Accurate traffic forecasting results are instrumental in guiding routing policies and resource allocation to ensure the quality of service in Low Earth Orbit (LEO) satellite networks. However, due to the variability and complexity of satellite traffic, as well as the constraints of onboard resources, improving the generalization ability of the traffic forecasting model while ensuring prediction accuracy becomes a key challenge. To overcome this problem, we propose a novel traffic forecasting (MG-TF) model that combines meta-learning with the gated recursive unit (GRU) to improve traffic prediction accuracy in the case of small data samples. Specifically, we apply the model-agnostic meta-learning (MAML) algorithm to optimize the initial parameters of the GRU model, thus enabling the model to quickly adapt to the new data. Our simulation results indicate the superiority and effectiveness of the MG-TF model over other baseline models in terms of prediction accuracy and learning speed in satellite traffic forecasting.