The proposal is an ML model which learns a mapping between the things that NWPs are good at (humidity, temperature) and clouds. The ultimate aim is to do better than NWP's "paremeterisations" of clouds.
Maybe start without PV. Then add PV. The idea being that combining PV, multi-level NWP and satellite makes sense because they're all different "views" of the clouds.
Use "Perceiver IO as RNN" where each cross-attention is a new timestep, and each timestep creates a new timestep.
Use Hierarchical Perceiver (issue #14) because there's a strong sense of locality: the most important NWP "pixel" should be the NWP pixel at the location of the satellite pixel. Although it'll also be useful to focus on what the clouds look like upwind of that pixel. Could imagine a simple CNN that maps a small area of NWPs to a satellite pixel (although the satellite would probably have to be reprojected). Although I still think I prefer using a Hierarchical Perceiver, because a CNN can't focus upwind etc. Could imagine the first level of the hierarchy literally just focuses on a single vertical column of NWPs.
Inputs
At least 2 timesteps of recent:
multi-level NWPs of humidity, temperature, wind speed & direction
Satellite
PV? (if, in issue #20, we can find a way to get multi-PV prediction working as well as single-PV prediction). Probably also feed in clearsky irradiance or the output of an ML model that learns the local features of each PV system, so the model can see what this PV system is telling us about the atmosphere, not local features or celestial mechanics. Maybe a single model can learn these features of each PV system, if the PV element also includes the absolute Sun angle and elevation (so the model can learn shading).
land / sea / land height (maybe from the UKV files?)
Maybe use timesteps on the hour (so all inputs are at the same time as the NWPs?)
Query
All queries include temporal encoding of the timestep to forecast.
Cloud prediction task: Multi-level NWP of humidity and temperature
The proposal is an ML model which learns a mapping between the things that NWPs are good at (humidity, temperature) and clouds. The ultimate aim is to do better than NWP's "paremeterisations" of clouds.
Maybe start without PV. Then add PV. The idea being that combining PV, multi-level NWP and satellite makes sense because they're all different "views" of the clouds.
Use "Perceiver IO as RNN" where each cross-attention is a new timestep, and each timestep creates a new timestep.Use Hierarchical Perceiver (issue #14) because there's a strong sense of locality: the most important NWP "pixel" should be the NWP pixel at the location of the satellite pixel. Although it'll also be useful to focus on what the clouds look like upwind of that pixel. Could imagine a simple CNN that maps a small area of NWPs to a satellite pixel (although the satellite would probably have to be reprojected). Although I still think I prefer using a Hierarchical Perceiver, because a CNN can't focus upwind etc. Could imagine the first level of the hierarchy literally just focuses on a single vertical column of NWPs.
Inputs
At least 2 timesteps of recent:
Maybe use timesteps on the hour (so all inputs are at the same time as the NWPs?)
Query
All queries include temporal encoding of the timestep to forecast.
Output