Abstract: Abstract: In order to obtain a feasible hourly temperature prediction method at night for solar greenhouses and improve the adaptability of the active heat collection and release system, this research attempted to obtain 72 h meteorological information of the test area from the weather forecast published by China Weather Net (It mainly included weather conditions, outdoor air temperature, wind speed, outdoor relative humidity, cloud fraction, visibility, and rainfall probability). The weather conditions that were crawled and screened were simplified into three categories: sunny, cloudy, and overcast. As a black box model, the statistical model did not need to accurately reflect the real mechanism. Usually, the measured data was used to obtain the functional relationship between variables by mathematical statistics, which greatly reduced the difficulty of modeling. Based on multiple linear regression, vector machine regression and random forest regression methods that require less modeling data, the hourly temperature prediction models for indoor nighttime were constructed according to the three conditions of sunny, cloudy and overcast. Mean square error (MSE) and coefficient of determination (R2) were used to evaluate the prediction accuracy of various models. The results showed that the three regression models had different accuracies under different weather conditions. On sunny and cloudy days, the random forest regression model had the highest accuracy, with MSE and R2 being 0.37, 0.94 and 0.05, 0.98, respectively. On cloudy days, the MSE and R2 of the support vector machine regression model were 0.40 and 0.87, respectively, which were better than the remaining two models. Therefore, the random forest regression model can be selected to predict the hourly temperature of greenhouses at night on sunny and cloudy days, and the vector machine regression model can be used on overcast days. The optimal model was selected to predict the hourly indoor air temperature at night, and compare the predicted value with the measured value. The verification results showed that the curves of predicted temperature and measured temperature were well-fitted, and one-way ANOVA showed no significant difference between them (P>0.05). On sunny, cloudy, and overcast days, the MSE was 0.19, 0.19, and 0.15, and the R2 was 0.98, 0.95, and 0.90, respectively. The mean absolute errors between the measured and predicted values of hourly air temperature at night were 0.1, 0.3, and 0.3 ℃, respectively. However, the large error between the predicted value and measured value mainly occurred at 18:00, because the external blanket was closed in advance, which reduced the heat loss of the solar greenhouse. The actual indoor temperature was still higher. Therefore, the model constructed in this paper had high prediction accuracy and would be used in production practice. The collected heat can be reasonably distributed for multiple consecutive nights according to the forecast and prediction results, so as to improve the adaptability of the active heat collection and release system.