Non-intrusive load identification method based on improved V-I track
Patent Information
- Application Number
- CN202410950303.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-12
AI Technical Summary
传统负荷辨识方法仍存在以下问题:1、采用单一的电气数据进行特征构造,负荷特征信息挖掘不足;2、直接利用图像分类模型作为负荷辨识模型,模型规模过大
[0038] 1. By pairing V-I f Trajectory, active current, and instantaneous power features are fused to increase the information dimension of the features and improve the distinction of the features;
Smart Images

Figure CN118940102A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of electric power equipment, relates to a non-intrusive load monitoring technology, and specifically relates to a non-intrusive load identification method based on an improved VI trajectory. Background Art
[0002] With the development of society, energy consumption and carbon emission issues are becoming increasingly serious. Load monitoring technology can effectively monitor the operation of each load of end users, help users better understand their own electricity consumption behavior, cultivate green energy consumption habits, improve energy utilization, and help achieve the "dual carbon" goal as soon as possible. Therefore, it is of great significance to realize the identification of different types of electrical equipment.
[0003] Domestic experts and scholars have conducted active research on the problem of load identification. LAM et al. used normalized voltage and current data to construct VI trajectories, and extracted corresponding shape features as load features for load identification. DU et al. mapped the VI trajectory to a grid to form a binary VI trajectory image feature to reduce computing costs. LIU used the binary VI trajectory image feature as an intermediate domain between visual recognition and NILM, and used the transfer learning method to apply the model in the field of visual recognition to load identification. The traditional VI trajectory is constructed directly using voltage and current, but because the power factor of the resistive load is large and the active component accounts for a high proportion, it is difficult to achieve the expected identification accuracy by directly using VI trajectory features. TESHOME et al. used reactive current and voltage to construct VI f Trajectory, increasing the characteristic distinction between different types of resistive loads.
[0004] Due to the excellent performance of deep learning in speech recognition, image classification and video segmentation, deep learning algorithms have also been widely used in load identification models. Chen Junfeng et al. used GAF to convert voltage, current and reactive current into image features, and then fused them through color coding technology. Finally, RBFNet and L2-AlexNet were used to realize collaborative load identification on the edge and cloud respectively; Cui Haoyang et al. used transfer learning to identify the features of visualized non-active current images with the pre-trained Inception-v3 model, and improved the learning efficiency of the model while maintaining a high recognition accuracy; Wang Shouxiang et al. respectively input power and VI trajectory features into the feature extraction model constructed by back propagation neural network and convolutional neural network (CNN), and then fused the extracted features to complete the identification. Xie Yang et al. used AlexNet to classify VI trajectories and achieved a high accuracy. Although many network models with excellent performance have appeared in the field of image recognition, and they have also achieved a good combination with the NILM field.
[0005] At present, load identification adopts the method of constructing features of load electrical data and classifying features through neural networks to identify load types. Traditional load identification methods still have the following problems: 1. Using single electrical data for feature construction, the load feature information mining is insufficient; 2. Directly using image classification model as load identification model, the model scale is too large. Summary of the invention
[0006] Purpose of the invention: In order to overcome the deficiencies in the prior art, a non-intrusive load identification method based on an improved VI trajectory is provided, which can accurately identify the type of electrical equipment and monitor the usage of various loads.
[0007] Technical solution: To achieve the above object, the present invention provides a non-intrusive load identification method based on improved VI trajectory, comprising the following steps:
[0008] S1: Extract single load current data;
[0009] S2: Decompose the single load current data into reactive current data and active current data;
[0010] S3: Construct color fusion features using active current, reactive current, voltage and power data;
[0011] S4: Construct a lightweight load identification model DSHDNet using depthwise separable convolution DSC and hybrid dilated convolution HDC;
[0012] S5: Divide the color fusion features into a training set and a test set, input the training set into the DSHDNet model for training, and use the test set to test and verify the DSHDNet model to obtain a trained lightweight load identification model DSHDNet;
[0013] S6: The load identification results are obtained through the output of the trained lightweight load identification model DSHDNet.
[0014] Furthermore, the method for extracting the single load current data in step S1 is: according to the time point of the load switching event, the voltage and current of the three cycles before and after are subtracted to obtain the single load current data.
[0015] Furthermore, in step S2, the current sequence is decomposed by Fryze power theory to obtain active current components and reactive current components.
[0016] Furthermore, the Fryze power decomposition theory calculation formula in step S2 is as follows:
[0017]
[0018] i f (t) = i(t) - i a (t)(4)
[0019] Where: T is the selected steady-state period; P is the active power of the steady-state period; V is the effective value of the voltage in the steady-state period; i(t) is the current value at time t in the steady-state period; v(t) is the voltage value at time t in the steady-state period; i a (t) is the active current component; i f (t) is the reactive current component.
[0020] Furthermore, the construction of the color fusion feature in step S3 includes the following steps:
[0021] A1: Use the normalized reactive current and voltage data as the horizontal and vertical coordinates to plot the VI f Trajectory;
[0022] A2: Convert active current and instantaneous power into image features through Gram angular field (GAF);
[0023] A3: VI f The trajectory, active current, and instantaneous power features are input into the R, G, and B color channels respectively to form color fusion features and increase the information dimension of the features.
[0024] Furthermore, in step A1, the steady-state current and voltage sequences are normalized, and the calculation formula is as follows:
[0025]
[0026] Where: v(t) is the voltage value at time t in the extraction period; v max and v min They represent the maximum and minimum voltage values in the extraction cycle respectively; i(t) is the current value at time t in the extraction cycle; i max and i min They represent the maximum and minimum current in the extraction cycle respectively.
[0027] Furthermore, in step A1, VI is constructed by trajectory mapping. f Trajectory features,Trajectory mapping specifically includes the following steps:
[0028] B1: Let v k =[v1,v2,…,v K ] and i k =[i1,i2,…,i K ] are voltage and current sequences of length K respectively;
[0029] B2: Normalize the voltage and current sequences to obtain sequences with values between 0 and 1 and
[0030] B3: Define the size of the grid as N×N, which means the generated VI f The trajectory map resolution is N×N, and a set of N×N zero matrices are created;
[0031] B4: Convert voltage and current according to the grid size. The specific conversion formula is as follows:
[0032]
[0033] Where: is the floor rounding function; v m and i m The value of is less than or equal to N;
[0034] B5: V m and i m Sampling from the first element in sequence, the i-th element in the matrix m Row v m The elements of the column are assigned a value of 1, and this process is repeated until all elements in the sequence are traversed.
[0035] Furthermore, in step S4, the expansion rates of the separation convolution parts in the three depthwise separable convolutions DSC are set to 1, 2, and 3, respectively, to achieve the fusion of the depthwise separable convolution DSC and the hybrid dilated convolution HDC, thereby improving the receptive field of the convolution layer and reducing the number of parameters and computation of the model.
[0036] Furthermore, in step S5, VI f The trajectory, active current, and instantaneous power images are input into the DSHDNet lightweight model in parallel. The feature fusion of the three images is achieved through separate convolution and point convolution with different dilation rates, which improves the recognition accuracy of the model and reduces the scale of the model.
[0037] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0038] 1. Through VI f The trajectory, active current, and instantaneous power features are integrated to increase the information dimension of the features and improve the distinguishability of the features;
[0039] 2. The load identification model is constructed using depthwise separable convolution (DSC) and hybrid dilated convolution (HDC), which maintains the load identification accuracy and reduces the model size. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1It is a schematic diagram of the process of the present invention;
[0041] Figure 2 A framework diagram of the lightweight load identification model provided by the present invention;
[0042] Figure 3 This is the color fusion feature image provided by the present invention. DETAILED DESCRIPTION
[0043] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0044] The present invention provides a non-intrusive load identification method based on improved VI trajectory, such as Figure 1 As shown, it includes the following steps:
[0045] S1: Extract single load current data:
[0046] The method for extracting single load current data is as follows: according to the time point when the load switching event occurs, the voltage and current of the three cycles before and after are subtracted to obtain the single load current data.
[0047] S2: Decompose the single load current data into reactive current data and active current data:
[0048] The power factor of resistive load is large, and the active component accounts for a high proportion. The traditional VI trajectory is drawn by load voltage and current data, which makes it difficult to accurately identify this type of load. Therefore, the present invention decomposes the current sequence through Fryze power theory to obtain active current components and reactive current components;
[0049] The Fryze power decomposition theory calculation formula is as follows:
[0050]
[0051] i f (t) = i(t) - i a (t)(4)
[0052] Where: T is the selected steady-state period; P is the active power of the steady-state period; V is the effective value of the voltage in the steady-state period; i(t) is the current value at time t in the steady-state period; v(t) is the voltage value at time t in the steady-state period; i a (t) is the active current component; i f (t) is the reactive current component.
[0053] S3: Construct color fusion features using active current, reactive current, voltage and power data:
[0054] Single Grayscale VI f The trajectory can maintain effective identification of other electrical equipment on the basis of improving the distinction between different resistive loads, but it has the problem of insufficient information mining and containing a large amount of invalid blank information. The present invention uses GAF to encode the power and active current waveforms, and uses color coding to compare them with the grayscale VI f The trajectories are effectively fused to form color fusion features, which increases the dimension of feature information and helps improve recognition accuracy;
[0055] Step S3 specifically includes the following steps:
[0056] A1: Use the normalized reactive current and voltage data as the horizontal and vertical coordinates to plot the VI f Trajectory:
[0057] The current waveforms of the same type of electrical equipment are roughly the same, but due to different power consumption, the current size will be different, which will lead to the increase in the difference in VI trajectory characteristics of the same type of load. In order to reduce this uncertainty, the steady-state current and voltage series need to be normalized, and the calculation formula is as follows:
[0058]
[0059] Where: v(t) is the voltage value at time t in the extraction period; v max and v min They represent the maximum and minimum voltage values in the extraction cycle respectively; i(t) is the current value at time t in the extraction cycle; i max and i min Respectively represent the maximum and minimum values of the current during the extraction cycle;
[0060] Use the normalized reactive current and voltage series as the horizontal and vertical coordinates to plot the VI f Trajectory diagram, but directly constructed VI f The trajectory features are large, which will generate a huge amount of calculation, which is not conducive to practical application. f The track is mapped to the grid. If the track passes through a cell, the cell is occupied and assigned a value of 1, otherwise it is assigned a value of 0. This processing method can preserve the VI to the greatest extent. f The information of the trajectory can significantly reduce the computational cost. Therefore, in step A1, VI is constructed by trajectory mapping. f Trajectory features,Trajectory mapping specifically includes the following steps:
[0061] B1: Let vk =[v1,v2,…,v K ] and i k =[i1,i2,…,i K ] are voltage and current sequences of length K respectively;
[0062] B2: Normalize the voltage and current sequences to obtain sequences with values between 0 and 1 and
[0063] B3: Define the size of the grid as N×N, which means the generated VI f The trajectory map resolution is N×N, and a set of N×N zero matrices are created;
[0064] B4: Convert voltage and current according to the grid size. The specific conversion formula is as follows:
[0065]
[0066] Where: is the floor rounding function; v m and i m The value of is less than or equal to N;
[0067] B5: V m and i m Sampling from the first element in sequence, the i-th element in the matrix m Row v m The elements of the column are assigned a value of 1, and this process is repeated until all elements in the sequence are traversed.
[0068] A2: Active current and instantaneous power are converted into image features through Gram angular field (GAF):
[0069] A3: VI f The trajectory, active current, and instantaneous power features are input into the R, G, and B color channels respectively to form color fusion features and increase the information dimension of the features.
[0070] In the Construction VI f Only reactive current is used in trajectory characteristics. The lack of active current will cause the load resistance information to be lost. At the same time, power, as an important characteristic of the load, cannot be used in VI. f To solve the above problem, step A2 of the present invention converts active current and instantaneous power into image features through Gram Angular Field (GAF), and then step A3 inputs the encoded active current and power information into G and B channels respectively in RGB color space, and converts VI fThe trajectory image is input into the R channel to form a color image that integrates different feature information.
[0071] In the Construction VI f Only reactive current is used in trajectory characteristics. The lack of active current will cause the load resistance information to be lost. At the same time, power, as an important characteristic of the load, cannot be used in VI. f To solve the above problem, the active current and instantaneous power can be converted into image features through the Gram Angular Field (GAF), and then the encoded active current and power information can be input into the G and B channels respectively in the RGB color space. f The trajectory image is input into the R channel to form a color image that integrates different feature information.
[0072] Step S3 of the present invention makes full use of the blank part in the trajectory image, reflects the difference in the value in the cell through the difference in color and brightness of each pixel, and well integrates different feature information, making up for the lack of single feature information and VI f The problem of information loss in the trajectory is solved, which significantly improves the utilization of feature information.
[0073] S4: Construct a lightweight load identification model DSHDNet using depthwise separable convolution DSC and hybrid dilated convolution HDC;
[0074] In order to increase the learning ability of features, the existing load identification model continues to expand the scale of the model, resulting in large model calculations and long training time. To solve the above problems, the hybrid dilated convolution and the depthwise separable convolution can be combined to reduce the model calculation amount while maintaining accurate load identification;
[0075] Reference Figure 2 In the present invention, the expansion rates of the separation convolution parts in the three depthwise separable convolutions DSC are set to 1, 2, and 3 respectively, so as to realize the fusion of the depthwise separable convolution DSC and the hybrid dilated convolution HDC, improve the receptive field of the convolution layer and reduce the number of parameters and the amount of calculation of the model;
[0076] The expansion rate of the existing convolution is 1, and HDC modifies the expansion rate parameter. DSC can be divided into two parts: separation convolution and point convolution. The expansion rate parameters of the three DSC modules are set to 1, 2, and 3 respectively, thereby realizing the fusion of DSC and HDC modules.
[0077] The dilated convolution module increases the receptive field by filling the hole in the middle of the convolution kernel, but stacking multiple dilated convolutions with the same dilation rate will lead to information loss. The model of the present invention selects dilated convolutions with dilation rates of 1, 2, and 3 to replace the stacking of multiple dilated convolutions with the same dilation rate, thereby enhancing the coherence of the field of view.
[0078] S5: Divide the color fusion features into a training set and a test set, input the training set into the DSHDNet model for training, use the test set to test and verify the DSHDNet model, use the precision, recall, and F1 score to evaluate the model output results, and use the number of parameters, calculation amount, and training time to evaluate the model scale, and obtain the trained lightweight load identification model DSHDNet;
[0079] S6: The load identification results are obtained through the output of the trained lightweight load identification model DSHDNet.
[0080] The existing load identification only considers a single feature as the model input and directly uses the image classification model as the load identification model. f The trajectory, active current, and instantaneous power images are input into the DSHDNet lightweight model in parallel. The feature fusion of the three images is achieved through separate convolution and point convolution with different dilation rates, which improves the recognition accuracy of the model and reduces the scale of the model.
[0081] The present invention collects new data in real time on site, imports the pre-processed new data into the training set, retrains the model, and continuously improves the prediction accuracy of the model.
[0082] Based on the technical solution of the present invention, this embodiment applies the method of the present invention as follows:
[0083] The data of this paper adopts the public dataset PLAID, whose sampling frequency is 30kHz. Considering the actual power consumption environment, this paper selects 1744 sets of voltage and current data of common electrical appliances in household scenarios, including refrigerators, hair dryers, microwave ovens, laptops, washing machines, vacuum cleaners, fluorescent lamps, incandescent lamps, air conditioners, fans, and heaters.
[0084] The PLAID public dataset contains large differences in the same type of equipment data, that is, the features between the same type of equipment are quite different, which is conducive to improving sample diversity, enabling better network training and increasing network stability. In the PLAID dataset, the number of load samples is unbalanced. The number of samples for hair dryers is 246, but the number of samples for vacuum cleaners is 73. Since the network learns more information on loads with more samples than on loads with fewer samples during training, the network recognition accuracy will decrease. Therefore, it is necessary to process the unbalanced dataset. The present invention attaches corresponding labels to the new samples generated by Borderline-SMOTE and adds them to the original sample set. There are a total of 2706 samples, of which the number of samples for different types of loads is expanded to 246.
[0085] The present invention adopts the constructed DSHDNet lightweight model to perform load identification. Figure 2 The network architecture of the load identification model built based on Pytorch is shown. After the data set is processed by Borderline-SMOTE, 50 samples of each category are randomly selected from the original real sample set in this example to form a test set with a total of 550 samples. The rest of the simulated samples generated by oversampling are used as training sets for training. In order to prevent the occurrence of overfitting, a Dropout layer is added after the first fully connected layer (FC), and the parameter is set to 0.15. Finally, the classification results are output through Softmax. The activation function of each layer uses the linear rectifier unit ReLU, the loss function of the model uses the cross entropy loss function, and Adam is used to optimize the network parameters. The number of training times is 500. The parameter settings of each model are shown in Table 1.
[0086] Table 1
[0087] Model parameter settings HDC-DSC1 Convolution kernel size = 3*3, dilation rate = 1 HDC-DSC2 Convolution kernel size = 3*3, dilation rate = 2 HDC-DSC3 Convolution kernel size = 3*3, dilation rate = 3 Fully connected layer 1 Number of neurons = 200 Fully connected layer 2 Number of neurons = 100 Output Layer Number of neurons = 11
[0088] according to Figure 3 It can be seen that the VI trajectory and the active current and instantaneous power features encoded by the GAF image are input into the R, G, and B color channels respectively, and the fusion features of different types of loads are obtained as a color image with a resolution of 64×64.
[0089] The present invention selects 11 types of loads in the data set as labels, uses the color fusion features of different types of loads as input, and uses the DSHDNet model to identify the types of loads. The number of parameters of the model proposed by the present invention is 42285, the amount of calculation is 372100, and the training time is 346.79 seconds. The precision rate of the present invention on the test set reached 0.953, the recall rate reached 0.954, and the F1 score reached 0.953, realizing high-precision identification of different types of loads at a low computational cost.
[0090] This paper uses the public dataset PLAID to establish a load identification method based on multi-feature fusion and lightweight model. The voltage, active current, reactive current and instantaneous power data of the load are converted into image features, and the load type is identified by using the lightweight model DSHDNet through sample data classification. Finally, the effectiveness of the method is verified by the public dataset, and the method can be applied to engineering practice.
Claims
1. A non-intrusive load identification method based on improved VI trajectory, characterized in that: The steps include: S1: Extract single load current data; S2: Decompose the single load current data into reactive current data and active current data; S3: Construct color fusion features using active current, reactive current, voltage and power data; S4: Construct a lightweight load identification model DSHDNet using depthwise separable convolution DSC and hybrid dilated convolution HDC; S5: Divide the color fusion features into a training set and a test set, input the training set into the DSHDNet model for training, and use the test set to test and verify the DSHDNet model to obtain a trained lightweight load identification model DSHDNet; S6: The load identification results are obtained through the output of the trained lightweight load identification model DSHDNet.
2. A non-intrusive load identification method based on improved VI trajectory according to claim 1, characterized in that: The method for extracting the single load current data in step S1 is: according to the time point when the load switching event occurs, the voltage and current of the three cycles before and after are subtracted to obtain the single load current data.
3. A non-intrusive load identification method based on improved VI trajectory according to claim 1, characterized in that: In step S2, the current sequence is decomposed by Fryze power theory to obtain active current components and reactive current components.
4. A non-intrusive load identification method based on improved VI trajectory according to claim 3, characterized in that: The Fryze power decomposition theory calculation formula in step S2 is as follows: i f (t)=i(t)-i a (t) (4) Where: T is the selected steady-state period; P is the active power of the steady-state period; V is the effective value of the voltage in the steady-state period; i(t) is the current value at time t in the steady-state period; v(t) is the voltage value at time t in the steady-state period; i a (t) is the active current component; i f (t) is the reactive current component.
5. The non-intrusive load identification method based on improved VI trajectory according to claim 1, characterized in that: The construction of the color fusion feature in step S3 includes the following steps: A1: Use the normalized reactive current and voltage data as the horizontal and vertical coordinates to plot the VI f Trajectory; A2: Convert active current and instantaneous power into image features through Gram angular field (GAF); A3: VI f The trajectory, active current, and instantaneous power features are input into the R, G, and B color channels respectively to form color fusion features and increase the information dimension of the features.
6. A non-intrusive load identification method based on improved VI trajectory according to claim 5, characterized in that: In step A1, the steady-state current and voltage sequences are normalized, and the calculation formula is as follows: Where: v(t) is the voltage value at time t in the extraction period; v max and v min They represent the maximum and minimum voltage values in the extraction cycle respectively; i(t) is the current value at time t in the extraction cycle; i max and i min They represent the maximum and minimum values of the current during the extraction cycle respectively.
7. The non-intrusive load identification method based on improved VI trajectory according to claim 5, characterized in that: In step A1, VI is constructed by trajectory mapping. f Trajectory features,Trajectory mapping specifically includes the following steps: B1: Let v k =[v1,v2,…,v K ] and i k =[i1,i2,…,i K ] are voltage and current sequences of length K respectively; B2: Normalize the voltage and current sequences to obtain sequences with values between 0 and 1 and B3: Define the size of the grid as N×N, which means the generated VI f The trajectory map resolution is N×N, and a set of N×N zero matrices are created; B4: Convert voltage and current according to the grid size. The specific conversion formula is as follows: Where: is the floor rounding function; v m and i m The value of is less than or equal to N; B5: V m and i m Sampling from the first element in sequence, the i-th element in the matrix m Row v m The elements of the column are assigned a value of 1, and this process is repeated until all elements in the sequence are traversed.
8. The non-intrusive load identification method based on improved VI trajectory according to claim 1 is characterized in that: In the step S4, the expansion rates of the separation convolution parts in the three depthwise separable convolutions DSC are set to 1, 2, and 3 respectively, so as to realize the fusion of the depthwise separable convolution DSC and the hybrid dilated convolution HDC, improve the receptive field of the convolution layer and reduce the number of parameters and the amount of calculation of the model.
9. A non-intrusive load identification method based on improved VI trajectory according to claim 8, characterized in that: In step S5, VI f The trajectory, active current, and instantaneous power images are input into the DSHDNet lightweight model in parallel. The feature fusion of the three images is achieved through separate convolution and point convolution with different dilation rates, which improves the recognition accuracy of the model and reduces the scale of the model.
Citation Information
Patent Citations
Load identification method based on feature fusion and deep learning and application thereof
CN110443302A
Power consumer load fine grit identification method and identification system
CN113036759A
Full convolutional neural network model for rapid classification
CN113920421A
Non-intrusive load identification method based on time sequence imaging and deep learning
CN117315434A
Non-intrusive load identification method based on binary V-I track color coding
CN117351259A
Cited By
Non-intrusive load monitoring system based on deep residual network
CN122652217A
