一种高精度的概率积分法参数预计方法

为使概率积分法参数预计具备更高的精度,在充分分析概率积分法参数与地质采矿条件间关系的前提下,整合遗传算法和BP神经网络的功能特性构建了一种新的网络模型,即NA(new approach)模型。该模型利用遗传算法的寻优能力获得网络最优输入自变量组合,并优化模型的权值和阈值。以中国45个典型的地表移动观测站数据作为训练和测试集,分别建立NA、BP网络和SVM模型,并将模型预测结果与实测数据做了对比分析。结果表明:对不同的概率积分法参数预计时,NA模型的预测精度都优于BP网络和SVM模型;且误差波动范围小、稳定性高。随预计参数的不同,BP网络和SVM模型预测精度各占优势,但两者预测效果相差甚微。

A high-precision calculation method of probability-integral method parameters

In order to make prediction parameters of probability-integral method with a higher precision, on the premise of intensive analyzing the relationship between probability-integral method parameters of surface subsidence and geologic and mining conditions, a new artificial neural network was put forward, named New Approach (NA) model, which is established by integrating the functional characteristics of genetic algorithm and BP neural network. And it is used the optimization ability of genetic algorithm for finding the optimal input and independent variables of the network, and the optimized weights and thresholds of the model as well. Then the NA model, BP neural network model and SVM model were established, respectively, based on the training and testing data which is derived from 45 typical data of surface moving observation stations. Analysis was conducted by comparing the predicted results and the observed values with each other. The results indicate that NA model has higher precision than BP neural network and SVM model during arbitrary parameter of probability-integral method prediction. It also has smaller error range and higher stability then others. Besides, when compared the BP model and SVM model prediction results, it can be find that their have tiny differences in accuracy.