LSTM Model and Data Segmentation Collaboration A Deep Learning Strategy for Traffic Flow Anomaly Detection Using SUMO Simulated Data

To address the escalating challenge of traffic congestion and the associated difficulties in obtaining specific congestion data sets, this paper explores the use of deep learning strategies to determine traffic flow conditions using traffic flow data generated by urban traffic simulation software (SUMO). This paper describes an approach that combines LSTM and data segmentation to convert SUMO data into a general model format. This conversion is compatible with LSTM-based models originally trained on SUMO traffic flows. The paper confirms through rigorous quantitative analyses that the use of SUMO-generated traffic flows in the field of traffic flow prediction has the advantages of data purity, ease of collection and ubiquity over the use of real traffic flows and verifies that there can be the same correspondence mapping between the real-world dataset and the SUMO dataset. We have also investigated the robustness and generality of the model, demonstrating its ability to adapt to different traffic scenarios, and its ability to identify anomalous traffic flows by 89.5% under experimental conditions, which is 8.1% higher than the model based on the NGSIM dataset (81.4%).This work highlights the promise of using simulation software to generate datasets, combined with deep learning approaches, to provide practical and scalable solutions to support data-driven traffic management decisions. In general, this paper propose the use of simulation software with data segmentation mapping to provide a solution for the research and analysis of regional traffic in circumstances where unavailability and realistic pristine datasets are not obtainable.