Bi-LSTM-Based Traffic Flow Prediction and Adaptive Signal Control via Gap-Statistic K-means++ Cross-Partitioning

Abstract

Accurate traffic state prediction and coordinated signal control are essential for improving urban traffic efficiency. This study proposes a hybrid framework that combines a bidirectional long short-term memory (Bi-LSTM) network for short-term traffic prediction with an enhanced clustering method for intersection coordination. The clustering module uses K-means++ initialization and the Gap Statistic to dynamically determine the optimal number of regions, enabling similar traffic patterns to be grouped adaptively. The framework is evaluated on two public datasets, PEMS-BAY and CityFlow. The Bi-LSTM model achieved a prediction accuracy of 97.5 percent and a root mean square error of 0.12 on the PEMS-BAY dataset, outperforming baseline methods, including standard LSTM and CNN-LSTM. The clustering module achieved a silhouette coefficient of 0.47 and a Calinski-Harabasz index of 1,154.8. These results indicated strong intra-cluster cohesion and inter-cluster separation. Furthermore, the accuracy of signal control, which was defined as the proportion of intersections that received the correct timing adjustments based on predicted flow levels, was 91.2 percent in the coordinated control simulation. The proposed method outperforms traditional fixed-control and non-coordinated strategies by reducing prediction error, lowering control latency, and improving network-wide adaptability. These results demonstrate that the integrated approach improves the accuracy of traffic forecasting and the effectiveness of region-based signal coordination. This makes it a robust solution for real-time urban traffic management.