Hybrid Optimization Prediction Of RBF Neural Network Based PCA

With the increasing demand for production quality in modern industrial processes,advanced sensor equipment and measurement technology have been introduced to obtain more and more detailed process data information,however,it is a pity that some key variables for improving production quality cannot be directly measured.Therefore,how to use existing process variables to predict these key variables has become an urgent problem.In this context,artificial neural network prediction methods are widely used,particularly in chemical processes.Radial Basis Function(RBF)neural network is the most commonly used prediction model among artificial neural networks.It is a three-layer feedforward neural network with a single hidden layer.Different from other feedforward neural networks,it seeks to find global optimal and optimal approximation characteristics,and can establish a relationship model between variables to predict other key variables.As the amount of data generated in the chemical process is getting larger and larger,nonlinearity is getting stronger and stronger,and the related parameters involved in the RBF neural network are not well optimized,resulting in that the predictive performance of key variables cannot meet this situation.In order to solve the above issues,this paper proposes to optimize traditional RBF neural networks and preprocess the process data to obtain more modeling accuracy.The main research contents are as follows:(1)Aiming at the parameters are not well optimized of RBF neural networks and leads to poor prediction performance,LM-GA optimized RBF neural network based on PCA is proposed.First,the principal component analysis(PCA)method is used to preprocess the process input variables of the neural network.Secondly,using Levenberg-Marquardt(LM)algorithm and genetic algorithm(GA)to train the RBF neural network twice to obtain the optimal RBF neural network prediction model.Finally,the strategy is applied to the pressure prediction on both sides of an industrial coke furnace system,and comparison is done with traditional RBF neural network and GA optimized RBF neural network using PCA,where root mean square error(RMSE)and mean absolute error(MAE)indices are calculated.It is proved that the prediction performance of the proposed method is better than the other two methods.(2)Aiming at the large amount of data and strong nonlinearity in chemical processes,and PCA cannot deal with the nonlinear relationship data well,LM-GA optimized RBF neural network based on serialized PCA(SPCA)is proposed.First,PCA and kernel principal component analysis(KPCA)are adopted to preprocess the input process variables to the neural network.Secondly,combined with preprocessed variables,the RBF neural network model is optimized using LM algorithm and GA algorithm.Finally,the strategy is applied to the Tennessee-Eastman(TE)process and compared with traditional RBF neural networks and LM-GA optimized RBF neural network using PCA.It is proved that the prediction performance of the proposed method is better than the other two methods.