Mathematical Model Solution And Development Of Composite Neural Network For 1780mm Hot Rolling Mill

With the development of green,information and intelligence in iron and steel industry,the traditional model self-learning method is not suitable for the customized production of hot rolled strip due to more unqualified products production caused by the lag of adjustment.Therefore,it is urgent to improve the accurate setting ability and dynamic adaptive ability of model parameters.Neural network has been widely used in the field of hot rolling due to its nonlinear mapping and good generalization ability.However,it is rare to combine neural network with hot-rolling mathematical model for on-line setting and adaptive calculation.With the development of intelligent rolling technology,it is necessary to combine neural network with the existing rolling model to establish the smart hot-rolling mathematical model with high adaptive ability.This paper is based on the research project called 'Improving the size accuracy of the first steel strip after specification change in a 1780mm hot rolling mill'.The purpose of this article is trying to create a complex neural network with the help of the relevant documents about on-line network,to crack the black box program of on-line neural network and to guide the debugging of network parameters to improve the size and temperature controlling ability for the first steel after specification change.The main research contents of this paper are as follows:(1)On the basis of the research of model principle and structure of traditional hot-rolling models,the friction model,forward slip model,deformation resistance model,rolling force model,rolling torque and power models were extracted from the on-line program and two methods about combinations of neural network and mathematical model were also analyzed.(2)The structure of the composite neural network which combines the adaptive linear neural network with the radial basis neural network was designed.Method of data processing and algorithm of network offline training were also designed.Then,the software named 'Complex neural network prediction software for prediction of hot-rolling model parameters' was successfully developed using Matlab software.The developed software has the predictive function for 10 key parameters of hot-rolling mathematical model.(3)Based on the 70,000 sets of training data,the hot rolling deformation resistance correction network with 11 inputs and 1 output was off-line trained,and then 8,750 sets of test data were used for prediction.The proportion of relative prediction errors in 5%,10%,and 15%is 43.39%,72%and 88.03%respectively.Compared with the prediction results of a single BP network and an independent RBF network,the mean square error(MSE)of the complex deformation resistance correction network is smaller,the coefficient of determination(R)is higher,and the average absolute percentage error(MAPE)and the standard deviation of the relative error(RESTD)is also smaller.The results show that the presented network performance is better,and its prediction accuracy is higher.(4)Based on the 30,000 sets of training data,the prediction network for width spread in finishing mill with 27 inputs and 1 output was also off-line trained,and 3,750 sets of test data were used for prediction.The proportion of relative prediction error in 5%,10%and 15%is 44.82%,73.36%and 85.7%respectively.Compared with the prediction results of a single BP network and an independent RBF network,the presented network has higher prediction accuracy and the network performance is better.