One of the keys to the safe operation of nuclear power plants is to achieve accurate and rapid prediction of their operating parameters. In this paper, a data fusion prediction method has been developed which fuses multi-fidelity simulation data with measurement data. First, a variety of simulation data similar to the measurement data are selected and sorted according to fidelity, to train the GRU neural network to build a pre-trained model. Second, some of the measurement data are used to fine-tune the model to improve the prediction accuracy. Finally, the fine-tuned model is used to predict the future state of operating parameters. The feasibility of this method is verified by using measurement data of a steam generator heat transfer tube rupture accident simulated by a PKL thermal hydraulic test bench and multiple sets of similar RELAP5 simulation data. The effectiveness of each part of the method is also illustrated by ablation experiments.