How Long is Long Enough? The Effect of Interannual Correlation on Wind Resource Assessment

Abstract:

Wind energy development relies on estimates of the wind resource at a development site. The estimates are expected to portray production and variability over the lifetime of the wind farm. Since long-term measurements at the development site are rarely available, correlations with nearby long-term measurements are often used as a proxy. This common practice assumes that the long-term reference measurements, typically over 10‒20 years, adequately represent the actual resource characteristics of both the reference site and the target site. However, current methodologies do not consider the possibility of significant year-to-year correlations in the wind resource, suggesting that 10 or 20 years may not be long enough – or may be too long – to assess development potential.

To explore the role of year-to-year correlations in impacting interannual variability, we have investigated 54 years of homogenized hourly mean wind speed data from 117 Canadian stations. After selecting only stations with significant wind resources, we have extrapolated the hourly 10-meter winds to an altitude of 80 meters and calculated hourly power production and then an annual capacity factor for a widely-deployed turbine, the GE 1.5 MW XLE, deployed at the site of the wind speed measurements. We evaluate the effect of correlations between consecutive years by randomly permuting a station’s sequence of annual capacity factors, and comparing the statistics of the randomized data to the statistics of the actual data.

Estimating the expected value of average production for a station’s final twenty years (a typical wind farm lifetime) from the immediately preceding years, we find for randomized data that the error decreases the more years are considered, as expected. However, surprisingly, for the actual time series the best estimate is obtained using just a short few-year segment immediately preceding the final 20-year target period. Further, using the sample standard deviation σ and mean μ of N-year data segments we find that interannual variability (σ/μ) for actual data continues to grow with N out to the 54-year limit of our records, while for randomized sequences it remains essentially constant. Our presentation will explore the effects of these correlation-induced errors and biases on P50 and P90 estimates.