| Real-time monitoring of certain key variables that are difficult to measure directly but are closely related to the modern industrial process is necessary to ensure product quality and safety of the production process.Soft sensor techniques are widely employed in various modern process industries and process control areas to indirectly accomplish effective estimation of the dominant variables by exploiting the relationship between easily accessible auxiliary variables and difficult-to-measure dominant variables to build mathematical models.However,in view of the problems of the multivariate,nonlinearity and dynamic nature of production processes in modern industrial process modeling lead to higher model complexity and lower modeling accuracy.A nonlinear optimization problem that combines penalized likelihood and long short-term memory（LSTM）neural networks based on sparse learning theory and optimization method in this paper is designed to develop research on dynamic soft sensor algorithms for complex industrial processes.The main research contents of the thesis are as follows:（1）A dynamic soft sensor algorithm（NNG-LSTM）based on LSTM neural network and its input variable selection is proposed by combining the nonnegative garrote（NNG）algorithm with long short-term memory（LSTM）neural network for the problem of modeling dynamic soft sensor of complex industrial processes.At first,the outstanding historical information memory capability of the LSTM neural network is employed to deal with the dynamic and time lag problems in industrial processes.Then,the NNG algorithm is applied to compress the input weights of the LSTM network to eliminate redundant variables and improve the accuracy of the model.Finally,experimental results show that the proposed algorithm can effectively eliminate redundant variables,reduce the complexity of the model and improve its predictive performance.（2）A dynamic soft sensor algorithm（ANNG-LSTM）based on adaptive NNG and LSTM neural networks is proposed to address shortcomings of biased coefficient estimation in the NNG algorithm for optimizing input weights of LSTM neural networks.First,the maximum mutual information coefficient algorithm is used to measure the degree of association between input variables and target variables to design an adaptive weight vector.Then,it is embedded into the NNG-LSTM algorithm constraints to enable the algorithm to impose different degrees of penalties according to the characteristic importance of different input variables,thus realizing its adaptive input variable selection and simultaneous parameter optimization.Numerical simulation results demonstrate that variable selection results of the model built by the ANNG-LSTM algorithm are more accurate and effective,with higher prediction accuracy and generalization performance.（3）The ANNG-LSTM algorithm is applied to model the soft sensor of SO2concentration in the flue gas emissions of a coal-fired power plant through an in-depth study of the process mechanism of limestone-gypsum wet flue gas desulphurization process,and the performance is compared with other advanced algorithms to verify the effectiveness of the proposed algorithm.The experimental results demonstrate that the model built by the ANNG-LSTM algorithm outperforms other algorithms in terms of all performance indicators,and its model has fewer input variables and higher prediction accuracy,which can accurately predict the dynamic changes of SO2 concentration in the emission of flue gas. |
