Scholarlink

<p>Considering that ozone is essential to understanding air quality and climate change, this study presents a deep learning method for predicting atmospheric ozone concentrations. The method combines an attention mechanism with a convolutional neural network (CNN) and long short-term memory (LSTM) network to address the nonlinear nature of multivariate time-series data. It employs CNN and LSTM to extract features from short time series, enhanced by the attention mechanism to improved short-term prediction accuracy. It takes eight meteorological and environmental parameters from 16,806 records (2018–2019) as input, which are selected principal component analysis (PCA). It features an attention-based CNN-LSTM hybrid deep learning model with specific settings: a time step of 5, a batch size of 25, 15 units in the LSTM layer, the ReLU activation function, 25 epochs, and an overfitting avoidance strategy with a dropout rate of 0.15. Experimental results demonstrate that this hybrid model outperforms individual models and the CNN-LSTM model, especially in forward prediction with a multi-hour time lag. The model exhibits a high coefficient of determination (R<sup>2</sup> = 0.971) and a root mean square error of 3.59 for a 1-hour time lag. It also exhibits consistent accuracy across different seasons, highlighting its robustness and superior time-series prediction capabilities for ozone concentrations.</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

Attention mechanism based CNN-LSTM hybrid deep learning model for atmospheric ozone concentration prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月1日

查看全文
  • Jiang Yuan,
  • Hua Dengxin,
  • Wang Yufeng,
  • Yang Xueting,
  • Di Huige,
  • Yan Qing

摘要

Considering that ozone is essential to understanding air quality and climate change, this study presents a deep learning method for predicting atmospheric ozone concentrations. The method combines an attention mechanism with a convolutional neural network (CNN) and long short-term memory (LSTM) network to address the nonlinear nature of multivariate time-series data. It employs CNN and LSTM to extract features from short time series, enhanced by the attention mechanism to improved short-term prediction accuracy. It takes eight meteorological and environmental parameters from 16,806 records (2018–2019) as input, which are selected principal component analysis (PCA). It features an attention-based CNN-LSTM hybrid deep learning model with specific settings: a time step of 5, a batch size of 25, 15 units in the LSTM layer, the ReLU activation function, 25 epochs, and an overfitting avoidance strategy with a dropout rate of 0.15. Experimental results demonstrate that this hybrid model outperforms individual models and the CNN-LSTM model, especially in forward prediction with a multi-hour time lag. The model exhibits a high coefficient of determination (R2 = 0.971) and a root mean square error of 3.59 for a 1-hour time lag. It also exhibits consistent accuracy across different seasons, highlighting its robustness and superior time-series prediction capabilities for ozone concentrations.