In this paper, we present an evolutionary approach for extracting a model of flood prediction from hydrological data observed timely on water heights in a river watershed. Since this kind of data recorded by sensors on river basins is highly scarce and hopefully much unbalanced between cases of floods and non-floods, we have adopted the notion of aggregate variables which values are computed as aggregates on raw data. An evolutionary algorithm is involved to allow selecting the best sets - juries of classifiers- of such variables as predictive variables. Two real hydrological data sets are trained and they both show the efficiency of the method compared to traditional solutions for prediction.