The characteristics of nonlinearity and time-varying changes in most industrial processes usually cripple the predictive performance of conventional soft sensors. In this article, moving-window Gaussian process regression (MWGPR) is proposed to effectively capture the process dynamics and to model nonlinearity simultaneously. Applications of the proposed MWGPR method to the modeling of the activity of a catalyst and an industrial propylene polymerization process are presented. The results clearly demonstrate that the MWGPR method effectively tracks the process changes to generate satisfactory predictive performance. Two modeling strategies, namely, dual updating and dual preprocessing, are applied to MWGPR, in an attempt to more efficiently track the process dynamics. Dual updating takes into account both the time-varying variance of the process and the bias between the actual measurement and the model prediction. The improvement in performance is illustrated by a case study on modeling catalyst activity. Simultaneous removal of embedded noise in both process parameters and process output variables by dual preprocessing could significantly improve the predictive capability of MWGPR, as illustrated by the performance of a modeling study of an industrial propylene polymerization process.