Particulate matter forecast and prediction in Curitiba using machine learning.

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Vehicle emissions are a major source of urban air pollution, particularly Particulate Matter (PM2.5).
• Understanding the interaction between emissions, meteorology, and PM2.5 is crucial for public health and policy.
• Existing methods for PM2.5 prediction and forecasting require enhancement.

Purpose of the Study:

• To analyze the relationship between vehicular emissions, meteorological variables, and PM2.5 concentrations.
• To develop and evaluate machine learning models for predicting and forecasting PM2.5.
• To provide insights for mitigating the impact of vehicle emissions in urban environments.

Main Methods:

• Utilized hourly and daily data on meteorological conditions, vehicle flow, and PM2.5 concentrations from Curitiba, Brazil (2020-2022).
• Employed Random Forest (RF) and Long Short-Term Memory (LSTM) neural networks for prediction and forecasting.
• Used Multiple Linear Regression (MLR) and naive estimation as baseline models.

Main Results:

• Random Forest achieved high prediction accuracy (99.37% daily), identifying planetary boundary layer height as a key factor.
• Long Short-Term Memory demonstrated superior forecasting accuracy (99.71% for 1-hour horizon).
• Both RF and LSTM models outperformed baseline MLR and naive methods.

Conclusions:

• Machine learning models significantly enhance the accuracy of PM2.5 prediction and forecasting.
• The study provides a foundation for understanding pollutant dispersion and informing urban air quality policies.
• Accurate forecasting can aid in developing strategies to mitigate health impacts from vehicle emissions.