Machine Learning for Air Quality Prediction : Comparative Analysis with Web Dashboard Implementation
Lertudomkitpaisan, Phornpailin (2026)
2026
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202604156483
https://urn.fi/URN:NBN:fi:amk-202604156483
Tiivistelmä
This research introduces a machine learning system which forecasts air quality in Delhi, India through its development and testing. The persistent presence of high pollution levels in Delhi makes precise PM2.5 forecasting tools essential for public health education.
The researcher tested two machine learning models through the training process which used archived air quality records from monitoring stations in Delhi between 2021 and 2024. The dataset required thorough preprocessing which included removing outliers and creating new features through lag var-iables and rolling averages and establishing a sequence of train-validation-test splits to safeguard against data leakage.
XGBoost achieved superior predictive performance with MAE of 5.81 μg/m³, RMSE of 11.04 μg/m³, and R² of 0.950 on the test set, outperforming LSTM which produced MAE of 12.19 μg/m³, RMSE of 15.92 μg/m³, and R² of 0.896. Feature importance analysis showed that the 7-day rolling average feature and previous day value feature together made up about 80% of model predictions which used engineered temporal features.
The XGBoost model operating system was connected to a Flask-React web dashboard which provides users with instant PM2.5 and Air Quality Index forecasts. The users enter current pollutant measurements to obtain AQI health categories which display color codes according to EPA standards. The dashboard presents data from 2021 to 2024 through historical trend visualization.
The researcher tested two machine learning models through the training process which used archived air quality records from monitoring stations in Delhi between 2021 and 2024. The dataset required thorough preprocessing which included removing outliers and creating new features through lag var-iables and rolling averages and establishing a sequence of train-validation-test splits to safeguard against data leakage.
XGBoost achieved superior predictive performance with MAE of 5.81 μg/m³, RMSE of 11.04 μg/m³, and R² of 0.950 on the test set, outperforming LSTM which produced MAE of 12.19 μg/m³, RMSE of 15.92 μg/m³, and R² of 0.896. Feature importance analysis showed that the 7-day rolling average feature and previous day value feature together made up about 80% of model predictions which used engineered temporal features.
The XGBoost model operating system was connected to a Flask-React web dashboard which provides users with instant PM2.5 and Air Quality Index forecasts. The users enter current pollutant measurements to obtain AQI health categories which display color codes according to EPA standards. The dashboard presents data from 2021 to 2024 through historical trend visualization.