面向配电网数字孪生模型构建的边缘协作方法

Edge Collaboration for Construction of Distribution Network Digital Twin Model

边缘计算通过计算资源的下沉为配电网数字孪生(digital twin,DT)模型训练提供算力支撑,但仍需对边缘层计算资源进行整合优化,提升模型构建效率与精度。针对此,该文提出面向配电网DT模型构建的边缘协作方法。首先,建立基于边缘协作的配电网DT模型构建框架,通过局部模型训练与全局模型整合实现配电网孪生。其次,建立协作训练损失函数,在保障DT模型长时同步率约束前提下,最小化模型损失。最后,借助李雅普诺夫优化构建与DT模型长时同步率约束相关的虚拟队列,针对无线信道时变性以及协作决策耦合性带来的信息不确定、不对称问题,提出基于深度强化学习的配电网DT模型边缘协作构建算法。仿真结果表明,相较于其他两种传统算法,所提算法在保障DT模型长时同步率约束前提下,可分别降低全局模型损失34.94%和55.93%,降低样本数据队列积压27.40%和19.68%。

The edge computing enables the training of the digital twin(DT)models for a distribution network by leveraging the edge computing resources.However,to enhance the efficiency and accuracy of the model construction,it is necessary to integrate and optimize the edge computing resources.This paper proposes an edge collaboration for constructing the DT models of a distribution network.First,an edge collaboration-based framework is established to achieve the twin synchronization through the local model training and the global model integration.Second,a collaborative training loss function is proposed with the aim of minimizing the model loss,all while ensuring the long-term synchronization rate constraint of the DT model.Finally,the Lyapunov optimization is used to construct a virtual queue related to the DT model's long-term synchronization rate constraint.To overcome the information uncertainty and asymmetry problems caused by the wireless channel variability and the collaborative decision coupling,a deep reinforcement learning-based edge collaboration algorithm is proposed for the DT model construction in the distribution network.Simulation results show that,compared to the baseline algorithms,the proposed algorithm is able to respectively reduce the global model loss by 34.94%and 55.93%,and decrease the sample data queue backlog by 27.40%and 19.68%,while ensuring the long-term synchronization rate constraint of the DT model.