Abstract: To build a more accurate model for predicting temperature and humidity in solar greenhouse. The microclimatic observing experiment was carried out at Baodi District of Tianjin from 2011 to 2013 in winter (January, February and December). The temperature and humidity prediction models in geenhouse by using stepwise regression and BP neural network were established at 3 periods (0:00-8:00, 8:00-17:00, 17:00-23:00) of 3 kinds of weather types (sunny, cloudy, overcast). The results showed that: (1) the average accuracy rate of the absolute error of simulated and actual values less than 3℃ was 88%, and the root-mean-square error (REMS) was 2℃ under 9 conditions in greenhouse by using stepwise regression model of temperature, and the average accuracy rate of the absolute error of simulated and actual values less than 3℃ was 94%, and the root-mean-square error (RMES) was 1.6℃ under 9 conditions in greenhouse by using BP neural network model of temperature; the temperature prediction model established by BP neural network was more accurate and stable; (2) the average accuracy rate of the absolute error of simulated and actual values less than 6% was 81%, and the root-mean-square error (REMS) was 5.7% under 9 conditions in greenhouse by using stepwise regression model of relative humidity, and the average accuracy rate of the absolute error of simulated and actual values less than 6% was 80%, and the root-mean-square error (RMES) was 6.7% under 9 conditions in greenhouse by using BP neural network model of relative humidity. Both 2 models are not suitable for predicting the relative humidity of solar greenhouse at 8:00-17:00, while the humidity prediction model established by stepwise regression at 17:00-23:00 and 0:00-8:00 is more accurate and stable.hour,but the humidity prediction model established by stepwise regression at 17-23 hour and 0-8 hour is more accurate and stable.