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

Abstract

The growing penetration of converter-interfaced generation, flexible demand, and prosumer assets is tightening the physical coupling between transmission and distribution networks. As a result, operators increasingly benefit from coordinated transmission-distribution (TSO-DSO) state estimation (SE) that provides a consistent view of bus voltages across voltage levels. In practice, however, TSOs and DSOs operate under strict data-governance constraints: measurements, historical archives, and detailed network information cannot be freely pooled across organizational boundaries. This limits the feasibility of centralized, data-hungry learning pipelines and complicates coordinated SE at scale.

This thesis proposes a federated learning (FL) framework for coordinated SE of an integrated TSO-DSO system. FL enables multiple operators to collaboratively train a shared neural estimator without exchanging raw data: each participant performs local training on its own measurement/state pairs and only model updates are aggregated on a central server. To evaluate the approach in a controlled setting, a steady-state digital twin of the Dutch transmission system is extended with three reduced-order distribution feeders (representing three DSO regions). A large synthetic dataset is generated through power-flow simulations, and measurement uncertainty is modeled via bounded uniform noise injected into the input features. The estimator predicts bus voltages in real and imaginary components rather than magnitude-angle form, avoiding phase-wrap discontinuities and improving numerical stability during training.

We establish a strong centralized baseline and study the effect of a warm-start strategy, where the model is initialized from a pre-trained checkpoint to accelerate convergence. We then implement a so called masked-target FL to reflect partial observability and data locality: each client trains on a restricted subset of inputs and supervises only its own target voltages. Three aggregation strategies are compared: FedAvg, FedProx, and FedAdam. This comparison will be in terms of convergence and global test accuracy over the full system state. The results show that FL achieves stable convergence and preserves data locality, while a substantial gap remains between federated and centralized test accuracy under partial labels and heterogeneous client distributions. Among the FL strategies, FedAdam provides the best global performance, indicating that server-side adaptive optimization can mitigate client drift in this setting.

Overall, this work demonstrates the feasibility of FL for multi-operator SE training in TSO-DSO systems and provides an end-to-end, reproducible benchmark covering digital-twin modeling, dataset generation, centralized baselines, and federated training. The findings also highlight key challenges, notably non-IID data, unequal target cardinality, and evaluation mismatch between masked validation and global testing, and motivate future work on personalization, hybrid neural WLS refinement, and privacy-preserving inference for real-time digital-twin applications.