Deep learning-based intelligent perception of dynamic voltage limits for transmission-distribution coordinated control under high renewable penetration

Abstract

High penetration of wind power, photovoltaic generation, energy storage and power-electronic interfaces makes modern voltage profiles stochastic and spatially coupled. Fixed voltage limits or locally tuned Volt/VAR rules cannot adapt when one transmission boundary bus coordinates multiple active feeders with heterogeneous renewable fluctuations. This paper proposes TDN-LimitNet, a deep learning based perception method for dynamic voltage limits in transmission-distribution coordinated control. Multi-source measurements, renewable forecasts and topology data are converted into a spatiotemporal graph state; a dilated temporal convolutional network extracts fluctuation features, and a physics-guided interval head estimates adaptive upper and lower limits learned from scenario-based AC optimal power-flow labels. A safety projection layer maps the perceived interval into feasible instructions for on-load tap changers, reactive devices, smart inverters and optional active-power curtailment. Studies on a transmission equivalent connected with modified IEEE 33- bus and 69-bus feeders cover renewable penetration from 30% to 75%. Relative to fixed limits, TDN-LimitNet raises voltage compliance from 88.3% to 96.7%, reduces the normalized violation index by 61.0%, and limits average inference time to 12.6 ms per control interval. Dynamic limit perception therefore provides a practical interface between system level voltage security and feeder-level voltage quality.