In next point-of-interest (POI) recommendation, some approaches rely on extracting static information from user check-in sequences to infer user preferences for POIs. However, these methods lack the capability to model users’ dynamic transitions between POIs. Moreover, some methods ignore user preferences for categories, which makes them unable to explore user major intentions and thereby results in limited recommendation performance. To address these two issues, this paper proposes a next POI recommendation algorithm, which leverages a spatial-temporal knowledge graph and models user dual-level preferences (KGDP). Firstly, KGDP constructs a spatial-temporal knowledge graph to uncover both user static check-ins and dynamic transition relationships between POIs. Then, user preferences for POIs are explored by generating user check-in representations through a knowledge graph and an attention mechanism. In addition, users’ higher-level preferences for POI categories are extracted from category sequences using a Transformer mechanism to enhance recommendation effectiveness. Finally, users’ dual-level preferences are linearly integrated to generate the final POI recommendation result for the next visit. Experimental results on four real-world datasets demonstrate that KGDP achieves significant improvements over baseline models in terms of both recall (Rec) and normalized discounted cumulative gain (NDCG).