The main steam parameters of the recovery steam generator (HRSG) are crucial for the healthy operation of the combined cycle unit. Aiming at the problems of many nonlinear operating variables inside the recovery steam generator and prolonged main steam state parameters, a prediction model of main steam parameters of the recovery steam generator incorporating the improved kernel principal component analysis (KPCA), the improved pelican optimisation algorithm (POA), and the bi-directional gated recirculation neural network (BiGRU) is proposed. Firstly, the SIS operation data of the combustion engine power plant were collected to determine the input variables by grey correlation analysis; secondly, the feature information of the input parameters was extracted by KPCA and the input dimensions were selected according to the principal component contribution ratio; finally, the Pelican Optimization Algorithm (POA) was improved and KPCA-IPOA-BiGRU was constructed to conduct the prediction test of the main steam parameters of the recovery steam generator for three pressure levels. The results show that for the outlet steam parameters of three pressure levels, the R2 of the proposed model is greater than 98%, which has better prediction effect compared with the control model.