An Expressway Traffic Flow Prediction Method Considering Inter-Lane Differences and Upstream and Downstream Cross-Section Correlations

In the existing traffic flow prediction studies, the differences of traffic flow between different lanes within a cross-section road and the correlation of traffic flow between upstream and downstream cross-sections are not fully considered. This research develops a framework for expressway traffic flow prediction by integrating principal component analysis (PCA) and long short-term memory (LSTM) neural networks, which satisfies the real-time and accuracy requirements of intelligent connected technologies. The traffic flow data of urban expressways are collected, and the fast Fourier transform (FFT) method is applied for data preprocessing to improve the predictability of the original data. PCA is used to fuse the features of lateral and longitudinal traffic flow between lanes, establishing correlated traffic flow data between lanes to reduce data dimensionality. The correlated data is then integrated into the LSTM model for lane-level traffic flow prediction, and the prediction results aggregated to obtain the cross-section traffic flow estimates. The proposed method is validated using traffic data from expressway checkpoint detectors along the Third Ring Road in Wuhan. The results demonstrate that the model considering inter-lane differences and upstream and downstream cross-section correlations can improve the accuracy of cross-section traffic flow prediction. Compared to prediction results that only consider temporal features, the mean absolute error, root mean square error, and mean absolute percentage error can be reduced by 6.66%, 6.23%, and 17.51%, respectively. When compared with models that account for either upstream-downstream correlations or inter-lane differences alone, the proposed model demonstrates superior performance, with reductions in mean absolute error, root mean square error, and mean absolute percentage error ranging from a minimum of 1.53% to a maximum of 12.88%. Additionally, the proposed model demonstrates higher prediction accuracy than support vector regression (SVR) and random forest (RF) algorithms. In time-segmented predictions, the model performs particularly well during the evening peak and off-peak hours.