Ramblings of a Graduate Student

Warning: Today’s post is a long one.

Imagine for a moment that you were asked to predict how your best friend would react in any given situation.  Chances are that if you knew this person for any appreciable amount of time you would have a sense of how your friend reacts to situations when he or she is tired, happy, sad, etc.  Chances are you would also have a sense of how your friend typically reacts to the hearing of good news, bad news, etc.  Combining all this with being able to read your friend’s body language, I would suspect you could make a fairly accurate prediction.

Now, what if your best friend was invisible?  You’d have to rely on second hand information, or in other words, information you did not observe yourself, in order to make that prediction.  Even though you might be able to make the correct prediction most of the time, I’m willing to be you’d be wrong more times than you were in the original scenario.

In reality, meteorologists are asked to do the latter on a daily basis.  We are asked to forecast an atmosphere that we have studied for years, but cannot directly “see”.  All of the information, or guidance, we have is second hand.  It comes from instruments that observe things for us.  It comes from computer models that are solving mathematical equations.  It comes from conceptual models that we learn about and develop through our studies. And, to a certain extent, it comes from our gut.  The best forecasters cannot always explain why they made the forecast they made; sometimes they just have to rely on their gut.

So, what happens when all of this guidance contradicts itself?  Or when the guidance is continually changing?  As you can imagine, forecasting the weather under these circumstances is extremely frustrating. This is exactly the type situation that forecasters in the area from Oklahoma City, Oklahoma (OKC) to Tulsa, Oklahoma (TUL) to Kansas City, Missouri (MKC) to Saint Louis, Missouri (STL) to Little Rock, Arkansas (LIT) back to OKC are facing today.

The sequence of images that follow are known as a D(prog)/Dt.  In words, this is a sequence of images that show how a numerical forecast (prog) of a given valid time has changed over the past few model simulations.  Specifically, in this case, all images below are forecasts for 03 UTC 5 April 2010 (10 PM CDT, 4 April 2010).  The first image is an 84-hour forecast, the next is a 78-hour forecast, followed by a 72-hour forecast, and so on to the last image, which is an 18-hour forecast.  Again, all images are forecasts for (in other words, “valid at”) 03 UTC 5 April (10 PM CDT, 4 April 2010).

Each image is a forecast of the 1000-500 mb thickness (blue dotted, except for 540 [green] and 558 [black]), 700-500 mb omega (thin contours), and accumulated precipitation over the previous three hours (color-filled).  These images come from an ensemble forecast system, meaning the output is generated from a combination of a lot of models that were run at the same time.

Imagine that it is Thursday evening (03 UTC / 10 PM CDT) and a forecaster on shift is trying to predict the probability of thunderstorms over the aforementioned area for Sunday night into Monday morning (essentially a 72-84 hour forecast).  The image above is an 84-hour forecast, meaning the model was run at 15 UTC / 10 AM CDT that morning.  Based on this guidance, he or she would be hard-pressed to think precipitation would occur, so the would most likely be for 0% chance of precipitation.

Still on shift Thursday evening, a forecaster would have access to yet another model forecast; those that were generated at 21 UTC / 4 PM CDT that afternoon (above).  This would give a forecaster even more confidence that precipitation isn’t very likely Sunday evening into Monday morning.

Model guidance from 03 UTC / 10 PM CDT Thursday evening (above) continue the trend of having no precipitation over the region on Sunday evening into Monday morning.  At this point a forecaster would be feeling extremely confident with a no precipitation forecast!

The next set of model guidance (above; 09 UTC / 4 AM CDT Friday Morning) begins to put doubt into the mind of a forecaster.  Although not a lot, nor is it widespread, it does appear that this model run begins to hint at the possibility of thunderstorms developing after sunset Sunday evening.

With the 15 UTC / 10 AM CDT Friday morning guidance (above), a forecaster would begin to really start to worry about the possibility of thunderstorms developing in the OKC to TUL area.

After examining the 21 UTC / 4 PM CDT Friday afternoon guidance (above), a forecaster would most likely begin to feel fairly confident that thunderstorms would be possible after dark Sunday evening and issue a 30-40% probability of thunderstorms with the thought that these probabilities would need to be increased as the event draws closer.

That evening (03 UTC / 10 PM CDT Friday) a forecaster would be less certain of thunderstorms in the TUL area, even less certain in OKC, and begin to worry about thunderstorms in the MKC to STL areas.  He or she would also begin to have concerns about thunderstorms developing in far SE OK and moving into LIT area overnight.  This is a 48-hour forecast and models typically do well with guidance in this time frame.

The 09 UTC / 4 AM CDT Saturday guidance continues the trend of developing thunderstorms in SE KS and MO, and trimming back on the thunderstorm chances in the OKC to TUL area.  However, models now indicate that there will be two separate areas of thunderstorms.  One stretching SW to NE from SE KS into northern Missouri with heavy precipitation and another area from far NE TX into NE AR.

The 36-hour guidance (15 UTC / 10 AM CDT Saturday) is reasonably similar to the previous guidance and so forecasters in MKC, STL, and LIT would have fairly high confidence of thunderstorms after dark on Sunday.  Forecasters in TUL might be worried about the possibility of thunderstorms developing in far NE OK early in the evening, and forecasters in OKC would have increasing confidence with once again going with a dry forecast.

The 30-hour guidance (21 UTC / 4 PM CDT Saturday; above) continues the trend of two main precipitation axes: one from SC KS northeast into northeast MO and another area in Arkansas.  At this point confidence would be fairly high in precipitation in MKC, STL, and LIT; confidence would be rather low in OKC; and a forecaster in TUL would be pulling his or her hair out.

After looking at the 24-hour guidance (03 UTC / 10 PM CDT Saturday; above), almost all forecasters would be pulling their hair out.  The area of heavy precipitation in the MKC area has now shifted northeast into NE MO and IL, a signal for precipitation has once again developed in the OKC and TUL areas, and the signal for precipitation in LIT has decreased some and/or moved into Tennessee.

Finally, this morning’s guidance (above; 18-hour guidance; 09 UTC / 4 AM CDT Sunday) continues the trend of moving the heaviest precipitation into NE MO and IL, increasing the precipitation signal in the OKC and TUL areas, and now reintroduces heavier precipitation in the LIT area.

Remember, all of these changes occurred over areas in which this suite of numerical models was not producing precipitation 72 hours ago!  Determining why the precipitation signal in NE KS shifted into NE MO and IL is fairly obvious at this hour; previous model forecasts were too slow in developing the precipitation.  Thunderstorms ended up developing earlier in the MKC area and will be in NE MO and IL after dark.   In other words, in this area, the model had to play catch up.  As to the other areas?  A quick perusal of regional radars indicates that there are no areas of precipitation nor are there any areas where precipitation is imminent over the OKC, TUL, and LIT areas.

So, in a more philosophical sense, why did the model forecasts change so much over the last 84 hours?

The honest answer lies somewhere between the fact that our models use too coarse of resolution to always be correct and the fact that our understanding of the atmosphere combined with our ability to create mathematical equations to represent the atmosphere is not as good as we would like.  Tangentially, this touches on the concept of “predictability”.  When a vast majority of our numerical models predict the occurrence of a similar event and/or the D(prog)/Dt shows relatively similar forecasts, we say the event was “highly predictable” and we have higher confidence in our forecasts.  When our models forecast different events and/or D(prog)/Dt shows constantly changing forecasts, we say the event had “low predictability” and our confidence is much lower.

The thunderstorm forecast for tonight has low predictability and anyone’s confidence in any single forecast is rather low.   However, the benefit of forecasting is that the answer will reveal itself in due time.