数学与统计学院知识库

短时电力负荷的准确预测在电网规划的技术层面和经济方面具有重要意义。本文基于近年来较流行的机器学习算法对短时电力负荷预测进行深入的研究：针对电力历史负荷的数据波动特点进行多时间尺度分析；根据静态和动态神经网络模型特点，在小波包分解的基础上，通过分析精度高、泛化能力强且适于处理时序数据的WNN、RBFNN、ENN和NAR神经网络模型，建立了基于GA遗传算法的组合模型和NAR回归型动态神经网络模型；为验证所提出理论的有效性，本文选择利用美国PJM区域2016.12.1~2017.11.30的小时电力负荷数据进行预测比较分析，另外选择澳大利亚新南威尔士州2016.12.1~2017.11.30的小时电力负荷数据证明本论文所提出模型的适用性。
本论文主要有以下几个方面的研究内容和结果：第一，根据美国PJM电力负荷数据特点，利用复小波分析理论进行多时间尺度分析。结果表明：非平稳的负荷时序中的存在4种变化周期，分别是以52天、33天、17天和7天为变化趋势；第二，由第一部分负荷多时间尺度分析的结果，以第一主周期52天的数据为例，建立基于GA的组合预测模型和NAR神经网络预测模型。第三，基于所提出的预测分析模型，以2017.12.01~2017.12.31的数据进行预测分析，并由实际发生的负荷数据进行验证。结果表明：PJM区域基于GA的组合模型和NAR神经网络预测模型的R值均为0.976，SMAPE分别为1.4%和1.7%；新南威尔士州基于GA的组合模型和NAR神经网络预测模型的R值分别为0.973和0.979，SMAPE分别为1.4%和1.8%。有效的短时电力负荷预测对于制定合理的电网规划和保证电网系统的稳定运行提供理论支持。在时序预测分析方面，本研究的结果表明，本文所提出的基于多时间尺度分析和GA的组合预测模型与NAR神经网络模型对负荷数据预测均具有较高精度。

The accurate short term load forecasting is significant for power grid planning and economic development. Based on the popular machine learning algorithms, this paper studies the short term load forecasting. Firstly, analyses the characteristics and fluctuations of the historical electric load data with the multi time scale analysis technique. Secondly, proposes a combined model based on genetic algorithm (GA), WNN, RBFNN and ENN for short-term power load forecasting. Thirdly, builds one NAR neural network for short-term power load forecasting. Finally, two hourly electricity load datasets (Dec. 01, 2016 to Nov. 30, 2017, Pennsylvania—New Jersey—Maryland, American; Dec. 01, 2016 to Nov. 30, 2017, the State of New South Wales, Australia) are selected to test the effectiveness and generalization ability of the combined and NAR models.
The main contents and results of this paper are as follows: First, according to the characteristics of PJM load data in American, analyses its periods by complex wavelet analysis theory. The results show that there are 4 kinds of period variations, which are 52 days, 33 days, 17 days and 7 days, respectively. Second, according to the result of multiple time scale analysis, the first period is used to build the combined forecasting mode, which is NAR neural network optimized by GA. Third, the combined model is used to analyze and forecast the Pennsylvania—New Jersey—Maryland, American load data from Dec. 01, 2016 to Nov. 30, 2017. And the forecasting result is tested by actual load data. The results show that the R values of combined forecasting model and NAR neural network model are both 0.976, SMAPE are 1.4% and 1.7%, respectively. For the State of New South Wales, R values of combined model and NAR neural network model are 0.973 and 0.979 respectively, and SMAPE are 1.4% and 1.8% respectively.The effective short term load forecasting provides theoretical support for formulating reasonable grid planning and ensuring stable operation of the grid system. In terms of time series analysis, the forecasting results show that the proposed combined forecasting model based on GA and NAR neural network model are effective for short term load forecasting.