基于SSA-BP神经网络的概率积分法预计参数求取研究

Estimated Parameter Extraction Research of Probabilistic Integration Method Based on SSA-BP Neural Network

为解决BP(Back-ProPagation,BP)神经网络求取概率积分法预计参数出现的局部最优解和收敛速度慢的问题,采用麻雀搜索算法(Sparrow Search Algortihm,SSA)优化BP神经网络的结构,得到最优的权重值和偏置项,建立了基于SSA-BP神经网络的概率积分法预计参数求取模型。结合50组典型的实测数据,随机抽取45组数据输入SSA-BP神经网络模型进行训练,剩余数据输入训练好的模型求取概率积分法预计参数,并与实测数据对比,分析SSA-BP神经网络模型和BP神经网络模型的优劣;通过改变训练样本和测试样本的数量,讨论模型精度与训练样本数量的关系。研究表明:①SSA-BP神经网络模型预计下沉系数q、水平移动系数b、开采影响传播角θ、主要影响角正切值tanβ和拐点偏移距s/H的平均绝对百分比误差分别为1.33%、3.48%、0.49%、3.86%和9.33%,BP神经网络模型的相应取值分别为8.05%、7.34%、3.33%、9.82%和19.60%,可见前者求解精度更高。②两种模型求取的预计参数均与实测数据较接近,SSA-BP神经网络模型最大相对误差为21.00%,最大迭代次数为89次,BP神经网络模型最大相对误差为35.00%,最大迭代次数为205次,说明优化后的BP神经网络模型预计结果精度更高,收敛速度快。③在样本总量不变的情况下,随着训练样本数量的增加,模型的预计精度和可靠性也相应增加。上述分析对于进一步提升概率积分法预计参数求取精度有一定的参考价值。

In order to solve the problem of local optimal solution and slow convergence speed of the expected parameters of the BP(Back-ProPagation,BP)neural network,the sparrow search algorithm(SSA)is adopted to optimize the structure of the BP neural network,obtain the optimal weight value and bias term,and establish a probabilistic integration method estimated parameter extraction model based on the SSA-BP neural network.Combined with 50 groups of typical measured data,45 groups of data were randomly selected and input into SSA-BP neural network model for training.The remaining data were input into the trained model to obtain the predicted parameters of probability integral method,and compared with the measured data,the advantages and disadvantages of SSA-BP neural network model and BP neural network model were analyzed.By changing the number of training samples and test samples,the relationship between model accuracy and the number of training samples was discussed.The study results showed that:①The average absolute percentage errors of the SSA-BP neural network model predicting the sinking coefficient q,the horizontal movement coefficient b,the mining influence propagation angleθ,the main influence angle tangent tanβand the inflection point offset distance s/H are 1.33%,3.48%,0.49%,3.86%and 9.33%,respectively,and the corresponding values BP neural network model are 8.05%,7.34%,3.33%,9.82%and 19.60%,respectively,which indicated that the solution accuracy of the former model is higher.②The estimated parameters obtained by the two models are close to the measured data,the maximum relative error of the SSA-BP neural network model is 21.00%,the maximum number of iterations is 89 times,while the maximum relative error of the BP neural network model is 35.00%,and the maximum number of iterations is 205 times,indicating that the prediction results of the optimized BP neural network model are more accurate and the convergence speed is fast.③When the total number of samples is unchanged,as the number of training samples increases,the prediction accuracy and reliability of the model also increase accordingly.The above analysis has a certain reference value for further improving the accuracy of predicting parameters of probability integral method.