风电功率预测可以提高电力系统接纳风电的能力，改善电力系统运行的安全性和经济性。在大规模集群风电并网的背景下，针对风电的波动性和不确定性，本文深入研究了区域多风电场功率的时空相关特性，并通过概率预测来量化区域风电功率的不确定性。主要工作及成果如下：(1) 提出了采用规则藤Copula模型拟合区域多风电场各随机变量间时空相关结构的研究方法。对比相关系数难以描述相关结构的局限，以及传统Gaussian copula模型建模精度不足的缺陷，规则藤Copula模型可以显著提升了相关结构的拟合精度。(2) 提出了基于多元分布的区域风电功率概率预测方法。根据Copula理论建立区域风电功率的多元分布模型，并采用随机模拟法和条件筛选法构建预测功率的条件概率分布，得到了预测效果稳健的概率预测结果。(3) 提出了区域风电功率多时间断面场景集的生成方法，提高了区域风电功率时空关联特性的刻画能力。(4) 建立了区域风电功率概率预测直接法的预测模型。提出了距离权重的核密度估计方法，以距离核函数描述样本相似度的不同，相对于多元核密度估计法计算更高效，提升了概率预测效果。建立了基于ADMM算法优化的Lasso分位数回归概率预测模型，相对于传统分位数回归模型，该方法计算效率高，模型泛化能力强，尤其适用于高维大数据的分布式计算框架。经算例验证，本文所提方法在风电概率预测、相关结构建模以及功率场景集生成上均取得了较好的结果，适用于电力系统中风电功率的不确定性分析，而概率预测的结果可以有效提高电力系统的决策优化水平。

Wind power forecasting can be used to solve the problems brought by the higher penetration of wind power in the power systems. The safety and economy of the power systems can also be improved with the forecasting results. In the context of the regional wind power forecasting, the uncertain nature of the variable wind power and the spatial-temporal correlations of the regional wind power are studied. Different probabilistic forecasting methods are proposed to describe the uncertainty information of wind power outputs. The main work and contributions of the thesis are as follows.(1) The regular vine (R-vine) copula model is used to describe the spatio-temporal dependence structure for the variables in the regional wind power forecasting model. Since the correlation coefficients are unable to describe the dependence structure, while the Gaussian copula model is not accurate enough, the fitting accuracy of the dependence structure can be significantly improved by the proposed R-vine copula model.(2) The regional wind power forecasting model based on the multivariate distribution is proposed to provide the probabilistic forecasting results. The stochastic simulation and the multi-condition selecting method is applied to extract the conditional distributions. Compared with the traditional methods, the proposed method is more robust in performance.(3) The method of producing the predicted wind power scenarios is proposed to describe the spatio-temporal dependence among different time points and different wind farms.(4) Two different models based on the direct method are proposed to provide the probabilistic forecasting results. The first model is the distance-weighted kernel density estimation (DWKDE). The DWKDE model is more efficient than the traditional multivariate kernel density estimation because the simplified distance kernels are used to describe the similarities. The performance of the evaluation proves the effectiveness of the method. The other model is the quantile regression model with Lasso regularization, which is optimized by the ADMM algorithm. High efficiency is achieved due to the distributed learning scheme. Overfitting is avoided because of the embedded variable selection. This method is suitable for the high-dimensional applications in the framework of paralleling computing.The case studies show that the methods proposed in this thesis performs well in the applications of forecasting, which include the probabilistic forecasting of wind power, the estimation of dependence structures and the forecasting of the wind power scenarios. The high-quality forecasting results can be used in the analysis of the uncertainty information of wind power in power systems. Improvements can be achieved for many decision-making problems considering the probabilistic forecasting results.