【作者】 刘文杰；

【导师】 魏海坤；

【作者基本信息】 东南大学 ， 模式识别与智能系统， 2021， 硕士

【摘要】 研究光伏发电功率预测对于保障电网稳定运行、提升电网光电消纳能力、减少光伏电站由于限电带来的经济损失、提高电站运营管理效率有着巨大帮助。但目前光伏功率预测系统的设计应用中存在着以下几个问题:一是模型无法针对现场工况及时更新导致预测精度下降;二是针对一些极端天气没有进行特殊设计,导致极端天气下的预测能力不足;三是一些光伏电站现场缺少气象观测站,无法获取实时更新的气象预报,在这种情况下进行未来四小时超短期预测的相关研究文献很少。解决这些问题对于提升光伏发电功率预测的精度有着重要意义。因此本文基于合作研究院提供的光伏电站历史运行数据,针对光伏功率预测的相关问题进行了深入研究。本文完成的主要工作如下:(1)确定了系统整体设计框架。调研了大型光伏发电系统运行原理。对光伏发电功率预测系统进行了需求分析,确定了预测的性能指标,并对系统的结构和主要功能进行了设计。最后阐述了系统的整体运行流程,明确了后续的研究内容。(2)建模前对现有数据进行了预处理与分析。利用孤立森林算法对光伏电站的历史数据进行了数据清洗。然后分析了光伏发电功率出力的相关影响因素。在进行不同天气条件下的气象数据分析后,利用分层划分的思想通过尽可能少的特征将天气类型分为了晴天、多云、阴雨三类,第一层采用了改进的K-Means算法、第二层采用了二阶差分的方法。(3)设计了光伏发电短期功率预测模型。针对目前短期预测模型无法适应现场工况环境变化导致预测精度会降低的问题,基于OS-ELM算法设计了能够进行在线学习更新的光伏短期功率预测模型。该模型在晴天、多云、阴雨三种天气类型下均取得了良好效果,短期预测的r RMSE分别为2.98%、5.32%、4.09%。针对雾霾天气对光伏功率预测的影响,以PM2.5为例设计了前置模型进行辐照度的预校正,实验结果显示在严重雾霾天气下预测误差能够降低约5个百分点。(4)设计了光伏发电超短期功率预测模型。分析了相似日模型、基于功率序列预测模型、基于误差序列的预测模型各自的预测效果与优劣势。针对传统KNN算法在处理大批量数据时耗时较长的问题,提出了基于Annoy算法的相似日预测方法。在保证精度的同时提升了效率。针对没有实时天气预报情况下的四小时超短期预测需求,区别于传统思路直接用功率序列预测,提出了基于GRU的误差序列超短期预测模型,保证了整体的预测精度和稳定性。最后设计的超短期功率预测模型将三个模型进行结合,发挥了三个模型各自的优势,避开了各个模型的缺点,在实验数据集上的超短期第16步预测精度达到了96.20%。(5)完成了光伏发电功率预测系统的开发与试运行。系统软件基于Java与Python语言进行开发,通过运行时调用的方法实现了两种语言的共存。系统具备良好的人机交互界面,在几个光伏电站提供的数据下均取得了良好的实际运行效果。本文通过调研实际光伏电站的运行需求展开系统设计、数据处理与建模预测工作,着重解决了短期模型自适应更新、极端天气下的功率预测、缺少气象观测站情况下的超短期功率预测等问题,对提升光伏发电功率出力的预测精度、提高发电并网的利用效率、维持光伏发电产业高速增长的发展态势具有积极意义。更多还原

【Abstract】 Research on photovoltaic power generation forecast is of great help to ensure stable operation and improve the photoelectric absorption capacity of the power grid.The research can also reduce the economic losses caused by photovoltaic power stations due to power rationing,and improve the efficiency of power station operation and management.However,there are several problems in the design and application of photovoltaic power prediction systems.First,the model cannot be updated in time according to the field conditions,which leads to the decline of prediction accuracy.Second,there is no special design for some extreme weather,resulting in the lack of prediction ability under extreme weather.Third,some photovoltaic power plants lack meteorological observation stations onsite and cannot obtain real-time updated weather forecasts.In this case,there is little relevant research literature on the ultra-short-term prediction of the next four hours.Solving these problems is of great significance to improve the accuracy of photovoltaic power forecasts.Therefore,based on the historical operation data of photovoltaic power stations provided by the cooperative research institute,this paper conducts an in-depth study on the related issues of photovoltaic power prediction.The main works completed in this paper are as follows:(1)The overall design framework of the system is determined.The operation principle of a largescale photovoltaic power generation system is investigated.The demand for a photovoltaic power prediction system is analyzed.The performance index of prediction is determined.And the structure and main functions of the system are designed.Finally,the overall operation process of the system is described,and the follow-up research content is clarified.(2)The existing data are preprocessed and analyzed before modeling.The isolated forest algorithm is used to clean the historical data of the photovoltaic power station.Then the influence factors of PV power output are analyzed.After analyzing the meteorological data under different weather conditions,the idea of hierarchical division is used to divide the weather types into sunny,cloudy,and rainy with as few features as possible.The first layer uses the improved k-means algorithm,and the second layer uses the second-order difference method.(3)The short-term power prediction model of photovoltaic power generation is designed.Aiming at the problem that the current short-term prediction model can not adapt to the changes of field conditions and environment,which leads to the decrease of prediction accuracy,an online learning and updating photovoltaic short-term power prediction model is designed based on the OSELM algorithm.The model has achieved good results in three weather types: sunny,cloudy,and rainy.The short-term prediction r RMSE are 2.98%,5.32% and 4.09%.Aiming at the influence of haze weather on photovoltaic power prediction,a pre-correction model for irradiance is designed with PM2.5 as an example.The experimental results show that the prediction error can be reduced by about 5% in severe haze weather.(4)The ultra short term power prediction model of photovoltaic power generation is designed.The advantages and disadvantages of similar day model,power series-based model,and error seriesbased model are analyzed.Aiming at the problem that the traditional KNN algorithm takes a long time to process large quantities of data,a similar day prediction method based on ANNOY algorithm is proposed.It improves efficiency while ensuring accuracy.To meet the demand of four-hour ultra short term forecasting without real-time weather forecast,an error series ultra-short-term prediction model based on GRU is proposed,which is different from the traditional methods of direct power series prediction,to ensure the overall accuracy and stability of the prediction.Finally,the ultra-shortterm power prediction model designed combines the advantages of the three models and avoids the disadvantages of each model.The prediction accuracy of step 16 on the experimental data set reaches96.20%.(5)The development and trial operation of the photovoltaic power generation prediction system is completed.The system software is developed based on JAVA and Python,and realizes the coexistence of the two languages through the method called at run time.The system has good manmachine interaction and has achieved good operation effect under the data provided by several photovoltaic power stations.In this paper,the system design,data processing,modeling and prediction are carried out by investigating the actual operation demand of photovoltaic power stations.The problems of short-term model self-adaptive updating,power prediction in extreme weather and ultra-short-term power prediction in the absence of meteorological observation stations are mainly solved.It is of great significance to improve the prediction accuracy of photovoltaic power output,improve the utilization efficiency of power generation grid-connection and maintain the development trend of high-speed growth of the photovoltaic industry.更多还原