Study Of Short-Term Power Prediction In Wind Farm

Under the background of promoting energy conservation and emissions reduction vigorously all over the world, the renewable energy represented by wind has been widespread concerned. China has abundant wind energy resources, and the wind power industry is flourishing in recent years. However, with the construction of large scale wind power base and the introduction of wind power to the power system, the problems brought by the volatility of wind power to the power system have become increasingly prominent, restricting the healthy development of wind power seriously, and a lot of areas even have to switch out power brownouts. In order to reduce the adverse impact to the power system because of the introduction of wind power to the power system and ensure that the wind power technology gets benign large-scale development, reasonable scheduling and stable operation, the wind power prediction technology is particularly important.With the prediction time increasing, the prediction models’accuracies decline rapidly which are based on traditional statistical prediction methods, and single machine learning method can not meet the requirement of accurate prediction, the prediction errors are too large at shock point. In order to improve the prediction accuracy of model, this paper presented a prediction model based on the combination of wavelet analysis and artificial neural network. First, use the wavelet analysis smart algorithm to decompose historic power series into approximation signal and detail signal, thus get some strong regularity subsequences. In order to retain useful information and remove high-frequency noise, use the wavelet analysis algorithm again to denoise the low-level detail signal components with more high-frequency noise. So far, we could establish the nonlinear auto-regressive dynamic neural network prediction model based on the high-level details of the signal, the denoised low-level details and the approximation signal. Finally, composite all layers’ prediction value and get the final prediction power value. The instance simulation’s error analysis results shew that this modle’s prediction accuracy was significantly improved compared with the single neural network model and the combination method without wavelet denoising.In recent years, support vector machine has become a new hot spot in machine learning, it is suited to deal with the small sample problem, its training speed is faster than the neural network, thus this article established another wind power short-term prediction model using support vector machine algorithm. However, the ultra parameters of support vector machine and kernel function determine the prediction ability of the prediction model, for this problem, this paper used grid search algorithm and particle swarm optimization algorithm to select the model’s ultra parameters. Taking into account that different ultra parameters could get the same training error, but when apply them to the test model the test error would be different, in order to overcome this problem this paper has made improvements to the grid search algorithm and particle swarm optimization. Because the power prediction sequence lags behind the actual power sequence in this modle, the prediction error got too large, to resolve this problem this model used the preprocessing results of the original nonlinear power signal in the first modle and eventual eliminated the lag. In addition, from the point of correcting the prediction results, the error correction strategy based on the error sequence of the prediction results was proposed in this paper. Most of the traditional prediction systems only considered preprocessing the model inputs-historical power sequence, but neglected the law of the prediction results’error sequence, this idea has a certain significance to the post-processing.