Scholarlink

<p>The short-term air quality forecasting models serve as an early warning system for local agencies, aiding in preparing mitigation strategies against severe pollution episodes. This paper explores the application of Transfer Learning to enhance short-term air quality forecasting model accuracy when labelled data is limited or missing, as often occurs with newly installed monitoring stations or due to sensor malfunctions. These monitoring stations are typically installed in areas of high exposure, like roads or urban/industrial areas, due to recurrent peak episodes or to monitor background pollutant levels generally. Forecasts with greater reliability, even when there is limited historical data available due to the recent installation of the monitoring station for example, are expected to enable the swift implementation of proactive measures to prevent significant pollution episodes from happening. The proposed method leverages knowledge from spatially neighbouring air quality monitoring stations to achieve the multi-modal spatial-temporal transfer learning to the target station, exploring multivariate time series data available from neighbouring monitoring stations. This study employed historical air quality data from spatially adjacent monitoring stations identified in South Wales, UK. The study evaluates the predictive capabilities of four base models and their corresponding transfer learning variants for estimating NO<sub>2</sub> and PM<sub>10</sub> pollutant levels, which are the most difficult pollutants to meet objectives and limit values in the UK’s air quality strategy. The paper highlights the importance of capturing spatial patterns from different monitoring stations along with temporal trends when it comes to air quality prediction. Our experiments demonstrate that transfer learning models outperform models trained from scratch on air quality multivariate time series prediction problems in a low data environment. The proposed hybrid Graph Convolutional-LSTM model, making use of a novel Granger causality-based adjacency matrix for the new site, has significantly outperformed other baseline models in predicting pollutants, achieving notable improvements in prediction accuracy of approximately 8% for PM<sub>10</sub> and 7% for NO<sub>2</sub> values, as reflected in the RMSE values. It has also demonstrated the potential for data-efficient approaches in spatial transfer learning by reducing the need for large datasets by incorporating prior causal information.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid graph convolutional LSTM model for spatio-temporal air quality transfer learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月5日

查看全文
  • Sooraj Raj,
  • Jim Smith,
  • Enda Hayes

摘要

The short-term air quality forecasting models serve as an early warning system for local agencies, aiding in preparing mitigation strategies against severe pollution episodes. This paper explores the application of Transfer Learning to enhance short-term air quality forecasting model accuracy when labelled data is limited or missing, as often occurs with newly installed monitoring stations or due to sensor malfunctions. These monitoring stations are typically installed in areas of high exposure, like roads or urban/industrial areas, due to recurrent peak episodes or to monitor background pollutant levels generally. Forecasts with greater reliability, even when there is limited historical data available due to the recent installation of the monitoring station for example, are expected to enable the swift implementation of proactive measures to prevent significant pollution episodes from happening. The proposed method leverages knowledge from spatially neighbouring air quality monitoring stations to achieve the multi-modal spatial-temporal transfer learning to the target station, exploring multivariate time series data available from neighbouring monitoring stations. This study employed historical air quality data from spatially adjacent monitoring stations identified in South Wales, UK. The study evaluates the predictive capabilities of four base models and their corresponding transfer learning variants for estimating NO2 and PM10 pollutant levels, which are the most difficult pollutants to meet objectives and limit values in the UK’s air quality strategy. The paper highlights the importance of capturing spatial patterns from different monitoring stations along with temporal trends when it comes to air quality prediction. Our experiments demonstrate that transfer learning models outperform models trained from scratch on air quality multivariate time series prediction problems in a low data environment. The proposed hybrid Graph Convolutional-LSTM model, making use of a novel Granger causality-based adjacency matrix for the new site, has significantly outperformed other baseline models in predicting pollutants, achieving notable improvements in prediction accuracy of approximately 8% for PM10 and 7% for NO2 values, as reflected in the RMSE values. It has also demonstrated the potential for data-efficient approaches in spatial transfer learning by reducing the need for large datasets by incorporating prior causal information.