Optimal transport (OT) as an efficient way to measure differences in distribution. provide new ideas for solving the unsupervised domain adaptation (UDA) problem. However, existing work based on OT ignores the multi-modal structure of the domain and does not adequately reflect the true data distribution. Therefore, this paper proposes a new deep UDA framework, namely attention-weighted optimal transport and cluster alignment (AWOT). Specifically, the model first proposes a weighting strategy based on a self-attentive mechanism to reduce the bias caused by mini-batch selection in training. The strategy calculates weights by correlating the iterative prediction results of the source and target domains to weight the optimal in-transmission cost function. A confidence thresholding technique is also proposed to filter out low confidence target samples to enhance the robustness of pseudo-label. Moreover, AWOT is also equipped with a clustering centroid alignment loss to learn discriminative features. Comprehensive experiments have shown that AWOT outperforms existing state-of-the-art methods on several domain adaptation benchmarks.