Addressing the strong nonlinearity of short-term power loads and the deficits of insufficient feature interaction and linear reconstruction errors in traditional decomposition-ensemble models, this paper proposes a short-term load forecasting model based on multi-scale parallel feature fusion and Kolmogorov-Arnold network (KAN) reconstruction. First, a feature optimization and signal decomposition strategy combining Pearson correlation coefficient-maximal information coefficient (PCC-MIC) correlation analysis and variational mode decomposition (VMD) is established to filter redundant meteorological features and mitigate sequence non-stationarity. Second, a dual-channel architecture is designed for parallel extraction using Informer and temporal convolutional network (TCN). Then, an adaptive gated fusion mechanism (AGFM) learns time-varying weights to dynamically regulate feature attention of different time steps for precise multi-scale fusion. Finally, KAN is introduced to replace conventional linear output layers. Leveraging learnable B-spline activation functions, KAN enables adaptive nonlinear mapping from high-dimensional fused features to load values. Case studies demonstrate that the proposed model significantly outperforms mainstream baselines in prediction accuracy, offering a reliable reference for power system planning and stable operation.