Diagnostics and Calibration

[anonymous34]

Is it possible to evaluate diagnostics for only a certain date range? In addition, is there a suggested warm-up time for Raven?

[rchlumsk]

Is it possible to evaluate diagnostics for only a certain date range? In addition, is there a suggested warm-up time for Raven?

Yes, a relatively new feature is the ability to include a time series weighting file, which can be used to not weigh specific data points in the diagnostics. For example, for two years’ worth of data, a time series weighting file with zeroes for the first year of data and ones for the second year of data would only weight the second year.

This can be used to specify a “warm-up” period for Raven, which is common in hydrologic modelling to reduce the impact of initial conditions. For hydrologic models in general, a warm-up period depends on the model complexity; for a relatively simple model, a warm-up period of one year is likely sufficient.

Likewise, diagnostics could be reported for only summer months (or Mondays, for that matter) by using non-zero weightings only during these months.

[anonymous34]

What is the purpose of a warm-up period? What kind of period should be used?

[rchlumsk]

What is the purpose of a warm-up period? What kind of period should be used?

The purpose of the warm-up period (or spin-up period, as it is also called) is intended to reduce the impact of initial conditions on model results. This is done by running the model for a sufficiently long time that the model physics can account for inaccuracies or errors in initial conditions; the length of time the model is run before results are used is called the spin-up period.

The length of period used for a warm-up is typically one year, although it might be wise to test how long the model is run for before changes to initial conditions become negligible.

[anonymous34]

I am in the process of setting up separate calibration and validation periods for my model. What is the best/easiest way to setup these runs in Raven for different periods?

[rchlumsk]

I am in the process of setting up separate calibration and validation periods for my model. What is the best/easiest way to setup these runs in Raven for different periods?

There are a few different ways to setup the runs. You could have:
• Separate model files for calibration and validation runs, although not the easiest way. Note that you can use a place holder value of -1.2345 in observed data for periods that should not be evaluated, that is the other way to control what periods the fit metric is evaluated for
• Single file for each gauge data with separate weight files for calibration and validation, so that the metrics are evaluated accordingly to the correct data set
• Have a separate start and duration setup in the rvi file (separate files or commented/uncommented system) to run the model for various times, i.e. starting with calibration period or validation period accordingly. Main trick here is to ensure the solution.rvc from the best calibration run is used to warm start the validation run (true in any validation run). Should be able to have all forcing and observed data in the files in this case, Raven can read the data and only use portion for after the validation start

[anonymous34]

The parameters.csv file is outputting -6666.66 and -7777.77 for values, what does this mean?

[rchlumsk]

The parameters.csv file is outputting -6666.66 and -7777.77 for values, what does this mean?

The -6666.66 code means the value was supplied but not required therefore autogenerated, and the -7777.77 value means the value was required and was autogenerated. Note that the parameters.csv file is becoming obsolete.

[anonymous77]

How do we define the time series we want for the calibration optimization? I understand that a file can be generated to identify which points will be used for Nash-Sutcliffe, but again I don’t know how to make that work.

[rchlumsk]

How do we define the time series we want for the calibration optimization? I understand that a file can be generated to identify which points will be used for Nash-Sutcliffe, but again I don’t know how to make that work.

You can specify an observation weights time series file, as documented on page 128 of the manual. Usually, you would set zero weights for unused observations (e.g., winter data) and a one for used observations. The length of the observation time series must be the same as the observation weights series.