Coupled multi-gas indicator model for predicting temperature of coal spontaneous combustion in goaf and its application

Technical Summary: Coupled Multi-Gas Indicator Model for Predicting Temperature of Coal Spontaneous Combustion in Goaf

Problem Statement
Coal mine fires, predominantly caused by spontaneous combustion in goaf (mined-out) areas, pose a severe threat to underground safety. While gas composition analysis is a widely used monitoring technique due to its sensitivity, the relationship between indicator gas concentrations and coal temperature is complex and nonlinear. Existing prediction methods, including traditional statistical approaches and single machine learning models (e.g., Support Vector Regression, standard Neural Networks), often suffer from high sensitivity to hyperparameters, a tendency to fall into local optima, and an inability to adequately capture the nonlinear coupling relationships and temporal dependencies among multiple gas indicators. Consequently, there is an urgent need for a model that can accurately predict coal temperature by effectively fusing multi-gas data and modeling temporal dynamics.

Methodology
The study proposes a novel optimization model named MEF-RG (Multi-Feature Fusion - Random Forest and Gradient Boosting Decision Tree) to predict coal spontaneous combustion temperatures. The methodology involves three primary stages:

1. Data Acquisition and Feature Engineering:

   • Source Data: The model was trained on a large-scale coal spontaneous combustion experimental dataset (2,976 samples) and validated against field data from the Weijiadi Coal Mine.
   • Input Variables: Five key indicator gases were identified: Carbon Monoxide (CO), Oxygen (O₂), Carbon Dioxide (CO₂), Methane (CH₄), and Ethane (C₂H₆).
   • Feature Construction: To capture temporal dynamics and interactions, the authors constructed a 255-dimensional feature matrix. This included:
     • Lag features: Time steps of 1, 2, 3, 6, 12, and 24.
     • Sliding statistical features: Mean, standard deviation, minimum, and maximum within windows of 6, 12, 24, and 48 time steps.
     • Interaction features: Ratios and products of gas concentrations (e.g., O₂/CO).
   • Preprocessing: Data underwent missing value imputation, denoising, and standardization.

2. Model Architecture (MEF-RG):

   • Base Learners: The model utilizes two base algorithms: Random Forest (RF) and Gradient Boosting Decision Tree (GBDT). RF is chosen for its stability and resistance to overfitting, while GBDT provides strong sequential learning capabilities.
   • Meta-Learner Strategy: A "meta-feature enhancement" strategy is employed. The predictions from the RF and GBDT base models, along with their quadratic terms and interaction terms, are combined with the original input features to form an augmented meta-feature vector.
   • Fusion: A linear model (using regularized least squares with options like Ridge or Lasso) acts as the meta-learner to adaptively weight and fuse the base model outputs, aiming to compensate for the inherent limitations of RF (such as static feature slicing) and GBDT.

3. Validation:

   • The model was compared against five base models (RF, GBDT, SVR, Linear Regression, LSTM) and two fusion models (BO-SVR, BPNN).
   • Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²).

Key Contributions

• Novel Fusion Architecture: The paper introduces the MEF-RG model, which systematically integrates GBDT and RF through a meta-learner to address the feature-capture limitations and generalization stability issues of standalone ensemble models.
• Advanced Feature Engineering: The study constructs a high-dimensional feature space (255 dimensions) that explicitly models lagged characteristics, sliding statistics, and gas-gas interactions, moving beyond treating gas concentrations as independent static variables.
• Mechanistic Interpretability: Using SHAP (SHapley Additive exPlana- tions) analysis, the study demonstrates that the model's predictions align with physical laws, such as the inverse relationship between the O₂/CO ratio and predicted temperature during different oxidation stages.

Results
The MEF-RG model demonstrated superior predictive performance compared to all benchmark models on both public datasets and independent field test data from the Weijiadi Coal Mine.

• Performance Metrics (Field Test Set): The MEF-RG model achieved an MAE of 0.8766 °C, an RMSE of 1.2069, and an R² of 0.8605.
• Comparative Improvement: Compared to the best-performing single model (RF), the MEF-RG model reduced MAE by approximately 49.2%, reduced RMSE by 41.9%, and improved R² by 45.6%.
• Generalization: The model maintained high accuracy on the public dataset (R² = 0.961), confirming its robustness and generalization capability across different data sources.

Significance and Limitations
The paper claims that the MEF-RG model provides a reliable technical tool for the earlier and more accurate early warning of coal spontaneous combustion hazards. By accurately capturing the nonlinear coupling relationships among multiple gas indicators, the model offers significant engineering application value for mine safety production.

However, the authors modestly acknowledge specific limitations:

• Data Dependency: The model is highly dependent on the quality of monitoring data for the five specific indicator gases.
• Computational Cost: Due to the 255-dimensional feature matrix and the GBDT-RF fusion architecture, the training time is approximately three times that of a single RF model.
• Real-Time Constraints: The current computational requirements fall short of the millisecond-level response time needed for real-time online early warning systems in goaf areas.

The authors conclude that future work will focus on introducing multi-physical field parameters, conducting research on feature dimensionality reduction, and exploring lightweight model deployment to address these computational constraints.