[Paper Review] Time-to-Green predictions for fully-actuated signal control systems with supervised learning

Alexander Genser, Michail Makridis+4 more|arXiv (Cornell University)|Aug 24, 2022

Traffic Prediction and Management Techniques54 references1 citations

TL;DR

This paper proposes a supervised machine learning framework to predict Time-to-Green (T2G) for fully-actuated traffic signals using aggregated loop detector and signal phase data. It demonstrates that Random Forest models outperform Linear Regression and LSTM in predicting the duration of the next red phase, achieving a mean absolute error (MAE) of 0.90–3.27 seconds on real-world Zurich data, making it suitable for practical speed-advisory and traffic optimization systems.

ABSTRACT

Recently, efforts have been made to standardize signal phase and timing (SPaT) messages. These messages contain signal phase timings of all signalized intersection approaches. This information can thus be used for efficient motion planning, resulting in more homogeneous traffic flows and uniform speed profiles. Despite efforts to provide robust predictions for semi-actuated signal control systems, predicting signal phase timings for fully-actuated controls remains challenging. This paper proposes a time series prediction framework using aggregated traffic signal and loop detector data. We utilize state-of-the-art machine learning models to predict future signal phases' duration. The performance of a Linear Regression (LR), a Random Forest (RF), and a Long-Short-Term-Memory (LSTM) neural network are assessed against a naive baseline model. Results based on an empirical data set from a fully-actuated signal control system in Zurich, Switzerland, show that machine learning models outperform conventional prediction methods. Furthermore, tree-based decision models such as the RF perform best with an accuracy that meets requirements for practical applications.

Motivation & Objective

• To improve the accuracy of Time-to-Green (T2G) predictions for fully-actuated signal control systems, which are complex and non-linear due to dynamic optimization and transit priority.
• To develop a generic, data-driven framework that does not require knowledge of the underlying signal control algorithm, enabling application in proprietary or unknown systems.
• To enhance SPaT message quality for vehicle motion planning by accurately forecasting the next red phase duration using real-time traffic signal and loop detector data.
• To incorporate physics-informed features from traffic flow theory to improve model performance, especially in multi-modal urban environments with public transport priority.
• To evaluate the feasibility of real-time T2G prediction for integration into speed-advisory and traffic-smoothing systems.

Proposed method

• The framework uses historical traffic signal and loop detector (LD) data from a fully-actuated intersection in Zurich, Switzerland, with public transport priority.
• Extensive feature engineering is performed using domain-specific concepts from traffic flow theory, such as vehicle density, queue length, and detection timing patterns.
• Three supervised learning models are trained: Linear Regression (LR), Random Forest (RF), and Long-Short-Term-Memory (LSTM) neural network, all regressing on the duration of the next red phase.
• A baseline 'no-change' model uses the last observed red phase duration to establish a performance benchmark.
• Model performance is evaluated using standard metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Explained Variance (EH), and Normalized Mean ratio (NM).
• The models are trained offline on two months of data and tested on unseen cycles to assess generalization and prediction accuracy.

Experimental results

Research questions

• RQ1 Can supervised machine learning models accurately predict the duration of the next red phase (Time-to-Green) in a fully-actuated signal control system with complex, dynamic behavior?
• RQ2 How do different machine learning models—Linear Regression, Random Forest, and LSTM—compare in performance when predicting T2G using real traffic signal and loop detector data?
• RQ3 To what extent do physics-informed features derived from traffic flow theory improve the predictive accuracy of T2G models in multi-modal urban intersections?
• RQ4 Can the proposed framework achieve sufficient accuracy for practical deployment in speed-advisory and traffic-smoothing systems, especially in the presence of public transport priority?
• RQ5 What are the limitations of current models in predicting T2G for signal phases with high variance, such as those serving trams?

Key findings

• The Random Forest (RF) model achieved the best performance, with a Mean Absolute Error (MAE) of 0.90 to 3.27 seconds across different signal phases, significantly outperforming the baseline and other models.
• The RF model demonstrated superior ability to capture non-linear patterns in signal phase durations, especially in complex scenarios involving public transport priority.
• The LSTM model showed limited improvement over the baseline, indicating that its sequential modeling capability was less effective than expected for this time series forecasting task.
• The baseline 'no-change' model performed poorly, with high errors, validating the need for advanced machine learning approaches in T2G prediction.
• The RF model struggled with high-variability T2G peaks (e.g., for tram-served phases), resulting in higher errors (MAE of 85.88 and 64.02 seconds) and low explained variance (below 2%), indicating a need for enhanced data or model adaptation.
• The study confirms that feature engineering based on traffic flow theory enhances model quality and enables effective prediction without prior knowledge of the signal control algorithm.