| Industrial time series modeling and forecasting is the basis of system optimization control and intelligent scheduling,which is of great significance for cement,steel,chemical and other enterprises to reduce carbon emissions and energy consumption,and improve product quality and production efficiency.Combined with the basic theory of industrial timing modeling and deep learning,this paper studies the industrial timing prediction algorithm from the aspects of model input selection,deep network architecture and integration with traditional models,aiming at the characteristics of production process such as multivariable,large lag,non-linearity and strong coupling.The main work is as follows:(1)Aiming at the problems of multivariable and large lag in the production process,this paper proposes a model input optimization that combines principal component analysis and moving grey relational analysis(PCA-MGRA).PCA selects key variables from the multivariate variables as the input for modeling,and better retains the timing information required for modeling.On the other hand,the time-lag dynamic changes of various variables are judged through MGRA,and the model-driven information is optimized.Applied to actual industrial data,the experimental results show that the PCA-MGRA method effectively solves the problems of data redundancy and timing mismatch,and improves model modeling and prediction accuracy.(2)Aiming at the strong coupling and nonlinear dynamic characteristics of continuous production processes,this paper proposes a hybrid deep neural network prediction model named deep convolutional neural network and gated recurrent unit network(DCGNet).The model is composed of three modules and aims to jointly solve the modeling problem of multivariable nonlinear dynamic coupling in complex industrial processes.First of all,multi-layer CNN in series with GRU focuses on extracting strong coupling features of multivariate variables.Secondly,the GRU network is used to accurately capture the long-term dynamic dependencies relationship of the variable to be measured.Finally,a fully connected layer(FC)is used to connect these two parallel modules,and single-step prediction is achieved through weighted fusion of features.Experimental results show that this method is superior to the traditional modeling and prediction algorithm and has robustness.This method of combining relevant feature selection to build a model can be a special reference for other time series modeling.(3)Aiming at the characteristics of periodic quasi-linear variation trend and local unstable fluctuation of signals in the intermittent production process,this paper proposes a hybrid model combined convolutional-gated recurrent unit networks and statistical autoregressive model(CG-AR).The model takes into account the influence of related variables,as well as the linear and non-linear characteristics of the signal to be measured.The traditional AR model simulates the linear trend of the change of the variable to be measured.Considering that the strong coupling in the production process will cause nonlinear fluctuations,CNN-GRU is used to characterize the nonlinear dynamic coupling relationship of multivariate variables.Combining theory with practice,the proposed method is applied to real industrial data sets,and the experimental results show that the proposed method has a good prediction effect. |
