Integrating GCN, BiLSTM, and Attention for Accurate Short-Term Traffic Forecasting on Urban Road Networks

Urban traffic congestion has become a global challenge, causing economic losses, environmental pollution, and reduced quality of life.Accurate short-term traffic flow prediction is essential for optimizing traffic management, improving travel efficiency, and supporting the development of intelligent transportation systems.When applied to large-scale traffic systems, conventional statistical models frequently fail to account for the highly nonlinear and interdependent spatio-temporal patterns.This paper presents a GCN-BiLSTM-Attention model for short-term traffic flow prediction.This model integrates graph convolution to capture spatial dependencies, bidirectional LSTM for temporal modeling, and an attention mechanism to enhance interpretability.Evaluated on the METR-LA dataset with 207 traffic sensors, the model achieves an MAE of 1.96 mph, RMSE of 4.25 mph, and ±10% accuracy of 91.58%.A comprehensive data preprocessing pipeline-including anomaly removal, imputation, and normalization-ensures high-quality input.Attention weight analysis shows a focus on recent time steps, aligning with traffic dynamics.Results demonstrate the model's effectiveness in learning complex spatiotemporal patterns and its potential for deployment in intelligent transportation systems.