Maybe a model that sees satellite imagery, NWP, and PV for entire GB, and directly predicts national PV. Maybe with inputs from each GSP. Downres using CNN? Could also use this model to predict each GSP. Label each pixel with its GSP. Is there a way to force the sum of each GSP to equal the National PV?
But how to predict future satellite imagery? Maybe a unet trained on downsampled satellite imagery. We'll probably need data of the full geographical extent to train that model, so we have enough examples.
Maybe need to use a Perceiver IO in order to fit the full image into the model. Or use a CNN to doensample.
Or a hierarchical model which predicts each GSP and then predicts entire country.
Or model which "just" adjusts GSP predictions, perhaps taking into account local correlations and uncertainty
All these models could be a useful step towards national demand forecasting, too
Maybe a model that sees satellite imagery, NWP, and PV for entire GB, and directly predicts national PV. Maybe with inputs from each GSP. Downres using CNN? Could also use this model to predict each GSP. Label each pixel with its GSP. Is there a way to force the sum of each GSP to equal the National PV?
But how to predict future satellite imagery? Maybe a unet trained on downsampled satellite imagery. We'll probably need data of the full geographical extent to train that model, so we have enough examples.
Maybe need to use a Perceiver IO in order to fit the full image into the model. Or use a CNN to doensample.
Or a hierarchical model which predicts each GSP and then predicts entire country.
Or model which "just" adjusts GSP predictions, perhaps taking into account local correlations and uncertainty
All these models could be a useful step towards national demand forecasting, too