Work

Graph foundation models for intelligent monitoring and control of renewable energy networks

Year: 2026
Type: article
Abstract: The rapid expansion of renewable energy resources has transformed modern power systems into highly interconnected and dynamic networks that require intelligent monitoring and adaptive control. Existing graph-based approaches mainly learn structural relationships from historical data and do not explicitly respect the electrical laws of the network. This paper proposes a physics-informed graph foundation model (PI-GFM) coupled with a digital twin (DT) for monitoring and control of renewable energy networks. PI-GFM is an impedance-aware graph encoder that is pre-trained in a self-supervised manner on pooled heterogeneous network data and then adapted through task heads for fault classification, bus-voltage estimation and one-hour-ahead DT forecasting. Electrical knowledge enters as soft penalties on nodal active and reactive power balance and on the linearised (LinDistFlow) branch voltage drop; the full nonlinear AC power flow is not enforced, and this scope is stated explicitly. The DT synchronises a latent network state over a four-hour window through a learned fusion, under explicitly defined operating limits, and its voltage estimates drive a reactive-power support controller. The framework is evaluated on a fully specified simulation of the IEEE 33-bus feeder with photovoltaic, wind and battery resources, solved with AC power flow and subjected to inverter faults, high-impedance line faults, voltage-instability events and reconfigurations, using chronological splits, 5 random seeds and DNN, LSTM, GCN, GAT and GraphSAGE baselines. On the unseen test season, PI-GFM attains a bus-voltage MAE of 3.31 × 10 −3 p.u., 17% lower than the best baseline (3.96 × 10 −3 p.u.), and a DT synchronisation accuracy of 84.9% of bus-hours within 0.005 p.u. (best baseline 75.9%). Its four-class accuracy is 96.39% (best baseline 97.12%), with a fault-detection false-alarm rate of 1.29%. Used as controller input, it keeps all buses within 0.95–1.05 p.u. in 88.2% of test hours, against 85.8% without control and 96.9% with perfect state knowledge. The ablation shows that both the physics terms and the DT reduce the estimation error, and pre-training improves low-label performance. The results are obtained on simulated data, on one feeder family, on a single CPU platform.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7233715616
FWCI:
Cites: 34
Cited by:
Related to: 10