Accurate influenza forecasting is critical for timely public health responses and resource allocation. To address challenges such as strong non-stationarity, spatiotemporal heterogeneity, and the prediction lag of traditional models during outbreak peaks , this study proposes a deep learning forecasting framework based on Variational Mode Decomposition (VMD) and Dual-Channel Feature Fusion (VMD-DCFF-IF). The framework first employs VMD to decompose influenza time series into Intrinsic Mode Functions (IMFs) with distinct frequency characteristics, thereby reducing nonlinear coupling. Subsequently, a parallel dual-channel feature extraction network is constructed, utilizing an improved Convolutional Neural Network (CNN) and a Spatio-Temporal Graph Convolutional Network (STGCN) to synergistically capture high-dimensional temporal patterns and spatial correlations. Finally, an adaptive fusion module comprising BiGRU and BiLSTM is designed to achieve dynamic integration of multi-source features for accurate prediction. Historical surveillance data from 2013 to 2023 from the Chinese National Influenza Center were used for validation, comparing the proposed method with mainstream models and advanced Transformer-based architectures such as Informer and Autoformer. Experimental results demonstrate that VMD-DCFF-IF significantly outperforms existing baselines in core metrics, achieving a MASE of 0.508. Phase-specific performance analysis confirms the model´s superior dynamic capturing capability during peak periods, effectively overcoming the overfitting issues common in pure attention mechanisms when processing small-sample, high-noise epidemiological data. Ablation studies further substantiate the unique contributions of each core module to enhancing predictive accuracy and system robustness. To facilitate reproducibility and foster further research, the source code and implementation details are publicly available at https://github.com/xue18334792279/VMD-DCFF-IF/tree/main.