Reduced Overdispersion in Stochastic Weather Generators for Statistical Downscaling of Seasonal Forecasts and Climate Change Scenarios Yongku Kim Institute for Mathematics Applied to Geosciences National Center for Atmospheric Research Joint work with R. W. Katz, B. Rajagopalan and G. Podestá Supported by NSF 2009 AGU Fall Meeting, 14-18 December, San Francisco 1
Outline Challenge/Motivation GLM Stochastic Weather Generator Downscaling of Seasonally Aggregated Climate Overdispersion Phenomenon Applications Downscaling Seasonal Forecasts Downscaling Climate Change Scenarios Possible Extensions 2
Utility of Weather Generators Historical Data Synthetic series 28.5 12.4 23.1 10.2 29.1 11.4 25.8 9.7 Process model Crop Watershed Erosion Frequency distribution of outcomes (Decision Making) 3
Challenge/Motivation Weather generation in the Pampas region of Argentina A wet season in the Southern Hemisphere summer impact to crop models. Interactions between changing climate and technological innovations in agricultural decision making. Study Region 4
GLM Stochastic Weather Generator Richardson stochastic weather generator (e.g., Richardson WRR 1981). Precipitation Occurrence: chain-dependent process (i.e. Markov chain for occurrence). Precipitation Intensity: independent, identically distributed conditional on occurrence. Min. and Max. temperature: bivariate first-order autoregressive process. Dependence between temperature and precipitation. Overdispersion phenomenon. Computationally intense and unwieldy to modify it to downscaling. 5
GLM Stochastic Weather Generator Furrer & Katz(Clim. Res. 2007). Introduction of covariates such as a seasonal cycle and/or an atmospheric index. e.g., The probability that it rains on a specific day is modeled depending on the day of the year and the ENSO. A separate model fit for each month of the year and/or for each ENSO category. All glm model selection via likelihood ratio test, AIC or BIC. A limitation of parametric stochastic weather generators: A marked tendency to underestimate the observed interannual variance of seasonally aggregated variables (e.g., Katz and Parlange (1998)). Webpage: Generalized linear model (GLM) weather generator, see http://www.image.ucar.edu/ eva/glmwgen/. 6
GLM Stochastic Weather Generator Downscaling of Seasonally Aggregated Climate To reduce overdispersion phenomenon, incorporate time series of seasonal total precipitation and seasonal mean minimum and maximum temperature in the GLM weather generator as covariates (Summer: October-March, Winter: April-September). These seasonal time series are smoothed using locally weighted scatterplot smoothing (LOESS) to avoid introducing underdispersion. 7
GLM Stochastic Weather Generator Precipitation occurrence: Logistic GLM for p t = P r{j t = 1}, ln[p t /(1 p t )] = β 0 + β 1 J t 1 homogeneous Markov chain + β 2 Cos t + β 3 Sin t parallel sine waves + β 4 J t 1 Cos t + β 5 J t 1 Sin t separate sine waves + β 6 I t S t summer total precip. (I t = 1) + β 7 (1 I t )W t winter total precip. (I t = 0) Treats both transition probabilities P 01 and P 11 simultaneously, incorporating common and separate signals. Precipitation intensity: Gamma GLM for mean precipitation intensity lnµ t = γ 0 + γ 1 Cos t + γ 2 Sin t + γ 3 I t S t + γ 4 (I t 1)W t. 8
GLM Stochastic Weather Generator Temperature: Coupled univariate models for minimum (X t ) and maximum (Y t ) daily temperature, essentially equivalent to bivariate AR(1) X t = α 0 + α 1 J t dependence on precipitation + α 2 X t 1 + α 3 Y t 1 autoregressive / crosscorrelation term + α 4 Cos t + α 5 Sin t seasonal cycle + α 6 I t S t + α 7 (I t 1)W t + ε t (X) seasonal mean minimum temp., error Y t = λ 0 + λ 1 J t dependence on precipitation + λ 2 Y t 1 + λ 3 X t autoregressive / crosscorrelation term + λ 4 Cos t + λ 5 Sin t seasonal cycle + λ 6 I t S t + λ 7 (I t 1)W t + ε t (Y ) seasonal mean maximum temp., error The normally distributed error terms ε t (X) and ε t (Y ) are uncorrelated, the two models are fit separately. 9
GLM Stochastic Weather Generator Overdispersion Phenomenon Standard deviation 100 200 300 400 I: original II:smoothed III:unsmoothed I II III Model boxplots of the inter-annual SDs for precipitation (red lines: observed inter-annual SDs) 10
Standard deviation 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 I: original II:smoothed III:unsmoothed IV:temporal trend I II III IV Model boxplots of the inter-annual SDs for minimum temperature (red lines: observed inter-annual SDs) 11
Standard deviation 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 I: original II:smoothed III:unsmoothed IV:temporal trend I II III IV Model boxplots of the inter-annual SDs for maximum temperature (red lines: observed inter-annual SDs) 12
Applications Downscaling Seasonal Forecasts Seasonal climate forecasts based on multiple global climate models (e.g., IRI: upto 6-9 months in 3-month moving windows). Historical years are classified into three categories (wet/normal/dry) based on the terciles of the smoothed historical seasonal precipitation. The years are resampled with replacement with the probabilistic forecast as the weight metric (e.g., 40(wet):35(normal):25(dry) forecast). The smoothed seasonal precipitation values of the resampled years are plugged in the GLM weather generator as covariates. IRI Forecast 13
Applications density 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 Model Climatology Seasonal Forecasts Observed density 0.0 0.1 0.2 0.3 0.4 0.5 Model Climatology Seasonal Forecasts Observed density 0.00 0.01 0.02 0.03 0.04 Model Climatology Seasonal Forecasts 200 400 600 800 1000 1200 1400 1600 12 13 14 15 16 17 18 0 20 40 60 precipitation minimum tempertature number of hot days Pilar region based on IRI seasonal forecast 2003 (dry year) 14
density 0.0000 0.0005 0.0010 0.0015 0.0020 0.0025 Model Climatology Seasonal Forecasts Observed density 0.0 0.1 0.2 0.3 0.4 0.5 Model Climatology Seasonal Forecasts Observed density 0.00 0.01 0.02 0.03 0.04 Model Climatology Seasonal Forecasts 200 400 600 800 1000 1200 1400 1600 12 13 14 15 16 17 18 0 20 40 60 precipitation minimum tempertature number of hot days Pilar region based on IRI seasonal forecast 2000 (wet year) 15
Applications Downscaling Climate Change Scenarios Climate change projections for the 21st century are only reliably available at monthly time scale and for a spatial scale given by the model grid-size based on an ensemble of global climate change models (IPCC AR4). Our approach can also be used for downscaling short term (i.e., decadal) projections. To demonstrate this, we select an earlier dry epoch (1931-1955). The smoothed precipitation and temperature for each year of these epochs were used a covariates to the GLM weather generator to produce daily weather sequences consistent with the decadal variability of these epochs. 16
Applications precipitation 200 400 600 800 1000 1200 Observed Mean Projection 95% Projection Band minimum temperature 11 12 13 14 15 16 Observed Mean Projection 95% Projection Band maximum temperature 25 26 27 28 29 30 Observed Mean Projection 95% Projection Band 1935 1940 1945 1950 1955 1935 1940 1945 1950 1955 1935 1940 1945 1950 1955 year year year Simulated means and 95th projection band during earlier dry epoch (1931-1955)for Pergamino region 17
density 0.0000 0.0010 0.0020 Model Climatology Scenario Projections density 0.0 0.1 0.2 0.3 0.4 0.5 Model Climatology Scenario Prjections density 0.00 0.02 0.04 0.06 Model Climatology Scenario Projections 500 1000 1500 2000 11 12 13 14 15 16 17 0 10 20 30 40 precipitation minimum temperature number of hot days Pergamino region during earlier dry epoch (1931-1955) 18
Possible Extensions Multisite: consider a spatial structure in the GLM. Weather variability: variance and autocorrelation constrained to be constant. Could be addressed by introducing winter/summer differences. Extremes: high precipitation amounts are not sufficiently reproduced by the model. Apply extreme value technique using, for example, the stretched exponential distribution. Hidden Markov modeling approach: Consider a hidden regime for each GLM stochastic process. Uncertainty analysis: assess and report uncertainty of the generated weather. 19