Developing and Testing Wind Velocity Retrieval Algorithms for Doppler Wind Lidar

A 3-dimensional wind lidar is being evaluated at the National Wind Technology Center (NWTC) for its applications in wind energy. The focus of the work described here is to develop algorithms that can increase data availability and accuracy in estimating wind velocity from the line of sight (los) velocity (V_los) from Plan Position Indicator (PPI) scans. The common algorithm (AL0) starts by removing V_los estimates that have low signal-to-noise ratio (SNR). Then, assuming a horizontally homogeneous wind field and zero vertical wind speed (w), the wind velocity is estimated by application of ordinary least square (OLS) fitting, and the results are averaged to produce the 10-minute mean wind velocity (scalar averaging) at each range-gate position. This approach has uncertainties because: (1) SNR is robust but conservative for quality control and use of any SNR threshold may result in exclusion of valid V_los values causing low data availability. (2) While 10-minute mean w = 0 is typically valid, assuming zero w for each individual V_los field may introduce biases. (3) The variance of V_los changes with azimuth angle as it is the projection of the variance of the wind vector on the los. This violates the equal variance assumption in OLS fitting. The two new algorithms are developed to increase data availability and the accuracy of 10-minute mean wind velocities. Both algorithms assume that the wind velocity is normally distributed and use the maximum likelihood estimator for which the variance of V_los changes with azimuth angle. The first algorithm (AL1) uses the 10-minute mean V_los to estimate the 10-minute mean wind velocity. In comparison to scalar averaging, AL1 can reduce the variation in V_los and the assumption of w = 0 is more likely to be valid. To increase data availability, V_los with low SNR is retained if its difference from the mean is smaller than three times the standard deviation of V_los. The second algorithm (AL2) uses the median of V_los over 10 minutes (as opposed to the mean value as in AL1). For a normal distribution, the sample median is a robust estimate of the mean and is insensitive to outliers (e.g. incorrect measurements associated with low SNR). Thus, using the sample median allows for the use of V_los with very low SNR and eventually increase data availability for AL2. A preliminary analysis of lidar data collected during February 15 to 26, 2013 shows that AL2 out-performs AL0 and AL1 when the resulting wind speed estimates are compared with independent data from a sonic anemometer (Table 1). Work is underway to test the performance of the three algorithms using a dataset of several months collected during spring/summer 2013 at NWTC, and the errors/uncertainties of each approach will be quantified in terms of their relationships with atmospheric conditions, such as wind shear and atmospheric stability, using the data from instrumentation deployed on the NWTC meteorological towers.Table 1 Summary of performance of the three lidar wind retrieval algorithms