Wind Farm Modeling Based On Artificial Neural Network And Its Simulation

With the clean, economic and other advantages, wind power industry as the vital part ofnew energy power generation, has been rapid development in all countries. But over the pasttwo years, the pace of development slowed down, which is due to technological progresscan’t catch the scale of development. However wind power model is the basis of the technicalstudy, this research was extremely important in the moment. Mechanism method is often usedas a current wind power modeling, but the strong nonlinearity and strong coupling of windfarm determines the low-precision, and complexity of mechanism modeling. So I consider thetest method for modeling based on artificial neural network, which can approximate anynonlinear function with the subsystem decoupling problem. It’s easy to implement, providingaccurate model to facilitate analysis and control. The main research contents and conclusionsof this paper are the following three parts：（1）First, the use of back-propagation (Back-Propagation, BP) neural network initiallyfit the static nature of the wind farm model. This neural network is suitable for modeling asthe basic network model, but standard BP network has many problems. Therefore, Elmanneural network which has more complicated structure is obtained with improved algorithm, asa dynamic real-time simulation model. This wind farm model is more accurate and can welldescribe the dynamics character, not limite to the role of a parameter.（2）Support Vector Machine (SVM) based on statistical theory is adopted to establishthe system model. This is not only a feed-forward network, but also as a dynamic model afterimproving the input-output model and adding the delay operator. The net with stronglocalized radial basis kernel function can predict the output more accurately. Simulationresults indicate that the final SVM network can use less data to establish more accuratemodels, and can avoid the complex structure of neural networks and the problem of easilytrapping in local minima.（3）The paper based on Bayesian statistics assessed the weights of uncertainty as anevaluation criteria for network performance. Because neural networks for modeling have awide range, there is no uniform standard for the evaluation of quality. Proposed a criterionbased on the qualitys of weights from the first two chapters was obtained by Bayesianstatistical methods. This value called posterior probability distribution fuction is a measure ofthe standard. Finally I found that the performance of SVM network is better than the twotraditional neural networks’.