Hybrid Deep Learning Architecture for Multivariate Time Series Forecasting: A Study on Urban Traffic Volume Prediction

Abstract

Time Series Forecasting (TSF) is a necessity in many areas, such as finance, meteorology, and Intelligent Transportation Systems (ITS). Despite that, conventional statistical and Machine Learning (ML) methods have been commonly used in these applications, but they regularly do not manage to seize the intricacy of the time-ordered dependencies and the inter-variable relationships, mainly in cases of Multivariate Time Series (MTS). This proposed approach evaluates a hybrid Deep Learning (DL) model that combines Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and an Attention Mechanism (AM) to improve forecasting accuracy. A comparative analysis is conducted, evaluating the hybrid architecture on both regression and classification frameworks. The proposed model is tested on the Metro Interstate Traffic Volume (MITV) dataset, which contains hourly traffic data along with weather and time-related features. The results show the model’s efficacy as a regression tool, achieving an R2 of 0.8734. However, the primary contribution is the demonstration of its superiority as a classifier. When framed as a multi-class classification problem to predict five distinct congestion levels, the same architecture achieves 89.2% accuracy. Overall, the technique offers a scalable and interpretable deep learning framework for intelligent traffic monitoring.