In order to decrease the data transmission frequency of the sensor nodes in greenhouse WSN system, a method based on two-level prediction was presented. Firstly, Letts’ criterion was imported to detect the sequence of outliers and the calculation method of sliding recursive sequence variance was proposed to facilitate real-time calculation of the nodes. Secondly, a piecewise linear regression equation combined with weighted adaptive algorithm was established to form two-level prediction models in sensor node and server. When forecasting error exceeded the set threshold, the sensor nodes uploaded the actual collection value. At other times, the server automatically triggered the linear regression prediction model to fill the partial data. At the same time, a variable error threshold determination method based on parabola was presented according to the characteristics of the automatic control of facility environment. The tests proved that the first order linear regression model approximated the raw data curve in prescriptive error threshold and the number of sending data of WSN sensor nodes could be reduced 93% by using two-level prediction algorithm （error threshold is 0.9）.