Figure 4 . Scatter plots showing the performance of NN1W and
NN2W against SA across all sites. Points above the grey zero line show
configurations where the NN configuration improved performance over SA.
The “Maximum improvement” line is based on the performance of the SA
simulations, and is simply (1-NSE) in subplots a and b, and (1-KGE) in
subplots c and d.
We also compared the KGE for different periods of temporal aggregation
to evaluate whether performance improvements of the NN configurations
persisted across timescales (Figure 5). The KGE score was chosen here
because it shows greater variability than the NSE score in Figure 3,
though the results are similar for NSE. We see that the sub-daily
aggregations, on average, showed better performance for both NN
configurations, demonstrating that they were able to capture the diurnal
cycle of turbulent heat fluxes. This is mostly due to the strong
dependence of turbulent heat fluxes on solar radiation, which we will
further explore in section 3.2. Both NN1W and NN2W were able to
outperform SA across all timescales for sensible heat.
However, at daily and longer temporal aggregations differences between
models were seen in latent heat performance. The NN1W configuration
performed better at sub-daily timescales than for daily or longer
aggregations, for which performance was similar to SA. In contrast, the
NN2W configuration performed better for latent heat than SA across all
timescales.