Research On Model Predictive Control Of Cement Rotary Kiln Based On Sliding Window Convolutional Neural Network

The control system of cement rotary kiln is a key link in cement technology,with the progress of the cement industry,the control technology of cement rotary kilns has been significantly improved.In the process of rotary kiln calcination,the stability of the kiln condition and the control of the firing zone temperature are closely related to the quality of cement clinker production,energy consumption and production cost.The study of the control method of key process parameters of cement rotary kiln is a guide for operators to adjust and set the production parameters of rotary kiln,which helps the automation technology of cement production and is one of the prerequisites to improve the production quality and energy saving of cement.The main research contents of the this paper are as follows.(1)The new dry cement production technology and the mechanism of cement calcination process in rotary kiln are discussed;the common problems of cement rotary kiln control system are analyzed and solutions are proposed;the control target quantities(kiln main motor current,secondary air temperature,smoke chamber nitrogen oxide concentration,smoke chamber oxygen concentration)and operating variables(kiln head coal feeding quantity,high temperature fan speed)are selected;the research object,multi-input and multi-output time delay system for cement rotary kiln calcination process is identified.(2)A sliding window convolutional neural network-based model is established to predict the key parameters(i.e.,control target quantities)in the cement rotary kiln.First,the correlation between the input and output of the control system on the time scale is studied,and the time series input layer is constructed and input to the neural network in the form of sliding window to reduce the influence of time-varying time delay on the control;then,the data features of the time series input layer are extracted using a multilayer convolution-pooling layer to eliminate data redundancy,And feature integration is performed using the fully connected layer;finally,the network is optimized by continuously correcting the prediction error through the adaptive moment estimation algorithm to achieve multi-step prediction of key parameters in the cement rotary kiln,laying the foundation for the control of key parameters.(3)The Convolutional Neural Network-Model Predictive Control(CNN-MPC)method is proposed.Based on the SW-CNN predictive model,the real data of the control target quantity(i.e.,the control system output)are fed back to correct the predicted values.The optimal control sequence is obtained by solving the objective function using the differential evolutionary algorithm,and the control quantity of the next moment is allowed to act on the cement rotary kiln calcination system to complete the control of the rotary kiln calcination process.(4)Experiments Experiments were conducted using production data of a cement company in Hebei Province to verify the SW-CNN prediction effect and the control effect of CNN-MPC.The experiments were designed to observe the control effects of CNN-MPC on several key parameters of the rotary kiln calcination system(kiln main motor current,secondary air temperature,flue chamber NOx concentration,flue chamber oxygen concentration),and compared with the traditional proportional-integral-derivative control(PID)control for the MIMO time delay system.