GC43A-0688:
Quantifying the Geographic Diversity Needed to Reduce Wind Power Variability

Thursday, 18 December 2014
Clara St. Martin, University of Colorado at Boulder, Boulder, CO, United States, Mark Handschy, Enduring Energy, LLC, Boulder, CO, United States and Julie K Lundquist, U. of Colorado at Boulder, Boulder, CO, United States
Abstract:
The variability in wind-generated electricity complicates its integration into the electrical grid at higher penetrations. This variability can be reduced by interconnecting wind farms across large regions. We investigate the effectiveness of this aggregation with two extensive data sets. The first consists of 44 years of hourly 10-m wind speeds from 117 stations spanning distances of 5000 km across Canada from the National Climate Data Archive of Environment Canada. The second dataset consists of four years of 5-minute averaged wind speeds from stations across 350 km of the northwestern US from the Bonneville Power Administration. This study seeks to quantify how time scale affects the drop-off of correlation with distance, and hence statistical independence and the degree of geographic diversity needed to provide smoothing on various time scales.

To eliminate non-stationarities, we filter both datasets to remove diurnal cycles and seasonal trends, after which we digitally high-pass filter the data on 0.25-2000 hour timescales relevant to power grid management. The wind speed data are then correlated between stations for each high-pass filter cut-off. In both datasets, the correlations between stations fall to zero with increasing station separation distance. Differences in inter-station azimuthal bearing account for a small but distinct fraction of the wide scatter in correlation vs. distance for the Canadian data set, with stations separated along a line 10° North of East-West being systematically less correlated than those perpendicular to that line.

Similarities between these two datasets reveal behavior that, if universal, could be particularly useful for grid management. Both datasets exhibit a correlation length that varies as l = l0(1+25/T)-2, where l0 is the correlation length without high-pass filtering and T is the high-pass cut-off (in hours). Since the inter-site separation needed for statistical independence falls for time scales shorter than 25 hours, faster fluctuations can be effectively smoothed by aggregating wind plants over smaller areas.