Power station boiler heated area ash is one of the important factors leading to the reduction of boiler thermal efficiency and even affecting boiler output, so accurate prediction of boiler heated area ash fouling condition is a prerequisite for optimizing and improving boiler energy efficiency. To address this problem, the paper takes the cleanliness factor of the economizer of a 300MW power station boiler as the research object, and proposes a combined model integrating timevarying filter-based empirical mode decomposition (TVF-EMD) and convolutionally gated neural unit (CNN-GRU) to predict the change of boiler heated surface energy efficiency. Firstly, the non-linear and non-smooth cleanliness factor raw data are preprocessed by the improved wavelet threshold method to remove noise and outliers; then the processed data are decomposed by TVF-EMD to obtain the preset intrinsic modal components, and the components with the threshold value greater than 0.2 are superimposed and reconstructed according to the autocorrelation coefficient, so as to reduce the impact of the low correlation components on the prediction accuracy while retaining the important features of the raw data; finally, the reconstructed model is combined with convolutional gated neural unit (CNN-GRU) to predict the energy efficiency change of the boiler heating surface. Finally, the reconstructed signal is used to establish the nonlinear relationship between input and output by using the powerful feature extraction capability of convolutional neural network (CNN) and the temporal memory capability of gated recurrent unit (GRU) to achieve more accurate time series prediction. The results show that the decomposition using TVF-EMD can improve the prediction accuracy by 9.628 67% compared with the direct prediction, which provides theoretical and technical support for the subsequent optimization and the formulation of more reasonable soot-blowing strategies.