4-KM WRF Model Is Great, When it Works
In "Severe Weather Forecasting Secrets: Pt. 2: The WRF" I sung the praises of the 4-KM WRF forecast model for forecasting storms in your area up to 48 hours in advance. The model has the ability to predict individual thunderstorms, and usually does a good job. In that blog entry I said: "I've been impressed with the 4-KM WRF here in PA this Spring. If it says that storms are going to break out in a region of the state, they usually do, and there have been cases where it looks very similar to the radar map for the forecast time."
I wanted to "prove" that comment by showing you a remarkable example of the model correctly predicting storms here in Pennsylvania last Friday 12-24 hours in advance (but I also wanted to show an example of the model busting). So without further ado, here are screen shots of the actual radar (left) and simulated WRF radar from the model run the night before (at right) for the storms that we had Friday. Because the radar is in Central time and in 30-minute increments, any of these screencaps are +/- 1 hour. The first shot is at 8 AM:
The model predicted a large stormy area over Lake Erie and is hinting at stuff starting to pop up in NW PA, which is happening, although not as aggressively as the model predicted. It tagged the storms in Delaware but was too far south with the New Jersey storms. Now, we flip to noon.
Incredibly (remember this forecast was made 15 hours ago) much is unfolding exactly as the model predicted. A line of storms is developing on the ridges of south central PA. Three other lines of storms are placed to the west (it underestimated the strength of these though). Scattered storms are stretching from northern New Jersey southward into Delaware (it missed the line developing on the eastern shore of Chesapeake Bay). The large area of heavy rain is moving onto Lake Erie's eastern shore. Storms are popping up east of Rochester, NY. In general, if you were using this hour of the model to plan outside activities or setup for a storm chaser, you'd be doing pretty well.
At 3 pm, it has the general idea: A strong line of storms stretches from Rochester through Central PA (the model has it too far east and separated from the rain shield coming off of Lake Erie), and a secondary line still exists behind it. What it overestimated here was the coverage of the storms in Eastern PA and New Jersey, which are still few and far between; it has generated the storms on Chesapeake Bay now, so if you expected those "early afternoon" you're still happy.
By 6 PM (and remember, this is one day in advance) it starts to have some trouble with the speed of the line of storms. It has picked up on the dissipation of the first line of storms (which would have been helpful to know if you saw that radar at 3 and got excited in Eastern PA), but has been too aggressive on forming two additional lines behind it. Still, the precipitation is in generally the correct areas. After this, the model wants to hold the rain shield back in western PA when in reality it moves forward, as you can see at the 11 PM shot:
Now here's an example of the WRF doing a terrible job, misleading Pennsylvania weather enthusiasts by stating that the storm would be covered by homegrown individual thunderstorms late in the afternoon, when in reality all we had was a shield of light rain (this was yesterday). Again this is the 00Z run from the night before on the right, radar on the left.
At 2 PM (shown above), the model has storms in the general area, but none are pinpointed and it has the light rain in western PA too far north. It says there are no storms in the state at all, missing the development off of eastern Lake Erie and over the ridges of the Laurel Highlands in SW PA, where it typically does a good job. This could be a sign that things aren't going well.
Two hours later, it's still hasn't picked up on the ongoing storms in NW PA, and has missed the storms in eastern Jersey. Still there is some hope becuase it has the storms in SE NY and northern Virginia in the right place. Unfortunately it has storms in the SC & SE part of PA that aren't there.
At 7 PM, it really starts to fall apart. The model is predicting widespread individual thunderstorms over most of the state that aren't happening; instead we have a light rain shield over much of the area and nothing at all happening in the eastern part of the state.
Three hours later, the general shape of the rain shield is correct, but it's in the wrong place and the individual storms never developed inside of it, which is something the model should be good at predicting. It's also raining heavily over West Virginia and Ohio, which should have been cleared out according to the model. NOTE: Sometimes it's a case of timing -- if you run the model and radar loops the storm areas may look similar but the model is off by a few hours. Sadly, this was not the case here, as the predicted storms never developed.
So what does this mean if the model can be sometimes remarkably accurate but sometimes wrong? Like all models, the WRF can pick up on incorrect atmospheric signals that can make a forecast bust, but I still think that the WRF wins more than it looses, and the high-resolution that enables you to see storms still makes it more useful than other models. The situation outlined above, which is a complete bust, is much more rare, I believe. I have no empirical evidence of this; this is just based on my observation of the model this Spring and Summer here in Pennsylvania. Certainly if the model is doing strange things early in the day, there is probably not much hope later.
By the way, you may be wondering: Did yesterday's 12Z model run correct this bad behavior? I didn't talk about the 12Z runs here because they don't come out until mid-afternoon, which is too late for most planning purposes. Sadly in this case, it did not. The model still predicted widespread storms that did not happen, in fact they were stronger. Below is the 6 PM radar versus prediction:
