Calibration and Combination of CHFP precipitation forecasts over South America using Ensemble Regression

MARISOL OSMAN; CAROLINA VERA

Orono, Maine

Workshop; 41st CPC Climate Diagnostic and Prediction Workshop; 2016

NOAA Climate Prediction Center

p { margin-bottom: 0.1in; direction: ltr; line-height: 120%; text-align: left; }Inthis study we calibrate and combine the climate predictions availableat the Climate Historical Forecast Project of the World ClimateResearch Program to develop better seasonal precipitation forecasttools over South America. Forecasts made with initial conditions ofNovember and May and valid at December-January-February (DJF) andJune-July-August (JJA), respectively (Lead 1) over the period1982-2006 are considered. A Multi-Model Ensemble (MME) using 11models, each one with around 10 ensemble members, was constructed andits performance was evaluated against CMAP database. The domain ofapplication spans [15°N-60°S;275°E-330°E] and it is also dividedin two subdomains: tropical South America (grid points to the northof 20°S) and extratropical South America (grid points to the southof 20°S).TheEnsemble Regression technique (EREG) applies a regression equationdeveloped for the ensemble mean to each ensemble member to obtain aprobability density function (PDF) which represents the ensembleprediction. EREG is first applied to each model and its ensemblemembers to calibrate them and to equalize models for differences inensemble members. A weight is determined for each model, whichreflects the likelihood of that model of being the best performingone and is considered to build the weighted MME.Then, two approachesare used to obtain the final consolidated PDF. The first one consistsin using the weighted MME in a new ensemble regression, resulting ina weighted super-ensemble regression (WSEREG) to get the finalconsolidated PDF. The other technique consists in obtaining theconsolidated PDF computing the normalized summing of the weightedmodels? PDF (weighted kernels, WKERNELS).The cross-validationapproach is used to estimate the regression parameters. Theconsolidated PDFs obtained are used to forecast the three equallyprobable categories below normal, near normal and above normal. Theseforecast are confronted against those obtained counting theproportion of ensemble members of the MME falling in each category,known as counting estimate technique (CE).Resultsshow that both WKERNELS and WSEREG outperform CE in terms of theRanked Probabilistic Skill Score (RPSS) and Brier Skill Score (BSS)in both seasons but especially in austral summer (DJF). However, onlyin northern South America the performance of both consolidationtechniques is slightly better than the climatological values of thepredictand (three categories equally probable); as BSS and RPSSvalues are lower than 0.3. In extratropical South America both RPSSand BSS values change from less than -0.5 for CE to near 0 for bothWKERNELS and WSEREG.On the other hand, reliability diagramscomputedover the entire domain shows that WKERNELS and WSEREGsubstantially improve the forecast in terms of reliability respect tothan obtained with CE. This improvement is somewhat higher for thebelow normal category than for the above normal category.Nevertheless, in reliability diagrams computed only overextratropical South America,WKERNELS is considerable better than CEwhile the performance of WSEREG is somewhat modest. Futureworks include testing different weighting techniques as well aschanging some parameter of the ensemble regression. Also theapplication of the ensemble technique to the models participating inthe North America Multi Model Ensemble project is consider for thecoming months.