【作者】 刘文杰；

【导师】 徐青山； 汪春；

【作者基本信息】 东南大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 综合能源负荷预测对于精益化综合能源系统管理,优化能源结构和推动“双碳”目标的实现有着重大意义。然而,目前实现精确高效的综合能源负荷预测仍面临挑战。一方面,综合能源系统存在着不同程度的多能源耦合,使得综合能源系统负荷预测变得更加复杂;另一方面,综合能源负荷影响因素多,使得现有负荷预测模型在处理大量特征时存在计算复杂度高以及过拟合风险等问题。针对以上问题,本文从特征集构建、特征提取、特征选择、数据降维和数值预测等方面开展研究,提出了一种基于特征降维的综合能源负荷预测方法,其主要研究内容如下:1)分析了综合能源负荷特性及其影响因素,为负荷预测模型输入特征集的构建提供理论依据。从气象、经济、日期类型、耦合负荷、政策五个维度进行分析,选择了十余种不同的影响因素作为输入特征集的待选变量,并给出了相应的量化指标和时序化处理的方法。其中,本文通过提出特定空气质量指标和经济指标,实现了相应政策类型强弱的量化,构造了政策维度的特征集输入,丰富了负荷预测特征集的选择。2)提出了综合能源负荷运行特征量提取方法,分别采用时域分析和时频域分析对特征集中的变量进行特征提取,达到判定负荷预测方法和提取时频域特征的目的。在时域分析中,本文采用的方法包括基于ADF检验的平稳性校验、基于自相关函数的自相关性分析和基于STL分解的周期性分析;在时频域分析中,采用了变分模态分解方法提取时频域特性,并结合FFT、PSD,提出了一种基于能量贡献率的变分模态分解模数确定的方法,实验证明其有效控制了特征的数量。3)提出了基于GSO-MIC的特征选择方法。该方法首先采用最大信息系数分析待预测负荷和特征变量之间的相关性,使用过滤法对特征变量进行预筛选。在此基础上,利用Gram-Schmidt正交化特性,对特征变量进行二次选取,减少了特征之间按的冗余性,提高特征选择的有效性。4)提出了基于深度稀疏自编码器非线性数据降维的方法。考虑到特征集在经过模态分解后规模增大,且特征之间还可能存在的非线性冗余,本文采用了深度稀疏自编码器对选择好的特征集进行降维,在保留关键信息的同时显著降低了特征的数量和复杂性。实验中,通过对比原始信号与重构信号的重构误差,验证了所提方法的有效性。5)提出了基于注意力机制的CNN-BiLSTM综合能源负荷预测方法,并在省级智慧能源服务平台上完成了部署。结合负荷特性和降维后的特征矩阵,该预测模型有效提高了综合能源负荷预测的效率和精确度。在本文各章节的实验算例中,利用了公开数据集和项目实验数据集验证了方法的有效性。更多还原

【Abstract】 Integrated energy load forecasting is of great importance for lean and integrated energy system management,optimizing the energy structure and promoting the goal of "double carbon".However,there are still challenges to achieve accurate and efficient integrated energy load forecasting.On the one hand,there are different degrees of multi-energy coupling in the integrated energy system,which makes the integrated energy system load forecasting more complex;on the other hand,the integrated energy system is affected by many factors,which makes the existing load forecasting models suffer from high computational complexity and overfitting risk when dealing with a large number of features.To address the above problems,this paper conducts research on feature set construction,feature extraction,feature selection,data downscaling and numerical prediction,and proposes a comprehensive energy load forecasting method based on feature downscaling,whose main research contents are as follows:1)The integrated energy load characteristics and their influencing factors are analyzed to provide a theoretical basis for the construction of the input feature set of the load forecasting model.Five dimensions are analyzed: meteorology,economy,date type,coupled load,and policy,and more than ten different influencing factors are selected as the variables to be selected for the input feature set,and the corresponding quantified indexes and methods of time-series processing are given.Among them,this paper achieves the quantification of the strength of the corresponding policy types by proposing specific air quality indicators and economic indicators,constructs the input feature set of the policy dimension,and enriches the selection of the load prediction feature set.2)An integrated energy load operation feature extraction method is proposed,and the variables in the feature set are extracted using time-domain analysis and time-frequency domain analysis,respectively,for the purpose of determining the load forecasting method and extracting time-frequency domain features.In the time-domain analysis,the methods used in this paper include smoothness check based on ADF test,autocorrelation analysis based on autocorrelation function and periodicity analysis based on STL decomposition;in the time-frequency domain analysis,the variational modal decomposition method is used to extract time-frequency domain features,and combined with FFT and PSD,a method of determining the modulus of variational modal decomposition based on energy contribution rate is proposed,and the experiment proves that it The number of features is effectively controlled.3)A GSO-MIC-based feature selection method is proposed.The method first uses the maximum information coefficient to analyze the correlation between the load to be predicted and the feature variables,and uses the filtering method to initially select the feature variables.On this basis,the Gram-Schmidt orthogonalization property is used to select the feature variables twice,which reduces the redundancy between features by and improves the effectiveness of feature selection.4)A method based on depth sparse self-encoder for dimensionality reduction of nonlinear data is proposed.Considering the increase in the number of features after modal decomposition and the possible nonlinear redundancy among the features,this paper adopts a deep sparse selfencoder to reduce the dimensionality of the selected feature set,which significantly reduces the number and complexity of features while retaining the key information.In the experiments,the effectiveness of the proposed method is verified by comparing the reconstruction error of the original signal with that of the reconstructed signal.5)A CNN-BiLSTM integrated energy load forecasting method based on attention mechanism is proposed and deployed on the provincial smart energy service platform.Combining the load characteristics and the reduced-dimensional feature matrix,the prediction model effectively improves the efficiency and accuracy of integrated energy load forecasting.In the experimental arithmetic of each chapter of this paper,the validity of the method is verified using public datasets and project experimental datasets.更多还原