Forecasting PM2.5 Concentrations in Nakhon Sawan Province Using Machine Learning, Deep Learning, and Prophet-Based Hybrid Models

Abstract

Particulate matter with an aerodynamic diameter of less than 2.5 microns (PM2.5) in Nakhon Sawan Province is an important environmental and public health issue. PM2.5 concentrations vary seasonally and are influenced by several air pollution and meteorological factors. This study aimed to compare the performance of baseline models, machine learning models, deep learning models, and Prophet-based hybrid models for daily PM2.5 forecasting in Nakhon Sawan Province. The study used secondary data collected from January 2020 to December 2025, consisting of PM2.5 concentrations and ten air pollution and meteorological variables. The data were divided chronologically into a training set from 2020 to 2023, a validation set from 2024, and a test set from 2025. Forecasting performance was evaluated for 1-, 3-, and 7-day-ahead horizons by comparing Persistence, Moving Average, Prophet, LSTM, GRU, XGBoost, Prophet-LSTM, Prophet-GRU, and Prophet-XGBoost models. The models were assessed using RMSE, MAE, MAPE, and , together with the classification of days when PM2.5 concentrations exceeded 50 µg/m³. The results showed that GRU achieved the best performance for 1- and 3-day-ahead forecasting, with RMSE values of 12.2476 and 18.4001, respectively. In contrast, Prophet achieved the best performance for 7-day-ahead forecasting, with an RMSE of 20.5216. These findings indicate that the most suitable forecasting model may vary depending on the forecasting horizon. Comparing multiple groups of models provides a clearer understanding of the strengths and limitations of each approach. The findings can support the development of forecasting tools and local PM2.5 monitoring systems.