Weather-Fused CNN-LSTM and GRU Ensemble Learning for Explainable Air Quality Index Forecasting Using the CPCB Methodology

Abstract

Air pollution is now one of the biggest environmental and public-health problems facing Indian cities. The Central Pollution Control Board (CPCB) has repeatedly reported that most monitored cities exceed the National Ambient Air Quality Standards for particulate matter, which makes reliable Air Quality Index (AQI) forecasting genuinely useful for public health advisories, traffic control, and simply deciding whether it is safe to step outside. An earlier version of this system predicted AQI from historical pollutant readings alone. It had no weather context, could not respond in real time, and gave no indication of which factor was actually driving a given forecast. This paper builds on that system by fusing raw pollutant concentrations (PM2.5, PM10, NO2, SO2, CO, O3) with weather variables (temperature, humidity, wind speed, atmospheric pressure) and cyclical time-of-day and day-of-week features. We also implement the official CPCB 2014 sub-index methodology ourselves, breakpoint table and max-operator rule included, so that the AQI values used for training and testing are computed rather than assumed from an existing label column. Forecasting is done with a two-branch ensemble: a CNN-LSTM hybrid that first extracts short, local patterns with 1-D convolutions and then models longer sequences with stacked LSTM layers, paired with a separately trained GRU network that is lighter and acts as a second opinion. The two outputs are simply averaged, which cuts down variance compared to relying on one model alone. On top of this, a permutation-importance step tells us which features the ensemble actually leans on when it makes a prediction. Testing on an hourly dataset with a proper chronological (non-shuffled) train/validation/test split, the ensemble reached a Mean Absolute Error of 16.89 AQI points, an RMSE of 23.19, an R² of 0.799, a MAPE of 13.19%, and, arguably the number that matters most in practice, a CPCB category accuracy of 82.06% — meaning that more than four out of five hourly forecasts land in the correct health-advisory bracket. Taken together, these results show that a fairly modest set of changes — weather fusion, a two-model ensemble, and a simple explainability layer — noticeably improves both the accuracy and the usefulness of a pollutant-only baseline.

Introduction

This study proposes an AI-based Air Quality Index (AQI) forecasting system that improves upon traditional AQI prediction by combining pollutant data, weather information, and explainable deep learning. While the Central Pollution Control Board (CPCB) AQI provides a standardized measure of current air quality, it does not predict future conditions needed for timely public health advisories and pollution control measures.

The existing system had several limitations: it relied only on historical pollutant concentrations, ignored weather factors such as wind speed and humidity, lacked explainability, and used a single recurrent model that struggled with sudden pollution spikes. To overcome these issues, the proposed system integrates pollutant concentrations, weather observations, and cyclical time features (hour and day) into a two-branch CNN-LSTM and GRU ensemble model. The system also implements the CPCB 2014 AQI calculation from scratch to ensure accurate and verifiable ground-truth labels and incorporates permutation importance to explain which features influence each prediction.

Objectives

The study aims to:

• Integrate pollutant, weather, and time-based features for improved AQI prediction.
• Implement CPCB's AQI computation independently.
• Develop a CNN-LSTM and GRU ensemble for one-hour-ahead AQI forecasting.
• Provide explainable predictions using feature importance analysis.
• Evaluate performance using MAE, RMSE, R², MAPE, and CPCB category accuracy.
• Deploy the model through a Streamlit dashboard for real-time use.

Literature Review

Previous AQI forecasting methods include:

• Statistical models (ARIMA, SARIMA): simple and interpretable but weak for nonlinear pollution patterns.
• Machine learning models (Random Forest, Decision Trees, SVR): more accurate but require manual feature engineering.
• Deep learning models (LSTM, GRU): effectively capture temporal patterns.
• Hybrid CNN-RNN models: combine local pattern extraction with long-term sequence learning and generally outperform earlier approaches.
• Recent attention-based hybrid models achieve very high accuracy but require large datasets and complex tuning. Most existing models, however, remain black boxes with limited interpretability.

Proposed Methodology

The system operates in four stages:

1. Preprocessing: Cleans data, scales features, and creates 24-hour sliding windows.
2. AQI Ground Truth: Calculates AQI using CPCB 2014 sub-index rules rather than relying on existing labels.
3. Ensemble Forecasting: Uses parallel CNN-LSTM and GRU models whose predictions are averaged for improved accuracy and robustness.
4. Explainability: Applies permutation importance to identify the pollutants, weather variables, or time features most responsible for each forecast.

The dataset includes:

• Pollutants: PM2.5, PM10, NO?, SO?, CO, and O?.
• Weather variables: Temperature, humidity, wind speed, and atmospheric pressure.
• Time features: Cyclical hour-of-day and day-of-week encodings.

Key Contributions

• Integrates weather data with pollutant measurements for better forecasting.
• Uses a CNN-LSTM + GRU ensemble to improve prediction accuracy and handle sudden pollution spikes.
• Implements a verified CPCB AQI computation to ensure reliable training labels.
• Adds feature-level explainability using permutation importance, making predictions more transparent and trustworthy.
• Provides a real-time Streamlit dashboard for practical deployment.

Conclusion

What we set out to do here was fix three specific gaps in the earlier version of this system — no weather context, no explainability, and poor handling of spikes — and we think the changes described in this paper do that reasonably well. Weather and cyclical time features are now fused in at the input; the CPCB 2014 sub-index rules are implemented and checked against CPCB\'s own worked example rather than assumed; the CNN-LSTM and GRU branches, combined by simple averaging, visibly smooth the response to spikes without losing the daily cycle; and the permutation-importance step correctly surfaces the dominant drivers behind a forecast, matching what other researchers have found on real Indian air-quality data. On the held-out test set, the ensemble reached an MAE of 16.89, RMSE of 23.19, R² of 0.799, MAPE of 13.19%, and CPCB category accuracy of 82.06%. If there\'s one takeaway, it\'s that you don\'t need an enormous architecture to meaningfully improve a forecasting system — weather fusion, a two-model ensemble, and a basic explainability layer got us a system that is more accurate in the categories that matter, more stable under noise, and considerably more transparent about why it predicts what it predicts

References

[1] \"Air quality index forecast in Beijing based on CNN-LSTM multi-model,\" ScienceDirect, 2022. [2] \"An air quality index prediction model based on CNN-ILSTM,\" Scientific Reports, Nature Publishing Group, 2022. [3] \"Geographically Aware Air Quality Prediction Through CNN-LSTM-KAN Hybrid Modeling with Climatic and Topographic Differentiation,\" MDPI Atmosphere, 2025. [4] \"Air-quality prediction based on the ARIMA-CNN-LSTM combination model optimized by dung beetle optimizer,\" Scientific Reports, Nature Publishing Group, 2023. [5] \"Spatially resolved air quality index prediction in megacities with a CNN-Bi-LSTM hybrid framework,\" ScienceDirect, 2024. [6] S. Hochreiter and J. Schmidhuber, \"Long Short-Term Memory,\" Neural Computation, vol. 9, no. 8, pp. 1735-1780, 1997. [7] K. Cho, B. van Merrienboer, C. Gulcehre, et al., \"Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation,\" Proc. EMNLP, 2014. [8] V. S. Bandari, \"Accurate hourly AQI prediction using temporal CNN-LSTM-MHA+GRU: A case study of seasonal variations and pollution extremes in Visakhapatnam, India,\" ScienceDirect, 2025. [9] G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time Series Analysis: Forecasting and Control, 5th ed., Wiley, 2015. [10] L. Breiman, \"Random Forests,\" Machine Learning, vol. 45, no. 1, pp. 5-32, 2001. [11] F. Pedregosa, G. Varoquaux, A. Gramfort, et al., \"Scikit-learn: Machine Learning in Python,\" Journal of Machine Learning Research, vol. 12, pp. 2825-2830, 2011. [12] M. Abadi, P. Barham, J. Chen, et al., \"TensorFlow: A System for Large-Scale Machine Learning,\" Proc. 12th USENIX OSDI, 2016. [13] M. T. Ribeiro, S. Singh, and C. Guestrin, \"\'Why Should I Trust You?\': Explaining the Predictions of Any Classifier,\" Proc. 22nd ACM SIGKDD, 2016. [14] Central Pollution Control Board (CPCB), \"National Air Quality Index,\" Ministry of Environment, Forest and Climate Change, Government of India, 2014.

Copyright

Copyright © 2026 M. Ankitha, Pravitha R. Prasad. 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.