A Machine Learning Framework for Short-Term Traffic Flow Prediction and Adaptive Traffic Signal Optimization Using Internet of Things (IoT) Sensor Data

Abstract

Traffic demand at urban intersections fluctuates drastically across different signal phases, while conventional fixed-time signal control strategies fail to dynamically adjust green light durations in response to short-term traffic variations. This study addresses two core research questions: whether higher single-point traffic prediction accuracy can guarantee superior signal control performance, and how phase-level prediction uncertainty can be integrated into adaptive green time optimization. One-minute aggregated traffic and signal operation data collected from an urban signalized intersection in Zurich are adopted for model training and verification. Lagged traffic volumes, rolling statistical features, temporal indicators, and signal control parameters are constructed as model input variables. Under consistent chronological train-validation-test data partitioning, four mainstream prediction approaches are comprehensively compared: XGBoost, BiLSTM, rolling-window ARIMA, and the proposed Hurdle-Based Uncertainty-Aware Multi-Task Adaptive Signal Control model (HUMAS-TSC). The proposed HUMAS-TSC model simultaneously predicts traffic demand for signal groups sg2, sg4 and sg6. It leverages hurdle modelling to handle sparse zero-flow observations and incorporates prediction uncertainty quantification to support stable green time allocation. Experimental results reveal that XGBoost achieves the optimal single-point prediction accuracy, with a Mean Absolute Error (MAE) of 0.2925 and a Root Mean Squared Error (RMSE) of 0.3885. In contrast, HUMAS-TSC delivers the best signal control performance: it cuts the average vehicle delay by 25.804% relative to fixed-time control, reduces the high saturation ratio to merely 0.107%, and completely eliminates saturation cap incidents. The core conclusion indicates that single-point prediction accuracy alone cannot determine the overall efficiency of signal control systems, instead, uncertainty-aware phase-level optimization can simultaneously enhance traffic operation efficiency and operational stability. Theoretically, this research clarifies the intrinsic correlation among single-point prediction precision, phase-level predictive uncertainty, and downstream signal control performance. From an engineering perspective, it provides an uncertainty-driven adaptive signal timing solution for urban intersections, which balances traffic passing efficiency, phase service reliability and system operational stability.