Optimizing hourly air quality index forecasting: a particle swarm optimization-enhanced hybrid approach combining convolutional and recurrent neural networks

Abstract

Air pollution is still a serious worldwide issue, and accurate air quality index (AQI) prediction is needed. This paper proposes a hybrid deep learning model integrating 1D convolutional neural networks (Conv1D) and long short-term memory (LSTM) networks, optimized with particle swarm optimization (PSO) to enhance AQI forecasting. The model was evaluated at six urban areas: Bandra, Thane, Mazgaon, Kurla, Nerul, and Malad, and compared with a single LSTM network. PSO adjusted hyperparameters like hidden units, batch size, epochs, and learning rate was used to improve predictive accuracy. The Conv1D+LSTM hybrid model drastically decreased RMSE by 49.19% (Bandra), 33.97% (Thane), 5.24% (Mazgaon), 20.52% (Kurla), 35.85% (Nerul), and 27.54% (Malad), and R² Score improvements up to 751.2%. Training logs indicated smoother convergence with loss decrease at faster rates compared to LSTM, showing better learning efficiency and generalization. By combining spatial and temporal feature extraction with automated hyperparameter tuning, this model captures sophisticated pollution patterns which increases the reliability of AQI prediction. Enhancements in the future can be adding regularization methods and more feature inputs to improve the accuracy.