导水裂隙带高度是顶板(涌)突水、地下水资源流失的重要影响因素之一，是矿井防治水研究的重点。

为了准确地预测煤层顶板导水裂隙带高度，选取开采深度、采高、煤层倾角、工作面斜长、硬岩岩性比例系数和开采方法作为导水裂隙带高度的主要影响因素，搜集200例导水裂隙带高度实测样本作为模型数据集。首先，采用自适应高斯噪声过采样方法(synthetic minority over-sampling technique for regression with Gaussian noise，SMOGN)对原始数据集进行过采样，结合8折交叉验证，将平均绝对误差(E_MA)、均方根误差(E_RMS)和决定系数(R²)作为回归模型评价指标，确定最优的BP神经网络结构，然后采用变异粒子群优化算法(mutation particle swarm optimization，MPSO)，对神经网络的初始权值和阈值进行优化，最后将优化后的预测模型进行工程现场应用。

结果表明：该数据集下，BP神经网络采用Huber loss和Adam一阶优化算法，训练速度和稳定性均得到提升，最优激活函数为Tanh，最优隐藏层节点数为12。当MPSO种群数量为50时，模型性能最好，经过SMOGN过采样和MPSO超参数优化，最终训练集的E_MA为0.163，E_RMS为0.216，R²为0.948，验证集的E_MA为0.260，E_RMS为0.341，R²为0.901。在现场应用中模型预测的相对误差均在9%以下。结果表明结合SMOGN技术和MPSO超参数优化技术，显著提高了模型的稳定性和泛化性能，改善了样本分布特征，提高了样本利用效率和模型预测效果，对导水裂隙带高度模型的训练和预测具有重要的借鉴意义。

The height of a hydraulically conductive fracture zone, a significant factor influencing roof water inrushes and groundwater resource loss, is identified as a research focus of the prevention and control of mine water disasters.

To accurately predict the heights of hydraulically conductive fracture zones in coal seam roofs, five parameters were selected as the primary factors influencing hydraulically conductive fracture zones the mining depth: mining height, coal seam inclination, the length of the mining face along its dip direction, proportional coefficient of hard rocks (i.e., the ratio of the cumulative thickness of hard rocks within the statistical height above the coal seam roof to the statistical height), and mining method. A total of 200 measured samples concerning the heights of hydraulically conductive fracture zones were collected as the model dataset. First, over-sampling of the original dataset was conducted using the synthetic minority over-sampling technique for regression (SmoteR) combined with the introduction of Gaussian Noise (SMOGN). In conjunction with 8-fold cross-validation, the optimal back propagation (BP) neural network structure was determined by using the mean absolute error (denoted by E_MA), root mean square error (denoted by E_RMS), and coefficient of determination (denoted by R²) as the assessment indices of the regression model. Then, the initial weights and thresholds of the BP neural network were optimized using the mutation particle swarm optimization (MPSO) algorithm. Finally, the optimized prediction model, i.e., the MPSO-BP model, was applied to the engineering field.

The results indicate that based on the original dataset, the BP neural network, using the Huber loss and Adam first-order optimization algorithm, enhanced the training speed and stability. Consequently, the optimal activation function was determined at Tanh and the optimal hidden layer node number at 12. The MPSO-BP model yielded the optimal performance where the MPSO population number was 50. After SMOGN and MPSO, the training set yielded an E_MA value of 0.163, an E_RMS value of 0.216, and an R² value of 0.948, and these values were 0.260, 0.341, and 0.901, respectively, for the validation set. The field application indicated that the MPSO-BP model yielded relative errors of below 9% in the prediction. Therefore, the integration of the SMOGN and MPSO can significantly enhance the stability and generalization capability of the prediction model, the sample distribution characteristics, the sample utilization efficiency, and the predicted effects of the model. This study can serve as a reference for the training and prediction of models for the heights of hydraulically conductive fracture zones.