Research

Physics-Informed Neural Networks with Federated Learning for Microgrid Management

Standard neural surrogates for power systems learn to fit data but happily violate physics. We constrain the network instead — Kirchhoff's laws are enforced directly in the loss function — and train it federated, so each edge site learns locally and never ships raw data anywhere.

• Proposed a physics-constrained federated learning framework enforcing Kirchhoff's laws directly in the loss function, enabling edge training without sharing raw data.
• Delivered an order-of-magnitude inference speedup and markedly faster reinforcement learning convergence, without any site ever exposing its raw data.