CN103942453A

CN103942453A - Intelligent electricity utilization anomaly detection method for non-technical loss

Info

Publication number: CN103942453A
Application number: CN201410191218.XA
Authority: CN
Inventors: 刘念; 刘杰
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2014-05-07
Filing date: 2014-05-07
Publication date: 2014-07-23

Abstract

The invention discloses an intelligent power consumption anomaly detection method aimed at non-technical losses and belongs to the technical field of power load analysis. The method is: 1) preprocessing the original data; 2) extracting features from the sample data; 3) dividing the sample into initial training samples and optimization samples; 4) sampling real-time data, extracting the characteristics of the samples, and generating test samples ;5) Use the GA algorithm to optimize the parameters and determine the optimal ELM parameter value; 6) Substitute the optimal ELM parameter value, training samples and test samples for abnormal detection; 7) If the detection time is an integer multiple of 72 hours, the statistical classification Accuracy and abnormal false detection rate; if the abnormal false detection rate exceeds the set limit, jump to 8), otherwise, jump to 4); 8) update user training samples, jump to 5). The method of the invention has clear physical concept, clear thinking, simple analysis and calculation, and can effectively solve the problem of on-line detection of non-technical loss for any electric load.

Description

An intelligent power consumption anomaly detection method for non-technical losses

技术领域technical field

本发明属于电力负荷分析技术领域，尤其涉及一种针对非技术性损失的智能用电异常检测方法。The invention belongs to the technical field of power load analysis, and in particular relates to an intelligent power consumption anomaly detection method aimed at non-technical losses.

背景技术Background technique

非技术性损失(Nontechnical Loss，NTL)是相对于技术性损失而提出来的概念，通常指已经被传送到用户侧使用但是未被计价的电能，与配网侧的用户窃电和一系列欺骗性用电行为有关。如今，非技术性损失已经成为影响电力公司收益的重要因素。Nontechnical loss (Nontechnical Loss, NTL) is a concept proposed relative to technical loss. It usually refers to the electricity that has been transmitted to the user side but has not been priced. related to electrical behavior. Today, non-technical losses have become an important factor affecting the power company's earnings.

针对非技术性损失，国外提出了许多关于NTL检测的方法，其检测方法非常丰富，包括统计学方法、决策树、人工神经网络、数据挖掘、知识发现以及最优路径树等，但都是离线检测。其历史数据仅由静态的历史负荷曲线提供，如果用户近期突然改变用电习惯，被误测为异常用电的可能性大大提升，同时对于历史数据有限甚至缺失的用户，不具有适用性。随着智能用电的发展，智能电表的普及使得在线检测作为未来异常检测的主要方式成为可能。同时，现有的研究多关注分类准确性，然而，在样本基数很大的情况下，误检率高的问题不容忽略。现场验证耗费人力物力，误检造成的浪费在样本基数很大的情况下，带来的损失同样很大。同时，异常误检率高，对环境的适用能力差的算法不易推广。For non-technical losses, many methods for NTL detection have been proposed abroad, and the detection methods are very rich, including statistical methods, decision trees, artificial neural networks, data mining, knowledge discovery, and optimal path trees, etc., but they are all offline detection . Its historical data is only provided by static historical load curves. If users suddenly change their electricity consumption habits in the near future, the possibility of being mistakenly detected as abnormal electricity consumption is greatly increased. At the same time, it is not applicable to users with limited or missing historical data. With the development of smart electricity consumption, the popularity of smart meters makes it possible to use online detection as the main way of abnormal detection in the future. At the same time, existing research focuses on classification accuracy. However, in the case of a large sample base, the problem of high false detection rate cannot be ignored. On-site verification consumes manpower and material resources, and the waste caused by false detection will also cause great losses when the sample base is large. At the same time, algorithms with high anomaly false detection rate and poor adaptability to the environment are not easy to promote.

发明内容Contents of the invention

针对上述现有技术存在的问题，本发明提出一种针对非技术性损失的智能用电异常检测方法，其特征在于，该检测方法的具体步骤为：Aiming at the problems existing in the above-mentioned prior art, the present invention proposes a method for detecting abnormality in intelligent power consumption aimed at non-technical losses, which is characterized in that the specific steps of the detection method are:

步骤1：对原始负荷数据进行预处理；Step 1: Preprocessing the original load data;

步骤2：对样本负荷数据进行特征提取；Step 2: Perform feature extraction on the sample load data;

步骤3：对样本进行划分，确定初始训练样本和寻优样本；Step 3: Divide the samples, determine the initial training samples and optimization samples;

步骤4：利用遗传算法GA进行离线参数寻优，确定最优极限学习机ELM参数值；Step 4: Use the genetic algorithm GA to optimize offline parameters and determine the optimal extreme learning machine ELM parameter values;

步骤5：采样实时数据，按步骤2的方法提取样本的特征，生成测试样本；Step 5: Sample real-time data, extract the characteristics of the sample according to the method of step 2, and generate a test sample;

步骤6：基于极限学习机ELM算法，代入最优极限学习机ELM参数值、初始训练样本和测试样本进行在线异常检测；Step 6: Based on the extreme learning machine ELM algorithm, substitute the optimal extreme learning machine ELM parameter values, initial training samples and test samples for online anomaly detection;

步骤7：如果检测时刻为72小时的整数倍，统计分类精度CA和异常误检率FDC；如果异常误检率超过设定限值FDC^lim跳转到步骤8，否则，跳转到步骤5；Step 7: If the detection time is an integer multiple of 72 hours, count the classification accuracy CA and abnormal false detection rate FDC; if the abnormal false detection rate exceeds the set limit value FDC ^lim , go to step 8, otherwise, go to step 5;

步骤8：更新用户训练样本，跳转到步骤4。Step 8: Update user training samples, skip to step 4.

所述步骤1具体为：The step 1 is specifically:

对原始负荷数据进行删选，删除非整点时刻的原始负荷数据，统一样本负荷数据采样频率；Delete and select the original load data, delete the original load data at off-hours, and unify the sampling frequency of the sample load data;

对于原始负荷数据缺失问题，采用同周期样本负荷数据的平均值进行填充或覆盖，其计算方法为：For the problem of missing original load data, the average value of sample load data in the same period is used to fill or cover, and the calculation method is:

${x x}_{i i} = = \frac{{x x}_{i i - - W W} + + {x x}_{i i + + W W}}{22};; W W = = 77 \times \times 24 twenty four;;$

其中，x_i表示第i个时刻的样本负荷数据；W为一周的理论采样样本数。Among them, x _i represents the sample load data at the i-th moment; W is the theoretical sampling number of samples in one week.

所述步骤2具体为：The step 2 is specifically:

对样本的原始特征集进行预处理，转化为输出特征集；Preprocess the original feature set of the sample and convert it into an output feature set;

所述原始特征集由基本数据元素组成，为原始特征集中第m个用户的第i个时刻的样本，每个都是一个包含H个特征值的向量：The raw feature set consists of basic data elements composition, is the sample of the i-th moment of the m-th user in the original feature set, each Both are a vector containing H eigenvalues:

${x x}_{i i}^{m m . . 00} = = {{{x x}_{h h}^{00};; h h = = 11 ~ ~ H h}},, H h = = 1515;;$

原始特征集包括采样时间、采样终端号、有功功率、无功功率、A相电流、B相电流、C相电流、A相电压、B相电压、C相电压、15分钟平均有功功率、功率因数、累计有功功率、逻辑地址和户号；The original feature set includes sampling time, sampling terminal number, active power, reactive power, phase A current, phase B current, phase C current, phase A voltage, phase B voltage, phase C voltage, 15-minute average active power, and power factor , cumulative active power, logical address and account number;

所述输出特征集由基本数据元素组成，为输出特征集中第m个用户的第i个时刻的样本，每个都是一个包含G个特征值的向量：The output feature set consists of basic data elements composition, is the sample of the i-th moment of the m-th user in the output feature set, each is a vector containing G eigenvalues:

${x x}_{i i}^{m m} = = f f (({x x}_{i i}^{m m . . 00})) = = {{{x x}_{g g};; g g = = 11 ~ ~ G G}};; G G = = 1616;;$

输出特征集包括有功功率、无功功率、功率因数、累计有功功率、日最大负荷、日最大负荷利用小时、用电系数、个人离散系数、公共离散系数、离散系数差值、前一小时用电量同期比值、昨天同一小时用电量同期比值、上周同一小时用电量同期比值、峰谷划分、峰谷连续数和电量波动；The output feature set includes active power, reactive power, power factor, cumulative active power, daily maximum load, daily maximum load utilization hour, power consumption coefficient, individual dispersion coefficient, public dispersion coefficient, dispersion coefficient difference, power consumption in the previous hour The ratio of electricity consumption over the same period, the ratio of electricity consumption in the same hour yesterday, the ratio of electricity consumption in the same hour last week, the division of peak and valley, the number of consecutive peaks and valleys, and the fluctuation of electricity;

所述用电系数的计算公式为：The power utilization factor The calculation formula is:

${PKAP PKAP}_{i i}^{m m} = = \frac{{CAP CAP}_{i i}^{m m} - - {CAP CAP}_{i i - - 11}^{m m}}{{SEC SEC}^{m m}};;$

其中，表示第m个用户第i个时刻样本的累计有功功率；表示第m个用户第i-1个时刻样本的累计有功功率；SEC^m表示第m个用户所在供电区域的公用变压器容量；in, Indicates the cumulative active power of the m-th user at the i-th time sample; Indicates the cumulative active power of the mth user at the i-1th time sample; SEC ^m represents the capacity of the public transformer in the power supply area where the mth user is located;

所述个人离散系数的计算公式为；The individual coefficient of dispersion The calculation formula is;

$\begin{matrix} {DC DC}_{i i}^{m m} = = \sqrt{\frac{11}{n no - - 11} {Σ Σ}_{t t = = i i - - n no}^{i i} (({AP AP}_{t t}^{m m} - - \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}}))} / / \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} \\ \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} = = avg avg (({AP AP}_{i i - - n no}^{m m},, \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot;,, {AP AP}_{i i}^{m m})) \end{matrix};;;$

其中，表示第m个用户第i个时刻样本的个人离散系数；表示第m个用户第t个时刻样本的有功功率；avg是计算平均值的函数，表示在分析时段内第m个用户的有功功率均值；n为24；in, Indicates the personal dispersion coefficient of the m-th user at the i-th time sample; Indicates the active power of the mth user at the tth time sample; avg is a function to calculate the average value, Indicates the average active power of the mth user within the analysis period; n is 24;

所述公共离散系数的计算方式为计算同一供电区域内所有用户的个人离散系数的均值；The public coefficient of dispersion The calculation method of is to calculate the mean value of the individual dispersion coefficient of all users in the same power supply area;

所述日最大负荷的计算公式为：The maximum daily load The calculation formula is:

${AP AP}_{i i}^{m m} {| |}_{max max} = = max max (({AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four}^{m m},, {AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four + + 11}^{m m},, . . . . . .,, {AP AP}_{t t}^{m m},, . . . . . .,, {AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four + + 23 twenty three}^{m m}));;$

其中，int是取整函数；Among them, int is the integer function;

所述日最大负荷利用小时t_hour的计算公式为：The formula for calculating the daily maximum load utilization hour t _hour is:

${t t}_{hour hours} = = \frac{{CAP CAP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four + + 23 twenty three}^{m m} - - {CAP CAP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four}^{m m}}{{AP AP}_{i i}^{m m} {| |}_{max max}};;$

所述离散系数差值的计算公式为：The coefficient of dispersion difference The calculation formula is:

${ΔDC ΔDC}_{i i}^{m m} = = | | D D. {C C}_{i i}^{m m} - - {PDC PDC}_{i i}^{m m} | |;;$

所述用电量同期比值的计算公式为：The ratio of electricity consumption over the same period The calculation formula is:

${PC PC}_{t t}^{m m} = = (({AP AP}_{{t t}_{c c}}^{m m} - - {AP AP}_{t t}^{m m})) / / {AP AP}_{{t t}_{c c}}^{m m};;$

其中，表示第m个用户第t_c个时刻样本的有功功率；t_c表示同期被比较的时段，t_c包括前一小时、昨天同一小时或上周同一小时；in, Indicates the active power of the m-th user at the tc- _th time sample; _tc indicates the period of comparison during the same period, and _tc includes the previous hour, the same hour yesterday or the same hour last week;

所述峰谷连续数QC为峰、谷时段长短的计数标记，峰谷连续数QC初始值设置为0；当时，峰谷划分的属性值为1，峰谷连续数QC加1；当时，峰谷划分的属性值为零；当峰谷划分的属性值为-1，峰谷连续数QC减1；若改变符号，峰谷连续数QC重新置零；The peak and valley continuous number QC is the counting mark of the length of the peak and valley periods, and the initial value of the peak and valley continuous number QC is set to 0; When , the attribute value of the peak-valley division is 1, the continuous number of peaks and valleys QC plus 1; when When , the attribute value of the peak-valley division is zero; when Attribute value for peak-valley division is -1, the continuous number of peaks and valleys QC minus 1; if Change the sign, and reset the peak-to-valley continuous number QC to zero;

所述电量波动的计算公式为：The power fluctuation The calculation formula is:

${VAR VAR}_{t t}^{m m} = = | | {AP AP}_{t t}^{m m} - - \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} | | . .$

所述步骤3具体为：The step 3 is specifically:

选取T时刻以后的样本作为参数优选的依据，则对应的寻优样本表示为：The samples after time T are selected as the basis for parameter optimization, and the corresponding optimization samples are expressed as:

${V V}_{T T}^{m m} = = {{(({l l}_{i i},, {x x}_{i i}^{m m}))}};; i i = = T T,, T T + + 11,, . . . . . .;;$

其中，为第m个用户异常检测模型的寻优样本；l_i取值为1或-1，分别表示第i个时刻样本正常用电或异常用电；in, is the optimization sample of the m-th user anomaly detection model; the value of l _i is 1 or -1, which respectively represent the normal or abnormal power consumption of the sample at the i-th moment;

对应选择T时刻的前驱样本作为异常检测模型的初始训练样本，其表示为：Correspondingly, the precursor sample at time T is selected as the initial training sample of the anomaly detection model, which is expressed as:

${M m}_{T T}^{m m . . 00} = = {{(({l l}_{i i},, {x x}_{i i}^{m m}))}};; i i = = 1,2 1,2,, . . . . . .,, T T - - 11;;$

其中，为第m个用户异常检测模型的初始训练样本。in, is the initial training sample for the mth user anomaly detection model.

所述步骤4中遗传算法GA离线参数寻优中的适应度函数取为异常检测准确率ADC，计算公式为：The fitness function in the genetic algorithm GA offline parameter optimization in the step 4 is taken as the abnormal detection accuracy rate ADC, and the calculation formula is:

$ADC ADC = = d d / / ((b b + + d d + + \frac{11}{22}));;$

其中，b表示被极限学习机ELM算法检测为正常但其实为异常的样本数量；d表示被极限学习机ELM算法检测为异常且确实为异常的样本数量。Among them, b represents the number of samples that are detected as normal by the extreme learning machine ELM algorithm but are actually abnormal; d represents the number of samples that are detected as abnormal by the extreme learning machine ELM algorithm and are indeed abnormal.

所述步骤7中的分类精度CA表示为：The classification accuracy CA in the step 7 is expressed as:

CA＝(a₁+d₁)/(a₁+b₁+c₁+d₁)；CA=(a ₁ +d ₁ )/(a ₁ +b ₁ +c ₁ +d ₁ );

异常误检率FDC表示为：The abnormal false detection rate FDC is expressed as:

FDC＝(c₁+d₁+1)/(b₁+d₁+1)；FDC=(c ₁ +d ₁ +1)/(b ₁ +d ₁ +1);

其中，在线阶段，a₁表示被极限学习机ELM算法检测为正常且确实为正常的样本数量；b₁表示被极限学习机ELM算法检测为正常但其实为异常的样本数量；c₁表示被极限学习机ELM算法检测为异常但其实为正常的样本数量；d₁表示被极限学习机ELM算法检测为异常且确实为异常的样本数量。Among them, in the online stage, a ₁ represents the number of samples that are detected as normal by the extreme learning machine ELM algorithm and are indeed normal; b ₁ represents the number of samples _that are detected as normal by the extreme learning machine ELM algorithm but are actually abnormal; The number of samples detected as abnormal by the ELM algorithm of the learning machine but actually normal; d ₁ represents the number of samples detected as abnormal by the ELM algorithm of the extreme learning machine and indeed abnormal.

发明的有益效果：本发明提出的异常检测方法，以异常误检率作为更新用户训练样本和ELM参数的触发机制，以分类精度作为评估检测效果的指标，物理概念明确，思路清晰，分析计算简便，可以有效解决对任意用电负荷进行非技术性损失的在线检测的问题。Beneficial effects of the invention: the anomaly detection method proposed by the present invention uses the abnormal false detection rate as the trigger mechanism for updating user training samples and ELM parameters, and uses the classification accuracy as the indicator for evaluating the detection effect. The physical concept is clear, the thinking is clear, and the analysis and calculation are simple. , can effectively solve the problem of on-line detection of non-technical loss for any power load.

附图说明Description of drawings

图1是针对非技术性损失的智能用电异常检测原理图；Figure 1 is a schematic diagram of abnormal detection of intelligent power consumption for non-technical losses;

图2为本发明提出的智能用电异常检测方法的流程图。FIG. 2 is a flow chart of the method for detecting abnormal intelligent power consumption proposed by the present invention.

具体实施方式Detailed ways

下面结合附图，对检测实施例作详细说明。应该强调的是，下述说明仅仅是示例性的，而不是为了限制本发明的范围及其应用。The detection embodiments will be described in detail below in conjunction with the accompanying drawings. It should be emphasized that the following description is only exemplary and not intended to limit the scope of the invention and its application.

本方法将离线检测和在线检测技术相结合，提出一种针对非技术性损失的智能用电异常检测方法，其检测原理如图1所示。This method combines offline detection and online detection technology, and proposes an intelligent power consumption anomaly detection method for non-technical losses. The detection principle is shown in Figure 1.

图2是本发明提出的智能用电异常检测方法的流程图，则具体步骤为：Fig. 2 is the flowchart of the abnormal detection method of intelligent power consumption proposed by the present invention, and the specific steps are:

步骤1：对原始负荷数据进行预处理。Step 1: Preprocessing the raw load data.

由于不同的用户数据采集方式不统一，例如商业用户的数据是每小时统计一次，而工业用户则是部分每小时统计一次，部分每半小时或每15分钟统计一次，因此，需要对原始负荷数据进行删选，删除非整点时刻的原始负荷数据，统一样本负荷数据采样频率为1小时/次，便于进一步分析。Since different user data collection methods are not uniform, for example, the data of commercial users is counted every hour, while the data of industrial users is counted once every hour, and some are counted once every half hour or every 15 minutes. Therefore, it is necessary to collect raw load data Carry out deletion, delete the original load data at off-hours, and unify the sample load data sampling frequency to 1 hour/time, which is convenient for further analysis.

步骤2：对样本负荷数据进行特征提取。Step 2: Perform feature extraction on the sample loading data.

对样本负荷数据进行特征提取，就是对样本的原始特征集进行预处理，然后转化为输出特征集。Feature extraction for sample load data is to preprocess the original feature set of the sample and then convert it into an output feature set.

原始特征集由基本数据元素组成，为原始特征集中第m个用户的第i个时刻的样本，每个都是一个包含H个特征值的向量：The original feature set consists of basic data elements composition, is the sample of the i-th moment of the m-th user in the original feature set, each Both are a vector containing H eigenvalues:

原始特征集包括采样时间、采样终端号、有功功率、无功功率、A相电流、B相电流、C相电流、A相电压、B相电压、C相电压、15分钟平均有功功率、功率因数、累计有功功率、逻辑地址和户号。The original feature set includes sampling time, sampling terminal number, active power, reactive power, phase A current, phase B current, phase C current, phase A voltage, phase B voltage, phase C voltage, 15-minute average active power, and power factor , cumulative active power, logical address and account number.

输出特征集由基本数据元素组成，为输出特征集中第m个用户的第i个时刻的样本数据，每个都是一个包含G个特征值的向量：The output feature set consists of basic data elements composition, is the sample data of the i-th moment of the m-th user in the output feature set, each is a vector containing G eigenvalues:

输出特征集包括有功功率、无功功率、功率因数、累计有功功率、日最大负荷、日最大负荷利用小时、用电系数、个人离散系数、公共离散系数、离散系数差值、前一小时用电量同期比值、昨天同一小时用电量同期比值、上周同一小时用电量同期比值、峰谷划分、峰谷连续数和电量波动。The output feature set includes active power, reactive power, power factor, cumulative active power, daily maximum load, daily maximum load utilization hour, power consumption coefficient, individual dispersion coefficient, public dispersion coefficient, dispersion coefficient difference, power consumption in the previous hour The ratio of electricity consumption over the same period, the ratio of electricity consumption in the same hour yesterday, the ratio of electricity consumption in the same hour last week, the division of peak and valley, the number of consecutive peaks and valleys, and the fluctuation of electricity.

用电系数的计算公式为：Electricity coefficient The calculation formula is:

其中，表示第m个用户第i个时刻用电量与所在供电区域的公用变压器容量的比值；表示第m个用户第i个时刻样本的累计有功功率；表示第m个用户第i-1个时刻样本的累计有功功率；SEC^m表示第m个用户所在供电区域的公用变压器容量；in, Indicates the ratio of the power consumption of the m-th user at the i-th moment to the capacity of the public transformer in the power supply area; Indicates the cumulative active power of the m-th user at the i-th time sample; Indicates the cumulative active power of the mth user at the i-1th time sample; SEC ^m represents the capacity of the public transformer in the power supply area where the mth user is located;

个人离散系数的计算公式为；individual coefficient of dispersion The calculation formula is;

其中，表示第m个用户第i个时刻样本的个人离散系数；表示第m个用户第t个时刻样本的有功功率；avg是计算平均值的函数，表示在分析时段内用户的有功功率均值；离散系数基于前24小时的数据求得，因此n取值为24。in, Indicates the personal dispersion coefficient of the m-th user at the i-th time sample; Indicates the active power of the mth user at the tth time sample; avg is a function to calculate the average value, Indicates the average active power of users within the analysis period; the dispersion coefficient is obtained based on the data of the previous 24 hours, so the value of n is 24.

公共离散系数的计算方式为计算同一供电区域内所有用户的个人离散系数的均值；common coefficient of dispersion The calculation method of is to calculate the mean value of the individual dispersion coefficient of all users in the same power supply area;

日最大负荷的计算公式为：Daily maximum load The calculation formula is:

其中，int是取整函数。Among them, int is the integer function.

日最大负荷利用小时t_hour的计算公式为：The calculation formula of daily maximum load utilization hour t _hour is:

离散系数差值的计算公式为：difference coefficient of dispersion The calculation formula is:

用电量同期比值的计算公式为：The ratio of electricity consumption over the same period The calculation formula is:

其中，表示第m个用户第t_c个时刻样本的有功功率；t_c表示同期被比较的时段，t_c包括前一小时、昨天同一小时或上周同一小时。in, Indicates the active power of the m-th user at the tc- _th time sample; _tc indicates the period to be compared in the same period, and _tc includes the previous hour, the same hour yesterday or the same hour last week.

峰谷连续数QC为峰、谷时段长短的计数标记，峰谷连续数QC初始值设置为0；时，峰谷划分的属性值为1，峰谷连续数QC加1；当时，峰谷划分的属性值为零；当峰谷划分的属性值为-1，峰谷连续数QC减1；若改变符号，峰谷连续数QC重新置零，如此往复。The peak and valley continuous number QC is the counting mark of the length of the peak and valley periods, and the initial value of the peak and valley continuous number QC is set to 0; When , the attribute value of the peak-valley division is 1, the continuous number of peaks and valleys QC plus 1; when When , the attribute value of the peak-valley division is zero; when Attribute value for peak-valley division is -1, the continuous number of peaks and valleys QC minus 1; if Change the sign, the peak and valley continuous number QC is reset to zero, and so on.

电量波动的计算公式为：Power fluctuations The calculation formula is:

步骤3：对样本进行划分，确定初始训练样本和寻优样本。Step 3: Divide the samples, determine the initial training samples and optimization samples.

步骤4：利用遗传算法GA进行离线参数寻优，确定最优极限学习机ELM参数值。Step 4: Use the genetic algorithm GA to optimize offline parameters, and determine the optimal extreme learning machine ELM parameter values.

基于遗传算法的参数优选过程可以归纳为：The parameter optimization process based on genetic algorithm can be summarized as:

(1)初始化：随机生成一组初始参数C⁰和L⁰，初始参数取值区间分别设定为[0.1，100]和[0.001，100]，并对每个初始参数进行编码，进而构造初始种群。(1) Initialization: Randomly generate a set of initial parameters C ⁰ and L ⁰ , set the value intervals of the initial parameters to [0.1, 100] and [0.001, 100] respectively, and encode each initial parameter, and then construct the initial population.

(2)适应度评估：将极限学习机参数代入求得相应的适应度函数，适应度函数取为异常检测准确率ADC，表示异常检测正确的样本数占所有检测为异常样本的比例。(2) Fitness evaluation: Substituting the parameters of the extreme learning machine into the corresponding fitness function, the fitness function is taken as the abnormality detection accuracy rate ADC, which indicates the proportion of the number of samples that are correctly detected as abnormal to all samples detected as abnormal.

按照下面公式求得异常检测结果：The anomaly detection result is obtained according to the following formula:

$CL CL = = ELM ELM (({M m}_{T T}^{m m . . 00},, {V V}_{T T}^{m m},, {C C}^{{g g}_{11}},, {L L}^{{g g}_{11}})),, {g g}_{11} = = 1,2,3 1,2,3,, . . . . . .;;$

其中，表示基于给定初始参数和寻优样本求出分类结果的函数，g₁表示遗传算法的训练代数，和表示训练g₁代后得到的参数，CL(Classification Labels)为相应的分类结果，CL取值为1或-1，分别表示被极限学习机ELM算法检测为正常或异常。in, Indicates the function of obtaining the classification result based on the given initial parameters and optimization samples, g ₁ indicates the training algebra of the genetic algorithm, and Indicates the parameters obtained after training g for ₁ generation, CL (Classification Labels) is the corresponding classification result, and the value of CL is 1 or -1, indicating that it is detected as normal or abnormal by the extreme learning machine ELM algorithm.

将样本的真实情况和检测出来的异常情况作比较，得出如表1所示的结果：Comparing the real situation of the sample with the abnormal situation detected, the results shown in Table 1 are obtained:

表1样本的真实情况和检测出来的异常情况比较结果Table 1 Comparison results between the real situation of the sample and the detected abnormal situation

令a表示被ELM算法检测为正常且确实为正常的样本数量；b表示被ELM算法检测为正常但其实为异常的样本数量；c表示被ELM算法检测为异常但其实为正常的样本数量；d表示被ELM算法检测为异常且确实为异常的样本数量。则适应度函数ADC的计算公式为：Let a represent the number of samples that are detected as normal by the ELM algorithm and are indeed normal; b represents the number of samples that are detected as normal by the ELM algorithm but are actually abnormal; c represents the number of samples that are detected as abnormal by the ELM algorithm but are actually normal; d Indicates the number of samples that are detected as abnormal by the ELM algorithm and are indeed abnormal. Then the calculation formula of the fitness function ADC is:

$ADC ADC = = d d / / ((b b + + d d + + \frac{11}{22})) . .$

(3)选择、交叉：选取若干适应度较高的个体作为父代种群，使用启发式交叉函数，依据比例系数向较差父代移动一段距离作为交叉形成的子代，即(3) Selection and crossover: Select a number of individuals with high fitness as the parent population, use the heuristic crossover function, and move a certain distance to the poorer parent according to the proportional coefficient as the offspring formed by crossover, that is

${P P}_{11}^{{g g}_{11} + + 11} = = {P P}_{11}^{{g g}_{11}} + + R R (({P P}_{22}^{{g g}_{11}} - - {P P}_{11}^{{g g}_{11}}));;$

其中，R为比例系数，取值为(1，2)范围内的随机数，为父代参数值，其中为适应度较高的父代，为子代参数值。Among them, R is a proportional coefficient, and the value is a random number in the range of (1, 2), is the parent parameter value, where For parents with higher fitness, is the descendant parameter value.

${P P}_{11}^{{g g}_{11} + + 11} = = {P P}_{11}^{{g g}_{11} + + 11} + + rand rand ((g g));;$

其中，rand函数是产生随机数的随机函数。Among them, the rand function is a random function that generates random numbers.

(4)检测：当种群代数超过最大代数或连续10代的适应度变化小于1％时，终止参数寻优过程，输出当前的最优参数值C^opt和L^opt；否则跳转步骤(2)。(4) Detection: When the population algebra exceeds the maximum algebra or the fitness change of 10 consecutive generations is less than 1%, the parameter optimization process is terminated, and the current optimal parameter values C ^opt and L ^opt are output; otherwise, jump to step (2) .

步骤5：采样实时数据，按步骤2的方法提取样本的特征，生成测试样本。Step 5: Sample real-time data, extract the characteristics of the sample according to the method in step 2, and generate a test sample.

在线采样的频率为每小时一次，在进行在线异常检测时，检测时刻T₀的原始特征样本为：The frequency of online sampling is once per hour. When performing online anomaly detection, the original feature samples at the detection time T ₀ are:

${TE TE}_{t t}^{m m . . 00} = = {{{x x}_{h h}^{00};; h h = = 11 ~ ~ H h}};; t t = = {T T}_{00},, H h = = 1515;;$

其中，为第m个用户的第t个时刻的原始特征样本。in, is the original feature sample of the mth user at the tth moment.

与步骤2中原始数据的处理方法一致，在线采样的T₀时刻的测试样本同样要将原始特征集转化为指定的输出特征集，即Consistent with the processing method of the original data in step 2, the test sample at time T ₀ of online sampling also needs to transform the original feature set into the specified output feature set, namely

${TE TE}_{t t}^{m m} = = f f (({TE TE}_{t t}^{m m . . 00})) = = {{{x x}_{g g};; g g = = 11 ~ ~ G G}};; t t = = {T T}_{00},, G G = = 1616;;$

其中，为第m个用户的第t个时刻的输出特征样本。in, is the output feature sample of the mth user at the tth moment.

步骤6：基于极限学习机ELM算法，代入极限学习机ELM最优参数值、训练样本和测试样本进行在线异常检测。Step 6: Based on the extreme learning machine ELM algorithm, the optimal parameter values of the extreme learning machine ELM, training samples and test samples are substituted for online anomaly detection.

基于极限学习机ELM算法，使用极限学习机ELM最优参数值、训练样本和测试样本建立异常检测模型，公式为：Based on the extreme learning machine ELM algorithm, use the extreme learning machine ELM optimal parameter value, training samples and test samples to establish an anomaly detection model, the formula is:

${\overset{^^}{l l}}_{t t} = = ELM ELM (({M m}^{m m . . k k},, {TE TE}_{t t}^{m m},, {C C}^{m m . . k k},, {L L}^{m m . . k k}));; k k = = 1,2,3 1,2,3 . . . . . .;;$

其中，为第t个时刻的检测结果，其取值为1或-1；k为在线更新次数；M^m.k为第m个用户第k次更新后的异常检测模型的训练样本；为实时数据的输出特征集中第m个用户第t个时刻的测试样本；C^m.k、L^m.k为遗传算法GA求得的对应于训练样本M^m.k的最优参数值。in, is the detection result at the tth moment, and its value is 1 or -1; k is the number of online updates; M ^mk is the training sample of the anomaly detection model updated by the mth user for the kth time; is the test sample of the mth user in the output feature set of real-time data at the tth moment; C ^mk and L ^mk are the optimal parameter values corresponding to the training sample M ^mk obtained by the genetic algorithm GA.

步骤7：如果检测时刻为72小时的整数倍，统计分类精度和异常误检率指标；如果异常误检率FDC突然大幅攀升超过设定限值FDC^lim，说明用户发现了重大变故，如生产经营方向调整等，此时跳转到步骤8；否则，跳转到步骤4。Step 7: If the detection time is an integer multiple of 72 hours, count the classification accuracy and abnormal false detection rate indicators; if the abnormal false detection rate FDC suddenly rises sharply and exceeds the set limit value FDC ^lim , it means that the user has discovered a major change, such as production and operation Direction adjustment, etc., skip to step 8 at this time; otherwise, skip to step 4.

分类精度CA作为传统评价标准已经被广泛用于相关研究，表示检测准确率，具体表达式为：Classification accuracy CA as a traditional evaluation standard has been widely used in related research, indicating the detection accuracy, the specific expression is:

由于检测到异常用电行为之后，被检测为异常的用电数据通常需要进行人工现场确认与处理，而现场检查需要消耗大量的人力与物力，故选择异常误检率FDC作为算法性能的另一个评价指标。该指标通过统计被错检为异常用户的样本比例，来反映异常检测算法带来的经济成本与人力资源价值，可表示为：After detecting abnormal power consumption behavior, the abnormal power consumption data usually needs to be manually confirmed and processed on site, and the on-site inspection needs to consume a lot of manpower and material resources, so the abnormal false detection rate FDC is selected as another indicator of algorithm performance. evaluation index. This indicator reflects the economic cost and human resource value brought by the anomaly detection algorithm by counting the proportion of samples that are wrongly detected as abnormal users, which can be expressed as:

FDC＝(c₁+d₁+1)/(b₁+d₁+1)。FDC=(c ₁ +d ₁ +1)/(b ₁ +d ₁ +1).

步骤8：更新用户训练样本，跳转到步骤5。Step 8: Update user training samples, skip to step 5.

当异常误检率FDC突然大幅攀升越过设定值，说明用户发生了重大变故，如经营方向调整等。此时，基于历史采样和验证后的测试样本在线更新训练样本，将最近期72小时的测试样本更新到训练样本，同时删除原训练样本中最前72小时的样本数据，以维持训练样本数稳定在一个固定的数值，具体公式为：When the abnormal false detection rate FDC suddenly rises sharply and exceeds the set value, it indicates that a major change has occurred to the user, such as an adjustment of the business direction. At this time, the training samples are updated online based on historical sampling and verified test samples, and the latest 72-hour test samples are updated to the training samples, and the sample data of the first 72 hours in the original training samples are deleted at the same time, so as to keep the number of training samples stable at A fixed value, the specific formula is:

M^m.k＝M^m.(k-1)+TE^m.k-M_k M ^mk ＝M ^m.(k-1) +TE ^mk -M _k

${TE TE}^{m m . . k k} = = {{(({l l}_{{t t}_{i i}},, {TM tm}_{{t t}_{i i}}^{m m}))}},, {t t}_{i i} = = t t - - 7171,, t t - - 7070,, . . . . . .,, t t;;$

${M m}_{k k} = = {{(({l l}_{i i},, {x x}_{i i}^{m m}))}} &Subset; &Subset; {M m}_{T T}^{m m};; i i = = 7272 k k - - 7171,, \cdot \cdot \cdot \cdot \cdot \cdot,, 7272 k k - - 1,72 1,72 k k$

其中，M^m.(k-1)为第m个用户第k-1次更新后的异常检测模型的训练样本；TE^m.k表示第m个用户第k次更新训练样本的过程中，所提取的最近期的72小时的测试样本的集合；表示第m个用户第t_i时刻的测试样本；M_k表示要删除的原训练样本中最前72小时的样本数据集，其是k-1次更新后所得的训练样本M^m.(k-1)的子集。Among them, M ^m.(k-1) is the training sample of the anomaly detection model updated by the m-th user for the k-1th time; TE ^mk represents the extracted training sample during the k-th update of the m-th user A collection of test samples from the most recent 72 hours; Indicates the test sample of the mth user at the time t _i ; M _k indicates the sample data set of the first 72 hours in the original training sample to be deleted, which is the training sample M ^{m obtained after k-1 updates.(k-1 )} subset.

本发明可以适用于各种类型的用户负荷数据的在线监测，包括而不局限于历史数据缺乏、用户突然改变用电习惯等类型的用电数据。The present invention can be applied to the online monitoring of various types of user load data, including but not limited to the lack of historical data, the user's sudden change of electricity consumption habits and other types of electricity consumption data.

以上所述，仅为本发明较佳的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到的变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应该以权利要求的保护范围为准。The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present invention can easily think of changes or Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims

1. A method for detecting abnormality in intelligent power consumption aimed at non-technical loss, characterized in that, the specific steps of the detection method are:

Step 1: Preprocessing the original load data;

Step 2: Perform feature extraction on the sample load data;

Step 3: Divide the samples, determine the initial training samples and optimization samples;

Step 4: Use the genetic algorithm GA to optimize offline parameters and determine the optimal extreme learning machine ELM parameter values;

Step 5: Sample real-time data, extract the characteristics of the sample according to the method of step 2, and generate a test sample;

Step 6: Based on the extreme learning machine ELM algorithm, substitute the optimal extreme learning machine ELM parameter values, initial training samples and test samples for online anomaly detection;

Step 7: If the detection time is an integer multiple of 72 hours, count the classification accuracy CA and abnormal false detection rate FDC; if the abnormal false detection rate exceeds the set limit value FDC ^lim , go to step 8, otherwise, go to step 5;

Step 8: Update user training samples, skip to step 4.

2. A method for detecting abnormalities in intelligent power consumption aimed at non-technical losses according to claim 1, wherein the step 1 is specifically:

Delete and select the original load data, delete the original load data at off-hours, and unify the sampling frequency of the sample load data;

For the problem of missing original load data, the average value of sample load data in the same period is used to fill or cover, and the calculation method is:

{x x}_{i i} = = \frac{{x x}_{i i - - W W} + + {x x}_{i i + + W W}}{22};; W W = = 77 \times \times 24 twenty four;;

Among them, x _i represents the sample load data at the i-th moment; W is the theoretical sampling number of samples in one week.

3. A method for detecting abnormalities in intelligent power consumption for non-technical losses according to claim 2, characterized in that, the step 2 is specifically:

Preprocess the original feature set of the sample and convert it into an output feature set;

The raw feature set consists of basic data elements composition, is the sample of the i-th moment of the m-th user in the original feature set, each Both are a vector containing H eigenvalues:

{x x}_{i i}^{m m . . 00} = = {{{x x}_{h h}^{00};; h h = = 11 ~ ~ H h}},, H h = = 1515;;

The original feature set includes sampling time, sampling terminal number, active power, reactive power, phase A current, phase B current, phase C current, phase A voltage, phase B voltage, phase C voltage, 15-minute average active power, and power factor , cumulative active power, logical address and account number;

The output feature set consists of basic data elements composition, is the sample of the i-th moment of the m-th user in the output feature set, each is a vector containing G eigenvalues:

{x x}_{i i}^{m m} = = f f (({x x}_{i i}^{m m . . 00})) = = {{{x x}_{g g};; g g = = 11 ~ ~ G G}};; G G = = 1616;;

The output feature set includes active power, reactive power, power factor, cumulative active power, daily maximum load, daily maximum load utilization hour, power consumption coefficient, individual dispersion coefficient, public dispersion coefficient, dispersion coefficient difference, power consumption in the previous hour The ratio of electricity consumption over the same period, the ratio of electricity consumption in the same hour yesterday, the ratio of electricity consumption in the same hour last week, the division of peak and valley, the number of consecutive peaks and valleys, and the fluctuation of electricity;

The power utilization factor The calculation formula is:

{PKAP PKAP}_{i i}^{m m} = = \frac{{CAP CAP}_{i i}^{m m} - - {CAP CAP}_{i i - - 11}^{m m}}{{SEC SEC}^{m m}};;

in, Indicates the cumulative active power of the m-th user at the i-th time sample; Indicates the cumulative active power of the mth user at the i-1th time sample; SEC ^m represents the capacity of the public transformer in the power supply area where the mth user is located;

The individual coefficient of dispersion The calculation formula is;

\begin{matrix} {DC DC}_{i i}^{m m} = = \sqrt{\frac{11}{n no - - 11} {Σ Σ}_{t t = = i i - - n no}^{i i} (({AP AP}_{t t}^{m m} - - \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}}))} / / \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} \\ \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} = = avg avg (({AP AP}_{i i - - n no}^{m m},, \cdot &Center Dot; \cdot \cdot \cdot \cdot,, {AP AP}_{i i}^{m m})) \end{matrix};;;

in, Indicates the personal dispersion coefficient of the m-th user at the i-th time sample; Indicates the active power of the mth user at the tth time sample; avg is a function to calculate the average value, Indicates the average active power of the mth user within the analysis period; n is 24;

The public coefficient of dispersion The calculation method of is to calculate the mean value of the individual dispersion coefficient of all users in the same power supply area;

The maximum daily load The calculation formula is:

{AP AP}_{i i}^{m m} {| |}_{max max} = = max max (({AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four}^{m m},, {AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four + + 11}^{m m},, . . . . . .,, {AP AP}_{t t}^{m m},, . . . . . .,, {AP AP}_{int int ((\frac{t t}{24 twenty four})) \cdot \cdot 24 twenty four + + 23 twenty three}^{m m}));;

Among them, int is the integer function;

The formula for calculating the daily maximum load utilization hour t _hour is:

{t t}_{hour hours} = = \frac{{CAP CAP}_{int int ((\frac{t t}{24 twenty four})) \cdot &Center Dot; 24 twenty four + + 23 twenty three}^{m m} - - {CAP CAP}_{int int ((\frac{t t}{24 twenty four})) \cdot \cdot 24 twenty four}^{m m}}{{AP AP}_{i i}^{m m} {| |}_{max max}};;

The coefficient of dispersion difference The calculation formula is:

{ΔDC ΔDC}_{i i}^{m m} = = | | D D. {C C}_{i i}^{m m} - - {PDC PDC}_{i i}^{m m} | |;;

The ratio of electricity consumption over the same period The calculation formula is:

{PC PC}_{t t}^{m m} = = (({AP AP}_{{t t}_{c c}}^{m m} - - {AP AP}_{t t}^{m m})) / / {AP AP}_{{t t}_{c c}}^{m m};;

in, Indicates the active power of the m-th user at the tc- _th time sample; _tc indicates the period of comparison during the same period, and _tc includes the previous hour, the same hour yesterday or the same hour last week;

The peak and valley continuous number QC is the counting mark of the length of the peak and valley periods, and the initial value of the peak and valley continuous number QC is set to 0; When , the attribute value of the peak-valley division is 1, the continuous number of peaks and valleys QC plus 1; when When , the attribute value of the peak-valley division is zero; when Attribute value for peak-valley division is -1, the continuous number of peaks and valleys QC minus 1; if Change the sign, and reset the peak-to-valley continuous number QC to zero;

The power fluctuation The calculation formula is:

{VAR VAR}_{t t}^{m m} = = | | {AP AP}_{t t}^{m m} - - \overset{&OverBar; &OverBar;}{{AP AP}_{t t}^{m m}} | | . .

4. A method for detecting abnormalities in intelligent power consumption for non-technical losses according to claim 3, characterized in that, the step 3 is specifically:

The samples after time T are selected as the basis for parameter optimization, and the corresponding optimization samples are expressed as:

{V V}_{T T}^{m m} = = {{(({l l}_{i i},, {x x}_{i i}^{m m}))}};; i i = = T T,, T T + + 11,, . . . . . .;;

in, is the optimization sample of the m-th user anomaly detection model; the value of l _i is 1 or -1, which respectively represent the normal or abnormal power consumption of the sample at the i-th moment;

Correspondingly, the precursor sample at time T is selected as the initial training sample of the anomaly detection model, which is expressed as:

{M m}_{T T}^{m m . . 00} = = {{(({l l}_{i i},, {x x}_{i i}^{m m}))}};; i i = = 1,2 1,2,, . . . . . .,, T T - - 11;;

in, is the initial training sample for the mth user anomaly detection model.

5. A kind of intelligent power consumption anomaly detection method aiming at non-technical loss according to claim 4, is characterized in that, in described step 4, the fitness function in genetic algorithm GA off-line parameter optimization is taken as abnormality detection accuracy ADC, the calculation formula is:

ADC ADC = = d d / / ((b b + + d d + + \frac{11}{22}));;

Among them, b represents the number of samples that are detected as normal by the extreme learning machine ELM algorithm but are actually abnormal; d represents the number of samples that are detected as abnormal by the extreme learning machine ELM algorithm and are indeed abnormal.

6. A method for detecting abnormalities in intelligent power consumption aimed at non-technical losses according to claim 5, wherein the classification accuracy CA in the step 7 is expressed as:

CA=(a ₁ +d ₁ )/(a ₁ +b ₁ +c ₁ +d ₁ );

The abnormal false detection rate FDC is expressed as:

FDC=(c ₁ +d ₁ +1)/(b ₁ +d ₁ +1);

Among them, in the online stage, a ₁ represents the number of samples that are detected as normal by the extreme learning machine ELM algorithm and are indeed normal; b ₁ represents the number of samples _that are detected as normal by the extreme learning machine ELM algorithm but are actually abnormal; The number of samples detected as abnormal by the ELM algorithm of the learning machine but actually normal; d ₁ represents the number of samples detected as abnormal by the ELM algorithm of the extreme learning machine and indeed abnormal.