With the rapid expansion of blockchain applications, the performance issues associated with its role as a distributed storage system are becoming increasingly prominent. The requirement for each node in the existing blockchain system to store complete transaction records has led to an exponential increase in data storage demand, which poses challenges related to storage costs and processing speed. This article proposes an optimization strategy for cold and hot hierarchical storage based on deep learning, marking the first application of artificial intelligence technology to the hierarchical storage of blockchain, thereby enhancing the storage efficiency and performance of blockchain systems. This strategy categorizes data into ‘hot’ and ‘cold’ based on access frequency and employs different storage solutions to mitigate the requirements for high-performance storage resources. Additionally, it uses deep learning algorithms to predict data access patterns, optimizing data classification and migration. The mechanism is implemented within the ChainMaker technology framework, and its effectiveness is verified through the monitoring of system performance indicators. The results indicate that the cold and hot grading strategy can significantly increase the data access speed and storage efficiency, thereby supporting the application of blockchain in big data environments.