结合卷积神经网络和注意力机制的LSTM采空区地表沉降预测方法

LSTM goaf surface subsidence prediction method combining convolutional neural network and attention mechanism

为解决采空区地表塌陷区域时序预测中存在的监测点空间特征难以提取的问题,本文提出了一种可以提取监测点关键空间特征的CNN-Attention-LSTM组合神经网络模型。首先,增加作为特征输入的邻近监测点个数,使用卷积神经网络(CNN)提取由多个监测点构成的多维时间序列的空间特征;其次,将提取后的多维特征时序输入多层感知器(MLP)中计算注意力权重,并与特征输入作Hadamard积,实现特征输入的注意力权重分配;然后,利用长短期记忆神经网络(LSTM)进行回归预测;最后,通过全连接层,整合输出目标监测点的预测值。本文以龙首矿西二采区地表塌陷区域为例,给出其地表沉降监测数据预测结果,并与实际采集的数据作对比。结果表明,引入注意力机制的CNN-Attention-LSTM的组合模型比CNN-LSTM模型和LSTM模型精度更高,且增加有效特征输入能够显著提升CNN-Attention-LSTM模型的预测精度。

In order to solve the problem of difficult extraction of spatial features of monitoring points in the time series prediction of surface collapse areas in the mining zone,a CNN-Attention-LSTM combined neural network model that can extract key spatial features of monitoring points is proposed.Firstly,the number of neighbouring monitoring points as feature input is increased,and the spatial features of the multidimensional time series composed of multiple monitoring points are extracted using convolutional neural network(CNN).Secondly,the extracted multidimensional feature time series are input into the multilayer perceptron(MLP)to calculate the attention weights and make Hadamard product with the feature inputs to achieve the allocation of the attention weights of the feature inputs.After that regression prediction is performed using long short term memory neural network(LSTM).Finally,through the fully connected layer,the predicted values of the target monitoring points are integrated and output.In this paper,we take the surface collapse area in the west second mining area of Longshou mine as an example to give the prediction results of its surface subsidence monitoring data and compare them with the actual collected data.The results show that the combined CNN-Attention-LSTM model with the introduction of the attention mechanism is more accurate than the CNN-LSTM model and the LSTM model respectively,and the addition of effective feature inputs can significantly improve the prediction accuracy of the CNN-Attention-LSTM model.