Deep belief network (DBN) is an effective deep learning model, which has addressed the deep training problem caused by vanishing gradients in traditional neural networks and been widely applied in multiple fields. However, like the other machine learning models, DBN also faces the problem of difficulty in regulating generalization ability, which poses a severe challenge to the application scopes and effects. This paper proposes a method for quantifying generalization error boundary of DBN (DBN-QG), aiming to solve the problem that the upper bound of generalization error is uncontrollable and improve the generalization ability. First, a DBN model is constructed, among which the correlation characteristics of training error and testing error are analyzed on the data sequence, and then providing a mathematical description of the generalization error. Second, by analyzing the Rademacher complexity and sparsity degree as well as the other characteristics, a quantifiable constraint theorem for the upper limit of the generalization error boundary is proposed and further proved. Finally, based on the upper bound constraint of the generalization error, the optimal parameter configurations are given in the training process, and the resulting DBN-QG is used to predict the CO2 state of the coastal water environment in Jiaozhou Bay. The experimental results show that the proposed DBN-QG is superior to other models in terms of prediction accuracy and generalization performance. Furthermore, the boundary quantification method of DBN generalization error not only enhances its interpretability in CO2 prediction, but also provides a relatively generalized theoretical basis for the quantitative analysis of machine learning models.