摘要
光伏发电呈现出时变性、波动性和随机性，对光伏发电的稳定性带来不利影响。当光伏电站大规模接入电网后，由于其波动性给电网的整体稳定运行带来了巨大挑战，降低了电网运行的可靠性，增加了维护电网波动所带来的的运行和管理成本。因此，光伏发电功率的准确且合理的预测对电网的安全调度、维护电网的稳定运行和提高光伏电站利用率具有重要意义。目前传统的机器学习对光伏发电预测大多都是通过直接预测的方式，本文本着提升精度的思想，首先通过结合光伏的特征融合选取和模型参数组合选取的方式设计了一种基于改进Xgboost算法的光伏发电功率的预测模型。考虑到这种方式的缺陷在于对极端天气的适用度较差，提出了一种基于天气类型的高斯混合聚类相似日筛选模型。最后，针对单个模型提升精度有限的问题设计了基于Stacking模型融合的算法。主要研究内容包括：
（1）基于改进的Xgboost的超短期光伏预测算法。目前传统的机器学习对光伏发电预测大多都是通过直接预测的方式，无法对特征进行有效的筛选。本文设计一种改进的Xgboost的超短期预测算法，通过一种特征融合的方式去有效的筛选模型和参数组合的方式去有效提高Xgboost对光伏发电功率的预测精度。
（2）基于高斯混合聚类算法的相似日筛选模型。为了解决光伏发电在一些极端天气(阴天、雨天等)下的光伏预测精度低的问题，在不同天气情况下基于高斯混合聚类算法设计了相似日的筛选模型。与传统的
筛选方式筛选出的样本在同一模型下进行对比分析，发现该方式筛选出的样本有利于模型在极端天气情况下的拟合。
（3）基于Stacking模型融合下的光伏发电功率预测算法。针对单一的预测模型预测精度提升有限的问题，引入了集成学习的思想和方法，提出一种基于Stacking方法来结合支持向量机、BP神经网络、线性回归、决策树、xgboost等模型的短期预测方式，通过与单一模型Xgboost、BP神经网络相对比，精度有了明显的提升，通过与光伏发电的实际功率对比，具有很好的吻合性，在实际应用中具有很高的工程推广价值。
关键词: 光伏发电短期预测，模型融合Stacking ，相似日的聚类，高斯混合聚类，Xgboost
Abstract
Photovoltaic power generation exhibits time-varying, volatility and randomness, which adverly affects the stability of photovoltaic power generation. When the photovoltaic power plant is connected to the power grid on a large scale, its volatility pos a huge challenge to the overall stable operation of the power grid, reduces the reliability of the power grid operation, and increas the operation and management costs of maintaining power grid fluctuations. Therefore, accurate and reasonable prediction of photovoltaic power generation is of great significance for the safe dispatch
of the power grid, maintaining the stable operation of the power grid, and improving the utilization rate of photovoltaic power plants. At prent, most of the traditional machine learning predictions for photovoltaic power generation are through direct prediction. This paper is bad on the idea of improving accuracy. First, a combination of photovoltaic feature fusion lection and model parameter combination lection is ud to design a method bad on improved Xgboost algorithm. Predictive model of photovoltaic power. Considering the shortcoming of this method is that the prediction accuracy of extreme weather is poor, a similar day screening model bad on Gaussian mixture clustering is further propod. Finally, an algorithm bad on the fusion of Stacking model is designed to solve the problem that the single model has limited lifting accuracy. The main rearch contents include:
(1) Ultra-short-term photovoltaic prediction algorithm bad on improved Xgboost. At prent, most of the traditional machine learning predictions of photovoltaic power generation are through direct prediction, which cannot effectively filter the features. This paper designs an improved Xgboost ultra-short-term prediction algorithm, through a feature fusion method to effectively filter the model and parameter combination to effectively improve the prediction accuracy of Xgboost for photovoltaic power generation.
(2) Similar day screening model bad on Gaussian hybrid clustering algorithm. In order to solve the problem of low photovoltaic forecast accuracy in some extreme weather (cloudy, rainy, etc.), a similar day screening model was designed bad on Gaussian hybrid clustering algorithm under different weather conditions. The samples lected by the traditional screening method were compared and analyzed under the same model, and it was found that the samples lected by this method were beneficial to the fitting of the model under extreme weather conditions.
(3) Bad on the fusion of Stacking model photovoltaic power prediction algorithm. Aiming at
the problem of limited prediction accuracy improvement of a single prediction model, the idea and method of integrated learning are introduced, and a short-term prediction method bad on Stacking method combining support vector machine, BP neural network, linear regression, decision tree, xgboost and other models Compared with the single model Xgboost and BP neural network, the accuracy has been significantly improved, and by comparing with the actual power of photovoltaic power generation, it has a good agreement and has high engineering promotion value in practical applications.
Keywords: short-term forecast of photovoltaic power generation, model fusion Stacking, clustering of similar days, Gaussian hybrid clustering, xgboost
目录
第一章绪论 (1)
1.1 研究背景与意义 (1)
1.2 光伏发电出力预测研究技术类别 (2)
卑微的近义词1.3 光伏短期功率研究现状 (4)
1.3.1 国内研究现状 (4)
1.3.2 国外研究现状 (5)
百折不回1.4 课题的主要研究内容 (6)
第二章光伏出力因素分析以及模型原理 (8)
2.1 天气因素的可视化分析 (8)桃花妖哪里多
2.1.1 辐照度 (8)
2.1.2 大气温度 (9)
喝黄酒的好处2.1.3 天气类型的影响 (9)
2.2 光伏预测基本模型的介绍 (10)
2.2.1 决策回归树 (10)
2.2.2 线性回归 (12)
2.2.3 BP神经网络 (13)
夸张句大全
2.2.4 支持向量机 (16)
2.3光伏发电功率误差评价指标 (17)
2.4 本章小结 (19)
第三章基于改进Xgboost的超短期光伏发电功率预测 (20)
3.1 Xgboost模型原理 (20)
咳嗽怎么治最有效
3.1.1 Xgboost (20)
3.1.2 Xgboost的模型参数 (21)
3.2 基于改进Xgboost的光伏预测 (21)
3.2.1 数据预处理过程 (21)
3.2.2 特征融合与测试 (26)
好听的群网名3.2.3 模型设计及流程图 (27)
3.3 算例分析 (28)
3.4 本章小结 (31)
第四章基于高斯混合聚类的相似日筛选模型 (33)
4.1 高斯混合聚类算法 (33)
4.2 基于高斯混合聚类的相似日筛选模型 (34)电路图讲解和实物图
4.2.1 模型设计和流程图 (34)
4.2.2 基于天气类型的光伏发电功率数据聚类 (35)
4.2.3 基于欧氏距离的相似日筛选 (36)
4.3 算例分析 (37)
4.4 本章小结 (39)
第五章基于Stacking模型融合的超短期光伏发电功率预测 (40)
5.1 Stacking融合模型原理 (40)
5.2 基于Stacking模型融合的光伏发电功率预测 (42)
5.2.1 Stacking模型融合设计 (42)
5.2.2 模型设计与流程图 (42)
5.3 算例分析 (44)
5.3.1 天气类型1的仿真分析 (44)
5.3.2 天气类型2的仿真分析 (45)
5.3.3 天气类型3的仿真分析 (47)
5.3.4 天气类型4的仿真分析 (48)
5.4 本章小结 (50)
第六章总结与展望 (51)
参考文献 (52)
附录1 攻读硕士学位期间撰写的论文 (55)
附录2 攻读硕士学位期间申请的专利 (56)
附录3 攻读硕士学位期间参加的科研项目 (57)
致谢 (58)
