Enhancing Air Quality Prediction Accuracy Using Hybrid Deep Learning

Abstract

With fine particulate matter (PM2.5) being one of the most dangerous pollutants affecting human health and ecological sustainability, air pollution has grown to be a major global environmental and public health concern. For efficient air quality monitoring, management, and early warning systems, PM2.5 concentrations must be predicted accurately and in real time. An intelligent hybrid deep learning framework for real time air quality prediction is presented in this paper, with a focus on PM2.5 periodicity analysis. The suggested system successfully captures the nonlinear connections and temporal dependencies found in meteorological and air quality datasets by integrating several deep learning techniques. The model can better comprehend daily, weekly, and seasonal fluctuations in pollutant behavior by including periodicity analysis to find recurrent patterns in PM2.5 concentration levels. By combining sophisticated neural network components, the hybrid architecture improves prediction accuracy and model reliability by extracting significant spatial-temporal characteristics. The predictive model is trained using historical air quality data and pertinent Meteorological parameters as input characteristics. According to experimental results, the suggested method performs noticeably better in terms of prediction accuracy, resilience, and flexibility than both standalone deep learning techniques and conventional machine learning models. The results demonstrate how using periodicity analysis improves forecasting performance and offers a dependable tool to assit environmental monitoring organizations and legislators in putting timely pollution control measure into place.

Introduction

Air pollution has become a major global environmental and public health issue, mainly due to rapid urbanization, industrialization, and increased vehicle emissions. Among pollutants, PM2.5 is particularly dangerous because its very small particles can penetrate deep into the lungs and bloodstream, causing serious health problems such as respiratory diseases, heart conditions, and premature death. Therefore, accurate monitoring and prediction of PM2.5 levels are essential for effective air quality management and early warning systems.

Traditional air quality prediction methods, including statistical and conventional machine learning models, often struggle to capture the complex nonlinear relationships and temporal patterns present in environmental data. Recent advancements in artificial intelligence and deep learning have improved prediction capabilities, but many existing models still fail to fully consider the periodic patterns of PM2.5 changes caused by daily activities and seasonal variations.

To address these limitations, the study proposes a hybrid deep learning framework that combines Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) along with periodicity analysis to improve real-time PM2.5 prediction accuracy. The model uses environmental and meteorological data such as temperature, humidity, wind speed, and pollutant concentrations collected from monitoring stations and public databases. The data undergo preprocessing, feature engineering, and periodicity analysis using methods like Fourier transformation and autocorrelation to identify daily, weekly, and seasonal pollution patterns.

The hybrid model captures both spatial and temporal relationships in air quality data and is trained using historical datasets. Its performance is evaluated using metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the R² score. Results show that the system successfully identifies pollution cycles and accurately predicts PM2.5 levels for 1-hour, 6-hour, and 24-hour forecasts.

Conclusion

The experimental findings demonstrate that the hybrid model performs better in terms of prediction accuracy, resilience, and real-time performance than conventional statistical and machine learning techniques. The model’s capacity to recognize recurrent pollution trends, such as daily, weekly, and seasonal cycles, is made possible by the use of periodicity analysis, which improves forecasting accuracy and lowers prediction error. Additionally, the framework effectively integrates historical pollution data and climatic variables, enabling the model to understand intricate correlations that affect PM2.5 concentration levels. When compared to traditional models, the hybrid model produces reduced prediction errors and greater correlation with observed air quality data, according to performance evaluation utilizing metrics like MAE,RMSE, and R2. The capacity of the suggested system to function in a real-time setting, which makes it appropriate for implementation in smart city air monitoring systems and environmental management platforms, is another significant contribution of this study. By offering quick and reliable air quality forecasts, the system can support early warning systems, help policymakers with pollution management measures, and contribute to public health protection.

References

[1] “Prediciton of PM2.5 ConcentrationBased onDeep Learning, Multi-ObjectiveOptimization, and Ensemble Forecast, ”Sustainability, vol. 16, no. 11, 2024,Z.Gao, X. Mo, and H. Li. [2] “Improving 3-day deterministic air pollution forecast using machine learning algsorithms,” Atmospheric Chemistry and Physics, 2024. Z. Zhang, C.Johansson, M.Engardt, M.Stafoggia, and X. Ma. [3] “Comparative Analysis of Multiple Deep Learning Models for Forecasting Monthly Ambient PM2.5 Concentrations, “Z. He and Q. Guo, Atmosphere, 2024. [4] CNN-RNN hybrid model for PM2.5 prediction, Atmosphere, 2025. This study demonstrates that prediction accuracy across several cities is increased when CNN and RNN models are combined. [5] AI for Cleaner Air: Deep Learning and Conventional Time-Series Methods for Predictive PM2.5 Modeling, Computer Modeling in Engineering & Sciences, 2025. [6] “E-STGCN: Extreme Spatiotemporal Graph Convolutional Networks for Air Quality Forecasting, “ M. Panja, T. Chakraborty, A. Biswas and S. Deb, 2024. In this work, spatial-temporalinteractions between monitoring stations are captured using graph convolution networks. [7] “PCDCNet: A Surrogate Model for Air Quality Forecasting with Physical-Chemical Dynamics and Constraints, “ S. Wang et al., 2025. For better forecasting, the model combines deep learning with physical Atmospheric processes. [8] “TopoFlow: Physics-Guided Neural Networks for High-Resolution Air Quality Prediction, “ A.Kheder et al., 2026. The approach improves PM2.5 forecasting accuracy by incorporating wind and terrain data into neural networks. [9] Cumulative Logit Models and Machine Learning Algorithms for Air Quality Prediction, Environemnt, Development, and Sustainability, 2025. [10] Fu.Xi-Air:Urban air quality forecasting with multimodal machine learning. arXivpreprint(2025)-multimodal hybrid model integrating emissions, meteorology, and pollutants.

Copyright

Copyright © 2026 P Poojitha, Dr. T. Senthil, Dr. R. Karunia Krishnapriya, Mr. Pandreti Praveen. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.