Financial risk control models are expected to adapt to diverse and dynamic scenarios automatically, but traditional methods often struggle with heterogeneous data and cross-scenario generalization. This paper proposes a novel method integrating Heterogeneous Graph Neural Networks (HGNN) and Reinforcement Learning (RL) to tackle these challenges. We design a scenario-model bipartite graph to model complex interactions between financial scenarios and risk control models, enabling effective feature representation and scenario-model self-adaptation. HGNN captures structural and relational patterns, while Proximal Policy Optimization Reinforcement Learning dynamically optimizes model adaptation across scenarios. Evaluated on real-world financial datasets, our approach outperforms several baselines, including XGBoost, Random Forest, and SVM etc. Experimental results show that HGNN + RL achieves an AUC of 0.89, an F1-score of 0.78, and an accuracy of 0.80, surpassing other algorithms such as XGBoost by 9.9% in AUC, 6.4% in F1-score, and 5.3% in Accuracy. This study provides a scalable and adaptive solution for multi-scenario financial risk control models.