混合人工智能颶風(fēng)預(yù)報(bào)系統(tǒng)：深度學(xué)習(xí)集成方法和卡爾曼濾波器

AhybridAIhurricaneforecastingsystem:deeplearningensembleapproachandKalmanfilter

EbrahimEslamiandAIteammembersPI:Dr.YunsooChoi

DepartmentofEarthandAtmosphericSciences

UniversityofHouston

April2019

Yanniclops

AhmedKhanSalman EbrahimEslami

AlqamahSayeed

Overview

2007ThomsonHigherEducation

3

Overview

Trackingthepathandforecastingtheintensityofhurricanesarechallenging:

Dynamicalmodels,likeHWRF,produceasignificantmodel-measurementerror.

Accurateforecastingisverydifficulttoachieveafterlandfall.

4

Machinelearningcanbeasupplementaryapproachtotunehurricaneforecasting.

Overview

TropicalCycloneHistoryPacific:since1949,Atlantic:since1851

5

Introduction

ArtificialIntelligence

MachineLearning

NeuralNetwork

DeepNeural

Network

Deep

ConvolutionalNeuralNetwork

DeepLearning

Regressive

DeepConvolutional

NeuralNetwork

6

PAGE

7

DeepLearning

RegressiveDeepConvolutionalNeuralNetwork:

Hurricanemodel1Hurricanemodel2

…

Hurricanemodeln

Predictedhurricanepathandintensity

…

…

PAGE

8

HurricaneForecasting

Introduction

Atropicalcycloneforecastinvolvesthepredictionofseveralinterrelated

features,including:

Track,intensity,rainfall,stormsurge,areasthreatened,etc.

NationalHurricaneCenter(NHC)normallyissuesaforecastevery6hoursandupto72hours.

Officialforecastisbasedontheguidanceobtainedfromavarietyofsubjectiveandobjectivemodels.

Ensemblemodelisamainstreamapproachinhurricaneforecasting.

Machinelearning(deeplearning)isprovenasapowerfulensembletechnique.

Introduction

Dynamicalmodel

Statistical

model

ML

Ensemblemodel

Bestforhurricane

intensityforecasting

Bestforhurricanetrackforecasting

HurricaneForecasting

PAGE

10

InputModels

Summaryofglobalandregionaldynamicalmodelsfortrack,intensity,andwindradii:

ATCF*ID

ModelName

HorizontalResolution

Cycle/RunPeriod

NHCForecastParameters

NVGM/NVGI

NavyGlobalEnvironmentalModel

Spectral(~31km)

6hr(144hr)

Trackandintensity

AVNO/AVNIGFSO/GFSI

GlobalForecastSystem

Spectral(~13km)

6hr(180hr)

Trackandintensity

EMX/EMXI/EMX2

EuropeanCentreforMedium-RangeWeatherForecasts

Spectral(~9km)

12hr(240hr)

Trackandintensity

EGRR/EGRI/EGR2

U.K.MetOfficeGlobalModel

Gridpoint(~10km)

12hr(144hr)

Trackandintensity

CMC/CMCI

CanadianDeterministicPredictionSystem

Gridpoint(~25km)

12hr(240hr)

Trackandintensity

HWRF/HWFI

HurricaneWeatherResearchandForecastsystem

NestedGridpoint(18-6-2km)

6hr(126hr)

Trackandintensity

CTCX/CTCI

NRLCOAMPS-TCw/GFS

initialandboundaryconditions

NestedGridpoint(45-15-5km)

6hr(126hr)

Trackandintensity

HMON/HMNI

HurricaneMulti-scaleOcean-coupledNon-hydrostaticmodel

NestedGridpoint(18-6-2km)

6hr(126hr)

Trackandintensity

*TheAutomatedTropicalCycloneForecastingSystem(ATCF)

/modelsummary.shtml

Methodology

Input

3sub-models

DNN

Model1

Model2

.

.

.

Modeln

Track(allmodels)

Intensity(allmodels)

UHMLEnsembleHurricaneForecastingSystem:

GlobalandRegionalDynamicalModels

Output

IBTrACS:

TropicalCycloneBestTrackData

HurricaneIntensityandTrack

Track

Intensity

DNNsmodelingtimeperiod:

Trainingdata: 2003–2016

Nextstepprediction: 2017(e.g.HurricaneHarvey)

PAGE

12

Methodology

Weusedthreesub-modelsinourensemblemodel:

Intensitypredictor

Directionpredictor

Traveldistancepredictor

RegressiveDeepConvolutionalNeuralNetworkwasusedforallDNNmodels.

Afterensembletrackmodel,anEnsembleKalmanfilter(EnKF)wasusedtobias-

correctthehurricane’spath.

CNN

Model1

Model2

.

.

.

Modeln

Hurricane

Models

Output

(BestTrack)

Input

(Models)

DNN3

DNN2

DNN1

TravelDistance

Direction

Intensity

TravelDistance

Direction

Intensity

EnKF

ForecastedHurricanePathandIntensity

Methodology

EnsembleKalmanFilter(EnKF)

Ensembletrack

model(DNN2+DNN3)

Step1

Step2

Trackforecastbiases

Bias-corrected

hurricanetracks

EnKFappliedto

storms

HurricaneHarvey(2017

HurricaneHarvey(2017)

PAGE

16

Results

AllTropicalcyclones(models&besttrack)fortheNorthAtlanticin2017:

RMSEforhurricanepositionandintensity:

UHMachineLearningEnsemble(UHMLE)HurricaneModelingSystemvs.NHCofficialforecast(above)andothermodels(right).

TCLPHCCAFSSECTCIEMXIGFSIHMNIHWFIOCD5

NHCOfficial

UHMLE

Guidancemodelerrors(nmi)Intensityforecasterrors(knots)

0 20 40 60 80 100 120

Summary

Wedevelopedahybridthree-stepDNN-basedensemblehurricaneforecastingmodelwithEnsembleKalmanfilter(EnKF)post-processing.Themodelusedtheoutputofeightdynamicalhurricanemodels.

WeusedalltropicalcyclonesinAtlanticOceanfrom2003-2016andtestedthemodelfor

thosein2017.

EnKFfurtherimprovedthehurricanetrackforecastingbyreducingthebias.

ThepreliminarilyresultsshowstatisticaladvantagesoverNHCofficialforecasts–~13%

improvementintrackforecastbiasesand~30%improvementinintensityforecastbiases.

Challenges:

Long-termforecastingandfloodingpredictioncouldbechallengingduetouncertain

trainingdatasets.

Acknowledgements

ThankstoEarthScienceInformationPartner(ESIP)forseedfunding

ThankstoDr.Young-JoonKim(NOAAAFS)forprovidingausefulsuggestiononthisstudy

On-goingHurricanestudy

Imageforecastingusingadvanceddeepneuralnetwork:

ForAODandHurricanetracking

PAGE

20

Motivation

Source:airchem.snu.ac.kr

AODprediction(left)andhurricanetracking(right)arebothimageforecastingproblems…

L

HurricaneKatia

HurricaneIrma,2017(source:GOES,NOAA)

PAGE

22

Methodology

TestingImageForecastingwithAI:

Question:CanAIpredictbasicmovementsfromjustreceivingpreviousstateswith

Observation

justimageasinput?

Methodology

TestingImageForecastingwithAI:

Question:CanAIpredictbasicmovementsfromjustreceivingpreviousstateswith

Observation

justimageasinput?

Prediction

YESITCAN!

Methodology

Observation

TestingImageForecasting:Part2

CantheAIfollowtwofeaturestravelingindependentlyand