Seasonality properties of four statistical-downscaling methods in central Sweden

Daily precipitation in northern Europe has different statistical properties depending on season. In this study, four statistical downscaling methods were evaluated in terms of their ability to capture statistical properties of daily precipitation in different seasons. Two of the methods were analogue downscaling methods; one using principal component analysis (PCA) and one using gradients in the pressure field (Teweles-Wobus scores, TWS) to select the analogues in the predictor field. The other two methods were conditional-probability methods; one using classification of weather patterns (MOFRBC) and the other using a regression method conditioning a stochastic weather generator (SDSM). The two analogue methods were used as benchmark methods. The study was performed on seven precipitation stations in south-central Sweden and the large-scale predictor was gridded mean-sea-level pressure over Northern Europe. The four methods were trained and calibrated on 25 years of data (1961-1978, 1994-2000) and validated on 15 years (1979-1993). Temporal and spatial limitations were imposed on the methods to find the optimum predictor settings for the downscaling. The quality measures used for evaluating the downscaling methods were the residuals of a number of key statistical properties, and the ranked probability scores (RPS) for precipitation and maximum length of dry and wet spells. The results showed that (1) the MOFRBC and SDSM outperformed the other methods for the RPS, (2) the statistical properties for the analogue methods were better during winter and autumn; for SDSM and TWS during spring; and for MOFRBC during summer, (3) larger predictor areas were needed for summer and autumn precipitation than winter and spring, and (4) no method could well capture the difference between dry and wet summers.