基于深度学习的风电功率预测方法研究

发布时间：2018-12-29 17:42

【摘要】：风电的随机波动性影响电力系统的安全、稳定和经济运行,是大规模风电并网的主要挑战。风电功率预测是解决该问题的必要手段之一。现阶段风电功率预测最主要的问题是精度不足,除数值天气预报(Numerical Weather Prediction,NWP)精度的原因外,预测模型也存在两方面问题:一是模型深度不足、映射结构简单;另一是训练样本维度少、规模小、未经严格清洗。上述问题导致预测模型学习能力不足、难以适应复杂风场景和描述复杂地形中流场关系,极大地限制了风电功率预测精度的提升。因此,本文中提出了基于堆叠降噪自动编码机(Stack Denoising Auto-Encoder,SDAE)的短期风电功率预测方法,建立了以多点数值天气预报为输入、多台机组功率为输出的深度学习模型。主要工作包括:1)提出了风电功率预测数据清洗方法通过对NWP、实测风速、实测功率等风电功率预测建模数据进行质量分析,提出了两种实测风速插补算法:基于流场相似状态的K最近邻插补(K Nearest Neighbor,KNN)和基于相关性排序的插补算法;提出了基于桨距角上临界曲线的功率数据筛选方法和全场功率数据筛选规则,实现对实测功率数据的筛选。结果表明:基于流场相似状态的KNN插补算法精度更高,更适合后续的建模工作;基于桨距角上临界曲线的功率数据筛选方法能快速准确的筛选出正常状态下的功率数据。2)提出了基于SDAE的多点NWP误差修正方法分析了NWP误差的时空分布模式;提出了一种基于SDAE的多点NWP误差修正方法,并建立了以多点NWP为输入、多台风电机组风速为输出的三隐层SDAE网络模型。结果表明:SDAE模型比3种现有模型的修正精度更高,且无需分月、分机组进行大批量建模。3)建立了基于NWP修正风速和SDAE的风电功率预测模型提出了基于多对多映射结构的风电功率预测建模方法,建立了基于修正NWP风速和SDAE的风电功率预测模型,模型以多点修正NWP风速作为模型输入、多台机组功率作为模型输出。结果表明:与8种主流的预测模型相比,NWP修正对功率预测精度的提升效果明显,采用多点NWP输入有助于提高预测精度,三隐层SDAE网络优于浅层网络。
[Abstract]:The stochastic volatility of wind power affects the safety, stability and economic operation of power system, which is the main challenge of large-scale wind power grid. Wind power prediction is one of the necessary methods to solve this problem. At present, the main problem of wind power prediction is lack of precision. Besides the accuracy of numerical weather forecast (Numerical Weather Prediction,NWP), the prediction model also has two problems: first, the depth of the model is insufficient, and the mapping structure is simple; The other is the training sample dimension is small, without strict cleaning. The above problems lead to a lack of learning ability of the prediction model, which is difficult to adapt to the complex wind scene and describe the relationship between the flow field in the complex terrain, which greatly limits the improvement of the prediction accuracy of wind power. Therefore, in this paper, a short-term wind power prediction method based on stack noise reduction automatic coding machine (Stack Denoising Auto-Encoder,SDAE) is proposed, and a depth learning model with multi-point numerical weather forecast as input and power output as output is established. The main works are as follows: 1) the cleaning method of wind power prediction data is put forward. The modeling data of wind speed and measured power of NWP, are analyzed by quality analysis. Two interpolation algorithms are proposed: K nearest neighbor interpolation (K Nearest Neighbor,KNN) based on similar state of flow field and interpolation algorithm based on correlation ranking. The power data screening method based on the critical curve of pitch angle and the full-field power data screening rule are proposed to screen the measured power data. The results show that the KNN interpolation algorithm based on the similar state of the flow field is more accurate and more suitable for the subsequent modeling work. The power data screening method based on the critical curve of pitch angle can quickly and accurately screen the power data under normal condition. 2) the multipoint NWP error correction method based on SDAE is proposed to analyze the temporal and spatial distribution of NWP error. A multi-point NWP error correction method based on SDAE is proposed, and a three-layer SDAE network model with multi-point NWP as input and wind speed as output is established. The results show that the correction accuracy of the SDAE model is higher than that of the three existing models, and it does not need to be divided into months. The wind power prediction model based on NWP modified wind speed and SDAE is established. The wind power prediction modeling method based on many-to-many mapping structure is proposed. The wind power prediction model based on modified NWP wind speed and SDAE is established. The model takes the multi-point modified NWP wind speed as the input and the power of several units as the model output. The results show that compared with the 8 main prediction models, the NWP correction can improve the power prediction accuracy obviously, the multi-point NWP input is helpful to improve the prediction accuracy, and the three-hidden layer SDAE network is better than the shallow one.
【学位授予单位】：华北电力大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM614;TP181

【参考文献】