A hybrid modeling scheme for electrostatic precipitators (ESPs) integrating particulate matter (PM) removal mechanism and the data-driven compensation method is presented in this article. The structure of the training and coupling method for the mechanism model, which considers electrodynamics, including corona discharging, particle charging, and removal, and the data-driven method that was trained based on operation data to compensate for the distortion and flaw of the mechanism model was constructed. Design parameters and operation data of an ESP in a coal-fired power plant were collected for model training, and prediction performances were compared. Different combinations of correction factors, including exponential, proportional, and bias, were studied. The form with the highest prediction R-2 of 0.889 on the training set and 0.867 on the validation set was chosen as the mechanism part of the model. Deep neural networks with the highest R-2 of 0.885 on the training set and 0.857 on the validation set were also trained as a comparison group. The performance of the hybrid model achieved a R-2 of 0.933 on the training set and 0.887 on the validation set and outperformed the mechanism and data-driven models, thereby indicating that the hybrid modeling method could enhance prediction performance while maintaining the generalization capability of the PM removal mechanism in ESP. (C) 2018 Elsevier B.V. All rights reserved.