基于改进RBF神经网络的巷道变形预测模型

Deformation Prediction Method of Roadway Based on the Improved RBF Neural Network

由于经典RBF神经网络中的隐含层节点数、连接权值等结构参数基本由经验获取,因此经典RBF神经网络模型的性能取决于建立模型专家的主观性,存在一定的盲目性和随机性,难以对巷道变形进行准确预测。为此,采用贝叶斯阴阳和谐学习算法对经典RBF神经网络模型的隐含层节点个数、连接权值等结构参数进行了优化,提出了一种基于改进RBF神经网络的巷道变形预测模型,即对角型广义RBF神经网络模型。采用潞安和兖州矿区的综放回采巷道的现场长期监测数据分别对经典RBF神经网络模型以及对角型广义RBF神经网络模型进行了试验分析,结果显示:1对巷道顶底板变形进行预测时,对角型广义RBF神经网络模型的准确率约92.2%,经典RBF神经网络模型的准确率约80.6%;2对煤帮变形进行预测时,对角型广义RBF神经网络模型的准确率约90.2%,经典RBF神经网络模型的准确率约78.6%。上述试验结果表明,对角型广义RBF神经网络模型对于巷道变形预测的精度明显优于经典RBF神经网络模型,对于高精度巷道变形预测有一定的参考价值。

The structural parameters of the hidden layer node number and connection weights of the classical RBF neural network model are obtained by experiences,so,the performance of the classical RBF neural network model is depends on the subjectivity of experts,that is to say,there are certain blindness and randomness are existed,which lead to predict the roadway deformation with great difficult. In order to improve the performance of the classical RBF neural network model,the structural parameters of the hidden layer node number and connection weights are optimized by adopting the Bayesian ying-yang harmonly learning algorithm,a new roadway deformation prediction model based on improved BRF neural network is proposed,the improved neural network model can be names as diagonal type generalized RBF neural network model. The experiment is done based on the on-site long-term monitoring data of the fully-mechanized sublevel stoping roadways of Lu ＇an mining area and Yanzhou mining area to analyze the performance of the classical RBF neural network model and diagonal type generalized RBF neural network model,the results show that：1 the prediction precise of roof and floor roadway deformation of the diagonal type generalized RBF neural network model is about 92. 2 %,while the prediction precise of the classical RBF neural network model is about 80. 6 %;2 the prediction precise of coal side wall deformation of the diagonal type generalized RBF neural network model is about 90. 2 %,while the prediction precise of the classical RBF neural network model is about 78. 6 %. The above experimental results further show that the prediction of roadway deformation can be done with high precise by the diagonal type generalized RBF neural network model,which can provide some reference for the prediction of roadway deformation with high precise.