| The mining subsidence prediction model and parameter system is the core theory underlying the design of "under three" coal mining.Accurate and reliable inverse probability integral parameters are a prerequisite for accurate surface subsidence prediction,and the construction of surface subsidence prediction models is a difficult task in mining,but is important for the prevention and control of geo-environmental disasters caused by mining subsidence.The current literature shows that scholars have established probability integral parameters and mining subsidence prediction models,but there are still some problems:(1)In the process of inversion of probability integral parameters from real measurements,the standard genetic algorithm(SGA)for probability integral parameters has the disadvantage of easy premature convergence and easy to fall into local optimal solutions at a later stage.(2)For mining areas where there are no observation stations,the BP neural network does not take into account the correlation of the probability integral parameters,resulting in a reduction in the accuracy of the probability integral parameters,which does not match the final subsidence and horizontal movement curves;(3)For the prediction of a single subsidence point,the BP neural network tends to fall into a local optimum solution,while not taking into account the residuals resulting in reduced accuracy.(4)The Huainan mine is a high water level mining area,and as mining proceeds,water will flood the monitoring points resulting in the monitoring points being unmeasurable.In the prediction process of underwater subsidence points by support vector machine(SVM),the single kernel function has low generalization ability and does not take into account the residuals,which leads to the reduction of accuracy.In response to the above problems,this paper has conducted some research and the following conclusions have been obtained.(1)A multiple population genetic algorithm(MPGA)is proposed to invert the probability integral parameter using the idea of weighting.In the process of inversion of the probability integral parameter from the measured values,in order to compensate for the shortcomings of early convergence of the standard genetic algorithm(SGA)for a single population to find the probability integral parameter and the tendency to fall into a local optimum solution at a later stage.A multi-population genetic algorithm(MPGA)is proposed to invert the probability integral parameter,and the accuracy and reliability of the algorithm are investigated.Simulation experiments show that the MPGA-based probabilistic integral parameter inversion model is not only able to accurately find the expected parameters,but also has strong resistance to random errors and missing monitoring points in the observatory data.Experiments on the probabilistic integration method for the 1222(1)working face of Zhujidong coal mine showed that the standard deviation of the fitted subsidence and horizontal movement values from MPGA was 31 mm,while the standard deviation of the fitted subsidence and horizontal movement values from SGA was 32 mm,after only 57 iterations of MPGA and 100 iterations of SGA,indicating that the MPGA-based model is more accurate than the SGA-based model.This shows that the MPGA-based parametric model has certain advantages over the SGA parametric model in terms of accuracy and number of iterations.(2)A new and improved neural network probabilistic integral parameter prediction algorithm is proposed.In order to improve the accuracy of predicting probability integral parameters for the Huainan mining area lacking observation stations,a new and improved neural network probability integral parameter prediction algorithm(NIBPNN)is proposed.It not only considers the correlation of probability integral parameters,but also considers the accuracy of probability integral parameters.The 105 mining areas in China were used as the training set and 4 Huainan mines were used as the test set to predict the parameters of the probabilistic integration method using the above model,and the accuracy of the model's predicted parameters,the median error of the subsidence and horizontal movement values were analyzed,and the results showed that the NIBPNN was a significant improvement over the BPNN and IBPNN.(3)A surface subsidence prediction model with a BP strong predictor(BP-Adaboost)incorporating chaotic residuals weighted by multiple BP neural networks is proposed using the idea of weighting.To improve the accuracy of the surface subsidence prediction results caused by underground mining,the BP-Adaboost surface subsidence prediction model with a BP strong predictor incorporating chaotic residuals was proposed due to the high accuracy of BP neural networks in single-point subsidence prediction but the tendency to fall into local optimal solutions.The BP-Adaboost model with fused chaotic residuals,the BP neural network model and the BP-Adaboost model were used to predict the maximum subsidence value point in the stable and active periods in a single step and a multi-step prediction,respectively.The results showed that the BP-Adaboost model with fused chaotic residuals has the highest accuracy in both single-step and multi-step prediction,especially in single-step prediction.(4)A multi-kernel support vector machine(CHAOS-GA-MK-&-SVM)underwater prediction model optimized by genetic algorithm with chaotic residuals using the idea of weighting is proposed for multiple kernel functions.As the Huainan mining area is a high water level mining area,water will flood the monitoring points as mining progresses leading to unmeasurable monitoring points and SVM is more accurate in underwater monitoring point subsidence prediction.A genetic algorithm optimized multi-kernel support vector machine(CHAOS-GA-MK-?-SVM)underwater prediction model with chaotic residuals is proposed.The CHAOS-GA-MK-?-SVM model was used to predict the underwater subsidence values based on the measured data in the without water subsidence area,and the accuracy and feasibility of the prediction results were analyzed.The 1312(1)first mining face of the Huainan Gubei coal mine showed that the CHAOS-GA-MK-?-SVM model was more accurate with more complete data from the monitoring points,four of the six predicted values in the active period were available,the mean relative error(MRE)of the first four periods was 1.7%,which was more accurate and could be used,while the MRE of the last two periods was 9.6%,which was less accurate and could not be used.Xieqiao 2111(3)showed that the CHAOS-GA-MK-?-SVM model has high accuracy and can be used in all six predicted periods of the stability period with an MRE of 0.9%for the six periods when the data at the monitoring sites are more complete.However,in the case of severe data deficiencies at the monitoring sites,the lack of training samples will lead to overfitting of the CHAOS-GA-MK-?-SVM model,and a genetic algorithm optimized support vector machine(GA-?-SVM)model should be used to predict the underwater subsidence values.The GA-?-SVM model is available for all six predicted periods,and the MRE for the six periods is 3.3%.Figure 23 table 27 reference 81... |
PDF Full Text Request