Predicting the Water-Conducting Fracture Zone (WCFZ) Height Using an MPGA-SVR Approach

Changfang Guo, Zhen Yang, Shen Li, Jinfu LouSustainability 2020 · DOI

TL;DR

This paper introduces a novel Multi-Population Genetic Algorithm-Support Vector Regression (MPGA-SVR) model to accurately predict the height of the water-conducting fracture zone (WCFZ) in coal mining. By optimizing SVR parameters more effectively than traditional genetic algorithms, the MPGA-SVR model demonstrates superior accuracy and stability, crucial for preventing dangerous mine water inrushes and protecting the environment.

What Problem Does This Paper Address?

Coal mining operations worldwide face a significant and persistent threat from mine water inrushes. These events occur when fractures in the overlying rock strata connect with water-bearing bodies, allowing water to flood into active mining areas. The height of the water-conducting fracture zone (WCFZ) is a critical parameter for assessing this risk. Accurately predicting the WCFZ height is essential for implementing effective prevention and control measures, thereby safeguarding mining safety and preventing environmental damage. Existing methods, such as field measurements, theoretical calculations, numerical simulations, and empirical formulas, have limitations including being time-consuming, overly simplistic, or difficult to parameterize accurately. While artificial intelligence methods like Support Vector Regression (SVR) show promise, their performance is highly dependent on optimal parameter selection, a challenge that traditional genetic algorithms (GA) struggle with due to issues like premature convergence and slow convergence.

Methods

The core of this research is the development and application of a Multi-Population Genetic Algorithm-Support Vector Regression (MPGA-SVR) model.

  • Support Vector Regression (SVR): SVR is a machine learning technique used for regression tasks. It maps input data into a high-dimensional feature space to perform linear regression, making it effective for complex, non-linear relationships. The paper details the mathematical formulation of SVR, including its primal and dual problems.
  • Multi-Population Genetic Algorithm (MPGA): This is an enhancement of the standard Genetic Algorithm (GA). While GA uses a single population to evolve solutions, MPGA employs multiple independent populations that periodically exchange individuals (immigration). This strategy is designed to overcome the limitations of GA, specifically premature convergence (where the algorithm gets stuck in a suboptimal solution) and slow convergence. By using diverse populations with different genetic parameters, MPGA aims to achieve a better balance between global and local search capabilities, leading to faster and more accurate convergence to optimal parameters.
  • MPGA-SVR Integration: The MPGA is used to optimize the key parameters of the SVR model: the penalty factor (C) and the kernel function parameter (g). The insensitive loss function parameter (ε) was kept at its default value. The MPGA iteratively searches for the optimal combination of C and g by evaluating the fitness of each individual (representing a set of SVR parameters) based on the mean squared error (MSE) of the SVR model's predictions.

Datasets

The study utilized a dataset comprising 69 samples of water-conducting fracture zone (WCFZ) height. These samples were collected from the 8101 working face of the No. 2 coal seam in the Selian Coal Mine, China. The data includes four key influencing factors:

  • Mining depth (H)
  • Hard rock proportional coefficient (c)
  • Mining thickness (d)
  • Length of working face (L)

For training and testing the MPGA-SVR model, 70% of the data (48 samples) was used for training, and the remaining 30% (21 samples) was used for testing. The paper also mentions applying the model to a field test.

Key Insights and Ideas

  • MPGA's Superiority over GA for Parameter Optimization: The most significant insight is the demonstrated advantage of MPGA over standard GA in optimizing SVR parameters. MPGA's multi-population approach effectively mitigates premature convergence and accelerates the search process, leading to more robust and reliable parameter selection.
  • Holistic Factor Consideration: The MPGA-SVR model's strength lies in its ability to simultaneously consider multiple influencing factors (mining depth, hard rock coefficient, thickness, and length) for WCFZ height prediction. This contrasts with traditional empirical formulas that often rely on a single factor, making them less representative of complex geological conditions.
  • Improved Prediction Accuracy and Stability: The MPGA-SVR model consistently outperformed traditional methods and GA-SVR in terms of both prediction accuracy (higher correlation coefficient) and stability (lower variance in results across multiple runs). This enhanced reliability is critical for practical safety applications.
  • Practical Applicability: The model's successful application to a real-world field test validates its engineering utility and potential for real-time risk assessment in mining operations.

Key Results

  • Parameter Optimization: The MPGA converged to optimal SVR parameters (C and g) in significantly fewer generations (7) compared to GA (32 generations). The MSE during optimization was reduced from 0.014 to 0.005 by MPGA, versus 0.018 to 0.007 by GA.
  • Prediction Accuracy:
    • For training samples, MPGA-SVR achieved a correlation coefficient (r) of 0.97 and an MSE of 7.29 m², compared to GA-SVR's r of 0.96 and MSE of 11.13 m².
  • For test samples, MPGA-SVR achieved an r of 0.96 and an MSE of 28.70 m², while GA-SVR had an r of 0.93 and an MSE of 34.91 m².
  • Prediction Stability: When tested over 10 repeated runs, the MPGA-SVR model maintained a correlation coefficient consistently above 0.95, demonstrating high stability. In contrast, the GA-SVR model showed significant variability, with one run yielding a correlation coefficient as low as 0.6, attributed to premature convergence.
  • Field Application: In a field test comparing MPGA-SVR, GA-SVR, and an empirical formula, the MPGA-SVR model provided predictions closest to the measured WCFZ height (approximately 50 m) with the lowest absolute relative error. The empirical formula's error was significantly higher.

Limitations and Future Work

  • Limitations:
    • The study used a relatively small dataset (69 samples). While SVR is known for its effectiveness with small datasets, a larger and more diverse dataset could further enhance model generalizability.
  • The paper acknowledges that the model will be "continuously enriched and updated in practice." This implies that the current model is a foundational step rather than a final, perfected solution.
  • Future Work:
    • The authors suggest exploring different migration mechanisms and hierarchical execution architectures for sub-populations within the MPGA framework to achieve even more accurate results.
  • Further research could focus on incorporating additional influencing factors or exploring different machine learning algorithms in conjunction with MPGA.
  • The paper implies that continuous refinement and updating of the model based on ongoing field data will be crucial for improving its accuracy and generalizability.

Summary updated

Abstract

Mine water that inrushes from coal-roof strata has always posed a substantial threat to mining activities every year. Therefore, an accurate prediction of the water-conducting fracture zone (WCFZ) height in the mining overburden strata is of great significance for the prevention and control of mine water accidents. The support vector regression (SVR) is proposed to predict the height of the WCFZ based on the mining depth, hard rock proportional coefficient, mining thickness and length of the working face. Simultaneously, the multi-population genetic algorithm (MPGA) is employed to search for the optimal SVR parameters. The MPGA-SVR model is trained and tested with a total of 69 collected data samples, and it is also applied to a field test. The accuracy and stability of the model were measured by the mean squared error and correlation coefficients. The obtained results show that the MPGA-SVR model achieves a higher accuracy and stability than the traditional empirical formula and genetic algorithm (GA)-SVR model. In terms of the process for optimizing the SVR parameters, the MPGA can find the optimal parameters more quickly and accurately, and it can effectively overcome the problem of premature and slow convergence of the genetic algorithm (GA). The proposed model improves the prediction accuracy and stability, which will help to avoid accidents caused by the inrush of water inrush in mining overburden strata and protect the ecological environment of the mining area.
BibTeX
@inproceedings{guo2020predicting,
  author = {Changfang Guo and Zhen Yang and Shen Li and Jinfu Lou},
  title = {{Predicting the Water-Conducting Fracture Zone (WCFZ) Height Using an MPGA-SVR Approach}},
  booktitle = {Sustainability},
  year = {2020},
  doi = {10.3390/su12051809},
}
Figure