Short-Term Wind Speed Forecasting Based On Data In Wind Farm

Energy is an essential material basis for human social progress and economic development. Traditional energy such as coal, oil and gases has not only limited quantity but also brings potential pollution to the human survival environment. In the past decades, the social development is affected by the oil crisis and various climate change factors, so the development of renewable energy ideas gradually becomes the consensus of the international community. Wind energy as a renewable energy is produced by the air flow acting. Our country is rich in wind energy resources according to the statistical results provided by the meteorological department the available wind energy has almost one billion kilowatts or even more. Accurate short-term wind speed forecasting (STWSF) can reduce the economic losses caused by grid integration at all levels of the transmission and distribution in grid. The distribution of short-term wind is randomness, instantaneity and seasonal, which has the inhibitory effect for the accurate wind speed forecasting. The sample used for this PhD thesis is based on the real data from a wind farm of East China. The main aspects of this paper are listed as follows:(1) Data preprocessing in wind farm. The noisy, missing values, inconsistent points and outliers are checked and processed to improve the accuracy and effectiveness of data analysis. In consistent with the problems in this data, such as how to select the proper threshold within the given windows to determine the analysis interval, eliminate the noisy data with less robustness meanwhile prevent the over-kill phenomenon in the data filter; how to use the performance parameters (including soft threshold, the filter scale, basis function and filtering level, etc.) to adjust the level of the Wavelet filter such that the Wavelet de-noising can be applied to eliminate the noise in a reasonable way.(2) Forecasting modeling framework. The forecasting modeling framework with respect to the time series based on the RBF network is analyzed. The partial correlation techniques are used to select the proper variables and eliminate the influence of the local extremes in order to obtain the variable with the proper combination structure which benefits the theoretical analysis. The optimization strategy based on structural risk minimization principle for the theoretical analysis and the implementation steps are given. Finally, the cross validation methods are used to verify the effectiveness of the proposed approach based on the RBF network.(3) Short-term wind speed forecasting modeling. The higher dimension sample result the analysis difficult and lower computational efficiency, the corresponding solution based on the forecasting modeling framework in combined with manifold algorithm is given. The model variable selection, model order estimation, parameter optimization, structure analysis and strategy optimization etc are applied for the STWSF. The model structure selection, computational efficiency, the sample complexity reduction, manifold algorithm for dimensionality reduction and structure optimization are discussed. The lengths of information criterion penalty term used for length design are applied to solve the practical issue, improve the foreasting efficiency, compact the structure and improve the generality ability.(4) Frequency decomposition and multi-model AdaBoost network ensemble. The wind speed can be divided into a variety of frequency superposition, so the Wavelet transformation based on the short-time Fourier analysis is used to decompose the wind speed. The proper frequency range of the wind speed will be established by the multi-resolution and multi-scales Wavelet transformation based on the elements transformed into time-frequency domain. The model structure with high accuracy and the high efficiency is founded. In addition, the AdaBoost algorithm based on the neural network are used to solve the ’disadvantages’ of the neural network such as over-fitting and easily to fall into local minimum. This benefits the generality ability improvement of the neural network. The finial experimental evaluation demonstrates the effectiveness and correctness of the utilized strategies.