This paper develops different approaches for synergistic integration of physics-based models with data-driven models for modeling of high temperature power plant superheaters. Two types of data-driven models are developed, namely all-nonlinear static-dynamic neural network models as well as models obtained using a Bayesian machine learning approach. Series, integrated, and parallel coupling of first-principles models and data-driven models are proposed. Algorithms for solving the training problems for the data-driven models and for simulation of the hybrid models are developed. The developed approach is applied to high temperature power plant superheaters that face considerable modeling and measurement challenges. Measurement challenges arise due to harsh operating conditions that not only make placement of sensors difficult, but life of the placed sensors very limited. Modeling challenges arise due to complex inhomogeneous spatial distribution of flue gas and steam flowrates, especially under load-following operation and uncertainties in heat transfer characteristics because of multiple factors such as stochastic spatio-temporal variation of ash deposits on the superheater tubes in coal-fired power plants, internal oxide scale formation in the tubes, etc. First-principles two-dimensional differential–algebraic equation models of two industrial superheaters, one from a natural gas combined cycle plant and another from a coal-fired power plant, are developed. The results from the models are compared against industrial operational data. For all cases studied in this work, it is observed that both for steady-state and dynamic data, the hybrid models have much higher accuracy than when only the respective first-principles models are used. For example, during prediction of the average outside tube wall temperature of superheater tubes at the combined cycle power plant, the root mean squared error observed for the standalone first-principles model improved from 12.3 °C to 0.5 °C post hybridization with data-driven models. Similarly, while predicting the main steam outlet temperature in the coal-fired plant, the hybrid models resulted in reduction of the root mean squared error value from 19.5 °C to around 2 °C. In addition, the hybrid models yield highly resolved spatio-temporal temperature profiles that can be helpful for developing monitoring approaches of these critical components under load-following operation.