Experiment with ensemble of models

[JackKelly]

Detailed Description

Combing multiple ML models often wins Kaggle competitions.

The ensemble could include:

• Model which takes single step of satellite imagery from optical flow and predicts PV yield.
• Multiple Optical flow predictions using flow computed at t-1, t-2, etc.
• Model which predicts PV yield from ML predictions of future imagery (from SatFlow).
• NWP to PV.
• Similar cloud sequences from the past.

Context

Whilst ensemble models work well in many instances, perhaps our Perceiver IO will already perform as well as an ensemble model?

Possible Implementation

Maybe our "joint model" (based on Perceiver IO) would be the "meta model" which combines the ensemble? Maybe the joint model directly just gets historical data (satellite, historical pv) and the predictions from component models as inputs.

From a software engineering perspective, it might be nice to have several discrete models.