In recent years, emotion recognition based on electroencephalogram (EEG) has garnered increasing attention due to its ability to capture subtle physiological signals beyond facial or voice cues. However, traditional models often struggle with incomplete feature extraction and inadequate fusion strategies. In this study, a novel framework named Dynamic Hypergraph-based Gated Feature fusion with ECA attention (DHGF-ECA Network) is proposed. This framework combines dynamic hyperedge construction, multi-scale feature extraction, and an innovative feature fusion mechanism. First, we exploit the time-varying characteristics of phase locking value (PLV) to build dynamic functional connectivity, and incorporate an attention mechanism to adaptively adjust hyperedge weights. Second, multi-scale feature extraction is utilized to capture both local and global EEG signal patterns. Third, we introduce time-varying nonlinear entropy (TVNE) as the principal feature, while differential entropy (DE) and the Hurst exponent serve as auxiliary features. A gated feature selection strategy then fuses these features, enhancing discriminative capability. This method achieved an average high accuracy of 96.3% in the three-class classification task on the SEED dataset, outperforming other existing methods.