The development of the data element market urgently calls for scientific valuation methods. Existing approaches rely on static asset logic and fail to accommodate the dynamic fluctuations in data value across scenarios and over time. This paper proposes a dynamic valuation framework for data assets that integrates multi-agent reinforcement learning with a dynamic Shapley value mechanism. The framework reformulates data asset valuation as a collaborative, sequential decision-making process involving multiple agents, in which each data asset is modeled as an autonomous agent that learns through trial-and-error interactions within a simulated business environment, thereby unlocking its informational value. The core innovation lies in a differentiable Shapley value approximation method that quantifies the marginal contribution of each data asset in real time and in an interpretable manner, reducing computational complexity from exponential to linear and overcoming the static limitations and computational bottlenecks of traditional methods. Experiments conducted on a proprietary dataset and two public datasets demonstrate that, compared with the optimal baseline, the proposed approach achieves a 6.5% improvement in valuation accuracy, a 22% enhancement in allocation fairness, and an approximately 23% increase in convergence efficiency, with all metrics significantly outperforming existing benchmarks. This framework provides a computable and scalable dynamic valuation tool for data element markets, supporting practical scenarios such as data transaction pricing, corporate asset accounting, and revenue distribution.