이산 웨이블릿 변환을 이용한 LSTM 기반 PM 2.5 시계열 예측 지연 완화 - 서울 중구 사례 연구 -

Abstract

Accurate forecasting of fine particulate matter (PM 2.5) concentrations is critical for air quality management and public health. Long Short-Term Memory (LSTM) networks have demonstrated strong capabilities in modeling nonlinear and temporal dependencies in time series data. However, a common challenge in LSTMbased PM 2.5 forecasting is the presence of prediction lag, where model outputs tend to delay behind rapid changes in actual pollutant levels, reducing forecast responsiveness. This study proposes a hybrid approach combining Discrete Wavelet Transform (DWT) with LSTM to mitigate the lag effect in PM 2.5 time series prediction. The DWT decomposes the original PM 2.5 data into multiple frequency components, allowing the LSTM models to capture both short-term fluctuations and long-term trends more effectively. Based on observed data from Jung-gu, Seoul, the proposed DWT-LSTM model is evaluated against a baseline LSTM model. Experimental results demonstrate that the hybrid model significantly reduces prediction lag and improves forecasting accuracy across multiple evaluation metrics, including RMSE, MAE, and correlation coefficients. These findings suggest that integrating wavelet-based preprocessing with deep learning architectures enhances the temporal alignment and precision of air quality forecasts, contributing to more timely and informed decision-making.