Machine learning-based short-term traffic speed prediction using commercial traffic information data

Abstract

Traffic congestion in Bangkok’s central business district (CBD) is concentrated at closely spaced signalized intersections, making accurate short-term traffic speed prediction essential for effective traffic management. This study evaluates short-term traffic speed forecasting for 28 inbound road segments (nodes) at eight interconnected signalized intersections along Rama IV Road and Sathon Road using commercial traffic data aggregated as 5-minute median weekday speeds collected in March 2026. Nine prediction models were compared: Support Vector Regression (SVR), k-Nearest Neighbors (kNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), LSTM-RNN, LSTM-DNN, and the Spatial-Temporal Graph Convolutional Neural Network (STGCNN), which incorporates a topology-based adjacency matrix. Model performance was evaluated for 30- and 60-minute forecasting horizons using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and the F1 score for traffic congestion detection. The results indicate that STGCNN achieved the best overall performance. At the 30-minute forecasting horizon, it attained an MAE of 3.35 km/h, a MAPE of 27.57%, and an F1 score of 0.844, representing an approximately 8% lower MAE than the RNN model. Temporal deep learning models consistently outperformed conventional machine learning models, while spatial information further improved prediction accuracy. Aggregating traffic data into 15-minute intervals reduced the STGCNN's MAE to 2.94 km/h. Congestion propagation analysis showed that morning congestion typically originated along Rama IV Road between Henri Dunant and Si Lom intersections. Si Lom Intersection exhibited the earliest median congestion onset time, at approximately 6:10 a.m., before congestion propagated to adjacent intersections.