Dynamic soft sensor algorithm based on nonnegative garrote and long short-term memory neural network

In modern industrial process modeling, the multivariable, nonlinearity and dynamics of the production process increase the model complexity and reduce the model accuracy. In response to this problem, a dynamic soft-sensing algorithm based on the long short-term memory (LSTM) neural network and its input variable selection is proposed by embedding the nonnegative garrote (NNG) into the LSTM neural network. First, a well-trained LSTM neural network is generated with parameter optimization, in which the dynamics and time-delay of industrial processes are handled by its excellent memory capacity of historical information. Then, the NNG algorithm is used to compress the input weights of the LSTM network to eliminate the redundant variables and improve the model accuracy. Grid search and blocked cross-validation are used to find the optimal hyperparameter of the algorithm. Finally, the algorithm is applied to the soft-sensing modeling of SO2 concentration in the flue gas that is discharged from the desulfurization process of a thermal power plant, and the performance of the algorithm is compared with other state-of-the-art algorithms. Experimental results demonstrate that the proposed algorithm can effectively delete the redundant variables, reduce the model complexity and improve the prediction performance of the model. © 2023 South China University of Technology. All rights reserved.