Limitations and Further Work

  • Relationship between GCM output and observation is influenced by the grid scale and bias of the GCM models. Investigation of the influence of reanalysis products (rather than GCM) deserve investigation and comparison.
  • 2-d convolution did not perform well, would larger input grid sizes improve performance?
  • Network architecture search in this study is rudimentary, further investigation into techniques for network architecture search would be of interest.
  • Explanation for decreases under RCP8.5 are not offered in this study. DNN methods seen as black boxes. Explore techniques for model interpretation such as LIME (Ribeiro, et al. 2016) [1].
  • Investigate are differences due to sampling of locations? Eg: Carpentaria Downs further away from other sites and isolated. Would spatially uniform distribution of locations affect performance?
  • Would ensemble of CNN-LSTM networks improve variability of performance across multiple sites?

References.

[1]
Ribeiro, M.T., S. Singh, and C. Guestrin (2016 ). "Why should i trust you?": Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery: San Francisco, California, USA. 1135–44

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.