CNN, LSTM, and Hybrid Models for IoT-Based Air Quality Prediction in Smart Cities : A Review

Abstract

Accurate air quality prediction remains a significant challenge in smart cities due to the noisy, drifting, and highly nonlinear data generated by IoT sensor networks. Traditional statistical models struggle to capture these spatiotemporal complexities. This review systematically evaluates the comparative performance of Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and hybrid CNN–LSTM models for air quality forecasting using IoT sensor data. A systematic literature review was conducted on peer-reviewed studies published between 2020 and 2025, retrieved from Scopus, IEEE Xplore, ScienceDirect, and SpringerLink, using predefined inclusion criteria. Model performance was synthesized using reported RMSE, MAE, and MAPE metrics. The analysis reveals that LSTM models perform well for single-site temporal forecasting, while CNNs are effective in extracting spatial features from multi-sensor data. Hybrid CNN–LSTM models consistently achieve lower prediction errors in multivariate, multi-site, and multi-step forecasting tasks, demonstrating superior spatiotemporal learning capability. However, these models face limitations related to computational cost, sensor density requirements, and real-time deployment. Overall, the review provides practical guidance for selecting deep learning models for IoT-enabled urban air quality management.