The content of this article comes from the "Surveying and Mapping Bulletin" No. 6 in 2024, drawing review number: GS Jing (2024) No. 1024
A prediction method of LSTM goaf surface subsidence combining convolutional neural network and attention mechanism
Gao Motong1,2,3, Yang Weifang1,2,3, Liu Zuyu4, Cao Xiaoshuang1,2,3, Zhang Ruiqi1,2,3, Hou Yuhao1,2,3
1. School of Geomatics, Surveying, Mapping and Geoinformatics, Lanzhou Jiaotong University, Lanzhou 730070, China; 2. National and Local Joint Engineering Research Center for Application of Geographic National Conditions Monitoring Technology, Lanzhou 730070, China; 3. Gansu Provincial Engineering Laboratory of Geographic National Conditions Monitoring, Lanzhou 730070, China; 4. Longshou Mine, Jinchuan Group Co., Ltd., Jinchang 737100, Gansu, China
Funds: National Natural Science Foundation of China (42061076); Excellent platform of Lanzhou Jiaotong University (201806)
Keywords: time series modeling, land subsidence prediction, deep learning, attention mechanism, long short-term memory
Citation format: Gao Motong, Yang Weifang, Liu Zuyu, et al. Surface Settlement Prediction Method of LSTM Goaf Combining Convolutional Neural Network and Attention Mechanism[J]. Bulletin of Surveying and Mapping, 2024(6): 53-58.doi: 10.13474/j.cnki.11-2246.2024.0610
Abstract:In order to solve the problem that the spatial features of monitoring points are difficult to extract in the time series prediction of surface subsidence areas in goafs, this paper proposes a CNN-Attention-LSTM combined neural network model that can extract the key spatial features of monitoring points. Firstly, the number of adjacent monitoring points as feature input was increased, and the convolutional neural network (CNN) was used to extract the spatial features of the multi-dimensional time series composed of multiple monitoring points. Secondly, the extracted multi-dimensional feature time series was input into the Multilayer Perceptron (MLP) to calculate the attention weight, and the Hadamard product was made with the feature input to realize the attention weight distribution of the feature input. Then, the long short-term memory neural network (LSTM) was used for regression prediction. Finally, through the fully connected layer, the predicted value of the output target monitoring point is integrated. In this paper, the surface subsidence area of the West No. 2 Mining Area of Longshou Mine is taken as an example, and the prediction results of the surface subsidence monitoring data are given and compared with the actual collected data. The results show that the combined model of CNN-Attention-LSTM with attention mechanism is more accurate than that of CNN-LSTM model and LSTM model, and the prediction accuracy of CNN-Attention-LSTM model can be significantly improved by adding effective feature input.
body
About author:Motong GAO (1999—), male, master student, main research direction is deep learning-based land subsidence prediction model. E-mail:[email protected] Corresponding author: Yang Weifang. E-mail:[email protected]
First instance: Yang Ruifang review: Song Qifan
Final Judge: Jin Jun
information