Abstract: Accurately predicting the history-dependent mechanical response of composites is crucial for engineering applications, as it requires considering the evolution of microscopic heterogeneous field variables. The complexity of this problem leads to prohibitively high computational costs in conventional finite element analysis, making it challenging to achieve micro-macro coupled multiscale analysis. To address this issue, this paper proposes a Temporal Convolutional Network (TCN) modeling method of composites based on clustered reduced-order homogenization. Based on the theory of Clustered Nonuniform Transformation Field Analysis (CNTFA) and a cyclic-random hybrid sampling strategy, the CNTFA model is used to generate datasets for training the TCN surrogate model for a significantly improved computational efficiency. To further reduce training costs, a transfer learning strategy is introduced for similar problems. By leveraging the pre-trained TCN model from the source problem and freezing certain residual blocks, only a small amount of target problem data is required to fine-tune the parameters of remaining residual blocks, yielding an accurate surrogate model for the target problem. The proposed method maintains a framework design independent of microscopic constitutive models, ensuring a wide generality. Finally, a numerical study is conducted on the macroscopic effective properties of fiber-reinforced composites. The results demonstrate that while maintaining high prediction accuracy, the online computational efficiency of the TCN surrogate model is improved by two orders of magnitude compared to the CNTFA model.