PM_2.5 Forecast in Korea using the Long Short-Term Memory (LSTM) Model

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• South Korea's Ministry of Environment forecasts particulate matter (PM2.5) concentrations using the AirKorea system.
• Current forecasts combine the Community Multiscale Air Quality (CMAQ) model, artificial intelligence (AI), and human forecasters.
• The subjective nature of current PM2.5 grade designation warrants an evaluation of objective forecasting methods.

Purpose of the Study:

• To evaluate the forecast skills of the Long Short-Term Memory (LSTM) algorithm for PM2.5 concentrations in Korea.
• To compare LSTM forecasts against CMAQ-only and the operational AirKorea forecasts using 2019 observational data.
• To assess the potential of AI models to enhance PM2.5 forecasting objectivity and accuracy.

Main Methods:

• Utilized observational data from 2019 for PM2.5 concentrations across 19 districts in Korea.
• Compared forecast performance metrics between CMAQ-only, LSTM, and AirKorea systems.
• Analyzed forecast skill levels and false alarm rates for different pollution grades (low, moderate, high, very high).

Main Results:

• LSTM and AirKorea forecasts demonstrated higher one-day PM2.5 forecast skills (72-79% and 73-80%) compared to CMAQ-only (39-70%).
• AI forecasts, particularly LSTM, showed comparable skills to human forecasters at AirKorea.
• CMAQ-only forecasts exhibited significantly higher false alarm rates (up to 86%) than LSTM and AirKorea (up to 58%).

Conclusions:

• The Long Short-Term Memory (LSTM) model, when applied to CMAQ forecasts, achieves forecast skill levels comparable to the operational AirKorea system.
• AI models offer a more objective approach to enhancing PM2.5 forecast accuracy in Korea.
• Implementing appropriate AI models can substantially improve the reliability and objectivity of air quality predictions.