Abstract: Using the minutely global solar radiation data from September 1, 2010 to August 31, 2011 at the Turpan solar energy station of center for wind and solar energy resources of China Meteorological Administration (CMA), two prediction methods of minutely global solar radiation were compared, namely, a statistical extrapolation (TRD) method and a dynamic neural network (NNN) method. The results show that for the annual average, the prediction effect is better by the TRD than by the NNN. The prediction accuracy rates of the two methods is related with weather situation, and it is higher in a sunny day than in a rainy day, while it is lower in a cloud day and a sand-storm day. The impact factors of radiation observation such as cloud, aerosol and dust and so on could be accurately expressed by the TRD method at the beginning time of prediction, which could be continued to the first or second step of rolling forecasts. The prediction effect is better by the NNN than by the TRD for the non-linear variation of global solar radiation under complex weather such as rainy day and cloud day. The predication accuracy of global solar radiation is higher by the TRD than by the NNN for three hours after sunrise and before sunset. The prediction effect is slightly better by the NNN than by the TRD for a sudden variation of global solar radiation.

Key words: Global solar radiation, Minutely prediction, Statistical extrapolation, Dynamic neural network