Wind turbine blade leading edge erosion is a major source of power production loss and early detection benefits optimization of repair strategies. Two machine learning (ML) models are developed and ...evaluated for automated quantification of the areal extent, morphology and nature (deep, shallow) of damage from field images. The supervised ML model employs convolutional neural networks (CNN) and learns features (specific types of damage) present in an annotated set of training images. The unsupervised approach aggregates pixel intensity thresholding with calculation of pixel-by-pixel shadow ratio (PTS) to independently identify features within images. The models are developed and tested using a dataset of 140 field images. The images sample across a range of blade orientation, aspect ratio, lighting and resolution. Each model (CNN v PTS) is applied to quantify the percent area of the visible blade that is damaged and classifies the damage into deep or shallow using only the images as input. Both models successfully identify approximately 65% of total damage area in the independent images, and both perform better at quantifying deep damage. The CNN is more successful at identifying shallow damage and exhibits better performance when applied to the images after they are preprocessed to a common blade orientation.
Two years of high-resolution simulations conducted with the Weather Research and Forecasting (WRF) model are used to characterize the frequency, intensity and height of low-level jets (LLJ) over the ...U.S. Atlantic coastal zone. Meteorological conditions and the occurrence and characteristics of LLJs are described for (i) the centroids of thirteen of the sixteen active offshore wind energy lease areas off the U.S. east coast and (ii) along two transects extending east from the U.S. coastline across the northern lease areas (LA). Flow close to the nominal hub-height of wind turbines is predominantly northwesterly and southwesterly and exhibits pronounced seasonality, with highest wind speeds in November, and lowest wind speeds in June. LLJs diagnosed using vertical profiles of modeled wind speeds from approximately 20 to 530 m above sea level exhibit highest frequency in LA south of Massachusetts, where LLJs are identified in up to 12% of hours in June. LLJs are considerably less frequent further south along the U.S. east coast and outside of the summer season. LLJs frequently occur at heights that intersect the wind turbine rotor plane, and at wind speeds within typical wind turbine operating ranges. LLJs are most frequent, intense and have lowest core heights under strong horizontal temperature gradients and lower planetary boundary layer heights.
Continued growth of wind turbine physical dimensions is examined in terms of the implications for wind speed, power and shear across the rotor plane. High-resolution simulations with the Weather ...Research and Forecasting model are used to generate statistics of wind speed profiles for scenarios of current and future wind turbines. The nine-month simulations, focused on the eastern Central Plains, show that the power scales broadly as expected with the increase in rotor diameter (D) and wind speeds at hub-height (H). Increasing wind turbine dimensions from current values (approximately H = 100 m, D = 100 m) to those of the new International Energy Agency reference wind turbine (H = 150 m, D = 240 m), the power across the rotor plane increases 7.1 times. The mean domain-wide wind shear exponent (α) decreases from 0.21 (H = 100 m, D = 100 m) to 0.19 for the largest wind turbine scenario considered (H = 168 m, D = 248 m) and the frequency of extreme positive shear (α > 0.2) declines from 48% to 38% of 10-min periods. Thus, deployment of larger wind turbines potentially yields considerable net benefits for both the wind resource and reductions in fatigue loading related to vertical shear.
Cold-season windstorms represent an important, and potentially changing, geophysical hazard in the Northeastern United States. Here we employ an integrated research methodology including both a ...storyline approach, where three intense windstorms from the current climate are subjected to pseudo-global warming (PGW) experiments, and a long-term transient simulation using the Weather Research and Forecasting (WRF) model. An ensemble of WRF simulations is built for each windstorm using different planetary boundary layer and microphysical parameterizations. The fidelity assessment suggests all ensemble members capture the windstorm evolution in contemporary climate. The configuration with highest fidelity is used in the PGW experiments performed with perturbed temperature fields, constant relative humidity, and deiced Great Lakes. These perturbation simulations indicate some evidence for a reduction of sea level pressure and increases in wind speed over and downwind of the Great Lakes and over the Atlantic Ocean plus an increase in precipitation accumulation but a reduction in snow coverage. These changes are spatially inhomogeneous in terms of magnitude and sign but are consistent with changes in potential vorticity. Alberta Clippers and Colorado Lows dominate the cyclones responsible for historical windstorms and thus are sampled in the PGW simulations. However, the transient simulation suggests an increasing role for tropical cyclones that undergo transition to extratropical cyclones. This reinforces the value of combining information from both PGW perturbation experiments within a storyline context and transient simulations when seeking to quantify the future risk associated with cold-season windstorms under changing climate.
Leading edge erosion (LEE) of wind turbine blades causes decreased aerodynamic performance leading to lower power production and revenue and increased operations and maintenance costs. LEE is caused ...primarily by materials stresses when hydrometeors (rain and hail) impact on rotating blades. The kinetic energy transferred by these impacts is a function of the precipitation intensity, droplet size distributions (DSD), hydrometeor phase and the wind turbine rotational speed which in turn depends on the wind speed at hub-height. Hence, there is a need to better understand the hydrometeor properties and the joint probability distributions of precipitation and wind speeds at prospective and operating wind farms in order to quantify the potential for LEE and the financial efficacy of LEE mitigation measures. However, there are relatively few observational datasets of hydrometeor DSD available for such locations. Here, we analyze six observational datasets from spatially dispersed locations and compare them with existing literature and assumed DSD used in laboratory experiments of material fatigue. We show that the so-called Best DSD being recommended for use in whirling arm experiments does not represent the observational data. Neither does the Marshall Palmer approximation. We also use these data to derive and compare joint probability distributions of drivers of LEE; precipitation intensity (and phase) and wind speed. We further review and summarize observational metrologies for hydrometeor DSD, provide information regarding measurement uncertainty in the parameters of critical importance to kinetic energy transfer and closure of data sets from different instruments. A series of recommendations are made about research needed to evolve towards the required fidelity for a priori estimates of LEE potential.
An automated wind turbine wake characterization algorithm has been developed and applied to a data set of over 19 000 scans measured by a ground-based scanning Doppler lidar at Perdigão, Portugal, ...over the period January to June 2017. Potential wake cases are identified by wind speed, direction and availability of a retrieved free-stream wind speed. The algorithm correctly identifies the wake centre position in 62 % of possible wake cases, with 46 % having a clear and well-defined wake centre surrounded by a coherent area of lower wind speeds while 16 % have split centres or multiple lobes where the lower wind speed volumes are no longer in coherent areas but present as two or more distinct areas or lobes. Only 5 % of cases are not detected; the remaining 33 % could not be categorized either by the algorithm or subjectively, mainly due to the complexity of the background flow. Average wake centre heights categorized by inflow wind speeds are shown to be initially lofted (to two rotor diameters, D, downstream) except when the inflow wind speeds exceed 12 ms−1. Even under low wind speeds, by 3.5 D downstream of the wind turbine, the mean wake centre position is below the initial wind turbine hub height and descends broadly following the terrain slope. However, this behaviour is strongly linked to the hour of the day and atmospheric stability. Overnight and in stable conditions, the average height of the wake centre is 10 m higher than in unstable conditions at 2 D downstream from the wind turbine and 17 m higher at 4.5 D downstream.
Short‐term forecasting of wind gusts, particularly those of higher intensity, is of great societal importance but is challenging due to the presence of multiple gust generation mechanisms. Wind gust ...observations from eight high‐passenger‐volume airports across the continental United States (CONUS) are summarized and used to develop predictive models of wind gust occurrence and magnitude. These short‐term (same hour) forecast models are built using multiple logistic and linear regression, as well as artificial neural networks (ANNs) of varying complexity. A suite of 19 upper‐air predictors drawn from the ERA5 reanalysis and an autoregressive (AR) term are used. Stepwise procedures instruct predictor selection, and resampling is used to quantify model stability. All models are developed separately for the warm (April–September) and cold (October–March) seasons. Results show that ANNs of 3–5 hidden layers (HLs) generally exhibit higher hit rates than logistic regression models and also improve skill with respect to wind gust magnitudes. However, deeper networks with more HLs increase false alarm rates in occurrence models and mean absolute error in magnitude models due to model overfitting. For model skill, inclusion of the AR term is critical while the majority of the remaining skill derives from wind speeds and lapse rates. A predictive ceiling is also clearly demonstrated, particularly for the strong and damaging gust magnitudes, which appears to be partially due to ERA5 predictor characteristics and the presence of mixed wind climates.
Plain Language Summary
Short‐term prediction of wind gusts is of high societal value but is challenging. Complex machine learning approaches offer potential for high‐fidelity short‐term forecasts. However, use of artificial neural networks offers only modestly increased skill increments over regression methods. Increased model complexity does not immediately translate to significantly improved forecasts.
Key Points
Artificial neural networks exhibit skill for short‐term forecasting of wind gust magnitude
Deep neural networks only slightly improve predictability of intense gusts at the cost of model overfitting and enhanced false alarm rates
Even high‐complexity models cannot overcome the predictability ceiling for high intensity wind gusts
“Wind Theft” from Onshore Wind Turbine Arrays Pryor, Sara C.; Shepherd, Tristan J.; Volker, Patrick J. H. ...
Journal of applied meteorology and climatology,
01/2020, Volume:
59, Issue:
1
Journal Article
Peer reviewed
Open access
High-resolution simulations are conducted with the Weather Research and Forecasting Model to evaluate the sensitivity of wake effects and power production from two wind farm parameterizations the ...commonly used Fitch scheme and the more recently developed Explicit Wake Parameterization (EWP) to the resolution at which the model is applied. The simulations are conducted for a 9-month period for a domain encompassing much of the U.S. Midwest. The two horizontal resolutions considered are 4 km × 4 km and 2 km × 2 km grid cells, and the two vertical discretizations employ either 41 or 57 vertical layers (with the latter having double the number in the lowest 1 km). Higher wind speeds are observed close to the wind turbine hub height when a larger number of vertical layers are employed (12 in the lowest 200m vs 6), which contributes to higher power production from both wind farm schemes. Differences in gross capacity factors for wind turbine power production from the two wind farm parameterizations and with resolution are most strongly manifest under stable conditions (i.e., at night). The spatial extent of wind farm wakes when defined as the area affected by velocity deficits near to wind turbine hub heights in excess of 2% of the simulation without wind turbines is considerably larger in simulations with the Fitch scheme. This spatial extent is generally reduced by increasing the horizontal resolution and/or increasing the number of vertical levels. These results have important applications to projections of expected annual energy production from new wind turbine arrays constructed in the wind shadow from existing wind farms.
This work presents a new observational wind atlas for the Great Lakes, and proposes a methodology to combine in situ and satellite wind observations for offshore wind resource assessment. Efficient ...wind energy projects rely on accurate wind resource estimates, which are complex to obtain offshore due to the temporal and spatial sparseness of observations, and the potential for temporal data gaps introduced by the formation of ice during winter months, especially in freshwater lakes. For this study, in situ observations from 70 coastal stations and 20 buoys provide diurnal, seasonal, and interannual wind variability information, with time series that range from 3 to 11years in duration. Remotely-sensed equivalent neutral winds provide spatial information on the wind climate. NASA QuikSCAT winds are temporally consistent at a 25km resolution. ESA Synthetic Aperture Radar winds are temporally sparse but at a resolution of 500m. As an initial step, each data set is processed independently to create a map of 90m wind speeds. Buoy data are corrected for ice season gaps using ratios of the mean and mean cubed of the Weibull distribution, and reference temporally-complete time series from the North American Regional Reanalysis. Generalized wind climates are obtained for each buoy and coastal site with the wind model WAsP, and combined into a single wind speed estimate for the Great Lakes region. The method of classes is used to account for the temporal sparseness in the SAR data set and combine all scenes into one wind speed map. QuikSCAT winds undergo a seasonal correction due to lack of data during the cold season that is based on its ratio relative to buoy time series. All processing steps reduce the biases of the individual maps relative to the buoy observed wind climates. The remote sensing maps are combined by using QuikSCAT to scale the magnitude of the SAR map. Finally, the in situ predicted wind speeds are incorporated. The mean spatial bias of the final map when compared to buoy time series is 0.1ms−1 and the RMSE 0.3ms−1, which represents an uncertainty reduction of 50% relative to using only SAR, and of 40% to using only SAR and QuikSCAT without in situ observations.
•We present the first full observational wind atlas for the Great Lakes.•We describe a method to integrate SAR and QuikSCAT wind speeds.•We describe a method to integrate satellite and in situ wind observations.•We present methods for correction of wind speeds for the ice season.•Wind prediction errors are reduced by integrating different data sets.