MOBT-1069: Using adjacent validity times in SAMOS

[bayliffe]

Additional KGO for samos coefficient estimation using adjacent validity times.

estimate-samos-coefficients-from-table/kgo_coordinates_adjacent.nc

Coefficient differences: Original (left) vs. Adjacent VTs (right)

DIFFER : VARIABLE : emos_coefficient_alpha : VALUES : 7.20529 <> 7.00847
DIFFER : VARIABLE : emos_coefficient_beta : VALUES : 0.678961 <> 0.683584
DIFFER : VARIABLE : emos_coefficient_gamma : VALUES : -0.406835 <> -0.402844
DIFFER : VARIABLE : emos_coefficient_delta : VALUES : 75.009 <> 75.7245

estimate-samos-coefficients-from-table/kgo_gam_cube_adjacent.nc

Coefficient differences: Original (left) vs. Adjacent VTs (right)

DIFFER : VARIABLE : emos_coefficient_alpha : VALUES : -1.19905 <> -1.36832
DIFFER : VARIABLE : emos_coefficient_beta : VALUES : 0.414136 <> 0.41368
DIFFER : VARIABLE : emos_coefficient_gamma : VALUES : 0.00087395 <> 0.00103704
DIFFER : VARIABLE : emos_coefficient_delta : VALUES : 0.536354 <> 0.439965

Additional KGO for SAMOS GAM training

The kgo_coords.pkl file contains gams for the historic forecast and truths. For each of these there is a gam for the means and standard deviations. Below I have plotted the means of both to try and demonstrate that the new KGO looks reasonable.

The first plot is of the existing kgo_coords.pkl file as usually configured in the acceptance test.

The second plot is the output for the same test that is returned simply by changing the forecast period to be 1-hour earlier (a different validity time).

Similarly here are the GAMS for a validity time 1-hour later.

The final plot shows the GAMS in the new KGO (kgo_coords_adjacent.pkl) that is created by looking at the original forecast period and the adjacent times +- 1 hour, which we've looked at individually above. The scales look to be a combination of the various validity times considered. The overall form is smoother which is also pleasing.

As is fairly obvious above the GAMS fitted to the historic forecasts and the truths are incredibly similar. I've checked the statistics_ data on the forecast and truth GAMs and they are incredibly similar (though not identical). I generated the same kind of plot after changing the diagnostic to wind speed which is also present in the acceptance test data tables. The plots still look very similar but the scale is very slightly different between the historic forecasts and truths. This suggests that for the temperature diagnostic (shown above) our forecasts and the truths are generally very close (as you might hope) and as a result the GAMs appear incredibly similar as well.