Scholarlink

<p>During oil shale pyrolysis, flue gas treatment via the Alberta Taciuk Process (ATP) is essential to meet environmental regulations. Accurate prediction of circulating airflow in flue gas coolers is critical for stable and safe system operation, yet remains challenging due to strong noise, fluctuations, and complex nonlinear dynamics. To address this, we propose a hybrid deep learning framework integrating convolutional neural networks (CNN), a self-organizing (SO) mechanism, bidirectional long short-term memory (BiLSTM) networks, efficient channel attention (ECA), and self-attention (SA). CNN captures intricate inter-variable dependencies under noisy, nonlinear conditions; BiLSTM models irregular temporal patterns; SO adaptively refines network structure to enhance flexibility and performance. The combined attention module–ECA for channel-wise relevance and SA for long-range contextual modeling–further improves feature representation. Experiments across multiple time scales demonstrate the proposed model’s superior predictive accuracy. Average MAPE values are: CNN-SObiLSTM-SAECA (0.002), CNN-biLSTM-SAECA (0.0027), CNN-SObiLSTM-AM (0.0024), CNN-SObiLSTM (0.0024), CNN-biLSTM (0.0026), LSTM (0.003), RNN (0.003), and CNN (0.0048). Results confirm the model’s robustness and efficacy for airflow forecasting, enhancing the reliability of ATP-based flue gas treatment systems.</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

Prediction of circulating airflow in flue gas cooler based on hybrid deep learning mode

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • J. Yang,
  • Y. Chi,
  • K. Cao,
  • S. Shan,
  • H. Xin,
  • Z. Xu,
  • S. Yang,
  • L. Zhao

摘要

During oil shale pyrolysis, flue gas treatment via the Alberta Taciuk Process (ATP) is essential to meet environmental regulations. Accurate prediction of circulating airflow in flue gas coolers is critical for stable and safe system operation, yet remains challenging due to strong noise, fluctuations, and complex nonlinear dynamics. To address this, we propose a hybrid deep learning framework integrating convolutional neural networks (CNN), a self-organizing (SO) mechanism, bidirectional long short-term memory (BiLSTM) networks, efficient channel attention (ECA), and self-attention (SA). CNN captures intricate inter-variable dependencies under noisy, nonlinear conditions; BiLSTM models irregular temporal patterns; SO adaptively refines network structure to enhance flexibility and performance. The combined attention module–ECA for channel-wise relevance and SA for long-range contextual modeling–further improves feature representation. Experiments across multiple time scales demonstrate the proposed model’s superior predictive accuracy. Average MAPE values are: CNN-SObiLSTM-SAECA (0.002), CNN-biLSTM-SAECA (0.0027), CNN-SObiLSTM-AM (0.0024), CNN-SObiLSTM (0.0024), CNN-biLSTM (0.0026), LSTM (0.003), RNN (0.003), and CNN (0.0048). Results confirm the model’s robustness and efficacy for airflow forecasting, enhancing the reliability of ATP-based flue gas treatment systems.