Abstract: For the sake of carrying out agrometeorological forecast service for the refined facilities, realizing off-season vegetable cultivation with high and stable yield and thus achieving the aim of disaster prevention and reduction and sustainable development, in this paper the observation data of microclimate and meteorological observation materials for the sunlight greenhouse in the national modern agricultural demonstrative zones, Datong County, Qinghai Province were applied to analyze the temperature variation laws in this region and build the corresponding forecast model. The results showed that the all daily variations of indoor annual average temperature appeared as a “1 rise, 2 falls” process. From 08:00 to 14:00, it was a temperature-rising phase, the maximal temperature-rising velocity was in an order as cloudy day (5.46℃) > sunny day (5.34℃) > rainy day (1.6℃); from 14:00 to 21:00, it was a rapid temperature-falling phase, the temperature-falling velocity was in an order as cloudy day (2.79℃) > sunny day (2.69℃) > rainy day (1.56℃); and from 21:00 to 08:00 (the next day), it was an slow temperature-falling phase, the temperature-falling velocity was in an order as sunny day (0.57℃) > cloudy day (0.53℃) > rainy day (0.49℃). A step-wise regression method was applied and the average temperature forecast model and the lowest temperature forecast model were built for the sunlight greenhouse in the plateau area. The average temperature forecast model was T_ave=3.991 0.688T₁ 0.113T₄, the correlation coefficient (R²) of predicted value and actually measured value reached 0.7062, the ABSe and RMSe were 1.45 and 1.85℃, the accuracy rate ≤2% was 75%, and the accuracy rate ≤4% was 96%. The lowest temperature forecast model was T_min=-3.05 0.366T₃ 0.692T₁-0.098T₂, the correlation coefficient (R²) of predicted value and actually measured value reached 0.8373, the ABSe and RMSe were 1.15 and 1.55℃, the accuracy rate ≤2% was 81%, and the accuracy rate ≤4% was 99%. The models simulated soundly and showed better practicability.