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

Area of Science:

• Mining Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate coal mine gas concentration prediction is vital for preventing disasters.
• Existing methods struggle with multivariate nonlinear spatiotemporal features and generalization.
• Challenges include low data utilization and poor prediction accuracy over longer durations.

Purpose of the Study:

• To propose a novel hybrid deep learning model for enhanced coal mine gas concentration prediction.
• To address limitations in existing methods regarding data utilization and feature integration.
• To improve the generalization capability of prediction models for longer time horizons.

Main Methods:

• Developed a hybrid CNN-BiLSTM-Attention deep learning model.
• Employed 1D-CNN for local spatial feature extraction and BiLSTM for temporal dependencies.
• Integrated an attention mechanism for dynamic feature weighting and used Sparse Row Search (SSA) for hyperparameter optimization.

Main Results:

• The proposed model achieved superior prediction accuracy with RMSE of 0.0171 and MAPE of 0.084.
• Demonstrated significant improvements over mainstream models like attention-LSTM, SSA-LSTM-Attention, and rTransformer-LSTM.
• Ablation experiments confirmed the necessity of the multi-module collaborative mechanism for prediction accuracy.

Conclusions:

• The hybrid CNN-BiLSTM-Attention model offers enhanced gas concentration prediction accuracy and generalization capability.
• The model provides a reliable basis for real-time gas disaster early warning systems.
• Findings contribute to improving safety by offering miners sufficient evacuation time through accurate predictions.