Abstract: The decay heat released by spent fuel assemblies is the main source of reactor core waste heat in PWR nuclear power plants. Accurate prediction of decay heat is crucial for the design and safety analysis of the nuclear power plant cooling system. Traditional methods calculate decay heat using nuclide decay simulation programs such as ORIGEN-S. However, calculating decay heat for a large number of fuel assemblies can incur high computational costs. In recent years, machine learning has been employed to predict decay heat. However, overfitting issues may arise due to insufficient data, leading to low prediction accuracy. This study establishes a co-training model based on Gaussian Process Regression (GPR) and Support Vector Regression (SVR) to generate high-quality virtual decay heat samples. These virtual samples, combined with measured decay heat data, form a mixed dataset used to train an Extreme Learning Machine (ELM) model for decay heat prediction. The results show that the co-training approach significantly enhances the stability and accuracy of decay heat predictions. After training on the mixed dataset, the prediction stability of the ELM model increased by 39.9%, and the RMSE of the predicted decay heat was 25.7% lower than that of the nuclide decay simulation program. This research provides new insights for addressing the small sample problem in the field of nuclear engineering.