基于MVMD与动态混合图注意力网络的多元负荷概率预测OA

Accurate load forecasting is essential for ensuring the economic and reliable operation of integrated en-ergy systems(IESs).However,the nonstationarity,dynamically time-varying coupling,and high stochasticity of multivariate loads pose significant challenges to load forecasting.To address these issues,this paper proposes a probabilistic forecasting method based on multivariate variational mode decomposition(MVMD)and a dynamic hy-brid graph attention network.First,MVMD is employed to synchronously decompose the original load sequences,enabling cross-series phase synchronization and modal alignment among multivariate loads.This process smooths the data while preserving the intrinsic temporal coupling characteristics between load series.Second,a dynamic hy-brid graph attention network is designed.Through a gating mechanism,physical prior static graphs and a meteorology-driven dynamic graphs are adaptively fused,and spatiotemporal graph convolution is incorporated to ex-tract higher-order time-varying coupling features.Finally,a quantile regression-based probabilistic output layer is constructed,extending conventional point forecasting to full probabilistic distribution forecasting.Experimental re-sults based on the dataset from Arizona State University demonstrate that the proposed method effectively quantifies load uncertainty and significantly outperforms several benchmark models in forecasting performance.