Table nbsp Optimal window length for the

Fig. 4. RMSE (top) and forecast skill (bottom) versus the forecast horizon for the GHI forecast.Figure optionsDownload full-size imageDownload as PowerPoint slide
Fig. 5. Same as Fig. 4 but for the DNI forecast.Figure optionsDownload full-size imageDownload as PowerPoint slide
These results show that GHI is much easier to forecast that DNI: the RMSE for all forecast horizons ranges between 30 and 45 Wm−2 for GHI, whereas for DNI, it BRD73954 is more than double, ranging from 50 to 100 Wm−2. The Figure shows that the optimized kNN models reduce the RMSE with respect to the baseline persistence model for all the cases studied. The reduction in the RMSE translates into significant forecast skills that range between 10 and 22%, and 10 and 26% for the GHI and DNI testing set, respectively.
Ideally there would be no degradation in the forecast performance for the testing set relative to the optimization set. In practice that is never the case, and the tables show a small increase in RMSE and a small decrease in the forecast skill when the models are applied to the testing set. This deterioration is clearly visible in Fig. 4 and Fig. 5. The inclusion of the variables derived from the images only slightly improves the forecast, with a bigger improvement in the case of DNI.