Performance Evaluation of PM2.5 Forecasting Using SARIMAX and LSTM in the Korean Peninsula

Abstract

Air pollution, particularly fine particulate matter (PM2.5), poses significant environmental and public health challenges in South Korea. The National Institute of Environmental Research (NIER) currently relies on numerical models such as the Community Multiscale Air Quality (CMAQ) model for PM2.5 forecasting. However, these models exhibit inherent uncertainties due to limitations in emission inventories, meteorological inputs, and model frameworks. To address these challenges, this study evaluates and compares the forecasting performance of two alternative models: Long Short-Term Memory (LSTM), a deep learning model, and Seasonal Auto Regressive Integrated Moving Average with Exogenous Variables (SARIMAX), a statistical model. The performance evaluation was focused on Seoul, South Korea, and took place over different forecast lead times (D00–D02). The results indicate that for short-term forecasts (D00), SARIMAX outperformed LSTM in all statistical metrics, particularly in detecting high PM2.5 concentrations, with a 19.43% higher Probability of Detection (POD). However, SARIMAX exhibited a sharp performance decline in extended forecasts (D01–D02). In contrast, LSTM demonstrated relatively stable accuracy over longer lead times, effectively capturing complex PM2.5 concentration patterns, particularly during high-concentration episodes. These findings highlight the strengths and limitations of statistical and deep learning models. While SARIMAX excels in short-term forecasting with limited training data, LSTM proves advantageous for long-term forecasting, benefiting from its ability to learn complex temporal patterns from historical data. The results suggest that an integrated air quality forecasting system combining numerical, statistical, and machine learning approaches could enhance PM2.5 forecasting accuracy.