Optimization For Very Short-term Wind Power Forecasting Algorithm

The intermittence and fluctuation of wind power have brought difficults and challenges to grid dispatch, and wind power forecasting is the effective soltuion to this problem. In this thesis, a wind farm in North China and a wind farm in Texas U.S.A. are taken as study objects for very short-term wind power forecasting algorithm optimization based upon historical operation data and meteorological data.Based on the theoretical study on power curve characteristics of wind turbine generator systems and the nonlinear abilities of support vector machine (SVM), Piecewise Support Vector Machine (PSVM) model is proposed for the intended application. According to the study results, the Mean Relative Error (MRE) of PSVM model is17.50%, which is4.76%lower than SVM model; the Genetic Algorithm (GA) is applied in the forecasting model with the aim of parameter optimization; wind power time series data is taken as signal, and with the decomposition process of wavelet transform, the WT-LSSVM model is built up. The case study show great performance of the proposed model while huge time loss for computing.Neural Network has the parallel processing of information and self-learning ability, which is suitable for forecasting. Based on those abilities, Back Propagation Neural Network (BPNN) model and Radial Basis Function Neural Network (RBFNN) model are developed for comparison purpose. The RBFNN forecasting model has better accuracy and wider applications than that of BPNN model. Compare to the forecasting error of BPNN, the error of RBFNN decreases0.30%(Mean Absolute Percentage Error, MAPE) and0.14%(Mean Square Percentage Error, MSPE) respectively. Hilbert Huang Transform (HHT) is utilized in wind power time series data for model training, and the HHT-ANN model is obtained based on HHT and RBF neural network, which show the potential of forecasting accuracy improvement with the MAPE decreasing from3.39%to1.89%.To realize the steady and efficient wind power forecasting along with the advantages of both individual forecasting models, which are Least Square Support Vector Machine (LSSVM) and Radial Basis Function Neural Network (RBFNN), a very short-term hybrid wind power forecasting model is developed by means of Grey Relational Analysis. Compared to LSSVM and RBFNN, the results show0.29%and0.4%reduction on MAPE respectively. Based upon the analysis of wind speed distribution features, the weights of each single model are pre-calculated according to different wind speed classifications. This forecasting model has wider application and faster computing speed. The case study results show that the proposed hybrid model has outperformed individual model approach. Its MAPE and RMSE are2.37%and 3.79%respectively.In addition, there are always errors in forecasting models. Therefore it is acceptable to develop the probability forecasting. The forecasting error distribution characteristics are studied and a Monte Carlo Method is proposed. Its practicability is verified. The wind power forecasting algorithm and the research results of uncertainty analysis are applied in wind power forecasting system.