Abstract: Intelligent Transportation Systems (ITS) are an essential component of smart cities, and traffic flow prediction, as a core technology, plays a vital role in enhancing urban operational efficiency and achieving sustainable development. To improve prediction accuracy, researchers have widely adopted deep learning methods based on attention mechanisms in recent years to model traffic data, aiming to more effectively capture the complex spatiotemporal dependencies. However, these models generally suffer from poor interpretability, which not only limits the optimization of model behavior by researchers but also affects the trust of traffic managers in the prediction results. To address this issue, this paper proposes TPMOVis (Traffic Prediction Model Optimization Visual Analysis System), a visual analysis system for attention-based traffic flow prediction models, designed to enhance model transparency, diagnostic capability, and credibility. The system not only supports intuitive visualization of traffic data with spatiotemporal attributes and attention weights but also integrates error analysis and model behavior diagnosis functions to help users identify prediction anomalies and model attention biases. Based on this, the system provides a variety of interactive methods, enabling users to explore and adjust the model's focus areas based on diagnostic results, thereby further improving prediction performance and credibility. Through three case studies involving experts and a user survey, the effectiveness of TPMOVis in enhancing model interpretability, diagnosability, and user trust has been verified.