An DAG-Based Resource Allocation Mechanism of Federated Learning for New Power Systems

The traditional federated learning framework heavily relies on a single central server, which leads to problems such as single-point failures and malicious attacks. The new-type power system brings diverse collaborative business needs of “generation-transmission-distribution-storage”. With the significant increase of sensing terminals of new-type power devices, the security protection of data generalization becomes more and more crucial, and the energy consumption of devices has become a critical bottleneck for current federated learning tasks. The DAG structure has inherent decentralization and asynchronous characteristics, which can greatly accelerate the speed of global aggregation in federated learning, and the complexity of the DAG network can ensure the security and reliability of the model. In this paper, we propose a DAG-based federated learning framework for energy-constrained new-type power systems. In order to solve the problems of energy loss and training delay in DAG-based federated learning, a resource allocation algorithm based on multi-objective differential evolution is proposed. The algorithm aims to consider the impact of device energy consumption on federated learning performance, so as to minimize the completion time and energy loss of federated learning tasks under the constraint of expected learning accuracy of edge devices in the smart grid.