C3-WP1 - Post-Processing

The aim of this work package is to identify the optimal post-processing of the operational regional decadal prediction system for Central Europe which removes systematic errors in the regional model output applying different approaches of bias and drift correction. The statistical post-processing represents an efficient and low-cost tool to calibrate climate model output before being communicated to users.

The work package is structured by several tasks: First, different approaches of statistical post-processing are applied based on observational data for Central Europe and the multi-decadal assimilation runs nudged to observations for bias correction or rather the decadal predictions for drift correction. The developed statistical transfer functions are compared for different variables and applied for the calibration of the decadal predictions which are then evaluated by cross validation with independent observations (C3-1.1, Fig. 1).

Fig. 1: Research concept and possible approaches of the statistical post-processing

The optimal combination of the individual approaches for the statistical post-processing of the decadal predictions is analysed in detail and separately for different seasons, regions in Central Europe and variables or climate phenomena (C3-1.2). The developed metric of combined statistical post-processing is integrated in the central evaluation system CES of MiKlip II in order to be applied for the calibration of different regional decadal predictions and for the comparison with that of global decadal predictions (C3-1.3). Finally, this metric is as well applied as efficient calibration approach with minor computational requirements for the operational regional decadal prediction system with CCLM (C3-1.4).

The deliverables of the work package are given by the milestones of the different tasks: the comparison of the individual approaches of the statistical post-processing (M1, month 24), the optimal combination of the individual approaches of the statistical post-processing for different seasons, regions and variables (M2, month 30) and the integration of the statistical post-processing in the central evaluation system (M3, month 33) and in the operational prediction system (M4, month 36).
 

Progress so far

We have developed a two-step bias-adjustment approach for removing systematic errors of decadal climate predictions. This method has been tested on the third version of CCLM decadal hindcasts (preop). In step one, we estimate model errors based on a long-term CCLM assimilation run and observational data, as well as a transfer function (TF) which can adjust model simulation towards reality. In step two, the TF is applied to the complete set of decadal hindcast simulations (270 individual runs). The bias-adjusted decadal predictions for monthly precipitation are evaluated both on leading-year-dependent multi-year timescales, which is commonly used for decadal hindcasts, and on the monthly timescale for the assessment of inter- and intra-annual variability within decadal forecasts. We obtain a maximum improvement of 30% explained variance (R2) and a maximum reduction of 40mm/month root mean square error (RMSE). An estimated mean square skill score (MSSS) of more than 0.5 is observed regionally. A paper of this study with title ‘Bias adjustment for decadal predictions of precipitation in Europe from CCLM’ is under review by co-authors. Another paper exploring the effects of various statistical method for this application is currently in preparation.