Remote power flow prediction using PMU data and machine learning for enhanced DER resilience in communication loss scenarios

This paper presents a novel approach to remote power flow prediction in distribution networks using Machine Learning (ML) models trained on Phasor Measurement Unit (PMU) data. The research addresses the challenge of maintaining Distributed Energy Resource (DER) operation during communication failures with centralized Active Network Management (ANM) systems. We evaluate three advanced machine learning models-LightGBM, Temporal Fusion Transformer, and TSMixer-against traditional baseline approaches using high-resolution PMU data collected over seven weeks. The models were tested for both short-term (1-hour) and medium-term (4-hour) prediction horizons, with the Temporal Fusion Transformer achieving the best accuracy (RMSE of 21.83A and MAPE of 3.89% for 1-hour predictions). However, considering the practical constraints of edge deployment in distribution networks, LightGBM emerges as the most suitable solution, offering comparable performance (RMSE of 22.59A) with significantly lower computational requirements. The results demonstrate that local intelligence powered by machine learning can be used to maintain DER operations during communication failures, potentially reducing unnecessary curtailment and supporting grid stability.