Traffic flow prediction is essential for Internet of Vehicles (IoV) systems as it provides accurate historical and real-time data for both past and future scenarios, enhancing transportation efficiency, reducing congestion, and improving road safety. However, real-world traffic flow prediction remains challenging due to the uncertainty of noise data. Most current graph convolution-based methods depend on heavily preprocessed, smoothed data, which can obliterate critical features, thus hampering accurate predictions in noisy settings. To address this challenge, a novel denoising attention enhanced recurrent multigraph convolutional network (DA-RMN) is proposed for traffic flow prediction. DA-RMN mainly consists of the denoising diffusion gated attention fusion (DDGAF) module and the recurrent multigraph convolution (RMGC) module. The DDGAF module, which includes the denoising diffusion probabilistic model (DDPM) and gated attention fusion (GAF) modules, is engineered to mitigate noise in traffic flow data and enhance the extraction of essential trend features. The DDPM module is tasked with restoring data to a noise-free state, effectively filtering out distortions and inaccuracies. Furthermore, the GAF module focuses on amplifying the detection of essential trend features, ensuring that subtle but critical patterns are not overlooked. The RMGC module consists of a multigraph convolution network based on the gated recurrent unit, residual connection, and fully connected layer to enhance the multigraph spatial-temporal feature of the DDGAF module. Extensive experiments on five real-world datasets demonstrate that DA-RMN outperforms the state-of-the-art model.