Airborne particulate matter measurement and prediction with machine learning techniques

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Air quality is crucial for human health, necessitating accurate monitoring and forecasting.
• Industrial zones often feature extensive air pollution monitoring networks.
• Collected data can assess current conditions and predict future air quality trends.

Purpose of the Study:

• To evaluate machine learning methods for predicting particulate matter (PM2.5 and PM10) concentrations.
• To compare the performance of different algorithms across various time scales.

Main Methods:

• Utilized data from the Ecolumn monitoring station, measuring temperature, pressure, humidity, and PM levels (PM1.0, PM2.5, PM10).
• Developed and tested models using Decision Tree, Random Forest, Recurrent Neural Network, and Long Short-Term Memory (LSTM) algorithms.
• Experimented with varying hyperparameters, network architectures, and time scales (10 min, 1 h, 24 h).

Main Results:

• The Long Short-Term Memory (LSTM) algorithm achieved optimal results for short averaging times.
• Decision Tree and Random Forest models showed strong performance for long averaging times, with minimal accuracy loss.
• Performance varied across algorithms and time scales, indicating specific strengths for different forecasting needs.

Conclusions:

• LSTM is highly effective for short-term air quality prediction.
• Decision Trees and Random Forests offer robust alternatives for longer-term forecasting.
• The study provides recommendations for applying these machine learning methods in air quality management.