Scholarlink

<p>Accurate gas concentration prediction is vital for preventing mining accidents. Traditional methods exhibit limited accuracy and computational inefficiency in handling complex time-series data. This paper proposes a CEEMDAN–SCSSA–CNN–BiLSTM hybrid model to enhance prediction performance. The framework integrates a trigonometric-enhanced Cauchy mutation sparrow search algorithm (SCSSA) to optimize hyperparameters, improving global convergence and nonlinear optimization capabilities. A convolutional neural network (CNN) extracts spatial features from decomposed signals, while a bidirectional long short-term memory (BiLSTM) captures bidirectional temporal dependencies, enhancing sensitivity to concentration variations. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) preprocesses raw data through multi-scale decomposition and noise reduction, followed by feature reconstruction. Comparative experiments showed that the proposed model achieved root mean squared errors (RMSEs) of 0.0026 and 0.0053 on the training and test sets, respectively; mean absolute errors (MAEs) of 0.0020 and 0.0043; mean absolute percentage errors (MAPEs) of 2.7% and 2.6%; mean squared errors (MSEs) of 7.58 × 10<sup>−6</sup> and 2.8 × 10<sup>−5</sup>; and coefficients of determination of 0.9983 and 0.9906. Compared with the closest competing method (SCSSA–CNN–BiLSTM), the proposed model reduced the test set RMSE, MAE, and MAPE by approximately 8.9%, 8.7%, and 9.9%, respectively; relative to the non-optimized CNN–BiLSTM, the test set RMSE was reduced by about 52.6%. The synergistic integration of signal decomposition, metaheuristic optimization, and deep learning components enables robust handling of complex temporal patterns while maintaining computational efficiency. This approach provides a reliable technical solution for real-time gas monitoring systems, significantly advancing mine safety management through improved early warning capabilities.</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

Time-Series Prediction of Mine Gas Concentration Based on CEEMDAN–SCSSA–CNN–BiLSTM

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Kangnan Li,
  • Yaqin Wu,
  • Feng Du,
  • Xiang Zhang,
  • Yiqiao Wang,
  • Jiazhi Sun,
  • Yangyang Guo,
  • Minghao Zhao

摘要

Accurate gas concentration prediction is vital for preventing mining accidents. Traditional methods exhibit limited accuracy and computational inefficiency in handling complex time-series data. This paper proposes a CEEMDAN–SCSSA–CNN–BiLSTM hybrid model to enhance prediction performance. The framework integrates a trigonometric-enhanced Cauchy mutation sparrow search algorithm (SCSSA) to optimize hyperparameters, improving global convergence and nonlinear optimization capabilities. A convolutional neural network (CNN) extracts spatial features from decomposed signals, while a bidirectional long short-term memory (BiLSTM) captures bidirectional temporal dependencies, enhancing sensitivity to concentration variations. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) preprocesses raw data through multi-scale decomposition and noise reduction, followed by feature reconstruction. Comparative experiments showed that the proposed model achieved root mean squared errors (RMSEs) of 0.0026 and 0.0053 on the training and test sets, respectively; mean absolute errors (MAEs) of 0.0020 and 0.0043; mean absolute percentage errors (MAPEs) of 2.7% and 2.6%; mean squared errors (MSEs) of 7.58 × 10−6 and 2.8 × 10−5; and coefficients of determination of 0.9983 and 0.9906. Compared with the closest competing method (SCSSA–CNN–BiLSTM), the proposed model reduced the test set RMSE, MAE, and MAPE by approximately 8.9%, 8.7%, and 9.9%, respectively; relative to the non-optimized CNN–BiLSTM, the test set RMSE was reduced by about 52.6%. The synergistic integration of signal decomposition, metaheuristic optimization, and deep learning components enables robust handling of complex temporal patterns while maintaining computational efficiency. This approach provides a reliable technical solution for real-time gas monitoring systems, significantly advancing mine safety management through improved early warning capabilities.