Post-traitement statistique des prévisions d'ensemble à Météo-France - Meteo France
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Post-traitement statistique des prévisions d’ensemble à Météo-France Maxime Taillardat maxime.taillardat@meteo.fr 12 mars 2019
Motivation of post-processing ensemble forecasts Run: 2012−10−26 Lead time: 12 h Rank of the observed temperature: 18 Météo-France’s forecast 35-members global observed ensemble system −10 0 10 20 30 (PEARP), 10 km Temperature (°C) resolution over France. Observations and forecasts of 2-m temperature (T2m) at Lyon-Bron for the run of 1800UTC (different lead times) Maxime Taillardat 1/16
Motivation of post-processing ensemble forecasts Météo-France’s 35-members global ensemble system (PEARP), 10 km resolution over France. Observations and forecasts of 2-m temperature (T2m) at Lyon-Bron for the run of 1800UTC (different ◮ Need of a simultaneous correction of bias and lead times) dispersion ◮ Whatever the quality of the raw ensemble, post-processing improves forecast attributes (Hemri, 2014) Maxime Taillardat 1/16
The state of the art ◮ Existing methods: Analogs, Perfect Prog., Rank-based matching, CDF-matching, Member dressing, Bayesian Model Averaging, SAMOS... ◮ A review of techniques: Gneiting, 2014 The most (famous // widely-used // efficient) post-processing method : ◮ Ensemble model output statistics (EMOS) (Gneiting, 2005) also called Non-homogeneous Regression (NR) ◮ fitting parameters linearly on predictors on some training period: X N f (y |x1 , . . . , xN ) = N (µ = a0 + ai xi , σ2 = b + cs2 ) i=1 y response variable, x1 , . . . , xN ensemble member values or any other predictor, s2 ensemble variance Maxime Taillardat 2/16
Benefits of non-parametric post-processing No assumptions on the weather variable you deal with ◮ Raw ensemble (not EMOS post-processed) : Non-parametric Maxime Taillardat 3/16
Benefits of non-parametric post-processing No assumptions on the weather variable you deal with ◮ Raw ensemble (not EMOS post-processed) : ◮ EMOS post-processing : Non-parametric Maxime Taillardat 3/16
Benefits of non-parametric post-processing No assumptions on the weather variable you deal with ◮ Raw ensemble (not EMOS post-processed) : ◮ EMOS post-processing : Non-parametric ◮ Post-processing possible results : Maxime Taillardat 3/16
Benefits of non-parametric post-processing No assumptions on the weather variable you deal with ◮ Raw ensemble (not EMOS post-processed) : ◮ EMOS post-processing : Non-parametric ◮ Post-processing possible results : Maxime Taillardat 3/16
Best-of post-processing possible outputs (initialization: 18 UTC) Maxime Taillardat 4/16
From binary regression trees to QRF ◮ Example: CART algorithm (Breiman, 1984) Taillardat, Maxime, Olivier Mestre, Michaël Zamo, and Philippe Naveau. "Calibrated ensemble forecasts using quantile regression forests and ensemble model output statistics." Monthly Weather Review 144, no. 6 (2016): 2375-2393. Maxime Taillardat 5/16
How to verify/evaluate ensemble forecasts ? ◮ Differ from deterministic forecasts ◮ A "point" (eg. for one day) verification is a nonsense ◮ It has to be statistical Several attributes sought ◮ Reliability ◮ Accordance between forecasted probabilities and observed frequencies of an event and/or exchangeability between observations and ensemble members ◮ Resolution/Discrimination ◮ Ability to differ from a climatological forecast ◮ Sharpness ◮ Getting the least dispersed forecast "the paradigm of maximizing the sharpness of the predictive distributions subject to calibration"(Gneiting et al. 2006) Maxime Taillardat 6/16
How to verify/evaluate ensemble forecasts ? ◮ Differ from deterministic forecasts ◮ A "point" (eg. for one day) verification is a nonsense ◮ It has to be statistical Several attributes sought ◮ Reliability ◮ Accordance between forecasted probabilities and observed frequencies of an event and/or exchangeability between observations and ensemble members ◮ Resolution/Discrimination ◮ Ability to differ from a climatological forecast ◮ Sharpness ◮ Getting the least dispersed forecast "maximizing the value of the predictive distributions subject to good overall performance" Maxime Taillardat 6/16
"Here comes the rain again..." Production constraints ◮ Relatively shallow datasets ( 1 up to 3 years) ◮ Costly reforecasts Possible alternatives ◮ Data-shifting: Work with anomalies (obs −→ obs − ensemble mean) ◮ Data-pooling: Regionalization/Neighborhood ◮ Data-boosting: Work with several observations/distribution Maxime Taillardat 7/16
"Here comes the rain again..." Example: post-processing RR6 (6-h rainfall) ◮ This variable is (by far !) the most difficult to calibrate whatever the method, a really hard nut to crack for statisticians ◮ A lot of zeros, a lot of "extremes" What we have tested For EMOS: ◮ Try different distributions "tailored" for extremes (Gamma, GEV...) ◮ Try different sets of predictors For QRF: ◮ Improve QRF for "quantiles": Gradient Forests (GF) (Athey et al., 2017) ◮ Extrapolation in distributions tails: semi-parametric approach Maxime Taillardat 8/16
A semi-parametric approach Goal: Be skillful for extremes without degrading overall performance Use QRF/GF outputs to fit a distribution which would: ◮ Model jointly low, moderate and heavy rainfall ◮ Be flexible ◮ Use of an Extended GP distribution (EGP3) (Papastathopoulos and Tawn, 2013 ; Naveau et al., 2016 ; Tencaliec’s PhD 2018) Maxime Taillardat 9/16
A semi-parametric approach Goal: Be skillful for extremes without degrading overall performance What we expect: ◮ QRF/GF possible output Maxime Taillardat 9/16
A semi-parametric approach Goal: Be skillful for extremes without degrading overall performance What we expect: ◮ QRF/GF possible output ◮ After "EGP TAIL" Maxime Taillardat 9/16
Results on RR6 Raw ensemble Analogs Analogs_C Analogs_COR 0.20 0.20 0.20 0.20 JPZ test : 0 % JPZ test : 85 % JPZ test : 100 % JPZ test : 99 % 0.15 0.15 0.15 0.15 0.10 0.10 0.10 0.10 0.05 0.05 0.05 0.05 0.00 0.00 0.00 0.00 1 3 5 7 9 11 Analogs_VSF EMOS CSG EMOS EGP EMOS GEV 0.20 0.20 0.20 0.20 JPZ test : 98 % JPZ test : 89 % JPZ test : 11 % JPZ test : 98 % 0.15 0.15 0.15 0.15 0.10 0.10 0.10 0.10 0.05 0.05 0.05 0.05 0.00 0.00 0.00 0.00 QRF QRF EGP TAIL GF GF EGP TAIL 0.20 0.20 0.20 0.20 JPZ test : 100 % JPZ test : 91 % JPZ test : 98 % JPZ test : 63 % 0.15 0.15 0.15 0.15 0.10 0.10 0.10 0.10 0.05 0.05 0.05 0.05 0.00 0.00 0.00 0.00 Maxime Taillardat 10/16
ROC curves ◮ Value: Focus on the upper left corner ◮ Taillardat, Maxime, Anne-Laure Fougères, Philippe Naveau, and Olivier Mestre. "Forest-based and semi-parametric methods for the postprocessing of rainfall ensemble forecasting" Accepted in Weather and Forecasting (2019). HAL preprint . ArXiv preprint 1711.10937 Maxime Taillardat 11/16
ROC curves ◮ Value: Focus on the upper left corner ◮ Taillardat, Maxime, Anne-Laure Fougères, Philippe Naveau, and Olivier Mestre. "Forest-based and semi-parametric methods for the postprocessing of rainfall ensemble forecasting" Accepted in Weather and Forecasting (2019). HAL preprint . ArXiv preprint 1711.10937 Maxime Taillardat 11/16
ROC curves ◮ Value: Focus on the upper left corner ◮ Taillardat, Maxime, Anne-Laure Fougères, Philippe Naveau, and Olivier Mestre. "Forest-based and semi-parametric methods for the postprocessing of rainfall ensemble forecasting" Accepted in Weather and Forecasting (2019). HAL preprint . ArXiv preprint 1711.10937 Maxime Taillardat 11/16
ROC curves ◮ Value: Focus on the upper left corner ◮ Taillardat, Maxime, Anne-Laure Fougères, Philippe Naveau, and Olivier Mestre. "Forest-based and semi-parametric methods for the postprocessing of rainfall ensemble forecasting" Accepted in Weather and Forecasting (2019). HAL preprint . ArXiv preprint 1711.10937 Maxime Taillardat 11/16
Post-processing strategy for hourly rainfall ◮ Work on 3 watersheds in Brittany (global size : 720 km2 ) ◮ Météo-France’s LAM-EPS PEAROME (12 members, grid scale : 2.5 km), base time : 21UTC, lead times from 1 to 45 hours ◮ Use of hourly calibrated (over rain gauges) radar precipitation observations ANTILOPE (grid scale : 1 km) ◮ Data spans from 12/2015 to 03/2016 and from 05/2016 to 06/2016 ◮ Predictors : statistics on the hourly rainfall ensemble, and on surface humidity and temperature Ensemble Copula Coupling (ECC) (Schefzik et al., 2013) ◮ Restores physical consistency between grid points and time steps For each time step, each watershed : ECC Result of calibration : 1 distribution for the watershed −−→ 12 members for each grid point Maxime Taillardat 12/16
ECC + post-processing visualization Maxime Taillardat 13/16
Future architecture ◮ Data pooling + Data boosting Raw PEAROME 12 mem 2.5 km Raw PEAROME* 300 mem 10 km (vs. 81 ANTILOPE obs) Post-P of PEAROME* on H.C.A 10km 300 PEAROME* quantiles 10 km ANTILOPE grid 1.25km PEAROME grid 2.5km PEARP grid 10km HOMOGENEITY CALIBRATION AREA 10km Calib PEAROME 12 mem 2.5 km Maxime Taillardat 14/16
The after sales service... Ensemble forecasting requires met. skill (as deterministic) + ◮ understanding of probabilities ◮ notions in game theory ◮ communication of uncertainty Maxime Taillardat 15/16
The after sales service... Ensemble forecasting requires met. skill (as deterministic) + ◮ understanding of probabilities ◮ notions in game theory ◮ communication of uncertainty One must provide statistical//vizualisation tools Roebber diagram RR24 > 150mm 10 5 3 2 1.5 1.3 1.0 1 0.9 0.8 0.8 0.8 0.7 Probability of Detection 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Success Ratio Maxime Taillardat 15/16
Other references Gneiting, T., A. E. Raftery, A. H. Westveld III, and T. Goldman, 2005: Calibrated probabilistic forecasting using ensemble model output statistics and minimum crps estimation. Monthly Weather Review, 133 (5), 1098–1118. Schefzik, R., T. L. Thorarinsdottir, T. Gneiting, and Coauthors, 2013: Uncertainty quantification in complex simulation models using ensemble copula coupling. Statistical Science, 28 (4), 616–640. Whan, K., and M. Schmeits, 2018: Comparing area-probability forecasts of (extreme) local precipitation using parametric and machine learning statistical post-processing methods. Monthly Weather Review, (2018). Maxime Taillardat 16/16
You can also read