Extant graph-based models for EEG emotion recognition have primarily focused on the direct interconnections between points within the EEG channel space, overlooking the deeper temporal relationships inherent in EEG signal sequences. This constraint results in less-than-ideal effectiveness for subject-independent EEG-based emotion recognition tasks. The current paper introduces an innovative approach that employs hypergraphs to grasp more advanced temporal connections within EEG signals. The proposed Temporal Hypergraph Convolutional Network (THGCN) model is designed to elucidate the intricate and dynamic interaction patterns of EEG signals across different emotional states by constructing temporal feature hypergraphs. This approach not only grasps the temporal features of EEG signals but also elucidates the complex interactions between signals across different emotional states. We demonstrate its effectiveness through comprehensive experiments on widely recognized emotion EEG datasets, such as SEED and SEED-IV. The results from the experiments suggest that the proposed method surpasses alternative approaches, achieving a recognition accuracy of 88.2% for the subject-independent experiment and 90.9% for the subject-dependent experiment on the SEED database. Similarly, the SEED-IV database results in an accuracy of 75.2% for the subject-independent experiment and 78.7% for the subject-dependent experiment.