A Pruning Regularization Stochastic Configuration Network Algorithm Based on Node Contribution and Its Application in Soft Sensor Development

Stochastic configuration network (SCN) has emerged as an efficient nonlinear regression modeling method. However, the traditional SCN has to set the model training stop criterion (e.g. the maximal hidden layer node number, or the expected error tolerance) in advance. The unreasonable parameter setting may lead to the over-fitting SCN model. To solve this problem, an adaptive pruning regularization SCN (PRSCN) algorithm based on the node contribution is proposed and its application to soft sensor modeling is demonstrated. First of all, this method builds the definition of node contribution and then assigns different weights to the hidden nodes according to the corresponding contributions. Secondly, an adaptive pruning strategy is presented to delete the redundant nodes so that an optimal network structure is obtained. Finally, the L2 regularization method is used to calculate the output weights for ensuring the good model generalization. The proposed method is applied to develop the soft sensor from one real sulfur recovery unit system, and the testing results show that PRSCN has a better predictive performance than the basic SCN method.