In Northwest China, a little change in temperature and precipitation would lead into a dramatically adverse disaster in the vulnerable ecological environment. However, due to the scarcity and uneven distribution of the meteorological stations in Northwest China, it is difficult to accurately assess the regional climate change in time and space. The Regional Climate Model (RCM) in state of art provides a powerful tool to simulate with highly horizontal resolution, which may supplement the existing observation. However, the poor performance of the model has slowed down the progress of regional climate change research in Northwest China. With the gridded dataset (CN05) and observation in situ, in this paper, simulation ability is assessed of the High Resolution WRF model′s output in the spatial pattern, extremes, and anomaly and so on. On the assumption that the temperature and precipitation′s biases keep in constant, the data in situ are used to train the Classification Regression Tree Model (CART) for the Bias Predication Model. The results show that the spatial patterns of annual mean temperature and annual precipitation from CN05 dataset consist with WRF model′s output, and WRF′s outputs illustrate more features in spatial pattern. However, the WRF overestimates precipitation in alpine and lakes. The temperature and precipitation anomalies in the Xinjiang perform more homogenously as compared with other regions. WRF′s 95% of the simulated precipitation in Gansu and Xinjiang is better than that in Ningxia, Qinghai and Shaanxi. For 5% and 95% points of the temperature, the WRF simulation has warm bias in most sites, while the CN05 simulation has cold bias. Temperature and precipitation′s bias have a certain relation among the five different geomorphic units. It is impossible to establish a uniform Bias Predication Model to apply in whole Northwest China, because of the height and terrain complexity, which has a large weight in Bias Predication Model.