UoP
Solar irradiance (GHI) varies sharply with sky condition - clear, cumulus, overcast - and a single regression network has to learn three different physics regimes at once, which it does poorly. Photovoltaic operators need short-horizon forecasts that are accurate across all three regimes, not just on average.
Approach
Built a two-stage system. First, a classical CV sky-condition classifier labels each sky image into one of the three regimes. Then a regime-specialized ResNet sub-model - trained only on its assigned regime - produces the GHI estimate. The split lets each ResNet specialize to a single physics regime. Achieved 10.80% nRMSE and 98.88% Pearson correlation. Published at IEEE ICIIS 2021.
A single regressor fits the average regime well and the tails badly. Routing first turns a hard regression problem into three easier ones, and the routing is interpretable - operators can see which regime each forecast came from, which makes the system trustworthy in a way an opaque end-to-end network is not.