Scheduled Downtime
On Friday 21 April 2023 @ 5pm MT, this website will be down for maintenance and expected to return online the morning of 24 April 2023 at the latest

How do use model spin up for forecast data

William.Hatheway

William.Hatheway

Active member
I have searched the WRF forum for different reccomendations about spin up time for model simulations. It seems most cases are talking about case studies and not using forecast data (from the GFS).

So for an example, I want to run a 72 hour forecast over a region of interest. I have two nested domains at 9x9km and 3x3km. According to the best practicies the model spin up should be somewhere between 6 and 12 hours. But since the GFS only comes out every 6 hours and is initialized once how would i do model spin up? Should I use the previous run or should I combine the previous run up to the new run?

Input data: 00Z GFS forecast data to GFS f72 forecast data
Runtime: 72-hrs into the future
Model spin up: 6-12 hours
But GFS 00Z doesn't have a way to spin up that I can figure out.

So would this work?

Input data: 18Z GFS to 00Z GFS (for spin up) -> 00Z GFS to f72 GFS forecast data?

Any input would be helpful.
 
kwerner said:
Okay, well then I think the answer to your initial question about whether this would work is "yes." Give it a try and see!
Click to expand...
So here's a few more questions,

Here's my setup: 12Z GFS forecast data for 24 hours.

Would GDAS/FNL data from the previous run (6Z) for 6hrs be better for model spin up instead of the previous 6z GFS forecast data?

Does the GDAS/FNL data for forecast time f06 the same as the initial data for 12z f00?


I was trying to simulate milton and it never got the central pressure right and severely underestimated it

IS GDAS the same as FNL?

is there a way the see the ungrib data pressure values to see if GFS data using as input from noaa nomads is good data?
 
Last edited:
William.Hatheway said:
Would GDAS/FNL data from the previous run (6Z) for 6hrs be better for model spin up instead of the previous 6z GFS forecast data?
Click to expand...
I can't say for sure which would be better. GDAS Final analysis data would include any updates to the data, while forecast data is just the raw input, without error adjustments. You could experiment with each to see what gives the best outcome.

William.Hatheway said:
Does the GDAS/FNL data for forecast time f06 the same as the initial data for 12z f00?
Click to expand...
These two datasets represent the same time period, but there are likely differences since the large scale models were initialized at different times.

William.Hatheway said:
IS GDAS the same as FNL?
Click to expand...
Assuming when you refer to 'FNL' you mean GFS/FNL, these are not the same datasets. GDAS Final Analysis is the 0.25 degree Global Tropospheric Analyses and Forecast Grids, while GFS FNL is the 1 degree Operational Model Global Tropospheric Analyses. They come from two different models, both run by NCEP.

William.Hatheway said:
is there a way the see the ungrib data pressure values to see if GFS data using as input from noaa nomads is good data?
Click to expand...
You could use some of the tools mentioned in this presentation to look at the data (start with slide 16). I'm not certain, but I believe the data from the different sources (e.g., RDA vs NOMADS) are the same, but just provided by different entities.
 
I've actually never used GDAS data, so I can't comment on that. You can always try both and see which gives the best results.
 
Hello,

I use ICON-EU for model on my website but for various clients whatever is most appropriate or requested. In almost any cases I recommend to anyone to discard first 6 hours of WRF output, and for that period of time they just use previous WRF forecast instead of the most recent one.
 
meteoadriatic said:
/Hello,

I use ICON-EU for model on my website but for various clients whatever is most appropriate or requested. In almost any cases I recommend to anyone to discard first 6 hours of WRF output, and for that period of time they just use previous WRF forecast instead of the most recent one.
Click to expand...
So I have adapted a different way to run it.
An example
Forecast start date: 12/ 29/2024 00Z
Forecast end date: 12/31/2024 00Z

Model spin up time: 12/28/2024 18Z (6 hours before the desired forecast time)

So model runglength is 54h ours using the previous 6 hours of forecast time to spin it up and then the 00Z starts the time of interest.

What do you think at for that method? @meteoadriatic
 
I understand what you try to achieve but I can't figure why, instead of just discarding first 6 or so hours of forecast data? Unless there is some very specific scenario.... Your method seems reasonable but still has 2 downsides in my opinion:
1) You're initializing model with older data which is in theory less accurate
2) You're running 6 hours more than you need data for
 
meteoadriatic said:
I understand what you try to achieve but I can't figure why, instead of just discarding first 6 or so hours of forecast data? Unless there is some very specific scenario.... Your method seems reasonable but still has 2 downsides in my opinion:
1) You're initializing model with older data which is in theory less accurate
2) You're running 6 hours more than you need data for
Click to expand...

1. The data i'm initializing with is the GFS GDAS reanalysis, which is the 18Z GFS with observations assimilated into it and run again. The idea is that it is more accurate than the non-assimilated data. The reason i want to do this is so that I don't have to toss the 00-06Z data at the beginning of the time period of interest and instead i can toss the assimilated data that noaa has done already.

2. jI thought model spin up was 6 hour reccomended?
 
If you think that this data is more accurate than the most recent forecast dataset, which could very well be, then your method seems pretty reasonable.
From what I understand, the length of spinup depends a lot on number of time steps; let's say you start fine simulation and have time step of few seconds, then spinup will be say only hour or two because you will have a lot of time steps within that period. On the other hand if you ran climate simulation with 30 km grid, you will probably need several days of spinup because of small number of timesteps in particular simulated time. I hope that makes sense :)
 
meteoadriatic said:
If you think that this data is more accurate than the most recent forecast dataset, which could very well be, then your method seems pretty reasonable.
From what I understand, the length of spinup depends a lot on number of time steps; let's say you start fine simulation and have time step of few seconds, then spinup will be say only hour or two because you will have a lot of time steps within that period. On the other hand if you ran climate simulation with 30 km grid, you will probably need several days of spinup because of small number of timesteps in particular simulated time. I hope that makes sense :)
Click to expand...
That does, i guess I need to test the method and report back and examine some local weather vs what the model says using the two different methods.

My thought is if I am tossing the first 6 data points anyways for using the most recent data, due to spin up, then the previous run with data assimilation should be the one to toss it. I don't know it's an interesting thought
 
@ William: I'm not sure I understand the problem. 0z GFS output files become available at ~ 4z. Add your own WRF forecast runtime and the spin-up period is already in the past (I think the model only needs significantly more than 6 hours for spin-up in severe convective cases depending on resolution, timestep and so on, but you can also experience with the DFI option to see if it reduces spin-up time in your case).
 
DanielF said:
@ William: I'm not sure I understand the problem. 0z GFS output files become available at ~ 4z. Add your own WRF forecast runtime and the spin-up period is already in the past (I think the model only needs significantly more than 6 hours for spin-up in severe convective cases depending on resolution, timestep and so on, but you can also experience with the DFI option to see if it reduces spin-up time in your case).
Click to expand...
I think I see your point, since the GFS data comes out at ~4Z any output from WRF using that data would already be in the past. Therefore tossing the first 6 hours would have the spin up issue resolve. See model spin up is not talked in depth in the tutorials so I am still trying to grasp it.

Also what is DFI? I don't recognize that acronym @DanielF
 
You must log in or register to reply here.
Top