Predicting Model Of Parameters Based On Non-linear Principal Component Analysis And Neural Network

In chemical processes, the prediction and control of product quality indexes and production process parameters play a very important role to steady the whole factory running production system and increase the output and quality.H₂/N₂ ratio is an important control index for ammonia synthesis. According to the characteristic of H₂/N₂ ratio such as big-lag, non-linear and variety-by-time, non-linear principal component analysis and the neural network are combined, the prediction model of H₂/N₂ ratio is confirmed base on large numbers of the measured data. In the process of modeling, the first step is to decrese the dimensions of variables that are the influential factors, and the result is that the two principal components has contained the information that is 94.85% of the information of the five variables. The principal components are chosen as input variables of the network model. After performing a large number of computing experiment and comparison like training, recalling and predicting data by the network, the model structure of it is confirmed, and the neural network model for predicting H₂/N₂ ratio is established. The result show that the average absolute value of the absolute error between the predicting data and the surveying data is 0.0352, and the average relative error is 1.6926%, these can satisfy the requirement of predicting H₂/N₂ ratio timely, and has training quickly, predicting rapidly characteristic.The water content of ammonium phosphate is an important quality index. According to the actuality of ammonium phosphate production such as multi -influential factors, predicting and controlling hardly of the water content, firstly, original data is treated by non-linear principal component analysis to lower the number of variables. The result is that the three principal components have contained the information that is 96.29% of the information of the nine variables. The principal components are chosen as input variables of the network model, the neural network model for predicting water content is established. The result show that the average absolute value of the absolute error between the predicting data and the surveying data is 0.2019, and the average relative error is 7.1%.