应用改进的神经网络学习方法预测储层参数

人工神经网络理论在石油科学的研究中具有重要的理论和现实意义。文章在分析了模拟退火算法和变尺度法各自的优势和原理基础上，针对前向网络反向传播算法（BP）收敛速度缓慢和易陷入局部极值点的缺点，将有全局寻优特性的模拟退火算法（SA）和快速收敛的局部寻优变尺度算法（BFGS）有效地结合，提出了一种快速、高效的前向网络混合学习策略，即SA-BFGS混合算法来训练网络。用它代替传统BP网络中的梯度下降法，通过训练网络权值，使网络具有较快的收敛速度和较高的逼近精度。在测井资料计算储层参数的实际应用中，该法能极大地改进前向网络的收敛速度与收敛性能，处理速度快、稳定性好、可信度高，具有较好的应用前景。

PREDICTING RESERVOIR PARAMETERS BY APPLYING THE MODIFIED NEURAL NETWORK LEARNING METHOD

The theory of artificial neural network is of great theoretic and practical importance to petroleum scientific research. On the basis of analyzing the advantages and principles of simulated annealing algorithm and variable schedule algorithm and in light of the forward network backpropagation algorithm’s defeats of slowly converging and easily sticking in local extreme point, a quick and effective forward network hybrid learning strategy, i.e. training up network by simulated annealing—variable schedule hybrid algorithm, was put forward by combining the overall searching simulated annealing algorithm with the quickly converging and locally searching variable schedule algorithm. The gradient falling algorithm in conventional BP neural network can be replaced by it, thus making the network being possessed of rapid convergence speed and high approximating accuracy by training up the network weight. Through the actual application to calculating reservoir parameters by use of log data, it was shown that this method could greatly improve the convergence speed and behavour of forward neural network, being quickly processing speed, good stability and high reliability, therefore it is of fairly good prospects.