RBF neural network modeling based on PCA clustering analysis

Abstract

Aiming at the problem of lower forecast accuracy of traditional RBF neural network model, we suggest a new modeling method. First, pretreatment data are sampled using SPSS software. Principal component analysis (PCA) was applied to original data correlation analysis to remove the correlations between attributes and find the main influencing indicator to reduce the number of input layer nodes. The following step is to apply the clustering analysis on data samples to choose the training data samples for neural network modeling and reduce the number of data samples. Then, after preprocessing, the data samples can be used to construct RBF neural network prediction model. The core problem solving at this stage is to determine the number of hidden layer nodes and the corresponding data center, so that design the network which its target error meet the requirement. We use cement clinker strength prediction as the research example to determine the network structure is 5-89-2. Finally, we implemented the neural network training and simulation work using the neural network toolbox provided in MATLAB. This modeling method can filter the raw data sample to fully reflect the characteristics of the problem in its field, effectively reduce RBF network structure and meanwhile improve the forecast precision of the model. The simulation results show that our model had higher prediction accuracy. The method of construction of prediction model provides a new idea and method for the study of prediction problems, and expands the thought for the research of other field modeling. 1