Dynamic forecast combinations of improved individual forecasts for the prediction of wind energy

We study the prediction performance of different improved individual wind energy forecasts in various static and dynamic combination processes. To this end, we develop a combined error minimization model (CEMM) based on nonlinear functions. This approach reflects the nonlinear nature of weather and especially of wind energy prediction problems. Based on the model, we construct significantly improved individual forecasts. The corresponding time dependent model coefficients are determined by dynamic OLS (ordinary least squares) regression and Kalman filter methods. The former method shows a slightly better performance than the Kalman filter based approaches. Further improvements can be achieved by a combination of these improved wind energy forecasts. In this case, the combination coefficients are calculated from a static and two dynamic OLS regressions. The resulting forecasts are characterized by a further increased prediction accuracy compared to the combination of the uncorrected forecast data and can outperform a given benchmark.