The Forecasting of PM2.5 Using a Hybrid Model Based on Wavelet Transform and an Improved Deep Learning Algorithm

TLDR

Haze from rising PM2.5 emissions has become a serious public health issue, and accurate prediction is essential for policymakers, yet the nonlinear nature of PM2.5 time series and unresolved LSTM gradient and wavelet selection challenges hinder reliable forecasting. This study proposes a hybrid WT‑SAE‑LSTM model to improve PM2.5 forecasting accuracy. The model decomposes PM2.5 series from six Chinese sites into low‑ and high‑frequency components via wavelet transform, predicts each component with a stacked autoencoder‑LSTM network, and reconstructs the full series from the predicted components. The hybrid model outperforms BP and other baselines, with optimal configurations identified for each site, demonstrating that the WT‑SAE‑LSTM approach can enhance PM2.5 prediction accuracy.

Abstract

In recent years, the haze has caused serious troubles to people's lives, with the continuous increase of PM2.5 emissions. The accurate prediction of PM2.5 is very crucial for policy makers to make predictive measures. Due to the nonlinearity of the PM2.5 time series, it is difficult to predict accurately. Despite some studies about PM2.5 being proposed, the problem of the LSTM (long short-term memory) gradient disappearance and random selection of wavelet orders and layers isn't still solved. In this study, a novel model based on WT (wavelet transform)-SAE (stacked autoencoder)-LSTM is proposed. Firstly, six study sites from China are taken as examples and WT is used to decompose PM2.5 time series into several low-and high- frequency components based on different samples. Secondly, the decomposed components are predicted based on SAE-LSTM. Finally, the predicted results are reconstructed in view of all low-and high-frequency components and the predicted results are obtained. The results imply that: (1) the forecasting performance of SAE-LSTM is better than that of other models (e.g., BP (back propagation)) used for comparison; (2) for six different PM 2.5 samples, four orders five layers, five orders six layers, five orders seven layers, three orders six layers, five orders seven layers, and five orders six layers are the most appropriate. The conclusion that such a novel model may help to enhance the accuracy of PM 2.5 prediction can be drawn.