Prediction Of Surface Settlement In Goaf Based On CNN-LSTM

The problem of land subsidence caused by large-scale underground mining in mines seriously affects the safety of mine production and personnel safety.Jinchang,Gansu is the largest nickel production base in my country,and its nickel ore resources account for 62% of the country's total nickel ore reserves.It is known as the "nickel capital of China".Since2016,continuous underground mining has led to the collapse and continuous subsidence of the stope in the second west mining area of Jinchang Longshou Mine.In order to ensure the production safety of the mining area and maximize the recovery of the remaining nickel ore resources,it is of great significance to predict the surface subsidence of the mining area.Most of the prediction objects in previous studies are for the surface subsidence of points,and the traditional models used have problems such as low prediction accuracy and difficulty in modeling.With the development of In SAR technology and artificial intelligence technology,based on the long-term,short-period,and large-area surface subsidence data of In SAR,combined with the powerful feature extraction and nonlinear fitting capabilities of deep learning to achieve accurate surface subsidence prediction in mining areas is the future.development direction.Based on the above background,this paper uses two deep learning models,CNN and LSTM,to form a CNN-LSTM model to study the surface subsidence of the second mining area in the west second mining area of Jinchuan copper-nickel mining area.The main contents are as follows:1.Based on SBAS-In SAR technology,the surface subsidence data of the study area during the period from 2019.3.22 to 2020.6.8 were obtained.2.Based on the K-means algorithm,the subsidence trend of the study area is analyzed by unsupervised clustering,and the study area is divided into: mild subsidence area(average subsidence rate is 0.6039mm/a),moderate subsidence area according to its different subsidence trends zone(average sedimentation rate is 1.9279mm/a)and heavy subsidence zone(average sedimentation rate is 3.5605mm/a).3.Use the deep learning framework of Matlab for model development and design,and build a surface subsidence prediction model based on CNN-LSTM.The model uses the data of the first 29 periods to predict the surface subsidence value of the next 8 periods.The main process factors are used as the model input and convoluted The neural network module performs feature extraction and then uses the LSTM network to predict the surface subsidence.4.Compare the three prediction methods of deep network prediction model based on CNN-LSTM hybrid,RNN prediction model and LSTM prediction model through MAE and MAPE two evaluation indicators.The experimental results show that: CNN-LSTM model,LSTM model and RNN model in the short-term validation set MAPE and MAE performance: 0.0250,0.995mm;0.0384,1.764mm;0.0502,2.422 mm.In the mid-term validation set,MAPE and MAE performance are: 0.0511,2.356mm;0.0656,3.708mm;0.0688,4.515 mm.In the long-term validation set,MAPE and MAE performance are: 0.0691,4.16mm;0.0839,6.313mm;0.1320,7.471 mm.In the overall validation set MAPE and MAE performance are: 0.0461,2.29mm;0.0638,3.630mm;0.0776,4.469 mm.Among them,the CNN-LSTM model performs the best in different validation sets and is the best one among the three models.5.The spatial and temporal distributions of the absolute and relative errors of the three models are analyzed,and the basic distribution laws of the relative and absolute errors of the models are obtained,as well as the main reasons for the abnormal distribution of errors.