Chapter 5: Uncertainty of Models
Summary
In the local downscaling task, the CNN-LSTML demonstrates a lower MPI for the 95% coverage probability than the other models. However, does exhibit some dependency between prediction and residual at the lower ranges of predicted radiation. Under the two climate warming scenarios RCP4.5 and RCP8.5 all models exhibit a higher uncertainty for the upper values of radiation, however the CNN-LSTML also demonstrates lower MPI for the 95% coverage probability. Indicating less uncertainty than the other models.
The universal downscaling task exhibits higher MPI in comparison to the local downscaling task, however the CNN-LSTMU has lower uncertainty at the 95% coverage probability than the GLM model. This is also consistent for the climate warming scenarios RCP4.5 and RCP8.5. Both models exhibit higher uncertainty for the upper values of radiation, although the CNN-LSTMU exhibits a lower range between the 97.5% and 5% quantiles of the bias. The GLM model exhibits higher uncertainty during the Summer months under the test set, whereas the CNN-LSTMU exhibits higher uncertainty during the Summer and early Autumn months. Under the RCP4.5 climate warming scenario the GLM exhibits higher uncertainty in winter and summer, while the CNN-LSTMU exhibits higher uncertainty in winter under the same profile. This is similar under the RCP8.5 scenario for the CNN-LSTMU model, whereas the GLM is more uncertain during the winter months under that profile.
Downscaling Global Climate Models with Convolutional and Long-Short-Term Memory Networks for Solar Energy Applications by C.P. Davey is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.