Electroencephalogram (EEG) signals exhibit nonstationary dynamics with high temporal resolution but limited spatial resolution. A critical challenge lies in identifying stable neural states during rapid emotional transitions and decoding dynamic interregional interactions. To address this, we propose a dynamic microstate temporal graph attention network (DMT-GAT) that integrates transient EEG microstates with brain functional networks. First, EEG signals are segmented into four prototypical microstates (labeled as MS1, MS2, MS3, and MS4) via global field power peak detection and K-means clustering. Emotion-related microstates (MS3/MS4) are then selected through independent t-tests based on valence and arousal ratings. Next, a brain functional network is constructed by calculating phase-locked value synchronization specifically on the time series of MS3/MS4 microstates, capturing millisecond-scale interregional dynamics during emotional shifts. Frequency-domain features are integrated into the network nodes, forming graph-structured data. Finally, a GAT with multi-head mechanisms classifies emotions by adaptively weighting node interactions. On the DEAP dataset, our method achieves average accuracies of 99.19% (valence) and 99.26% (arousal). For the SEED dataset, it maintains a robust accuracy of 95.29%. Crucially, the DMT-GAT uniquely reveals prefrontal-amygdala interactions during emotional regulation, bridging the gap between dynamic brain networks and rapid neurodynamics. This work provides a novel framework for high-resolution emotion recognition and advances understanding of neural mechanisms underlying affective transitions.