It was vitally important to effectively obtain the spatial distribution of temperature and humidity of the greenhouse mushroom house in advance for optimizing environmental stress of edible fungi, early warning of disease and pre-regulation of the mushroom house. The traditional single-point prediction could not well meet the demand of evaluation of overall environmental performance for the mushroom house. According to the characteristics of time series, non-linear and different spatial distribution of temperature and humidity in mushroom house, a multi-point prediction method of temperature and humidity for the mushroom house based on convolutional neural network (CNN) and gated recurrent unit neural network (GRU) was proposed, which took the historical outdoor meteorological data of the greenhouse, the indoor microclimate environmental data, environmental distribution characteristics, the ventilation information and the humidification information as input by constructing a two-dimensional matrix according to the time series. Firstly, CNN was used to mine the effective information contained in the data to extract the high-dimensional features reflecting the interrelation of greenhouse environmental data, and then the extracted feature vectors were constructed as time series and input to the GRU network for multi-point prediction of temperature and humidity. The prediction model proposed was applied to the multi-point prediction of temperature and humidity in the mushroom house of a solar greenhouse in Beijing Academy of Agricultural and Forestry Sciences, and the experimental results showed that the averaged RMSE and MAE were 0.211℃ and 0.140℃, respectively, for the temperature prediction at each point in the mushroom house, and the average proportion of error control within ±0.5℃ was 97.57%. For the humidity prediction at each point in the mushroom house, the averaged RMSE and MAE were 2.731% and 1.713%, respectively, and the average proportion of error control within ±5% was 92.62%. Comparing with traditional BP neural network, long short-term memory neural network (LSTM), and gated recurrent unit neural network (GRU), the prediction model proposed had higher prediction accuracy.