Customized LSTM Models with Bi-Directional Architectures

This research investigates the efficacy of bi-directional Long Short-Term Memory (LSTM) networks in traffic flow prediction (TFP) for urban transportation systems. It addresses the critical need for precise TFP models to alleviate traffic congestion and optimize transportation infrastructure efficiency. By leveraging LSTM networks with bi-directional architectures, capable of capturing temporal dependencies in sequential data, the study aims to enhance prediction accuracy and robustness. Using real-world traffic data from the Metro Interstate Traffic Volume Dataset, the research develops and evaluates multiple predictive models, including dense neural networks (DNNs), convolutional neural networks (CNNs), and traditional recurrent neural networks (RNNs), alongside customized LSTM models. The findings reveal that the bi-directional LSTM models outperform baseline models, demonstrating substantial improvements in prediction accuracy. Comparative analysis highlights the novel contribution of bi-directional LSTM architectures in enhancing TFP accuracy, particularly in dynamic urban traffic environments. The experimental results underscore the potential of advanced deep learning techniques in developing data-driven solutions for urban transportation management, holding implications for urban planners and policymakers aiming to alleviate traffic congestion and enhance mobility in urban areas.