Graph-structured data provides an intuitive representation of complex relationships among entities, enabling a wide range of applications across diverse domains. As a powerful tool for processing such data, Graph Neural Networks (GNNs) have emerged as a prominent research focus. However, similar to most deep learning methods, GNNs face significant challenges in out-of-distribution (OOD) generalization. Moreover, due to the inherent complexity of graph-structured data, existing OOD generalization techniques are not directly applicable. To address this issue, this paper investigates the OOD generalization problem in GNNs from a causal perspective and propose a novel plug-and-play method named Causal Stabilization Learning (CSL). It firstly introduces a feature-space data augmentation framework that projects samples into a high-dimensional space, combined with causal clustering to infer latent causal environment labels. These latent labels achieve learning of feature-level causal stabilization, through which cluster-invariant causal features are identified. This mechanism guides the GNNs to give higher prioritization to stable causal features that demonstrate consistent patterns across different environments. Experimental results demonstrate that the OOD generalization of GNNs is significantly enhanced by the proposed CSL method in tasks involving complex structures and heterogeneous data distributions.