基于相邻风场大数据的风电短期功率预测研究

发布时间：2019-04-27 18:44

【摘要】：在能源匮乏和环境问题的大背景下,世界各国开始寻求低碳发展道路,竞相发展可再生能源,风电便是其中之一。风电资源丰富、装机灵活,风电技术相较其他可再生能源发电而言也更成熟,效率更高,能很好地替代化石能源,保证能源供应的前提下更好地保护好环境。经过近些年的快速发展,我国已经成为世界上风电装机最大的国家。但是,风电具有的随机性、间歇性和反调峰特性,严重影响了我国风电的大规模并网消纳,导致了严重的弃风问题。因此,对风电短期功率预测的研究可以弥补风电不稳定的缺点,有利于电网更加合理地安排调度计划,使得更多的风电得到消纳,有效地缓解弃风问题,对我国风电产业健康持续的发展具有重要意义。另一方面,随着风场大数据的逐步崛起,利用大数据进行风电功率预测是未来发展的一个趋势。而深度学习在大数据的挖掘中正在发挥着越来越突出的贡献。其中,卷积神经网络(CNNs)发展最为成熟,在图像识别、模式识别等领域取得了成功。本文首先基于相邻风场大数据的结构特点,通过真实的数据构建了三维的实验数据集,并运用统计分布、动态关联性分析等方法,研究了实验数据集的数据特点,为后续预测建模打下基础。接着,建立了风电短期功率CNNs预测模型,利用多个CNNs网络独立运行,实现模型多输出的效果;通过重点阐释风电短期功率CNNs预测模型建立的全过程,详细分析模型的预测效果,验证了风电短期功率CNNs预测模型的实用性和可靠性。结果显示,CNNs预测模型在误差控制上有较好的效果,在整体预测精度提升的同时,对不同时间节点、不同功率样本的预测效果较传统方法而言,也更为平均。最后,通过建立CNNs预测模型和物理预测模型的组合预测模型,采用分类式的结构权重,充分发挥两种方法在不同样本中的优势,进一步降低风电短期功率预测误差。实际工作中,将组合模型权重确定问题转化成参数优化问题,利用遗传算法(SC)快速求解,效率高。从实验结果看,组合预测模型误差较CNNs预测模型降低约5%,分类式的结构权重也较单一权重下的误差要略小。通过本文的研究,一定程度上证明了CNNs网络方法在处理风电短期功率预测问题上的大数据时,有较好的应用前景。
[Abstract]:In the background of energy shortage and environmental problems, countries in the world began to seek low-carbon development path, and competing for renewable energy, wind power is one of them. Compared with other renewable sources, wind power technology is more mature and more efficient. It can replace fossil energy and protect the environment better on the premise of ensuring energy supply. After the rapid development in recent years, China has become the largest wind power installed country in the world. However, the randomness, intermittency and anti-peak-shaving characteristics of wind power seriously affect the large-scale grid-connected dissipation of wind power in China, resulting in a serious wind abandonment problem. Therefore, the research on short-term power forecasting of wind power can make up for the shortcomings of instability of wind power, help the power grid to arrange the dispatching plan more reasonably, make more wind power be absorbed, and effectively alleviate the problem of wind abandonment. It is of great significance to the healthy and sustainable development of wind power industry in China. On the other hand, with the gradual rise of wind field big data, using big data to forecast wind power is a trend in the future. And in-depth learning in big data's excavation is playing a more and more prominent contribution. Among them, convolutional neural network (CNNs) is the most mature and has been successful in image recognition and pattern recognition. Firstly, based on the structural characteristics of the adjacent wind field big data, a three-dimensional experimental data set is constructed through real data, and the data characteristics of the experimental data set are studied by means of statistical distribution, dynamic correlation analysis, and so on. It lays a foundation for the following prediction modeling. Then, a short-term wind power CNNs prediction model is established, which uses multiple CNNs networks to run independently to realize the effect of multi-output of the model. The whole process of the establishment of short-term wind power CNNs prediction model is explained, and the forecasting effect of the model is analyzed in detail. The practicability and reliability of the wind power short-term CNNs prediction model are verified. The results show that the CNNs prediction model has a good effect on error control. While the overall prediction accuracy is improved, the prediction effect of different time nodes and different power samples is more average than that of the traditional method. Finally, the combination forecasting model of CNNs prediction model and physical prediction model is established, and the classification structure weight is adopted to give full play to the advantages of the two methods in different samples, so as to further reduce the short-term power prediction error of wind power. In practical work, the weight determination problem of combinatorial model is transformed into parameter optimization problem, and genetic algorithm (SC) is used to solve the problem quickly, which has high efficiency. The experimental results show that the error of the combined prediction model is about 5% lower than that of the CNNs prediction model, and the error of the structure weight of the classification formula is slightly smaller than that of the single weight. Through the research in this paper, it is proved that the CNNs network method has a good application prospect in dealing with big data in the short-term power prediction of wind power.
【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

【参考文献】