2008). The blades were modeled with a uniform set of scatterers, each with a specific backscattering cross-section value such that the reflected signal resembled a turbine blade. The hub echo was simulated as a random process with a varying mean around zero Doppler velocity, which mimics actual data quite well. The tower signal was simulated using a standard weather signal simulator with zero velocity and a model spectrum width of 0.3 m s−1. The three signals were coherently added to simulate the WTC baseband in-phase and quadrature signals. One cycle of the simulated spectral evolution was used in the experiment to limit the amount of data generated. This cycle contains 12 spectra with hub contamination and 2 spectra with flash contamination and is a fair representation of the entire spectral evolution. Ground clutter and clear-air spectra (i.e., no hydrometeors present in the radar volume) were included in the set of simulated data to analyze the features on signals typically observed by radar for a total of 16 simulated spectra.

To properly analyze the distinguishing features, simulated weather signals were combined with the 16 simulated spectra. For this experiment, weather signals were simulated using the method developed by Zrnić (1975). This allows signal power, mean radial velocity, and spectrum width to be varied systematically such that the contamination effects of WTC could be studied for a wide variety of weather signals. It should be noted that the interaction between WTC and weather was not considered in the simulated data. The range of weather signal parameter values is given in Table 1. Spectra of weather signals containing ground clutter were used as a reference for BiasSP, as the spectra were free of any WTC effects. In other words, these cases had a delta bias of zero. Likewise, the weather-only cases were free from any WTC effects and were forced to have a delta bias of zero since the GCF effects were absent from this dataset. We computed BiasTotal by using the 14 WTC signals in the spectral evolution cycle combined with a variety of simulated weather signals.

The delta-bias values were computed over four dimensions: WTC spectrum set, CSR, weather mean radial velocity, and weather spectrum width. To represent the delta bias in a manageable manner, the dimensionality of the dataset was reduced. The delta-bias values over all simulated velocities and spectrum widths were aggregated together for each spectrum set and CSR because the values did not vary considerably from one set of simulation parameters to the next. Contamination was defined to occur where the delta bias exceeded a given threshold. Classifications for specific CSR values and WTC spectra were selected such that 99% of all the delta-bias values were lower than the selected threshold. It was observed that the radial velocity and spectrum width were more susceptible to contamination at lower CSR values than signal power. The delta bias for radial velocity with a threshold of 1 m s−1 was selected to determine the ground truth because the spectrum width estimator has several limitations (Meymaris and Williams 2007). The two-dimensional plot in Fig. 4 shows the delta-bias values for a specific CSR and spectrum set of the WTC evolution for radial velocity as well as the ground truth classification based on the delta bias threshold (shown in white).

Using delta bias to determine the contamination truth, the feature values were divided into the following six signal classes: contaminated spectra containing a flash or hub signal, noncontaminated spectra with a flash or hub signal, and noncontaminated spectra with ground clutter or weather signal only. Histograms were plotted to determine how well the four features could distinguish between contaminated and noncontaminated data. An ideal feature would have histograms with complete separation between contaminated and noncontaminated classes. The histograms for each feature are shown in Fig. 5, from which it is evident that no feature alone can effectively identify WTC contamination. Although each feature itself may not be useful, they can be combined in a fuzzy logic system (FLS) to detect weather signals contaminated by WTC.

5. Detection algorithm

Fuzzy logic is effective when there is no distinct partitioning of the inputs of a classification algorithm as is the case for WTC detection. Moreover, fuzzy logic is relatively simple to implement and is currently in use on operational weather radars to classify hydrometeors (Liu and Chandrasekar 2000; Zrnić et al. 2001) and to identify ground clutter (Hubbert et al. 2009), so it was a natural choice for detecting WTC contamination.

The inputs to the FLS WTC detection algorithm were the features discussed previously: CPA, σS−1, μ4, and HWR. Each input was “fuzzified” using a membership function to determine a degree of contamination based on that feature only. Fuzzified values were weighted according to how effective the feature was in identifying WTC contamination. Finally, the weighted values were aggregated and compared against a threshold to determine whether data were contaminated by WTC. A diagram of the FLS for WTC detection is shown in Fig. 6.

To assess the performance of the algorithm, receiver operating characteristics (ROCs) were used. ROC analysis has its origin in statistical decision theory developed during World War II for the analysis of radar images (Green and Swets 1966) and has recently been applied in a similar context to evaluate the performance of a ground clutter detection algorithm for weather radars (Hubbert et al. 2009). The ROC curves show the fraction of true identified WTC contamination (i.e., the probability of detection, Pd) versus the fraction of incorrectly identified WTC contamination (i.e., the probability of false alarm, Pfa) as a function of the FLS threshold.

a. Data scaling and the 2030 scenario

In a realistic radar environment, most of the resolution volumes do not contain WTC contamination but instead will be dominated by clear air, weather echoes, or weather with ground clutter echoes. Without any scaling of the simulated data distribution, the results of the algorithm would have been heavily influenced by the WTC-contaminated spectra. To adjust the distribution of hub and flash contaminations and the weather and weather with ground clutter cases to reflect a realistic distribution while stressing the importance of detecting WTC, a worst-case scenario was created. Several assumptions were made for the so-called 2030 scenario; these are discussed next.

The total land area of the United States is approximately 9.83 million km2 and it is assumed from wind resource maps that approximately 10% of the contiguous United States (CONUS) is suitable for wind energy. Offshore wind turbines were not considered for this scenario as the complexity of mixed contamination from WTC and sea clutter is beyond the scope of this work and deferred to future research. The land area required to construct enough wind turbines to fulfill the goal of 20% wind power by 2030 is 50 000 km2 (Department of Energy 2008). Using these land estimates, it can be shown that approximately 5% of the CONUS land area is usable for wind power and will be contaminated by WTC; that is,
i1520-0426-27-11-1868-e7
For the radar environment at the lowest antenna elevation angle used by the Weather Surveillance Radars-1988 Doppler (WSR-88Ds) of the Next Generation Weather Radar (NEXRAD) network, it was assumed that only resolution volumes out to a range of 150 km could possibly be contaminated by WTC. Under normal propagation conditions, wind farms cannot be seen at that distance but the assumption takes into account possible AP conditions. At ranges greater than 150 km, the height of the beam would be high enough such that contamination would be unlikely. Of the resolution volumes within 150 km of the radar, it was assumed that the volumes contain hydrometeors with or without ground clutter. Under the worst possible AP conditions, it was assumed that all of the resolution volumes within 30 km of the radar would be contaminated by ground clutter with the likelihood of contamination decreasing with range out to 150 km. An ad hoc function was created to determine the worst-case scenario fraction of the lowest elevation data that were contaminated by ground clutter. The function is defined as
i1520-0426-27-11-1868-e8
Using (8), the fraction of ground clutter contamination for the radar data at the lowest antenna elevation angle within a range of 150 km can be calculated as
i1520-0426-27-11-1868-e9
Within the set of radar volumes with ground clutter, some correspond to WTC and it was shown previously that WTC occupied 5% of the area within 150 km. It was assumed that WTC contamination could be equally likely due to the hub or flash returns. The distribution of the types of spectra for the data before weighting and for the 2030 scenario is shown in Table 2. Using the percentages in Table 2, the proper scaling of the statistical results obtained from the simulated data can be performed. To convert from the unweighted data to the 2030 scenario, one needs only to multiply the unweighted value by the ratio between the 2030 scenario and unweighted data. For example, for the weather-only case, proper scaling to the 2030 scenario requires multiplying by 9.44. However, a restriction was placed on the weighting such that the sum of the weights was equal to one (see Table 2 for a complete list of weights). Using these weights, a more representative Pd and Pfa can be computed for the 2030 scenario. The weighted probability of detection and probability of false alarm are calculated as
i1520-0426-27-11-1868-e10
i1520-0426-27-11-1868-e11
where TP is the number of correct detections, FP the number of false detections, C the total number of contaminated datasets, and NC the total number of noncontaminated datasets. The subscripts F, H, wx, and gc denote the flash, hub, weather, and ground clutter cases, respectively. After the cost function and proper scaling were established, the optimization of the algorithm parameters was performed to improve the performance of the FLS.

b. Algorithm optimization

The initial selection of the membership functions was done by visual inspection of the feature histograms discussed in section 4 and feature weights were selected to be uniform. Hence, the performance of the FLS could be improved by optimizing these parameters through the use of a cost function. The cost function was selected to emphasize the importance of detecting WTC contamination while minimizing the false alarm rate and is defined as
i1520-0426-27-11-1868-e12
where β is the desired minimum probability of detection. When the number of missed detections was less than the number of false alarms for the minimized probability of false alarm, J was minimized. Note that this cost function is analogous to the well-known Neyman–Pearson test (Skolnik 1962); however, we have chosen a criterion that ensures a desired level of detection performance while minimizing the number of false alarms. This cost function allows for optimization of all the FLS parameters to maximize the performance of the WTC contamination detection algorithm through an exhaustive search.

Each membership function was varied by shifting the function to the left or right by 5% of the largest value for each feature. The maximum values for CPA, σS−1, μ4, and HWR were 1.0, 0.5, 500 000, and 100, respectively. The slope for each function was fixed whenever possible but two exceptions to this restriction were made for μ4 and HWR. Moving the membership function left by 5% for μ4 would have resulted in a membership function with a degree of membership of 1.0 for a μ4 value of 0. To avoid this unrealistic membership function for both features, the slope was modified such that one inflection point would be adjusted by 5% while the other would stay constant to provide a more realistic membership function. The weights were varied so every combination would be possible without duplication of relative weights. The weighting was also restricted such that the sum of the weights had a maximum value of 1.0. To equally weight any combination of two or more features, an increment of was used. The last parameter that was optimized in the FLS was the final threshold used to determine whether WTC contamination exists. To minimize the number of threshold values for the optimization, the range of values was set as 0.15–0.50 in increments of 0.05. It was determined through an analysis on a large sample set of data that any threshold value above 0.50 would likely have a low probability of detection and any threshold less than 0.15 would result in an unacceptable probability of false alarm.

The optimized membership functions as well as the original membership functions are shown in Fig. 7. The membership function for CPA was shifted left by 20% of the maximum feature value and for σS−1, the membership function was shifted left by 5%. A comparison of the optimized and original membership functions for the two features reveals that the initial membership functions were too conservative. The membership functions were originally designed to minimize false alarms while ensuring that detections were correct but were not aggressive enough for the algorithm. In contrast, the original membership functions for μ4 and HWR were too aggressive. Both membership functions were optimized to minimize the number of false alarms at the expense of a reduced number of detections. The optimized feature weights of the algorithm for CPA, σS−1, μ4, and HWR were , respectively. It is somewhat surprising that CPA has such a small weight as it was thought to discriminate well between contaminated and noncontaminated data. However, the weather with the ground clutter case, which is not contaminated by WTC, has high CPA values similar to contaminated spectra. To minimize this high false alarm rate, the weighting for CPA was reduced through the optimization process. However, CPA did not get a 0 weight, which was a valid option, signifying that the feature does add to the effectiveness of the detection algorithm. The low weighting for HWR was not as surprising as the weighting for CPA since a high probability of false alarm was a possible limitation of the feature when a slow-moving weather signal overpowered the WTC signal. Through the optimization process, the final threshold for the FLS was found to be 0.25.

The ROC plots for the algorithm before and after optimization of the algorithm parameters are shown in Fig. 8, where improvement in the performance of the algorithm is easily visualized. After the optimization (shown as a solid line), the curve shifts to the left and up toward the corner, indicating an improvement in performance. The probability of detection and probability of false alarm before optimization were 86.86% and 1.84% and after the optimization were 89.03% and 0.59%, respectively. The optimization process improved the detection rate by about 2% and also decreased the false alarm rate by a factor of 3.

It has been shown that the detection algorithm is able to successfully identify contaminated data 89.03% of the time while misclassifying data only 0.59% of the time for simulated weather and WTC signals. While the performance of the algorithm is acceptable for simulated signals, it is important to show the same for real signals. An example application of the automatic detection algorithm on real signals is given next.

6. Example application of the detection algorithm

WTC signals for this example application were collected with the Dodge City, Kansas (KDDC), radar by Isom et al. (2009) on 30 March 2006 using volume coverage pattern 21 (VCP 21) (National Oceanic and Atmospheric Administration 2006). Although time series data for weather and WTC were recorded, the collected data did not contain any cases of weather and WTC in the same resolution volume. To properly test the algorithm on cases other than the clear-air case, test sets of data created by Isom et al. (2009) were used in this example, as is explained next.

a. Clear-air case

Initially, the algorithm was applied to WTC signals in clear air where the algorithm identified which volumes were contaminated. Contamination flags were censored on volumes where the signal-to-noise ratio (SNR) was less than 10 dB. Hence, for this real application, SNR became another parameter in the algorithm used as a binary decision to minimize the number of false alarms. The preliminary result of the detection algorithm applied to the clear-air case of WTC is shown in Fig. 9.

An outline of the wind farm location used by Isom et al. (2009) (shown in white in Fig. 9) was used to easily identify the volumes that have the potential to be labeled as contaminated. It can be seen from the detection plot that the algorithm is able to successfully identify WTC contamination within the area of the wind farm while only a few false alarms exist outside this region. This is promising as it demonstrates that contamination can be successfully identified when WTC returns appear in clear-air conditions.

b. Weather cases

Two real datasets were combined to test the algorithm with a mixture of WTC and weather signals. The WTC signals used for the clear-air case were combined with a separate set of weather signals that were recorded at a similar range from the radar using the same radar parameters. Three weather cases were “simulated” by using different relative scalings to observe the performance of the algorithm at different levels of weather-to-WTC power ratios. The first case involved decreasing the weather signal powers by 10 dB from the nominal value to observe the performance of the algorithm for weak weather data. The second case used weather signals as recorded by the radar (no scaling). For the third case, the weather signal powers were increased by 10 dB. As was done for the clear-air case, SNR censoring was included to minimize the number of false alarms. The censoring was the same for all three cases and was based on the weather data with no scaling.

The results of the algorithm for these three weather cases are also shown in Fig. 9. As expected, the detection plots show that the contamination effects of WTC are lessened as the weather signals become stronger. For the case where the weather signal powers are reduced by 10 dB, all of the volumes labeled as contaminated are also labeled as contaminated for the clear-air case. However, there are a few volumes for this case that are not labeled as contaminated that were labeled as such for the clear-air case, indicating the contamination effect of WTC is reduced when the return signal is a combination of weather returns and WTC. As the weather power is increased for the other two weather cases, the lessened effect of WTC can be observed by the reduced number of detections within the boundary of the wind farm.

It is encouraging to see that the detection algorithm is able to identify the presence of WTC contamination with few false alarms. One possible solution to reducing the number of false alarms in both cases is to use the location of the wind turbines as an input to the FLS. These locations do not change, but possible AP conditions should be taken into account. Thus, a broader area around each wind farm could be identified as having a high likelihood of WTC contamination. This additional input could be added into the algorithm using a map (similar to a clutter map) specific to each radar without too much effort to lower the occurrence of false alarms. However, the feasibility of this approach depends on having precise and up-to-date location information for every wind farm in the LOS of every radar.

7. Conclusions and future work

The objective of this work was to develop an automatic algorithm to identify weather radar signals that are contaminated by wind turbine clutter (WTC). Previous studies of the characterization of WTC were used as a starting point to understand the contamination of weather radar signals. Using the observed characteristics, temporal and spectral features were developed and tested to detect the contamination due to the flash and hub returns. The features used in the detection algorithm were clutter phase alignment (CPA), spectral flatness (σS−1), fourth central spectral moment (μ4), and hub-to-weather ratio (HWR). To determine the ground truth, which provides objective data classification into contaminated and not contaminated categories, the concept of delta bias was used, which without the accuracy of the detection algorithm could not be measured. The delta bias of the velocity was used to divide the feature values into separate classes and the histograms for each feature were the basis for the initial membership functions in the fuzzy logic detection algorithm. An optimization of all the fuzzy logic parameters was performed using an exhaustive search and the optimal algorithm parameters were determined by a cost function based on statistics for a worst-case “2030 scenario,” which maximized the probability of detection while minimizing the probability of false alarm for the contaminated weather signal.

An example application of the detection algorithm on real data was performed for a clear-air case and for three weather cases with varied weather signal power. For the clear-air case, it was shown that the algorithm was able to identify WTC contamination within the wind farm with relatively good accuracy with a minimal number of false alarms outside of it. When the algorithm was applied to a mixed return of WTC and weather signals, it was able to detect contamination within the boundary of the wind farm without an increase in the number of false alarms. The diminishing effect of WTC contamination was also observed as the power of the weather signal increased. This is an important point as the existence of a wind turbine in a resolution volume does not necessarily translate into contamination of weather signal parameter estimates.

A limitation to this work was the focus on the simulation of a single wind turbine in a resolution volume. It is possible that the radar may observe more than one wind turbine in a resolution volume and this scenario should be considered in future work. However, with the increasing size of individual wind turbines, this limitation may become less of an issue as the spacing of wind turbines will also increase to accommodate the larger structures. With greater spacing between individual turbines, placement of wind turbines will eventually become such that only one turbine is located within a typical weather radar resolution volume.

Further research can be performed to improve the algorithm so it can be implemented in an operational setting. WTC data from an operational radar can be used to assess the sensitivity of the algorithm with a smaller number of samples. The computational complexity of the features in the algorithm should also be taken into consideration when implementing a real-time algorithm. In addition to assessing the algorithm for a realistic set of radar parameters, estimates of the probability of detection and false alarm rate using recorded data should be computed to validate the algorithm performance from the simulated data.

With the ongoing upgrade of the WSR-88D to include dual-polarization capability, the polarimetric variables should be analyzed to determine their effectiveness in identifying WTC contamination. Any extra information provided by additional features can only improve the performance of the algorithm.

With the explosive growth of the wind power industry, concerns have grown with regard to the negative impacts of wind turbines on weather radar, and the problems will only escalate over time with the continued growth of this industry. This research has laid a foundation for the detection of weather signals contaminated by WTC through the use of real and simulated data. The real-time detection of weather signals contaminated by WTC will greatly aid users of weather radar data. Forecasters issuing warnings will benefit from the knowledge that the data available to them are free from contamination. Also, automatic algorithms will benefit by using only clean data, thus lowering the bias of the estimates and reducing the likelihood of false alarms. Finally, this algorithm is an important step toward the successful mitigation of the detrimental effects of WTC on weather radar, which will allow for the recovery of weather signal parameters from WTC-contaminated data.

Acknowledgments

The authors thank the three anonymous reviewers for their useful suggestions and the Radar Operations Center of the National Weather Service for their work in studying the impacts of WTC and their assistance in carrying out experiments to collect WTC data. Additional thanks go to Khoi Le and Boon Leng Cheong for their work in simulating WTC; Brad Isom, who prepared and provided the Dodge City, Kansas, radar data; and Feng Nai for his comments regarding the analysis of the results. Funding was provided by the NOAA/Office of Oceanic and Atmospheric Research under NOAA–University of Oklahoma Cooperative Agreement NA17RJ1227, U.S. Department of Commerce.

REFERENCES

  • Burgess, D. W., , Crum T. D. , , and Vogt R. J. , 2008: Impacts of wind farms on WSR-88D radars. Preprints, 24th Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, New Orleans, LA, Amer. Meteor. Soc., 6B.3. [Available online at http://ams.confex.com/ams/pdfpapers/128810.pdf].

    • Search Google Scholar
    • Export Citation
  • Cheong, B. L., , Palmer R. D. , , and Xue M. , 2008: A time-series weather radar simulator based on high-resolution atmospheric model simulations. J. Atmos. Oceanic Technol., 25 , 230243.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cornwall, J. M., 2007: Mitigation of wind turbine effects on FAA and military radars. Adapted from Jason Rep. JSR-07-126, MITRE Corporation, 8 pp.

    • Search Google Scholar
    • Export Citation
  • Department of Defense, 2006: The effect of windmill farms on military readiness. Office of the Director of Defense Research and Engineering, 62 pp. [Available online at http://www1.eere.energy.gov/windandhydro/federalwindsiting/pdfs/dod_windfarms.pdf].

    • Search Google Scholar
    • Export Citation
  • Department of Energy, 2008: 20% wind energy by 2030. Energy Efficiency and Renewable Energy Rep. DOE/GO-102008-2567, 229 pp.

  • Gallardo, B., , Perez F. , , and Aguado F. , 2008: Characterization approach of wind turbine clutter in the Spanish Weather Radar Network. Proc. Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, European Meteorological Society, 9.8.

    • Search Google Scholar
    • Export Citation
  • Green, D. M., , and Swets J. A. , 1966: Signal Detection Theory and Psychophysics. 1st ed. Robert E. Krieger Publishing, 455 pp.

  • Greving, G., , and Malkomes M. , 2006: On the concept of the radar cross section RCS of distorting objects like wind turbines for the weather radar. Proc. Fourth European Conference on Radar in Meteorology and Hydrology, Barcelona, Spain, European Meteorological Society, P5.8.

    • Search Google Scholar
    • Export Citation
  • Greving, G., , and Malkomes M. , 2008: Weather radar and wind turbines—Theoretical and numerical analysis of the shadowing effects and mitigation concepts. Proc. Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, European Meteorological Society, P2.12.

    • Search Google Scholar
    • Export Citation
  • Hubbert, J. C., , Dixon M. , , and Ellis S. M. , 2009: Weather radar ground clutter. Part II: Real-time identification and filtering. J. Atmos. Oceanic Technol., 26 , 11811197.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Isom, B. M., 2007: Characterization and mitigation of wind turbine clutter for the WSR-88D network. M.S. thesis, School of Electrical and Computer Engineering, University of Oklahoma, 93 pp.

  • Isom, B. M., and Coauthors, 2009: Detailed observations of wind turbine clutter with scanning weather radars. J. Atmos. Oceanic Technol., 26 , 894910.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Liu, H., , and Chandrasekar V. , 2000: Classification of hydrometeors based on polarimetric radar measurements: Development of fuzzy logic and neuro-fuzzy systems, and in situ verification. J. Atmos. Oceanic Technol., 17 , 140164.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Meymaris, G., , and Williams J. K. , 2007: Spectrum width estimators for the NEXRAD ORDA: Evaluation and recommendation. Preprints, 23rd Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, San Antonio, TX, Amer. Meteor. Soc., 5B.5. [Available online at http://ams.confex.com/ams/pdfpapers/119628.pdf].

    • Search Google Scholar
    • Export Citation
  • National Oceanic and Atmospheric Administration, 2006: WSR-88D Products and Algorithms. Part C, Doppler Radar Meteorological Observations, Federal Meteorological Handbook 11, FCM-H11C-2006, Office of the Federal Coordinator for Meteorological Services and Supporting Research, 390 pp.

    • Search Google Scholar
    • Export Citation
  • Papoulis, A., , and Pillai S. U. , 2002: Probability, Random Variables, and Stochastic Processes. McGraw-Hill, 852 pp.

  • Perry, J., , and Biss A. , 2007: Wind farm clutter mitigation in air surveillance radar. IEEE Aerosp. Electron. Syst. Mag., 22 , 3540.

  • Siggia, A. D., , and Passarelli R. E. , 2005: Gaussian model adaptive processing (GMAP) for improved ground clutter cancellation and moment calculation. M S-NR:ERAD3-P-00117, Sigmet, Inc., 8 pp.

    • Search Google Scholar
    • Export Citation
  • Skolnik, M. I., 1962: Introduction to Radar Systems. 3rd ed. McGraw-Hill, 648 pp.

  • Vogt, R. J., , Reed J. , , Crum T. , , Snow J. T. , , Palmer R. D. , , Isom B. M. , , and Burgess D. W. , 2007: Impacts of wind farms on WSR-88D operations and policy considerations. Preprints, 23rd Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, San Antonio, TX, Amer. Meteor. Soc., 5B.7. [Available online at http://ams.confex.com/ams/pdfpapers/120352.pdf].

    • Search Google Scholar
    • Export Citation
  • Vogt, R. J., and Coauthors, 2009: Continued progress in assessing and mitigating wind farm impacts on WSR-88Ds. Preprints, 25th Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Phoenix, AZ, Amer. Meteor. Soc., 11B.6. [Available online at http://ams.confex.com/ams/pdfpapers/150646.pdf].

    • Search Google Scholar
    • Export Citation
  • Yu, T-Y., , Wang Y. , , Shapiro A. , , Yeary M. B. , , Zrnić D. S. , , and Doviak R. J. , 2007: Characterization of tornado spectral signatures using higher-order spectra. J. Atmos. Oceanic Technol., 24 , 19972013.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zrnić, D. S., 1975: Simulation of weatherlike Doppler spectra and signals. J. Appl. Meteor., 14 , 619620.

  • Zrnić, D. S., , Ryzhkov A. V. , , Straka J. , , Liu Y. , , and Vivekanandan J. , 2001: Testing a procedure for automatic classification of hydrometeor types. J. Atmos. Oceanic Technol., 18 , 892913.

    • Crossref
    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

Temporal evolution of WTC spectra. The large power return at 0 Doppler velocity is attributed to the tower of the wind turbine. The flashes, which appear periodically, are from the rotating blades of the turbine when in the vertical position, and the oscillation around 0 velocity is attributed to the slowly moving structures near the rotor hub of the wind turbine.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 2.
Fig. 2.

A hub-contaminated weather spectrum. The maximum value of the spectrum is used to compute the power of the “weather” signal. In this case, the power of the actual weather signal was computed, and when compared to the power of the hub contamination, it was determined that the weather signal was contaminated.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 3.
Fig. 3.

Plot of the delta bias. The bias due to clutter filtering can be seen near 0 Doppler velocity. The increase in the delta bias near 9 m s−1 indicates hub contamination.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 4.
Fig. 4.

Plots of (left) the temporal evolution of WTC spectra with (middle) corresponding delta biases for radial velocity and (right) truth data. Classifications for each CSR and WTC spectra were selected such that 99% of the delta-bias values were lower than the selected threshold of 1 m s−1 (shown in white). The truth data (right) illustrate what range of CSR values are considered contaminated for each WTC spectra.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 5.
Fig. 5.

Histograms for (top) (left) CPA and (right) σS−1; (bottom) (left) μ4 and (right) HWR. Each feature has a range of overlapping values for contaminated (solid) and noncontaminated (dashed) data, making identification of WTC difficult using a single feature. However, the four features can be combined in an FLS to effectively detect WTC contamination.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 6.
Fig. 6.

Diagram of the fuzzy logic system for WTC detection. Inputs to the FLS are the detection features. Each feature is fuzzified using a membership function. The fuzzy feature values are then weighted and aggregated before being compared against a threshold.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 7.
Fig. 7.

Original (dashed) and optimized (solid) membership functions for the FLS: (top) (left) CPA and (right) σS−1; (bottom) (left) μ4 and (right) HWR. Membership functions for CPA and σS−1 were more aggressive in the detection algorithm after the optimization while the membership functions for μ4 and HWR were too aggressive before optimization.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 8.
Fig. 8.

ROC plot before (dashed) and after (solid) optimization. FLS threshold values before optimization range from 0.0 to 1.0 in increments of 0.05 and correspond to the points beginning at the upper-right corner and follow the dashed line to the bottom-left corner. Threshold values after optimization range from 0.15 to 0.5 in increments of 0.05 and correspond to the points on the solid line from right to left. The optimal thresholds before and after optimization are denoted by a circle and cross, respectively. It can be seen that the probability of detection for the optimized algorithm increases while the probability of false alarm is reduced.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Fig. 9.
Fig. 9.

Plot of radar images of WTC and (top) weather and (bottom) corresponding detections (yellow) from the fuzzy logic algorithm. The detection algorithm is able to effectively identify WTC contamination within the wind farm, outlined in white, with very few false alarms outside of it. As the power of the weather signal is increased, there are fewer resolution volumes labeled as contaminated, indicating a diminished effect of WTC contamination for weather.

Citation: Journal of Atmospheric and Oceanic Technology 27, 11; 10.1175/2010JTECHA1437.1

Table 1.

Range of weather simulation parameters.

Table 1.
Table 2.

Distribution of data before and after weighting with scaling weights.

Table 2.
Save