Model Spinup

[Michael Wilkinson]

In the Forecasting chapter of Tim Vasquez’s book, Storm Chasing Handbook, (p.119) there’s mention of “model spinup” time.

I’m not sure that I understand the concept since the way I’m interpreting it makes it very counterintuitive to me.

Here’s a quote from the book, “within a couple of hours of the model start time, model accuracy is poor due to the effects of ‘model spinup’, the temporary mathematical noise which quickly dampens after the first few hours. This can produce spurious artifacts in the model charts.”

It sounds like that a model such as the HRRR is producing more accurate forecasts for a few hours in the future than it is for an hour in the future.

Does that mean that the model is doing something different in obtaining an hour 4 forecast than it is an hour 1 forecast?

Is the hour 4 forecast based on the model’s hour 3 forecast which is based on the hour 2 forecast which is based on the hour 1 forecast? If so, how does the hour 4 forecast become more accurate than hour 1?

Where can I find a beginner level, in-depth explanation of model spin up?

Tim’s book was published in 2009. Some resources I found online that mentioned model spin up were from 2011 and 2015. Is model spin up still a factor to consider with the newest versions of weather models?

 

[Paul Knightley]

Others will be much better placed than me to comment on this, but I think it (at least partially) depends on how the model is started up - if there are no spin-up processes in the initial condition (e.g. various motions) then this might be an issue - but if the model starts with these processes already 'spun up', it eliminates it.

 

[Jeff Duda]

Model "spin-up" refers to the initiation of grid-scale circulations that are not present in the initial conditions of a model run. Nearly all high-resolution model runs are initialized from a coarser-resolution field, which means the circulations that the model grid can represent do not exist in the initial conditions. "Spin-up" refers to the process of those circulations coming into being from nothing. The larger the difference in grid spacing between the source of the initial conditions and the model resoution, the longer the 'spin-up' process takes. If you can start a model integration using initial conditions on the same scale as the model domain, then there is essentially no spin-up.

The reason the HRRR, in particular, does not deal with "spin up" problems, is because it uses the output from a 1-hour 'pre-forecast' that does have spin-up issues as its own initial conditions. This helps alleviate the traditional-spin up problems. Other models do not do this.

 

[Tim Vasquez]

One of my readers mentioned the discussion so I figured I'd take a look. I haven't been on the forum much lately due to all the data projects I have going on.

This is a good discussion and Jeff and Paul posted some good info. I'll likely update all this in the next version of Storm Chasing Handbook, as much of the older text is about 15 years old. I'm planning on at least one more update for next winter (Storm Chasing Handbook 2020) and that may be the final edition.

 

[Steve Miller]

Tim, I'm thrilled to know there is another version on the way! That book and your others were an important part of my early chase days.

 

[Paul Knightley]

Tim, I'm thrilled to know there is another version on the way! That book and your others were an important part of my early chase days.

Agreed - indeed, I'm about to start my annual habit of reading the book, to really get me in the mood!

 

[Michael Wilkinson]

I’m very happy to hear about the 2020 edition. I will be among the first to purchase it.

The book has greatly helped my understanding of the behavior of the atmosphere. Last year, before my first real serious effort at chasing, I watched Rich Thompson’s lecture series on YouTube. This year, I’m in the process of watching it again. Because I’ve read Tim’s book, I’m getting a lot more value out of the lectures. I’m finding that I’m understanding concepts that I didn’t have any frame of reference for the first time I heard them.

I highly recommend the book.

 

[Jeff House]

Also why some days the late morning runs do better than the early afternoon runs. Really we need to be nowcasting ourselves after lunch anyway. I try to cut myself off models around 15-16Z. But I sneak in the 18Z on late day setups.

 

[Jeff Duda]

Also why some days the late morning runs do better than the early afternoon runs. Really we need to be nowcasting ourselves after lunch anyway. I try to cut myself off models around 15-16Z. But I sneak in the 18Z on late day setups.

That generally should not be the case. What I suspect is that on some more critical days some later HRRR forecasts just happened to perform more poorly in your specific area of interest than a previous forecast that morning. Reflectivity and precipitation verifications do not bear out this result in any systematic way.

Feel free to provide me with some case studies illustrating this. As I work on the HRRR and closely with the verification staff, I would be very interested in looking deeper into cases when this occurs.

 

[Jeff House]

Hi Jeff. Like you said the HRRR does not really have the problem. I concede it's anecdotal, and mostly NAM. With the information you share, I will be more confident in later runs esp the HRRR. Thank you!

Still, for everyone, good old fashioned nowcasting is required. Model is just a tool.

In 2016 we blew a TX PH tornado that had showed up on the 14Z HRRR (supercell) but not later runs. Boundary had washed out on surface and visible, so it was probably nebulous to both humans and machines. However Storm Track members documented a tornado in the location.

I also use it for work, but again I do not have the data Jeff D does. Moving forward I will trust it again.

 

[Ashwin D]

@Jeff Duda in case the global model resolution is not the same as the mesoscale model resolution is there a way to calculate the approximate spin up time it takes so that those frames can be discarded ? So in my case it is exactly one half. The global model resolution is 13 kms. The mesoscale model is 7 kms. Usually I discard the first few hours of data and then use the rest.

One thing that I do want to add is that the temporal resolution of the two is also not identical. The global model outputs data every 3 hours. The mesoscale model can be made to output data every 15 minutes. As we all know a lot can happen in meteorology in 3 hours(The surface pressure can drop dramatically in that period). Does this affect the quality of simulations ?

 

[Jeff Duda]

First of all, output frequency is not the same thing as temporal resolution. You need to consider the model time step to ascertain the latter. I do not know what the time steps are for various global and mesoscale models, but I know they are near the limits of what computer technology can allow for good forecast output latency.

To determine spin-up time, you would need to calculate kinetic energy spectra and see how long it takes for the near-grid-scale energy to approach the proper spectral slope. In the synoptic scale and above, kinetic energy follows a k^-3 spectrum (for wavelength k), whereas in the mesoscales and below, we have found that the KE spectral slope is k^-5/3. However, all models have filters near the grid scale which decrease KE starting around 8*delta_x to 4*delta_x, so you should not expect to see k^-5/3 near the grid scale.

In my previous work initializing 3 km WRF simulations off of 12 km NAM analyses, spin-up was mostly complete by about 3 hours. But again, the time it takes depends a lot on the scale separation between your IC data and your model grid spacing. The bigger that discrepancy, the longer it will take.

 

[Ashwin D]

First of all, output frequency is not the same thing as temporal resolution. You need to consider the model time step to ascertain the latter. I do not know what the time steps are for various global and mesoscale models, but I know they are near the limits of what computer technology can allow for good forecast output latency.

To determine spin-up time, you would need to calculate kinetic energy spectra and see how long it takes for the near-grid-scale energy to approach the proper spectral slope. In the synoptic scale and above, kinetic energy follows a k^-3 spectrum (for wavelength k), whereas in the mesoscales and below, we have found that the KE spectral slope is k^-5/3. However, all models have filters near the grid scale which decrease KE starting around 8*delta_x to 4*delta_x, so you should not expect to see k^-5/3 near the grid scale.

In my previous work initializing 3 km WRF simulations off of 12 km NAM analyses, spin-up was mostly complete by about 3 hours. But again, the time it takes depends a lot on the scale separation between your IC data and your model grid spacing. The bigger that discrepancy, the longer it will take.

Yes thank you very much for the in depth post.

Regarding your first point you are absolutely correct that output frequency is not the same thing as temporal resolution. What I meant was the following
1) Global model time step is usually around 15 minutes if I am not mistaken. Mesoscale model time step is around 5 minutes.
2) Global model output is usually around 3 hours.
3) One can interpolate the global model 3 hour output to create a time series of say an interval of 5 minutes or 15 minutes.
4) But will the interpolated time series input to the mesoscale model output contain meaningful information on that scale ?

 

[Quincy Vagell]

Sometimes, the HRRR is not particularly useful in the afternoon during a chase, as it may be struggling to resolve ongoing convection. There are also cases in which the HRRR is blowing up robust convection at hours 1-2, but nowcasting suggests otherwise. I know the HRRR has gotten better over time and the high reflectivity bias has been improved.

One specific example is the HRRR consistently blowing up a discrete supercell (near a quasi-triple point) over southwestern Kansas on March 27th of this year. The atmospheric setup was marginal and it wasn't going to take much of a model forecast error to cause such a prediction. (convection vs. no convection) The verification was that some transient/marginal supercells did briefly form over the Texas panhandle, while a small shower that quickly dissipated was all that managed to form in Kansas. Interestingly enough, only the 12z model run backed off on the supercell, while most of the remaining morning runs continued to show it.

The HRRR isn't perfect. Yesterday, it had some similar issues resolving a decaying MCS in central Oklahoma and how it thought the atmosphere would quickly recover, allowing for discrete storms to form by afternoon. There was partial air mass recovery, but not to the extent that the morning runs had implied.

It's meteorology over modelogy. It helps to know model biases and to be able to qualitatively assess initial mesoscale conditions to see if model output makes sense, given current/near-term observations.

 

[Jeff Duda]

Yes thank you very much for the in depth post.

Regarding your first point you are absolutely correct that output frequency is not the same thing as temporal resolution. What I meant was the following
1) Global model time step is usually around 15 minutes if I am not mistaken. Mesoscale model time step is around 5 minutes.
2) Global model output is usually around 3 hours.
3) One can interpolate the global model 3 hour output to create a time series of say an interval of 5 minutes or 15 minutes.
4) But will the interpolated time series input to the mesoscale model output contain meaningful information on that scale ?

Temporally, yes, as mesoscale processes don't evolve fast enough for a 15-minute interval to severely hamper or truncate mesoscale-resolvable features. Spatially, however, the input to a mesoscale model from a global model would be missing all features below the scales resolvable by the global model. Granted, many mesoscale processes are driven by synoptic scale processes that are resolved on global model grids, but not all of them (espeically those that are orographically forced). Anything related to gravity wave drag would also likely be missing in the mesoscale model input.