Tuesday, May 28, 2019

Tools for Crucial Weather Routing: With an ongoing comparison of GFS and FV3-GFS.

We will add new data at the end here until we figure out which model wins out!

It can happen that crucial decisions depend on details of a forecast. Usually this will be for a longer term forecast, since one or two days out are generally pretty good.  We have a specific case at hand now, which is a classic, real-world example.  The vessel has two optional routes, on each a crucial turn must be  made roughly 96 hr from now, meaning we are in interested in what happens on Friday, May 31.

The two main primary sources we have are the GFS and the new FV3 GFS, which will replace the former on June 12.  We see below that these two do not agree at all for Friday.

Again, Friday shown here is 4 days out from these  Monday 18z forecasts.  The displays are the meteograms from LuckGrib.  The standard GFS we have used for years shows roughly 20 kts from the east, or just above east.  The in-principle-better model FV3 GFS shows 30 kts from the SE.  For the vessel at hand here, and almost all vessels for that matter, this is a huge difference.

These LuckGrib meteograms are plotted with the vessel following a specific course at a specific speed. Thus the output applies to the time and location of the vessel.  Below we see these two data images with the vessel in the location shown by the blue line in each.

For a specific comparison at 12z on Friday, the GFS is 18.3 kts at 077 whereas the FV3 GFS is 27.5 kts at 139.

So what can we do faced with this? GFS itself offers no probabilities or other qualifications.  We have in our textbook Modern Marine Weather several ways to evaluate a forecast using the 500-mb maps and other means, but this situation is rather different from the examples in the book as this is for the Southern Hemisphere, off the coast of Queensland.

We got the above data from the powerful LuckGrib app, which offers many model options and very neat presentation of the data. Two models offered in LuckGrib that help us look at probabilities are the standard deviation outputs from the Global Ensemble Forecast System (GEFS) model and the probabilistic wind forecasts from the Oceanic National Blend of Models (NBM).

Below are the same meteograms (route and dates) for these two models, followed as above with the actual plot of the wind showing values at 12z on Friday.

And now the specific values at the vessel at 12z Friday.

These data require a bit of explanation.  The GEFS ensemble data comes from starting a computation with a base model forecast, called the Control run, which is similar but not identical to the base GFS forecast. Then they repeat this forecast 20 times, each time making some likely variation, such as changing the input data or trying different physics solutions. The idea is that the spread in these results is a measure of the uncertainty in the forecast. We could actually look at each of the 20 runs to see how widely they vary, or as LuckGrib has decided we look at the mean values of all of the runs along with the standard deviation of this set of runs.

The standard deviation (SD) is a measure of the spread in uncertainty assuming the different runs introduce more or less random differences. In the most basic sense, SD is defined in statistical terms with the curve below, from our book Mariners Pressure Atlas, in which we use these ideas for storm forecasting in the tropics based on pressure values.

The basic idea is the larger the SD, the broader the distribution of winds, with 68% of the observed winds expected to be within 1 SD of the mean, as shown below.

In the presentation of this GEFS data, we might expect the forecasted spread of wind speeds to be the mean wind ± 1 SD, but this is not what we see. Referring to the actual output of the GEFS (above) we see

which is not the same as 20.5 ± 6.2, which equals 14.3 to 26.7 kts, instead we see 16.2 to 25.5 kts.

This is a subtlety that might not matter too much in many practical situations, but since this output itself is a new concept, it might help to know the source of the apparent discrepancy. The issue is the GEFS model presents the SD as a vector quantity, meaning they do not give it for the wind speed alone, but rather they give separate values for the E-W component and the N-S component of the wind vector. Thus we have a situation like the one shown below.

In this sketch, the SD is called "sigma," shown as a dashed line.   u is the E-W SD and v is the N-S SD. The quoted value of SD in LuckGrib is sqrt (sima-u^2 + sigma-v^2).  The red vectors mark the spread of the wind speeds, and the blue lines mark the spread of wind angles reported.

Here is the latest data from Tuesday morning:

First we see that the GFS and FV3 GFS are now closer with the former at 14.3 @ 105 and the latter still higher at 22.1 @ 123.  The directions are closer, but speeds are still notably different at 3.5 days out.


The GEFS give us a spread of 15.0 to 23.5 kts, with directions varying from E to SE. So according to the ensemble analysis, this forecast is still up in the air.

The NBM calls for a mean wind of 21.1 kts from the SE, with 25% chance of being > 23.1 kts and 25% chance of being <17.1 kts. Again, no reassurance from this at all.

That is where we stand. No confidence for crucial decisions on or near Friday.

I will contact our meteorologist friends in Australia to see if they have any insights on what is taking place and what they think might happen.  This could all be tied to pressure systems moving across the continent of Australia which they will likely understand.

In the meantime the vessel (Jacob Adoram's Emerson) is going as slow as possible waiting for some guidance in the forecasts.

With the GEFS data we were only looking at the standard deviations, but this model also computes probabilistic winds, but the public presentation does not reach down to Australia.

Following on to see how these forecasts pan out...

======= New data from Wednesday morning =======

Update times from 06z to 18z, as explained below.

These two models are definitely closer now. We will also see  below that the ECMWF is right between these two forecasts.  We still have the FV3 calling for much stronger wind on Sat than GFS.

We can compare these to the ECMWF, which can be read from windy.com.

This is the summary:

GFS 13.6 @ 105
FV3 17.0 @ 120
EC    16 @ 105
GEFS 18.2 @115  with range  15.3 to 22.2 @ E to SE
NBM 21.6 @ 132  with a 25% chance of < 17.5

So both the probabilistic models (see below) cover the predictions of the 3 direct models.  The NBM looks to be the most off, but we have to keep this in perspective. The NBM for US waters updates every hour, but this oceanic version updates on a strange schedule. There is a new run at 00, 05, 07, 12, 17, and 19z. It then has a short latency and is available in about 1 hr after each run.  The GFS, for comparison, is run every 6 hr at the synoptic times, but it has a total latency of about 5 hr. GEFS has a latency of about 6 hr and FV3 is about 7 hr. So in principle, the Oceanic NBM could be the newest data... but this remains unclear till we learn more. It is blending models that are much older, so the full implications are not clear yet.

Now that we have this data so conveniently from LuckGrib, the next step will be to learn more about how to use it, which in a sense we are struggling with at the moment.

We will continue to add  data when we get it. In the meantime, we look at the overall weather patterns that are leading to this in another post.

====== Thursday morning ==========

For our 12z Friday test case and adding a Sun 12z new test case, we have:

FV3 19.9 @ 093       Sun 12z = 20 @ SE
GFS 16.1 @ 107       Sun 12z = 22 @ SE

EC1 17.5 @ 090       Sun 12z = 14.5 @ SE
EC2 23.5 @ 132       Sun 12z = 19.7 @ 107

We do not have direct access to ECMWF data as this is licensed data (ie the UK sells it), but we have two ways to access it, at least in principle.  We have an account with PredictWind, which offers a model called PWE that is based on the ECMWF data and we have the windy.com output, which they claim is the ECMWF. These are shown below.

Again, this is AU time, so 12z on Fri is at 22z, which is 17.5 @ E, with a GFS report of 15.5 @ 110, which is close to what we read directly for the GFS. The EC forecasts 20+ winds all the way through next Tuesday.

This PredictWind version of EC that we access via the nav program Expedition. This gives 23.5 @ 132 for Friday.  These two presumed sources of EC wind are so different we have to conclude we do not know what the official EC forecast is. Below are the actual numbers for the PredictWind version:

30-May-19 0300,,107,7.2
30-May-19 0600,,120,10.0
30-May-19 0900,,118,13.2
30-May-19 1200,,115,15.9
30-May-19 1500,,121,16.8
30-May-19 1800,,138,18.4
30-May-19 2100,,136,20.7
31-May-19 0000,,129,19.7
31-May-19 0300,,122,19.2
31-May-19 0600,,132,21.4
31-May-19 0900,,131,23.8
31-May-19 1200,,132,23.5  (windy.com value 17 E-SE)
31-May-19 1500,,130,23.2
31-May-19 1800,,127,22.9
31-May-19 2100,,121,23.6
01-Jun-19 0000,,111,23.3
01-Jun-19 0300,,107,22.3
01-Jun-19 0600,,110,22.4
01-Jun-19 0900,,118,25.3
01-Jun-19 1200,,121,27.1
01-Jun-19 1500,,121,26.0
01-Jun-19 1800,,123,25.6
01-Jun-19 2100,,118,26.6
02-Jun-19 0000,,111,25.1
02-Jun-19 0300,,111,21.5
02-Jun-19 0600,,112,21.9
02-Jun-19 0900,,111,21.5
02-Jun-19 1200,,108,20.0  (windy.com value = 15 SE)
02-Jun-19 1500,,105,18.8
02-Jun-19 1800,,109,19.0
02-Jun-19 2100,,114,18.6
03-Jun-19 0000,,127,18.2
03-Jun-19 0300,,134,19.3
03-Jun-19 0600,,136,20.1
03-Jun-19 0900,,138,22.4
03-Jun-19 1200,,142,25.0
03-Jun-19 1500,,153,26.7
03-Jun-19 1800,,156,26.8
03-Jun-19 2100,,162,28.0
04-Jun-19 0000,,160,27.8

With these model forecasts at hand (GFS, FV3, and some sort of EC) we can look at the probabilistic winds from this morning:

We see here a still surprising result, namely for tomorrow, less than 24h away, the uncertainty is still large. We have mean forecast of 17.7 @ 111, but with a spread of 15.4 to 20.4 over E to SE.  This range is significant to almost all sailboats, 15 kts vs 20 kts.

We get a bit more specification from NBM. A median wind of 19.3 from the SE, but only 10% chance of less than 14.7; 25% chance less than 16.2. So low wind side is unlikely, whereas we still have a high wind side.

At this point, I do not see any justification for gambling on the lighter side of the forecast, and indeed we should likely count on more rather than less.

With all this said, and still more to say, we can pause an look at how these models have done so far in predicting the winds that Jacob is actually reporting and we can measure with ASCAT.

This article will soon be a live link:  Compare ASCAT, Observations, and Model Forecasts.

Monday, May 27, 2019

Compare ASCAT and WindSAT Scatterometer Wind Data

We have two sources of scatterometer data these days, the 3 Metop satellites with ASCAT data and the US Navy Coriolis satellite with WindSAT data. The latter does not get much mention in official NWS forecast discussions, whereas ASCAT is referred to frequently. The reason for this is not clear, so in  a first step toward trying to understand this, we will compare the data whenever we can spot passes at about the same time, over the same region.  It could be we can calculate that conjunction of passes, which if so we will add to our satellite prediction time article, which is underway.

The OSWT site presents the ASCAT data in a 10º x 15º Lat-Lon grid and the WindSAT in a 20 x 30 grid. They both have in principle 25-km resolution, so the reason for this is not completely clear. It could be simply that the single WindSAT data swath is about twice as wide as one of the halves of the two ASCAT swaths, so it would likely take a different file layout to account for it.

If the WindSAT data are consistent with ASCAT, then this would be to many mariners a "new" and useful tool, because it has a wider swath of data without a nadir gap like ASCAT has. In that sense, the WindSAT scatterometer is more like the old QuikSCAT that provided so much valuable data before it failed—long after its design lifetime.

Below are several examples, the newest will be at the end where we are adding them. There is there a May 31 example that shows a larger difference than earlier ones... and we know what the answer is!

Example 1. Time difference 3h 05m

Figure 1. ASCAT valid 1153z on May 27.

Figure 2. WindSAT valid 0848z on May 27, about 3 hr earlier than the ASCAT pass

The data are not exactly at the same time, which has to be considered, but the reported winds are generally similar, though not at all identical. Yellow means above 20; green means below 20. Blue means below 15.  

So from this comparison we can say, if we assume the ocean had not changed, the wind data agree to within about ± 2.5 kts, which is all it takes to change the colors of the barbs. There is a yellow patch in WindSAT that is green seen in ASCAT, but the the green is a dark green, which is closer to yellow.  

In the next two examples, we put position markers at the colorbar wind-speed boundaries on one image and then read the wind at the corresponding positions on the other image.

Example 2. Time difference 3h 04m

Example 3. Time difference 3h 26m

Several Examples to be inserted here.

Example x. Time difference 2h 58m
This is one of particular interest, because we have been tracking Jacob Adorams's  ocean rowing vessel Emerson in this region at this time... and we know from him what the actual winds were at the time.  It is also the case where we see the largest difference so far between ASCAT and WindSAT.

Here is where Emerson was located (yellow stared marked) at the time of these passes:

And here is what the actual winds were on this day:

At the time of both satellite passes, the wind was 20 kts from the SE.

We are looking specifically at the region in the bottom of the outlined section, where ASCAT reports 20-22 kts from the S-SE (152T), whereas WindSAT shows rain contamination over this region but wind vectors from many directions. The ASCAT is consistent with observations well within the uncertainty of the observations.

At first glance this seems a big difference, but the WindSAT is more sensitive to rain than ASCAT and there are indeed many rain squalls in this region at this time.

To give WindSAT the benefit of the doubt, at first we thought maybe WindSAT was trying to show the unusual observation we got from the boat of 10 kt winds from 070 for Friday morning and much of Thursday.  But on the 30th, we could spot a local High in the region that would have accounted for this. It showed up nicely the BOM "gradient wind" streamline maps—we will be discussing these in another blog post.

With that in mind, we looked to the surface analysis maps shown below; one before these two satellite passes (18z), and one afterward (00z, the next day).

The magenta region shows the overlap in scatterometer data. We see the gradient actually weakening in the southern part where the winds differ most during this period, whereas the later ASCAT shows higher winds. Keep in mind this is SH, winds going counterclockwise around the High and out of the High into the Low, meaning this wind direction is SE, tending toward S-SE.  

In short, the maps do not support the WindSAT observations.  At this point we need to ask the experts what they can tell us about this. 

We will report back here when we learn more... and we will add several other examples above this one. They are done, just not posted yet.

Friday, May 24, 2019

Compact Day of the Year Calendar

There are several occasions in navigation when we need to know the total number of days between two dates, or a new date after so many days from another date.

One example is when planning a long distance coastal or ocean route which yields the total duration, and we know the starting date and time, and want to know the date of arrival.

Another example is when figuring the watch error in cel nav for a watch set some time ago. We might know the watch gains 0.335 seconds every 10 days and it was set on April 14. It is now July 22 and we want to know the watch error.

We could of course start out with "Thirty days has September, April, June and November..." to find number of days for each month (if we did not know this already), and then count out the days. But we don't want to make mistakes at the Nav Table, so it is much easier to look up that (in say 2019) April 14 is day 104 and July 22 is day 203. This difference is  99 days and (99 /10) x 0.335 = 3.3 seconds.  The watch error at hand is -3s.

If you had a Nautical Almanac onboard, you would find that there is a Day of the Year (DOY) Calendar in every issue, because they know this comes up in ocean navigation. If you do not have an Almanac, you can use our new compact DOY calendar, which we will keep updated and available online.

Note added Nov 29, 2023: we never did follow up on this project, and instead just recommend the DOY calendar in the Nautical Almanac....  In part because you can just google DOY and get one.

One way to use this is to email it to yourself and then open in any ebook reader, such as Kindle or Apple Books. That will store the document in the library of that reader so it is always easy to find. It is designed for this purpose, being all on one page, with small type that can be zoomed.