A non-intrusive load anomaly identification method and system for a power utilization system

Abstract

The application provides a kind of non-invasive load anomaly identification method of power utilization system, including real-time acquisition electrical load signal and noise reduction;Based on the electrical load signal after noise reduction, detect whether switching event occurs;If it occurs, after the load working state reaches stability, periodically collect corresponding load voltage and load current and pretreatment;Using grid mapping method, each cycle V-I trajectory is respectively mapped into RGB color image three-channel pixel matrix according to reactive power, power factor and current sequence distribution numerical characteristics;Extract each cycle load current each high harmonic amplitude and after binary conversion according to the size of pixel matrix, with the RGB color image three-channel pixel matrix after mapping fusion into color mixed image;Color mixed image is input into the convolutional neural network trained in advance to identify load anomaly.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load identification, and particularly relates to a non-intrusive load anomaly identification method and system of a power system. BACKGROUND

[0002] In recent years, with the high popularity of electrified equipment, the probability of electrical fire is also increasing. Among them, one-third of the larger fire is caused by electrical reasons, and most of them are electrical line faults. Through these data, the severe situation of power safety is revealed. As a crowd gathering place, the college dormitory often has electrical fires, so it is necessary to study the prevention of electrical fires in college dormitories.

[0003] Nowadays, the protection of electrical fires is mainly through existing hardware facilities, such as various smoke alarms, which need to sense related physical phenomena, and then rely on manual methods to avoid the occurrence of fires, but this method must have the premise of the occurrence of burning phenomenon. Therefore, it is necessary to reduce the probability of fire from the perspective of electrical line safety risk assessment before the occurrence of electrical fires.

[0004] At present, an important means of electrical line safety risk assessment is non-intrusive load monitoring (NILM) of the power system. However, this method has the following shortcomings, including: (1) As the amount of raw data of load basic characteristics gradually increases, the identification process becomes more complex, resulting in a decrease in identification efficiency; (2) Due to the high similarity of some electrical power, current and other load characteristics, the feature overlap phenomenon occurs, and the identification effect is poor. Even if the V-I trajectory characteristics of the load in the steady state are used to improve the load identification, only the shape characteristics of the voltage-current signal are retained, but the power difference between devices cannot be presented, a large amount of valuable information is lost, and the identification accuracy is limited.

[0005] Therefore, it is necessary to propose a new non-intrusive load monitoring method, which can not only distinguish between loads with similar V-I trajectories and small power size differences to improve identification accuracy, but also reduce the computational complexity of the identification algorithm and improve accuracy. SUMMARY

[0006] The technical problem to be solved by the embodiments of the present application is to provide a non-intrusive load anomaly identification method and system of a power system, which can not only distinguish between loads with similar V-I trajectories and small power size differences to improve identification accuracy, but also reduce the computational complexity of the identification algorithm and improve accuracy.

[0007] In order to solve the above technical problems, the embodiments of the present application provide a non-intrusive load anomaly identification method of a power system, which comprises the following steps:

[0008] collecting an electrical load signal on a load bus of an electricity utilization system in real time and performing noise reduction processing;

[0009] detecting whether a switching event occurs based on the electrical load signal after noise reduction processing;

[0010] If it is detected that a switching event occurs, periodically collect the corresponding load voltage and load current after the load operating state reaches stability, and pre-process the load voltage and load current to obtain a load voltage and load current whose V-I trajectory in each cycle is represented by a continuous grid;

[0011] Based on the load voltage and load current whose V-I trajectory in each cycle is represented by a continuous grid, use a grid mapping method to map the V-I trajectory in each cycle according to the reactive power, power factor, and current sequence distribution numerical characteristics into a preset RGB color image three-channel pixel matrix, respectively;

[0012] Extract the amplitude of each high-order harmonic of the load current in each cycle, and after binary conversion according to the size of the pixel matrix, further fuse with the mapped RGB color image three-channel pixel matrix to form a color mixed image;

[0013] Input the color mixed image into a pre-trained convolutional neural network to identify load abnormalities.

[0014] The step of performing noise reduction processing on the electrical load signal is specifically:

[0015] The electrical load signal is selected for wavelet packet decomposition with a wavelet packet basis and a certain number of decomposition layers, and the decomposition coefficients are subjected to corresponding threshold processing, and the low-frequency coefficients and high-frequency coefficients of the nth layer are further reconstructed to obtain the electrical load signal after noise reduction processing; wherein the number of decomposition layers is 6, and the wavelet packet function is sym12.

[0016] The detection of whether a switching event occurs based on the electrical load signal after noise reduction processing is determined by a double-sided sliding window algorithm; wherein

[0017] The specific steps of the double-sided sliding window algorithm are to set two sliding windows, calculate the difference between the average values of the two windows, if the difference is greater than a set first threshold value, it is considered that a switching event occurs, and the time when the event occurs is recorded; continue to detect from the time when the event occurs, if the difference is less than a set second threshold value, it is considered that the switching event ends, the time when the event ends is recorded, and the load is in steady state operation.

[0018] The load voltage and the load current are pre-processed by a bilinear interpolation method; wherein the specific steps of the bilinear interpolation method include:

[0019] First step, set the resolution of the grid to N*N, calculate the size of each cell by the following formula (1):

[0020]

[0021] Where, i min , i max are the minimum and maximum values of current sampling value, v min , v max are the minimum and maximum values of voltage sampling value, Δi, Δv are the size of each cell (pixel point);

[0022] Second step, calculate the distance D m between two adjacent sampling points after mapping according to the following formula (2) m ; If the distance D m is greater than 1, it indicates that the distance between two points is greater than the length or width of the cell, and the interval between two points needs to be interpolated to complete, the specific interpolation process is shown in formula (3) and formula (4);

[0023]

[0024]

[0025]

[0026] Where, v m and i m are the voltage and current values of the mth sampling point of the sampling data; is the number of interpolation points needed to be filled between the mth and (m+1)th sampling points, (v' m+k , i' m+k ) is the kth interpolation point filled, k = 1, 2, …, K m .

[0027] Wherein, the step of mapping each period V-I trajectory to the preset RGB color image three channel pixel matrix according to the voltage, current and power factor value characteristics specifically includes:

[0028] First step, construct three 1 matrices with N*N dimensions, respectively R, G and B matrices, calculate the mapping coordinates of each sample point after interpolation according to formula (5);

[0029]

[0030] Second step, according to formula (6), (7) and (8), respectively, assign values to the coordinates of the corresponding sample points in the R, G and B matrices;

[0031]

[0032]

[0033]

[0034]

[0035] Wherein, T m is the number of occurrences of each coordinate point; i fm is the reactive current; v m is the voltage of the sample point; P is the active power; i rms , v rms are the current and voltage effective values of the collected data respectively; I e indicates the distribution of the current sequence; M is the number of collected data in the stable period.

[0036] Wherein, the amplitude of each high harmonic of the load current in each cycle is extracted, and after binary conversion of the corresponding number of bits according to the size of the pixel matrix, further fusion is carried out with the mapped RGB color image three-channel pixel matrix to form a color mixed image. The specific steps include:

[0037] The amplitudes of the fundamental wave, third harmonic, fifth harmonic, seventh harmonic, ninth harmonic and eleventh harmonic of the load current waveform are extracted using the FFT transform algorithm, and are rounded;

[0038] The amplitudes of the above harmonics are converted into binary numbers, and the number of binary digits is equal to the size of the pixel matrix;

[0039] The converted binary numbers are arranged by column, and after being divided into three groups of N*2 matrices, they are respectively merged with R, G and B three-channel pixel matrices; wherein the combination of the fundamental wave and the third harmonic is merged with the R matrix; the fifth harmonic and the seventh harmonic are merged with the G channel pixel matrix; the ninth harmonic and the eleventh harmonic are merged with the B channel pixel matrix;

[0040] The generated three-channel fusion pixel matrix is superimposed to obtain the mixed color image.

[0041] Wherein, the convolutional neural network includes three convolutional layers, two pooling layers and one fully connected layer.

[0042] The embodiment of the application also provides a non-intrusive load anomaly identification device for a power utilization system, which comprises:

[0043] A signal acquisition and processing unit is configured to acquire electrical load signals on a load bus of the power utilization system in real time and perform noise reduction processing.

[0044] A switching event detection unit is configured to detect whether a switching event occurs based on the electrical load signals after noise reduction processing.

[0045] a V-I network processing unit, configured to periodically collect the load voltage and the load current after the load operating state reaches stability if it is detected that the switching event occurs, and to pre-process the load voltage and the load current to obtain the load voltage and the load current of each cycle V-I trajectory characterized by a continuous grid;

[0046] a V-I trajectory mapping unit, configured to map each cycle V-I trajectory according to the reactive power, the power factor and the current sequence distribution value features into a preset RGB color image three-channel pixel matrix by using a grid mapping method based on the load voltage and the load current of each cycle V-I trajectory characterized by a continuous grid;

[0047] a color mixed image acquisition unit, configured to extract the high-order harmonic amplitude of each cycle load current, and to further fuse the extracted high-order harmonic amplitude with the mapped RGB color image three-channel pixel matrix after binary conversion of corresponding bits according to the pixel matrix size to form a color mixed image;

[0048] a load anomaly recognition unit, configured to input the color mixed image into a pre-trained convolutional neural network to recognize the load anomaly.

[0049] The convolutional neural network comprises three convolutional layers, two pooling layers and one fully connected layer.

[0050] The embodiment of the present application has the following beneficial effects:

[0051] The RGB color mixed image constructed by the present application can fully utilize the V-I trajectory features, the reactive power size difference, the power factor size difference, the current distribution difference and the harmonic amplitude difference, so that the load feature information is more abundant, the load with similar V-I trajectories and small power size difference can be distinguished to improve the recognition accuracy, and the operation complexity of the recognition algorithm is reduced to improve the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings obtained according to these drawings without creative labor are still within the scope of the present application.

[0053] Figure 1 A flowchart of a non-intrusive load anomaly recognition method of a power utilization system according to an embodiment of the present application is provided.

[0054] Figure 2A structural diagram of a convolutional neural network model in a non-intrusive load anomaly identification method of an electricity utilization system provided by an embodiment of the present application is provided.

[0055] Figure 3 An effect diagram of load bilinear interpolation processing in a non-intrusive load anomaly identification method of an electricity utilization system provided by an embodiment of the present application is provided.

[0056] Figure 4 An effect diagram of a three-channel pixel image of a load V-I trajectory and a synthesized color image in a non-intrusive load anomaly identification method of an electricity utilization system provided by an embodiment of the present application is provided.

[0057] Figure 5 An effect diagram of a three-channel mixed image of a load and a synthesized color mixed image in a non-intrusive load anomaly identification method of an electricity utilization system provided by an embodiment of the present application is provided.

[0058] Figure 6 A structural schematic diagram of a non-intrusive load anomaly identification device of an electricity utilization system provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings.

[0060] As shown in Figure 1 A non-intrusive load anomaly identification method of an electricity utilization system provided by an embodiment of the present application is provided, and the method comprises the following steps:

[0061] Step S1, real-time collection of electrical load signals on a load bus of an electricity utilization system and noise reduction processing;

[0062] Step S2, detection of whether a switching event occurs based on the electrical load signals after noise reduction processing;

[0063] Step S3, if it is detected that a switching event occurs, then after the load working state reaches stability, the corresponding load voltage and load current are periodically collected, and the load voltage and load current are preprocessed to obtain a load voltage and a load current each of which is characterized by a continuous grid for each V-I trajectory;

[0064] Step S4, based on the load voltage and the load current each of which is characterized by a continuous grid for each V-I trajectory, a grid mapping method is used to map each V-I trajectory according to the reactive power, the power factor and the current sequence distribution numerical characteristics into a preset RGB color image three-channel pixel matrix, respectively;

[0065] Step S5, extract the amplitude of each high-order harmonic of the load current of each cycle, and after binary conversion according to the size of the pixel matrix, further fuse with the mapped RGB color image three-channel pixel matrix to form a color mixed image;

[0066] Step S6, input the color mixed image into the pre-trained convolutional neural network to identify the load abnormality.

[0067] The specific process is, before step S1, actually collect single load data of commonly used equipment and two-by-two combined mixed load equipment, then extract the color mixed image of each load data of each cycle to form a training set and a test set, and train a convolutional neural network model to obtain a pre-trained convolutional neural network as a subsequent load abnormality recognition network. It can be understood that the convolutional neural network structure is relatively simple and can be directly run on an embedded device.

[0068] In one embodiment, the convolutional neural network model is as shown in Figure 2 includes three convolutional layers, two pooling layers, and one fully connected layer.

[0069] In step S1, the electrical load signal on the load bus of the power utilization system is collected in real time, that is, the load voltage data and the load current data are collected.

[0070] Then, the electrical load signal is denoised, and the steps are as follows:

[0071] For the electrical load signal, wavelet packet basis and a certain number of decomposition layers are selected for wavelet packet decomposition, and the decomposition coefficients are subjected to corresponding threshold processing, and the low-frequency coefficients and high-frequency coefficients of the nth layer are further reconstructed to obtain the denoised electrical load signal; wherein the number of decomposition layers is 6, and the wavelet packet function is sym12.

[0072] In one example, the voltage and current data of the power utilization end are collected in real time and denoised using wavelet packets, the sampling frequency is 25kHz, each cycle contains 500 sampling points, the wavelet packet denoising is a hard threshold method, the number of decomposition layers is 6, and the wavelet packet function is sym12.

[0073] In step S2, whether a switching event occurs is determined by a double-sided sliding window algorithm. The specific steps of the double-sided sliding window algorithm are as follows: two sliding windows are set, the difference between the average values of the two windows is calculated, if the difference is greater than a first threshold value, it is considered that a switching event occurs, and the event occurrence time is recorded; continue to detect from the event occurrence time, if the difference is less than a second threshold value, it is considered that the switching event ends, the event end time is recorded, and the load is in steady state operation.

[0074] In step S3, first, it is detected that the switching event occurs, and then the corresponding load voltage and load current are periodically collected after the load working state reaches stability;

[0075] Secondly, the load voltage and load current are preprocessed by the bilinear interpolation method to ensure the continuity of the V-I trajectory mapping image; wherein the specific steps of the bilinear interpolation method include:

[0076] Firstly, the resolution of the grid (image) is set to N*N, and the size of each cell (pixel point) is calculated by the following formula (1):

[0077]

[0078] Wherein, i min , i max are the minimum and maximum values of the current sampling value, v min , v max are the minimum and maximum values of the voltage sampling value, and Δi, Δv are the size of each cell (pixel point).

[0079] Secondly, the distance D m between the two adjacent sampling points after mapping is calculated according to the following formula (2): m If the distance D m is greater than 1, it indicates that the distance between the two points is greater than the length or width of the cell, and the discontinuity phenomenon occurs, and the interpolation is needed to complete the interval between the two points. The specific interpolation process is shown in formula (3) and formula (4).

[0080]

[0081]

[0082]

[0083] Wherein, v m and i m are the voltage and current values of the mth sampling point of the sampling data; is the number of interpolation points needed to be filled between the mth and (m+1)th sampling points, (v’ m+k , i’ m+k ) is the kth interpolation point filled, k=1,2,…,K m .

[0084] In step S4, the grid mapping method is used to map each period V-I trajectory into the preset N*N*3 RGB color image three-channel pixel matrix according to the reactive power, power factor and current sequence distribution numerical characteristics, with voltage as the horizontal coordinate and current as the vertical coordinate. The steps include:

[0085] Firstly, three all-1 matrices of N*N dimension are constructed, which are R, G and B matrices respectively, and the mapping coordinates of each sample point after interpolation are calculated according to formula (5);

[0086]

[0087] Secondly, the coordinates of the corresponding sample points in the R, G and B matrices are respectively assigned values according to formula (6), (7) and (8);

[0088]

[0089]

[0090]

[0091]

[0092] Wherein, T m is the number of occurrences of each coordinate point; i fm is the reactive current; v m is the voltage of the sample point; P is the active power; i rms , v rms are the current and voltage effective values of the collected data; I e represents the distribution of the current sequence; and M is the number of collected data in a stable period.

[0093] In step S5, firstly, the FFT transform algorithm is used to extract the amplitudes of the fundamental wave, the third harmonic, the fifth harmonic, the seventh harmonic, the ninth harmonic and the eleventh harmonic of the load current waveform, and rounding is performed;

[0094] Secondly, the amplitudes of the above harmonics are converted into binary numbers, and the number of binary digits is equal to the size of the pixel matrix;

[0095] Then, the converted binary numbers are arranged by column, and after being divided into three groups of N*2 matrices, they are respectively merged with the R, G and B three-channel pixel matrices; wherein the combination of the fundamental wave and the third harmonic is merged with the R matrix; the fifth harmonic and the seventh harmonic are merged with the G channel pixel matrix; and the ninth harmonic and the eleventh harmonic are merged with the B channel pixel matrix;

[0096] Finally, the generated three-channel fusion pixel matrix is superimposed to obtain a mixed color image.

[0097] It can be seen that the color mixed image in the embodiment of the application has included the information such as the V-I trajectory feature, the reactive power size difference, the power factor size difference, the current distribution difference and the harmonic amplitude difference, so the recognition effect is much better than that of the method using any of the above information alone.

[0098] In step S6, the color mixed image obtained in step S5 is input to the pre-trained convolutional neural network, and the recognition result is obtained. When the identified load type is a violation load, the corresponding alarm is started, so as to achieve the purpose of real-time monitoring of the power load.

[0099] As Figures 3 to 5 shown, the application scenario of the non-intrusive load abnormality identification method provided by the embodiment of the application is further described, as follows:

[0100] The single load data of three commonly used devices in dormitories and the mixed load devices combined in pairs are actually collected, and then the color mixed image of each load data per cycle is extracted to form a training set and a test set, and a convolutional neural network model is trained as an identification network.

[0101] The load voltage data and current data are collected: the voltage and current data of the power consumption at the household end are collected in real time, and wavelet packets are used for denoising, with a sampling frequency of 25 kHz and 500 sampling points per cycle. The wavelet packet denoising is a hard threshold method, the decomposition layer is 6, and the wavelet packet function is sym12.

[0102] Whether a switching event occurs is determined by a double-sided sliding window algorithm, two average power sliding windows are set, each window contains the collected data in one cycle, the difference between the average values of the two windows is calculated, if the difference is greater than the set first threshold value, it is considered that a switching event occurs, and the event occurrence time is recorded; continue to detect from the event occurrence time, if the difference is less than the set second threshold value, it is considered that the switching event ends, the event end time is recorded, and the load is in steady state operation.

[0103] After detecting that a switching event occurs, the collected load voltage data and current data are subjected to bilinear interpolation processing to ensure the continuity of the V-I trajectory mapping image. As Figure 3 shown, after the interpolation processing of the load, the V-I trajectory mapping image eliminates the discontinuity.

[0104] Using the grid mapping method, the voltage is taken as the horizontal coordinate and the current is taken as the vertical coordinate, and the V-I trajectory of each cycle is mapped into a 64*64*3 RGB color image three-channel pixel matrix according to the voltage, current and power factor value characteristics, including the following steps:

[0105] A 3-dimension 64*64 all-1 matrix is constructed, which is R, G and B matrix respectively, and the mapping coordinates of each sample point after interpolation are calculated according to formula (5).

[0106] According to formulas (6), (7) and (8), the coordinates (pixel values) of the corresponding sample points in the R, G and B matrices are assigned, asFigure 4 The load V-I is mapped to the three-channel pixel image and the synthesized color image.

[0107] The amplitude of each high-order harmonic of the load current in each cycle is extracted, and the corresponding binary conversion of the grid size is performed, and is fused with the RGB three-channel pixel matrix to form a mixed color image, and the specific steps include the following steps:

[0108] The amplitudes of the fundamental wave, the third harmonic, the fifth harmonic, the seventh harmonic, the ninth harmonic and the eleventh harmonic of the current waveform are extracted using the FFT transform algorithm, and are rounded.

[0109] The amplitudes of the above harmonics are converted into binary numbers, and the number of binary digits is equal to the size of the pixel matrix. The converted binary numbers are arranged in columns and divided into three groups of 64*2 matrices, so as to be combined with the R, G and B three-channel pixel matrices respectively, wherein the combination of the fundamental wave and the third harmonic is combined with the R matrix, the fifth harmonic and the seventh harmonic are combined with the G channel pixel matrix, and the ninth harmonic and the eleventh harmonic are combined with the B channel pixel matrix, and the combination effect is as shown in Figure 5 .

[0110] The obtained color mixed image is input into the convolutional neural network trained in advance, and the recognition result is obtained. When the recognized load type is a violation load, the corresponding alarm is started, so as to achieve the purpose of real-time monitoring of the power load.

[0111] The effects of the non-intrusive load abnormality identification method provided by the embodiments of the present application on several violation loads are verified by experiments:

[0112] The experimental data is collected from the load equipment on site, and the current and voltage data characteristics of the hot water kettle, hair dryer, dryer and several mixed conditions are collected to verify the results as follows:

[0113] Load type Load state Recognition rate Hot kettle Heating 100％ Hair dryer Hot air 100％ Dryer Hot air 100％ Hot kettle + dryer Heating + hot air 100％ Hair dryer + dryer Hot air + hot air 100％ Hot kettle + hair dryer Heating + hot air 100％

[0114] As shown in Figure 6 , it is a non-intrusive load abnormality identification device of an electricity using system provided in the embodiments of the present application, which comprises;

[0115] The signal acquisition and processing unit 110 is used for real-time acquisition of the electrical load signal on the load bus of the electricity using system and noise reduction processing;

[0116] The switching event detection unit 120 is used for detecting whether a switching event occurs based on the electrical load signal after noise reduction processing;

[0117] The V-I network processing unit 130 is configured to periodically collect the load voltage and the load current after the load working state reaches stability if the switching event is detected, and pre-process the load voltage and the load current to obtain the load voltage and the load current of each cycle V-I trajectory characterized by a continuous grid.

[0118] The V-I trajectory mapping unit 140 is configured to map each cycle V-I trajectory according to the reactive power, the power factor and the current sequence distribution value features into a preset RGB color image three-channel pixel matrix by using a grid mapping method based on the load voltage and the load current of each cycle V-I trajectory characterized by a continuous grid.

[0119] The color mixed image acquisition unit 150 is configured to extract the high-order harmonic amplitude of each cycle load current, and after the corresponding binary conversion of the pixel matrix size is performed, the color mixed image is further fused with the mapped RGB color image three-channel pixel matrix to form a color mixed image.

[0120] The load anomaly recognition unit 160 is configured to input the color mixed image into a pre-trained convolutional neural network to recognize the load anomaly.

[0121] The convolutional neural network includes three convolutional layers, two pooling layers and one fully connected layer.

[0122] The embodiment of the present application has the following beneficial effects:

[0123] The RGB color mixed image can fully utilize the V-I trajectory features, the reactive power size difference, the power factor size difference, the current distribution difference and the harmonic amplitude difference, so that the load feature information is more abundant, the power value difference between devices can be presented, the recognition accuracy is improved, the operation complexity of the recognition algorithm is reduced, and the accuracy is improved.

[0124] It should be noted that the units included in the above system embodiment are only divided according to the function logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only for easy mutual differentiation, and are not used to limit the protection scope of the present application.

[0125] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc.

[0126] The above merely provides the preferred embodiment of the present application, and cannot allude the protection scope of the present application, therefore, any equivalent changes made according to the claims of the present application shall be within the scope of the present application.

Claims

1. A non-intrusive load anomaly identification method of an electrical power system, characterized by, The method comprises the following steps: Real-time acquisition of the electrical load signal on the load bus of the power utilization system and noise reduction processing; Based on the electrical load signal after noise reduction processing, it is detected whether a switching event occurs; If it is detected that a switching event occurs, the corresponding load voltage and load current are periodically acquired after the load working state reaches stability, and the load voltage and load current are preprocessed to obtain a load voltage and a load current, each cycle V-I trajectory of which is represented by a continuous grid; Based on the load voltage and the load current, each cycle V-I trajectory of which is represented by a continuous grid, a grid mapping method is used to map each cycle V-I trajectory according to the reactive power, the power factor and the current sequence distribution numerical characteristics into a preset RGB color image three-channel pixel matrix; Each high-order harmonic amplitude of the load current of each cycle is extracted, and after being converted into a binary number according to the size of the pixel matrix, the binary number is further fused with the mapped RGB color image three-channel pixel matrix to form a color mixed image; The color mixed image is input into a previously trained convolutional neural network to identify load abnormalities. The step of mapping each cycle V-I trajectory according to the reactive power, the power factor and the current sequence distribution numerical characteristics into a preset RGB color image three-channel pixel matrix specifically comprises: The first step, constructing three full 1 matrices with 3 dimensions N*N R, G, and B matrices, respectively, and calculating the mapping coordinates of each sample point after interpolation according to formula (5); （5）； In the second step, the coordinates of the corresponding sample points in the R, G and B matrices are assigned values according to formulas (6), (7) and (8). wherein, is the number of occurrences for each coordinate point; is the reactive current; is the voltage of the sample point; is the active power; , are the current and voltage effective values of the collected data, respectively; represents the distribution of the current sequence; M is the number of collected data in the stable period.

2. The non-intrusive load anomaly identification method of the power utilization system according to claim 1, wherein, The step of performing noise reduction processing on the electrical load signal specifically comprises: For the electrical load signal, wavelet packet decomposition is performed on a wavelet packet basis and a certain decomposition level, and corresponding threshold processing is performed on the decomposition coefficients, and the low-frequency coefficients and the high-frequency coefficients of the nth layer are further reconstructed to obtain the electrical load signal after noise reduction processing; wherein the decomposition level is 6, and the wavelet packet function is sym12.

3. The non-intrusive load anomaly identification method of the power utilization system according to claim 1, wherein, The load voltage and the load current are preprocessed by a bilinear interpolation method; wherein the specific steps of the bilinear interpolation method comprise: The first step, set the resolution of the grid to N*N The size of each cell is calculated by the following equation (1): （1）； wherein, , are minimum and maximum values of the current sample values, respectively, , are minimum and maximum values of the voltage sample values, respectively, , is the size of each cell pixel point; Second step, calculate the distance between two adjacent sampling points after mapping according to the following formula (2) If the distance is greater than 1, it indicates that the distance between the two points is greater than the length or width of the cell, and discontinuity occurs, at which time interpolation is needed to complete the interval between the two points. The specific interpolation process is shown in formulas (3) and (4). ​ （2） （3） （4） wherein, and are the voltage and current values of the mth sampling point of the sampling data, respectively; is the number of interpolation points needed to be supplemented between the mth and (m+1)th sampling points, , is the kth interpolation point filled in, k = 1, 2, …, .

4. The non-intrusive load anomaly identification method of the power utilization system according to claim 3, wherein, The specific steps of extracting each high-order harmonic amplitude of the load current of each cycle, converting the high-order harmonic amplitude into a binary number according to the size of the pixel matrix, and further fusing the binary number with the mapped RGB color image three-channel pixel matrix to form a color mixed image comprise: The amplitudes of the fundamental wave, the third harmonic, the fifth harmonic, the seventh harmonic, the ninth harmonic and the eleventh harmonic of the load current waveform are extracted using the FFT transformation algorithm, and are rounded. The amplitudes of the above harmonics are converted into binary numbers, and the number of binary digits is equal to the size of the pixel matrix. The converted binary numbers are arranged in columns and are to be divided into three groups N*2 After the matrix of the converted binary numbers is divided into three groups, the three groups are respectively merged with R, G and B channel pixel matrices; the fundamental wave and the third harmonic are merged with the R matrix; the fifth harmonic and the seventh harmonic are merged with the G channel pixel matrix; the ninth harmonic and the eleventh harmonic are merged with the B channel pixel matrix. The generated three-channel fused pixel matrix is superimposed to obtain the color mixed image.

5. The non-intrusive load anomaly identification method of the power utilization system according to claim 1, wherein, The convolutional neural network comprises three convolutional layers, two pooling layers and one fully connected layer.

6. A non-intrusive load anomaly identification device of a power utilization system, characterized by, It comprises; A signal acquisition and processing unit for real-time acquisition of the electrical load signal on the load bus of the power utilization system and noise reduction processing; A switching event detection unit for detecting whether a switching event occurs based on the electrical load signal after noise reduction processing; A switching event detection unit for detecting whether a switching event occurs based on the electrical load signal after noise reduction processing; A V-I network processing unit is configured to periodically collect the load voltage and the load current after the load working state reaches stability if a switching event is detected, and to pre-process the load voltage and the load current to obtain the load voltage and the load current each characterized by a continuous grid of V-I trajectories; A V-I trajectory mapping unit is configured to map each V-I trajectory according to the reactive power, the power factor and the current sequence distribution value features into a preset RGB color image three-channel pixel matrix by using a grid mapping method based on the load voltage and the load current each characterized by a continuous grid of V-I trajectories; A color mixed image acquisition unit is configured to extract each high-order harmonic amplitude of the load current, to perform binary conversion according to the pixel matrix size, and to further fuse the RGB color image three-channel pixel matrix after mapping to form a color mixed image; A load abnormality identification unit is configured to input the color mixed image into a pre-trained convolutional neural network to identify the load abnormality. The step of mapping each V-I trajectory according to the reactive power, the power factor and the current sequence distribution value features into a preset RGB color image three-channel pixel matrix specifically includes: The first step, constructing three full 1 matrices with 3 dimensions of N*N R, G, and B matrices, respectively, and calculating the mapping coordinates of each sample point after interpolation according to formula (5); （5）； In the second step, the coordinates of the corresponding sample points in the R, G and B matrices are respectively assigned values according to the formulas (6), (7) and (8). wherein, is the number of occurrences for each coordinate point; is the reactive current; is the voltage of the sample point; is the active power; , are the current and voltage effective values of the collected data, respectively; represents the distribution of the current sequence; M is the number of collected data in the stable period.

7. The non-intrusive load anomaly identification device of the power utilization system according to claim 6, wherein, The convolutional neural network includes three convolutional layers, two pooling layers and one fully connected layer.

Images