Research On Prediction Of Surface Subsidence In Mining Area Based On Machine Learning Combination Model

The practice of China’s rapid economic development in the past few decades shows that coal-based fossil energy will still occupy a dominant position in China’s energy structure for a long time.In order to prevent the surface subsidence caused by large-scale coal mining,which affects the safety production in the mining area and the lives of the surrounding people,it is particularly important to obtain accurate and reliable information on the surface deformation in the mining area and further construct the mining subsidence model.Although some scholars have established various subsidence prediction models,most of the methods are to establish a single prediction model,which makes it difficult to effectively analyze the change trend of subsidence process,resulting in great differences in prediction performance.Instead,the combined models can fully consider the advantages,disadvantages and complementarity of each single prediction model,and the models are combined according to different types of settlement data to improve the prediction performance.Therefore,based on comprehensive analysis of the prediction performance,advantages and disadvantages of a single model,this thesis introduces the idea of combined model,and two new combined prediction models are proposed correspondingly in terms of two groups of data of mining subsidence monitoring.The main research contents and results of this thesis are as follows:（1）Based on the GNSS automatic monitoring data of the 110801 working face in the east 1^st area of Banji Coal Mine in Bozhou,a combined prediction model of GA-KF and PSR-BP-Adaboost based on wavelet analysis is established in this thesis combined with the idea of model parameter optimization and data preprocessing.Firstly,due to the good time-frequency capability with multi-resolution,the theory of wavelet transform is used to decompose complex settlement data into random sequence and trend sequence with different time-frequency characteristics.Then,in order to overcome the influence of unclear statistical characteristics of noise in the process of surface subsidence,the genetic algorithm（GA）is introduced to optimize the Kalman filter,and a GA-KF model is constructed to predict the trend sequence.Besides,in order to overcome the problems of poor generalization ability and easily falling into local minimum point of BP neural network,the Adaboost algorithm is introduced to improve the BP neural network and the PSR-BP-Adaboost model is established to predict the random sequence on the basis of constructing the training and prediction matrix by using the phase space reconstruction method（PSR）.Finally,the one-step prediction value of the combined prediction model of the original monitoring data is obtained by superposition of the results of trend sequence and random sequence.The results of the experiment show that the combined model fully extracts the complicated variation features of settlement data,integrates the advantages of single GA-KF model and BP-Adaboost model,effectively improving the prediction accuracy.Meanwhile,it is found that the modeling prediction effect varies with different lengths of monitoring sequence.Specifically,the combined model constructed in this chapter has the highest accuracy when the length of modeling sequence is 100,and the mean absolute error（MAE）and the root mean square error（RMSE）are only 0.138mm and 0.223 mm respectively,which is much smaller than the other two models and has a certain ability of anti-interference.（2）Based on the measured leveling data of No.1613（1）working face in the South3^rd area of Guqiao Coal Mine in Huainan,a combined prediction model of ARIMA modified by PSR-ISSA-SVM is established in this chapter combined with the concept of model parameter optimization and error correction.The model includes two key research contents:Firstly,the ARIMA model is used to predict the settlement data,and the nonlinear prediction residual sample is obtained.In the second place,the idea of error correction is introduced,and the nonlinear regression and correction on the prediction errors are carried out by using support vector machine（SVM）.Secondly,in order to improve the prediction performance of the model,the improved sparrow algorithm（ISSA）is adopted to optimize the super parameters of SVM model,and PSR-ISSA-SVM model is established to correct the prediction residual on the basis of constructing the input and output mapping matrix of SVM with PSR technology.The experimental results show that the prediction accuracy of PSR-ISSA-SVM model is higher than that of the PSR-BP model and PSR-SSA-SVM model,which also shows that the ISSA algorithm can effectively improve the search ability of the population and has better optimization performance.In the active period of mining subsidence,the settlement process is extremely unstable.The PSR-ISSA-SVM modified ARIMA combined model established in this thesis has greatly improved the prediction accuracy compared with other models.In the recession period,as the amplitude of surface subsidence becomes smaller and tends to be stable,the overall prediction accuracy of each model is relatively high,and the prediction accuracy of this combined model is not significantly improved,on the whole,however,the prediction performance of the combined model is optimal no matter in the active or recession period,and the prediciton accuracy improves significantly especially in the active period.Figure 37 Table 17 Reference 90...