Using genetic algorithms to explore new predictor variables for statistical precipitation downscaling with analog methods.

Analog methods (AMs) allow for the prediction of local meteorological variables of interest(predictand), such as the daily precipitation, on the basis of synoptic variables (predictors).They rely on the hypothesis that similar situations at the synoptic scale are likely to result insimilar local weather conditions. AMs can rely on outputs of numerical weather predictionmodels in the context of operational forecasting or outputs of climate models in thecontext of climate impact studies. The predictors archive is usually a reanalysisdataset. Different meteorological variables from the NCEP reanalysis 1 were assessed after itsrelease to identify the best predictors for daily precipitation. This former work provided abasis on the top of which more complex methods were developed by adding additionalvariables in a stepwise way. However, the first predictors of the method often remain thesame, and the selection of new predictors is done manually. Nowadays, several newreanalysis datasets are available and were proven more skilful for analog methods than theNCEP reanalysis 1. The accuracy of several variables has significantly improved and morevariables are now available than before. Therefore, the former selection of predictor variablesmight not be optimal anymore. Different variables from various reanalyses shouldbe assessed, which can turn out to be a cumbersome task if done manually andextensively. Genetic algorithms (GAs) were shown to successfully optimize the parameters of theAMs, such as the spatial domain on which the predictors are compared, the selection of thepressure levels and the temporal windows of the predictors, a weighting between predictors,and the number of analog dates to select. GAs can jointly optimize all parameters of AMs andget closer to a global optimum by taking into account the dependencies between parameters.Moreover, GAs can objectively infer parameters that were previously assessed manually, andcan take into account new degrees of freedom. The mutation operator of GAs was identifiedas a key element for this application, and new variants were developed that provedefficient, such as the chromosome of adaptive search radius, which takes no controlparameter. Therefore, we propose using GAs to explore the variables from three reanalyses(MERRA-2, ERA-interim, CFSR) and select the most relevant ones, along with theappropriate analogy criteria. Although the expert’s expertise remains necessary to supervisethe selection of predictors, GAs facilitate the exploration of large datasets. The first testsproved the potential of this approach with the selection of unexpected – but yet relevant –combinations of variables and analogy criteria