Selective recursive partial least squares modeling and its application

The selective recursive partial least squares (SR-PLS) soft sensor modeling is proposed and its application is discussed in the paper. Aiming at the issues of high updating frequency and relatively poor real-time performance of the basic moving window recursive partial least squares (MW-RPLS) model, the selective sparse strategy of modeling samples is proposed. As a result, the SR-PLS model is then established. An absolutely relative prediction error bound (ARPEB) is set for the controlling condition of the modeling sample sparseness. A new modeling sample is selectively introduced and the oldest one is discarded. The parameters of the model are re-estimated by updating the mean and variance of the samples recursively. The developed model is then applied to the industrial process, C8-aromatics isomerization, for on-line estimation of para-xylene concentration. The simulation results show that the slow time-varying process can be dealt with effectively by the proposed SR-PLS model. The good predictive capability and real-time performance of the model is illustrated as well. The model updating frequency is reduced with the selective sparse strategy of the modeling samples and the calculation performance is then improved.