The experiment setting is similar to Twin Experiment 2, but with a different truth (at the upstream location, CHC and CHG are non zero in the summer).
Results
At assimilation station
Time series:
RMSE (computed over 51 analyses cycles over 2015):
At validation stations
Time series (Location 5006A40):
RMSE (Location 5006A40):
Over all locations:
Discussions
- Discuss with NIER the reasoning behind the DA (error modelling) settings. Perhaps we can check the resulting error statistics against observed error statistics. We can also take a pragmatic approach: check if the filter performance is sensitive to the specified error statistics.
- Are we satisfied with the error modelling (std of Observation & settings of noise statistics & which parameters (and tributaries) to include)?
- EnKF works well in improving forecast accuracy at the assimilation station and all down stream stations, but not the upstream stations.
- Do we need to check other water quality variables as well? which ones are of importance?
- The RMSE of the deterministic run from the assimilation station to down stream is more or less uniform. Yet the EnKF impact gets bigger in the downstream direction.
- Can we explain this?
- The RMSE at the most downstream location is especially small, even the deterministic one, as if it is not really affected by the noise defined upstream. The model has apparently a different dynamic at this downstream reach.
- Can we explain this?
- Do we need to add noise to downstream tributaries? Why did we choose to only add noise to upstream tributaries (see first point)?
- Water temperature is not included in the noise model definition nor in the state definition of the filter. Yet it is affected as it has quite a significant RMSE. The filter has also a slightly positive effect downstream on water temperature. It is likely due to the inclusion of solar radiation in both the noise model and state definition of the filter.
- Are we satisfied with water temperature settings and results?
- Goal is to improve forecasts and find the best DA settings
- Can we say anything about optimality of the filter?
- Have we reached the maximum possible improvement? what can be done to improve the accuracy even more? (1) use a larger ensemble, (2) use more stations for assimilation, i.e. upstream (3) check if other algorithms with the same settings used here can produce more improvement
- There are multiple experiments possible (e.g. (a) comparison DA algorithms, (b) comparison different noise settings, (c) comparison different assimilation settings (which locations do we assimilate), (d) comparison nr. of ensembles (comparison EnKF-GS limited nr vs. ensembles with EnKF)
- What experiments to include in paper?
- In case we see a systematical difference between Deterministic run and observations, discuss next steps:
- Should we advise to increase nr./frequency of observations to enhance forecast using DA
- Should we improve internal EFDC parameter settings to reduce the systematical error? Is this feasible?
