The Application Research Of Supervised Learning Method In Short-Term Wind Power Prediction

Random, intermittent and fluctuation of the wind determine the randomness and volatility of wind power. When large-capacity wind power merge into the grid, if scheduling improperly, there is a great impact on the security and stability of power grid. While accurate wind power prediction is one of the key factors to ensure the stable operation of the grid. Short-term wind power forecasting referring to the prediction of the future power generation of 0-72h which has a great significance for power system scheduling, production and maintenance. Base on above, this article mainly fuses on the following aspects for further study of the show-term wind power prediction. They are wind speed and other meteorological data preprocessing, single power prediction method, combination forecasting method, and quantization of the uncertainty of power prediction, and so on.Firstly, Gaussian fitting, Fourier fitting and Least Square Support Vector Machine(LSSVM) were used to create a conversion relationship between wind and power. For over-fitting problem, method of increasing the amount of training data, data de-nosing process and regularization technology were introduced, and they effectively avoid the over-fitting phenomenon occur. Kernel method was introduced for the issue of the basic functions selection, and it solve the problem of difficult or impossible to choose the basic functions.Secondly, because the ridge regression based on kernel does not have a sparseness, support vector machine with sparse nature was introduced. At the same time, to get a more accurate description of the correspondence between wind speed and power, wind speed was divided into high, medium and low segment according to Weibull distribution of wind speed and wind power curve, then an integrated method for wind power prediction based on Gaussian model and LSSVM was proposed with considering the feature of each segment of wind speed.The distribution characteristics of scatterplot of wind speed and power show that, for the same wind speed at different time, the wind power distribute in a range rather than a fixed value. So Gaussian Process Regression (GPR) was introduced to forecast the distribution of wind power. Due to the instability and computationally intensive features of GPR, Bootstrap Aggregation (Bagging) and Fully Independent Training Conditional (FITC) Approximation method was raise to weaken the shortcoming of GPR. Simultaneously a new weight portfolio strategy based on Bayesian Committee Machine (BCM) was posed to combine the models which generated by Bagging, and the strategy effectively improve the accuracy of the combining algorithm.Supervised learning methods and statistical methods were applied for further research of the relationship between wind speed and power in the above problems. Binding experiments can draw the following conclusions, that is combination forecasting method can eliminate some of the deficiencies of the single method and effectively improve the power prediction accuracy. Moreover, the GPR method can produce the distribution of value of the power, in other word, it can either give an accurate prediction value, but also accurately quantify the uncertainty of the value.At last, the proposed methods based on the above studies were also integrated into the actual wind power prediction system and verified the effectiveness of the raised method. Moreover, a flexibility, scalability architecture of wind power prediction system and platform was proposed and designed on the basis of the existing wind power prediction system.