基于非侵入式负荷检测与分解的电力数据挖掘

1基于非侵入式负荷检测与分解的电力数据挖掘摘要：能源的合理利用对缓解我国所面临的能源短缺以及减少碳排放具有十分重大的意义。智能用电是坚强智能电网的重要环节之一，是互动服务体系的核心。本文旨在研究构建非侵入式负荷分解与辨识的数学模型和计算方法为前提的适用优化模型，通过对数据的深入挖掘，得出准确高效的辨识决策方法，以及相应用电设备的实时用电量。首先，根据非侵入式负荷监测系统的工作原理，提取电流稳态特征和稳态电流的谐波含有率，从宏观上分为两大类用电器来计算一系列指标参数，如峰值、均值、均方根等，对负荷类别进行区分。大功率设备在启停切换时功率差异较为明显，而低功率设备的功率变化较为接近，其设备启停在高频采样中较为直观。当多种设备发生功率混叠时，低功率设备的识别则具有一定难度，可以通过增加负荷特征提高识别准确度。即引入一个以正态分布形式的隶属度函数，从而赋值给解权重w，通过度量辨识算法得到的结果与当前用电设备投切状态的逼近程度，从而得到用电设备的投切状态。最后依次求出各用电设备每秒的实时用电量。然后，本文根据用电器在启动时会产生一定特征的负荷信号，采用MATLAB自带的神经网络工具箱，通过专门的模式识别神经网络模型Patternnet,训练算法采用量化共轭梯度法trainscg实现了一种基于神经网络的模式识别方法。该算法可以实时监测家用电器的运行及用电情况。最终挂式空调的识别率为82.57%，九阳热水壶的识别率为87.69%，其余用电器的识别率均≤10%，故确认设备1为YD3九阳热水壶，设备2为YD9挂式空调。最后依次求出每个用电设备的实时用电量。其次，本文将附件3中3个设备组中的4种有限多状态设备(YD1、YD2、YD7、YD8)和3种启/停二状态设备(YD3、YD5、YD6)的稳态时的有功功率划分为若干个聚类点，实现将设备的负荷离散化处理得到有限个状态的目的，提高后续负荷分解计算的可操作度和结果精度。根据观察分析确定样本设备的功率区段代表值，据此得到样本设备到稳态的有功功率变化量。最后通过MATLAB在对样本设备功率情况聚类分区后，以各功率聚类中心作为该工作段功率特征，通过遗传优化求解负荷识别模型求出设备组4—6的用电设备操作记录以及每个用电设备的实时用电量。最后，在负荷数据分析与特征提取的研究基础上，建立遗传算法多特征优化的目标函数模型，通过遗传迭代实现不同电器状态变化的精确分解与识别。首先要对种群个体进行基因编码。因为在实际生活中存在一定噪声的情况下，功率较为接近的低功率负荷更加难以识别。为解决该问题，引入活性电流谐波特征值，从而增加识别算法的准确性和抗干扰能力。故对单功率目标寻优函数模型进行优化，在进行多目标函数寻优之前，对不同特征值进行去量纲处理。综上所述，本文提出的多特征优化目标函数对功率、谐波采样值及其相应特征向量进行归一化处理：当λ=0时，目标函数表示使用功率特征的单目标寻优；λ=1时，表示使用谐波特征的单目标寻优。通过多特征遗传目标函数优化模型，识别出附件四中的用电设备以及每个用电设备的实时用电量。关键词：非侵入式负荷监测BP人工神经网络聚类分析遗传算法2PowerDataMiningBasedonNon-intrusiveLoadDetectionandDecompositionAbstract:Thispaperaimstostudythemathematicalmodelandcalculationmethodfornon-invasiveloaddecompositionandidentification.Theoptimizedmodelisbasedonthepremiseofdatamining,andanaccurateandefficientidentificationanddecisionmethodisderivedtheamount.Firstly,accordingtotheworkingprincipleofthenon-intrusiveloadmonitoringsystem,thesteady-statecharacteristicsofthecurrentandtheharmoniccontentofthesteady-statecurrentareextracted,andaredividedintotwomajorcategoriesofelectricalappliancestocalculateaseriesofindexparameterstodistinguishtheloadcategory.Thepowerdifferenceofthehigh-powerequipmentduringthestart-stopswitchingisobvious,whilethepowerchangeofthelow-powerequipmentisrelativelyclose.Itismoreintuitivetostartandstoptheequipmentinhigh-frequencysampling.Whenmultipledeviceshavepoweraliasing,theidentificationoflow-powerdevicesisdifficult,andtheidentificationaccuracycanbeimprovedbyincreasingtheloadcharacteristics.Thatis,introducingamembershipfunctionintheformofanormaldistribution,whichisassignedtothesolutionweightw,throughthemeasurementoftheresultsobtainedbythealgorithmandtheapproximationofthecurrentstateofthepowerequipmentswitching,soastoobtaintheswitchingstateoftheelectricalequipment.Then,accordingtotheloadsignalwhichwillproducecertaincharacteristicswhenstartingtheappliance,thispaperusestheneuralnetworktoolboxwhichcomesfromMATLAB,throughthespecialpatternrecognitionneuralnetworkmodelPatternnet,andthetrainingalgorithmusesthequantitativeconjugategradientmethodtrainingscgtorealizeakindofNeuralnetworkbasedpatternrecognitionmethod.Secondly,thispaperpresentsthesteady-staterealpoweroffourlimitedmulti-statedevicesandthreekindsofstart/stoptwo-statedevicesinthethreedevicegroupsinAppendix3.Thepowerisdividedintoseveralclusteringpointstoachievethepurposeofobtainingafinitenumberofstatesbydiscretizingtheloadoftheequipment,therebyimprovingtheoperationdegreeandtheresultaccuracyofthesubsequentloaddecompositioncalculation.Accordingtotheobservationanalysis,thepowersectionrepresentativevalueofthesampledeviceisdetermined,andtheactivepowerchangeamountfromthesampledevicetothesteadystateisobtainedaccordingly.Finally,afterpartitioningthepowerofthesampledevicebyMATLAB.Finally,basedonthestudyofloaddataanalysisandfeatureextraction,amulti-featureoptimizationobjectivefunctionmodelofgeneticalgorithmisestablished.Throughgeneticiteration,theaccuratedecompositionandrecognitionofdifferentelectricalstatechangesareachieved.Thefirststepistogeneticallyencodeindividualindividuals.Becausethereisacertainamountofnoiseinreallife,itismoredifficulttoidentifylowpowerloadswithrelativelyclosepower.Inordertosolvethisproblem,activecurrentharmoniccharacteristicvaluesareintroducedtoincreasetheaccuracyandanti-jammingcapabilityoftherecognitionalgorithm.Therefore,thesingle-powertargetoptimizationfunctionmodelisoptimized.Beforethemulti-objectivefunctionisoptimized,thedifferenteigenvaluesaresubjectedtodimensioning.Insummary,themulti-featureoptimizationobjectivefunctionproposedinthispapernormalizesthepowerandharmonicsamplingvaluesandtheircorrespondingeigenvectors.Keywords:Non-intrusiveloadmonitoring,BPartificialneuralnetwork,Clusteranalysis,Geneticalgorithm3目录1.挖掘目标..................................................42.分析方法与过程............................................42.1.总体流程.........................................................42.2.具体步骤.........................................................52.3.结果分析........................................................353.结论.....................................................354.参考文献.................................................3641.挖掘目标1.1.挖掘背景在能源消费比重中，电能是整个社会生产和生活中重要的能源消费形式。为此，倡导节能减排是当前开展节能工