What type of study is this?

This is a Experimental Study study.

October 3, 2025Open Access

Gas concentration prediction based on SSA algorithm with CNN-BiLSTM-attention

WXWenjing XingShandong University YYYanguo YangTaiyuan University of Technology YZYanxin ZhangFirst Affiliated Hospital of GuangXi Medical University

Key Points

The proposed model achieves a low RMSE of 0.0171 and MAPE of 0.084, indicating effective gas concentration predictions.
Compared to other models, this approach improves RMSE by up to 30.2% and MAPE by up to 43.62%, showcasing superior accuracy.
Using a combination of CNN, BiLSTM, and attention mechanisms, the model effectively captures spatiotemporal features for predictions.
Generalization capability is enhanced as prediction horizons increase, enabling reliable real-time gas monitoring and safety.

Abstract

Abstract Accurate prediction of coal mine gas concentration is a crucial prerequisite for preventing gas exceed and disasters. However, the existing methods still suffer from issues such as low data utilization, difficulty in effectively integrating multivariate nonlinear spatiotemporal features, and poor generalization capability when achieving relatively high prediction accuracy but requiring longer prediction durations. To address these challenges, this study focuses on a tunneling face in a Shanxi coal mine and proposes a novel hybrid deep learning model (CNN-BiLSTM-Attention). The model employs a 1D-CNN to extract local spatial features of gas concentration, temperature, wind speed, rock pressure, and CO concentration, utilizes BiLSTM to model bidirectional temporal dependencies, and incorporates an attention mechanism to dynamically weight critical features, such as sudden changes in gas concentration. Additionally, the sparserow search algorithm (SSA) was applied to automatically optimize hyperparameters, including the number of CNN filters and BiLSTM hidden units. The results demonstrate that the proposed SCBA model achieves an RMSE of 0.0171 and MAPE of 0.084. Compared to mainstream models such as attention-LSTM, SSA-LSTM-aAttention, and rTransformer-LSTM, the RMSE was improved by 23.3%, 4.4%, and 30.2%, respectively, while the MAPE was improved by 38.7%, 14.3%, and 43.62%, respectively, indicating superior prediction accuracy. To validate the contribution of each module, ablation experiments were conducted by sequentially removing the CNN, BiLSTM, and Attention components for error analysis, confirming the necessity of the multi-module collaborative mechanism. In multi-step predictions, the proposed model exhibits better generalization capability as the prediction horizon increases, with the slowest error growth, providing insights for enhancing gas concentration prediction generalization. The model achieves the smallest error at 20 time steps, laying a foundation for subsequent anomaly analysis and offering miners sufficient safety evacuation time, thereby establishing a reliable basis for real-time gas disaster early warning.

KI fragen

Bookmark

View Full Paper