Soft sensor modeling based on two-stage sparsity and normalized long short-term memory

To address the multivariate and data redundancy issues in modern industrial process modeling, a two-stage sparse normalized long short-term memory neural network model is proposed. First, in the model construction phase, a feature traversal sparsity module is designed, which combines L1 regularization to perform sparse optimization on input features, thereby mitigating the interference of redundant and irrelevant features. The structure of the long short-term memory network is improved by introducing a normalized state, which dynamically mitigates the interference of input redundancy and noise on gating units and avoids the accumulation of low-value information in the state. Second, in the model optimization phase, a regularization strategy fusing Smoothly Clipped Absolute Deviation and Spearman's rank correlation coefficient is proposed to select key features and control redundant features, enhancing the model's sparsity and robustness. Finally, the effectiveness and superiority of the proposed model are verified through numerical simulations and soft sensor experiments on the industrial process of a thermal power plant.