Federated learning, as a distributed machine learning approach, protects data privacy by training shared models locally across multiple nodes, and has been widely adopted in privacy-sensitive collaborative scenarios. However, it faces two key challenges in practical applications: First, while differential privacy mechanisms are introduced to enhance privacy protection, the added random noise leads to both reduced model accuracy and slower convergence. Second, the imbalanced feature distribution across clients makes model optimization more challenging. To address these challenges, we propose PD-ADPVFL (Performance-Driven Adaptive Differential Privacy Vertical Federated Learning), which introduces minimal clipping strategies and threshold mechanisms. Combined with three different performance evaluation methods, it selectively preserves certain model iteration states and dynamically adjusts client privacy budgets, effectively mitigating performance degradation caused by differential privacy mechanisms and significantly improving model accuracy. Additionally, by adaptively adjusting client model iteration counts based on feature distributions, it effectively addresses convergence difficulties caused by non-IID data. We provide privacy analysis of the proposed algorithm, proving its reliability and privacy guarantees. Experimental results on multiple public datasets demonstrate that our algorithm achieves high model accuracy while effectively protecting data privacy, and shows strong resilience to non-IID features.