Study On Very Short Term Wind Speed And Short Term Wind Power Prediction Based On Error Analysis

In recent years, with the rapid growth of new energy which mainly is wind power, large-scale wind power integration to a certain extent has eased the increasingly tense energy situation and environment pollution problem due to abuse of the traditional power to a certain extent. But with the constant increase of the penetration to the power grid, wind power has become a threat to the cooperative and safe operation of power grid because of its significant volatility and randomness. Wind power abandonment becomes more and more serious. Therefore, the high-accuracy wind power prediction is of vital importance. The study of wind power prediction at home and abroad is generally forcused on the model and algorithm, which can be divided into two categories:physical and statistics. The former uses the aerodynamic theory model to describe the variation of wind speed on a wind farm. The later extract information from the historical data and use it to predict. But in the existing methods, single-model-prediction is of low accuracy and not stable. Combined model depends on the proper weight coefficient to sum up each model result to ensure accuracy.The weight coefficients are calculated by individual algorithm or fitting model. At the same time, analysis and research on the wind power prediction error is rare, the distribution of wind power prediction error needs further research. To solve the above problems, this paper proposes a method of building the combination forecast model based on the error analysis and feedback:on one hand, to study the wind power prediction error, on the other hand, to propose a good way to build a combined model to improve the prediction accuracy.First, this paper introduced a predict model based on the popular support vector machine (SVM) algorithm, combined with the wavelet transform for data processing, the grid search method for parameters optimization and the SMO algorithm for rapidly SVM solving. The prediction model achieved an MAE 11.8% for very short-term wind speed prediction and 10.72% for short-term wind power prediction, which can meet the requirements of engineering application. But the algorithm principle of support vector machine (SVM) inevitably lead to instability of prediction accuracy. Neglect of the support vector reduces its ability to predict the extremum values.Then, using SVM model to predict to history data to get the very-short-term wind speed prediction error sequence and short-term wind power prediction error sequence. Analyze the time-variant characteristics and correlation characteristic of very-short-term wind speed prediction error, the orrelation characteristic and distribution characteristics of short-term wind power prediction error to find out the chanding laws of prediction errors based on time and other relevant factors and get the probability of error distribution.Finally, error analysis feedback prediction models was built based on Neural Network for very-short-term and Gaussian Process Regression for short-term. Feedback models used SVM prediction results and historical datas as inputs according to the error analysis conclusion. As is tested in the simulation example, very-short-term wind speed prediction error was reduced by 15.8% on average, short-term wind power prediction error was reduced by 11.3% on average. The promotion effect is remarkable. Models ased on Gaussian process regression can output probability prediction results at the same time. The error analysis results can help the power grid with operation scheduling and security analysis.