Summary

Overall both of the local CNN-LSTM_L and CNN-LSTM_L DENSE models outperform the other local models. Each of the model residuals exhibit a degree of non-uniform variance, and the autocorrelation of the residual indicates that there is still some unexplained information in the radiation given the model parameters. Each of the models appear to underestimate the radiation signal to some degree with the CNN-LSTM_L providing better estimates for the extremes in the signal. Although, the estimates appear to smooth the variation between the months providing better indication of the overall seasonality of the radiation. Both CNN-LSTM model variations exhibited highest average error for estimates during the months of October and December, at both the example sites. The GLM model also exhibited larger average errors during January and February at both sites. Indicating that there may be some level of uncertainty in estimating extremes of radiation during the summer months.

Interestingly the best performing sites for the universal models differed to those of the local model. The 2-dimensional case introduced a higher degree of variation for both the input signal and the output signal due to the increase in the number of grid points, which may account for this difference. The CNN-LSTM_U model exhibited higher average error for the months in spring and autumn, as opposed to the GLM which exhibited higher average bias in the summer. The CNN-LSTM_L model demonstrated better performance on the majority of sites in the local downscaling task. While the KGE indicated the CNN-LSTM_U exhibited a difference between the predicted and observed distributions for radiation in the universal case (also indicated in the Taylor diagram), the CNN-LSTM_U exhibited good performance for the majority of sites as indicated by the other metrics.

Chapter 4: Model Evaluation and Comparison

Summary