Clustering Federated Learning (CFL) is a distributed learning paradigm that groups clients with similar characteristics to perform model training within clusters, addressing the performance decline of Federated Learning (FL) models due to data heterogeneity. However, existing CFL methods still face several challenges, including the diversification of data heterogeneity, the dynamic adaptability of clustering algorithms, and client selection. To address these challenges, this paper proposes an Adaptive Clustering Federated Learning Framework with Loss-Driven Biased Client Selection (LDBCS-ACFL). The framework comprehensively considers the complexity of diversification of data heterogeneity and dynamic client changes. It proposes an adaptive clustering algorithm specifically designed for the diversification of data heterogeneity, and utilizes a dual-clustering strategy to achieve fine-grained client clustering. Additionally, a clustering update strategy based on Jensen-Shannon (JS) divergence is proposed to respond to dynamic client changes. To further enhance model stability and training accuracy, the framework proposes a loss-driven biased client selection method, which dynamically adjusts the client participation ratio in training based on cluster loss. Experiments conducted on four public datasets under different data heterogeneity scenarios demonstrate that the LDBCS-ACFL framework outperforms baseline algorithms in terms of model performance.