Data pre-processing and NaNs filling

[zannads]

@PansosPonso said:

• Using logarithms improves things. (@eldemet confirmed)
• When using features among DMAs (for example the minimum value across DMAs for each day), use some kind of normalization before (because for example, this feature will tend to result in the value of DMA F or A).
• Some interesting features: days_since_rain, real_feel, dew_point, heat_index, ewm of DMAs. (see also his commit on dev)
• Instead of forecasting the value of some DMAs directly, it is beneficial to first add some random noise, forecast that value and then take the average.