Proposed Models

  • Initial attempts leveraged 2-d convolutional networks following approach of SRCNN and Conv-LSTM (Shi et al).
  • Grid search several architecture configurations (filters, depth, etc).
  • \(3 \times 3\) grid appears not to be large enough input to support learning radiation signal.
  • Resulted in low \(R^2\) and negative efficiency, predictions resulted in network appearing to learn mean of distribution rather than variation.

Focus on CNN-LSTM

Worked well for single grid point output would it work for 2d grid output?

Creative Commons License
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.