动态概率神经网络及油气概率分布预测

Dynamic probabilistic neural network and predicting probabilistic distribution of oil/gas.

传统概率神经网络(PNN)采用了前向径向基函数神经网络的局部化的、高斯型作用函数,虽然具有较好的分类能力,但也存在以下局限性:①如果学习样本增多,模式层的神经元个数也增多,导致运算矩阵的增大而使之几乎丧失处理大数据量实际资料的能力;②模式层到求和层的权值固定为常数,必然要求学习集合中各类样本的数量相等,从而会影响其接受实际数据的能力。为此文中引入了动态概率神经网络(DPNN),它与传统PNN在结构上的区别在于:①模式层到求和层不以等权连接,权值由学习样本集的概率分布所确定;②学习样本集中各类样本的数量(各类在模式层的神经元个数)可以不相等。文中还介绍了DPNN算法。理论数据测试充分展示了DPNN结构具有动态调整、学习收敛快速、分类识别能力强大等特点。通过选取G油田实际资料的22种属性作为网络的输入向量,用DPNN进行分类识别得到了含油气概率分布图,可为预测有利油气圈闭及油水分布规律提供依据。

Traditional probabilistic neural network (PNN) adopted localized and Gauss-style's operational function of Radial Basis Function that has better categorizing ability but exists the following limitations:①the number of neural cells of model increased if studying samples increased, which leads operational matrix to increase and makes it almost lose the ability of large data volume processing ;② The weight from model to summation is fixed as constant that needs the number of all kinds of samples in studying sets to be equal,which affects the ability of receiving real data. In this reason, the paper introduced the Dynamic Probabilistic Neural Network (DPNN),which is different from traditional PNN in structure as follows:①using unequal weight to connect the model and summation, the weight is determined by probabilistic distribution of studying sample sets;②the number of different kinds of samples in studying sample sets (the number of neural cells lain in model for different kinds) can be unequal. The paper also introduced DPNN algorithm. The theoretic data tests showed the DPNN structure is characterized by rapid dynamic adjustment and studying convergence as well as strong ability of identifying categories. The 22 attributes of real data in G oilfield are selected as input vectors of network, the oilbearing probabilistic distribution map is obtained by using categorized identification, which can provide the basis of predicting prospective oil/gas traps and distributing rule of oil/water.