Accurate short-term wind power forecasts become increasingly important as the penetration of wind generation on the electric grid increases. When wind power makes up more than a minor portion of power generation, accurate forecasting of significant changes in wind speed that leads to large changes in power production (known as “ramps”) are needed several hours in advance to effectively and economically manage the electric grid. This is particularly important on small isolated electric grids, such as on the big island of Hawaii.

With this in mind, an observation targeting study was done using a 48-member Ensemble Kalman Filter. The goal was to identify locations and variables to observe in order to improve the short-term wind forecast for wind farms on the Island of Hawaii. The results of this research laid the groundwork for the current WindNET project that aims to test the impact of targeting observations on short-term wind power forecasts. Three sodars and one radiometer were sited using the observation targeting results in order to test their impact on zero to three hour wind power forecasts for the Apollo wind farm near the southern tip of the island.

Two approaches were taken to attempt to extract the useful information and improve the zero to three-hour forecast of wind power at Apollo, particularly the forecast of ramps:

1. Assimilation of the observations into a mesoscale model using observation nudging, and 2. The use of time-lagged correlations between the observations and wind power production under different weather regimes to predict the probability of ramps.

In the first approach, the WindNET observations were assimilated into the Mesoscale Atmospheric Simulation System (MASS) using observation nudging. Two-kilometer resolution simulations were run every 3 hours. The simulations were run for 15 hours, which consisted of a 3-hour nudging period and a 12-hour forecast period. Outer grid simulations with resolutions of 32 and 8 km were run for 18 hours every 6 hours to provide external boundary conditions. Forecasts were produced for the months of December 2010 and April 2011. These months were selected because of the prevalence of significant wind power ramps at the Apollo wind farm. In the control experiment, rawinsonde and METAR observations were assimilated on the outer grids using the MASS optimal interpolation scheme. METAR data was also assimilated on the 2-km inner grid using observation nudging during the 3-hour nudging period. The WindNET data experiment included all of the data assimilated by the control experiment. In addition, it assimilated the WindNET sodar and radiometer data through observation nudging. Preliminary results show a modest improvement in the short-term forecast through the assimilation of the WindNET observations.

In the second approach, time lagged correlations were computed between prior changes in WindNET observations and future changes in wind power production at the Apollo wind farm under various weather regimes. This approach has a much lower computational cost and may be more suitable for look-ahead periods of one hour or less. This approach may be useful in reducing the overall short-term forecast errors. However, since the goal of WindNET is to improve the short-term prediction of wind power ramps, another approach was tried as well. Quantile regression was used as a tool to assess the probability of a significant power ramp or the occurrence of high amplitude sub-hourly power variability in the upcoming three hours based on both observed and model-predicted data. The data were classified into regimes in order to customize the probabilistic forecast to the prevailing weather, time of day and season.

At the conference presentation, we will review the project in greater detail and present results of the forecasting experiments with a focus on the impact of the WindNET data on short-term wind power ramp forecasts.