Accurate prediction value of room cooling load is the basis data for energy conservation in air conditioning operating process. Firstly, according to the room energy balance equation, with analysis of relationship among cooling load, heat extraction and heat storage, the room cooling load prediction model is put forward. The frequency domain decomposition method is employed to realize the calculation of heat storage, while the deep recurrent neural network is employed to realize the prediction of room cooling load under room air constant condition. Finally, combining the heat storage under room air temperature fluctuation condition and room cooling load under room air constant condition, the room cooling load prediction value under temperature control mode can be obtained. In order to improve the learning efficiency of deep recurrent neural networks, an adaptive switching method between Gauss-Newton method and Levenberg-Marquardt (LM) method is introduced. The building energy consumption simulation toolkits and real-test based experiments show that the proposed method can realize the prediction of hourly room cooling load quickly and efficiently. The proposed model and method realizes accurate prediction for room cooling load under temperature control mode, and can be utilized to realize the quantitative analysis for building passive thermal energy storage, as well as provide references for the direct load control in power grid demand side management.