It is difficult to determine hidden layer nodes and select parameters randomly when neural network is applied to complex systems. To solve this problem, a new rule modeling method is explored. Firstly, the incremental dynamic rule extraction is realized by the rough set & decision tree coupling algorithm. Secondly, the obtained dynamic rules are used to calculate the minimum rule number, which can be used as the basis to determine the hidden layer nodes and realize network rule modeling. Thirdly, the initial network parameters are optimized to avoid local minimum and improve the training speed and precision of the model. Finally, the optimization model is applied to the air quality forecast. Performance test and comparative analysis show that the model has a fast convergence speed and controls the error within 4%, and its learning speed and prediction accuracy are obviously better than the traditional model. The research results realize the rule model construction and optimization under dynamic incremental mode and provide a new research idea for dynamic data processing.