Research On Some Issues Of Short-term Wind Speed Forecasting For Wind Farms

With the continuous decrease of the traditional fossil energy and the gradual deterioration of ecological environment, it has been crucial to develop renewable energy to cope with the energy crisis, improve the environment, and realize the sustainable development of human society. Among the various renewable energy choices, wind energy is under rapid development due to its irreplaceable features and advantages. The development and utilization of wind resources is meaningful since it can effectively improve the energy crisis and reduce the environmental pollution. At present, wind energy is used by integrating wind power into power grid. However,due to the intermittency, volatility and randomness of wind energy, the output of wind power will become unstable, which poses severe challenges to the safe and stable operation of power system. In order to solve this problem, one of the effective ways is to forecast wind speed/power for wind farms. According to the research status of short-term wind speed forecasting, this paper studies methods for reducing the error of point prediction and quantifying the uncertainty of the prediction,using real wind speed data from wind farms in Jiangsu, Ningxia and Yunnan provinces. Specifically,the main research work of this thesis include:(1) Researches on modeling and forecasting of real wind speed data are conducted. Three mainstream models in the field of short-term wind speed forecasting are used, including time series models, neural networks and support vector machines.The results of one-step-ahead and multi-step-ahead prediction are compared and analyzed. Two mechanisms, namely, iterative and direct for multi-step-ahead forecasting are adopted and compared. The experimental results indicate the limitations of the single forecasting models and the necessity to study combined forecasting methods.(2) A novel weighed combination forecasting method is proposed to reduce the risk of model selection and improve the accuracy of prediction. The proposed method is based on ensemble learning, which combines time series models, neural networks and support vector machines. The weights for the component models are determined by considering the historical performance of each model and the generalization capability of the ensemble. The effectiveness of the proposed method is assessed on the real wind speed data sets.(3) A new hybrid model, which is based on signal decomposition and machine learning, is proposed to improve the acccuracy of point prediction. This method first uses the empirical mode decomposition algorithm to decompose the original wind speed series into several sub-series. And then the original features are constructed from these sub-series. After a feature selection process,relevant and informative features are selected to build a predictive model. The experimental results indicate that compared with the single forecasting models, the proposed hybrid model can improve the forecasting accuracy significantly.(4) In order to quantify the uncertainty associated with point predictions,a novel method based on radial basis function neural networks and non-dominated sorting genetic algorithm-? is proposed to directly generate the prediction intervals. The construction of prediction intervals is formulated in a multi-objective optimization framework, which aims at concurrently minimizing the width and maximizing the coverage probability of the constructed intervals. In order to reduce the dimensionality of the parameter space and simplify the complexity of the search process, a two-step method is proposed to train the predictive model. The performance of the proposed method is verified on actual wind speed data sets.(5) A new Gaussian process regression based hybrid approach is proposed for short-term wind speed probabilistic forecasting. In the proposed approach, the autoregressive model is employed to capture the overall structure from wind speed series, and the Gaussian process regression is adopted to model the residuals.Different types of covariance functions are combined to capture the characteristics of the data and automatic relevance determination is used to take into account the relative importance of different inputs. The forecasting results indicate that the proposed method can not only improve point forecasts, but also generate satisfactory prediction intervals.