Dynamic Task Offloading in Power Grid Internet of Things: A Fast-Convergent Federated Learning Approach

With the unprecedented development of power grid Internet of Things (IoT) devices, such as the smart meter and the smart substation, various electrical applications have emerged to provide intelligent load-monitoring services. However, these computation-intensive applications significantly raise security concerns and have energy-efficient computation requirements. To tackle the problem, this article introduces a novel prediction-assisted task offloading scheme for power grid IoT to determine the optimal offloading decisions without compromising privacy. First, a federated learning approach is designed to train the task prediction model with fast training speed by selecting local servers intelligently. Then, according to the traffic loads of computation tasks at edge computing nodes and electrical users' behavioral features, a dynamic prediction-assisted task offloading method is proposed. The evaluation results demonstrate that the proposed prediction-assisted approach could efficiently ensure the power grid IoT network's loading balance and reduce the network overhead.