Multi-fidelity graph neural networks for efficient power flow analysis under high-dimensional demand and renewable generation uncertainty

Abstract

The modernization of power systems faces uncertainties due to fluctuating renewable energy sources, electric vehicle expansion, and demand response initiatives. These uncertainties require consideration in power flow analyses by calculating power flow across a spectrum of generation and load values. Traditional power flow methods, which rely on iterative solutions of non-linear equation sets, become computationally burdensome under these uncertain conditions. Surrogate models have been used to reduces the computational cost of power flow analysis under uncertainty. Recently, graph neural networks (GNNs) have gained increasing attention as surrogates for power system simulations. However, GNNs, similar to other machine learning models, require significant amounts of training data. This paper proposes an innovative approach combining a multi-fidelity methodology with a GNN-based surrogate model for efficient power flow calculations. This model is trained using both high-fidelity power flow simulations and low-fidelity power flow simulations. Our multi-fidelity GNN model not only reduces the cost of generating training data, but also outperforms both single-fidelity GNN models trained solely on high-fidelity data and conventional neural network models. Additionally, the proposed model exhibits robustness to minor topology changes and achieves reasonable performance with unseen topologies. The model’s effectiveness is validated through simulations on standard IEEE systems.