Solar radiation data can be used to simulate surface dynamic and thermal process. Solar radiation data is the important input parameter of the models in ecology, hydrology, crop, solar radiation transmission, global circulation and so on. Quantitative simulation of solar radiation is important for understanding climate change in Northwest China. However, the solar radiation stations are sparse in Northwest China, so using small amount of radiation site data interpolating or extrapolating is difficult to obtain the spatial distribution of solar radiation data. There are more many weather stations in Northwest China, so it is one of the best methods to simulate the solar radiation by using a large number of meteorological observations. In this article, the LM(Levenberg-Marquardt) algorithm is used to optimize the BP neural network (LM-BP neural network is abbreviation of BP neural network for the optimization). This article simulates solar radiation using LM-BP neural network, H-S and A-P climate models at Urumqi, Kashi, Hami, Xining and Guyuan radiation stations and uses MPE, MBE and RMSE indexes of accuracy assessment to test the three models. The results indicate that LM-BP neural network has the highest accuracy in model simulations, showing satisfactory performance compared with the simulation results of traditional two climate models, simulated and observed values of fitting degree model is superior to H-S and A-P climate models. So we selects the LM-BP neural network model to simulate solar radiation in Northwest China. Basing on the meteorological data from 159 weather stations in Northwest, we apply the BP neural network optimized LM (Levenberg - Marquardt) algorithm to simulate the total month solar radiation during 1990-2012 in these meteorological observation stations. Then the solar radiation value of the 159 weather stations and the measured radiation data of the 25 radiation observation station to obtain the spatial-temporal distribution of annual average solar radiation by interpolation, and analyzes. These results indicate that average annual total radiation in 1990-2012 in Northwestern ranges from 262 MJ/m~2 to 643 MJ/m~2, presenting the distribution pattern of high in the middle, low on both end. LM neural network is a promising method for solar radiation simulation, which can be used in the simulation of solar radiation in the area of no radiation observation.