Theory And Method Of Probabilistic Wind Power Generation Forecast

The large-scale wind power integrated into grid eases the pressure on China’s energy and brings huge economic and environmental benefits. Wind power is currently the renewable energy with most mature technologies and most large-scale development conditions. However, as a kind of intermittent power, the integration of wind power into grid is bound to increase the difficulty of power system operation and control, as well as the heavier burden of spinning reserve. Therefore, the forecast of wind farm’s output or a group of wind farms’output is of great importance.The short-term wind power forecast is to predict the output of wind turbine or wind power in the future24hours～72hours. The forecast results can be used for optimizing the output of conventional unit and system reserve, improving the safety and economy of system operation. According to the forecast results, the short-term wind power prediction methods can be divided into single point prediction and probabilistic forecast. Single point prediction methods provide the forecast wind power with the maximum probability for the future time periods. The prediction error (taking the normalized average of the48hours average absolute error as an example) is ranging from15%to40%. Since point forecast cannot provide the uncertain information of wind power forecast, in recent years, probabilistic forecast method has received increasing attention and research. This kind of methods can forecast not only the expectation value of wind power also the distribution information of the prediction error, which provides an important reference for the operation risk and the risk decision of power system integrated with wind power.Currently, there are several following issues of short-term probabilistic wind power generation:①the need to further improve the prediction accuracy, including expectation prediction accuracy, distribution prediction accuracy and so on;②The existing methods are mostly for each look-ahead time only, without considering the correlation information between the time periods is ignored from forecast results;③in a region with multiple wind farms, the forecast of wind power generation ignores the characteristics of dynamic temporal and spatial correlation among wind power generation forecast errors of multiple wind farms.In this paper, Support Vector Machine is firstly used to forecast single point value of wind generation for each wind farm, and the probability density function (PDF) of prediction error is forecasted by Sparse Bayesian Learning, then one parametric approach is proposed for forecasting the marginal PDF of wind generation with error correction. Secondly, the statistical property of prediction error of wind power generation is analyzed, finding the existence of time domain correlation information of wind generation. Then the correlation coefficient matrix of wind generation forecast errors during forecast periods is estimated for quantitative analysis of the time domain correlation information of wind generation prediction errors. Combined with the marginal PDF of wind generation forecast, an approach for short-term joint PDF forecast of multi-time-interval wind generation is also provided. With further analysis of forecast errors of multiple wind farms outputs, the existence of the spatio-temporal cross correlation among the wind generation forecast errors has been found. And a dynamic conditional correlation regressive model is used to estimate the dynamic conditional correlation matrix which can describe the quantitative relation of temporal and spatial correlation. Finally, with the combination of PDF of each wind farm output power and dynamic conditional correlation matrix, the joint PDF of multiple wind farms output power is formed, and it is further transformed into multi-dimensional scenarios using multivariate random variable sampling skill. Test results illustrate the efficiency of the method.