Prediction Method of Batch Pulp Cooking End Point
Xiang-yang LI,Xue-feng ZHU,Dao-ping Huang,Huan-bin
LIU LIU
DOI: https://doi.org/10.16866/j.com.app.chem2001.02.005
2001-01-01
Abstract:Based on the analyses of conventional cooking models including Hatton model, MODOCell model, Kerr model and Chari model, this paper presents a new modeling method, the neural network modeling method. Compared to the BP neural network, the RBF neural network (RBFNN) has the best approximation ability, higher training speed and no local minima. So, the RBF neural network is an effective modeling tool and this paper uses it. When choosing the input variables and the output variable of the RBFNN model, this paper exploits the measurable variables in the field and the knowledge of cooking process. The input variables of RBFNN model include Kappa number, effective alkali, sulfidity and chip quality. The later two input variables are not included in conventional cooking models.Among the conventional models, Hatton model, which reveals there is a linear relationship between the Kappa number and the logarithm of H-factor, is relatively easily applicable. According to this revelation, the output variable of RBFNN model is the logarithm of H-factor of cooking end point instead of the H-factor itself. The advantage of the choice of the output variable of RBFNN is that it can reduce the degree of nonlinearity which the RBFNN model need express and so that it reduces the node number of the hidden layer of RBFNN and enhances the training speed. Both the RBFNN model and Hatton model are applied to the cooking data sampled at a cooking plant of a large paper-making factory in Fujian, a southeast province of China. The software development platform is MATLAB. During the processes of model training and model predicting, shift data window technology and one-step prediction technology are applied. The application results reveal that RBFNN model is superior to Hatton model in prediction precision. The reason why the RBFNN model has higher prediction precision than conventional cooking models is that the RBFNN model can use more input information and has the ability to approximate any nonlinear functions. The neural network modeling method, which exploits the measurable variables in the field of factories and the knowledge of industrial process, has generalization value in other industrial modeling processes.