Uncertainty Pattern Recognition of Multi-Type Renewable Power Plants and the Coupling Mechanism with Multi-Level Voltage Sensitivity

Abstract

With high penetrations of renewable power plants, increased output volatility and uncertainty exacerbate local voltage-stability risks. Targeting heterogeneous renewable facilities, this paper proposes an integrated analysis framework that couples deep feature disentanglement, joint forecasting, and multi-level voltage sensitivity. First, using event logs as the driver, we combine variational mode decomposition (VMD) with a residual convolutional autoencoder (Res-CAE) to extract multi-band disturbance modes from power time series. An adversarial mechanism is then employed to disentangle meteorological disturbances from device-driven behaviors, yielding canonical uncertainty labels dominated by high-/mid-frequency components. Next, a spatio-temporal attention–driven multi-task sequence model is designed to jointly forecast active/reactive power and nodal voltage, providing stable inputs for subsequent sensitivity assessment. Finally, dynamic voltage sensitivity is estimated via sliding-window regression, and the disturbance–response chain and its propagation scope are quantified across local, dynamic, and system-level tiers. Event-driven simulations on the SoCal 28-Bus benchmark indicate that, under high- and mid-frequency dominated scenarios, dynamic voltage sensitivity increases by approximately 58.3% relative to low-volatility regimes, revealing a nonlinear transition by which renewable uncertainty is amplified through multi-frequency modes toward critical nodes. The proposed paradigm offers a transferable pathway for modeling and predicting complex source–network coupling in renewable-dominated power systems and provides data support for proactive voltage-control strategy design.