Abstract: Overheating and the resultant tube failures of boiler high temperature heating surfaces such as platen superheater is one of the sore major problems affecting safe operation of coal-fired power generating units. Accurate prediction of metal tube temperature distribution improves operation safety and reduces the risk of tube-overheating. Due to the limitation of measurement method and the huge time cost of numerical method, artificial neural network was used to predict the tube temperature distribution. However, the traditional artificial neural network methods lack interpretability and overly rely on training samples, leading to poor generalization ability under insufficient data scenarios. Therefore, this paper proposed a novel KEE-PINN （Knowledge Extraction and Embedding – Physical Informed Neural Network） method for platen superheater tube temperature distribution, which integrates the data mining algorithms with physical informed neural networks. Firstly, in order to find out the boiler operating parameters that affect the tube temperature distribution of the platen superheater, a data mining model based on association rules was established to analyze the historical operating data of boiler. And, a CFD typical operating condition database was established based on the obtained association rules. Subsequently, a deconvolution neural network that uses the boiler operating parameters as the input and the platen superheater tube distribution as the output was constructed. And, the monotonic relationships obtained by the data-mining model were embedded into the deconvolution neural network model, prompting the model to obey the mechanisms and effectively inhibiting the overfitting or underfitting issues. Taking platen superheater of a 600 MW super-critical wall-fired boiler, the results shows that the KEE-PINN model can maintains adherence to the verified monotonic physical relationship between parameters even in unknown operating conditions, and can realize the real-time prediction of tube temperature distribution. The MSE error of PINN model is only 0.023. Both the accuracy and the response speed of the model meet the demand of coal-fired plants for real-time monitoring of tube temperature distribution. In addition, the model adopts the confidence of the association rules as the weights of physical loss terms, instead of obtaining it from adaptive algorithms due to the lack of physical interpretation. This makes the weights of physical loss highly interpretable and statistically dependent, and solves the drawback of not being able to balance each physical loss term in traditional PINN models.