To address the dynamic voltage variations in distribution networks caused by the integration of a high proportion of renewable energy and demand response
a convolutional neural network (CNN)-graph neural network (GNN)-Koopman-based distribution network voltage dynamic assessment method considering demand response is proposed. Firstly
Latin hypercube sampling is employed to construct wind-photovoltaic-load uncertainty scenarios. These are combined with demand response peak-shaving and valley-filling strategies
and power flow calculations are performed to build a dynamic voltage sequence dataset. Secondly
CNN is utilized to extract the spatial local features of node voltages
while GNN is used to model the topological correlations within the distribution network. The Koopman operator is introduced to achieve global linear dynamic evolution in a high-dimensional embedding space
thereby constructing an end-to-end voltage time-series dynamic assessment model. Simulation results demonstrate that the proposed method achieves significantly higher assessment accuracy compared to methods such as CNN-Koopman and GNN-Koopman. Moreover
it can accurately capture the regulating characteristics of demand response on node voltages during load valley and peak periods.
HAN F,LI J,QIAO J,et al. Framework design of power system situation cognition and operational risk assessment technology based on probability statistics and deep learning [C ] //2025 10th Asia Conference on Power and Electrical Engineering(ACPEE).Beijing,China,2025:900-904.
LEI C,BU S,CHEN Q,et al. Decentralized optimal power flow for multi-agent active distribution networks:a differentially private consensus ADMM algorithm [J ] .IEEE Transactions on Smart Grid,2024,15(6):6175-6178.
XIAO F,YIN X,TANG J,et al. Research on networked protection scheme of active distribution networks with doubly-fed wind turbines and electric vehicles [J ] .IEEE Transactions on Industry Applications,2025,61(4):5946-5961.