A method and system for non-intrusive load identification of homogenized industrial loads

Abstract

The application discloses a homogenized industrial load non-intrusive load identification method and system, and the method comprises the following steps: performing multi-dimensional segmented aggregation approximation processing on current data in high-frequency data of the homogenized industrial load; constructing a distance similarity matrix by using the multi-dimensional segmented aggregation approximation processed data; converting the distance similarity matrix into a weighted recursive matrix through normalization and threshold control; constructing a discretized V-I trajectory based on high-frequency current and voltage data in the high-frequency data of the industrial load directly obtained; performing weighted mapping on the discretized industrial load V-I trajectory based on the weighted recursive matrix; training a load classifier by using different categories of homogenized industrial load data to obtain the mapped recursive V-I trajectory feature map; and classifying the industrial load to be classified by using the load classifier. The application can effectively identify the industrial load with high homogenization, and the accuracy is high.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of load monitoring and analysis in industrial power systems, and relates to a method and system for non-intrusive load identification of homogeneous industrial loads. BACKGROUND

[0002] With the continuous development of digital power grid technology, energy visualization, reduction of comprehensive energy consumption and peak load shifting are urgent needs for the construction of new power systems, and online monitoring of power load is the key to realization. The online monitoring methods of power load mainly include invasive load monitoring (ILM) and non-invasive load monitoring (NILM). Non-invasive load identification has a simple hardware structure, is easy to install and maintain, and is more in line with the actual needs of smart grid construction.

[0003] Current non-invasive load identification methods mainly focus on the residential side, and industrial loads are less researched due to difficulties in data collection, but the stability, controllability and high specific gravity of industrial loads are important supports for the construction of new power systems. Existing research mainly relies on low-frequency data combined with event synthesis to judge load type, which has the problems of long interval time between load events, loss of load information, difficulty in extracting effective event occurrence and end markers when load events occur intensively, and identification errors. Moreover, low-frequency data has time lag effect, which is difficult to cope with the fine scheduling of current complex power systems. Industrial loads have the characteristics of high homogeneity, and there are a large number of motor loads, so the discrimination degree of homogenized loads is small in research based on high-frequency data, which is prone to misjudgment.

[0004] In existing research, the feature expression based solely on voltage-current (V-I) trajectory cannot fully capture the time series characteristics of voltage and current, especially when dealing with loads with similar V-I trajectories, which shows great limitations. In addition, although some researches have adopted algorithms specifically used for time series dynamic feature extraction, such as weighted recurrent graph, Markov transition field and Gram angle field, and proposed load feature expression and identification methods that integrate these methods, these technologies may cover or partially lose the features of homogenized loads due to high information dimension in the feature information extraction process. SUMMARY

[0005] In order to solve the problems existing in the prior art, the present application provides a method for non-intrusive identification and classification of industrial loads based on high-frequency current and voltage data. On the basis of traditional V-I trajectory features, time series features are further introduced, and the method of high-dimensional space mapping is used to avoid the covering and loss problems caused by redundant feature information. Specifically, the method maps the features into an RGB color space to generate a new feature map, thereby realizing more accurate load identification, especially for accurate monitoring and classification of homogenized industrial loads in new power systems.

[0006] In order to achieve the above object, the technical scheme adopted by the present application is as follows: a homogenized industrial load non-intrusive load identification method, comprising the following steps:

[0007] The current data in the high-frequency data of the high-frequency centering processed industrial load is subjected to multi-dimensional segmented aggregation approximation processing;

[0008] The data subjected to the multi-dimensional segmented aggregation approximation processing is used to construct a distance similarity matrix , and the similarity of the load characteristics is quantified by calculating the distance between the data points;

[0009] Through normalization and threshold control, the distance similarity matrix is converted into a weighted recursive matrix , and the key information of the load characteristics is extracted;

[0010] Based on the high-frequency current and voltage data in the high-frequency data of the directly obtained industrial load, a discretized V-I trajectory is constructed;

[0011] Based on the weighted recursive matrix , the discretized industrial load V-I trajectory is subjected to weighted mapping to form a mapped recursive V-I trajectory feature map;

[0012] The mapped recursive V-I trajectory feature map of different categories of homogenized industrial load data is used for training to obtain a load classifier;

[0013] Based on the industrial load to be classified, the load classifier is used to classify the industrial load to be classified.

[0014] Further, the high-frequency data of the industrial load includes voltage type data and current type data, and the storage matrix is and ; the matrix and is subjected to centering processing, comprising:

[0015] The minimum value , and the maximum value , of the current and voltage signals are calculated;

[0016] The step size of the data is calculated: , wherein is , ;

[0017] According to the step size The current and voltage signals are adjusted to make their data distribution more uniform.

[0018] Further, the current data is processed by multi-dimensional segmented aggregation approximation,

[0019] The mean value of each segment signal is calculated as follows:

[0020]

[0021] wherein , is the segment index, is the length of the original signal, is the target embedding size.

[0022] Further, a distance similarity matrix is constructed, including constructing a distance measurement matrix between each time series point, wherein the matrix measurement adopts the Euclidean distance; for any two points, the distance is defined as follows:

[0023] wherein is the power parameter in distance calculation.

[0024] Further, by normalization and threshold control, the distance similarity matrix is converted into a weighted recursive matrix , and the key information of the load characteristics is extracted, including:

[0025] The normalized distance matrix is calculated as: ;

[0026] The distance value is limited within the threshold : ;

[0027] The distance similarity matrix is converted into a weighted recursive graph , and the selection of parameters and satisfies , wherein .

[0028] Further, based on the weighted recursive matrix , the discretized industrial load V-I trajectory is weighted mapped to form a mapped recursive V-I trajectory feature map, including: the main features are mapped from a low-dimensional space to a high-dimensional space by a high-dimensional mapping technology, the features are decomposed, the most different features in the high-dimensional space are grabbed for low-dimensional projection, the projection data is normalized, the difference features are visualized by color mapping, and an improved VI trajectory graph capable of distinguishing homogeneous industrial loads is obtained.

[0029] Further, by generating five kinds of load feature data sets, the feature data sets are proportionally divided into training sets, test sets and validation sets, each feature data set contains a plurality of feature vectors, and the feature vectors represent the feature information of different loads; based on the classic neural network residual network, the load classifier is trained through the training set, the test set and the validation set.

[0030] Based on the concept of the method, the application also provides a homogeneous industrial load non-intrusive load identification system, a data processing module, a similarity calculation module, a feature extraction module, a V-I trajectory acquisition module, a feature map construction module, a classifier training module and a classification module;

[0031] The data processing module is used for multi-dimensional segmented aggregate approximation processing on the current data in the high-frequency data of the industrial load after high-frequency centering processing ;

[0032] The similarity calculation module constructs a distance similarity matrix using the data processed by the multi-dimensional segmented aggregate approximation processing, and quantifies the similarity of the load features by calculating the distance between the data points;

[0033] The feature extraction module converts the distance similarity matrix into a weighted recursive matrix by normalization and threshold control, and extracts the key information of the load features;

[0034] The V-I trajectory acquisition module is used for constructing a discretized V-I trajectory based on the high-frequency current and voltage data in the high-frequency data of the industrial load directly acquired;

[0035] The feature map construction module is used for weighting and mapping the discretized industrial load V-I trajectory based on the weighted recursive matrix , to form a mapped recursive V-I trajectory feature map;

[0036] The classifier training module trains the load classifier using the mapped recursive V-I trajectory feature map of the homogeneous industrial load data of different categories;

[0037] The classification module classifies the industrial load to be classified using the load classifier based on the industrial load to be classified.

[0038] The application can also provide a computer device comprising a processor and a memory, the memory being used to store a computer executable program, the processor reading the computer executable program from the memory and executing, and the processor executing the computer executable program can realize the homogeneous industrial load non-intrusive load identification method.

[0039] Meanwhile, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the homogeneous industrial load non-intrusive load identification method.

[0040] Compared with the prior art, the present application has at least the following beneficial effects:

[0041] The present application can reveal the periodicity and complexity of industrial load time series data while maintaining intuitiveness, effectively identify industrial loads, has high accuracy, and enriches the application field and range of load monitoring. On the basis of traditional V-I trajectory features, time series feature extraction is further introduced, and the method of capturing main features through high-dimensional space mapping avoids the coverage and loss problems caused by redundant feature information. Specifically, through the multi-dimensional PAA processing technology, the present application can significantly reduce the data dimension while maintaining the main feature information, and improve the data processing efficiency. By constructing a weighted recursive matrix, the key feature information of the load is further extracted, so that the highly homogeneous industrial load can be more accurately distinguished. The present application reprojects the key information to a high-dimensional space to decompose its dimensional features, captures the maximum difference dimension of the features, and uses color mapping difference to construct an improved VI trajectory diagram to intuitively display the dynamic characteristics of the load. This visualization method helps to more intuitively observe and analyze the characteristics of the load, and assists users in effectively managing and monitoring the load in a complex power system. The VI trajectory diagram is mapped to the RGB color space to generate a new feature diagram, thereby realizing more accurate load identification, and is particularly suitable for accurately monitoring and classifying highly homogeneous industrial loads in a new power system. BRIEF DESCRIPTION OF DRAWINGS

[0042] The present application will become apparent and easy to understand from the following description of embodiments taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 is a homogeneous industrial load classification flowchart.

[0044] Figure 2 is a comparison diagram of original VI trajectories of five types of homogeneous industrial loads according to the present application.

[0045] Figure 3 is a comparison diagram of improved VI trajectories of the first type of industrial load.

[0046] Figure 4 is a comparison diagram of improved VI trajectories of the second type of industrial load.

[0047] Figure 5 is a comparison diagram of improved VI trajectories of the third type of industrial load.

[0048] Figure 6 Fourth type of industrial load improvement VI trajectory comparison chart.

[0049] Figure 7 Fifth type of industrial load improvement VI trajectory comparison chart. DETAILED DESCRIPTION

[0050] Embodiments of the present application are described in detail below, in which like or similar designations refer to like or similar elements throughout.

[0051] The present application provides a homogenized industrial load non-invasive load identification method, comprising the following steps:

[0052] Step 1, collect high-frequency data of industrial load from the system, store as a matrix and then perform centering processing.

[0053] In this step 1, high-frequency current data and voltage data are obtained from the industrial power system, and these data are first decomposed periodically and then stored in matrices and respectively, each row of the matrix represents a cycle, and each column corresponds to a different time point. The obtained voltage and current data are centered to ensure the balance of the data in the subsequent processing. The centering step is as follows:

[0054] (1) Calculate the minimum value , and maximum value , of the current and voltage signals.

[0055] (2) Calculate the step size : where is , .

[0056] (3) Adjust the current and voltage signals according to the step size to make the data distribution more uniform.

[0057] Step 2, perform multi-dimensional piecewise aggregate approximation (PAA) processing based on the high-frequency centered current data of the industrial load . In this step, the centered current signal is processed by PAA to compress the dimension of the signal and reduce the amount of data while retaining the main features of the load. Let the length of the original signal be , and the target embedding size be , then the mean value of each segment of the signal is calculated as follows:

[0058]

[0059] where, , is the segment index.

[0060] Step 3, constructing a distance similarity matrix using the PAA-processed data by calculating the distance between data points to quantify the similarity of load characteristics;

[0061] In this step, the PAA-processed current signal is used to calculate the distance between each pair of data points and construct a distance similarity matrix . For any two points, the distance is defined as follows:

[0062] wherein, is the power parameter in distance calculation, used to control the measurement method of distance. The final distance similarity matrix is obtained.

[0063] Step 4, converting the distance similarity matrix to a weighted recurrent matrix by normalization and threshold control to extract key information of load characteristics;

[0064] In this step, the distance similarity matrix is used to construct a weighted recurrent matrix by setting appropriate parameters and to improve the discrimination of load characteristics. The specific weighting steps are as follows:

[0065] First, calculate the normalized distance matrix : ;

[0066] Then limit the distance value within the threshold : .

[0067] Convert the distance similarity matrix to a weighted recurrent graph by the above method, wherein the selection of parameters and must satisfy , wherein, .

[0068] Step 5, constructing the corresponding V-I trajectory from the collected high-frequency current and voltage data of industrial load;

[0069] In this step, the collected high-frequency current and voltage data of industrial load are data that have not been processed by PAA, and the dimension is 1 x .

[0070] Step 6: Based on the weighted recursive matrix The discrete industrial load V-I trajectory is difference mapped to form a new recursive V-I trajectory feature map;

[0071] In this step, the discrete industrial load V-I trajectory map has a dimension of , wherein the V-I trajectory is composed of high-frequency voltage and current discrete points, the weighted recursive matrix has a dimension of m x m, and the difference mapping is performed by capturing the maximum difference feature in the high-dimensional space.

[0072] The mapping process is as follows:

[0073]

[0074] wherein the high-dimensional mapping has a mapping process:

[0075] (1) In order to facilitate calculation, the mapping function is defined as ;

[0076] (2) The centralized is subjected to eigenvalue decomposition ;

[0077] (3) The principal eigenvalue in the eigenvalue space is calculated , wherein ;

[0078] (4) The projection of the final low-dimensional space is obtained , and the low-dimensional principal eigenvalue after projection is obtained.

[0079] The low-dimensional principal eigenvalue after projection is normalized:

[0080]

[0081] The normalized principal eigenvalue of each level is assigned to the RGB channel:

[0082]

[0083] Finally, the improved VI trajectory map that can distinguish homogeneous industrial loads is obtained.

[0084] Step 7, using the feature map obtained from the homogeneous industrial load data of each category to train the load classifier;

[0085] In this step, five load feature data sets are generated by steps 1-6 to train the load classifier, reference 2, and these feature data sets are proportionally divided into training sets, test sets and validation sets, each feature data set contains a plurality of feature vectors representing the feature information of different loads, and the load classifier is trained by the training set and the test set, wherein, in order to highlight the universality of the method, a classic neural network residual network is used.

[0086] Step 8, based on the industrial load to be classified, the load classifier is used to classify the industrial load to be classified.

[0087] Figure 2 The voltage-current (V-I) trajectory feature comparison chart of File1 to File5 five industrial loads is shown. It can be clearly observed that the voltage and current curves of these industrial loads at each time show high similarity, and their change trends are almost consistent. Therefore, it can be concluded that the traditional V-I trajectory chart has obvious limitations in distinguishing industrial loads with high homogeneity.

[0088] Figure 3 Figure 4 Figure 5 Figure 6 And Figure 7 The application effect of the method of the present application is shown. The method first extracts time series features from voltage and current data through distance similarity algorithm, and then maps the extracted three main features to RGB color channels through high-dimensional space mapping. In this process, the R channel corresponds to the most important feature, with 255 color values. The larger the color value, the greater the volatility of the feature, the farther the distance, and the lower the similarity. The G and B channels correspond to the second and third important features, respectively, and the principle is the same as the R channel. Over time, the color change in the figure intuitively reflects the dynamic difference of each load feature. The feature volatility that cannot be expressed in the V-I trajectory is amplified through color change, effectively improving the clarity of feature expression and the accuracy of load identification.

[0089] The present application provides an industrial load non-intrusive identification and classification method based on high-frequency current and voltage data, which aims to solve the problem of low identification accuracy of homogenized industrial loads in the prior art. Through multi-dimensional data processing, feature extraction and recursive graph construction, and a deep learning classification model, accurate identification and classification of complex industrial loads are achieved, which has the following beneficial effects:

[0090] (1) High-precision load feature extraction

[0091] ​​​By the multi-dimensional PAA processing technology, the application can significantly reduce the data dimension while keeping the main feature information, and improve the data processing efficiency. By constructing a weighted recursive matrix, the key feature information of the load is further extracted, so that the highly homogeneous industrial load can be more accurately distinguished.

[0092] (2) Intuitive visualization of dynamic characteristics

[0093] The application extracts the key information of time series data in the weighted recursive matrix, projects the key information to a high-dimensional space to decompose its dimensions, captures the dimension with the maximum feature difference, and uses color mapping difference to construct an improved VI trajectory diagram to intuitively display the dynamic characteristics of the load. This visualization method helps to more intuitively observe and analyze the characteristics of the load, and assists users in effectively managing and monitoring the load in a complex power system.

[0094] (3) Precise classification of homogeneous industrial loads

[0095] The application can effectively cope with the identification challenge of homogeneous industrial loads by designing and training a classic neural network residual network, significantly improving the classification accuracy, and helping to realize fine scheduling and energy visualization management of the power system.

[0096] In addition, the same technical concept as the method provides a homogeneous industrial load non-intrusive load identification system, a data processing module, a similarity calculation module, a feature extraction module, a V-I trajectory acquisition module, a feature map construction module, a classifier training module and a classification module.

[0097] The data processing module is used for multi-dimensional segmented aggregate approximation processing on the current data in the high-frequency data of the industrial load after high-frequency centralization processing.

[0098] The similarity calculation module constructs a distance similarity matrix using the data processed by the multi-dimensional segmented aggregate approximation processing , and quantifies the similarity of load characteristics by calculating the distance between data points.

[0099] The feature extraction module converts the distance similarity matrix to a weighted recursive matrix by normalization and threshold control, and extracts the key information of the load characteristics.

[0100] The V-I trajectory acquisition module is used to construct a discretized V-I trajectory based on the high-frequency current and voltage data in the high-frequency data of the industrial load directly acquired.

[0101] The feature map construction module is used to construct a feature map based on the weighted recursive matrix ​The discrete industrial load V-I trajectory is weighted and mapped to form a mapped recursive V-I trajectory feature map;

[0102] The classifier training module trains the load classifier by using the mapped recursive V-I trajectory feature maps of the homogeneous industrial load data of different categories.

[0103] The classification module classifies the to-be-classified industrial load by using the load classifier.

[0104] The application further provides a computer device comprising a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes, and the processor can realize the homogeneous industrial load non-intrusive load identification method when executing the computer executable program.

[0105] In another aspect, the application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program can realize the homogeneous industrial load non-intrusive load identification method when executed by a processor.

[0106] The computer device can be a notebook computer, a desktop computer or a workstation.

[0107] The processor can be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC) or a ready programmable gate array (FPGA).

[0108] The memory can be an internal storage unit of a notebook computer, a desktop computer or a workstation, such as a memory or a hard disk, or can be an external storage unit, such as a mobile hard disk or a flash card.

[0109] The computer readable storage medium can comprise a computer storage medium and a communication medium. The computer storage medium comprises volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules or other data. The computer readable storage medium can comprise a read only memory (ROM), a random access memory (RAM), a solid state disk (SSD) or an optical disk. The random access memory can comprise a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM). The computer readable storage medium can comprise a computer storage medium and a communication medium. The computer storage medium comprises volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules or other data. The computer readable storage medium can comprise a read only memory (ROM), a random access memory (RAM), a solid state disk (SSD) or an optical disk. The random access memory can comprise a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM).

[0110] The above merely illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made according to the technical idea of the present application, on the basis of the technical scheme, falls within the protection scope of the present application.

Claims

1. A non-intrusive load identification method for homogeneous industrial loads, characterized in that: The following steps are involved: Current data in high-frequency data of industrial load after high-frequency centralization processing Perform multi-dimensional segmented aggregation approximation processing; The distance similarity matrix is ​​constructed using the data after the multi-dimensional segmented aggregation approximation processing , quantifying the similarity of load characteristics by calculating the distance between data points; Through normalization and threshold control, the distance similarity matrix Convert to a weighted recursive matrix , extract key information of load characteristics; Constructing discretized VI trajectory based on high-frequency current and voltage data directly obtained from high-frequency data of industrial loads; Based on the weighted recursive matrix , the discretized industrial load VI trajectory is weighted mapped to form a recursive VI trajectory feature map after mapping; The load classifier is obtained by training the recursive VI trajectory feature map after mapping using homogeneous industrial load data of different categories. Based on the industrial load to be classified, the industrial load to be classified is classified using the load classifier.

2. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: The high-frequency data of the industrial load includes voltage data and current data, and the storage matrix is and ; For the matrix and Centralized processing, including: Calculate the minimum value of current and voltage signals 、 and maximum value 、 ; Calculate the step size of the data : ,in for 、 ; According to the step length Adjust the current and voltage signals to make their data distribution more uniform.

3. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: Current data When performing multi-dimensional segmented aggregation approximation processing, The mean of each segment of the signal is calculated as follows: in , is the segment index, is the length of the original signal, Embedding size for the target.

4. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: Constructing a distance similarity matrix This involves constructing a distance measurement matrix between each time series point, where the matrix measurement uses Euclidean distance; the distance between any two points is defined as follows: ,in, is the power parameter in the distance calculation.

5. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: Through normalization and threshold control, the distance similarity matrix Convert to a weighted recursive matrix ,The key information of load characteristics extraction includes: Calculate the normalized distance matrix : ; Clamp distance values ​​to a threshold Within: ; The distance similarity matrix Convert to weighted recursive graph ,parameter and The selection satisfies ,in, .

6. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: Based on the weighted recursive matrix The discretized industrial load VI trajectory is weighted mapped to form a recursive VI trajectory feature map after mapping, including: the main features are mapped from low-dimensional space to high-dimensional space through high-dimensional mapping technology, feature decomposition is performed, the maximum feature of feature differentiation in the high-dimensional space is captured for low-dimensional projection, the projection data is normalized, and color mapping is used to realize the visualization expression of differential features, thereby obtaining an improved VI trajectory map that can distinguish homogeneous industrial loads.

7. The non-intrusive load identification method for homogeneous industrial loads according to claim 1 is characterized in that: By generating five load feature data sets, the feature data sets are divided into training set, test set and validation set in proportion. Each feature data set contains several feature vectors, which represent the characteristic information of different loads. Based on the classical neural network residual network, the load classifier is trained through the training set, test set and validation set.

8. A non-intrusive load identification system for homogeneous industrial loads, characterized in that: Data processing module, similarity calculation module, feature extraction module, VI trajectory acquisition module, feature map construction module, classifier training module and classification module; The data processing module is used to process the current data of the high-frequency data of industrial load after high-frequency centralized processing. Perform multi-dimensional segmented aggregation approximation processing; The similarity calculation module uses the data after the multi-dimensional segmented aggregation approximation processing to construct a distance similarity matrix , quantifying the similarity of load characteristics by calculating the distance between data points; The feature extraction module uses normalization and threshold control to transform the distance similarity matrix Convert to a weighted recursive matrix , extract key information of load characteristics; The VI trajectory acquisition module is used to construct a discretized VI trajectory based on high-frequency current and voltage data in the high-frequency data of the industrial load directly acquired; The feature map construction module is used based on the weighted recursive matrix , the discretized industrial load VI trajectory is weighted mapped to form a recursive VI trajectory feature map after mapping; The classifier training module uses the recursive VI trajectory feature map after mapping of homogeneous industrial load data of different categories to train the load classifier; The classification module classifies the industrial load to be classified based on the industrial load to be classified by using the load classifier.

9. A computer device, characterized in that: The invention comprises a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the non-intrusive load identification method of homogeneous industrial loads as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, the non-intrusive load identification method for homogeneous industrial loads as claimed in any one of claims 1 to 7 can be implemented.

Images