Accurate hourly AQI prediction using temporal CNN-LSTM-MHA+GRU: A case study of seasonal variations and pollution extremes in Visakhapatnam, India

This research presents an innovative hybrid deep learning architecture, CNN-LSTM-MHA+GRU, aimed at precise hourly predictions of the Air Quality Index (AQI) utilizing dataset derived from Visakhapatnam, India (October 2022–October 2024). The framework amalgamates one-dimensional Convolutional Neural...

Full description

Saved in:
Bibliographic Details
Main Authors: Sreenivasulu T, Mokesh Rayalu G
Format: Article
Language:English
Published: Elsevier 2025-09-01
Series:Results in Engineering
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2590123025023758
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This research presents an innovative hybrid deep learning architecture, CNN-LSTM-MHA+GRU, aimed at precise hourly predictions of the Air Quality Index (AQI) utilizing dataset derived from Visakhapatnam, India (October 2022–October 2024). The framework amalgamates one-dimensional Convolutional Neural Networks (1D CNN) for the extraction of short-term patterns, Long Short-Term Memory (LSTM) integrated with Multi-Head Attention (MHA) to encapsulate long-term dependencies, and Gated Recurrent Unit (GRU) for the refinement of residual errors. The model, trained on a dataset comprising 14,073 records and optimized through Bayesian parameter tuning, demonstrated robust performance on the test dataset (R² = 0.9757, RMSE = 6.29, MAPE = 7.07 %). The proposed model consistently surpassed six benchmark models by an impressive margin of 6.6–15.4 % in terms of R². The application of five-fold cross-validation substantiated the model’s stability (mean R² = 0.9551 ± 0.0052). Statistical analyses, including MANOVA, ANOVA, and t-tests, uncovered seasonal pollution patterns, notably peaks during the winter and reductions during the monsoon. The model exhibited commendable generalizability when applied to the cities of Delhi and Mumbai (R² > 0.97) without necessitating retraining, and it showcased real-time applicability (0.08s/sample) even amidst high-AQI occurrences (MAPE = 4.58 % for AQI > 150). Interpretability driven by SHAP reinforces the significance of features, thereby rendering the model beneficial for the formulation of targeted emission control strategies. This framework presents a scalable, interpretable, and transferable approach for urban air quality forecasting and informed decision-making.
ISSN:2590-1230