由于环境污染问题在全球范围的日益加剧，以风能、太阳能等可再生能源为先导的新能源发电技术已经成为能源革命的发展趋势。得益于风能自身清洁和资源充足的特性，风力发电技术已经对原有污染较大的发电技术产生了强有力地冲击。然而，风资源也存在着波动性、随机性、不确定性的缺陷，极易造成风机运行中的脱网、弃风现象，给风电并网、风机的稳定运行带来了极大的挑战。在此应用背景下，本论文计及风资源的不确定性对风机稳定、安全运行所产生的影响，在风机运行数据异常识别、多时间尺度风速预测、多时间尺度风电功率预测、风机健康状态评估和多风场集中监控与预测平台的搭建五个方面进行了研究。具体研究内容如下： (1)开展了风机运行数据异常识别的研究。首先针对风机结构复杂、运行数据维度较大，易造成维度灾难的问题，结合深度置信网络(Deep Belief Network，DBN)在处理非线性映射问题的优势，提出了一种基于改进深度置信网络的特征提取方法；在降低属性维度后，建立了基于密度聚类(Clustering In Quest，CLIQUE)算法的无监督异常数据识别模型，通过比较网格节点密度与设定阈值之间的大小，将风机运行数据分为正常运行数据、缺失数据、堆栈异常数据以及离散异常数据。 (2)开展了多时间尺度的风速预测的研究。首先，对风速时间序列进行波动性分析，实现风速时间序列的最优分组；之后鉴于栈式降噪自动编码机(Stacked Denoising Autoencoder，SDAE)具有强大地处理噪声、异常特征提取的能力，利用SDAE学习原始风速时间序列下的隐藏特征；最后，单回归器在预测的过程中存在泛化能力较差的问题，采用Bootstrap方式抽取样本，搭建集成极限学习机预测模型，实现多时间尺度的风速预测。 (3)开展了多时间尺度的风电功率预测的研究。在上一节风速预测的基础上，构建风电监测数据特征优选策略，将原始输入空间映射到低维表达空间，避免过拟合问题；引入随机梯度下降法和误差权值调整方法，建立了变权随机森林的预测模型，对预测结果进行误差反馈，实时更新误差权值，提高风功率预测的准确率。 (4)开展了风机运行健康度评估的研究。由于风机结构的复杂性、模糊性，风机健康度的变化具有多态渐进性的特点。因此本文选取特定的风机监控数据，将变点算法嵌入到循环神经网络(Recurrent Neural Network，RNN)中，基于“离线分类，在线匹配”的思想，在避免过拟合的基础上挖掘原始的传感器序列数据变化规律，进而预测下一时刻的属性数据；建立了加权评估模型，结合多个重要部件的预测残差评估风机的健康状态。 (5)搭建了多风场集中监控与预测平台。本文依据不确定性特征挖掘的风机运行分析及预测关键问题的理论研究成果，采用Java开发平台、Oracle数据库等工具开发多风场集中监控与预测平台，促进了本文所研究的不确定性特征挖掘的风机运行分析及预测理论工作向工业化应用的转化。 关键词：异常数据识别；风速预测；风电功率预测；健康度评估；深度学习

 As environmental pollution problems are increasing worldwide, new energy generation technologies led by renewable energy sources such as wind and solar energy have become the developing trend of the energy revolution. Due to the characteristics of cleaning and abundance, wind power technology has already had a strong impact on the original power generation technologies. However, wind resources also have the defects of volatility, randomness and uncertainty, which can easily cause the phenomenon of off-grid and wind abandonment in the operation of the power grid and bring great challenges to the stable operation of wind power grid-connected and wind turbines. In this application context, this dissertation takes into account the uncertainty of wind resources and its impact on the stable and safe operation of wind turbines. Four aspects including abnormal identification of wind turbine operating data, multi-period scale wind speed prediction, multi-period scale wind power prediction,wind turbine health assessment and wind turbine condition monitoring and prediction system are studied. The specific research contents are as follows: (1)An outlier detection model for wind turbine operating data is established. Firstly, aiming at solving the problems of the complex structure of the wind turbine, the large dimension of the operational data and the dimensional disaster, combined with the advantage of the optimized deep belief network(DBN) in dealing with the nonlinear mapping problem, a feature extraction method based on deep belief network is proposed. After reducing the attribute dimension, the unsupervised anomaly data recognition model of Clustering in quest (CLIQUE) algorithm is established. By comparing the size of the grid node density with the set threshold, the SCADA data are divided into normal operated data, missing data, stacked anomaly data and the discrete anomaly data. (2)A model for multi-period wind speed prediction is proposed. Firstly, the variance analysis of the wind speed time series is carried out to realize the optimal grouping of the wind speed series. With the ability of handling the noise and extracting the abnormal features, the stacking denoising autoencoder (SDAE) is performed to capture the hidden characteristics within the wind speed series. Finally, owing to the poor generalized ability of single learner, Bootstrap method is used to extract samples, and the integrated extreme learning machine prediction model is built to conduct the multi-period wind speed prediction. (3)A model for multi-period wind power forecasting is proposed. Based on the multi-period wind speed prediction in the previous section, the optimal strategy for selecting features is constructed to map the original input space to the low-dimensional expression space to avoid the over-fitting. The stochastic gradient descent method and the error weight adjustment method are introduced to establish the variable weights. The prediction model of random forests which feeds the error back to the forecasting results is able to update the error weights in real time and improve the accuracy of wind power forecasting. (4)A model for evaluating the health status of the wind turbine is proposed. Due to the complexity and ambiguity of the structure, the changes in the health of the wind turbine are gradual rather than sudden. Therefore, this dissertation selects the specific monitoring data and embeds the change point algorithm into the Recurrent Neural Network (RNN). This dissertation builds the idea of "offline classification, online matching" to mine the underlying trend in the sequence data, and then performs the predicting procedure. Finally, the weighted evaluation method is used to evaluate the health status of the wind turbine by combining the prediction residuals of several important components. (5)The research on wind turbine condition monitoring and prediction system is conducted. Based on the theoretical research of the key problems of analysis and prediction on the operation of wind turbines considering the uncertainty mining, the dissertation uses Java platform and Oracle database to develop wind turbine condition monitoring and prediction system, which promotes the shift from theoretical work to industrial applications. Key words: outliers detection; wind speed prediction; wind power forecasting; evaluation of health status; deep learning