Variable air volume systems are nonlinear, time-varying and multivariable with large time delay. Model predictive control can achieve satisfactory stability and energy-saving, the performance of which depends on precision and generalization capability of the predictive model. To overcome difficulties in modeling by mechanism, this paper proposes a modeling method of multi-zone VAV systems based on neural networks. The factors influencing on the sensible cooling load and coupling between zones are analyzed and consequently the structure of the neural network model is determined. In order to fully demonstrate the dynamic characteristics of the VAV system, neural network training samples cover all the VAV dynamic range. To increase generalization capability, Bayesian regularization algorithm is used to train the network. Experimental results show that the neural network predictive model has satisfactory accuracy and good generalization performance.