nitrogen dioxide measured in mol/m^2 was converted into kilograms via multiplying the initial values by the number of m^2 in a rectangular raster cell and by 46.006 g/mol -- the molar mass of NO2.

https://www.kaggle.com/katemelianova/ds4g-spatial-panel-data-modeling

If a gross_load_mw prediction results in a capacity factor over 100, it is a bad prediction.

Cap it? Can we cap predictions for other ys too?

from merged_df?

time slider / layer selector for ys?

Bayesian Optimization?

https://registry.opendata.aws/sentinel5p/

How big would all the rasters be? What would download time be?

How does the time to open files, mask them, process statistics, and output take? Is it shorter than earth engine?

Currently, a statistic (mean / std / sum) of each feature is taken from a circle of n radius (10km, 20km) around the lat/lon of the power plant.

This doesn't account for wind, and has no (endogenous) method of considering background noise (i.e. car pollution) besides also including those as features.

What if we take observations at 2 (3?) points... Using daily weather data can't we take an observation 10km upwind (i.e. a sample of background noise) and 10km downwind (i.e. where pollution from the powerplant should have been blown, and difference them.

Advanced version: rather than just using a circle or a point upwind/downwind, what about an 'ice cream cone' stretching back to the powerplant.

right now we write wrappers manually. It would be good to use regex to beable to automatically interpret pandas offset codes such as '3D'

Look at recent days?

by fuel type / distance?

TODO: scaling

TODO: Add date features

Currently clipping based on daily percentiles, take whole year's worth of values to make sure we're not clipping too much.