CN103208091A

CN103208091A - Electric larceny preventing method based on data mining of electric load management system

Info

Publication number: CN103208091A
Application number: CN2013101481057A
Authority: CN
Inventors: 罗祾; 张浙波; 金家培; 童旭
Original assignee: State Grid Corp of China SGCC; Shanghai Municipal Electric Power Co; East China Power Test and Research Institute Co Ltd
Current assignee: State Grid Corp of China SGCC; Shanghai Municipal Electric Power Co; East China Power Test and Research Institute Co Ltd
Priority date: 2013-04-25
Filing date: 2013-04-25
Publication date: 2013-07-17
Anticipated expiration: 2033-04-25
Also published as: CN103208091B

Abstract

The invention relates to an anti-stealing electricity method based on data mining of an electricity load management system, the method comprising the following steps: 1) Obtaining user historical load data in a database of an electricity load management system, and performing newton interpolation on missing data Rounding up and supplementing; 2) Arranging the historical load data samples after rounding up and supplementing according to the time rule, establishing an autoregressive sliding average model through model identification and parameter estimation, obtaining a short-term load forecasting equation, and calculating short-term load forecasting values; 3) comparing The load prediction value and the actual collected value are used to determine whether the current user has the possibility of stealing electricity. If so, an alarm of "possible electricity stealing" will be issued. If not, the actual collected value will be saved to the database of the electricity load management system, and return to step 1. ). Compared with the prior art, the present invention has the advantages of high prediction accuracy and high screening accuracy of electric stealing suspects.

Description

An anti-stealing method based on data mining of electric load management system

技术领域technical field

本发明涉及一种反窃电方法，尤其是涉及一种基于用电负荷管理系统数据挖掘的反窃电方法。The invention relates to an anti-stealing method, in particular to an anti-stealing method based on data mining of an electric load management system.

背景技术Background technique

窃电行为一直是电力企业在营销方面遇到的重要问题，它在扰乱供用电秩序的同时，也对电力企业造成了巨大的经济损失。当前绝大部分的反窃电手段都集中在计量回路和电能表计里设置防治措施，或者是通过经验数据的粗糙判据进行比较，缺乏较严格可信的数学模型作为基础。而且，供电企业面对的用户众多，理论上判定的窃电行为必须在实践中进行深入调查才可确定，因此，如果能基于当前用电负荷管理系统的数据大幅度缩小窃电用户的嫌疑名单，将对减轻反窃电工作的压力带来重大利好。因此对反窃电方法的要求是：1.能基于用电负荷管理系统的数据进行电力用户的窃电嫌疑判断；2.能快速准确地筛选出窃电嫌疑名单并提出告警；3.能通过对有窃电可疑数据进行判断来完善用电负荷管理系统数据的可信度。Stealing electricity has always been an important problem in the marketing of electric power companies. While disrupting the order of power supply and consumption, it has also caused huge economic losses to electric power companies. Most of the current anti-stealing methods focus on setting prevention and control measures in metering circuits and electric energy meters, or make comparisons based on rough criteria of empirical data, lacking stricter and more credible mathematical models as the basis. Moreover, power supply companies face a large number of users, and theoretically determined electricity theft must be investigated in practice. Therefore, if the data of the current electricity load management system can be used to greatly reduce the list of suspected electricity theft , will bring great benefits to alleviating the pressure of anti-stealing work. Therefore, the requirements for the anti-stealing method are: 1. Be able to judge the power user’s suspected electricity theft based on the data of the power load management system; 2. Be able to quickly and accurately screen out the list of suspected electricity theft and raise an alarm; 3. Can pass Judging the suspicious data of electricity theft to improve the credibility of the data in the power load management system.

窃电的最终目的是通过各种方法使得电能表计量数字小于实际用电量，从而减少所应缴纳的电费，得到实际的经济利益，而不同的窃电方式对于用户负荷或电量的变化是有一定差别的。目前，为了减少电能表的计量数字，比较常见的窃电方法，主要有以下几种。The ultimate goal of stealing electricity is to make the measured number of the electric energy meter smaller than the actual electricity consumption through various methods, thereby reducing the electricity fee that should be paid, and obtaining actual economic benefits. There must be a difference. At present, in order to reduce the measurement figures of the electric energy meter, the more common methods of stealing electricity mainly include the following.

1)使计量回路断路、短路、短接。1) Make the metering circuit disconnect, short circuit, or short circuit.

比如切断3相电压中的其中一相，从而将减少1/3的实际用电量。For example, cutting off one of the phases in the 3-phase voltage will reduce the actual power consumption by 1/3.

2)改变二次接线。2) Change the secondary wiring.

比如改变二次电压电流接入电压互感器(PT)、电流互感器(CT)中的位置，将导致计算所得电量反向，从而减小了总的用电量。For example, changing the position where the secondary voltage and current are connected to the potential transformer (PT) and the current transformer (CT) will cause the calculated power to be reversed, thereby reducing the total power consumption.

3)人为使表慢步、停走。3) The watch is artificially slowed down or stopped.

比如人为在电能表走圈处插入小针，使得转速变慢甚至停止，从而减小电量计数。For example, artificially insert a small needle at the circle of the electric energy meter, so that the speed will slow down or even stop, thereby reducing the power count.

4)给CT倒灌电流使表倒转以降低表指示数等。4) Feed the current back to the CT to make the meter reverse to reduce the indicator number of the meter, etc.

所有的窃电行为表现出来的或是使电能表感应不了电功率，或是使电能表感应到的电功率与该负荷下正常计量时的电功率相比，产生一个下降的跃变或者使表指示数异常下降。在以上比较普遍的窃电方法中，有些将会导致固定比例的负荷降低。其中最为常见的第一种方法，往往导致负荷降低为0或者实际负荷的2/3。All the acts of stealing electricity show that the electric energy meter cannot sense the electric power, or the electric power sensed by the electric energy meter is compared with the electric power of the normal measurement under the load, resulting in a drop jump or making the meter indication abnormal. decline. Some of the more common stealing methods above will result in a fixed percentage load reduction. The first method, which is the most common, often causes the load to be reduced to 0 or 2/3 of the actual load.

另外还存在一些计量异常的原因，比如电压互感器回路熔断器熔断、端钮螺丝松动、断线，或者表计内部故障等，虽不是用户主观上的窃电行为，但实际产生的结果也是负荷的降低，依然会对电力企业造成经济上的损失。In addition, there are some reasons for measurement abnormalities, such as blown voltage transformer circuit fuses, loose terminal button screws, broken wires, or internal faults in meters, etc. Although it is not the subjective behavior of the user to steal electricity, the actual result is also a load. The reduction will still cause economic losses to power companies.

目前，电力企业查处窃电行为的方法比较有限，只有当用户用电负荷有十分明显的变化时，电力企业才能根据相应经验和实际勘察来对窃电行为进行判断和查处，但仍有许多的窃电行为逃脱了处罚，从而给电力公司造成了经济损失。目前较常见的且有实际效果的方法只有以下几种。At present, the methods for power companies to investigate and deal with electricity theft are relatively limited. Only when the user’s electricity load has a very obvious change, the power company can judge and investigate the power theft based on corresponding experience and actual investigation, but there are still many Electricity theft escaped punishment, resulting in economic losses to the power company. At present, there are only the following methods that are more common and have practical effects.

1)采用计量回路和模拟回路相互配合的方法。1) Adopt the method that the metering circuit and the analog circuit cooperate with each other.

对于总受、总计的用户，在总计量柜以外还有一个总受柜，从总受柜中分立出一路模拟测量回路，与计量回路相互配合，相互监视，通过比较来得到电量差错，从而获取窃电信息。但是这种方法，需要额外增加计量回路，不适用于只有总计的用户，比如上海电网目前的情况。而且当发生改变电力线路结构的窃电行为时，这种方法将失效。For the total receiving and totaling users, there is a total receiving cabinet besides the total metering cabinet, and an analog measurement circuit is separated from the total receiving cabinet, which cooperates with the metering circuit, monitors each other, and obtains the power error by comparison, so as to obtain Stealing information. However, this method requires an additional metering circuit, and is not suitable for users who only have totals, such as the current situation of Shanghai Power Grid. And when there is a stealing behavior that changes the structure of the power line, this method will fail.

2)采用智能远程抄表系统的方法。2) Adopt the method of intelligent remote meter reading system.

这种方法主要基于新型的智能电表和远程抄表系统，需要电表能就地检测电路是否断相、PT是否短接、CT是否分流等等，基于这些电表上的高级应用来较准确地检测到窃电行为。但是这种方法所耗费的金额较高，若进行电表更换，不仅需要大量的资金投入，还需要将负荷管理系统的后台进行改造；而且在今后随着用户对智能电表的认识，又可能会有新的窃电方式的产生，从而导致此方法的失效。This method is mainly based on a new type of smart meter and remote meter reading system, which requires the meter to be able to detect whether the circuit is out of phase, whether the PT is shorted, whether the CT is shunted, etc., based on the advanced application of these meters to more accurately detect Electricity theft. However, this method consumes a relatively high amount of money. If the meter is replaced, not only a large amount of capital investment is required, but also the background of the load management system needs to be transformed; and in the future as users understand smart meters, there may be The emergence of new ways of stealing electricity leads to the failure of this method.

3)基于电量分析的方法。3) The method based on electricity analysis.

这种方法主要是将用户现有电量与前一时间短或历史同期电量进行相应比较，得到异常电量数值，来检测窃电行为。但是这种方法目前选取的判据较为粗糙，需要人为判断，且会受一些特殊因素的影响，而导致判断失误。该方法使用的是人工经验猜测，一来判据过于粗糙，二来对负荷的随机性没有任何考虑。This method is mainly to compare the user's current power with the power of the previous short period or the same period of history to obtain abnormal power values to detect power theft. However, the currently selected criteria for this method are relatively rough, requiring human judgment, and will be affected by some special factors, resulting in misjudgment. This method uses artificial experience guessing. First, the criterion is too rough, and second, there is no consideration of the randomness of the load.

γ表示自协方差函数， $γ_{k} = Σ_{t = 1}^{N - k} Y (t) Y (t + k) / N;$ γ represents the autocovariance function, $γ_{k} = Σ_{t = 1}^{N - k} Y (t) Y (t + k) / N;$

106)令

的协方差函数为

通过矩法估计计算出滑动平均系数的估计值：106) order

The covariance function of is

Estimates of the moving average coefficients are calculated by method of moments estimation:

$\begin{matrix} {\overset{&OverBar; &OverBar;}{γ γ}}_{00} = = {σ σ}_{α α}^{22} ((11 + + {θ θ}_{11} + + \cdot &Center Dot; \cdot &Center Dot; \cdot \cdot + + {θ θ}_{q q})) \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{11} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{11} + + {θ θ}_{22} {θ θ}_{11} + + \cdot &Center Dot; \cdot \cdot \cdot &Center Dot; + + {θ θ}_{q q} {θ θ}_{q q - - 11})) \\ \cdot \cdot \cdot \cdot \cdot \cdot \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{q q} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{q q})) \end{matrix}\}$

表示自噪声序列的方差。

Indicates the variance of the self-noise sequence.

5.根据权利要求4所述的一种基于用电负荷管理系统数据挖掘的反窃电方法，其特征在于，所述的短期负荷预测方程为：5. a kind of anti-stealing method based on electricity load management system data mining according to claim 4, is characterized in that, described short-term load forecasting equation is:

所述的计算短期负荷预测值的公式为：

The formula for calculating the short-term load forecast value is:

6.根据权利要求5所述的一种基于用电负荷管理系统数据挖掘的反窃电方法，其特征在于，所述的步骤3)中，判断当前用户是否存在窃电可能行为的具体步骤为：6. A kind of anti-stealing method based on electricity load management system data mining according to claim 5, is characterized in that, in described step 3), the specific step of judging whether the current user has the possible behavior of stealing electricity is :

301)设置报警因子r和报警窗口长度W；301) Setting the alarm factor r and the alarm window length W;

302)根据负荷预测值y计算报警阈值y_th：y_th＝ry；302) Calculating the alarm threshold y _th according to the load forecast value y: y _th = r y ;

303)比较同一时间段的实际负荷曲线与报警阈值曲线，若从n时刻起的实际负荷值y(t)满足以下条件：303) Comparing the actual load curve and the alarm threshold curve in the same time period, if the actual load value y(t) from time n satisfies the following conditions:

$y the y ((t t)) < < {y the y}_{th the th},, &ForAll; &ForAll; t t &Element; &Element; [[n no,, n no + + W W]]$

则判断当前用户存在窃电可能行为，发出“窃电可能”的告警。Then it is judged that the current user has a possible behavior of stealing electricity, and an alarm of "possible electricity stealing" is issued.

7.根据权利要求6所述的一种基于用电负荷管理系统数据挖掘的反窃电方法，其特征在于，所述的报警因子r为：1/2≤r≤2/3。7. An anti-stealing electricity method based on data mining of an electricity load management system according to claim 6, wherein the alarm factor r is: 1/2≤r≤2/3.

为自回归系数和滑动平均系数，i＝1，2，...，p，j＝1，2，...，q。are autoregressive coefficients and moving average coefficients, i=1, 2, ..., p, j = 1, 2, ..., q.

所述的模型辨识和参数估计具体为：The model identification and parameter estimation are specifically as follows:

101)对按照时间排列的历史负荷数据样本进行零变换：101) Perform a zero transformation on the time-arranged historical load data samples:

$Y Y ((i i)) = = y the y ((i i)) - - \overset{&OverBar; &OverBar;}{y the y},, i i &Element; &Element; [[11,, N N]]$

y(i)为第i个样本值，Y(i)为相应的变换后的值，

为所有样本的平均值，N为样本容量且N≥50；y(i) is the i-th sample value, Y(i) is the corresponding transformed value,

is the average value of all samples, N is the sample size and N≥50;

102)计算自相关函数ρ_k的估计值：102) Calculate the estimated value of the autocorrelation function ρ _k :

${ρ ρ}_{k k} = = \frac{\frac{11}{N N - - k k} {Σ Σ}_{i i = = 11}^{N N - - k k} Y Y ((i i)) Y Y ((i i + + k k))}{\frac{11}{N N} {Σ Σ}_{i i = = 11}^{N N} Y Y {((i i))}^{22}}$

k表示自相关函数的项数，若满足条件

的频率小于95％，则自相关函数ρ_k具有截尾性，若随着k的增大，ρ_k趋向于0，则自相关函数ρ_k具有拖尾性；k represents the number of items of the autocorrelation function, if the condition is satisfied

If the frequency of is less than 95%, the autocorrelation function ρ _k has truncation, if as k increases, ρ _k tends to 0, then the autocorrelation function ρ _k has tailing property;

103)通过下述递推公式计算出偏相关函数a_kk的估值：103) Calculate the estimated value of the partial correlation function a _kk by the following recursive formula:

$\begin{matrix} {a a}_{1111} = = {ρ ρ}_{11} \\ {a a}_{k k + + 11,, k k + + 11} = = (({ρ ρ}_{k k + + 11} - - {Σ Σ}_{j j = = 11}^{k k} {a a}_{kj kj} {ρ ρ}_{k k + + 11 - - j j})) {((11 - - {Σ Σ}_{j j = = 11}^{k k} {a a}_{kj kj} {ρ ρ}_{j j}))}^{- - 11} \\ {a a}_{k k + + 11,, j j} = = {a a}_{kj kj} + + {a a}_{k k + + 11,, k k + + 11} {a a}_{k k,, k k - - j j + + 11} \end{matrix}\}$

j＝1，2，...，k，若满足条件

的频率小于95％，则偏相关函数a_kk具有截尾性，若随着k的增大，a_kk趋向于0，则偏相关函数a_kk具有拖尾性；j=1, 2, ..., k, if the conditions are met

The frequency of is less than 95%, then the partial correlation function a _kk has truncation, if as k increases, a _kk tends to 0, then the partial correlation function a _kk has tailing;

104)根据自相关函数和偏相关函数的性质计算自回归阶数p和滑动平均阶数q：104) Calculate the autoregressive order p and the moving average order q according to the properties of the autocorrelation function and the partial correlation function:

若ρ_k具有拖尾性且a_kk具有截尾性，则最大的满足

条件的k值就是模型的p的值，q＝0；If ρ _k has tailing property and a _kk has truncated property, then the maximum satisfying

The k value of the condition is the p value of the model, q=0;

若ρ_k具有截尾性且a_kk具有拖尾性，则最大的满足

条件的k值就是模型的q的值，p＝0；If ρ _k is truncated and a _kk is tailed, then the maximum satisfying

The k value of the condition is the value of q of the model, p=0;

若ρ_k、a_kk均具有截尾性，则最大的满足

条件的k值就是模型的p的值，最大的满足

条件的k值就是模型的q的值；If both ρ _k and a _kk are truncated, then the maximum satisfying

The k value of the condition is the p value of the model, and the maximum satisfying

The k value of the condition is the value of q of the model;

105)通过Yule-Walker方程计算自回归系数的估计值：105) Estimates of the autoregressive coefficients are calculated by the Yule-Walker equation:

106)令

的协方差函数为

通过矩法估计计算出滑动平均系数的估计值：106) order

The covariance function of is

$\begin{matrix} {\overset{&OverBar; &OverBar;}{γ γ}}_{00} = = {σ σ}_{α α}^{22} ((11 + + {θ θ}_{11} + + \cdot \cdot \cdot \cdot \cdot &Center Dot; + + {θ θ}_{q q})) \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{11} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{11} + + {θ θ}_{22} {θ θ}_{11} + + \cdot \cdot \cdot \cdot \cdot &Center Dot; + + {θ θ}_{q q} {θ θ}_{q q - - 11})) \\ \cdot \cdot \cdot \cdot \cdot &Center Dot; \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{q q} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{q q})) \end{matrix}\}$

表示白噪声序列的方差。 Represents the variance of the white noise sequence.

所述的短期负荷预测方程为：The short-term load forecasting equation is:

所述的计算短期负荷预测值的公式为：

The formula for calculating the short-term load forecast value is:

所述的步骤3)中，判断当前用户是否存在窃电可能行为的具体步骤为：In the described step 3), the specific steps for judging whether the current user has a possible behavior of stealing electricity are:

301)设置报警因子r和报警窗口长度W；301) Set the alarm factor r and the alarm window length W;

所述的报警因子r为：1/2≤r≤2/3。The alarm factor r is: 1/2≤r≤2/3.

与现有技术相比，本发明具有以下优点：Compared with the prior art, the present invention has the following advantages:

1、本发明利用历史数据进行窃电监测，包括历史数据归整和短期负荷预测，提高窃电嫌疑筛选的准确度，实现对电力用户窃电嫌疑的精确判断；1. The present invention uses historical data to monitor electricity theft, including historical data consolidation and short-term load forecasting, to improve the accuracy of screening suspicions of electricity theft, and to realize accurate judgments on suspicions of electricity theft by power users;

2、本发明通过短期负荷预测的方法对用户数据进行科学建模，有效地利用了历史数据库，较大地提高了预测的精确度，并且为反窃电智能化奠定了基础；2. The present invention scientifically models the user data through the method of short-term load forecasting, effectively utilizes the historical database, greatly improves the accuracy of forecasting, and lays the foundation for anti-stealing intelligence;

3、本发明时间序列预测方法进行短期负荷预测，计算相对简单且预测结果较准确。3. The time series forecasting method of the present invention performs short-term load forecasting, the calculation is relatively simple and the forecasting result is more accurate.

附图说明Description of drawings

图1为本发明的流程示意图；Fig. 1 is a schematic flow sheet of the present invention;

图2为实施例中某用户第51天的用电负荷预测示意图；Fig. 2 is a schematic diagram of electricity load forecasting of a certain user on the 51st day in the embodiment;

图3为实施例中某用户第52天的用电负荷预测示意图。Fig. 3 is a schematic diagram of a customer's electricity load forecast on the 52nd day in the embodiment.

具体实施方式Detailed ways

下面结合附图和具体实施例对本发明进行详细说明。本实施例以本发明技术方案为前提进行实施，给出了详细的实施方式和具体的操作过程，但本发明的保护范围不限于下述的实施例。The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

实施例Example

如图1所示，一种基于用电负荷管理系统数据挖掘的反窃电方法，该方法包括以下步骤：As shown in Figure 1, a kind of anti-stealing electricity method based on data mining of electric load management system, this method comprises the following steps:

1)获取用电负荷管理系统数据库中的用户历史负荷数据，并对缺失数据通过牛顿插值法进行归整补充；1) Obtain the user's historical load data in the database of the electric load management system, and supplement the missing data by Newton interpolation method;

2)按照时间规律对归整补充后的历史负荷数据样本进行排列，通过模型辨识和参数估计建立自回归滑动平均模型，获得短期负荷预测方程，计算短期负荷预测值；2) Arrange the adjusted and supplemented historical load data samples according to the time rule, establish an autoregressive sliding average model through model identification and parameter estimation, obtain the short-term load forecast equation, and calculate the short-term load forecast value;

3)比较负荷预测值与实际采集值，判断当前用户是否存在窃电可能行为，若是，则发出“窃电可能”的告警，若否，则将实际采集值保存至用电负荷管理系统数据库，返回步骤1)。3) Compare the load prediction value with the actual collected value to determine whether the current user has the possibility of stealing electricity. If so, issue an alarm of "possible electricity stealing". If not, save the actual collected value to the database of the electricity load management system. Go back to step 1).

用电负荷管理系统的特征是每隔一个固定时间t对电力用户的负荷进行一次采集，一般来说，很多用电负荷管理系统的t是15分钟，也即一小时采集4个点，一天采集96个点，然后将这些采集到的数据归整到历史数据库。在用电负荷管理系统数据库中的所有用户数据，由于可能存在的通讯中断等原因，将有一些坏数据、空数据等不完整的情况，这些数据缺陷在一定程度上会影响到数据挖掘的准确性，因此需要采用合适的数值计算方法对其进行完善和补充。The characteristic of the power load management system is to collect the load of power users every fixed time t. Generally speaking, the t of many power load management systems is 15 minutes, that is, to collect 4 points in an hour, and to collect in a day 96 points, and then organize the collected data into the historical database. All user data in the database of the power load management system, due to possible communication interruptions and other reasons, will have some incomplete data such as bad data and empty data. These data defects will affect the accuracy of data mining to a certain extent. Therefore, it is necessary to use appropriate numerical calculation methods to improve and supplement it.

步骤1)中，通过牛顿插值法进行归整补充具体为：In step 1), the rounding supplement by Newton interpolation method is specifically:

x₀，x₁，...，x_n为已知的n个采样时间点，f(x₀)，f(x₁)，...，f(x_n)为相应的采样值，在时间点x上缺失了采样值f(x)，则x ₀ , x ₁ , ..., x _n are the known n sampling time points, f(x ₀ ), f(x ₁ ), ..., f(x _n ) are the corresponding sampling values, in The sampled value f(x) is missing at time point x, then

f(x)＝f(x₀)+f[x₀，x₁](x-x₀)+…+f[x₀，x₁，…x_n](x-x₀)(x-x₁)…(x-x_n)+P_n(x)f(x)=f(x ₀ )+f[x ₀ , x ₁ ](xx ₀ )+…+f[x ₀ , x ₁ ,…x _n ](xx ₀ )(xx ₁ )…(xx _n )+P _n (x)

其中，f[x₀，x₁，…x_n]为函数f的n阶差商，Among them, f[x ₀ , x ₁ ,…x _n ] is the nth order difference quotient of function f,

f[x₀，x₁，…x_n]＝(f[x₀，x₁，…x_n-1]-f[x₁，…x_n])/(x₀-x_n)f[x ₀ , x ₁ ,...x _n ]=(f[x ₀ , x ₁ ,...x _n-1 ]-f[x ₁ ,...x _n ])/(x ₀ -x _n )

${P P}_{n no} ((x x)) = = f f [[x x,, {x x}_{00},, {x x}_{11},, \cdot &Center Dot; \cdot &Center Dot; \cdot \cdot {x x}_{n no}]] {Π Π}_{i i = = 00}^{n no} ((x x - - {x x}_{i i})) . .$

在归整后的用电负荷管理系统的数据上，本方法通过选定的负荷预测方法和模型，对用户进行短期负荷预测，将预测值与观测值进行比较，通过一定的比较手段，设立合理的判定阈值与报警窗口大小，对用户的窃电嫌疑进行合理判断，产生的嫌疑名单能大幅度降低人工筛选窃电用户的工作量，并且能通过科学的数据分析提高筛选的准确度。Based on the data of the consolidated electricity load management system, this method uses the selected load forecasting method and model to perform short-term load forecasting for users, compares the predicted value with the observed value, and establishes a reasonable The judgment threshold and the size of the alarm window can be used to reasonably judge the user's suspicion of stealing electricity. The generated suspect list can greatly reduce the workload of manual screening of electricity stealing users, and can improve the accuracy of screening through scientific data analysis.

传统的负荷预测方法主要有动力学方法和数理统计学方法，这些方法共同特点是先建立数据序列的主观模型，然后进行计算和预测。方法主要有以下几种：最小二乘拟合法、回归分析法、专家系统法、灰色预测法、时间序列方法。最小二乘拟合法和回归分析法并不适合短期负荷预测；人工神经网络法过于复杂，在统一的系统内不易产生标准解；专家系统法虽然在短期预测上较为精确，但需要耗费太多的人力物力，且无法实现自动化，不适合应用在面向大量用电客户的反窃电系统上；灰色预测法适合大量随机量的和值，不适合单一用户的随机量预测。因此，本发明选择了计算相对不复杂、短期预测相对准确的时间序列预测方法，通过用电负荷管理系统挖掘后的历史数据，对单个用户的负荷进行建模和预测，再进行反窃电嫌疑判断。Traditional load forecasting methods mainly include dynamical methods and mathematical statistics methods. The common feature of these methods is to first establish a subjective model of the data sequence, and then perform calculation and prediction. There are mainly the following methods: least squares fitting method, regression analysis method, expert system method, gray prediction method, time series method. The least squares fitting method and regression analysis method are not suitable for short-term load forecasting; the artificial neural network method is too complicated, and it is difficult to produce a standard solution in a unified system; although the expert system method is more accurate in short-term forecasting, it takes too much time. Manpower and material resources, and cannot be automated, it is not suitable for anti-stealing systems for a large number of electricity customers; the gray prediction method is suitable for the sum of a large number of random quantities, but not suitable for the random quantity prediction of a single user. Therefore, the present invention chooses a time series forecasting method with relatively uncomplicated calculation and relatively accurate short-term forecasting. Through the historical data mined by the power load management system, the load of a single user is modeled and forecasted, and then the anti-stealing suspicion is carried out. judge.

由于用电客户有一定的生产周期或是工作生活习惯，因此在用户负荷数据处理的工作按照时间规律对历史数据进行了排列，对一些特定的生产厂商划分为“工作日负荷”和“非工作日负荷”。Since electricity customers have a certain production cycle or work and living habits, the historical data are arranged according to the time rule in the work of user load data processing, and some specific manufacturers are divided into "working day load" and "non-working load". daily load".

在时间序列预测方法中，一般情况下数据符合的模型是AR(自回归)模型、MA(滑动平均)模型，或者ARMA(自回归滑动平均)模型，其中AR和MA模型都是ARMA模型的特例。典型的自回归滑动平均模型为：In the time series forecasting method, in general, the model that the data conforms to is the AR (autoregressive) model, the MA (moving average) model, or the ARMA (autoregressive moving average) model, where the AR and MA models are special cases of the ARMA model . A typical autoregressive moving average model is:

其中，y(t-p)，y(t-p+1)，...，y(t-1)，y(t)是连续一段时间的负荷数据值；α_t，α_t-1，...，α₀是白噪声序列(具体概率分布要视数据而定，一般假设为高斯白噪声，

p、q分别为自回归阶数和滑动平均阶数；和θ_j分别为自回归系数和滑动平均系数，i＝1，2，...，p，j＝1，2，...，q。Among them, y(tp), y(t-p+1), ..., y(t-1), y(t) are load data values for a continuous period of time; α _t , α _t-1 , .. ., α ₀ is a white noise sequence (the specific probability distribution depends on the data, it is generally assumed to be Gaussian white noise,

p and q are the autoregressive order and the moving average order respectively; and θ _j are autoregressive coefficients and moving average coefficients respectively, i=1, 2,..., p, j=1, 2,..., q.

在将数据平稳化后，即可对其进行ARMA模型的辨识和参数估计，具体为：After the data is stabilized, the identification and parameter estimation of the ARMA model can be carried out, specifically:

y(i)为第i个样本值，Y(i)为相应的变换后的值，

is the average value of all samples, N is the sample size and N≥50;

k表示自相关函数的项数，若满足条件

j＝1，2，...，k，若满足条件

若ρ_k具有拖尾性且a_kk具有截尾性，则最大的满足条件的k值就是模型的p的值，q＝0；If ρ _k has tailing property and a _kk has truncated property, then the maximum satisfying The k value of the condition is the p value of the model, q=0;

若ρ_k具有截尾性且a_kk具有拖尾性，则最大的满足

The k value of the condition is the value of q of the model, p=0;

若ρ_k、a_kk均具有截尾性，则最大的满足条件的k值就是模型的p的值，最大的满足

条件的k值就是模型的q的值；If both ρ _k and a _kk are truncated, then the maximum satisfying The k value of the condition is the p value of the model, and the maximum satisfying

The k value of the condition is the value of q of the model;

γ表示自协方差函数，为 $γ_{k} = Σ_{t = 1}^{N - k} Y (t) Y (t + k) / N;$ γ represents the autocovariance function, which is $γ_{k} = Σ_{t = 1}^{N - k} Y (t) Y (t + k) / N;$

106)令的协方差函数为

通过矩法估计计算出滑动平均系数的估计值：106) order The covariance function of is

$\begin{matrix} {\overset{&OverBar; &OverBar;}{γ γ}}_{00} = = {σ σ}_{α α}^{22} ((11 + + {θ θ}_{11} + + \cdot \cdot \cdot &Center Dot; \cdot \cdot + + {θ θ}_{q q})) \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{11} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{11} + + {θ θ}_{22} {θ θ}_{11} + + \cdot \cdot \cdot \cdot \cdot \cdot + + {θ θ}_{q q} {θ θ}_{q q - - 11})) \\ \cdot &Center Dot; \cdot \cdot \cdot \cdot \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{q q} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{q q})) \end{matrix}\}$

表示白噪声序列的方差。

Represents the variance of the white noise sequence.

至此完成ARMA模型的建立，得出短期负荷预测的方程，再通过用电负荷管理系统的历史数据库进行电力用户的短期负荷预测。So far, the establishment of the ARMA model is completed, and the equation of short-term load forecasting is obtained, and then the short-term load forecasting of power users is carried out through the historical database of the power load management system.

所述的计算短期负荷预测值的公式为：

The formula for calculating the short-term load forecast value is:

如果用户负荷数据无法满足自相关函数和偏相关函数的“拖尾”与“截尾”性，则可以对数据进行周期性差分，再判断差分后数据的自相关函数和偏相关函数的“拖尾”与“截尾”性。对差分后的数据再观察自相关函数和偏相关函数的“拖尾”与“截尾”性，然后再通过时间序列方法建模，并完成此电力用户的短期负荷预测。If the user load data cannot meet the "tailing" and "truncating" properties of the autocorrelation function and partial correlation function, the data can be periodically differentiated, and then the "dragging" of the autocorrelation function and partial correlation function of the data after the difference can be judged. tail" and "truncated". Observe the "tailing" and "truncation" of the autocorrelation function and partial correlation function on the differenced data, and then use the time series method to model and complete the short-term load forecast of the power user.

假设当前采集的数据序列为y(i)，i∈[1，T+t]，要预测的值为y(T+t+1)则以T为周期项的差分方法如下：Assuming that the currently collected data sequence is y(i), i∈[1, T+t], and the value to be predicted is y(T+t+1), then the difference method with T as the periodic term is as follows:

z(i)＝y(T+i)-y(i)，i∈[1，t]z(i)=y(T+i)-y(i), i∈[1,t]

z(i)，i＝1，2，...，t即为差分后的数据，通过时间序列对z(i)进行建模，预测出z(t+1)，然后再通过以下计算返回y(T+t+1)的预测值：z(i), i=1, 2,..., t is the data after difference, model z(i) through time series, predict z(t+1), and then return through the following calculation Predicted value of y(T+t+1):

y(T+t+1)＝y(t+1)+z(t+1)。y(T+t+1)=y(t+1)+z(t+1).

对于单一用户来说，窃电行为有两个明显的指标，其一是窃电量，其二是窃电时长。所述的判断当前用户是否存在窃电可能行为的具体步骤为：For a single user, there are two obvious indicators of electricity theft, one is electricity theft, and the other is the duration of electricity theft. The specific steps for judging whether the current user may steal electricity are as follows:

则判断当前用户存在窃电可能行为，发出“窃电可能”的告警。一旦系统出现“窃电可能”的告警，则当天数据将被认为是“不可信”，即非真实的负荷数据，不会被录入到预测用的数据库；反之，“可信”的数据则在每天结束后录入预测用的数据库，用来预测下一天的用电负荷。Then it is judged that the current user has a possible behavior of stealing electricity, and an alarm of "possible electricity stealing" is issued. Once the system has an alarm of "possible electricity theft", the data of the day will be considered "unreliable", that is, unreal load data will not be entered into the database for forecasting; on the contrary, "credible" data will be in After the end of each day, it is entered into the database for forecasting, which is used to predict the electricity load of the next day.

根据电力用户的实际情况对报警因子r进行一个合理的设置，如1/2≤r≤2/3。报警窗口的设置是因为单一用户的用电负荷具有较大的随机性，如果每次实际值低于报警阈值就进行“窃电可能”的报警，将可能导致系统频繁报警，降低窃电预测的准确度。一般来说，窃电用户因为窃电工作的繁琐性，不会只窃电很短的时间，所以反窃电工作者也可以根据当地的实际情况及历史数据对报警窗口W进行一个合理的设置，W过小，则系统的反窃电错误预报率将会降低，W过大，则系统的反窃电漏预报率将会增大。Set a reasonable alarm factor r according to the actual situation of power users, such as 1/2≤r≤2/3. The setting of the alarm window is because the electricity load of a single user has a large randomness. If the "possible electricity theft" alarm is issued every time the actual value is lower than the alarm threshold, the system may cause frequent alarms and reduce the probability of electricity theft prediction. Accuracy. Generally speaking, due to the cumbersome work of stealing electricity, users will not only steal electricity for a short period of time, so anti-stealing workers can also set a reasonable alarm window W according to the local actual situation and historical data , W is too small, the system's anti-stealing error prediction rate will decrease, and W is too large, the system's anti-stealing leakage prediction rate will increase.

基于用电负荷管理系统里的某用户连续50天的实际用电负荷，得到起第51天的用电负荷预测如图2，并与第51天的实际负荷曲线进行比较，取r＝2/3，W＝8(该用电负荷管理系统每15分钟进行一次采样，一天共采集96个点，采样窗口长度为8即代表2小时)，预测负荷数据与实际负荷数据较为相近。第51天的实际负荷数据没有达到引起“窃电可能”告警的要求，据此系统认为第51天是可信的，并将第51天的数据归入到历史数据集中。Based on the actual electricity load of a user in the electricity load management system for 50 consecutive days, the electricity load forecast on the 51st day is shown in Figure 2, and compared with the actual load curve on the 51st day, r=2/ 3. W=8 (the power load management system samples every 15 minutes, a total of 96 points are collected in one day, and the sampling window length is 8, which means 2 hours), the predicted load data is relatively close to the actual load data. The actual load data on the 51st day did not meet the requirement to cause the alarm of "possible electricity theft". Based on this, the system considered the 51st day to be credible and included the data on the 51st day into the historical data set.

根据更新后的历史数据集，再进行短期负荷预测，得到第52天的预测负荷预测曲线，如图3所示。通过之前的设置便可以发现，实际负荷曲线中的前24个采样点都明显低于红色的告警阈值曲线，达到了“窃电可能”告警的要求，而之后的52个采样数据则与预测值基本接近。通过比较分析得知，该用户在第52天的早上0点至6点的6小时内，存在较大的窃电嫌疑，因此本日的数据被列为“不可信”，不归入到历史数据集中。根据数值计算的结果，将此用户在这段时间内的用电行为锁定为“窃电嫌疑”。According to the updated historical data set, the short-term load forecast is carried out again, and the forecast load forecast curve of the 52nd day is obtained, as shown in Figure 3. Through the previous settings, it can be found that the first 24 sampling points in the actual load curve are significantly lower than the red alarm threshold curve, which meets the alarm requirement of "possible electricity theft", while the subsequent 52 sampling data are consistent with the predicted value Basically close. Through comparative analysis, it is known that the user was suspected of power theft within 6 hours from 0:00 am to 6:00 am on the 52nd day, so today's data is listed as "unreliable" and will not be included in historical data concentrated. According to the result of numerical calculation, the user's electricity consumption behavior during this period is locked as "suspected electricity stealing".

Claims

1. an anti-stealing electricity method based on electricity load management system data mining, is characterized in that, the method comprises the following steps:

1) Obtain the user's historical load data in the database of the electric load management system, and supplement the missing data by Newton interpolation method;

2) Arrange the adjusted and supplemented historical load data samples according to the time rule, establish an autoregressive sliding average model through model identification and parameter estimation, obtain the short-term load forecast equation, and calculate the short-term load forecast value;

3) Compare the load prediction value with the actual collected value to determine whether the current user has the possibility of stealing electricity. If so, issue an alarm of "possible electricity stealing". If not, save the actual collected value to the database of the electricity load management system. Go back to step 1).

2. A kind of anti-stealing electricity method based on the data mining of electric load management system according to claim 1, is characterized in that, described by Newton's interpolation method to round up and supplement specifically:

x ₀ , x ₁ , ..., x _n are the known n sampling time points, f(x ₀ ), f(x ₁ ), ..., f(x _n ) are the corresponding sampling values, in The sampled value f(x) is missing at time point x, then

f(x)=f(x ₀ )+f[x ₀ , x ₁ ](xx ₀ )+…+f[x ₀ , x ₁ ,…x _n ](xx ₀ )(xx ₁ )…(xx _n )+P _n (x)

Among them, f[x ₀ , x ₁ ,…x _n ] is the nth order difference quotient of function f,

f[x ₀ , x ₁ ,...x _n ]=(f[x ₀ , x ₁ ,...x _n-1 ]-f[x ₁ ,...x _n ])/(x ₀ -x _n )

{P P}_{n no} ((x x)) = = f f [[x x,, {x x}_{00},, {x x}_{11},, \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot; {x x}_{n no}]] {Π Π}_{i i = = 00}^{n no} ((x x - - {x x}_{i i})) . .

3. a kind of anti-stealing electricity method based on electricity load management system data mining according to claim 1, is characterized in that, described autoregressive sliding average model is specifically:

Among them, y(tp), y(t-p+1), ..., y(t-1), y(t) are load data values for a continuous period of time; α _t , α _t-1 , .. ., α ₀ is a white noise sequence; p, q are autoregressive order and moving average order respectively;

and θ _j are autoregressive coefficients and moving average coefficients respectively, i=1, 2,..., p, j=1, 2,..., q.

4. A kind of anti-stealing electricity method based on the data mining of electric load management system according to claim 3, is characterized in that, described model identification and parameter estimation are specifically:

101) Perform a zero transformation on the time-arranged historical load data samples:

Y Y ((i i)) = = y the y ((i i)) - - \overset{&OverBar; &OverBar;}{y the y},, i i &Element; &Element; [[11,, N N]]

y(i) is the i-th sample value, Y(i) is the corresponding transformed value, is the average value of all samples, N is the sample size and N≥50;

102) Calculate the estimated value of the autocorrelation function ρ _k :

{ρ ρ}_{k k} = = \frac{\frac{11}{N N - - k k} {Σ Σ}_{i i = = 11}^{N N - - k k} Y Y ((i i)) Y Y ((i i + + k k))}{\frac{11}{N N} {Σ Σ}_{i i = = 11}^{N N} Y Y {((i i))}^{22}}

k represents the number of items of the autocorrelation function, if the condition is satisfied

103) Calculate the estimated value of the partial correlation function a _kk by the following recursive formula:

\begin{matrix} {a a}_{1111} = = {ρ ρ}_{11} \\ {a a}_{k k + + 11,, k k + + 11} = = (({ρ ρ}_{k k + + 11} - - {Σ Σ}_{j j = = 11}^{k k} {a a}_{kj kj} {ρ ρ}_{k k + + 11 - - j j})) {((11 - - {Σ Σ}_{j j = = 11}^{k k} {a a}_{kj kj} {ρ ρ}_{j j}))}^{- - 11} \\ {a a}_{k k + + 11,, j j} = = {a a}_{kj kj} + + {a a}_{k k + + 11,, k k + + 11} {a a}_{k k,, k k - - j j + + 11} \end{matrix}\}

j=1, 2, ..., k, if the conditions are met The frequency of is less than 95%, then the partial correlation function a _kk has truncation, if as k increases, a _kk tends to 0, then the partial correlation function a _kk has tailing;

104) Calculate the autoregressive order p and the moving average order q according to the properties of the autocorrelation function and the partial correlation function:

If ρ _k has tailing property and a _kk has truncated property, then the maximum satisfying

The k value of the condition is the p value of the model, q=0;

If ρ _k is truncated and a _kk is tailed, then the maximum satisfying

The k value of the condition is the value of q of the model, p=0;

If both ρ _k and a _kk are truncated, then the maximum satisfying

The k value of the condition is the p value of the model, and the maximum satisfying The k value of the condition is the value of q of the model;

105) Estimates of the autoregressive coefficients are calculated by the Yule-Walker equation:

γ represents the autocovariance function,

γ_{k} = Σ_{t = 1}^{N - k} Y (t) Y (t + k) / N .

106) order

The covariance function of is

\begin{matrix} {\overset{&OverBar; &OverBar;}{γ γ}}_{00} = = {σ σ}_{α α}^{22} ((11 + + {θ θ}_{11} + + \cdot &Center Dot; \cdot \cdot \cdot \cdot + + {θ θ}_{q q})) \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{11} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{11} + + {θ θ}_{22} {θ θ}_{11} + + \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot; + + {θ θ}_{q q} {θ θ}_{q q - - 11})) \\ \cdot &Center Dot; \cdot \cdot \cdot &Center Dot; \\ {\overset{&OverBar; &OverBar;}{γ γ}}_{q q} = = {σ σ}_{α α}^{22} ((- - {θ θ}_{q q})) \end{matrix}\}

Represents the variance of the white noise sequence.

5. a kind of anti-stealing method based on electricity load management system data mining according to claim 4, is characterized in that, described short-term load forecasting equation is:

The formula for calculating the short-term load forecast value is:

6. A kind of anti-stealing method based on electricity load management system data mining according to claim 5, is characterized in that, in described step 3), the specific step of judging whether the current user has the possible behavior of stealing electricity is :

301) Setting the alarm factor r and the alarm window length W;

302) Calculating the alarm threshold y _th according to the load forecast value y: y _th = r y ;

303) Comparing the actual load curve and the alarm threshold curve in the same time period, if the actual load value y(t) from time n satisfies the following conditions:

y the y ((t t)) < < {y the y}_{th the th},, &ForAll; &ForAll; t t &Element; &Element; [[n no,, n no + + W W]]

Then it is judged that the current user has a possible behavior of stealing electricity, and an alarm of "possible electricity stealing" is issued.

7. An anti-stealing electricity method based on data mining of an electricity load management system according to claim 6, wherein the alarm factor r is: 1/2≤r≤2/3.