Improving Traffic-Flow Prediction Using Proximity to Urban Features and Public Space

Accurate traffic prediction and planning help alleviate congestion and facilitate sustainable traffic management through short-term traffic controls and long-term infrastructure design. While recent uptake on Machine Learning (ML) approaches helps refine our ability to predict the traffic flow, proximity to landmarks and public spaces are often overlooked, thus undermining the impact of location-specific traffic patterns. Using traffic-flow estimates from London, this study incorporates the proximity to urban features approximated with Kernel Density Estimation (KDE) and compares the performance of models with and without such features. They are also tested using classic spatial/non-spatial regression models and ML-based regression models. Results suggest that adding urban features considerably improves the performance of the ML models (Fine tree yielding R2 = 0.94, RMSE = 0.129, and MAE = 0.069), which compares favourably against the best performing non-ML model (the spatial error model returning R2 = 0.448, RMSE = 0.358, and MAE = 0.280). Sensitivity of the KDE is tested across different bandwidths for including urban features. The ML classification approach was also applied for estimating the traffic density and achieved high accuracy, with Fine KNN achieving 98.7%. They offer a robust framework for accurate traffic projection at specific locations, thus enabling road infrastructure designs that cater to the specific needs of the local situations.