关键词: BP神经网络模型, 统计降尺度, 黄河上游, 温度, 降水

Abstract: At present，sources of information on the climate of the last millennium include paleoclimatic proxies and models. Proxies of the last climate are natural archives that have，in some way，incorporated a strong climatic signal into their structure. But they exit several problems，such as spatial representativeness，uncertainty. What’s more，they are far from achieving season-specific histories for different climate variables at the regional scale.  Atmosphere-ocean coupled global circulation models could provide relatively high temporal resolution data. However，they also could not provide the regional scale information for their low spatial resolution. The regional climate response to global change has been an urgent issue to resolve. To understand regional climate information，this article tries to build the relationship between the global climate model outputs and the regional climate by using back propagation （BP） artificial neural network，a widely recognized statistical downscaling method. The global climate data is the output of ERIK simulation （covering the period 1000-1990） with the coupled atmosphere-ocean global climate model ECHO-G，which is forced by three external forcing factors: solar variability，greenhouse gas concentrations in the atmosphere including CO2 and CH4，and the effective radiative effects from stratospheric volcanic aerosols. The meteorological observation data are used for regional climate information during the past 50 years. The desert valley of the upper reaches of the Yellow River，located in arid and semi-arid areas，is much more sensitive to global change. Therefore，the temperature and precipitation series in this area were reconstructed during the last millennium. The results were well proved by tree-ring data and historical documents，with correlation coefficient being 0.58 and 0.57. Compared with model data，fitting series reproduced the regional characteristics of the interannual and interdecadal climatic variations. Whereas the extreme values were not well reconstructed，especially the extreme values of the winter temperature and the summer precipitation. The reconstructed series showed that there were three typical periods over this region during the last millennium，i.e. the Medieval Warm Period （MWP），the Little Ice Age （LIA） and the Present Warm Period （PWP）. The difference of winter temperature between MWP and LIA was 2 . The amplitude of precipitation variation was smaller than that of the temperature，especially in winter. The spatial pattern showed that PWP was the warmest period during the last millennium，but the precipitation in PWP was less than that in MWP. Back propagation （BP） artificial neural network is a helpful tool to remedy the shortage of proxies and models. It provides useful information about regional response to climate change，such as global warming. It also has several problems to be resolved，such as extreme values bias，single-point operation and lacking dynamic mechanism. Consequently，new downscaling techniques by combining statistical and dynamical downscaling techniques would be developed as the main next plan.

Key words: BP neural network, statistical downscaling, the upper reaches of Yellow River basin, temperature, precipitation