Please support this Forum and Neven's Blog

Author Topic: Arctic sea ice forecast model  (Read 1738 times)

Tealight

  • ASIF Citizen
  • Posts: 243
    • View Profile
    • CryosphereComputing
Arctic sea ice forecast model
« on: June 05, 2016, 08:12:47 PM »
This thread will be used to discuss and improve my arctic sea ice loss forecast model, which is based on my albedo calculations.

Albedo thread: http://forum.arctic-sea-ice.net/index.php/topic,1543.msg79289.html#new


Question : what is that "outgoing kWh" variable and how did you determine it ?
Also, which period did you analyze to determine "the best results so far" ?


So far I only compared my model to recent years, 2007 and later.
The outgoing kWh is based on my heat loss to space calculations. (10 degrees Celsius black body radiation with 35% back radiation) The actual values don't matter so much as the time when this sub-part becomes relevant. It had to match the June cliff losses from end of May to July.

I did not find any significant change in fit by just assuming that thickness does not change.

I think its very important to prevent the model from shooting towards zero area. If for example the area losses are overestimated for a week, then the average thickness will increase and reduce future area losses, because the ice is thicker.  At the moment this function isn't implemented because all losses are based on historic area data and not predicted ones. Maybe thats a big part of the rather large errors for 2008 and 2011.

I'm confused about your 0.000001 factor for snow cover compared to 0.75 for extent-area.
In my experience the snow cover factor should be about 1/4 of the extent-area factor.
Or am I misinterpreting the units on each variable ?

Ha, I made a mistake. All my ice data is in million km2, but my snow area is in km2. So all my land energy calculations were too high by 6 orders of magnitude, which is exactly the number of zeros in the factor.

------------------------
Edit1: I changed the model to use average thickness calculated with my own area values instead of historic area. This reduced the error by around 50%.

Edit2: By using calculated values for (extent - area) instead of historic ones the error reduces even further.

new errors for each year
year   2007   2008   2009   2010   2011   2012   2013   2014   2015
diff   0.35   0.74   0.46   -0.13   -0.27   0.29   -0.33   0.32   0.05
%   12%   25%   13%   -4%   -9%   13%   -9%   9%   2%
« Last Edit: June 05, 2016, 11:51:08 PM by Tealight »

mmghosh

  • ASIF Lurker
  • Posts: 43
    • View Profile
Re: Arctic sea ice forecast model
« Reply #1 on: June 06, 2016, 01:04:02 AM »
Following the development, great work!

Tealight

  • ASIF Citizen
  • Posts: 243
    • View Profile
    • CryosphereComputing
Re: Arctic sea ice forecast model
« Reply #2 on: June 07, 2016, 01:14:49 AM »
Following the development, great work!

Thanks, it nice to hear my work getting appreciated.

I did some research into the 2008 error and found out that PIOMAS just overestimated the ice volume, which led to low area losses in my model. After the big 2007 melt season 2008 should have surely started with a lower ice volume and not a higher one. Over most of the season 2008 volume is close to 2006 and experiences an atypical drop in Aug-Sep. (72% more losses than 2012)

If I use 2007 volume for the 2008 season the error reduces to 350,000 km2 or 12%.
« Last Edit: June 07, 2016, 02:34:43 AM by Tealight »

Rob Dekker

  • ASIF Citizen
  • Posts: 283
    • View Profile
Re: Arctic sea ice forecast model
« Reply #3 on: June 09, 2016, 07:26:35 AM »
Tealight,
Thanks for this interesting forecast model.
Couple of things :

1) What is your forecast method projecting for September 2016 ?

2) How did you determine land snow extent on a daily basis ? After all, Rutgers only provides monthly or weekly data.

3) What is the formula you use to estimate future ice loss at any given day ? And does that formula change over the melting season ?

4) When you state :
The outgoing kWh is based on my heat loss to space calculations. (10 degrees Celsius black body radiation with 35% back radiation) The actual values don't matter so much as the time when this sub-part becomes relevant. It had to match the June cliff losses from end of May to July.

what does that mean ?

5) You use the 2007 - 2015 period to calibrate your forecast. That is 9 data points.
But you also use a good number of variables where you mention 4 already : snow cover, extent, area, PIOMAS volume.
If you have 9 data points and 4 variables, then you run the danger of artificial fitting. Like Enrico Fermi stated : "With four parameters I can fit an elephant, and with five I can make him wiggle his trunk."


Tealight

  • ASIF Citizen
  • Posts: 243
    • View Profile
    • CryosphereComputing
Re: Arctic sea ice forecast model
« Reply #4 on: June 09, 2016, 09:40:58 PM »
What is your forecast method projecting for September 2016 ?

Very unreliable before July. (more at the end)

2) How did you determine land snow extent on a daily basis?
Linear interpolation

What is the formula you use to estimate future ice loss at any given day ? And does that formula change over the melting season ?
Only daily kWh change. The latitude changes follow roughly the ice edge and main melting areas in the 2007-2015 period. The further back you go in time, the more inaccurate it gets. The first attachment shows that the model still correlates with the actual area, but consistenly predicts a higher minimum.

what does that mean ?
Example: The daily outgoing longwave radiation per m2 is 5.686kWh at 10C (283.15K)
In my Land Albedo calculations (incoming radition) this 5.686kWh is first reached on 14. May. On 20. June  it is 6.54 kWh. On 27. July the incoming radiation goes below that value.
As long as incoming radiation is bigger than outgoing radiation then the difference is considered to be energy melting sea ice.

As formula: If(incoming>outgoing , incoming-outgoing , 0)

As I said before this isn't reality and was just used to (curve)fit the June volume loss. If you have  much better idea I will try it out.

5) "With four parameters I can fit an elephant, and with five I can make him wiggle his trunk."
I would add even more variables to control his legs, tail and ears. Only then he can be considered an elephant.

OK, seriously what I have actually done is fit my energy calculations to the PIOMAS volume change, since volume change is the heart of my model. My volume calculations also include average temperatures above the ice, for the 2007-2015 period. If I want to model other years I need to adjust these as well.

I never tried to fit my calculations to area change. It just turned out to be pretty accurate in late season if I divide the volume change by average thickness. As you might have seen in spring my model has most of the time a few weeks below actual area and a few weeks above actual area. Only if these average out by July the final area is quite accurate.

If I use my current volume model and the 2015 extent/area ratios then 2016 will reach a 2.66 million km2 minimum. However the June cliff is the most uncertain time and I rate the current error at at least 500k km2.

-----------------
Edit:
If you have 9 data points and 4 variables
To be more precise: I modeled volume change for every day during the melting season with 5 variables.Thats 185 data points per year or 1480 for the entire 2007-2015 period.
« Last Edit: June 10, 2016, 12:21:33 AM by Tealight »

Tealight

  • ASIF Citizen
  • Posts: 243
    • View Profile
    • CryosphereComputing
Re: Arctic sea ice forecast model
« Reply #5 on: June 11, 2016, 10:17:25 PM »
I produced a chart showing a breakdown of the combined volume change from my previous post.

The temperature curve is based on modified DMI 80N average temperatures to better model temperatures over the entire icepack and not just north above 80N.

The snow cover curve models the June cliff, mostly heat from continents blown over the ice.

The albedo curve is the (extent-area) calculations, which I have shown in previous posts.

What I'm still missing is a curve for bottom melt when ice drifts over warm open ocean. This should be more and more important as the ice gets more mobile and the open water area increases.

Tealight

  • ASIF Citizen
  • Posts: 243
    • View Profile
    • CryosphereComputing
Re: Arctic sea ice forecast model
« Reply #6 on: July 04, 2016, 02:20:55 PM »
July Update

new month = better model

Main model changes:
1. Switch from Cryosphere Today area to NSIDC area calculated by Wipneus
2. Switch from fixed latitude changes to a daily latitude calculation 

Wipneus data: https://sites.google.com/site/arctischepinguin/home/sea-ice-extent-area/data

The first change makes model more variable. With CT area the model is too averaged out.

The second change calculates the daily latitude of main melting areas for the ice pack based on sea ice extent. The range is between 60N and 85N in 0.25 decimal steps and allows each year to progress individually. This mainly reduces the error of pre 2007 years or very cold years like 2013.

At the end of the melting season 2012 reached 79.75N and 2009 reached 76.75N. The 1980s never exceed 74N at the end of the season. With these changes I achieved my goal of under 10% error for an individual year and under 5% for the 2007-2015 period.

Year      Error
2007   -2.4%
2008    6.5%
2009    8.6%
2010   -1.9%
2011   -7.2%
2012    6.3%
2013   -9.2%
2014   -0.4%
2015    1.6%
average: 4.9%




« Last Edit: July 04, 2016, 07:19:57 PM by Tealight »