How to omit a process without providing the relevant output for downstream processes?

[cneyens]

I have a model that includes the SolarGeometry, Atmosphere, Canopy, Runoff, and Soilzone processes. The Snow process is omitted because snow is negligible in my area of interest. I know I can omit the Snow process from the supplied process list, but downstream processes still expect the relevant fluxes and state variables, e.g., from files on disk. I can create these files and fill them with dummy data representing a system without snow, but that is cumbersome and error-prone. Also, for our large model (30 years, daily time step, with over 100,000 HRUs), this is very costly.

Is there an easier way to tell Pywatershed to internally use dummy data representing an omitted process, as required by downstream processes? I saw in the code that there is a PRMSSnow.set_snow_zero() function, which seems to do this, but it isn't used anywhere.

[jmccreight]

Cas, this is a good question. I was a bit off in my previous response, which you can ignore. I forgot that, while Adapters can be passed to individual processes, using Model really only allows files for inputs at the current time. I happened to be working on something else that was a perfect starting point for doing a hack based on what you are looking for. It was pretty straight forward. I have some comments in the preamble to the notebook gist that I'll copy here:

---

Following on Cas' question, here is a proof of concept about how to fake required process inputs.

While Adapters or inputs to Adapters can be passed to individual processes, using Model really only allows files for inputs at the current time. Since the design to allow data/Adapters to be passed to Model will take some thought, here is the hack to do it now. The following does require a minor PR to the develop branch for the control.edit_init_start_times method, however. That should be merged very soon.

Below, the full NHM model is run on the Delaware River Basin in June. It is checked that there is no snow. Then a submodel which drops PRMSSnow (and also PRMSGroundwater, and PRMSChannel) is setup in a way where we can add the input as Adapters after the fact. Of course, the original climate/weather forcing variables are set as input Adapters. Additional inputs are needed because PRMSRunoff and PRMSSoilzone processes have inputs which are not satisfied when PRMSSnow is omitted. These missing variables (in PRMSSnow) are set to 1-D arrays of length nhru full of zeros with the exception of the through_rain. The through_rain varible matches the PRMSCanopy variable net_ppt when snow is not present and this is set using numpy ref/pointer to this array in the PRMSCanopy conveniently already existing in memory. All the dummy PRMSSnow variables are 1-D array inputs to the Adapter class.

THe submodel is run and its outputs are verified to match the outputs of the full model run, proving that the we can quite easily fake required inputs for model processes not provided elsewhere in the model or externally.

This example is intended to be run in the examples/ directory of the pywatershed repository.

[cneyens]

Thank you for the reply and the great example gist @jmccreight! I've tried this approach with our model and it works like a charm.