Simplified stochastic configuration network-based optimised soft measuring model by using evolutionary computing framework with its application to dioxin emission concentration estimation

Abstract

With a supervisory mechanism to randomly assigning input weights and biases, the prediction performance of the soft measuring model of industrial process has been improved by stochastic configuration networks (SCNs). Although SCNs theoretically exhibit a universal approximation capability, the learning parameters generate considerable fluctuation for the evaluated performance. Hence, addressing the parameters by an intelligent optimisation method is necessary. Thus, this study investigates the parameter optimisation of soft measuring model based on simplified SCN (SSCN) by using the evolutionary computing (EC) framework. A searching strategy based on EC theory is used to optimise jointly the input features and learning parameters of the soft measuring model. Moreover, sensitivity and robust analysis of key learning parameters are performed. Experiments on benchmark datasets and dioxin emission datasets from municipal solid waste incineration with different sizes and dimensions are conducted to validate the proposed strategy.