With the emergence of cyber threats and growing complexity of networks, advanced firewall management strategies are essential for maintaining network security. This paper proposes a pioneering deep-learning framework for firewall access policy optimization, formulating them into common problems such as outdated, misconfigured, and permissive rules that degrade network integrity. Through GNNs and DBNs based prediction of firewall rule behavior at the hierarchical, global, and regional policy levels, the proposed methodology can analyze the aspects of firewall access rule behavior. With a simulated packet flow within the firewall access ruleset, this method improves the visibility of anomalies, misconfigurations of the rules, shadowing policies, access rules that are no longer in use, and problems that are rarely fixed by traditional approaches. The results highlight the significant benefits in terms of the effectiveness of implementing security system functionality. The deep learning model achieved 91.7% training and 89.3% validation accuracies over 50 epochs. It outperformed traditional methods in detecting misconfigurations in rules (92.5% accuracy, 24.2% improvement), shadowed policies (88.3% accuracy, 25.8% improvement), and outdated rules (85.6% accuracy and 26.2 improvement). By employing this innovative method, we enhanced the optimization of firewall access policies and offered a scalable and flexible solution to reduce cybersecurity threats, minimize unauthorized access, and enhance the decision-making processes.