Interpretable regional meteorological feature extraction enhances deep learning for extended 120-h PM2.5 forecasting

While deep learning models perform well in short-term PM2.5 forecasting, their performance tends to decay significantly with increasing forecast spans. This study proposes a “Domain-to-Point” approach to enhance deep learning models for extended 120-h PM2.5 forecasting. We used convolutional autoenc...

Full description

Saved in:
Bibliographic Details
Published in:Journal of cleaner production Vol. 483; p. 144287
Main Authors: Liu, Xinyi, Pu, Xueting, Lu, Chengwei, Zhang, Han, Li, Tao, Grieneisen, Michael L., Li, Jucheng, Ma, Ning, Yan, Chang, Zhan, Yu, Yang, Fumo
Format: Journal Article
Language:English
Published: Elsevier Ltd 10.12.2024
Subjects:
Convolutional autoencoder
Deep learning
External feature engineering
Meteorological feature extraction
PM2.5 forecasting
Deep learning
Meteorological feature extraction
PM2.5 forecasting
External feature engineering
Convolutional autoencoder
ISSN:0959-6526
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:While deep learning models perform well in short-term PM2.5 forecasting, their performance tends to decay significantly with increasing forecast spans. This study proposes a “Domain-to-Point” approach to enhance deep learning models for extended 120-h PM2.5 forecasting. We used convolutional autoencoders (ConvAEs) to extract regional meteorological features (RMFs) from both surrounding and future meteorological data provided by an advanced weather forecasting system. The RMFs and temporal indicators were used as predictor variables in a time-attention-based bidirectional long short-term memory neural network to forecast the next 24- to 120-h PM2.5. For four megacities in China, the ConvAEs compressed the high-dimensional meteorological data into a lower-dimensional representation, while preserving most of the relevant information (R2 > 0.75). Using the RMFs as inputs substantially mitigated the performance decay of the PM2.5 forecasting models over extended forecast spans, enabling accurate longer-term forecasting, with R2 improvements of 14.7–79.2% for the 120-h forecasts. This study further revealed that errors in the meteorological forecasts are the primary cause of performance decay, suggesting that improving the accuracy of meteorological forecasts could be an effective approach for enhancing PM2.5 forecasting performance. With Beijing as a case study, model interpretation using the Shapley-additive-explanations method demonstrated the crucial role played by RMFs in accurately forecasting a typical air pollution episode and identifying the main meteorological factors affecting pollution. The proposed “Domain-to-Point” approach enhances the capability of deep learning models for longer-term PM2.5 forecasting, thus facilitating the development of preventative measures to safeguard public health against air pollution. [Display omitted] •Regional meteorological features can be extracted via convolutional autoencoders.•External feature engineering enhances deep learning in forecasting long-term PM2.5•Model interpretation reveals the extracted features dominate PM2.5 forecasting.•Improving meteorological forecasting is crucial for enhancing PM2.5 forecasting.
ISSN:0959-6526
DOI:10.1016/j.jclepro.2024.144287