【作者】 张弛；

【导师】 王宝华；

【作者基本信息】 南京理工大学 ， 电气工程（专业学位）， 2021， 硕士

【摘要】 随着国家“碳中和,碳达峰”战略目标的提出,电力系统将迎来新一轮的改革浪潮,产业的升级要求配电网做到合理分配电能、减小能源损耗,进一步地给配电网的负荷预测工作提出了更高的标准。精准的电力负荷预测将为电能生产、输送、分配提供巨大的帮助,在传统预测方法难以满足新型配电网精确化、智能化要求的背景下,深度学习给配电网的负荷预测提供了新的思路。在对现有的短期负荷预测方法进行归纳总结后,分析了配电网用户用电行为特征,结合屡攀高峰的深度学习技术,提出了一种基于深度学习的配电网短期负荷预测方法,具体研究内容如下:首先,针对配电网用户用电行为特征,选择了自组织映射(Self-Organizing Map,SOM)神经网络建立用户负荷聚类模型,使用配电网采集到的真实负荷数据进行仿真分析,获得用户分类结果和聚类中心,为预测模型的结构选择提供了依据。经实际配电网用户信息验证及与其他聚类模型的对比分析,确认了选择SOM神经网络的正确性。其次,提出了一种基于鸟群算法的短期负荷预测组合模型,使用鸟群算法对深度神经网络(Deep Neural Network,DNN)、长短期记忆(Long Short Term Memory,LSTM)神经网络两种预测模型的结果加权取平均,获得初步预测结果。经过算例分析发现,本文提出的基于鸟群算法组合模型的用户负荷初步预测方法能够有效提高预测精度,相比于DNN、LSTM神经网络等单一预测模型及PSO-Elman(Particle Swarm OptimizationElman,PSO-Elman)神经网络、WT-BP(Wavelet Transform-Back Propagation)神经网络等组合模型,该方法准确度更高、泛化能力更强,适合配电网的短期负荷预测工作。最后,基于同类型用户平均误差的短期负荷预测补偿算法对初步预测结果进行修正。真实数据的仿真结果表明,这种“聚类-预测-补偿”结构的组合模型与众多现有的预测模型相比性能更为突出,能够在一定程度上减小用户用电特征不同带来的影响,在一定范围内不断提升预测精度,具有重要的工程应用价值。更多还原

【Abstract】 With the national strategic goal of "carbon neutrality,emission peak" proposed,the power system will be involved in a new round of reform,industrial upgrading requires distribution network to achieve reasonable distribution of electricity and reduced energy consumption.Furthermore,a higher standard is proposed for load forecasting of distribution network.Accurate power load forecasting will provide great help for production,transmission and distribution of electric energy.Under the background that traditional prediction methods cannot meet the requirements of precision and intelligence of new distribution network,deep learning provide a new idea for load forecasting of distribution network.After summarizing the existing methods of short-term load forecasting,the characteristics of power consumption of distribution network users are analyzed.Combined with the deep learning technology which repeatedly climbing peaks,a method of short-term load forecasting of distribution network based on deep learning is proposed.The specific research contents are as follows:Firstly,aim at the characteristics of power consumption of distribution network users,SelfOrganizing Map(SOM)neural network is selected as the principal part of user load clustering model.The real load data collected in distribution network is used for simulation analysis to obtain user clustering results and clustering centers,which provides a basis for choosing the structure of prediction model.The correctness of SOM neural network is verified by the actual information of distribution network users and comparing with other clustering models.Secondly,a combination model of short-term load forecasting based on Bird Swarm Algorithm(BSA)is proposed.The prediction results of Deep Neural Network(DNN)and Long Short-Term Memory(LSTM)neural network are weighted and averaged by BSA,which provides the preliminary prediction results.Through example analysis,it is found that the user load preliminary prediction method based on the combined model of BSA proposed in this paper can effectively improve the prediction accuracy,compared with DNN,LSTM neural network and other single prediction models and Particle Swarm optimization-Elman(PSO-Elman)neural network,Wavelet Transform-Back Propagation(WT-BP)neural network and other combined models.This method has higher accuracy and stronger generalization ability,and is suitable for short-term load forecasting of distribution network.Finally,the short-term load forecasting compensation algorithm based on the average error of historic data of the same type of users is used to modify the preliminary prediction results.The simulation results of real data show that this combination model with structure of "Clustering-Forecasting-Compensation" is more outstanding in performance compared with a number of existing forecasting models.It can reduce the effects of different power characteristics to a certain extent,improve forecasting accuracy within a certain scope,and have the important value of engineering application.更多还原