Options
2013
Conference Paper
Titel
Robust spatio-temporal wind power forecasts via nowcasting of low quality inputs
Abstract
The goal for this work was to reduce wind power forecast error using data from locations distant from the target wind farm being forecasted. In particular, we wanted to see if a mix of off and onshore wind data would be useful for this purpose. This data was of quite problematic so it was necessary to develop an algorithm based on bootstrapped expectation maximization (EM) to replace low quality or missing data with estimated values derived from numerical weather prediction (NWP) "nowcasted " values and other distant measurements. We then developed a feature selection algorithm which reduced the highdimensional data to sizes optimal for a neural network forecasting algorithm. We show that our missing data fill-in and feature selection algorithms can individually improve the forecasting performance of a commercial baseline system. Then we show that the distant measurements offer further improvement, but only if both missing data fill-in and feature selection are employed.