关键词: 神经网络, 因素有效性, 结构, 腐蚀, 预测

Abstract: Wavelet neural network model was obtained to predict tubular corrosion rate for oil-gas gathering and transferring under multiphase dynamics condition. The laboratory multiphase dynamics corrosion experiment was adopted firstly in order to gain coupon corrosion rate in conditions of various operating situation for learning and testing of wavelet neural network prediction model. Then, a multi-factors analysis of variance was used to research the influence on corrosion rate of temperature, pressure, velocity, hydrogen sulfide content, carbon dioxide content, dissolved oxygen content, moisture content, salt content and pH to fulfill factor validity filtration. Finally, the suitable wavelet neural network prediction model was established by the way of learning and testing in the basis of recognition of neuron number of hidden layer. The model reliability was further verified. The result showed that above factors except pressure had great impact on corrosion rate and were considered as effective input signals. 8-17-1 type of wavelet neural network structure exhibited the favorable accuracy and stability when neuron number hidden layer was 17. Levenberg Marquardt optimization algorithm was chosen to train model repeatedly until its root mean square error less than convergence tolerance 0.001. The predicted value was approximately linear with the experimental value. The determination coefficient of learning stage and testing stage was 0.9992 and 0.9967, respectively, and demonstrated the superior correlation. There was also no significant difference between model predicted value and verified one. Therefore, the prediction model of wavelet neural network possessed well capacity of tubular corrosion rate for oil-gas gathering and transferring under multiphase dynamics condition.

Key words: neural network, factor validity, structure, corrosion, prediction