Three of the partners are non-

Climate Adaptation Modelling at The Bjerknes Centre

Geophysical Institute Meeting, June 26, 2014

Dr. Michel d. S. Mesquita

michel.mesquita@uni.no

1. Uni Research Climate, Bergen Norway 2. The Bjerknes Centre for Climate Research

! 

The Bjerknes Centre is an umbrella organization made of four partners

! 

Three of the partners are non-

profit resaerch and one is a public Univeristy

! 

Collaborative research is then termed a Bjerknes joint

! 

Each institute maintains own research portfolio and

independent identity

The$structure$of$the$Bjerknes$Center$partnership$

Bjerknes The Center Research Uni

Nansen Center

Institute Marine Research

Univeristy of Bergen

! 

All four partners are actively engaged in regional climate research and actively

collaborate

! 

Uni Research has capabilities in hydrology and capactiy

building as well in addition to climate dynamics

! 

Uni Research is also one of three founding parnters of the Norwegian Centre for Climate Services (http://

klimaservicesenter.no/)

Regional$Climate$Research$at$Bjerknes$

!  Global&Modelling&and&Climate&Dynamics&

!  “Big”&science&

!  Global&projec:ons&

!  Basic&research&

!  Regional&Climate&Lab&

!  Regional&downscaling/analyses&and&

uncertainty&quan:fica:on&

!  LocalCregional&scale&process&studies&

!  SynthesizeCinterpretCintegrate&(i.e.&

develop&“regional&narra:ves”)&

!  Contextualize&around&realCworld&

problems&

!  Climate&OutREach&(CORE)&

!  Communicate&climate&science&to&

broad&audiences&

!  Engage&in&educa:on&&&capacity&

building&

Regional$Climate$Lab$&$Climate$Services$at$Uni:$

Structure$

Regional Climate Lab

Climate OutRE (CORE)

Global Modeling and Climate Dynamics

User Communities

5

Motivation

›  In Europe, intense precipitation has become more severe and more frequent

›  Winter rainfall has decreased over

Southern Europe and the Middle East, and has increased further north

›  In some parts of Europe, observed trends to more and longer heat waves and fewer extremely cold days and nights have been observed

›  The risk of and vulnerability to floods have increased over many areas in Europe

›  The insurance industry reports a pronounced increase in the number of weather-related events, which have caused significant losses

›  In some regions, low-lying coastal zones are considered to be particularly vulnerable to climate change, especially through sea-level rise, changes in wave climate and in

storminess

6

Extreme Weather Events in Europe:

preparing for climate change adaptation

October 2013

ISBN (print) 978-82-7144-100-5 ISBN (electronic) 978-82-7144-101-2 This report can be found at www.dnva.no

Produced by the Norwegian Meteorological Institute

In cooperation with

Present Climate

›  An estimated loss of € 415 billion since 1980 (2010 values)

›  Most costly hazards have been storms and floods, amounting to a combined total of almost € 300 billion

›  In Europe, it is estimated that 140 000 lives were lost since 1980 due to weather events

›  Largest impacts on life have come from heat waves

7

Extreme Weather Events in Europe:

preparing for climate change adaptation

October 2013

ISBN (print) 978-82-7144-100-5 ISBN (electronic) 978-82-7144-101-2 This report can be found at www.dnva.no

Produced by the Norwegian Meteorological Institute

In cooperation with

Economic and Life Losses

8

Land surface temperature anomaly (ºC)

-20 0 20 Below normal Above normal Average surface temperatures

Figure 3.1 The temperature difference in Europe between 11–18 December 2009 and the 2000–2008 average (left), and the average temperatures for 20–27 July 2010 compared to the normal for the period (right)

Source: (left) http://en.wikipedia.org/wiki/Winter_of_2009–10_in_Europe; (right) http://blogs.agu.org/wildwildscience/

2010/08/11/amazing-nasa-images-of-russian-heat-and-smoke (image from NASA Terra satellite).

28 | October 2013| Extreme Weather Events in Europe The scientific background to the analysis of extreme weather is to a very large extent driven by observations.

These give evidence for recent trends in our climate and suggest how it might change in future. An understanding of phenomena, processes, and how different aspects of the climate interact provides essential information for interpreting these observations. In this chapter, the current state of understanding about the principal extreme

climatic phenomena is described and assessed. A reliable analysis of past trends, an understanding of the factors that affect the trends and the way in which these will change in future because of global warming are all required to develop science-based adaptation strategies.

The immediate and underlying causes of changes in extreme temperatures are discussed in the following sections, including the way in which extremes are affected by global warming.

3.1 Extreme heat and cold In summary:

• Observations show a trend to more warm days, hot days and heat waves and fewer cold days over most parts of Europe since the mid 20^th century.

• Most places in Europe are very likely to experience more hot and fewer cold extremes as the global temperature increases.

• The magnitude of hot and cold extremes is expected to increase faster than mean temperatures over large parts of Europe.

• The probability of occurrence of heat waves, such as those in 2003 in Europe or 2010 in Russia, is expected to increase substantially. For example, what is now a 1 in 50-year event may become a 1 in 5-year event by the end of the 21^st century.

Introduction

Recent European summer heat waves and winter cold spells have had severe socio-economic and ecological impacts. The record-breaking 2003 and 2010 heat waves led to tens of thousands of heat-related deaths across Europe (Robine et al., 2008; Barriopedro et al., 2011), crop shortfalls, extensive forest fires and record high prices on the energy market amongst many other effects (Figure 3.1;

Schär et al., 2004; García-Herrera et al., 2010). In the winters of 2005/2006 and 2009/2010, parts of Europe experienced unusually cold temperatures that caused travel disruption, cold-related mortality and high energy consumption (Cattiaux et al., 2010).

Since adaptation to extremes of heat and cold is

particularly difficult, changes in their frequency, duration or spatial extent and extremes of intensities never

CHAPTER 3 EXTREME WEATHER PHENOMENA AND THEIR CONSEQUENCES:

THE SCIENTIFIC BACKGROUND

›  More frequent heat waves, with increased duration and intensity

›  Increase of summer dryness in Central and Southern Europe " enhanced risk of

drought, longer dry spells, and larger soil moisture deficits

›  More frequent droughts throughout Europe

›  High intensity and extreme precipitation are expected to become more frequent within the next 70 years

›  Increase in wind-storm risk over

Northwestern Europe, leading to higher storm damage when there is no adaptation

9

Extreme Weather Events in Europe:

preparing for climate change adaptation

October 2013

ISBN (print) 978-82-7144-100-5 ISBN (electronic) 978-82-7144-101-2 This report can be found at www.dnva.no

Produced by the Norwegian Meteorological Institute

In cooperation with

Future Climate

10

Norway’s strategy

11

›  Responding to local climate change is a responsibility of the municipalities

›  For a robust and resilient municipality, there is a need to understand how climate may affect the various elements the municipality is responsible for

›  More specialised knowledge on climate needed " creation of Climate Services

›  Pilot project with Hordaland County and municipalities (HORDAKLIM)

Climate Services

Climate Science

Decision makers

12

Prosjektbeskrivelse: Klimaservice i Hordaland - HORDAKLIM

side 2/10

2. Innovasjonsgrad

En forutsetning for å tilpasse seg klimaendringene er å forstå hvordan klimaendringene påvirker ens virksomhet og ta hensyn til klimaendringer i planleggings- og beslutningsprosesser, og det er viktig at alle har et felles kunnskapsgrunnlag og kollektiv bevissthet når man skal legge til rette for planer og beslutninger. Per i dag finnes ingen arena for kunnskapsoverføring fra forskningsmiljøene til lokalt- og regionalt myndighetsnivå. HORDAKLIM vil legge til rette for optimalisering av kunnskapsoverføring. Dette vil skje gjennom dialog mellom forskning og kommuner gjennom etablering av temagrupper der man adresserer spesifikke utfordringsområder (se figur 1).

Innovasjonen i dette prosjektet handler derfor spesifikt om kunnskapsoverføring og nettverkslæring i praksis:

1) Nasjonale myndigheter har lagt føringer og forventninger til klimatilpasning på regionalt og kommunalt nivå. I dette prosjektet skal Hordaland fylkeskommune i tett dialog med Fylkesmannen i Hordaland, stå sentralt i samordning og nettverksbygning med klart mål om å gjøre kommuner i Hordaland best mulig rustet til å møte klimautfordringer i sin region.

Vi vil bygge på eksisterende nettverk og etablere nye interkommunale ressurs(tema)grupper på områder som regnes som særlig utfordrende knyttet til klimatilpasning.

2) Nettverkslæring og samordning: Først og fremst er det et behov å skape gode felles arenaer der forskningsgrunnlaget for forventede klimaendringer presenteres enhetlig for alle berørte sektorområder (lokalt og regionalt myndighetsnivå). Uavhengig av hvordan kommunenes tilpasningsarbeid organiseres er det grunnleggende å ha kunnskap om hvordan klimaendringene vil slå ut i eget lokalmiljø. Et overordnet mål er å formidle kunnskapen om de fysiske projiserte klimaendringer direkte til kommunene. Overføring av kunnskap om klimaendringer er viktig, men samtidig ligger kunnskapen om praktiske løsninger i kommunene.

Det er også kommunene som best kjenner sine utfordringer, og dette må formidles til forskernettverket for at disse kan svare direkte på individuelle behov fra kommunene;

informasjonen må være relevant. Gjennom dialog mellom forskning og kommuner (samordnet via Fylkeskommunen) vil man identifisere de mest sentrale utfordringsområdene. Dette kan være innen ras/flom, landbruk/skogbruk, energi/vannkraft, turisme, havbruk osv. Hver kommune plasserer sine identifiserte utfordringer innen temagrupper. Dette vil motvirke at øykommuner ikke må bruke ressurser på fokusområdet ”snøskred” eller at jordbrukskommuner ikke vil måtte implementere planer for stormfloproblematikk Dette vil også skape felles erfaringsplattform mellom kommuner med lik tematisk utfordring. Dialogen mot kunnskapsmiljøet vil skje direkte opp mot disse temagruppene. Dette vil sikre at kunnskapsmiljøene får optimalisert bestillingen basert på reelle lokale/regionale behov og dermed kan skreddersy produktene etter behovene som er identifisert.

Gjennom temagruppene (?) Her ønsker vi å skape arena for gode nettverk og regionalt samarbeid som kan styrke den enkelte kommunes evne til å tilpasse seg de forventede endringene, og de naturvitenskaplige miljøenes evne til å levere tilpasset og relevant informasjon. Dette vil også

Nasjonale)myndigheter) og)føringer)(i.e.)St.)Meld.)33))

Direktorater)(MD,)DSB)) Fylkesmannen)Hordaland)

Hordaland)Fylkeskommune)

K1) K2) K3) K4) K5) K6)…..)K28) K29) K30) K31) K32) K33)

Ras/)

Flom) Kyst)og)

hav) UOordr.)

x₁Qx_n) Landbruk

skogbruk) Energi/

strøm)

Kommuner)–)33)i)Hordaland)

UOordringsområder)

K1)K3) K5) K9) K12) K22)K30)

K2)K4) K8) K13) K18) K29)K31)

K1) K3) K5) K9) K12) K22)K30)

K1) K3) K5) K9) K12) K22)K30)

INTERKOMMUNALE)RESSURSGRUPPER)

Norsk)klimaservice)senter)(Met.no,)NVE,)Uni)Research)Klima)) Q  Kompetanse)overføring/kapasitetsoppbygging)

Q  Levere)data)e`er)spesifikke)behov) Q  Nedskalerte)data)og)frembdsscenarier) Q  Risikovurdering)og)usikkerhetsvurdering)

To)veis)kommunikasjon)

Figur 1. Flytdiagram som illustrer hvordan prosjektet er oppbygd for å sikre god dialog for kunnskapsoverføring

13

What we do

14

Wind Energy

Present and future offshore wind power potential in northern Europe based on downscaled global climate runs with adjusted SST and sea ice cover

Idar Barstad^a,*, Asgeir Sorteberg^b,c, Michel dos-Santos Mesquita^c,d

aUni Computing, Uni Research, Allegt 55, Bergen, Norway bGeophysical Institute, University of Bergen, Bergen, Norway cBjerknes Centre for Climate Research, University of Bergen, Bergen, Norway dUni Bjerknes Centre, Bergen, Norway

a r t i c l e i n f o

Article history:

Received 28 February 2011 Accepted 5 February 2012 Available online 29 February 2012 Keywords:

Future wind resource Downscaling of climate model results IPCC-AR4

a b s t r a c t

Coupled global climate models coarse results have been downscaled to produce future wind power maps for northern Europe. The downscaling method utilizes a global, stretched atmospheric numerical model with sea-surface temperature (SST) as the main forcing. The model has horizontal grid spacing equivalent to about 30 km in the area of interest. As the climate models have often problems with the sea ice cover and storm tracks in vicinity of the sea ice, an alternative SST approach has been used. The SST signal from climate model runs under the A1B scenario has been added to the Era40 reanalysis data set, and used as lower boundary forcing. A 30-year control period (1972e2001) is compared to a future period (2020e2049) of equal length. Four realisations of the future period constitute the ensemble, which the future wind power potential is estimated from.

The results show that a weak reduction of wind power production is expected in the future period. The reduction of the power potential is in the range from 2 to 6% in most areas. The spread in the model ensemble is large and consequently the reduction becomes relative small. Regional pockets of increased potential appear in vicinity of high terrain. These results are regarded as uncertain as a little shift in storm tracks will lead to very different mountain shadow effects and alter the picture drastically.

!2012 Elsevier Ltd. All rights reserved.

1. Introduction

High ambitions for renewable energy and scarcity of suitable onshore areas for wind energy in Europe encourage large offshore wind farm installations. The plans for offshore wind parks for the next 10e15 years are formidable[1]. The depreciation time for such large parks can be comparable to the rapid man-made climate changes, which have already made a foot print on the Earth[2]. The infrastructure of such large offshore installation is typically more expensive than onshore counterparts, relying on somewhat even longer depreciation time - at least in a socio-economical context.

Coupled global climate models (CGCMs) are the only viable tools for addressing future changes on a time scale of decades. Natural variability along with man-made climate forcing determines the future state. CGCMs project changes in climate convincingly if used in an appropriate manner (e.g[3].). Nevertheless, models are not perfect due to coarse resolution and shortage in the physical and numerical treatment. Unsystematic errors in models and natural

variability leading to divergence among model results are not necessary a hindrance to reliable projections. Ensemble means have proven to be more accurate than individual models in reproducing the instrumental observational period (e.g[4].). This gives hope for the separation of the climate signal from that of noise providing enough ensemble members are used. Even systematic errors can, to some extent, be dealt with by combining models of different design. However, the use of model results in a relative sense (future estimate as a fraction of the present) is probably among the better method to reduce systematic errors. Not yet mentioned, the different emission scenarios introduce additional uncertainties to a future projection, requiring additional model realisations of a future state.

Downscaling of CGCMs is typically motivated by the desire of more details for some future time period. Dynamical downscaling is perhaps the most promising method for such refinement ([5]). In extra-tropical areas, the weather systems are highly advective of nature, and limited area numerical models with small model domains, are strongly influenced by the lateral boundaries and their placement. Models with larger model domain arefirst of all influenced by the lower boundary, i.e., properties of sea-surface temperature (SST) and sea ice.

*Corresponding author.

E-mail address:idar.barstad@uni.no(I. Barstad).

Contents lists available atSciVerse ScienceDirect

Renewable Energy

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / r e n e n e

0960-1481/$esee front matter!2012 Elsevier Ltd. All rights reserved.

doi:10.1016/j.renene.2012.02.008

Renewable Energy 44 (2012) 398e405

The fractional change in averaged annual power production for the period (2020–2049) versus (1972–2001). Unity indicated by black contour also shown over land for clarity.

Above unity means increased future power potential.

Latitudes and longitudes are indicated every 10°. Stations in Table 2 are indicated.

NORINDIA$Project:$Climate$Change$and$Water$Resources$in$Northwest$India

Fig: The Boundary of homogeneous

regionzone of India (Sontakke et al. 2008) Fig : Beas sub-basin up to Pandoh

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

NORINDIA$Project:$Modeling$ObjecIves$

Hydrology

›  Investigate the spatial and temporal variability of high-

resolution simulated precipitation (WRF) in Beas river basin compared with gauge data

›  Set up a calibrated hydrological model and a fully coupled atmos-hydro model in Beas river basin

›  Quantify changes in future streamflow and water availability

Source: Dr. Lu Li (lu.li@uni.no)

17

›  WRF:&Weather&Research&and&Forecas:ng&Model&

›  NDHMS:&the&NCAR&Distributed&hydrological&Modeling&System&

(Noah&based)&

›  WRFChydro&:&&NDHMS&coupling&with&WRF&&

WRFKhydro$

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

NORINDIA$Project:$Domain$Design$

›  ERA-Interim 6 hourly

›  27km: 125*125

›  9km: 100*100

›  3km: 118*118

›  Routing subgrid: 300m

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

Geogrid terrain

WRF-hydro routing channels

$Hydro$grids$

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

20

Monthly$precipitaIon:$Pandoh$

Under-estimates;

Over-estimates;

Captured variability;

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

Simulated$vs.$Observed$Discharge &

NS efficiency = 0.4 volume error 36%

Hydrology

Source: Dr. Lu Li (lu.li@uni.no)

Climate$Services:$Bangladesh$and$the$importance$of$

local$percepIons$(PhD.$cand.$Mathew$Reeve)$

mathew.reeve@uni.no!

22

›  A meteorological phenomenon in N.E.

Bangladesh has dramatic implications for monsoon definitions at the local scale

›  Parts of Bangladesh experience early onset in April that can result in up to 35%

of annual rainfall before the official start of monsoon season

›  Thus, monsoon forecasts such as those from IITM are essential useless for this region

›  Can we come up with more locale

appropriate definitions of the Monsoon?

WIND&DEFINITION&

RAIN&DEFINITION&

Society

Source: PhD Cand. Mathew Reeve (mathew.reeve@uni.no)

86 E 88 E 90 E 92 E 94 E

Apr 6 Apr 16 Apr 26 May 6 May 16 May 26 June 5 June 15 June 25 July 5 July 15

WIND RAIN COMMUNITY

21 N 24 N 27 N

a). b). c).

86 E 88 E 90 E 92 E 94 E

86 E 88 E 90 E 92 E 94 E

Climate$Services:$Local$PercepIons$MaQer!!

Society

Source: PhD Cand. Mathew Reeve (mathew.reeve@uni.no)

24

Ecology

Source: Dr. Michel d. S. Mesquita (michel.mesquita@uni.no)

25

Ecology

Source: Erikstad et al. (2013) Mar Ecol Prog Ser 475: 267–276, 2013

0.27) noise (statistics provided are for the top-ranked model). Population models were run in the R environment (R Develop- ment Core Team 2011).

Modelled fish prey availability

Whereas observational juvenile distribu- tions of fish prey were only available from autumn international 0-group fish surveys in the Barents Sea, a coupled ocean model and an IBM for ichthyo plankton (eggs, lar- vae and pelagic juveniles) enabled a contin- uous spatio-temporal description. We used the model setup for dispersal of ichthyo - plankton of northeast Arctic cod and Nor- wegian spring-spawning herring as de- scribed by Vikebø et al. (2011). Key elements are an IBM for early stages of fish forced by

the daily mean ocean weather forecast by the Norwe- gian Meteorological Institute produced with the 3-di- mensional ocean model MI-POM (described by En- gedahl 1995). In the model, fish larvae are represented by particles, and daily spatio-temporal distributions are available for overlap analysis with common guille- mot feeding areas. We defined a box centred around Hornøya of 200 × 200 km (i.e. within a reasonable for- aging range of common guillemots; Burke & Monte - vecchi 2009) and quantified the number of particles inside the box that originated from different spawning grounds along the coast at different times of the year.

RESULTS

Between 1986 and 1987, there was a very large (80%) decline in the population of common guille-

mots on Hornøya (Fig. 1). Since then, the popula- tion has steadily increased and surpassed the 1983 level in 2003 (Fig. 1). The yearly population growth rate varied between years, but was positive or close to 0 for all years except 1987, the collapse year (Fig. 1). Coincident with the collapse in guillemot population were very low levels of all prey species (Fig. 2).

Since the crash year would have represented an extremely influential outlier, we modelled the growth of the guillemot population from 1987 on - wards. Unlagged 0-group cod was the only well- supported 1-parameter model explaining guillemot population dynamics (Table 1) and accounted for 40% of the temporal variation in the population growth rates. Most of the other covariates (except for herring lagged by 6 yr) were poorer than the null model without prey covariates (Table S2 in the

270

1981 1984 1987 1990 1993 1996 1999 2002 2005 2008 2011Year

Number of birds

0 2000 4000 6000 8000 10000 12000 14000

Annual population growth rate ln(Nt/Nt–1)

–1.4 –1.2 –1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2

Population size growth rate

Fig. 1. Uria aalge. Annual variation in the approximate breeding popula- tion (no. of individuals) and in the population growth rate (ln [N_t /N_t–1])

of common guillemots breeding at Hornøya, NE Norway

1983 1987 1991 1995 1999 2003 2007 1983 1987 1991 1995 1999 2003 2007 1983 1987 1991 1995 1999 2003 2007

Abundance index (no. x 106 )

10²

10² 10¹ 10⁰ 10³

10³ 10⁴

10⁴

10⁵ 10⁶

10⁵

10³ 10⁴ 10⁶ 10⁵

0-group cod Capelin

Year

0-group herring

A B C

Fig. 2. Gadus morhua, Mallotus villosus, and Clupea harengus. Annual variation in fish stock size indices (estimates of num- bers, N × 10⁶) of appropriate age classes of fish species in the Barents Sea area known as important prey species for common guillemots breeding at Hornøya, NE Norway. (A) 0-group cod, (B) all age classes of capelin, (C) 0-group herring. The shaded

area indicates the year when the common guillemot population collapsed (see ‘Results’ for details)

EuroKCORDEX:$Europe$at$+2C$Global$Warming$

(IMPACT2C,$Stefan$Sobolowski)$

27

›  Initial findings

published in Vautard et al. (2014) used

ENSEMBLES

›  Very simple measure for robustness (model agreement)

Future Projections

Source: Vautard et al. (2014)

Climate$Services$and$InnovaIve$Methods:$Combined$StaIsIcal$

Dynamical$Downscaling$for$the$Hardanger$Bridge$and$Extreme$

Winds

28

›  The problem: high winds in fjords are not resolved even in high-resolution RCMs

›  Need LES scales (e.g., 300m)

›  However, extremes can be related to direction and windspeed upstream at the coast

›  Use emprical statistcal models and flow regime identification to estimate extreme winds in the future

Engineering

Source: Dr. Stephen Outten (stephen.outten@nersc.no)

Regional Climate Impacts

Hardanger$

bridge&

Utsira&

Current (1961-1990)

Future (2070-2099)

WS₅₀ = 37.9 ms^-1

WS₅₀ = 38.2 ms^-1

WS₅₀ = 38.2 ms^-1

WS₅₀ = 38.6 ms^-1

WS₅₀ = 38.1 ms^-1

Map&from:&hWp://www.kart.gulesider.no&

Photo&from:&&Norwegian&Public&Roads&Administra:on&

&

&

&&

&&

&&

Hardanger$Bridge$and$Extreme$Winds$

Source: Dr. Stephen Outten (stephen.outten@nersc.no)

Engineering

More on extreme winds in complex terrain

›  10 min wind velocity estimates for Hardanger bridge for the reference and future periods (source: Instanes and Outten in prep.)

30 35 40 45

0 20 40 60 80 100

Wind velocity (m/s)

Return period (years)

10 min max (2071-2100) 10 min mean (2071-2100) 10 min (1961-1990)

Source: Dr. Stephen Outten (stephen.outten@nersc.no)

Engineering

31

Education

Source: Dr. Michel d. S. Mesquita (developer of m2lab.org)

›  Course: Regional Climate Modelling Using WRF

•  463 participants

›  Course: Introduction to Applied Bayesian Statistics for Climate Research

•  328 participants

›  Course in collaboration with Plymouth State University (USA):

Weather Applications Using WRF

›  New Course in collaboration with the University of Bergen:

WIMEA-ICT e-learning: WRF training for Eastern Africa

›  Collaboration with Oxford University " Publication on framework for teaching climate modelling online

Collaborators:

32

Education

Source: Dr. Michel d. S. Mesquita (developer of m2lab.org)

e-WRF

33

Uncertainty: Bayesian Analysis

Source: Mesquita et al. (2014)

arXiv:1405.7447v1 [stat.ME] 29 May 2014

Horizontal resolution in a nested-domain WRF simulation: a Bayesian analysis approach

Michel d. S. Mesquita^∗

Uni Climate, Uni Research and Bjerknes Centre for Climate Research, Bergen, Norway

*Corresponding author email: michel.mesquita@uni.no

Bjørn ˚Adlandsvik

Institute of Marine Research, Bergen, Norway

Cindy Bruy`ere

National Center for Atmospheric Research, Boulder,CO, USA

Anne D. Sandvik

Institute of Marine Research, Bergen, Norway

ABSTRACT

The fast-paced development of state-of-the-art limited area models and faster computational resources have made it possible to create simulations at increasing horizontal resolution. This has led to a ubiquitous demand for even higher resolutions from users of various disciplines. This study revisits one of the simulations used in marine ecosystem projects at the Bjerknes Centre.

We present a fresh perspective on the assessment of these data, related more specifically to: a) the value added by increased horizontal resolution; and b) a new method for comparing sensitivity studies. The assessment is made using a Bayesian framework for the distribution of mean surface temperature in the Hardanger fjord region in Norway. Population estimates are calculated based on samples from the joint posterior distribution generated using a Monte Carlo procedure. The Bayesian statistical model is applied to output data from the Weather Research and Forecasting (WRF) model at three horizontal resolutions (9, 3 and 1 km) and the ERA Interim Reanaly- sis. The period considered in this study is from 2007 to 2009, for the months of April, May and June.

1.Introduction

The need for high-resolution data has become impor- tant in several disciplines. These data provide added infor- mation, as for example, to the study of complex topogra- phy regions such as the Norwegian fjords (??). However, producing such data, using a limited area model, can still be constrained by the computing resources available. For example, in order to make inferences about a model simu- lation, one needs a large sample to produce robust statis- tics (?). Producing large samples at high-resolution can become computationally expensive. This is especially the case when testing different combinations of parameteriza- tion schemes or a different model setup.

In this study, we present an alternative approach to an- alyzing output from limited area models based on Bayesian probability. Bayesian probability theory has been increas- ingly applied to regional climate modeling experiments in the past few years (??). The approach presented here al- lows one to make use of small samples to make inferences about the statistical population. The use of probability dis-

tributions also provides a richer view of the data for com- parison against observations. The next session will discuss the data, methods and the Bayesian approach. Section 3 will present the results, which will be followed by the con- clusion in Section 4.

2. Data and Methods

The experiments were made using the Weather Re- search and Forecasting (WRF) model version 3.1. Figure 1 shows the domain configuration, which consisted of a par- ent domain at 9 km resolution and two nested domains at 3 km and 1 km, respectively (with feedback=1, two-way nesting). They were run using 31 vertical levels. The mi- crophysical scheme chosen was the WRF Single-Moment 3- class scheme (mp physics=3). The cumulus parameteriza- tion option was turned off(cu physics=0). The planetary boundary layer scheme was the Yonsei University scheme (bl pbl physics=1). The longwave radiation scheme used was the RRTM scheme (ra lw physics=1) and the short- wave radiation was the Dudhia scheme (ra sw physics=1).

1

Fig.1. WRF model domain setup: parent domain at 9 km (outer domain), nest at 3 km (d02) and nest at 1 km (d03).

ECMWF ERA-interim Re-Analysis was used as the lat- eral boundary condition data. These data have been ob- tained from the ECMWF Data Server. The simulation was run from 2007 to 2009. The months of April, May and June of 2008 and 2009 were retained for the analy- sis. Here, results will be shown with respect to the three- hourly 2 m temperature in the Hardanger fjord region for the month of April. The box selected for the spatial av- eraging is located between 59.32^◦N, 60.75^◦N and 5.05^◦E, 7.90^◦E. From the timeseries created, we have randomly se- lected 200 timesteps for calculating the sample mean and variance.

An informative prior was selected based on the Kvamsøy weather station located at 60.358^◦N and 6.275^◦E. These data were obtained from the Norwegian Meteorological In- stitute data server ateklima.no. The Kvamsøy weather station has been operational since November 2003. The average surface temperature for April is 7.48±1.27^◦C for the years of 2003 to 2011.

a. The Bayesian model

In this study, the Bayesian model is applied to the 2m temperature in the Hardanger fjord region. It considers the case in which the mean (θ) and variance (σ²) are unknown (??). For the joint prior distributionp(θ,σ²) forθandσ², the posterior inference will use Bayes’ rule, as shown in Equation 1:

p(θ,σ²|y1, . . . , yn) =p(y1, . . . , yn|θ,σ²)p(θ,σ²) p(y1, . . . , yn) (1)

wherey1, . . . , yn, represent the data. Since the joint dis- tribution for two quantities can be expressed as the product of a conditional probability and a marginal probability, the posterior distribution can likewise be decomposed (Eq. 2):

p(θ,σ²|y1, . . . , yn) =p(θ|σ², y1, . . . , yn)p(σ²|y1, . . . , yn) (2) where the first part of the equation is the conditional probability of θ on the variance and the data; and the second part is the marginal distribution ofσ². The condi- tional probability part of the equation can be determined as a normal distribution:

{θ|y1, . . . , yn,σ²}∼normal(µn,σ²/κn) (3) Whereκn=κ0+nrepresents the degrees of freedom (df) as the sum of the prior df (κ0) and that from the data (n). µn is given by: µn = ^(κ0/σ

2)µ0+(n/σ²)y

κ0/σ²+n/σ² = ^κ0µ_κ^0+ny

n , whereyis the sample mean taken from the WRF simula- tion. The prior mean is given byµ0. The calculation ofσ² is explained next.

The second part of equation 2, the marginal distribu- tion ofσ², can be obtained by integrating over the unknown value of the mean,θ, as follows:

p(σ²|y1, . . . , yn)∝p(σ²)p(y1, . . . , yn|σ²) (4)

=p(σ²)

!

p(y1, . . . , yn|θ,σ²)p(θ|σ²)dθ (5) Solving the integral, and considering the precision (1/σ²) such that the distribution is conjugate, gives the following gamma distribution:

{1/σ²|y1, . . . , yn}∼gamma(νn/2,νnσ_n²/2) (6) Whereνn =ν0+nis the sum of degrees of freedom of the prior (ν0) and of the data (n). σ_n²is given byσ²_n=

1

νn[ν0σ²0+ (n−1)s²+^κ_κ⁰_nⁿ(y−µ0)²], whereyis the sample mean ands²is the sample variance, both taken from the WRF simulation.σ²0is the prior variance.

b. Monte Carlo sampling

Samples ofθandσ²can be generated from their joint posterior distribution using the following Monte Carlo pro- cedure (?):

σ²⁽¹⁾∼inv gamma(νn

2,σn²νn

2 ), θ⁽¹⁾∼normal(µn,σ²⁽¹⁾ κn

) ...

σ^2(S)∼inv gamma(νn

2,σ²_nνn

2 ), θ^(S)∼normal(µn,σ^2(S) κn

)

2

3.0 3.5 4.0 4.5 5.0 5.5 6.0

68101214

ERAi

θ σ2 +^{4.3 10}+^{4.3 10}

3.0 3.5 4.0 4.5 5.0 5.5 6.0

68101214

9 km

θ

σ2 +^{4.3 10}

+

4.8 7.1

3.0 3.5 4.0 4.5 5.0 5.5 6.0

68101214

3 km

θ

σ2

4.2 8.1+

4.3 10+

3.0 3.5 4.0 4.5 5.0 5.5 6.0

68101214

1 km

θ

σ2

+

4.6 9.2 4.3 10+

Fig.2. Monte Carlo samples from the joint distributions of the population mean (θ) and variance (σ²) for ERA Interim (ERAi) and for the different domains. The values in black show the mean value of the population mean (right side) and of the population variance (left side). Accordingly, the mean values ofθandσ²for the ERA Interim are indicated in red. Temperature given in degrees Celsius.

whereσ²is estimated using an inverse-gamma distri- bution (inv gamma). Eachθ^(S)is sampled from its con- ditional distribution given the data andσ²=σ^2(S). The simulated pairs of{(σ²⁽¹⁾,θ⁽¹⁾), . . . ,(σ^2(S),θ^(S))}are inde- pendent samples of the joint posterior distribution, i.e.:

p(θ,σ²|y1, . . . , yn). The simulated sequence{θ⁽¹⁾, . . . ,θ^(S)} can be seen as independent samples from the marginal pos- terior distribution ofp(θ|y1, . . . , yn), and so this sequence can be used to make Monte Carlo approximations to func- tions involvingp(θ|y1, . . . , yn). Whileθ⁽¹⁾, . . . ,θ^(S)are each conditional samples, they are also each conditional on different values ofσ². Together, they make up marginal samples ofθ.

3. Results

Monte Carlo samples from the joint distributions of the population mean and variance are shown in Figure 2. The ERA Interim distribution (ERAi), on the top left, shows larger spread both for the mean and the variance as com- pared to the three domains. The distribution for the 9 km domain seems to be offand does not match the ERA In- terim data. The 3 km nest shows the closest approximation to the mean of the ERA Interim, whereas the 1 km nest approximates the variance more closely.

3.5 4.0 4.5 5.0

01234

ERAi

θ p(θ, |, y1, …, yn)

4.2 4.4 4.6 4.8 5.0 5.2 5.4

01234

9 km

θ p(θ, |, y1, …, yn)

3.63.84.0 4.24.44.64.8

01234

3 km

θ p(θ, |, y1, …, yn)

4.0 4.2 4.4 4.6 4.8 5.0 5.2

01234

1 km

θ p(θ, |, y1, …, yn)

Fig.3. Monte Carlo samples from the marginal distri- bution ofθfor ERA Interim (ERAi) and for the different domains. The blue vertical lines give a 95% quantile-based posterior bound. In red, the mean value of the ERA In- terim posterior marginal distribution. Temperature given in degrees Celsius.

Figure 3 shows the marginal distribution of the mean, based on the Monte Carlo sampling. The red line indicates the mean value of the marginal distribution for the ERA Interim. The posterior bounds of the 9 km parent domain do not contain the mean value of the ERA Interim. Table 1 shows that even though there is some overlap between the posterior bounds of ERA Interim and the 9 km domain, this overlap is minimum. The 3 km and 1 km nests show a closer overlap with the ERA Interim data. The 3 km resolution domain is able to approximate the mean more realistically, also confirmed by the posterior bound overlap with ERA Interim (Table 1).

The marginal distribution of the ERA Interim variance is approximated more closely by the 1 km resolution do- main, as shown in Figure 4. The mean value of the ERA In- terim marginal distribution is within the posterior bounds for that resolution. In contrast, the 9 km and 3 km do- mains have posterior bounds outside of the ERA Interim mean value. There is, however, a better overlap between the ERA Interim and the 3 km posterior distribution, com- pared to the 9 km one (Table 1).

4. Conclusion

This study has used a Bayesian statistical model ap- plied to output data from the Weather Research and Fore- 3

Climate Services: Robust decreases in storm number and intensity over North/Norwegian Seas by Mid-century

›  A delta-change approach is applied to remove

systematic SST and sea-ice biases.

›  By any measure one chooses significant

reductions in the number and intensity of storms impacting the Norwegian coast are seen; > 25%

compared to present day

›  But are these findings useful for the user?

−25

−20

−15

−10

−5 0 5 10 15 20 25

−25

−20

−15

−10

−5 0 5 10 15 20 25

−25

−20

−15

−10

−5 0 5 10 15 20 25

−25

−20

−15

−10

−5 0 5 10 15 20 25

A) DJF %Change Storm Density

C) JJA %Change Storm Density

B) DJF %Change Storm Intensity

D) JJA %Change Storm Intensity

source: Sobolowski et al., 2014 (in prep, special issue Environmental Perspectives)

Storms

35

Collaboration

Greg Holland Cindy Bruyère

James Done Sherrie Fredrick

Tom Galarneau Ming Ge Abby Jaye Mari Jones Heather Lazrus Rebecca Morss Debasish PaiMazumder

Erin Towler Bill Skamarock

Michael Duda Laura Fowler

NCAR%Earth%System%Laboratory%

Na3onal%Center%for%Atmospheric%Research%

NCAR

Source: Dr. Cindy Bruyère (NCAR) – bruyerec@ucar.edu

37

COAWST MODELING SYSTEM

Coupled MCT

Ocean ROMS

Atmosphere WRF

Wave SWAN

Sediment Transport CSTMS

NCAR

Source: Dr. Cindy Bruyère (NCAR) – bruyerec@ucar.edu

WRF Grid

ROMS and SWAN Grid(s)

38

COAWST MODELING SYSTEM

NCAR

Source: Dr. Cindy Bruyère (NCAR) – bruyerec@ucar.edu

COAWST MODELING SYSTEM

39

NCAR

Source: Dr. Cindy Bruyère (NCAR) – bruyerec@ucar.edu

40

THE CYCLONE DAMAGE POTENTIAL (CDP) INDEX

CDP = 4

[( v

65 )

+ 5( R

50 )]

v

,

For v

> 65; if v

< 5, set v

= 5

Current Climate Future Change

NCAR

›  V

, the amount of energy dissipated at the surface by maximum winds;

›  R

, the radial extent of the surface wind field;

›  V

, the translational speed of the hurricane.

Source: Dr. Cindy Bruyère (NCAR) – bruyerec@ucar.edu

41

Thank you!

Dr. Michel d. S. Mesquita michel.mesquita@uni.no +47 55 58 38 18