Aiming at the problems of various influencing factors, complex modeling and insufficient prediction accuracy, a combined forecasting method of energy demand based on variational modal decomposition and deep belief network is proposed. Firstly, the periodicity and randomness of the cooling and heating load series are analyzed, and the variational modal decomposition method is proposed to decompose the cooling and heating load. Secondly, based on the fact that modal decomposition are prone to be redundant after decomposition, the sample entropy is used to reconstruct the decomposed mode components to reduce the degree of redundancy. Finally, because the initial weights of the deep belief network are too random, crisscross optimization algorithm is adopted. The cooling and heating load is predicted by optimizing the deep belief network, and the prediction results are analyzed according to the simulation example. The example shows that the proposed prediction method can effectively improve the prediction accuracy and is practical.