Federated Learning-Based State Estimation for Integrated Transmission–Distribution Networks

Abstract

The growing interdependence between transmission and distribution networks, combined with increasing data privacy constraints, necessitates state-estimation methods that facilitate multi-operator coordination without requiring centralized data sharing. This paper presents a federated learning (FL) framework for neural-network-based state estimation in integrated transmission–distribution systems. Each operator trains a local multilayer perceptron (MLP) using only its own measurements, while a TSO-level server aggregates model updates to obtain a global estimator. The approach is demonstrated on a steady-state model of the Dutch transmission system coupled with three reduced-order distribution feeders representing distinct DSO regions. Three FL algorithms are compared in terms of accuracy and convergence: FedAvg, FedProx, and FedAdam. Results show that FL achieves competitive estimation performance relative to a centralized benchmark while preserving data ownership, with FedProx providing the fastest convergence and FedAdam obtaining the most accurate predictions. The study highlights the potential of FL as a foundation for privacy-preserving coordination in future digitalized and multi-operator power systems.