Long-term records of solar UV radiation reaching the Earth's surface are scarce. Radiative transfer calculations and statistical models are two options used to reconstruct decadal changes in solar UV radiation from long-term records of measured atmospheric parameters that contain information on the effect of clouds, atmospheric aerosols and ground albedo on UV radiation. Based on earlier studies, where the long-term variation of daily solar UV irradiation was derived from measured global and diffuse irradiation as well as atmospheric ozone by a non-linear regression method [Feister et al. (2002) Photochem Photobiol 76:281-293], we present another approach for the reconstruction of time series of solar UV radiation. An artificial neural network (ANN) was trained with measurements of solar UV irradiation taken at the Meteorological Observatory in Potsdam, Germany, as well as measured parameters with long-term records such as global and diffuse radiation, sunshine duration, horizontal visibility and column ozone. This study is focussed on the reconstruction of daily broad-band UV-B (280-315 nm), UV-A (315-400 nm) and erythemal UV irradiation (ER). Due to the rapid changes in cloudiness at mid-latitude sites, solar UV irradiance exhibits appreciable short-term variability. One of the main advantages of the statistical method is that it uses doses of highly variable input parameters calculated from individual spot measurements taken at short time intervals, which thus do represent the short-term variability of solar irradiance.