2. HOWGH

Stochastic Hourly Weather Generator HOWGH: validation and its Use in Pest Modelling under

Present and Future Climates

Martin Dubrovsky 1,4, Martin Hirschi 2,3, Christoph Spirig 3

Acknowledgements: The first version of HOWGH weather generator was developed within the frame of bilateral collaboration between IAP and MeteoSwiss. Now it is being improved and validated within the frame of WG4VALUE

(project LD12029 sponsored by the Ministry of Education, Youth and Sports of CR) and VALUE (COST ES 1102 action).

(1) Institute of Atmospheric Physics ASCR, Prague, Czech Republic (dub@ufa.cas.cz)(2) Institute for Atmospheric and Climate Science, ETH Zurich, Switzerland

(3) Federal Office for Meteorology and Climatology MeteoSwiss, Zurich, Switzerland(4) CzechGlobe - Center for Global Climate Change Impacts Studies, Brno, Czech Republic

1. IntroductionTo quantify impact of the climate change on a specific pest (or any other weather-dependent process) in a specific site, we may use a site-calibrated pest (or other) model and compare its outputs obtained with site-specific weather data representing present vs. perturbed climates. The input weather data may be produced by the stochastic weather generator. Apart from the quality of the pest model, the reliability of the results obtained in such experiment depend on an ability of the generator to represent the statistical structure of the real world weather series, and on the sensitivity of the pest model to possible imperfections of the generator.

This contribution deals with the multivariate HOWGH weather generator, which is based on a combination of parametric and non-parametric statistical methods.

The contribution presents results of the direct and indirect validation of HOWGH. In the direct validation, the synthetic series generated by HOWGH (various settings of its underlying model are assumed) are validated in terms of (i) mean daily cycle of temperature and precipitation, and (ii) exceedances of hourly values over given temperature and precipitation thresholds. In the indirect validation, we assess the generator in terms of characteristics derived from the outputs of dynamic pest model SOPRA (simulating the development of codling moth) fed by the observed vs. synthetic series.

To demonstrate a potential of HOWGH for climate change impact experiments, the results of codling moth simulations for future climate are shown.

GC23D-0662 *** 16.12.2014 (Tuesday) *** AGU2014

madu1110@gmail.com this poster: www.ufa.cas.cz/dub/prasce/2014-agu-howgh~martin.pdf

Fig.1 Scheme of HOWGH

learning series consistingof daily values + hourly values

1981-01-01: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..231981-01-02: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..231981-01-03: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..231981-01-04: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..231981-01-05: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..231981-01-06: [SRAD,TAVG,PREC], [SRAD,TAVG,PREC]i=0..23

………

2009-12-30: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-31: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-30: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-31: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-30: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-31: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..23

calibrated M&Rfi WG= WG parameters:

PREC occurrence: Markov chainPREC amount: Gamma distribution

(SRAD, TMAX, TMIN) ~ AR(1)

observed hourly series1981-01-01:00 [SRAD, TAVG, PREC]1981-01-01:02 [SRAD, TAVG, PREC]

…2009-12-31:22 [SRAD, TAVG, PREC]2009-12-31:23 [SRAD, TAVG, PREC]

daily series:1981-01-01: [SRAD, TAVG, PREC]1981-01-02: [SRAD, TAVG, PREC]

…2009-12-31: [SRAD, TAVG, PREC]

synthetic daily series1981-01-01: [SRAD TAVG PREC]1981-01-02: [SRAD TAVG PREC]

…2009-12-31: [SRAD TAVG PREC]

synthetic daily + hourly series1981-01-01: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..231981-01-02: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..231981-01-03: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..23

……

2009-12-30: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..232009-12-31: [SRAD TAVG PREC], [SRAD TAVG PREC]i=0..23

synthetic hourly series1981-01-01: [SRAD TAVG PREC]i=0..231981-01-02: [SRAD TAVG PREC]i=0..231981-01-03: [SRAD TAVG PREC]i=0..23

……

2009-12-30: [SRAD TAVG PREC]i=0..232009-12-31: [SRAD TAVG PREC]i=0..23

2. sampling (adding hourlies to dailies)

3. adjustinghourlies to dailies

• is based on a combination of parametric and non-parametric statistical approaches

• it allows to generate optional number of weather variables

• here it is used it to generate hourly series of three weather variables required by SOPRA pest model:

- solar radiation- temperature- precipitation

The algorithm consists of three steps:(1) Generating daily series by M&Rfi weather

generator, which is based mostly on parametric modelling:§ PREC occurrence ~ Markov model§ PREC amount ~ Gamma distribution

§ SRAD + 2 temperature variables ~ 1st-order AutoRegressive model]§ 3 versions of 2 temperature variables:

- TMAX & TMIN- TAVG & DTR- TMID & DTR

- [TMID = (TMAX+TMIN)/2]- DTR is optionally normalized by quantile-mapping transformation

(2) For each single day generated by M&Rfi, the hourly values are sampled from the observational hourly database

(3) The sampled hourly values are fitted to driving daily values generated by M&Rfi

3. Observed weather data: Wädenswil (at Zurichsee), Switzerland; 1981-2009

4. Finetuning the weather generatorThe performance of the whole algorithm may be adjusted through various settings of driving WG M&Rfi (e.g. order of the Markov chain), settings of the resampling algorithm (e.g. definition of the distance), and the hourlies-to-dailies fitting algorithm. Here we test:

4.1 Three representations of temperature in M&Rfi:“XN” : TMAX & TMIN“TD”, “TDq” : TAVG & DTR (DTR = TMAX-TMIN); “TDq” indicates

use of “normalized” DTR obtained by quantile mapping of DTR

“MD”, “MDq”: TMID & DTR [TMID = (TMAX+TMIN)/2]

4.2 Various settings of the procedure for adjusting hourlies to dailies (see Fig. 2)a) temperature (not all options listed below are available for all temperature representations; e.g. AVG fit is not applicable with XN representation)

NoFit: hourly values are NOT fitted to daily valuesAVG-fit: hourly values are additively modified to fit TAVG generated

by daily WG (dWG)DTR-fit: daily cycle is scaled to fit DTR generated by dWG(A+D)-fit: both averages and daily temperature range are fittedsteps: interdiurnal steps are reducedALL-fit: (AVG + DTR + steps) fitted

b) precipitationNoFit: hourly values are NOT fitted to daily values A-fit: P’(h) = P(h) * k …[adjustment is proportional to P(h)]C-fit: P’(h) = P(h) + k2*Pclim(h) [adjustment proportional to Pclim(h)]Q-fit: P’(h) = P(h) * [1 + k3*Pclim(h)] …[adjustment is proportional to

P(h)*Pclim(h)]where: P(h) and P’(h) is hourly precipitation before and afteradjustment and Pclim(h) is climatological value of the hourly

precipitation

Figs. 3a-b: Effect of the HOWGH settings on mean daily cycle of PREC and TEMP

(based on 29-year observed series vs. 30-year synthetic series)

Fig.2: Fitting sampled hourly temperatures to driving daily valuesmotivation: Daily averages/sums of the sampled hourly values do not fit exactly values produced by daily WG. In result, the structure of the daily series aggregated from the hourlies might be misreproduced. The fitting procedure is applied to sampled hourly series to improvereproduction of driving daily series.

DWG: daily mean temperatures generated by daily WGsampled: sampled hourly series with no fittings appliedAVG fit, AVG + DTR fit, AVG + steps, ALL fitted: various degrees of

hourly-to-daily fitting

Fig.4b Exceedances of hourly precipitation sums above thresholdA: avg ( N ) B: Prob (day with N>1 ) C: Prob (day with Spell )

Fig.4a: Exceedances of hourly temperatures above / below thresholdA: avg ( N ) B: Prob (day with N ≥ 1 ) C: Prob (day with Spell )

PREC

≥10

mm

PREC

> 0

mm

PREC

≥1

mm

PREC

≥3

mm

TEM

P ≥

25 °C

TEM

P ≥

-30

°CTE

MP

≥32

°CTE

MP

≤-5

°C

ConclusionsThe HOWGH generator produces synthetic single-site hourly weather series. It is based on the parametric daily weather generator M&Rfi, whose output is disaggregated into hourly values using the resampling approach (Fig.1).The poster focuses on effects of (i) representing temperature in the daily WG, and (ii) adjustments to hourly values [Fig.2] which aim to fit sampled hourly values to daily values produced by M&Rfi. The present results indicate:

(i) Representation of temperature in M&Rfi has only minor effect on the mean daily cycles of hourly values of TEMP and PREC (Fig.3a-b). However, the use of TAVG & DTR combination shows better results than TMAX &TMIN in reproducing exceedances of high temperature thresholds (Fig.4a).

(ii) The hourly-to-daily adjustments slightly improve reproduction of the daily cycle of temperature (Fig.3a) and high temperature threshold exceedances (Fig.4a). On the other hand, these adjustments have negative effect on reproduction of daily precipitation cycle (Fig.3b) as well as on the exceedances of the precipitation thresholds (Fig.4b).

The experiments made with the pest model run with the synthetic weather series produced by HOWGH generator show good concordance between pest model output obtained with observed vs. synthetic weather series [Figs. 5a-b].

•representation of temperature in the daily WG: XN: TMAX & TMIN, TD, TDq ~ TAVG & DTR; MD, MDq ~ TMID & DTR (see section 4 for details) •hourly-to-daily adjustments: NoFit ~ no fitting procedure applied; OptFit ~ we did our best to fit what was possible to fit!

legend

•representation of temperature in the daily WG: XN: TMAX & TMIN, TD ~ TAVG & DTR (see section 4 for details) •hourly-to-daily adjustments: NoFit ~ no fitting procedure applied; A-fit, C-fit: see Section 4 for definition

legend

6. HOWGH & pest model

Fig.5a: Indirect validation of HOWGH via modelled Flight start of codling moth

observed and synthetic weather today (mean dates*)

Fig.5b: Indirect validation of HOWGH in terms of modelled Codling moth life phases

5. Validation of HOWGH in terms ofsub-daily spells

Figures 4a and 4b show 3 statistics (~ 3 columns) related to over-threshold exceedances of hourly temperatures and precipitation based on 29year observed (OBS) and 30-year synthetic weather series generated with various temperature representation and various settings of hourly-to-daily fitting procedure (see Box 4 for explanation of XN, TD, TDq, MD and MDq).

A. avg(N) = mean number (per-day) of hourly values exceeding the threshold

B. Prob (day with N ≥1) = probability that at least one hourly value exceeds the threshold during a single day

C. Prob (day with Spell) = probability that a single day includes a spell, which is defined here as a series of at least 3 exceedances of hourly values

Fig.6: Risk of the 3rd codling moth generation under reference (1980–2009) and future climates

Input weather data for this CC impact experiment were created by modifying the synthetis series generated by HOWGH using the CH2011 climate change scenarios for RCP3PD (yellow), A1B (grey), and A2 (purple) greenhouse gas scenarios..

referencesCH2014-Impacts (2014), Toward Quantitative Scenarios of Climate Change

Impacts in Switzerland, published by OCCR, FOEN, MeteoSwiss, C2SM, Agroscope, and ProClim, Bern, Switzerland, 136 pp. ISBN 978-3-033-04406-7; http://ch2014-impacts.ch/index.php?lang=en&id=report

Hirschi M. et al (2012) Downscaling climate change scenarios for apple pest and disease modeling in Switzerland. Earth Syst. Dynam., 3, 33–47, 2012 (www.earth-syst-dynam.net/3/33/2012/) (doi:10.5194/esd-3-33-2012)

Wädenswil

Magadino