Research On Modeling Of Dynamic Soft Sensor And Predictive Control Based On Recurrent Neural Network

Soft sensor is a method of selecting other easy-to-measure variables for process parameters that are difficult or costly to measure by hardware sensors,and constructing some kind of mathematical model to achieve prediction.Its features of convenient application,rapid response,easy realization of real-time monitoring,and control of key production indicators of industrial processes,have made it become a research hotspot in the field of process control in the process industry.Neural networks have excellent nonlinear mapping capability,self-learning,and self-organizing ability,which makes them widely used in the soft sensor of complex processes.Modern industrial processes often have multivariate,nonlinear and unknown time delay problems,which make the modeling accuracy of feedforward neural network insufficient.Recurrent Neural Network(RNN)has good ability to extract temporal information,which can effectively solve the process time delay problem.However,the plethora of measurable variables provided by modern industrial process measurement systems have resulted in a large amount of information redundancy.At the same time,due to the complexity of its loop structure,the network structure of RNN also has certain redundancy,resulting in poor generalization ability and insufficient modeling accuracy.How to filter the main variables and optimize the model structure in the massive data become the key to improve the prediction accuracy and model robustness.To address the above problems,this thesis uses RNN to model the process of time series with soft sensor,optimizes the input and loop structures of the network simultaneously with the help of the Nonnegative Garrote(NNG)algorithm based on a penalty function,and applies the proposed algorithm to the construction of models for predictive control.(1)A soft sensor modeling algorithm based on NNG for variable selection and structure optimization is proposed to address the nonlinear,time-delay and multivariate characteristics of data in complex industries and the redundancy of model structure.Firstly,the ability of RNN to extract relevant information between samples at different moments in the time-series data is exploited to obtain a well-trained dynamic soft sensor model.Secondly,the penalty function-based NNG compression algorithm is used to add penalty parameters to the input weights and loop structure weights simultaneously,and rolling cross validation is used to determine the optimization of the strangulation parameters to construct a simultaneous optimization algorithm for the input variables and loop structure under strangulation parameter constraints,which achieves variable selection.At the same time,the compression of the loop structure parameters reflects the importance of different time delay information in the data on the impact of the target output.The proposed algorithm shows the highest prediction accuracy compared to other algorithms tested on artificial data sets.(2)The proposed NNGSO-RNN algorithm is applied to the modeling and prediction of boiler inlet circulating air in the dry coke quenching industrial process.Firstly,the dry coke quenching characteristics,system structure and process flow are described in detail.In order to maximize the steam production in production target,after analyzing the data from monitoring point,select the boiler inlet circulating air temperature,which is directly related to the steam yield and involved in heat exchange,to build a soft sensor model to facilitate the control of the temperature within the maximum permissible range.Experimental results show that the algorithm can successfully predict the dynamic changes in circulating air temperature,and the importance of the different variables provides theoretical and technical support for optimization and control system design.(3)Based on the soft sensor model established by the proposed NNGSO-RNN algorithm,the RNN-based model prediction control algorithm is investigated and a typical Continuous Stirred Tank Reactor(CSTR)process is used as a typical case study to verify the effectiveness of the control algorithm.The CSTR process simulation module is first built using Simulink to generate the process reaction data,then the system is modelled data-driven using NNGSO-RNN to obtain an accurate predictive model,and finally the rolling optimization method and feedback correction method are investigated to achieve predictive control of the target variables of the studied process.Compared with PID and feedforward neural network-based model predictive controllers,the established predictive controller has superiority in terms of smoother adjustment of operating quantities and more accurate tracking of set values.