Stochastic Weather Generators
Transcription
Stochastic Weather Generators
Stochastic Weather Generators From WGEN to BayGEN Will Kleiber Department of Applied Mathematics University of Colorado Boulder, CO Workshop on Stochastic Weather Generators, Vannes, France May 17, 2016 Acknowledgements Balaji Rajagopalan Andrew Verdin Rick Katz Guillermo Podestá Federico Bert Branden Olson Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 1 / 21 Motivation “It is arguable that the artificiality of agricultural production systems make them less flexible, and therefore more vulnerable to climatic change than the naturally occurring species of the ecosystem within which they fit, and that the more unstable the climate the greater this vulnerability is likely to be.” - Oram (1985) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 2 / 21 Agriculture: Colorado Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 2 / 21 Agriculture: France near Paris Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 2 / 21 The Pampas: 750,000 km2 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN I Productive region: soybean, cereal, maize, wheat I Very flat: little subsurface flow, lack of drainage I Strong human systems and natural systems coupling SWG 2016 3 / 21 The Pampas: 750,000 km2 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 3 / 21 Pampas climate & land use change Alternating wet and dry epochs: I Early 1900s floods. I 1930-50 drought. I Increasing precipitation 1960-2000. I Agricultural expansion: 51 million metric tons of soy per year. I 1997-2003 flooding. I 2008 drought. 1200 1000 800 600 Total Precip (mm) → Extremes & frequency. 1960 1970 1980 1990 2000 2010 Year Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 4 / 21 Agricultural planning in the Pampas I Crop simulation models (DSSAT). I Analyze different management strategies. . . . the need for data. I I I Observed record limited. Respond to seasonal forecasts. All at ungauged locations. → Stochastic weather generation! Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 5 / 21 −33 Study region & data 3 5 2 −34 ● ● ● 6 4 ● 7 8 ● lat −35 ● ● 11 ● −36 ● 9 1 10 ● ● ● 16 ● 12 −37 13 ● 14 ● 17 ● −65 −64 −63 15 ● −62 −61 lon 1961–2013 daily precipitation, maximum and minimum temperature (area ≈ 200,000 km2 ). Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 6 / 21 Stochastic Weather Generators I Agricultural, ecological, hydrological models often require daily weather (e.g. precipitation, minimum/maximum temperature, solar radiation) I I On grid In the future I Stochastic Weather Generators (SWGs) can be used to produce infinitely long series of synthetic weather, for observation network infilling, or climate model downscaling I SWGs are statistical models whose simulated values “look like” observed weather I I I Daily statistics Interannual statistics SWGs are not forecast models Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 7 / 21 Short History Gabriel and Neumann (1962): First SWG for rainfall occurrence (persistence via Markov chain) p01 = P(rain today | no rain yesterday) p11 = P(rain today | rain yesterday) Buishand (1977): Geometric distributions for spell lengths Todorovic and Woolhiser (1975): Rainfall amounts (skewed distribution via exponential pdf) f (y|rain) = λ exp(−λy) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN for y>0 SWG 2016 8 / 21 Short History Richardson (1981), Richardson and Wright (1984): WGEN (precipitation, minimum temperature, maximum temperature, solar radiation) Given precipitation (non)-occurrence, TX (t)−µX (t) σX (t) TN (t)−µN (t) σN (t) R(t)−µR (t) σR (t) ∼ MVNormal(0, Σ) Ailliot et al. (2015) overview: focus on hierarchical models. Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 9 / 21 Multi-site weather generators Resampling: [Yates et al., 2003; Apipattanavis et al., 2007; Sharif & Burn, 2007] I Resample observed vector of variables. I Preserves (exact) spatial dependence and climatological statistics. I Restricted to historical extremes. I Difficult to simulate at ungauged locations. Model-based: [Wilks, 1998, 1999; Qian et al., 2002; Baigorria & Jones, 2010; Khalili et al., 2009] I Climate variables modeled consistent with single site. I Spatially-varying parameters. I Can be difficult to preserve space-time dependence. I Extendable to simulate at ungauged locations. Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 10 / 21 The Research Thread I WGEN (Richardson 1981; Richardson and Wright 1984) I Covariates via generalized linear models (Stern and Coe 1984; Chandler 2005; Furrer and Katz 2007) I Marginal spatial GLMs (Kleiber et al. 2012, 2013) I Joint SWGs (Verdin et al. 2015) I Actual seasonal/multidecadal forecasting (Verdin et al. 2016) I Fully Bayesian implementation (Verdin et al., in prep) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 11 / 21 Generic Methodology Precipitation occurrence O(s, t) = 1[WO (s,t)≥0] WO (s, t) ∼ GP(XO (s, t)T β O (s), CO ) Precipitation amount A(s, t) ∼ Gamma(αA (s), αA (s)/µA (s, t)) µA (s, t) = exp(XA (s, t)T β A ) Temperature ZN (s, t) = XN (s, t)T β N (s) + WN (s, t) ZX (s, t) = XX (s, t)T β X (s) + WX (s, t) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 12 / 21 Seasonal Forecasting Goal: Generate space-time weather scenarios consistent with seasonal forecasts and multidecadal trends, for resources planning and management. Include seasonal climate information in the form of additional covariates. Downscaling coarse scale information. Enables translation of seasonal forecasts or climate model projections for decision support systems. Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 13 / 21 Seasonal forecast (International Research Institute) 1. Sample ensemble of climatatology OND, A:N:B as weights. OND 2010 Precipitation: (15:35:50) (A:N:B) OND 2010 Temperature: (40:35:25) (A:N:B) 2. SWG simulation per ensemble member (covariates). 3. Ensemble of weather reflects uncertainty. Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 14 / 21 Seasonal forecast Differences in ensemble mean (unconditional minus conditional): Precip Max Temp Min Temp 0.6 100 0.4 50 0.5 0.2 0 0.0 0.0 −0.2 −50 −0.5 −0.4 −100 −0.6 Differences in 95% ensemble spread (unconditional minus conditional): Precip Will Kleiber (CU Applied Mathematics) Max Temp Min Temp 1.5 200 2 100 1 0.5 0 0 0.0 −100 −1 −0.5 −200 −2 1.0 −1.0 −1.5 WGEN to BayGEN SWG 2016 15 / 21 Seasonal forecast Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 16 / 21 BayGEN BayGEN: Incorporate parametric uncertainty in weather generation. I Variability between ensemble members I Better estimates of crop production risk I Better interannual statistics I Local climate estimates at ungauged locations Treat model parameters as random variables, posterior sampling propagates variability to weather simulations. β j (s) ∼ GP(β̂j (s), Cβj ) for j = O, A, N, X αA (s) ∼ GP(α̂A (s), CαA ) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 1 − Total Precipitation October 1 − Minimum Temperature October 1 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 2 − Total Precipitation October 2 − Minimum Temperature October 2 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 3 − Total Precipitation October 3 − Minimum Temperature October 3 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 4 − Total Precipitation October 4 − Minimum Temperature October 4 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 5 − Total Precipitation October 5 − Minimum Temperature October 5 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 6 − Total Precipitation October 6 − Minimum Temperature October 6 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 7 − Total Precipitation October 7 − Minimum Temperature October 7 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 8 − Total Precipitation October 8 − Minimum Temperature October 8 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 9 − Total Precipitation October 9 − Minimum Temperature October 9 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 10 − Total Precipitation October 10 − Minimum Temperature October 10 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 11 − Total Precipitation October 11 − Minimum Temperature October 11 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 12 − Total Precipitation October 12 − Minimum Temperature October 12 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 13 − Total Precipitation October 13 − Minimum Temperature October 13 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 14 − Total Precipitation October 14 − Minimum Temperature October 14 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 15 − Total Precipitation October 15 − Minimum Temperature October 15 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 16 − Total Precipitation October 16 − Minimum Temperature October 16 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 17 − Total Precipitation October 17 − Minimum Temperature October 17 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 18 − Total Precipitation October 18 − Minimum Temperature October 18 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 19 − Total Precipitation October 19 − Minimum Temperature October 19 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 20 − Total Precipitation October 20 − Minimum Temperature October 20 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 21 − Total Precipitation October 21 − Minimum Temperature October 21 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 22 − Total Precipitation October 22 − Minimum Temperature October 22 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 23 − Total Precipitation October 23 − Minimum Temperature October 23 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 24 − Total Precipitation October 24 − Minimum Temperature October 24 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 25 − Total Precipitation October 25 − Minimum Temperature October 25 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 26 − Total Precipitation October 26 − Minimum Temperature October 26 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 27 − Total Precipitation October 27 − Minimum Temperature October 27 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 28 − Total Precipitation October 28 − Minimum Temperature October 28 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 29 − Total Precipitation October 29 − Minimum Temperature October 29 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 30 − Total Precipitation October 30 − Minimum Temperature October 30 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Gridded simulation October 31 − Total Precipitation October 31 − Minimum Temperature October 31 − Maximum Temperature 30 162 40 92 25 35 53 20 30 17 30 15 25 10 10 6 4 20 5 15 2 0 10 1 Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 17 / 21 Extremes (a) Domain max of max temps (b) Domain min of max temps (c) Domain min of min temps (d) Domain max of min temps Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 18 / 21 Assessments of Risk I 100 weather trajectories from both GLMGEN and BayGEN I Propagated through DSSAT for soybean yield Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 19 / 21 Assessments of Risk I 100 weather trajectories from both GLMGEN and BayGEN I Propagated through DSSAT for soybean yield Probability of < 2500 kg ha−1 (break even yield): I ≈ 17% (GLMGEN) I ≈ 32% (BayGEN) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 19 / 21 Discussion I GLM space-time weather generator (GLMGEN).1 I I Conditional GLMGEN.2 I I Complete spatial simulations of “weather” Downscaling and seasonal forecasts Bayesian space-time weather generator (BayGEN).3 I 1 Verdin More comprehensive uncertainty quantification et al., (2015). Coupled stochastic weather generation using spatial and generalized linear models. Stochastic Environmental Research and Risk Assessment, 29(2), 347–356. 2 Verdin et al., (2016). A conditional stochastic weather generator for seasonal to multi-decadal simulations. Journal of Hydrology. 3 Verdin et al., (2016, In Preparation). BayGEN: A Bayesian space-time stochastic weather generator. In preparation. Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 20 / 21 What might the future hold? I Incorporate other variables (e.g., wind, solar, relative humidity) I High time resolution simulation (see later: Koch, Sun) I Physically constrained simulation (see later: Groyer, Guilloteau, Bessac, Atencia) I Large-domain applications (see later: Sommer) Will Kleiber (CU Applied Mathematics) WGEN to BayGEN SWG 2016 21 / 21