Scholarlink

<p>Releasing combustion gases containing particulate matter (PM) can cause air pollution. To mitigate health concerns, it is essential to have accurate and reliable methods for predicting PM<sub>10</sub> levels. The ability to predict PM<sub>10</sub> enables the potential development of an early warning system to implement preventive actions and avoid negative impacts on ambient air quality and public health. This study proposes a method for predicting PM10 concentration after 24&#xa0;h using the discrete wavelet transform (DWT), multi-layer perceptron (MLP), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM) models, which involve hourly historical air pollutant concentrations. Models are compared from two different perspectives. First, good performance models are identified based on wavelet types, and then, models that are successful in deep learning methods among these models are determined. This study presents a novel comprehensive prediction analysis of the multivariate time series for PM<sub>10</sub> concentration data in Istanbul, Türkiye. The hybrid model DWT-LSTM with the Db10 wavelet type outperforms other prediction models, with a root mean square error (RMSE) of 1.959, a mean absolute error (MAE) of 1.273, and a coefficient of determination (R²) of 99.2%. The findings provide strong empirical support for the air quality of cities. They can be utilized to assist institutions and organizations in taking preventative measures at the appropriate time, such as having the potential to be used directly in early warning systems related to air pollution, preparedness for hospital emergency services, information platforms for people with health conditions, smart home systems, and even for implementing traffic restrictions.</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

Advancing multivariate air pollution prediction using wavelet-based smoothing and deep learning methods

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Gökçe Nur Taşağıl Arslan,
  • Serpil Kılıç Depren

摘要

Releasing combustion gases containing particulate matter (PM) can cause air pollution. To mitigate health concerns, it is essential to have accurate and reliable methods for predicting PM10 levels. The ability to predict PM10 enables the potential development of an early warning system to implement preventive actions and avoid negative impacts on ambient air quality and public health. This study proposes a method for predicting PM10 concentration after 24 h using the discrete wavelet transform (DWT), multi-layer perceptron (MLP), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM) models, which involve hourly historical air pollutant concentrations. Models are compared from two different perspectives. First, good performance models are identified based on wavelet types, and then, models that are successful in deep learning methods among these models are determined. This study presents a novel comprehensive prediction analysis of the multivariate time series for PM10 concentration data in Istanbul, Türkiye. The hybrid model DWT-LSTM with the Db10 wavelet type outperforms other prediction models, with a root mean square error (RMSE) of 1.959, a mean absolute error (MAE) of 1.273, and a coefficient of determination (R²) of 99.2%. The findings provide strong empirical support for the air quality of cities. They can be utilized to assist institutions and organizations in taking preventative measures at the appropriate time, such as having the potential to be used directly in early warning systems related to air pollution, preparedness for hospital emergency services, information platforms for people with health conditions, smart home systems, and even for implementing traffic restrictions.