In recent decades, there has been a growing interest among scholars in the research on brain dynamics. The fusion of features from different brain atlases has proven to be beneficial for the study of brain dynamics. However, existing studies neglected the differences in atlas types during feature fusion and have failed to effectively learn their internal relationships when integrating atlases of different spatial scales. This study proposes complementary fusion to learn distinct feature representations from anatomical and functional atlases, thereby capturing complementary knowledge across different atlases by maximizing entropy. Additionally, we introduce consistency fusion to explore the intrinsic connections between atlases of different spatial scales, thereby capturing consistent knowledge across different scales through contrastive learning. Classification experiments conducted on the ABIDE I dataset demonstrate the superiority of our proposed method over other baseline models, achieving an accuracy of 72.1% and an F1 score of 75.9%.