Scholarlink

<p>Accurate prediction of gas concentration is critical for coal mine safety. However, under complex underground conditions, including ventilation disturbances, geological heterogeneity, and uncertainties in mining activities, gas concentration time series are often contaminated by strong noise and exhibit multi-level nonlinear temporal characteristics. These challenges limit the predictive performance of existing models. To address this issue, this paper proposes GMKNet, a multi-scale and multi-factor gas concentration prediction model designed to capture complex temporal structures and multi-factor coupling relationships. First, a GASFT module is developed by integrating a dynamic frequency cutoff mechanism with an intelligent noise detection mechanism. This module denoises and reconstructs the original gas concentration sequence, thereby improving input data quality. Second, an MDPM module is constructed within the encoder–decoder architecture. Temporal dependencies at different time granularities are extracted through multi-scale convolution and deformable receptive fields, while key time steps are emphasized using the ProbSparse attention mechanism, enabling multi-level modeling of complex temporal features. Finally, the KAN module is incorporated to capture nonlinear relationships in gas concentration data, further enhancing the model’s representation capability and prediction accuracy. Experimental results on three real-world datasets, namely MM264, JN-GZM, and JN-HFL, demonstrate that GMKNet achieves the best overall performance on mine data with different distribution characteristics and influencing factors. Compared with the baseline models, GMKNet reduces MSE, MAE, and RMSE by an average of 15.53%, 18.33%, and 8.24%, respectively, thereby significantly improving prediction accuracy. These results indicate that the proposed model effectively addresses several limitations of existing models, including insufficient denoising capability, inadequate utilization of multi-level temporal information, and limited ability to capture the intrinsic nonlinear characteristics of the data.</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

Gas Concentration Prediction Model with the Multi-scale and Multi-factor Based on GMKNet

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • XingYu Gong,
  • RongKun Jiang,
  • Yu Guan,
  • Yi Yang,
  • Na Li

摘要

Accurate prediction of gas concentration is critical for coal mine safety. However, under complex underground conditions, including ventilation disturbances, geological heterogeneity, and uncertainties in mining activities, gas concentration time series are often contaminated by strong noise and exhibit multi-level nonlinear temporal characteristics. These challenges limit the predictive performance of existing models. To address this issue, this paper proposes GMKNet, a multi-scale and multi-factor gas concentration prediction model designed to capture complex temporal structures and multi-factor coupling relationships. First, a GASFT module is developed by integrating a dynamic frequency cutoff mechanism with an intelligent noise detection mechanism. This module denoises and reconstructs the original gas concentration sequence, thereby improving input data quality. Second, an MDPM module is constructed within the encoder–decoder architecture. Temporal dependencies at different time granularities are extracted through multi-scale convolution and deformable receptive fields, while key time steps are emphasized using the ProbSparse attention mechanism, enabling multi-level modeling of complex temporal features. Finally, the KAN module is incorporated to capture nonlinear relationships in gas concentration data, further enhancing the model’s representation capability and prediction accuracy. Experimental results on three real-world datasets, namely MM264, JN-GZM, and JN-HFL, demonstrate that GMKNet achieves the best overall performance on mine data with different distribution characteristics and influencing factors. Compared with the baseline models, GMKNet reduces MSE, MAE, and RMSE by an average of 15.53%, 18.33%, and 8.24%, respectively, thereby significantly improving prediction accuracy. These results indicate that the proposed model effectively addresses several limitations of existing models, including insufficient denoising capability, inadequate utilization of multi-level temporal information, and limited ability to capture the intrinsic nonlinear characteristics of the data.