关键词: 系统辨识, 人工神经网络, 检验法, 多步广义梯度法, 遗传算法

Abstract: Partial differential equation models of some oil reservoir systems,such as well test interpretation could be transformed into series composed of non-linear function terms.Every term was a complex non-linear function of stratigraphic parameters θ.The number of term n known as structure parameter of the model was related to model structure.Functions as non-linear neural units was viewed to establish function link artificial neural network models of oil reservoir systems.F-test in system identification theory was used to determine the structure parameter n,and multistep generalized gradient learning algorithms to estimate the weighting coefficients of the network.Stratigraphic parameters were the foundation of well test interpretation,their unique estimate values were requested.The problem became more difficult because the above-mentioned aforesaid functions were multimodal functions and very sensitive in extreme point about the change of θ.Neither iteration methods nor genetic algorithms can work effectively.A new pattern of genetic algorithms was developed to solve the problem.Application has shown that the precision is high.Moreover unique estimate values of stratigraphic parameters were obtained.

Key words: system identification, artificial neural network, test, multi-step generalized gradient method, genetic algorithms