关键词: 城市交通, 动态通行能力, 高斯过程回归, 自适应信号控制, 交通流理论, 双向长短记忆神经网络

Abstract: Traditional capacity estimation methods based on Fundamental Diagram (FD) theory are inadequate to capture the dynamic characteristics of adaptive signal control and the inherent instability of traffic flow states. To accurately estimate the capacity of adaptive signal control systems, this paper proposes a dynamic capacity estimation method based on Gaussian Process Regression (GPR). To address the complex traffic data generated by adaptive control, this study develops a heterogeneous multi-dimensional feature space, which encompasses the traffic state x(t) , signal control strategy Φ(t) ,and decision time series H(t) . A bidirectional Long Short-Term Memory (LSTM) neural network is then introduced to encode the temporal data and extract relevant features, and a multi-head attention mechanism is applied to identify critical decision points within the adaptive control strategy. A Gaussian Process (GP) model is then used to learn the functional mapping from multi- dimensional input features to traffic volume. By leveraging the nonparametric Bayesian properties of GPs and their inherent capacity for uncertainty quantification, the proposed model flexibly accommodates the complex nonlinear relationships between signal control strategies and traffic capacity, thereby overcoming the static modeling limitations of conventional FD theory under fixed signal timing plans. Simulation results demonstrate that the proposed model outperforms not only the conventional FD-based static capacity model and a static GP model, but also four additional baselines—Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), LSTM, and Transformer—achieving optimal performance across the evaluation metrics. The proposed GPR method with extended input features effectively reflects the impact of adaptive control decision changes on traffic capacity, with improved capacity prediction accuracy ( EMAPE is 3.96% , R2 is 0.951 0). The findings demonstrate the practical value of the proposed method for the performance evaluation and systematic optimization of adaptive signal control systems.

Key words: urban transportation, dynamic capacity, Gaussian Process regression, adaptive signal control, traffic flow theory; bidirectional Long Short-Term Memory (LSTM) network