Non-intrusive load identification method and system based on feature optimization and GOA-SOM

Abstract

The invention discloses a non-intrusive load identification method and system based on feature optimization and GOA-SOM (Generic Object Architecture-Self-Organizing Map), and solves the problems of large workload and low identification accuracy caused by large difference between load and feature parameters in the existing non-intrusive load monitoring. The method comprises the steps of collecting electricity consumption information in real time, extracting typical features according to a load switching event, performing dimensionality reduction screening on the typical features by using a Fisher algorithm, optimizing an SOM by using a GOA algorithm, establishing an identification model, inputting the screened typical features into the identification model for iterative identification, and obtaining a judged load type. According to the method, standard characteristic data can be provided for load identification of the detection terminal under the condition that users are not interfered, so that the accuracy of load identification is ensured, the controllable load detection requirements of residents are met, and the method has applicability to load data in different users.

Description

Technical Field

[0001] The present invention relates to the technical field of load identification, and in particular to a non-intrusive load identification method and system based on feature optimization and GOA-SOM. Background Art

[0002] In order to balance the power fluctuations of the new power system with a high proportion of renewable energy, it is necessary to give full play to the regulatory role of demand-side resources. The consumption of clean energy can be promoted through reasonable load scheduling and coordinated optimization of resources on both the supply and demand sides. Demand-side power management is an important part of ensuring that the demand side has the ability to adjust flexibly. Power load power monitoring is the basis of power management. Users can also reduce their unnecessary energy expenses by analyzing data.

[0003] Traditional load monitoring is usually based on an invasive approach, installing different types of data collection devices on electrical equipment, increasing investment, limiting the scalability of the monitoring system, and causing poor user satisfaction. Non-invasive power load decomposition and identification, on the other hand, is to install non-invasive power monitoring devices at the user's power entrance to collect overall data such as current and voltage, without interfering with the user's normal power consumption.

[0004] Non-intrusive load monitoring has received extensive attention and research from scholars. The key to non-intrusive load monitoring technology is to extract effective, independent, and highly distinguishable load features from load electrical information. The accuracy of the priori correspondence between the load and its electrical information determines the accuracy of the signal identification algorithm. However, due to the fact that there are many brands and models of user loads, and the characteristic parameters are also different, the workload of pre-registering each load of the user is huge. Therefore, it is necessary to use a unified load feature applicable to the same type of load to identify the load, which requires correctly finding the load feature that can efficiently distinguish each load category. In an independent user, when the working state of the electrical appliance changes, a series of load characteristics such as the current, active power, and reactive power of the electrical equipment change in a fixed manner. Therefore, it is feasible to first extract the characteristics of the load event from the voltage and current signals collected at the user's power entrance, and then use the corresponding relationship between the characteristics and the load to identify the user's load. . Summary of the invention

[0005] The present invention mainly solves the problem that the existing intrusive load monitoring has limited system scalability and interferes with the user's electricity consumption, and the existing non-intrusive load monitoring has a large workload and low identification accuracy due to the large difference in load and characteristic parameters. It provides a non-intrusive load identification method and system based on feature optimization and GOA-SOM. The present invention uses the Fisher algorithm to screen the extracted typical residential load characteristics, reduce the redundancy of the load characteristics, and convert the waveform data into identifiable structured data. The present invention proposes a load identification algorithm based on the locust optimization algorithm to improve the self-organizing neural network under non-intrusive load monitoring. Considering the differences in physical characteristics of different load devices, the general load characteristics of electrical equipment are extracted, and an improved self-organizing neural network system with low time complexity and high identification accuracy is constructed to complete the determination of the load type and realize the effective identification of the load type of residential users.

[0006] The above technical problem of the present invention is mainly solved by the following technical solution: a non-intrusive load identification method based on feature optimization and GOA-SOM, comprising the following steps:

[0007] S1. Real-time collection of electricity consumption information;

[0008] S2. Extract typical features based on load shedding events;

[0009] S3. Use Fisher algorithm to reduce the dimension of typical features;

[0010] S4. Use GOA algorithm to optimize SOM and establish recognition model;

[0011] S5. Input the selected typical features into the identification model for iterative identification to obtain the determined load type.

[0012] The present invention screens the extracted typical characteristics of the resident load through the Fisher algorithm, and on this basis constructs a self-organizing neural network model (SOM) improved by the GOA algorithm (locust optimization algorithm), i.e., an identification model, and identifies the type of resident load through the identification model, thereby achieving high-precision identification of loads of different users. The present invention can provide standard characteristic data for load identification of the detection terminal without interfering with the user, thereby ensuring the accuracy of load identification. Based on the steady-state characteristics and transient characteristics of resident users in non-invasive load monitoring, the present invention establishes a complete processing method for non-invasive load identification, which meets the requirements for controllable load detection of residents and is applicable to load data of different users.

[0013] As a preferred solution, the collected electrical information includes current signal, voltage signal, power value and power factor.

[0014] In this solution, a high-frequency acquisition device is used, which is set at the user bus entrance, and the sampling frequency K is set to obtain the power consumption information at the user bus entrance. As a preferred solution, the frequency is set to K = 15Hz to obtain the user's real-time power information, including current signal, voltage signal, power value and power factor.

[0015] As a preferred solution, the waveform caused by the load switching event is separated and extracted from the acquired power consumption information, and typical features are calculated. The typical features include steady-state features and transient features.

[0016] This solution detects load switching events by detecting whether the average power entering the user's home has a sudden change. The waveform caused by the sudden change caused by the switching event is separated, the signal of the waveform is analyzed, and the typical characteristics are calculated. Typical characteristics include steady-state characteristics and transient characteristics. Steady-state characteristics include active / reactive power, active / reactive harmonics, current-voltage trajectory, fundamental component of voltage / current and its harmonic components, etc.; transient characteristics include instantaneous current / power, voltage noise, etc.

[0017] As a preferred solution, calculating the steady-state characteristics includes:

[0018] Calculate active and reactive power based on current and voltage;

[0019] Calculate the current assignment, effective value, and crest factor value;

[0020] Calculate the DC and harmonic components of current harmonics.

[0021] Specifically, the active power is calculated based on the product of the steady-state voltage effective value, the steady-state current effective value, and the cosine value of the voltage and current phase difference. The reactive power is calculated based on the product of the steady-state voltage effective value, the steady-state current effective value, and the sine value of the voltage and current phase difference.

[0022] The active and reactive power are calculated as follows:

[0023]

[0024] In the formula, V and I are the effective value of steady-state voltage and current respectively, k is the harmonic order, φ k is the phase difference between voltage and current.

[0025] Select the maximum value of the current value at the sampling point as the current amplitude; average the square sum of the current values ​​at the sampling points and take the square root to obtain the effective value of the steady-state current; calculate the crest factor based on the ratio of the current amplitude to the effective value of the steady-state current. The specific current amplitude, effective value, and crest factor values ​​are calculated as follows:

[0026]

[0027] Where IP ,I rms ,I CF They represent the current amplitude, current effective value and crest factor respectively. N is the number of sampling points. i(k) is the current value at the kth sampling point.

[0028] Calculate the frequency domain of current harmonic characteristics I(ω):

[0029]

[0030] Where ω is the current frequency, t is the current value at time t, and e is a constant.

[0031] As a preferred solution, calculating the transient characteristics includes: calculating the current pulse peak value and the concave-convex coefficient.

[0032] Specifically, the maximum value of the current value at the sampling point is selected to calculate the current pulse peak value; and the concave-convex coefficient is calculated according to the ratio of the current sum at the sampling point of the transient process to the current sum at the sampling point of the steady-state process.

[0033] The peak current pulse is calculated as follows: I P =max[i(k)]

[0034] Calculate the bump coefficient:

[0035] C=∑I TK / ∑I SK

[0036] In the formula, I TK is the current at the kth sampling point in the transient process, I SK is the current at the kth sampling point in the steady-state process.

[0037] As a preferred solution, the specific process of step S3 includes:

[0038] The load samples are classified in pairs. Each load sample includes typical features. The score of each typical feature is calculated according to the Fisher algorithm. The typical features are arranged in descending order according to the score, and the first N typical features are selected as the optimal solution of feature combination.

[0039] This solution uses the original load samples and their labels to select features through the Fisher algorithm, reducing the feature dimension to improve the efficiency of load feature identification. The specific steps include:

[0040] For the sample set X = {x 1 , x 2 , ..., x n} belongs to ω 1 and ω 2 There are two categories, each sample is represented by k features, and

[0041] The within-class covariance is calculated as:

[0042]

[0043] In the formula, m i is the category mean.

[0044] Total intra-class divergence S inc Expressed as: S inc =S 1 +S 2 .

[0045] Inter-class dispersion S exc Expressed as: S exc =(m 1 -m 2 )(m 1 -m 2 ) T

[0046] The samples are divided into j =ω T x j , j = 1, 2, ..., n for projection processing, projecting to the ω direction to obtain the new sample means of each category: Where ω represents the optimal projection direction.

[0047] The new intra-class divergence is:

[0048] The total intra-class divergence is: S inc-inc =S 11 +S 22 ;

[0049] The inter-class divergence is: S exc-exc =(m 11 -m 22 ) 2 ;

[0050] The objective function is constructed as:

[0051] Calculate the best projection direction:

[0052] Selecting the threshold For any k-dimensional sample X to be tested, the linear discriminant function is:

[0053] g(X)=ω T X+ω 0

[0054] Obtain the distance between g(X) and the new sample within ω, and classify the samples to be tested into categories with close distances.

[0055] Fisher score is used to score various load characteristics. The calculation formula of Fisher score Fs is:

[0056]

[0057] The load characteristics are arranged in descending order according to the scores, and the first N typical characteristics are selected as the optimal solution of feature combination to be input into the identification model established subsequently.

[0058] As a preferred solution, the GOA algorithm is used to optimize the input and output layer mapping weights of the SOM network.

[0059] As a preferred solution, the mapping weights of the input and output layers of the SOM network are optimized to establish a recognition model. The specific process includes:

[0060] S41. Initialize parameters, set the maximum number of iterations and the objective function, the number of independent variables and the number of locusts;

[0061] S42. define a fitness function, take the optimal solution of the locust individual as the fitness value, compare it with the historical global optimal solution, and iterate and update it;

[0062] S43. Update the locust position and speed, build an identification model, and solve the fitness value of each locust to determine whether the iterative update is completed. If the iterative update is completed, the global best fitness value is output as the global optimal solution;

[0063] The updated formula for locust speed and position is:

[0064] V id =βV id +C 1 random(0,1)(P id -X id )

[0065] +C 2 random(0,1)(P gd -X gd )

[0066] X id =X (i-1)d +V id .

[0067] S44. Take the global optimal solution as the weight W ij ;

[0068] S45. Randomly select a load characteristic sample X i ;

[0069] S46. traverse each node in the competition layer, calculate the similarity between the load feature sample and the node, and select the node with the smallest distance as the winning node;

[0070] The similarity is calculated as:

[0071]

[0072] S47. Determine the nodes to be included in the winning neighborhood according to the neighborhood radius σ, and calculate the magnitude of each update through the nearest neighbor function;

[0073]

[0074] Where: f(x) is the Gaussian approximation function, a, b, c are the Gaussian approximation function model coefficients.

[0075] S48. Update the weight W of the node in the winning neighborhood according to the updated amplitude and the distance of the load characteristic sample ij ;

[0076] W_v(s+1)=W_v(s)+γ(u,v,s)[D(t)-W_v(s)]

[0077] Where W_v(s) is the current weight of node v, γ(u, v, s) is the update constraint, and D(t) is the sample distance at time t.

[0078] S49. Return to step S45 until the number of iterations is reached to obtain the optimized weight.

[0079] As a preferred solution, the identification model is constructed as follows:

[0080] V id =βV id +C 1 random(0,1)(P id -X id )

[0081] +C 2 random(0,1)(P gd -X gd )

[0082] X id =X (i-1 ) d +V id

[0083] Among them, β is the inertia factor, C 1 and C 2 is the acceleration constant, random(0,1) is a random number in the interval [0,1], P id is the d-dimensional individual extreme value of the i-th weight, Pgd is the dth dimension of the global optimal solution, X id is the location parameter of the search space, V id For speed.

[0084] A non-intrusive load identification system based on feature optimization and GOA-SOM comprises a monitoring unit and an identification unit. The monitoring unit collects user power consumption information in real time, and the identification unit adopts a non-intrusive load identification method based on feature optimization and GOA-SOM.

[0085] The monitoring unit is a non-intrusive power monitoring device that collects current and voltage as a whole and does not interfere with the normal power consumption of users. The identification unit extracts typical features from the collected power consumption information according to load switching events, and uses the Fisher algorithm to reduce the dimension of the typical features; the GOA algorithm is used to optimize the SOM, establish an identification model, input the selected typical features into the identification model, perform iterative identification, and obtain the determined load type.

[0086] Therefore, the advantages of the present invention are: the typical characteristics of the extracted resident loads are screened by the Fisher algorithm, and on this basis, a self-organizing neural network model (SOM) improved by the GOA algorithm (locust optimization algorithm), i.e., an identification model, is constructed, and the resident load type is identified by the identification model, thereby achieving high-precision identification of loads of different users. The present invention can provide standard feature data for load identification of the detection terminal without interfering with the user, thereby ensuring the accuracy of load identification. The present invention establishes a complete processing method for non-invasive load identification based on the steady-state characteristics and transient characteristics of resident users in non-invasive load monitoring, which meets the requirements for detection of controllable loads of residents and is applicable to load data of different users. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 It is a schematic flow chart of the method of the present invention;

[0088] Figure 2 It is a graph of the iteration result of the GOA-SOM competition layer in the present invention;

[0089] Figure 3 This is a comparison chart of the recognition accuracy of the method of the present invention and other methods;

[0090] Figure 4 This is a comparison chart of the computational efficiency of the method of the present invention and other methods. DETAILED DESCRIPTION

[0091] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0092] Example:

[0093] This embodiment is a non-intrusive load identification method based on feature optimization and GOA-SOM. Figure 1 As shown, the following steps are included:

[0094] S1. Real-time collection of electricity consumption information;

[0095] S2. Extract typical features based on load shedding events;

[0096] S3. Use Fisher algorithm to reduce the dimension of typical features;

[0097] S4. Use GOA algorithm to optimize SOM and establish recognition model;

[0098] S5. Input the selected typical features into the identification model for iterative identification to obtain the determined load type.

[0099] The method of the present invention screens the extracted typical characteristics of the resident load through the Fisher algorithm, and on this basis constructs a self-organizing neural network model (SOM) improved by the GOA algorithm (locust optimization algorithm), that is, an identification model, and identifies the type of resident load through the identification model, thereby achieving high-precision identification of loads of different users. The present invention can provide standard feature data for load identification of the detection terminal without interfering with the user, thereby ensuring the accuracy of load identification. Based on the steady-state characteristics and transient characteristics of resident users in non-invasive load monitoring, the present invention establishes a complete processing method for non-invasive load identification, which meets the requirements for detection of controllable loads of residents and is applicable to load data of different users.

[0100] As a preferred solution of this embodiment, in step S1, the power consumption information is collected in real time, specifically by adopting a non-intrusive high-frequency collection device for sampling, by setting a high-frequency sampling device at the user bus entrance, setting the sampling frequency to K, and obtaining the user's real-time power consumption information. In this embodiment, the sampling frequency is 15Hz, and the collected power consumption information includes current signal, voltage signal, power value and power factor.

[0101] Step S2 extracts typical features according to the load switching event. As a preferred solution of this embodiment, the waveform caused by the load switching event is separated and extracted from the acquired power consumption information, and the typical features are calculated. This embodiment detects the load switching event by detecting whether the average power of the user entering the house changes suddenly, separates the waveform according to the sudden change caused by the ventilation event, and analyzes the power consumption information of the waveform to calculate the typical features. Typical features include steady-state features and transient features. Steady-state features include active / reactive power, active / reactive harmonics, current-voltage trajectory, fundamental component of voltage / current and its harmonic components, etc.; transient features include instantaneous current / power, voltage noise, etc. The calculation of steady-state features in this embodiment includes calculating active and reactive power based on current and voltage; calculating current assignment, effective value, and peak factor value; and calculating current harmonic DC component and harmonic component. The specific calculation method is as follows:

[0102] The active power is calculated by the product of the steady-state voltage RMS value, the steady-state current RMS value, and the cosine value of the voltage-current phase difference. The reactive power is calculated by the product of the steady-state voltage RMS value, the steady-state current RMS value, and the sine value of the voltage-current phase difference. The specific formula is as follows:

[0103]

[0104] In the formula, V and I are the effective value of steady-state voltage and current respectively, k is the harmonic order, φ k is the phase difference between voltage and current.

[0105] The maximum value of the current value at the sampling point is selected as the current amplitude; the square root of the average of the current values ​​at the sampling points is taken to obtain the effective value of the steady-state current; the crest factor is calculated based on the ratio of the current amplitude to the effective value of the steady-state current. The specific calculation formula is as follows:

[0106]

[0107] Where I P ,I rms ,I CF They represent the current amplitude, current effective value and crest factor respectively. N is the number of sampling points. i(k) is the current value at the kth sampling point.

[0108] Calculate the frequency domain of the current harmonic characteristic I(ω), and the calculation formula is as follows:

[0109]

[0110] Where ω is the current frequency, t is the current value at time t, and e is a constant.

[0111] The calculation of transient characteristics in this embodiment includes calculating the current pulse peak value and the concave-convex coefficient. The specific calculation method is as follows:

[0112] Select the maximum current value of the sampling point to calculate the current pulse peak value. The calculation formula is as follows:

[0113] I P =max[i(k)]

[0114] I P is the peak value of the current pulse.

[0115] Calculate the concave-convex coefficient, the formula is as follows:

[0116] C=∑I TK / ∑I SK

[0117] In the formula, I TK is the current at the kth sampling point in the transient process, I SK is the current at the kth sampling point in the steady-state process.

[0118] Step S3 uses the Fisher algorithm to reduce the dimension of typical features. The process includes:

[0119] The load samples are classified in pairs. Each load sample includes typical features. The scores of each typical feature are calculated according to the Fisher algorithm. The typical features are arranged in descending order according to the scores, and the first N typical features are selected as the optimal solution of feature combination. This step is to use the original load sample and its label to select features through the Fisher algorithm, reduce the feature dimension and improve the efficiency of load feature identification.

[0120] As a preferred solution of this embodiment, the specific process includes:

[0121] For the sample set X = {x 1 , x 2 , ..., x n} belongs to ω 1 and ω 2 There are two categories, each sample is represented by k features, and

[0122] The within-class covariance is calculated as:

[0123]

[0124] In the formula, m i is the category mean.

[0125] Total intra-class divergence S inc Expressed as: S inc =S 1 +S 2.

[0126] Inter-class dispersion S exc Expressed as: S exc =(m 1 -m 2 )(m 1 -m 2 ) T

[0127] The samples are divided into j =ω) T x j , j = 1, 2, ..., n for projection processing, projecting to the ω direction to obtain the new sample means of each category: Where ω represents the optimal projection direction.

[0128] The new intra-class divergence is:

[0129] The total intra-class divergence is: S inc-inc =S 11 +S 22 ;

[0130] The inter-class divergence is: S exc-exc =(m 11 -m 22 ) 2 ;

[0131] The objective function is constructed as:

[0132] Calculate the best projection direction:

[0133] Selecting the threshold For any k-dimensional sample X to be tested, the linear discriminant function is:

[0134] g(X)=ω T X+ω 0

[0135] Obtain the distance between g(X) and the new sample within ω, and classify the samples to be tested into categories with close distances.

[0136] Fisher score is used to score various load characteristics. The calculation formula of Fisher score Fs is:

[0137]

[0138] The load characteristics are arranged in descending order according to the scores, and the first N typical characteristics are selected as the optimal solution of feature combination to be input into the identification model established subsequently.

[0139] Step S4 uses the GOA algorithm to optimize the SOM and establish a recognition model. This step mainly uses the GOA algorithm to optimize the input and output layer mapping weights of the SOM network to build an improved self-organizing neural network system with low time complexity and high recognition accuracy. The specific process includes:

[0140] S41. Initialize parameters, set the maximum number of iterations and the objective function, the number of independent variables and the number of locusts;

[0141] S42. define a fitness function, take the optimal solution of the locust individual as the fitness value, compare it with the historical global optimal solution, and iterate and update it;

[0142] S43. Update the locust position and speed, build an identification model, and solve the fitness value of each locust to determine whether the iterative update is completed. If the iterative update is completed, the global best fitness value is output as the global optimal solution;

[0143] The updated formula for locust speed and position is:

[0144] V id =βV id +C 1 random(0,1)(P id -X id )

[0145] +C 2 random(0,1)(P gd -X gd )

[0146] X id =X (i-1)d +V id .

[0147] Among them, β is the inertia factor, C 1 and C 2 is the acceleration constant, random(0,1) is a random number in the interval [0,1], P id is the d-dimensional individual extreme value of the i-th weight, P gd is the dth dimension of the global optimal solution, X id is the location parameter of the search space, V id For speed.

[0148] S44. Take the global optimal solution as the weight W ij ;

[0149] S45. Randomly select a load characteristic sample X i ;

[0150] S46. traverse each node in the competition layer, calculate the similarity between the load feature sample and the node, and select the node with the smallest distance as the winning node;

[0151] The similarity is calculated as:

[0152]

[0153] S47. Determine the nodes to be included in the winning neighborhood according to the neighborhood radius σ, and calculate the magnitude of each update through the nearest neighbor function;

[0154]

[0155] Where: f(x) is the Gaussian approximation function, a, b, c are the Gaussian approximation function model coefficients.

[0156] S48. Update the weight W of the node in the winning neighborhood according to the updated amplitude and the distance of the load characteristic sample ij ;

[0157] W_v(s+1)=W_v(s)+γ(u,v,s)[D(t)-W_v(s)]

[0158] Where W_v(s) is the current weight of node v, γ(u, v, s) is the update constraint, and D(t) is the sample distance at time t.

[0159] S49. Return to step S45 until the number of iterations is reached to obtain the optimized weight.

[0160] Step S5 inputs the selected typical features into the identification model for iterative identification to determine whether the classification identification is completed. If not, iterative identification continues until the classification identification is completed, and the determined load type is output.

[0161] The present embodiment scheme is further described below with reference to actual examples, and a standard data set is used for demonstration. The power decomposition reference data set includes energy consumption data of 6 different households in a few weeks, as well as high-frequency current / voltage data of the main energy-consuming equipment of two of the households, providing 15Hz high-frequency data and second-level low-frequency data. The steady-state and transient process data of 8 types of household loads of one of the households: lighting, electric fans, computers, televisions, heaters, microwave ovens, air conditioners, and washing machines are selected. Seven typical features are selected, including the effective value of current, crest factor, DC component, third harmonic, and fifth harmonic in the steady-state characteristics, and the current pulse peak value and concave-convex coefficient in the transient characteristics. The load switching event is detected by detecting whether the average power of the user entering the home has a sudden change, and the waveform caused by the sudden change caused by the switching event is separated. The power consumption information of the waveform is analyzed, and the above-mentioned 7 typical features are used to calculate the typical characteristic value according to the power consumption information.

[0162] The current effective value and crest factor are calculated as follows:

[0163]

[0164] Where I P ,I rms ,I CF They represent the current amplitude, current effective value and crest factor respectively. N is the number of sampling points. i(k) is the current value at the kth sampling point.

[0165] The frequency domain calculation of the current harmonic characteristic I(ω) is as follows:

[0166]

[0167] Where ω is the current frequency, t is the current value at time t, and e is a constant.

[0168] The current pulse peak value is calculated as follows: I P =max[i(k)]

[0169] The concave-convex coefficient is calculated as follows:

[0170] C=∑I TK / ∑I SK

[0171] In the formula, I TK is the current at the kth sampling point in the transient process, I SK is the current at the kth sampling point in the steady-state process.

[0172] The calculation results of some typical eigenvalues ​​according to the above calculation method are shown in Table 1.

[0173]

[0174] Table 1 Some typical characteristic values ​​of typical loads

[0175] The Fisher algorithm is used to reduce the dimension of typical features. Based on the above 8 types of loads, they are classified in pairs. Each load sample includes typical features. The Fisher-score of different feature combinations in each classification is calculated according to the Fisher algorithm. The Fisher-score of 10 original input variables, including current effective value, active power, reactive power, crest factor, concave-convex factor, current third harmonic, current fifth harmonic, DC component, current steady-state amplitude, and current pulse peak, is ranked from high to low. The scores of the 10 typical load features are shown in Table 2.

[0176] Steady-state characteristics Fisher-score Steady-state characteristics Fisher-score Crest Factor 0.26 DC component 0.07 Reactive power 0.21 Current RMS 0.06 Active Power 0.15 Current steady-state amplitude 0.03 Current third harmonic 0.13 Current pulse peak 0.01 Current fifth harmonic 0.07 Concave and convex coefficient 0.01

[0177] Table 2 Scoring of 10 typical load characteristics

[0178] Among them, the typical features with lower Fisher scores have poor separability. In order to reduce calculation, the scores of each typical feature in the optimal combination are comprehensively considered, and finally the typical features with higher scores are selected. For example, the top 4 typical features in Table 2 use crest factor, reactive power, active power, and third harmonic of current as the optimal solution of feature combination.

[0179] The GOA algorithm is further used to optimize the input and output layer mapping weights of the SOM network and construct an identification model, including:

[0180] S41. Initialize parameters, set the maximum number of iterations and the objective function, the number of independent variables and the number of locusts;

[0181] S42. define a fitness function, take the optimal solution of the locust individual as the fitness value, compare it with the historical global optimal solution, and iterate and update it;

[0182] S43. Update the locust position and speed, build an identification model, and solve the fitness value of each locust to determine whether the iterative update is completed. If the iterative update is completed, the global best fitness value is output as the global optimal solution;

[0183] The updated formula for locust speed and position is:

[0184] V id =βV id +C 1 random(0,1)(P id -X id )

[0185] +C 2 random(0,1)(P gd -X gd )

[0186] X id =X (i-1)d +V id .

[0187] S44. Take the global optimal solution as the weight W ij ;

[0188] S45. Randomly select a load characteristic sample X i ;

[0189] S46. traverse each node in the competition layer, calculate the similarity between the load feature sample and the node, and select the node with the smallest distance as the winning node;

[0190] The similarity is calculated as:

[0191]

[0192] S47. Determine the nodes to be included in the winning neighborhood according to the neighborhood radius σ, and calculate the magnitude of each update through the nearest neighbor function;

[0193]

[0194] Where: f(x) is the Gaussian approximation function, a, b, c are the Gaussian approximation function model coefficients.

[0195] S48. Update the weight W of the node in the winning neighborhood according to the updated amplitude and the distance of the load characteristic sample ij ;

[0196] W_v(s+1)=W_v(s)+γ(u,v,s)[D(t)-W_v(s)]

[0197] Where W_v(s) is the current weight of node v, γ(u, v, s) is the update constraint, and D(t) is the sample distance at time t.

[0198] S49. Return to step S45 until the number of iterations is reached to obtain the optimized weight.

[0199] The identification model is established by inputting the peak factor, reactive power, active power and third harmonic of current, and the determined load type is output.

[0200] The GOA-SOM neural network is used to classify the sample data, and the network topology size is adaptively adjusted according to the number of identified load types to further enhance the classification effect and recognition efficiency of the network. The iterative results of the neurons in the competition layer of the GOA-SOM neural network are shown in Figure 2. Figure 2 As shown in Figure 2, it is found through statistical observation that the five types of loads have been completely separated. Figure 3 and Figure 4 As shown in FIG. , a comparison curve of the load identification accuracy and operation efficiency of the method of the present invention and other related identification methods is shown. Figure 3 The recognition accuracy curve of the algorithm is given as the load type increases. When the load gravity increases, the algorithm of the present invention is less affected and has a higher accuracy. Figure 4 For the calculation efficiency curve, the load identification time of the method of the present invention does not increase with the increase of the number of loads. The method of the present invention can accurately identify the operating status of the resident load, thereby obtaining its distribution law, laying a foundation for the scheduling plan to consider the load response effect and the effective control of the resident controllable load.

[0201] This embodiment also discloses a non-intrusive load identification system based on feature optimization and GOA-SOM, including a monitoring unit and an identification unit.

[0202] Monitoring unit, collects user electricity usage information in real time;

[0203] The monitoring unit is a non-intrusive power monitoring device and a high-frequency acquisition device. It is installed at the user bus entrance, sets the sampling frequency, and collects the current and voltage at the user bus entrance as a whole, which will not interfere with the user's normal power consumption. It obtains the user's real-time power information, including current signal, voltage signal, power value and power factor.

[0204] The identification unit is connected to the detection unit, receives the power consumption information collected by the detection unit, extracts typical features of the collected power consumption information according to the load switching event, and uses the Fisher algorithm to reduce the dimension of the typical features; the GOA algorithm is used to optimize the SOM, and an identification model is established. The screened typical features are input into the identification model for iterative identification to obtain the determined load type.

[0205] The method of detecting whether the average power of users entering the house has a sudden change is used to detect the load switching event. The waveform caused by the sudden change caused by the switching event is separated, the signal of the waveform is analyzed, and the typical characteristics are calculated. Typical characteristics include steady-state characteristics and transient characteristics. Steady-state characteristics include active / reactive power, active / reactive harmonics, current-voltage trajectory, fundamental component of voltage / current and its harmonic components, etc.; transient characteristics include instantaneous current / power, voltage noise, etc. The load samples are classified in pairs, and the scores of each typical feature are calculated according to the Fisher algorithm. The typical features are arranged in descending order according to the scores, and the first N typical features are selected as the optimal solution of the feature combination to reduce the feature dimension to improve the efficiency of load feature identification. The input and output layer mapping weights of the SOM network are optimized by the GOA algorithm to establish an identification model.

[0206] The present invention proposes a load identification algorithm based on a self-organizing neural network improved by a locust optimization algorithm under non-invasive load monitoring. Taking into account the differences in physical characteristics of different load devices, the universal load characteristics of electrical equipment are extracted, and an improved self-organizing neural network system with low time complexity and high identification accuracy is constructed to complete the determination of the load type and realize the effective identification of the load type of residential users. The present invention can provide standard feature data for load identification of the detection terminal without interfering with the user, thereby ensuring the accuracy of load identification. Based on the steady-state characteristics and transient characteristics of residential users in non-invasive load monitoring, the present invention establishes a complete processing method for non-invasive load identification, which meets the requirements for controllable load detection of residents and is applicable to load data of different users.

[0207] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0208] Although this article uses more terms such as typical characteristics, identification model, monitoring unit, identification unit, etc., it does not exclude the possibility of using other terms. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional restrictions is contrary to the spirit of the present invention.

Claims

1. A non-intrusive load identification method based on feature optimization and GOA-SOM, characterized in that: The following steps are involved: S1. Real-time collection of electricity consumption information; S2. Extract typical features based on load shedding events; S3. Use Fisher algorithm to reduce the dimension of typical features; S4. Use GOA algorithm to optimize SOM and establish recognition model; S5. Input the selected typical features into the identification model for iterative identification to obtain the determined load type.

2. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 1, characterized in that The collected electrical information includes current signal, voltage signal, power value and power factor.

3. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 1 is characterized in that The waveform caused by the load switching event is separated and extracted from the acquired power consumption information, and typical features are calculated. The typical features include steady-state features and transient features.

4. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 3 is characterized by: Computational steady-state features include: Calculate active and reactive power based on current and voltage; Calculate the current assignment, effective value, and crest factor value; Calculate the DC and harmonic components of current harmonics.

5. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 3 is characterized by: Calculating transient characteristics includes calculating current pulse peak value and concave-convex coefficient.

6. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 1, characterized in that The specific process of step S3 includes: The load samples are classified in pairs. Each load sample includes typical features. The score of each typical feature is calculated according to the Fisher algorithm. The typical features are arranged in descending order according to the score, and the first N typical features are selected as the optimal solution of feature combination.

7. A non-intrusive load identification method based on feature optimization and GOA-SOM according to any one of claims 1 to 6, characterized in that The GOA algorithm is used to optimize the input and output layer mapping weights of the SOM network.

8. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 7 is characterized in that The SOM network input and output layer mapping weights are optimized to establish a recognition model. The specific process includes: S41. Initialize parameters, set the maximum number of iterations and the objective function, the number of independent variables and the number of locusts; S42. define a fitness function, take the optimal solution of the locust individual as the fitness value, compare it with the historical global optimal solution, and iterate and update it; S43. Update the locust position and speed, build an identification model, and solve the fitness value of each locust to determine whether the iterative update is completed. If the iterative update is completed, the global best fitness value is output as the global optimal solution; S44. Taking the global optimal solution as a weight; S45. Randomly select a load characteristic sample; S46. traverse each node in the competition layer, calculate the similarity between the load feature sample and the node, and select the node with the smallest distance as the winning node; S47. Determine the nodes to be included in the winning neighborhood according to the neighborhood radius, and calculate the magnitude of each update through the nearest neighbor function; S48. Update the weights of the nodes in the winning neighborhood according to the updated amplitude and the distance of the load characteristic sample; S49. Return to step S45 until the number of iterations is reached to obtain the optimized weight.

9. The non-intrusive load identification method based on feature optimization and GOA-SOM according to claim 8, characterized in that The identification model is constructed as follows: V id =βV id +C1random(0,1)(P id -X id )+C2random(0,1)(P gd -X gd ) X id =X (i-1)d +V id Among them, β is the inertia factor, C1 and C2 are acceleration constants, random(0,1) is a random number in the interval [0,1], P id is the d-dimensional individual extreme value of the i-th weight, P gd is the dth dimension of the global optimal solution, X id is the location parameter of the search space, V id For speed.

10. A non-intrusive load identification system based on feature optimization and GOA-SOM, characterized in that: It comprises a monitoring unit and an identification unit, the monitoring unit collects user electricity consumption information in real time, and the identification unit adopts the method described in any one of claims 1-9.

Images