A Distributed and Secure Resource Allocation Method for Power Communication Networks Based on Policy Distillation

INTRODUCTION: In the next-generation smart grid communication architecture, how to achieve secure, dynamic, and fine-grained network resource allocation to ensure differentiated QoS for various services has become a key challenge. OBJECTIVES: Therefore, this study proposes a lightweight resource allocation method based on constrained policy distillation to address the challenge of balancing lightweight deployment with strong security assurance in power communication networks. METHODS: By integrating Graph Neural Networks (GNNs) and Bidirectional LSTM (Bi-LSTM), the model extracts three-dimensional features of topology, service, and resources to construct a 128-dimensional joint state representation. Moreover, a multi-objective reward function is designed that employs a double Q-network to mitigate value overestimation and generate a high-fidelity decision trajectory library. Through service-constrained policy distillation, the model innovatively combines KL divergence loss, a squared hard-constraint loss, and a soft-constraint L2 loss to compress the teacher model into a student model, subsequently compiled and deployed at the edge. Finally, a rule engine layer dynamically adjusts priorities for intercepting critical violations and ensures the security of the power system. RESULTS: Experimental results based on real-world power grid datasets demonstrate that our model achieves superior performance in resource efficiency, security, and edge effectiveness, effectively balancing lightweight deployment with strong security assurance in resource allocation for power communication networks. CONCLUSION: It can be seen that this method enables distributed and secure resource allocation in power communication network environments, thus providing reliable QoS guarantees for new-type power systems.