Daily precipitation performances of regression-based statistical downscaling models in a basin with mountain and semi-arid climates

The impacts of climate change on current and future water resources are important to study local scale. This study aims to investigate the prediction performances of daily precipitation using five regression-based statistical downscaling models (RBSDMs), for the first time, and the ERA-5 reanalysis dataset in the Susurluk Basin with mountain and semi-arid climates for 1979–2018. In addition, comparisons were also performed with an artificial neural network (ANN). Before achieving the aim, the effects of atmospheric variables, grid resolution, and long-distance grid on precipitation prediction were holistically investigated for the first time. Kling-Gupta efficiency was modified and used for holistic evaluation of statistical moments parameters at precipitation prediction comparison. The standard triangular diagram, quite new in the literature, was also modified and used for graphical evaluation. The results of the study revealed that near grids were more effective on precipitation than single or far grids, and 1.50° × 1.50° resolution showed similar performance to 0.25° × 0.25° resolution. When the polynomial multivariate adaptive regression splines model, which performed slightly higher than ANN, tended to capture skewness and standard deviation values of precipitations and to hit wet/dry occurrence than the other models, all models were quite well able to predict the mean value of precipitations. Therefore, RBSDMs can be used in different basins instead of black-box models. RBSDMs can also be established for mean precipitation values without dry/wet classification in the basin. A certain success was observed in the models; however, it was justified that bias correction was required to capture extreme values in the basin.