Abstract: The neural network training is hard to convergence and the accuracy is not exact, because of the existence of a complex relationship between the input coupling factor and the condition attribute redundancy in establishing model for predicting temperature of sunlight greenhouse. To solve the above problems, this article chose the principal component analysis to treat the samples by dimensionality reduction and decoupling. Using the treated data as input, the humidity of sunlight greenhouse as output, the model of NARX neural network was established by the Bayesian regularization algorithm to predict the humidity of sunlight greenhouse. The simulation result showed that the model had strong nonlinear dynamic description ability, and was able to predict indoor humidity accurately.