Many domains feature a dynamic characteristic like logistics, sports, or medicine and it would be useful to learn about frequent patterns, e.g., in what situations a traffic jam is likely to happen. Additionally, we are facing complex situations with many objects and relations that might change over time. The learning approach developed in this work identifies frequent temporal patterns out of qualitative, interval-based descriptions of dynamic scenes by extending the Apriori algorithm and combining ideas from relational as well as sequential association rule mining approaches. Temporal relations in patterns are represented by qualitative interval relations as they have been introduced by Allen and Freksa. The search for frequent patterns is a top-down approach starting by the most general (empty) pattern and performing specialization steps by applying an optimal refinement operator. In a second step, prediction rules are generated by splitting the identified patterns into precondition and consequence parts. The developed concepts are implemented in the MiTemP system and are evaluated on synthetic data and soccer matches of the RoboCup simulation league.