A method for detecting abnormal power consumption behavior in new energy distribution networks

Abstract

The present invention discloses a method for detecting abnormal electricity consumption behavior applicable to a new energy distribution network. Based on historical data of user stations, feature extraction is performed to obtain a sample set of a network for screening suspicious electricity consumption behavior. Three-layer deep learning screening networks are respectively constructed according to the coupling degree of data in a training set, and training of the suspicious behavior screening network is completed based on the sample set. The historical abnormal data is used as a sample set for false alarm identification, and an abnormal false alarm identification model is established. Data in an electricity collection system is input into a suspicious behavior screening network after feature extraction, and when the output abnormal state value exceeds a threshold, the user data is transmitted to an abnormal false alarm identification model. The abnormal false alarm identification model identifies whether there is an abnormal false alarm in the suspicious electricity consumption behavior of the user according to the characteristics of the electricity consumption data, and if so, the user is included in a monitoring list. The method is executed again for users in the monitoring list, and if the output abnormal state value still exceeds the threshold, it is determined that the user has abnormal electricity consumption behavior.

Description

Technical Field

[0001] The present invention belongs to the technical field of power detection, and specifically relates to a method for detecting abnormal power consumption behavior applicable to a new energy distribution network. Background Art

[0002] With the increasing penetration of new energy, traditional distribution networks are gradually transforming to new energy distribution networks. Compared with traditional distribution networks, new energy distribution networks have the characteristics of complex structure, strong distribution, and large temporal and spatial fluctuations. Traditional power consumption detection methods rely on manual investigation and on-site inspections. Such methods have many problems such as low detection efficiency and high detection delay, and are difficult to adapt to the characteristics of new energy distribution networks. In recent years, with the rise of artificial intelligence algorithms, various intelligent detection methods based on machine learning for abnormal power consumption behavior have gradually begun to be introduced into distribution networks, but the current mainstream methods still have certain defects when applied to new energy distribution networks: 1. The correlation between various data used to identify anomalies varies, and intelligent algorithms cannot adjust the correlation adaptively, which easily causes the model to have overfitting problems, resulting in a waste of computing power and resources; 2. Due to the differences in the characteristics of new energy distribution networks and traditional distribution networks, the power consumption recognition model trained based on the historical data of traditional distribution networks has a relatively serious misjudgment situation, and the complex and changeable characteristics of the new energy distribution network structure aggravate the misjudgment situation, which has caused a large number of abnormal power consumption false alarms in the power consumption collection system. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method for detecting abnormal electricity consumption behavior applicable to a new energy distribution network. The detection method constructs an abnormal false alarm recognition model and uses the model recognition results to determine whether there is an abnormality in electricity consumption.

[0004] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0005] A method for detecting abnormal power consumption behavior applicable to a new energy distribution network, the specific steps are as follows:

[0006] 1) Based on the historical data of the user's area, feature extraction is performed to obtain a sample set of the suspicious electricity behavior screening network;

[0007] 2) Construct three layers of deep learning screening networks according to the coupling degree of the data in the training set, and complete the training of the suspicious behavior screening network based on the sample set;

[0008] 3) Taking historical abnormal data as the sample set for false police identification, an abnormal false police identification model is established based on SVDD;

[0009] 4) After feature extraction, the data in the electricity consumption collection system is input into the suspicious behavior screening network. When the output abnormal state value exceeds the threshold, the user data is transmitted to the abnormal false alarm recognition model;

[0010] 5) The abnormal false alarm identification model identifies whether there is an abnormal false alarm for the user's suspicious electricity usage behavior based on the electricity usage data characteristics. If not, the user is determined to have abnormal electricity usage behavior; if yes, the user is included in the uncertainty monitoring list;

[0011] 6) After 24 hours, step 4) is executed again for the users in the monitoring list in step 5). If the output abnormal state value still exceeds the threshold, it is determined that the user has abnormal electricity consumption behavior.

[0012] Furthermore, the specific method for classifying the electricity consumption behavior characteristics of the historical data in step 1) is as follows:

[0013] 11) The historical data of the user area used to extract features includes current: A phase current I A 、B phase current I B 、C phase current I C , voltage: A phase voltage U A , B phase voltage U B 、C phase voltage U C , active power: total active power P Z , Phase A active power P A , B phase active power P B , C phase active power P C , reactive power: total reactive power Q Z , A phase reactive power Q A , B phase reactive power Q B , C phase reactive power Q C , power factor: active power factor Reactive power factor Event status: Undervoltage event of the main meter battery in the area;

[0014] 12) The network sample set data_example for screening suspicious electricity consumption behaviors obtained by feature extraction includes an output set Y and an input set X. The output set Y is a set of one-dimensional series, including abnormal state values ​​Y S The input set X is six sets of one-dimensional series, including voltage characteristic quantity U unb , current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State .

[0015] Preferably, the specific method of feature extraction in step 12) is as follows:

[0016] 121) Extract various electrical data in the collection system from 0 o'clock, with an extraction interval of 15 minutes, and extract 96 data in total;

[0017] 122) For the extraction of the feature quantity of the input set X, the expression formula is as follows:

[0018]

[0019] in, U A,i represents the voltage of the ith data point of phase A every day, U B,i Indicates the voltage of the ith data point of phase B per day, U C,i represents the voltage of the ith data point of phase C every day;

[0020]

[0021] in, I A,i represents the current of the ith data point per day in phase A, I B,i represents the current of the ith data point of phase B per day, I C,i represents the current of the ith data point of phase C per day;

[0022]

[0023] Among them, P A,i represents the active power of phase A at the ith data point per day, P B,i represents the active power of phase B at the ith data point per day, P C,i represents the active power of the i-th data point of phase C every day;

[0024]

[0025] Among them, Q A,i represents the active power of phase A at the ith data point per day, Q B,i represents the active power of phase B at the ith data point per day, Q C,i represents the active power of the i-th data point of phase C every day;

[0026]

[0027] in, represents the reactive power factor of the ith data point per day, represents the active power factor of the ith data point per day;

[0028]

[0029] Among them, the total meter voltage in the substation is abnormal, and the undervoltage event state characteristic quantity U State 1; the total meter voltage in the substation is normal, and the undervoltage event state characteristic quantity U State is 0.

[0030] Furthermore, in step 2), three layers of deep learning screening networks are constructed respectively according to the coupling degree of the data in the training set, and the specific method is as follows:

[0031] 21) Performing coupling degree association grouping for the characteristic quantities: First association level: voltage characteristic quantity U unb and current characteristic quantity I unb ; Second association level: active characteristic quantity P Min / Max and reactive characteristic quantity Q Min / Max ; Third correlation level: power factor characteristics And the area master meter battery undervoltage event status U State ;

[0032] 22) The deep learning screening network structure includes an input layer, a hidden layer and an output layer, wherein the input layer has 1 layer and 6 neurons; the hidden layer has 3 layers and 7-6-5 neurons; the output layer has 1 layer and 1 neuron;

[0033] 23) Hidden layer neuron categories: including learning neurons and transmission neurons:

[0034] The learning neuron connection mode is to keep connected with all neurons in the previous level and the next level, and the connection weight w ii is [-1,1], the transfer function is a tansig symmetric S-shaped function, and its expression formula is as follows:

[0035]

[0036] The output expression formula of the learning neuron is as follows:

[0037]

[0038] Among them, w ii represents the weight of the connection with the previous neuron, x i It represents the input of the i-th neuron in the previous layer of network to this neuron, and n represents the number of neurons in the previous layer of network;

[0039] The transfer neuron connection mode is that only the same neuron is connected at the previous level, and several neurons are connected at the next level. ii is always 1, and the transfer function is the del function, which is expressed as follows:

[0040] del(x)=x (9)

[0041] The output expression formula of the transfer neuron is as follows:

[0042] y=x (10)

[0043] 24) Connection method of neurons in each layer of network

[0044] There are 6 neurons in the input layer, x1, x2, x3, x4, x5, x6 correspond to the voltage characteristic quantity U respectively unb , current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State ;

[0045] The first hidden layer contains 7 neurons, of which h 11 ,h 12 ,h 13 For learning neurons, h 14 ,h 15 ,h 16 ,h 17 To transmit neurons, x1, x2 and h 11 ,h 12 ,h 13 Fully connected, x3, x4, x5, x6 and h 14 ,h 15 ,h 16 ,h 17 The second hidden layer contains 6 neurons, among which h 21 ,h 22 ,h 23 ,h 24 For learning neurons, h 25 ,h 26 For transmitting neurons, h 11 ,h 12 ,h 13 ,h 14 ,h 15 With h 21 ,h 22 ,h 23 ,h 24 Fully connected, h 16 ,h 17 With h 25 ,h 26 Connect in sequence; the third hidden layer contains 5 neurons, h 31 ,h 32 ,h 33 ,h 34 ,h 35 For learning neurons, h 21 ,h 22 ,h 23 ,h 24 ,h 25 ,h 26 With h 31 ,h 32,h 33 ,h 34 ,h 35 Fully connected;

[0046] The output layer contains a learning neuron y, h 31 ,h 32 ,h 33 ,h 34 ,h 35 Fully connected with y;

[0047] 25) Network training parameter setting

[0048] 85% of the data in the sample set of the screening network is randomly selected as the training set, and the remaining 15% is used as the test set; the learning rate interval is [0.01, 0.05], and the target error interval is [0.002, 0.01].

[0049] Furthermore, in step 3), the historical abnormal data is used as a sample set and an abnormal false alarm recognition model is established based on SVDD. The specific method is as follows:

[0050] 31) Extract and filter the arrays whose output set Y is 1 in the network sample set data_example. A total of n groups of data constitute the abnormal sample set data_fault, and its expression formula is as follows:

[0051] data_fault={s1,s2,...,s n} (11)

[0052] Among them, s i Contains 6 data,

[0053] 32) Construct an abnormal hypersphere F(R) based on the abnormal sample set data_fault 2 ,a,C), where R represents the radius of the hypersphere, a represents the center of the sphere, the coordinates are (η1,η2,η3,η4,η5,η6), 0≤η1,η2,η3,η4,η5,η6≤1, C represents the out-of-bounds penalty parameter, and the Gaussian kernel function K is introduced. The expression formula is as follows:

[0054]

[0055] Where σ represents the Gaussian parameter, σ>0, Represents Gaussian high-dimensional mapping function;

[0056] 33) Abnormal false police identification model

[0057] Output samples s whose abnormal state values ​​exceed the threshold f Input, calculate the distance d between it and the center a of the abnormal hypersphere, the expression formula is as follows:

[0058]

[0059] If d>R, the output result is no, the user is suspected of abnormal electricity usage, and a false alarm occurs in the suspicious behavior screening network. If d≤R, the output result is yes, and the user is suspected of abnormal electricity usage.

[0060] Preferably, the abnormal hypersphere parameter acquisition step in step 23) is as follows:

[0061] 321) Parameter solution model

[0062]

[0063] Among them, ξ i represents the slack variable, C>0;

[0064] 322) Parameter solution method

[0065] The expression formula after dual transformation of the optimization problem in the above formula (14) is as follows:

[0066]

[0067]

[0068] Among them, α1, α2, …, α n represents the Lagrange multiplier,

[0069] The expression formula of the center a of the abnormal hypersphere is as follows:

[0070]

[0071] The expression formula of the radius R of the abnormal hypersphere is as follows:

[0072]

[0073] Furthermore, the method for determining whether the output abnormal state value of step 4) exceeds the threshold value is as follows:

[0074] Output abnormal state value threshold S f :

[0075] S f =max(Y S1 ,Y S2 ,...,Y SN ) (18)

[0076] Among them, Y Sn Represents the output abnormal state value calculated by the nth group of normal feature data in the test set, and there are a total of N groups of normal feature data in the test set.

[0077] Compared with the prior art, it has the following advantages and beneficial effects:

[0078] The present invention uses the historical voltage, current, power, power factor, event reporting mark and abnormal state value of the user area as a training set, and uses ladder training to obtain a suspicious electricity behavior screening network; by summarizing historical abnormal data and building an abnormal false alarm recognition model based on support vector data description (SVDD). The electricity collection system transmits electricity information and reported events to the suspicious electricity behavior screening network in real time, and the data group determined to be suspicious behavior will be transmitted to the abnormal false alarm recognition model, and whether there is abnormality in electricity consumption will be determined based on the model recognition result.

[0079] The present invention is suitable for detecting abnormal power consumption behavior in new energy distribution networks. It effectively considers the correlation between various data used to identify anomalies, introduces the method of optimizing neural networks by using transfer neurons, improves the efficient use of computing power, and reduces the dependence of online abnormal power consumption behavior detection on hardware. By introducing abnormal false alarm recognition based on SVDD, through the highly sensitive suspicious screening of neural networks and the secondary verification of abnormal false alarms, while taking into account the success rate of power consumption anomaly recognition, the incidence rate of abnormal false alarms is reduced, and the accuracy rate of abnormal power consumption behavior investigation and handling is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a flow chart of a method for detecting abnormal power consumption behavior applicable to a new energy distribution network in the present invention;

[0081] Figure 2 It is a structure diagram of a neural network for screening suspicious behaviors in the present invention;

[0082] Figure 3 It is a diagram of the iterative process of the suspicious behavior screening neural network in the present invention;

[0083] Figure 4 It is a two-dimensional plane projection diagram of the abnormal hypersphere in the present invention. DETAILED DESCRIPTION

[0084] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The schematic implementation mode of the present invention and its description only use a method for detecting abnormal power consumption behavior applicable to a new energy distribution network to explain the present invention, and are not intended to limit the present invention.

[0085] like Figure 1 As shown, a method for detecting abnormal power consumption behavior applicable to a new energy distribution network is provided, and the specific steps are as follows:

[0086] 1) Based on the historical data of the user's area, feature extraction is performed to obtain a sample set of the suspicious electricity behavior screening network;

[0087] 2) Construct three layers of deep learning screening networks according to the coupling degree of the data in the training set, and complete the training of the suspicious behavior screening network based on the sample set;

[0088] 3) Taking historical abnormal data as the sample set for false police identification, an abnormal false police identification model is established based on SVDD;

[0089] 4) After feature extraction, the data in the electricity consumption collection system is input into the suspicious behavior screening network. When the output abnormal state value exceeds the threshold, the user data is transmitted to the abnormal false alarm recognition model;

[0090] 5) The abnormal false alarm identification model identifies whether there is an abnormal false alarm for the user's suspicious electricity usage behavior based on the electricity usage data characteristics. If not, it is determined that the user has abnormal electricity usage behavior; if yes, the user is included in the monitoring list;

[0091] 6) After 24 hours, step 4) is executed again for the users in the monitoring list in step 5). If the output abnormal state value still exceeds the threshold, it is determined that the user has abnormal electricity consumption behavior.

[0092] Furthermore, the specific method for classifying the electricity consumption behavior characteristics of the historical data in step 1) is as follows:

[0093] 11) The historical data of the user area used to extract features includes current: A phase current I A 、B phase current I B 、C phase current I C , voltage: A phase voltage U A , B phase voltage U B 、C phase voltage U C , active power: total active power P Z , Phase A active power P A , B phase active power P B , C phase active power P C , reactive power: total reactive power Q Z , A phase reactive power Q A , B phase reactive power Q B , C phase reactive power Q C , power factor: active power factor Reactive power factor Event status: Undervoltage event of the main meter battery in the area;

[0094] 12) The network sample set data_example for screening suspicious electricity consumption behaviors obtained by feature extraction includes an output set Y and an input set X. The output set Y is a set of one-dimensional series, including abnormal state values ​​Y S The input set X is six sets of one-dimensional series, including voltage characteristic quantity U unb, current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State .

[0095] Preferably, the specific method of feature extraction in step 12) is as follows:

[0096] 121) Extract various electrical data in the collection system from 0 o'clock, with an extraction interval of 15 minutes, and extract 96 data in total;

[0097] 122) For the extraction of the feature quantity of the input set X, the expression formula is as follows:

[0098]

[0099] in, U A,i represents the voltage of the ith data point of phase A every day, U B,i Indicates the voltage of the ith data point of phase B per day, U C,i represents the voltage of the ith data point of phase C every day;

[0100]

[0101] in, I A,i represents the current of the ith data point per day in phase A, I B,i represents the current of the ith data point of the day in phase B, I C,i represents the current of the ith data point of phase C per day;

[0102]

[0103] Among them, P A,i represents the active power of phase A at the ith data point per day, P B,i represents the active power of phase B at the ith data point per day, P C,i represents the active power of the i-th data point of phase C every day;

[0104]

[0105] Among them, Q A,i represents the active power of phase A at the ith data point per day, Q B,i represents the active power of phase B at the ith data point per day, Q C,i represents the active power of the i-th data point of phase C every day;

[0106]

[0107] in, represents the reactive power factor of the ith data point per day, represents the active power factor of the ith data point per day;

[0108]

[0109] Among them, the total meter voltage in the substation is abnormal, and the undervoltage event state characteristic quantity U State 1; the total meter voltage in the substation is normal, and the undervoltage event state characteristic quantity U State is 0.

[0110] Furthermore, in step 2), three layers of deep learning screening networks are constructed respectively according to the coupling degree of the data in the training set, and the specific method is as follows:

[0111] 21) Performing coupling degree association grouping for the characteristic quantities: First association level: voltage characteristic quantity U unb and current characteristic quantity I unb ; Second association level: active characteristic quantity P Min / Max and reactive characteristic quantity Q Min / Max ; Third correlation level: power factor characteristics And the area master meter battery undervoltage event status U State ;

[0112] 22) The deep learning screening network structure includes an input layer, a hidden layer and an output layer, wherein the input layer has 1 layer and 6 neurons; the hidden layer has 3 layers and 7-6-5 neurons; the output layer has 1 layer and 1 neuron;

[0113] 23) Neuron type: The hidden layer contains two types of neurons: learning neurons and transmission neurons:

[0114] The learning neuron connection mode is to keep connected with all neurons in the previous level and the next level, and the connection weight w ii is [-1,1], the transfer function is a tansig symmetric S-shaped function, and its expression formula is as follows:

[0115]

[0116] The output expression formula of the learning neuron is as follows:

[0117]

[0118] Among them, w ii represents the weight of the connection with the previous neuron, x i It represents the input of the i-th neuron in the previous layer of network to this neuron, and n represents the number of neurons in the previous layer of network;

[0119] The transfer neuron connection mode is that only the same neuron is connected at the previous level, and several neurons are connected at the next level. ii is always 1, and the transfer function is the del function, which is expressed as follows:

[0120] del(x)=x (9)

[0121] The output expression formula of the transfer neuron is as follows:

[0122] y=x (10)

[0123] 24) Connection method of neurons in each layer of network

[0124] There are 6 neurons in the input layer, x1, x2, x3, x4, x5, x6 correspond to the voltage characteristic quantity U respectively unb , current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State ;

[0125] The first hidden layer contains 7 neurons, of which h 11 ,h 12 ,h 13 For learning neurons, h 14 ,h 15 ,h 16 ,h 17 To transmit neurons, x1, x2 and h 11 ,h 12 ,h 13 Fully connected, x3, x4, x5, x6 and h 14 ,h 15 ,h 16 ,h 17 The second hidden layer contains 6 neurons, among which h 21 ,h 22 ,h 23 ,h 24 For learning neurons, h 25 ,h 26 For transmitting neurons, h 11 ,h 12 ,h 13 ,h 14 ,h 15 With h 21 ,h 22 ,h 23 ,h 24 Fully connected, h 16 ,h 17 With h 25,h 26 Connect in sequence; the third hidden layer contains 5 neurons, h 31 ,h 32 ,h 33 ,h 34 ,h 35 For learning neurons, h 21 ,h 22 ,h 23 ,h 24 ,h 25 ,h 26 With h 31 ,h 32 ,h 33 ,h 34 ,h 35 Fully connected;

[0126] The output layer contains a learning neuron y, h 31 ,h 32 ,h 33 ,h 34 ,h 35 Fully connected with y;

[0127] 25) Network training parameter setting

[0128] 85% of the data in the sample set of the screening network is randomly selected as the training set, and the remaining 15% is used as the test set; the learning rate interval is [0.01, 0.05], and the target error interval is [0.002, 0.01].

[0129] Furthermore, in step 3), the historical abnormal data is used as a sample set and an abnormal false alarm recognition model is established based on SVDD. The specific method is as follows:

[0130] 31) Extract and filter the arrays whose output set Y is 1 in the network sample set data_example. A total of n groups of data constitute the abnormal sample set data_fault, and its expression formula is as follows:

[0131] data_fault={s1,s2,...,s n} (11)

[0132] Among them, s i Contains 6 data,

[0133] 32) Construct an abnormal hypersphere F(R) based on the abnormal sample set data_fault 2 ,a,C), where R represents the radius of the hypersphere, a represents the center of the sphere, the coordinates are (η1,η2,η3,η4,η5,η6), 0≤η1,η2,η3,η4,η5,η6≤1, C represents the out-of-bounds penalty parameter, and the Gaussian kernel function K is introduced. The expression formula is as follows:

[0134]

[0135] Where σ represents the Gaussian parameter, σ>0, Represents Gaussian high-dimensional mapping function;

[0136] 33) Abnormal false police identification model

[0137] Output samples s whose abnormal state values ​​exceed the threshold f Input, calculate the distance d between it and the center a of the abnormal hypersphere, the expression formula is as follows:

[0138]

[0139] If d>R, the output result is no, the user is suspected of abnormal electricity usage, and a false alarm occurs in the suspicious behavior screening network. If d≤R, the output result is yes, and the user is suspected of abnormal electricity usage.

[0140] Preferably, the abnormal hypersphere parameter acquisition step in step 32) is as follows:

[0141] 321) Parameter solution model

[0142]

[0143] Among them, ξ i represents the slack variable, C>0;

[0144] 322) Parameter solution method

[0145] The expression formula after dual transformation of the optimization problem in the above formula (14) is as follows:

[0146]

[0147] Among them, α1, α2, …, α n represents the Lagrange multiplier, which is the parameter to be solved;

[0148] The expression formula of the center a of the abnormal hypersphere is as follows:

[0149]

[0150] The formula for the radius R of the abnormal hypersphere is as follows: α that satisfies 0<α<C k There are m in total, and the average radius of the anomalous hypersphere is R:

[0151]

[0152] Furthermore, the method for determining whether the output abnormal state value of step 4) exceeds the threshold value is as follows:

[0153] Output abnormal state value threshold S f :

[0154] S f =max(Y S1 ,Y S2 ,...,Y SN )(18)

[0155] Among them, Y Sn Represents the output abnormal state value calculated by the nth group of normal feature data in the test set, and there are a total of N groups of normal feature data in the test set.

[0156] Example

[0157] The specific implementation mode of the present invention is described in combination with the actual case data of a 35kV new energy distribution network in Sichuan. The method includes the following two parts: power consumption anomaly network model training and power consumption anomaly detection.

[0158] Step S1: Power consumption anomaly network model training:

[0159] A new energy distribution network originally had 193 ordinary users. After September 2021, 60 new energy users and 17 online new energy power sources (photovoltaic) were added. The metering method is high supply and low metering, and all users are three-phase power supply. There are 2154 sets of data, which are daily data randomly selected from 2017 to 2022. Each set of data includes 96 time series points. The historical data (including the data before the new energy power source and load are connected) are used for feature extraction, and the feature data set is used as the training sample and test sample to carry out network model training. The specific process is as follows:

[0160] Step S11 is feature extraction of historical data, which is divided into measurement value feature extraction (such as formulas 1 to 5) and state value feature extraction (such as formula 6), taking a group of historical data in the new energy distribution network as shown in Table 1.

[0161] Table 1 A group of historical data of new energy distribution network

[0162]

[0163] The characteristic data of the measured values ​​extracted from Table 1 according to Formulas 1 to 5 are shown in Table 2.

[0164] Table 2 Measurement value feature extraction value

[0165]

[0166] According to the state value feature data X6 extracted by formula 6=0, according to the historical audit record query, this group of users has no abnormal electricity consumption behavior on that day, so the neural network training set output Y=0 corresponding to this group of data; therefore, the neural network training feature data group corresponding to the historical data of this group in Table 1 is:

[0167] {x1,x2,x3,x4,x5,x6,Y}={0.0097,0.0074,0.5697,0,0.2398,0,0}

[0168] This is extended to 2154 groups of historical data in sequence, and a complete 2154×7 sample set is obtained: 1041 groups of neural network training feature data sets with Y=1 are abnormal feature data; 1113 groups of neural network training feature data sets with Y=0 are normal feature data.

[0169] Step S12: suspicious behavior screening network model construction, based on the 2154×7 sample set in step S11, the suspicious behavior screening network training and model verification are performed, and the specific steps are as follows:

[0170] Step S121 divides the sample set into a training set and a validation set according to a ratio of 85%-15%, wherein the ratio of the number of groups of abnormal feature data to normal feature data in the training set and the validation set is the same as that in the sample set, that is, the 2154×7 sample set is divided into a 1831×7 training set (including 885 groups of abnormal feature data and 946 groups of normal feature data) and a 323×7 validation set (including 156 groups of abnormal feature data and 167 groups of normal feature data);

[0171] Step S122 constructs a suspicious behavior screening neural network, whose network structure is as follows: Figure 2 As shown, x1, x2, x3, x4, x5, x6 in the training set and validation set correspond to the six input neurons in the input layer, Y corresponds to the output neuron in the output layer, and the training method is error back propagation;

[0172] Step S123 sets the training parameters: the learning method is LM method, the maximum number of iterations (net.trainParam.epochs) is 1000, the target error (net.trainParam.goal) is 0.002, and the learning rate (net.trainParam.lr) is 0.02;

[0173]

[0174] Step S124 starts network training, and starts training according to the training parameters in step S123. When the iteration error of the output result is less than the target error, the training ends, and the network performance verification is performed in parallel during the training. Figure 3As shown, when the training error reaches the 30th iteration, it is 1.92912×10 -3 , and the verification error is 1.14246×10 -3 , the number of iterations is less than the maximum limit, the error is less than the target error, and the suspicious behavior screening network is completed.

[0175] Step S125 abnormal state value threshold determination, according to formula (18), S f =0.784.

[0176] Step S13: abnormal false alarm recognition model is constructed, and an abnormal hypersphere F(R) is constructed based on the abnormal feature data in the 1041×7 sample set in step S11. 2 ,a,C), and solve the abnormal hypersphere F(R 2 ,a,C) Partial projection on the two-dimensional plane Figure 4 shown.

[0177] Step S2 imports real-time electricity consumption data, including 96 points of data starting from 0:00 and the battery undervoltage event of the total meter in the area.

[0178] Step S3 uses the real-time electricity data to perform feature extraction to obtain the user's daily feature array {x1, x2, x3, x4, x5, x6};

[0179] Step S4: Import the feature array {x1,x2,x3,x4,x5,x6} into the suspicious behavior screening network model trained in step S11 and calculate Y S , determine whether it exceeds S f =0.784, if it does not exceed the threshold, there is no abnormal electricity usage behavior, if it exceeds the threshold, it is determined whether it is in the monitoring list, if it exceeds the threshold and is in the monitoring list, it is determined that there is no abnormal electricity usage behavior, if it exceeds the threshold and is not in the monitoring list, it enters the abnormal false alarm identification;

[0180] In step S5, the abnormal false alarm recognition model determines whether there is a false alarm, that is, whether the feature array exceeding the threshold is within the abnormal hypersphere F. If it is within the abnormal hypersphere F, it is judged that there is no abnormal false alarm, and it can be judged that there is abnormal electricity consumption behavior; if it is outside the abnormal hypersphere F, it is judged as an abnormal false alarm, added to the uncertainty monitoring list and re-executed steps S3 to S4 with new daily electricity consumption data after 24 hours.

[0181] The above description is only an implementation mode of the present invention. It is stated again that for ordinary technicians in this technical field, several improvements can be made to the present invention without departing from the principle of the present invention. These improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A method for detecting abnormal power consumption behavior in a new energy distribution network, characterized in that: The specific steps are as follows: 1) Based on the historical data of the user's area, feature extraction is performed to obtain a sample set of the suspicious electricity behavior screening network; 2) Construct three layers of deep learning screening networks according to the coupling degree of the data in the training set, and complete the training of the suspicious behavior screening network based on the sample set; 3) Taking historical abnormal data as the sample set for false police identification, an abnormal false police identification model is established based on SVDD; 4) After feature extraction, the data in the electricity consumption collection system is input into the suspicious behavior screening network. When the output abnormal state value exceeds the threshold, the user data is transmitted to the abnormal false alarm recognition model; 5) The abnormal false alarm identification model identifies whether there is an abnormal false alarm for the user's suspicious electricity usage behavior based on the electricity usage data characteristics. If not, it is determined that the user has abnormal electricity usage behavior; if yes, the user is included in the monitoring list; 6) After 24 hours, step 4) is executed again for the users in the monitoring list in step 5). If the output abnormal state value still exceeds the threshold, it is determined that the user has abnormal electricity consumption behavior; In step 2), three layers of deep learning screening networks are constructed according to the coupling degree of the data in the training set, and the specific method is as follows: 21) Coupling degree association grouping for characteristic quantities: First association level: voltage characteristic quantity U unb and current characteristic quantity I unb ; Second association level: active characteristic quantity P Min / Max and reactive characteristic quantity Q Min / Max ; Third correlation level: power factor characteristics And the area master meter battery undervoltage event status U State ; 22) The deep learning screening network structure includes an input layer, a hidden layer and an output layer, wherein the input layer has 1 layer and 6 neurons; the hidden layer has 3 layers and 7-6-5 neurons; the output layer has 1 layer and 1 neuron; 23) Neuron type: The hidden layer contains two types of neurons: learning neurons and transmission neurons: The learning neuron connection mode is to keep connected with all neurons in the previous level and the next level, and the connection weight w ii is [-1,1], the transfer function is a tansig symmetric S-shaped function, and its expression formula is as follows: The output expression formula of the learning neuron is as follows: Among them, w ii represents the weight of the connection with the previous neuron, x i It represents the input of the i-th neuron in the previous layer of network to this neuron, and n represents the number of neurons in the previous layer of network; The transfer neuron connection mode is that only the same neuron is connected at the previous level, and several neurons are connected at the next level. ii is always 1, and the transfer function is the del function, which is expressed as follows: del(x′)=x′ (9) The output expression formula of the transfer neuron is as follows: y′=x′ (10) 24) Connection method of neurons in each layer of network The input layer has 6 neurons, x in1 ,x in2 ,x in3 ,x in4 ,x in5 ,x in6 Corresponding to voltage characteristic quantity U unb , current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State ; The first hidden layer contains 7 neurons, of which h 11 ,h 12 ,h 13 For learning neurons, h 14 ,h 15 ,h 16 ,h 17 To transmit neurons, x in1 ,x in2 With h 11 ,h 12 ,h 13 Fully connected, x in3 ,x in4 ,x in5 ,x in6 With h 14 ,h 15 ,h 16 ,h 17 The second hidden layer contains 6 neurons, among which h 21 ,h 22 ,h 23 ,h 24 For learning neurons, h 25 ,h 26 For transmitting neurons, h 11 ,h 12 ,h 13 ,h 14 ,h 15 With h 21 ,h 22 ,h 23 ,h 24 Fully connected, h 16 ,h 17 With h 25 ,h 26 Connect in sequence; the third hidden layer contains 5 neurons, h 31 ,h 32 ,h 33 ,h 34 ,h 35 For learning neurons, h 21 ,h 22 ,h 23 ,h 24 ,h 25 ,h 26 With h 31 ,h 32 ,h 33 ,h 34 ,h 35 Fully connected; The output layer contains a learning neuron y, h 31 ,h 32 ,h 33 ,h 34 ,h 35 Fully connected with y; 25) Network training parameter setting The sample set of the screening network randomly selects 85% of the data as the training set and the remaining 15% as the test set; The learning rate interval is [0.01, 0.05], and the target error interval is [0.002, 0.01].

2. A method for detecting abnormal power consumption behavior applicable to a new energy distribution network as claimed in claim 1, characterized in that: The specific method for classifying the electricity consumption behavior characteristics of the historical data in step 1) is as follows: 11) The historical data of the user area used to extract features includes current: A phase current I A 、B phase current I B 、C phase current I C , voltage: A phase voltage U A , B phase voltage U B 、C phase voltage U C , active power: total active power P Z , Phase A active power P A , B phase active power P B , C phase active power P C , reactive power: total reactive power Q Z 、A phase reactive power Q A , B phase reactive power Q B , C phase reactive power Q C , power factor: active power factor Reactive power factor Event status: Undervoltage event of the main meter battery in the area; 12) The network sample set data_example for screening suspicious electricity consumption behaviors obtained by feature extraction includes an output set Y and an input set X. The output set Y is a set of one-dimensional series, including abnormal state values ​​Y S The input set X is six sets of one-dimensional series, including voltage characteristic quantity U unb , current characteristic quantity I unb , Active characteristic quantity P Min / Max , reactive characteristic quantity Q Min / Max , Power factor characteristics Area master meter battery undervoltage event status U State .

3. A method for detecting abnormal power consumption behavior applicable to a new energy distribution network as claimed in claim 2, characterized in that: The specific method of feature extraction in step 12) is as follows: 121) Extract various electrical data in the collection system from 0 o'clock, with an extraction interval of 15 minutes, and extract 96 data in total; 122) For the extraction of the feature quantity of the input set X, the expression formula is as follows: in, U A,i represents the voltage of the ith data point of phase A every day, U B,i Indicates the voltage of the ith data point of phase B per day, U C,i represents the voltage of the ith data point of phase C every day; in, I A,i represents the current of the ith data point per day in phase A, I B,i represents the current of the ith data point of phase B per day, I C,i represents the current of the ith data point of phase C per day; Among them, P A,i represents the active power of phase A at the ith data point per day, P B,i represents the active power of phase B at the ith data point per day, P C,i represents the active power of the i-th data point of phase C every day; Among them, Q A,i represents the reactive power of phase A at the ith data point per day, Q B,i represents the reactive power of phase B at the ith data point per day, Q C,i represents the reactive power of the ith data point of phase C every day; in, represents the reactive power factor of the ith data point per day, represents the active power factor of the ith data point per day; Among them, the total meter voltage in the substation is abnormal, and the undervoltage event state characteristic quantity U State 1; the total meter voltage in the substation is normal, and the undervoltage event state characteristic quantity U State is 0.

4. The abnormal power consumption behavior detection method applicable to the new energy distribution network as claimed in claim 1 is characterized by: The step 3) uses the historical abnormal data as a sample set and establishes an abnormal false alarm recognition model based on SVDD, and the specific method is as follows: 31) Extract and filter the array of output set Y as 1 in the network sample set data_example. A total of n1 groups of data constitute the abnormal sample set data_fault, and its expression formula is as follows: data_fault={s1,s2,...,s n1 } (11) Among them, s i Contains 6 data, 32) Construct an abnormal hypersphere F(R) based on the abnormal sample set data_fault 2 ,a,C), where R represents the radius of the hypersphere, a represents the center of the sphere, the coordinates are (η1,η2,η3,η4,η5,η6), 0≤η1,η2,η3,η4,η5,η6≤1, C represents the out-of-bounds penalty parameter, and the Gaussian kernel function K is introduced. The expression formula is as follows: Where σ represents the Gaussian parameter, σ>0, Represents Gaussian high-dimensional mapping function; 33) Abnormal false police identification model Output the sample S whose abnormal state value exceeds the threshold f Input, calculate the distance d between it and the center a of the abnormal hypersphere, the expression formula is as follows: in, represents the Lagrange multiplier, If d>R, the output result is no, the user is suspected of abnormal electricity usage, and a false alarm occurs in the suspicious behavior screening network. If d≤R, the output result is yes, and the user is suspected of abnormal electricity usage.

5. A method for detecting abnormal power consumption behavior applicable to a new energy distribution network as claimed in claim 4, characterized in that: The steps for obtaining the abnormal hypersphere parameters in step 32) are as follows: 321) Parameter solution model Among them, ξ i represents the slack variable, C>0; 322) Parameter solution method The expression formula after dual transformation of the optimization problem in the above formula (14) is as follows: The expression formula of the center a of the abnormal hypersphere is as follows: The expression formula of the radius R of the abnormal hypersphere is as follows: Furthermore, the method for determining whether the output abnormal state value of step 4) exceeds the threshold value is as follows: Output abnormal state value threshold S f : S f =max(Y S1 ,AND S2 ,...,AND SN )(18) Among them, Y SN Represents the output abnormal state value calculated by the Nth group of normal feature data in the test set, and there are a total of N groups of normal feature data in the test set.

Images