The fouling state of radiant heat absorption surface in power station is dynamic, changing with the load and fuel and so on. Traditional modeling method for fouling state such as linear regression and ANN is used to establish the off-line static model. But this offline static model must constantly correct with online data to guarantee long-term application. If the model only uses new data to modeling then it will lose the useful information of the dynamic process. It is difficult to calculate and store large data sets with new data and old data combined. This paper present a method based on nonlinear regression PLS, taking into consideration not only the present state of process, but also the information extracted from the old data. Then the model can be update with the changes of operating conditions, automatically. A simulation for fouling state of radiant heat absorption surface, in 300MW boiler, using the presented method is carried out. The results show that predictive model can adapt to the dynamic process.