资源环境科技发展态势分析平台

In this article, we proposed a simple neural network (NN) technique for estimating erythemal ultraviolet (UV) radiation in Novi Sad (Serbia) using available input parameters. The technique implies the use of one of two models depending on the availability of the input parameter: (a) NN model 1 (NNM1) which uses global solar radiation, clearness index, cloudiness and air mass; and (b) NN model 2 (NNM2) which adds total ozone content (TOC) to the NNM1 inputs. The three feed-forward NNs with different internal structures and back propagation learning method for each NN model were used in modelling. The parallel calculation was used for learning each NN. The results showed that the NNM1 provides satisfactory estimate (R=0.975, MBE=-0.614%, MAPE=12.580%, RMSE= 17.716%) and that additional use of TOC NNM2 considerably improves the results (R=0.982, MBE=-0.726%, MAPE=10.161%, RMSE=14.509%). The performances of developed NNMs become significantly better if warm part of the year is isolated (MAPE=10.981 and 8.958; RMSE=13.889 and 11.709, for NNM1 and NNM2, respectively). Variations of reconstructed annual averages of daily doses in the period 1949-2012 indicate ability of the technique to model the relationship between erythemal UV radiation and the affecting atmospheric factors. The analysis showed that the increasing trend during the warm part of the year in the period 1981-1996 was mainly caused by TOC, while the increase after 1996 was to a greater extent caused by cloudiness.