Forecasting particulate matter concentration using deep learning and weather data

Abstract

Particulate matter (PM2.5) concentration is a critical air quality indicator with severe health impacts. Predicting its dynamics is challenging due to complex interactions with meteorological variables. This study evaluates long short-term memory (LSTM) and multilayer perceptron (MLP) networks to identify the most accurate forecasting architecture. We integrated continuous data from automatic weather stations (AWS) and air quality sensors. Data preprocessing involved temporal synchronization, outlier removal, and multivariate imputation by chained equations (MICE). Key input features past particulate matter, temperature, and relative humidity were selected via Pearson correlation and trained using a 24-hour sliding window. Results demonstrate that the single LSTM model outperforms both the MLP and hybrid architectures. The LSTM achieved the lowest root mean square error (RMSE) of 6.862 µg/m³ and a mean absolute percentage error (MAPE) of 37.38%. Although the MLP yielded a slightly higher coefficient of determination (0.810), it exhibited significant magnitude bias by treating sequential data as static features. Ultimately, this research provides a robust, data-driven framework for real-time air quality early warning systems in tropical regions.