A Traffic Flow Data Quality Repair Model Based on Spatiotemporal Correlation

Abstract

To address the data quality issues caused by environmental changes and other factors, this paper proposes a method for repairing missing traffic flow data from loop detectors, leveraging the spatiotemporal characteristics of traffic flow. Then, a Bi-directional Long Short-Term Memory with an Improved Genetic Algorithm (IGA-Bi-LSTM) model and an improved Deep Forest (DF) traffic flow data imputation model are constructed. By combining the advantages of these two models, the improved DF model is used to extract spatiotemporal characteristics and impute sequential data to obtain temporal features. These features are coupled with spatiotemporal characteristics and input into the IGA-Bi-LSTM neural network to establish the Spatiotemporal Imputation Model (STIM), ultimately enhancing data quality. To verify the reliability of the results, the experimental data used the PORTAL public dataset and compared the performance of Historical Average (HA), Autoregressive Integrated Moving Average (ARIMA) model, Random Forest (RF), and Bi-directional Long Short-Term Memory with Genetic Algorithm (GA-Bi-LSTM) models. The results indicate that the STIM model has more advantages compared to other methods. Finally, traffic flow theory is used for validation, and the results confirm that the imputed traffic flow data are reliable, demonstrating the significant importance of this research for traffic flow data analysis.

Keywords

• Traffic data quality repair
• cylinder multi-granularity input
• improved genetic algorithm
• Bi-LSTM
• deep forest model