The rapid digitalization of modern power systems has triggered exponential growth in critical data assets, ranging from real-time Phasor Measurement Unit (PMU) readings to equipment maintenance records. As these datasets increasingly drive grid operations and decision-making, ensuring data integrity and traceability has become paramount for power system security. Digital watermarking serves as an effective method for data provenance and tamper detection, yet existing techniques exhibit limitations when applied to heterogeneous smart grid data. Traditional watermarking fails to address the multi-modal nature of power system datasets, where numerical measurements (e.g., voltage values), categorical codes (e.g., equipment identifiers), and spatiotemporal sequences coexist. To bridge these gaps, this paper proposes a novel adaptive watermarking framework featuring three key innovations: (1) a hierarchical architecture decoupling the rule layer (feature extraction/structure adjustment) from the location layer (physical/logical embedding); (2) a random forest-driven decision engine that dynamically maps data characteristics to optimal watermarking schemes; (3) a four-dimensional optimization model evaluating data features, application scenarios, processing preferences, and watermark content requirements. Rigorous testing on real-world grid datasets demonstrates significant advantages over state-of-the-art methods: 31.5% reduction in Bit Error Rate (BER) under compression attacks, 92.7% scheme selection accuracy, and sub 100 ms decision latency for terabyte-scale datasets.