Dynamic graphs are widespread in social networks, biological networks, and recommendation systems. Community detection in dynamic graphs presents several challenges because these graphs continually change over time. Traditional methods often tackle the problem by clustering graphs separately at each timestep and matching these communities. Unfortunately, this strategy does not account for the temporal continuity and can lead to instability in community detection. Moreover, most current methods do not consider the scenario where nodes have attributes. To address these challenges, we introduce a multi-level contrastive graph clustering approach for Dynamic Graphs (MLCDG), a novel methodology employing deep clustering in the context of dynamic graph neural networks. MLCDG innovatively incorporates contrastive learning to generate clustering-oriented latent representations that effectively capture both node-level and temporal-level community structures. Our approach stabilizes community detection by maintaining temporal coherence and minimizing clustering disruptions caused by dynamic changes. The methodology outperforms existing state-of-the-art techniques, demonstrated by our experiments on real-world and synthetic datasets, validating our approach’s effectiveness in enhancing community detection in attributed dynamic graphs.