ATL: INVEST 94L - Models

[cycloneye]

Only model runs here.

[SFLcane]

Yea no….

[cycloneye]

By the way, is that NNIC a new model?

[AutoPenalti]

Yikes.

[Chris_in_Tampa]

By the way, is that NNIC a new model?

Here is the information I found about it:


PDF page 17:
https://www.icams-portal.gov/meetings/T ... rokova.pdf



Neural Network Intensity Consensus (NNIC) Model
- Model Input: DSHIP, LGEM, HWFI and AVNI
- Other Input: Vertical shear, SST, persistence
- Forecasts:
--- 1) NNIB - Simple average of 4 models
--- 2) NNIC - Neural network with one hidden layer
- Trained on 2014-2019
- Run operationally in 2020
- Evaluation: Is NNIC more accurate than the equally weight average NNIB?

---

PDF page 9:
https://www.icams-portal.gov/meetings/T ... linsky.pdf


"CNN" stands for "convolutional neural networks"

Pilot projects:
- Rapid intensification forecast with lightning input (CNN method)
- Neural Network Intensity Consensus (NNIC) model
--- Neural net with 1 hidden layer with HWFI, AVNI, DSHP, LGEM inputs + SST, shear, lat, persistence

NNIB = Equal weight of input models
NNIC = Neural network weight of input models

---

PDF page 14:
https://hfip.org/sites/default/files/ev ... lin-sm.pdf



NNIC Model
- Start with HWFI, AVNI, DSHP, LGEM intensity forecasts.
- Calculate evenly weighted intensity consensus.
- Use deviations from consensus as 4 predictors.
- For prototype add 3 basic SHIPS predictors.
--- Initial intensity, previous 12-h intensity change, forecast latitude.
- Use basic neural network (1 hidden layer) with 7 inputs.
- Train on 2014-2019 operational data.
- Real time processing set up for 2019 season.

[Kingarabian]

NNIC seems too bullish on many systems so far.

[Chris_in_Tampa]

By the way, is that NNIC a new model?

For the Neural Network Intensity Consensus (NNIC) I found this:

Code: Select all

! Write NN baseline consensus forecast to ATCF file
  aidlab='NNIB'
  call writeaidlocal2(ludat,atcfidl,aymdh,aidlab,&
                     &latdum,londum,ivmaxnc,stype,iftime,nft,minc)

! Write NN forecast to ATCF file
  aidlab='NNIC'
  call writeaidlocal2(ludat,atcfidl,aymdh,aidlab,&
                     &latdum,londum,ivmaxnn,stype,iftime,nft,minc)

From:
https://www.nco.ncep.noaa.gov/pmb/codes ... module.f90
Archived: https://web.archive.org/web/20210615095 ... module.f90

From that I would say likely this:

NNIC: Neural Network Intensity Consensus
NNIB: Neural Network Intensity Baseline Consensus

I'm going to add them into my model system this morning.

[cycloneye]

Thank you Chris for posting about the model.

[ronjon]

Looks NNIC is an ensemble approach combining 4 model results with a statistical training factor applied (i.e. neural networks). May be latest state of the art but like any approach has to compared to observations for robustness. A long tracker...interesting most intensity models slowly organize the system over the next 5-7 days. Hints that conditions even early in the season may not be too hostile.