Varying-coefficient space-time models for short-term wind forecasting

Short range forecasts of the wind vector one to three hours ahead are useful for scheduling and transmitting wind power used in generating electricity. Statistical models tend to be the most effective for capturing the dependencies among important variables within this time horizon, but most models usually focus on producing only a wind speed forecast. Both speed and direction directly influence wind power output. In this work, we develop a vector autoregressive (VAR) model in which the horizontal and vertical components of the wind vector at each of several spatial locations are both the input and the response. This model produces both wind speed and wind direction forecasts simultaneously at each of the spatial locations, but the number of locations for which parameters can be accurately estimated in the model and good forecasts are produced is limited. The dependencies in space and time change based on the direction of the prevailing winds, so a key improvement in this VAR model is to allow the coefficients to vary continuously based on a wind direction variable. The performance of the model is tested with three years of hourly wind vector data at four locations in the Pacific Northwest. It is useful not only in producing forecasts but also in simulation of long series of wind vectors that can be incorporated in utility system experiments.