1. Attributing Variation in a Regional Climate-change Modelling Experiment Chris Ferro Centre for Global Atmospheric Modelling Department of Meteorology University of Reading, UK CRU Seminar, Norwich, 16 December 2004

2. PRUDENCE Statistical model Illustrative example Grid-point analysis Discussion

3. The PRUDENCE Project European climate simulations 9 limited-area RCMs, driven by various global GCMs Control 1961–1990 Scenarios 2071–2100 http://prudence.dmi.dk Orography at 50km resolution

4. Experimental Design Observed SST/SICE AGCM RCM Observed Emissions SRES A2 Emissions AOGCM RCM AGCM RCM ControlScenario

5. Aims Characterise projected climate change effects of model differences relative importance of different model components: emissions, GCM, RCM and initial conditions

6. HADCM3 /OPYC3 / HADAM3HECHAM4 RCMCACAGroup CHRM11ETHZ CLM11GKSS HADRM11HC HIRHAM1111DMI PROMES11UCM RACMO11KNMI RCAO1111SMHI REGCM11ICTP REMO11MPI Selected Experiments

7. Land-averaged annual mean 2m air temperature interpolated to CRU grid ECHAMHADAM HIRHAM RCAO HIRHAM RCAO Temperature (°C) Year

8. General Linear Model T i j k l is temperature for GCM i, RCM j, Period k, Year l T i j k l =  i j k +  i j k t l +  i j k t l 2 + Z i j k l Residuals Z i j k l ∼ N(0,  2 i j k ), cor(Z i j k l, Z i j k l ) =  i k ( j, j )

9. Model Components T i j k l =  i j k +  i j k t l +  i j k t l 2 + Z i j k l  i j k =  +  i G +  j R +  k P +  i j GR +  i k GP +  j k RP +  i j k GRP  overall mean  i G effect of GCM i  j R effect of RCM j  k P effect of Period k  i j GR effect of combining GCM i with RCM j  i k GP effect of GCM i in Period k  j k RP effect of RCM j in Period k  i j k GRP effect of combining GCM i with RCM j in Period k

10. Simplified Model T i j k l =  +  i G +  j R +  k P +  i j GR +  i k GP + (  +  i G +  j R +  k P ) t l + (  +  j R ) t l 2 + Z i j k l  12.1(0.03)  -0.3(0.15) 1G1G 0.07(0.04) 1G1G 0.10(0.04) 1R1R -0.33(0.01) 1R1R 0.11(0.05) 1P1P -2.22(0.04) 1P1P -0.26(0.04)  11 GR -0.13(0.02)  0.21(0.05)  11 GP -0.31(0.05) 1R1R -0.03(0.02)  i j k CA HIRHAM-E0.560.62 RCAO-E0.550.51 HIRHAM-H0.500.43 RCAO-H0.460.52  i k CA ECHAM0.92-0.10 HADAM0.880.82

11. Diagnostic Plots ECHAM HADAM HIRHAM RCAO QQ PP

12. Variance Partition % Trend91.0 GCM0.1 RCM2.0 G x T1.8 R x T0.0 G x R0.3 G x R x T– Z4.7 ECHAM HADAM HIRHAM RCAO

13. Temperature Contrasts (  C) HIRHAM RCAO HADAM ECHAM HIRHAM RCAO HADAM warmer than ECHAM at 1961 becomes progressively cooler – warm bias greater for HIRHAM RCAO warmer than HIRHAM – warm bias greater for ECHAM

14. Response Contrasts HADAM – ECHAM response -0.19  C/decade (-0.37, -0.02) HADAM response lower than ECHAM independently of RCM, Period, Year RCAO response lower than HIRHAM at start of integration, becomes progressively higher independently of Period, GCM Year  C / decade

15. Grid-point Analysis Fit model separately at each grid point and plot maps: Proportion of variation explained by each model term Evolution of GCM temperature contrasts for each RCM Evolution of RCM temperature contrasts for each GCM Evolution of GCM response contrasts for each RCM Evolution of RCM response contrasts for each GCM

16. Variation Explained (%) model T Z GR GR GTRTGRT

17. Changes in Mean Temperature (°C) HIRHAM RCAO ECHAMHADAM

18. Temperature Contrasts: HADAM – ECHAM 1990196120712100 HIRHAM RCAO

19. Response Contrasts: HADAM – ECHAM 1990196120712100 HIRHAM RCAO

20. Temperature Contrasts: RCAO – HIRHAM 1990196120712100 ECHAM HADAM

21. Response Contrasts: RCAO - HIRHAM 1990196120712100 ECHAM HADAM

22. HADCM3 /OPYC3 / HADAM3HECHAM4 RCMCACAGroup CHRM11ETHZ CLM11GKSS HADRM11HC HIRHAM1111DMI PROMES11UCM RACMO11KNMI RCAO1111SMHI REGCM11ICTP REMO11MPI Selected Experiments

23. Residual Standard Deviation (⁰C) CLMRACMOCHRMREMOPROMESHIRHAMRCAOREGCMHADRM 0.380.39 0.400.440.450.460.48

24. Temperature Comparisons (⁰C)

25. Response Comparisons (⁰C/century)

26. Review Method: statistical model synthesises output and enables inference on climate changes, model differences, and the relative importance of model components. Results: warming signal dominates, choice of RCM affects temperature in mountainous regions, choice of GCM affects temperature and response in Atlantic and response over land, inter-annual variation is greater than GCM and RCM effects over land, effects of model differences change over space and time, other RCMs have different behaviour Extensions: RCMs, annual cycle, model resolution, EOFs, multivariate, spatial, ensemble generation

27. Discussion Danger of over-simplified summaries of model output Features can be hidden by poor experimental design If you were a politician… http://www.met.rdg.ac.uk/~sws02caf c.a.t.ferro@reading.ac.uk