Storage optimization method for lakehouse architecture in education systems with periodic workloads

Educational Lakehouse architectures face the dual challenges of growing data volume and multi-scale periodic access loads driven by pedagogical operations (such as academic calendars and course schedules); however, existing storage management methods lack an understanding of such operational semantics and the capability for proactive decision-making. To address this, this paper proposes an intelligent storage optimization method. First, it constructs a Multi-Scale Teaching Period Load Model (MSTP-LM), through semantic encoding and periodic embedding, translates structured operational information into machine-interpretable, multi-granularity temporal features, thereby achieving a deep integration of business rules and load patterns. Second, it designs a Multi-Factor Dynamic Value Evaluation Model (MFD-VE), which integrates access frequency, recency, and pedagogical relevance to quantitatively assess the time-varying characteristics of data across different educational stages in real time. Finally, it establishes an Intelligent Storage Migration Decision Framework (DQN-SM) based on Deep Q-Networks, utilizing a dual-stream network architecture to process the aforementioned heterogeneous features, and incorporating curriculum learning and reward shaping strategies, this framework enables the dynamic and adaptive migration of data across hot, warm, and cold storage tiers. Experimental results demonstrate that the MSTP-LM reduces the Mean Absolute Percentage Error (MAPE) to 5.8%, while the MFD-VE achieves a value Evaluation accuracy of 93.2%. Furthermore, compared to a rule-based baseline, the DQN-SM migration strategy reduces average access latency by 71.1% and storage costs by 24%, while boosting the Service Level Agreement (SLA) compliance rate to 99.1%.