1 / 56

The covariation of windstorm frequency, intensity and loss over Europe with large-scale climate diagnostics

The covariation of windstorm frequency, intensity and loss over Europe with large-scale climate diagnostics. A collaboration between SwissRe, MeteoSwiss, FP6 ENSEMBLES and NCCR Climate. 15.05.2008. Outline. The PreWiStoR project Predictability of European winter storminess

Samuel
Download Presentation

The covariation of windstorm frequency, intensity and loss over Europe with large-scale climate diagnostics

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. The covariation of windstorm frequency, intensity and loss over Europe with large-scale climate diagnostics A collaboration between SwissRe, MeteoSwiss, FP6 ENSEMBLES and NCCR Climate 15.05.2008

  2. Outline • The PreWiStoR project • Predictability of European winter storminess • Improved estimates of the European wind storm climate • Storm selection method • Improved estimates of loss due to European wind storms • The Swiss Re loss model • Calibration of ERA40, s2d and SwissRe storms • The covariation of wind storm frequency, intensity and loss over Europe with large-scale climate diagnostics • A bivariate extreme value peak over threshold model for wind storm intensity and loss

  3. PreWiStoR: Prediction of winter Wind Storm Risk • Problem: Observed records of wind storms are not long enough • Solution: ~150 storms based on observations. • Use probabilistic modelling to generate synthetic storms based on perturbed statistics • Calculate losses

  4. PreWiStoR: Prediction of winter Wind Storm Risk • Problem: Observed records of wind storms are not long enough • Solution: ~150 storms based on observations. • Use probabilistic modelling to generate synthetic storms based on perturbed statistics • Calculate losses

  5. PreWiStoR: Prediction of winter Wind Storm Risk • Problem: Observed records of wind storms are not long enough • Solution: ~150 storms based on observations. • Use probabilistic modelling to generate synthetic storms based on perturbed statistics • Calculate losses • New approach to use ENSEMBLE prediction systems (seasonal to decadal, s2d) • Replace statistical perturbation with physics • Utilise around ~500 seasons of S2D data • Obtain a better estimate of wind storm risk and losses See van den Brink et al. IJC (2005)

  6. PreWiStoR: Data • Seasonal to decadal (s2d) climate prediction models • Using the seasonal forecasting model of the ECMWF • A coupled ocean-atmosphere Global Circulation Model • 6-7 month forecast • Separate ocean analysis system to initiate the seasonal forecasts • ENSEMBLE prediction system: Model is run many times  Initial conditions are perturbed  Probabilistic Forecasts

  7. Monthly mean Geopotential Height @850hPa (m) ONDJFMA ERA40 Difference SYS 3

  8. Data Quality: Intercomparison of the 99th %-tile wind climate ERA40 ECMWF System 2 ECMWF System 3 Wind Gust WG Geostr. wind @ 850hPa GWS

  9. An Extreme Wind Index (EWI) • Spatial 95th percentile (calculated every 6 hours) • A measure of the extremity of lower bound of the spatial top 5% of wind • Applied to 850hPa Geostrophic Wind Speed (GWS) • Monthly averages taken for NDJFMA • Applied to ERA40 and Seasonal Forecasts

  10. Probabilistic prediction skill: ECMWF Sys2 • Ranked Probability Skill Score(terciles) • Bootstrap confidence intervals Little evidence of Predictabilty Initial Condition Pred. Nov Dec Jan Feb Mar Apr May

  11. Improved estimates of the European wind storm climate • Lack of predictability is disappointing, but the Seasonal Forecast data is still useful for risk assessment! • Remove first month from seasonal forecasts  independence of ensemble members • Join multiple forecasts together to form an ONDJFMA season

  12. Storm Selection Method Index: Q95 95% threshold Winter 1999/2000

  13. Number of wind storms identified in ERA-40 and s2d Example ERA-40 wind storm climatology

  14. Comparison of wind storm frequency • Wind storm climatologies are different in magnitude and shape • All s2d models seem to have a less negative shape than ERA-40

  15. Improved estimates of wind storm frequency and magnitude uncertainty 95% Confidence interval (profile log-likelihood) Return Level Return Period

  16. How can we compare the different climatologies? • Apply a calibration technique to the Q95 relying on different assumptions • Percentile based • A high threshold based • Mean based

  17. Example: percentile calibration curves SYS 3 SYS 2 DEMETER

  18. Frequency calibration: aliasing the data… • Each s2d dataset has a different temporal resolution of the Q95 • Has an effect on storm frequency, independent of model bias • Solution: Alias ERA-40 to the same temporal res. ERA-40, 6hr SYS3, 12hr SYS2, 12hr DEMETER, 24hr

  19. Percentile calibration and Aliasing

  20. 95th Percentile calibration and Aliasing

  21. GPD Parameters after calibration • Shape parameter is less negative • Aliasing has helped the frequency of occurrence (lambda)

  22. Summary: Storm intensity and storm frequency comparison • Large differences in storm intensities between SwissRe, ERA40 and s2d  need a calibration method... Necessarily a comprimise • -or- you believe the raw output of GCMs • Overall agreement in storm frequency between ERA40 and s2d, however, as shown before, aliasing of the signal is possible.

  23. Swiss Re Wind Storm Loss Model(catXos) • Vulnerability curve shows a cubic relation which is capped • Portfolio value distribution is inhomogeous

  24. The need for Calibration.... • ERA40 850hPa Geostrophic wind fields are different from SwissRe wind fields • SwissRe loss model is calibrated for use with SwissRe wind fields

  25. CALIB1: ERA40 GWS  SwissRE (*me2) • Adjustment curve: CDF(SwissRE)-CDF(ERA40) • Set to values greater than zero to zero

  26. CALIB2: Sys3 GWS  ERA40 GWS • Adjustment curve: CDF(ERA40)-CDF(Sys3 GWS)

  27. Comparison of Loss Return Periods • Calibrated wind storm wind fields including information on their duration is used as input to catXos • Error estimates from the calibration methodology can be used to estimate errors in loss • All loss return periods are expressed in %Total Insured Value (%TIV)

  28. Summary: Comparison of Loss Return Periods • All s2d datasets and ERA40 tend to indicate that the SwissRe underestimated the return period of loss between 1-5 years • For return periods > 40 years there is a tendency for SwissRe to overestimate the risk of loss • Uncertainty in the calibration estimates leads to large uncertainties in loss  bypass calibration by altering the vunerabilty in catXos • However, the use of s2d data has replaced statistical perturbation of storms (SwissRE) with dynamical perturbations (s2d)

  29. The covariation of wind storm frequency, intensity and loss over Europe with large-scale climate diagnostics • Hypothesis: Large-scale atmospheric state has an influence the frequency and magnitude of wind storms • As prediction of large-scale circulation improves in seasonal forecast models  improved estimates of storminess, a type of potential predictabilty... • S2d data maybe useful to determine the relationships since these relationships are determined using ERA40 or e.g. HadSLP i.e. Shorter than s2d • The chicken or the egg?  circular arguments

  30. Monthly mean Geopotential Height @850hPa (m) ONDJFMA ERA40 Difference SYS 3

  31. Parameters of the PCA • Performed on anomalies  monthly mean (previous slides) subtracted • Grid-points latitude weighted by the • Covariance matrix • pcaXcca CATtool • Five PCs chosen (will perform a Rule N check later) • PC loadings (EOFs) are scaled such that: • The length of the eigenvectors = eigenvalues • The PCs have mean of zero and a s.d of 1

  32. PC Loadings (EOF) GPH@850hPa anomalies ONDJFMA ERA40 SYS 3 Difference

  33. Vector Generalised Linear Models (VGLMs) • Extension of GLMs in that multivariate responses can be used • Allows modelling of the parameter of a chosen distribution as a function of the covariates • Applicable to distributions such as: Poisson, Gamma, GEV and GPD • R package VGAM, Yee & Stephenson (2007)

  34. A VGLM model of applied to the r-th largest  GEV distribution • ERA40 data • Could be used to explore observed variability (EMULATE) and decadal variability in s2d or C20C

  35. Exploratory analysis using Vector Generalised Additive Models (VGAMs) • Fit a smooth function in the vector generalised linear model • Allows non-linearity in relationships to be seen VGAM model VGLM model

  36. Storm Frequency Model: ERA40 D.F. Smoother = 1 D.F. Smoother = 2

  37. Storm Frequency Model: SYS 3 D.F. Smoother = 1 D.F. Smoother = 2 ?

  38. Storm Frequency Model: ERA40 Call: vglm(formula = COUNT ~ PC1 + PC2 + PC3 + PC4 + PC5 + SEAS.CYC, family = poissonff, data = datadf) Pearson Residuals: Min 1Q Median 3Q Max log(mu) -1.569 -0.5951 -0.07564 0.4592 2.612 Coefficients: Value Std. Error t value (Intercept) -0.744843 0.16803 -4.4329 PC1 0.291205 0.04045 7.1993 PC2 0.038500 0.04053 0.9499 PC3 0.237290 0.04117 5.7643 PC4 0.008085 0.04215 0.1918 PC5 0.023258 0.04248 0.5475 SEAS.CYC 0.647527 0.07881 8.2163

  39. Storm Frequency Model: SYS 3 Call: vglm(formula = COUNT ~ PC1 + PC2 + PC3 + PC4 + PC5 + SEAS.CYC, family = poissonff, data = datadf) Pearson Residuals: Min 1Q Median 3Q Max log(mu) -1.771 -0.6984 -0.1361 0.5543 4.712 Coefficients: Value Std. Error t value (Intercept) -1.01632 0.06254 -16.2516 PC1 0.14956 0.01766 8.4693 PC2 0.01769 0.01646 1.0744 PC3 0.18474 0.01682 10.9811 PC4 -0.01294 0.01722 -0.7512 PC5 0.00913 0.01679 0.5436 SEAS.CYC 0.82837 0.03274 25.3049

  40. Storm Frequency Model: ERA40 • Conditional frequency plots: Number of wind storms per month • Seasonal cycle held constant

  41. Storm Frequency Model: ERA40 • Conditional frequency plots: Number of wind storms per month • Remaining variable held constant • Given it is January: mean occurrence is ~2.4 • If PC1 is forecasted to be +2 • Then number of wind storms is likely to be ~ 4

  42. Storm Frequency Model: SYS 3 • Conditional frequency plots: Number of wind storms per month

  43. Summary: Storm frequency models • The NAO and the EAL are important for wind storm frequency • SYS 3 EAL is more strongly connected with storm freq. than ERA40 • SYS 3 NAO is less strongly connected with storm freq. than ERA40 • Formal likelihood ratio tests show that the seasonal cycle improves models • In the literature there is no framework on how to measure the “explained variance” of a GLM and VGLM/VGAM models, will investigate further  cross-validation • Calculation of conditional exceedance probabilities • Storm seriality: over-dispersion parameter of the Poisson GLM • Reperform calculations with the new storm selection (next section) • Adjust storm selection parameters so that ERA40 does not have as many storms (due to the 6hour time resolution)

  44. Storm Instensity Model: ERA40 Gamma Generalised Linear Model • Y= Monthly mean wind storm Q95 Gamma distribution VGLM model VGAM model

  45. Storm Intensity Model: ERA40 Call: vglm(formula = INTENSITY ~ PC1 + PC2 + PC3 + PC4 + PC5 + SEAS.CYC, family = gamma2, data = datadf) Pearson Residuals: Min 1Q Median 3Q Max log(mu) -1.978 -0.6599 -0.1958 0.5165 5.133 log(shape) -14.816 -0.1006 0.4248 0.6416 0.707 Coefficients: Value Std. Error t value (Intercept):1 2.393119 0.121266 19.7345 (Intercept):2 5.641005 0.088683 63.6085 PC1 0.014859 0.003678 4.0397 PC2 0.004446 0.003686 1.2060 PC3 0.004940 0.003784 1.3054 PC4 -0.010739 0.003703 -2.9004 PC5 -0.001216 0.003713 -0.3276 SEAS.CYC 0.033483 0.004177 8.0156 PC4: Negative influence of blocking

  46. Storm Intensity Model: SYS 3 Call: vglm(formula = INTENSITY ~ PC1 + PC2 + PC3 + PC4 + PC5 + SEAS.CYC, family = gamma2, data = datadf) Pearson Residuals: Min 1Q Median 3Q Max log(mu) -1.996 -0.7339 -0.1486 0.5368 7.478 log(shape) -29.975 -0.1507 0.3969 0.6383 0.707 Coefficients: Value Std. Error t value (Intercept):1 2.443144 0.037800 64.634 (Intercept):2 5.642231 0.035290 159.880 PC1 0.004222 0.001478 2.856 PC2 -0.003797 0.001438 -2.641 PC3 0.007121 0.001451 4.907 PC4 -0.001576 0.001490 -1.058 PC5 -0.002902 0.001459 -1.989 SEAS.CYC 0.031910 0.001200 26.593 PC3: EAL significant PC4: not significant (blocking biases in SYS3?)

  47. Storm Intensity Model: ERA40 • Conditional intensity plots: Monthly average Q95 (ms^-1) of wind storms

  48. Summary: Storm intensity models • In ERA40: +NAO and -blocking pattern are related to + storm intensity • In SYS 3: +NAO and +EAL pattern are related to + storm intensity • Differences could be due to longer dataset or biases in SYS 3? • Generally the statistical significance of intensity models is lower than with the frequency models

  49. Storm Loss Model: ERA40 Gamma Generalised Linear Model • Express total monthly loss as %TIV • Transform the loss data by the cube root (very long tailed dist) • Apply Gamma Generalised Linear Model

  50. Storm Loss Model: ERA40 & SYS 3 • Conditional loss plots: Monthly total cube-root of %TIV Lower influence of NAO on loss in SYS 3 (right) compared with ERA40 (left)

More Related