Federated Learning-Based Decentralized Optimization for Privacy-Preserving and Efficient Power Distribution in Nsukka and its Environs
TL;DR
This research introduces a decentralized optimization framework for improved operational performance of 11kV, 32-bus Nsukka Radial Distribution Network (RDN), using a privacy-preserving decentralized optimization framework known as Federated Learning (FL).
Abstract
Modern power distribution grids have become more sophisticated with the incorporation of Distributed Energy Resources (DERs), advanced metering infrastructure and intelligent monitoring technologies. However, traditional centralized optimization methods are communication-intensive, exposed to cybersecurity risks, and less desirable for data privacy, thus limiting their effectiveness in a smart distribution network. This research introduces a decentralized optimization framework for improved operational performance of 11kV, 32-bus Nsukka Radial Distribution Network (RDN), using a privacy-preserving decentralized optimization framework known as Federated Learning (FL). The system combines the Federated Averaging (FedAvg) algorithm with the Backward/Forward Sweep (BFS) power flow technique to optimize network operation while avoiding the sharing of raw operational data. A multi-objective optimization model was formulated, which minimizes the losses in the active power and also improves the voltage profiles, reduces communication overhead, and enhances network reliability without compromising customer data privacy. The framework was implemented using MATLAB under various loading conditions. The results confirmed that the proposed method achieved 21.0% reduction in active power losses, an increase in bus voltage by 3.9% (0.921- 0.958 p.u), 54.3% reduction in voltage deviation, 66.7% reduction in communication overhead, 38.9% accelerated convergence, 97.1% accuracy in prediction and 99.1% improvement in network reliability compared with conventional centralized optimization.
