Persistent memory (PM) poses challenges for B+-Tree indexing due to its read-write asymmetry and durability requirements. Existing PM-optimized B+-trees employ static storage layouts (either fully ordered or unordered), which lack adaptability to dynamic workloads and lead to suboptimal performance. This paper proposes DHB+Tree, a dynamic hybrid-storage B+-Tree that adaptively balances read and write performance. DHB+Tree partitions each leaf node into ordered and unordered zones, dynamically adjusting their ratio in real time based on workload patterns. A lightweight workload-aware parameter prediction model is introduced to guide this adjustment, minimizing overall operation latency. DHB+Tree ensures crash consistency and avoids physical data movement during restructuring. Experimental results demonstrate that DHB+Tree consistently outperforms state-of-the-art static designs in both throughput and adaptability under diverse workloads.