Scholarlink

<p>Reliable prediction of strata-pressure evolution is essential for intelligent longwall mining, but steeply inclined panels show spatially heterogeneous support loading that challenges short-term warning. Here we analyse two months of hydraulic-support data from Panel II1013 in the Huaibei mining area and develop a local prediction workflow for support-pressure states. Missing and zero values were repaired, random measurement noise was reduced using a one-dimensional Kalman filter, and normalized sliding-window samples were used to compare CNN, LSTM, CNN-LSTM, Transformer and CNN-LSTM-Attention models against persistence and BP neural network baselines. Data analysis revealed a persistent high-pressure concentration from the middle to upper face. Under matched data partitioning, CNN-LSTM-Attention achieved the lowest test RMSE and highest R<sup>2</sup> (RMSE 0.8632 ± 0.0615; MAE 0.4233 ± 0.0902; MAPE 2.5194 ± 0.4429%; R<sup>2</sup> 0.9891 ± 0.0016), reducing RMSE by 1.80% relative to BP and 9.60% relative to persistence. In a held-out 1000-sample window, the model achieved a relative regression accuracy of 98.17% (1—MAPE). A bounded multi-step validation on four representative supports, using a 24-point input window (approximately 2&#xa0;h), yielded valid forecast horizons of 5–10&#xa0;h when both horizon-level and farthest-step MAPE were ≤ 10%. These results support CNN-LSTM-Attention as a local single-support prediction module for graded warning assistance. Broader deployment will require multi-support spatiotemporal modelling and field validation of closed-loop support-control decisions.</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

Deep learning-based intelligent prediction of strata pressure in a steeply inclined fully mechanized longwall face

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Zhuocheng Ding,
  • Naizhong Xu,
  • Chang Su

摘要

Reliable prediction of strata-pressure evolution is essential for intelligent longwall mining, but steeply inclined panels show spatially heterogeneous support loading that challenges short-term warning. Here we analyse two months of hydraulic-support data from Panel II1013 in the Huaibei mining area and develop a local prediction workflow for support-pressure states. Missing and zero values were repaired, random measurement noise was reduced using a one-dimensional Kalman filter, and normalized sliding-window samples were used to compare CNN, LSTM, CNN-LSTM, Transformer and CNN-LSTM-Attention models against persistence and BP neural network baselines. Data analysis revealed a persistent high-pressure concentration from the middle to upper face. Under matched data partitioning, CNN-LSTM-Attention achieved the lowest test RMSE and highest R2 (RMSE 0.8632 ± 0.0615; MAE 0.4233 ± 0.0902; MAPE 2.5194 ± 0.4429%; R2 0.9891 ± 0.0016), reducing RMSE by 1.80% relative to BP and 9.60% relative to persistence. In a held-out 1000-sample window, the model achieved a relative regression accuracy of 98.17% (1—MAPE). A bounded multi-step validation on four representative supports, using a 24-point input window (approximately 2 h), yielded valid forecast horizons of 5–10 h when both horizon-level and farthest-step MAPE were ≤ 10%. These results support CNN-LSTM-Attention as a local single-support prediction module for graded warning assistance. Broader deployment will require multi-support spatiotemporal modelling and field validation of closed-loop support-control decisions.