An Integrated AI Approach for Real-Time Traffic Forecasting and Congestion Management in Intelligent IoT-Based Networks

Abstract

Traffic congestion remains an unresolved problem in today's urban transportation, particularly in smart infrastructure contexts with the requirement of accurate and timely traffic forecasting to augment realistic congestion improvement. Traditional forecast models such as Auto Regressive Integrated Moving Average (ARIMA), SVR, or shallow neural networks do not grasp the non-linear, spatiotemporal dynamics of real- time traffic and hence lead to inferior route decisions and increased congestion. In this study, a new hybrid artificial intelligence (AI) approach combining Bidirectional Encoder Representations from Transformers (BERT) and Convolutional Long Short-Term Memory (ConvLSTM) is presented to improve the accuracy of traffic prediction and enable proactive congestion control in intelligent Internet of Things (IoT) networks. The architecture has four steps: input representation via real-time sensor data, contextual embedding via BERT to capture temporal and semantic patterns, spatiotemporal modeling via ConvLSTM, and decision generation via reinforcement-based routing adaptation. The model was trained and evaluated using the METR-LA dataset, a real-world spatiotemporal traffic dataset for Los Angeles, and demonstrated improved performance compared to baseline models. Exactly, the proposed model registered a Mean Absolute Error (MAE) of 2.13, a Root Mean Square Error (RMSE) of 3.54, and an R² Score of 0.91, beating LSTM-only, CNN-LSTM, and Transformer-only models. The outcomes confirm the effectiveness of the integrated deep contextual learning and spatiotemporal memory in predicting traffic. The proposed system offers a scalable and sensible paradigm for future traffic management in smart cities.