Urban traffic flow prediction is a critical task in intelligent transportation systems (ITS), yet it faces challenges such as complex spatiotemporal dependency modeling and incomplete observation data. To address the widespread issue of point-wise missing data and the limited robustness of existing methods, this paper proposes a conditional diffusion-based traffic flow prediction framework, referred to as RDPFlow. The proposed method integrates a missing-aware masking mechanism with a heterogeneous external condition guidance strategy, enabling unified modeling of both traffic flow prediction and data imputation. By explicitly indicating missing regions, RDPFlow guides the model to distinguish between actual zero values and unobserved entries, and introduces external factors such as holidays and weather conditions for semantic enhancement, thereby improving the model’s adaptability to realistic missing patterns. Experimental results on TaxiBJ dataset show that RDPFlow consistently outperforms State-of-The-Art (SOTA) methods under both complete and missing data scenarios, achieving lower prediction errors and stronger generalization robustness.