Feedforward Fuzzy Neural Network Improvements In The Oil Industry

Feedforward fuzzy neural network is the result of organically integrating fuzzy technology and feedforward neural network. To deal with uncertain complicate problems, it behaves the strong superiorities. Rut there also exist some imperative problems which haven't been resolved, such as the problem of optimizing network structure, the design of learning algorithms, the improvement of convergence speed and error precision, the decrease of anthropic factors and so on. Firstly, this thesis systematically analyzes all aspects of feedforward fuzzy neural network and puts forward some improved measures. Then two improved feedforward fuzzy neural networks are presented and applied to resolving the practical proWem in petroleum industry. One is that a RFB(Radial Basis Function) fuzzy neural network is improved. Ry analyzing the characteristic of RFR and the essence of adjusting the weights with the gradient descending method, a new learning algorithm that can produce the hide layers' nodes with a self-organized mode is presented. Tn its applications, a simulative system based on RFB fuzzy neural network is accomplished for predicting product rate of catalytic cracking. A group experimental data coming from heavy oil's catalytic cracking reactive experiments on the mini-type catalytic cracking equipment are applied to analyzing the system capability. Observing the total error curve, the convergent speed is faster, and the error precision is higher Then a group untrained experimental data are input the simulative system in turn. And their predicting results are satisfying. The other is an improved fuzzy neural network that is joined a compensative operation layer and adopts the improved center-or-gravity defuzzication in feedforward fuzzy neural network based on Gaussian fuzzy logical system. Particularly, a mechanism that can dynamically adjust the learning step is adopted. So the network can avoid the strong sway phenomenon and the convergence speed is also speeded in the learning process. The method is applied toconstructing the prediction model of the reservoir water saturation in logging interpretation of oil field. When analyzing network's capability, the new method and the method based on conventional BP algorithm are respectively applied to predicting reservoir water saturation in Lin Pan-Shui oil field. The result shows that the former is more improved than the latter in the convergent speed and the error precision. To summarize, two kinds of improved feedforward fuzzy neural networks have some values in theory and application and is worth further extending.