A non-intrusive load identification method based on binarization V-I trajectory color coding

Abstract

A non-intrusive load identification method based on binary V-I trajectory color coding includes the following steps: extracting the steady-state interval of the device from the collected high-frequency voltage and current data; using Fryze power theory to decompose the current into active and reactive current; normalizing the obtained voltage and reactive current data and drawing a V-I trajectory curve; after obtaining the V-I trajectory, performing binary trajectory mapping and HSV color coding; then constructing a ResNet50 neural network, taking the image as input, adjusting the fully connected layer number of the neural network to be consistent with the number of device categories for feature extraction and load identification. By converting the V-I trajectory into a visual representation through HSV color coding, the knowledge between different fields is linked, which not only enhances the uniqueness of the load features, but also applies the excellent performance of the ResNet50 neural network in the image recognition field to the non-intrusive load identification field, and finally verifies it using the PLAID public data set.

Description

Technical Field

[0001] This invention relates to the field of power load identification technology, specifically to a non-intrusive load identification method based on binary VI trajectory color coding. Background Technology

[0002] Traditional load monitoring employs an invasive method, installing sensors on each piece of electrical equipment to record its power consumption and energy usage. While this method yields relatively accurate data, the cost of deploying sensors is prohibitively high. Therefore, non-intrusive load monitoring (NILM) has gradually gained attention. Compared to traditional invasive load monitoring, NILM only requires acquiring current and voltage data from the user's bus. This data is then processed and combined with intelligent algorithms to identify the type and operating status of the load within the user's premises. Because NILM requires less investment, is more practical, and facilitates refined demand-side management of the power grid, it helps optimize grid operation and dispatch, achieving peak shaving and valley filling. Consequently, NILM has received extensive and in-depth research from experts both domestically and internationally.

[0003] Non-intrusive load identification (NILM) is a crucial issue in the field, encompassing two main stages: feature extraction and device identification. Load characteristics are typically categorized into electrical and non-electrical features. Electrical features include: 1) power-related features such as active power, reactive power, and power factor; 2) current-related features such as current waveform, harmonic components, and their content; and 3) VI trajectory features such as VI trajectory diagrams. Non-electrical features, such as high-frequency electromagnetic interference, lack significant advantages in feature specificity and are therefore less frequently used. Among these, VI trajectory features have received considerable attention and research due to their ability to reflect the dynamic impedance characteristics of the device, their high correlation with device function, and their strong feature specificity.

[0004] In existing research, most scholars use normalized high-frequency voltage and current data to plot VI trajectories. This method has the following drawbacks: First, it is difficult to capture the fine-grained differences between appliances with similar trajectories, resulting in inaccurate classification. Second, ordinary VI trajectory diagrams mainly reflect the impedance characteristics of the load and cannot reflect the power characteristics of the equipment, making it difficult to quickly distinguish equipment with large power differences. Third, the plotting process is computationally intensive and time-consuming, which is not conducive to its widespread application. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a non-intrusive load identification method based on binary VI trajectory color encoding, which can not only transfer the trajectory into a binary matrix without sacrificing computational complexity, but also reduce small fluctuations in equipment load; at the same time, the binary VI trajectory map after HSV color encoding can better reflect various electrical characteristics and effectively improve the accuracy of load identification.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A non-intrusive load recognition method based on binary VI trajectory color encoding includes the following steps:

[0008] Step 1: Use high-frequency sampling equipment to acquire the voltage and current data of the home bus in real time;

[0009] Step 2: Assuming that only one device is turned on or off at a time, use an event detection algorithm to determine whether the device has changed its state. Once the device's operating state reaches a steady state, extract its steady-state current and voltage data.

[0010] Step 3: Preprocess the original current waveform according to Fryze power theory, and decompose the load current into active current and reactive current.

[0011] Step 4: Normalize the voltage and reactive current data obtained in Step 2 and 3, plot the VI trajectory curve, and perform binary trajectory mapping on it to enhance the uniqueness of the VI trajectory.

[0012] Step 5: Perform HSV color encoding on the binary VI trajectory obtained in Step 4 to obtain a binary color-encoded VI trajectory image;

[0013] Step 6: Construct a ResNet50 neural network model, using the binarized color VI trajectory image as input to classify the devices, thus completing non-invasive load identification.

[0014] In Step 2 above, when extracting high-frequency data from the device, the voltage waveform is collected starting from the zero-crossing period, and the voltage and current data within one steady-state period are collected.

[0015] The steps in Step 3 above for decomposing the load current using Fryze power theory are as follows:

[0016] i(t) = i a (t)+i r (t) (1)

[0017]

[0018]

[0019]

[0020] Among them, i a (t) represents the active current; i r (t) reactive current; P active Active power; V rms denoted as RMS voltage; T is the power supply cycle; v(t) is the voltage at time t in steady-state high-frequency sampling data; i(t) is the current at time t in steady-state high-frequency sampling data.

[0021] The specific steps of the binarization mapping method for the VI trajectory composed of voltage and reactive current in Step 4 above are as follows: Step 4.1: Obtain the waveforms of voltage v and reactive current i when the load is in steady state at a high sampling rate;

[0022] Step 4.2: Normalize the voltage and reactive current using the maximum absolute values ​​of the voltage and current waveforms:

[0023] i max =max(|i|), v max =max(|v|) (5)

[0024] Step 4.3: Divide the two-dimensional plane of voltage and reactive current into a uniformly sized cell grid, and calculate the length and width of each cell:

[0025]

[0026] In the formula: N∈Z + Thus, the 2N×2N elements in the matrix are obtained;

[0027] Step 4.4: Interpolate the original trajectory to ensure the continuity of the curve after mapping the binary trajectory. This is because the distance D between adjacent points... k Sometimes it is larger than the length or width of a single cell:

[0028]

[0029] In the formula: v k and i k These are the k-th data points of v and i, respectively;

[0030] Step 4.5: The interpolated trajectory points (I) j V j ),j=1,2,......J, mapped to 2N×2N grid cells.

[0031] In Step 5 above, the binarized VI trajectory is processed using HSV color encoding. The trajectory direction, power factor, and average pixel value over M periods are integrated into the H, S, and V channels, respectively. The specific method is as follows:

[0032] Step 5.1: To obtain the motion information of the VI trajectory, the hue H is used to represent the direction of the trajectory, and denoted as:

[0033]

[0034] In the formula: the function arg is the arctangent of the four quadrants, calculates the phase angle between two consecutive points in the trajectory, and the obtained angle ranges from 0° to 360°. Then, the hue of the j-th point is stored in a new 2N×2N matrix.

[0035]

[0036] In the formula: A = {a | passing through each point of the grid (nI)} j ,nV j )};

[0037] Step 5.2: When the load current is divided into active and reactive components, using only the reactive current to plot the VI trajectory will result in the loss of information about the relationship between the active and reactive components. Therefore, the saturation degree S is used to represent the ratio of active power to reactive power over multiple cycles, i.e., the power factor.

[0038]

[0039] In the formula: K is the total number of sampling points, P apparent For apparent power, V rms I rms These are the effective values ​​of the load voltage and load current, respectively.

[0040] Step 5.3: Brightness V, another color-generating attribute of the HSV color space, is also used to represent the repeatability of the VI trajectory; a binary image W with M cycles is generated using voltage and current waveforms. m (m=1,2,...,M), the brightness V is obtained by averaging the pixel values:

[0041]

[0042] The specific method for constructing the ResNet50 neural network in Step 6 above is as follows:

[0043] Step 6.1: Construct a ResNet50 neural network, consisting of 1 + 3 × (3 + 4 + 6 + 3) = 49 convolutional layers, plus the final fully connected layer, for a total of 50 layers. The ResNet50 network structure can be divided into seven parts. The first part does not contain residual blocks and mainly performs convolution, regularization, activation function, and max pooling calculations on the input. The second, third, fourth, and fifth parts all contain residual blocks. The 1×1 convolutional kernels in each part do not change the size of the residual blocks, but only change the dimensions of the residual blocks.

[0044] Step 6.2: To adapt to the actual number of classifications in the load recognition task, the number of neurons in the last fully connected layer was adjusted from 1000 to 11, which is the actual number of device categories.

[0045] Step 6.3: When training the ResNet50 model, set the initial learning rate to 10. -4 During training, if the loss value does not decrease within 10 epochs, the learning rate will be lowered to optimize the model. At the same time, to adapt to the computer processor and memory resources, the batch size during training will be configured to 8, and the number of model iterations will be configured to 300.

[0046] The sampling frequency of the data obtained in Step 1 above needs to be 30kHz, that is, 30,000 points are sampled per second. At the same time, the power frequency is 60Hz, that is, there are 60 cycles in 1 second, and one cycle contains 30,000 / 60 = 500 sampling points.

[0047] In Step 2 above, an event detection algorithm is used to determine whether the electrical appliance has undergone a state change, extract its steady-state current and voltage data, and determine whether the electrical appliance is operating in a steady-state phase by observing whether the voltage and current trajectories in several consecutive cycles overlap.

[0048] This invention provides a non-intrusive load identification method based on binary VI trajectory color encoding, which can transfer the trajectory into a binary matrix without sacrificing computational complexity, and can also mitigate small fluctuations in equipment load; at the same time, the binary VI trajectory map after HSV color encoding can better reflect various electrical characteristics and effectively improve the accuracy of load identification. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0050] Figure 1 This is a flowchart illustrating the overall framework of a non-intrusive load recognition method based on binarized VI trajectory color coding.

[0051] Figure 2(a) High-frequency sampling data of the air conditioner; (b) Steady-state sampling data of the air conditioner; (c) Voltage waveform of the air conditioner during operation; (d) Voltage waveform of the air conditioner during steady state; (e) Current waveform of the air conditioner during operation; (f) Current waveform of the air conditioner during steady state.

[0052] Figure 3 The active and reactive current waveforms extracted from the original current waveform are shown.

[0053] Figure 4 The VI trajectory diagram of 11 electrical appliances after binarization and color coding;

[0054] Figure 5 The specific steps of the RsNet50 neural network load identification algorithm are as follows. Detailed Implementation

[0055] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0056] like Figure 1 As shown, a non-intrusive load recognition method based on binary VI trajectory color encoding includes the following steps:

[0057] Step 1: Use high-frequency sampling equipment to acquire the voltage and current data of the home bus in real time;

[0058] Step 2: Assuming that only one device is turned on or off at a time, use an event detection algorithm to determine whether the device has changed its state. Once the device's operating state reaches a steady state, extract its steady-state current and voltage data.

[0059] Step 3: Preprocess the original current waveform according to Fryze power theory, and decompose the load current into active current and reactive current.

[0060] Step 4: Normalize the voltage and reactive current data obtained in Step 2 and 3, plot the VI trajectory curve, and perform binary trajectory mapping on it to enhance the uniqueness of the VI trajectory.

[0061] Step 5: Perform HSV color encoding on the binary VI trajectory obtained in Step 4 to obtain a binary color-encoded VI trajectory image;

[0062] Step 6: Construct a ResNet50 neural network model, using the binarized color VI trajectory image as input to classify the devices, thus completing non-invasive load identification.

[0063] In Step 2 above, when extracting high-frequency data from the device, the voltage waveform is collected starting from the zero-crossing period, and the voltage and current data within one steady-state period are collected.

[0064] The steps in Step 3 above for decomposing the load current using Fryze power theory are as follows:

[0065] i(t) = i a (t)+i r (t) (1)

[0066]

[0067]

[0068]

[0069] Among them, i a (t) represents the active current; i r (t) reactive current; P active Active power; V rms denoted as RMS voltage; T is the power supply cycle; v(t) is the voltage at time t in steady-state high-frequency sampling data; i(t) is the current at time t in steady-state high-frequency sampling data.

[0070] The specific steps of the binarization mapping method for the VI trajectory composed of voltage and reactive current in Step 4 above are as follows:

[0071] Step 4.1: Obtain the waveforms of voltage v and reactive current i under steady-state load conditions at a high sampling rate;

[0072] Step 4.2: Normalize the voltage and reactive current using the maximum absolute values ​​of the voltage and current waveforms:

[0073] i max =max(|i|), v max =max(|v|) (5)

[0074] Step 4.3: Divide the two-dimensional plane of voltage and reactive current into a grid of uniformly sized cells, and calculate the length and width of each cell:

[0075]

[0076] In the formula: N∈Z + Thus, the 2N×2N elements in the matrix are obtained;

[0077] Step 4.4: Interpolate the original trajectory to ensure the continuity of the curve after mapping the binary trajectory. This is because the distance D between adjacent points... k Sometimes it is larger than the length or width of a single cell:

[0078]

[0079] In the formula: v k and i k These are the k-th data points of v and i, respectively;

[0080] Step 4.5: The interpolated trajectory points (I) j V j ),j=1,2,......J, mapped to 2N×2N grid cells.

[0081] In Step 5 above, the binarized VI trajectory is processed using HSV color encoding. The trajectory direction, power factor, and average pixel value over M periods are integrated into the H, S, and V channels, respectively. The specific method is as follows:

[0082] Step 5.1: To obtain the motion information of the VI trajectory, the hue H is used to represent the direction of the trajectory, and denoted as:

[0083]

[0084] In the formula: the function arg is the arctangent of the four quadrants, calculates the phase angle between two consecutive points in the trajectory, and the obtained angle ranges from 0° to 360°. Then, the hue of the j-th point is stored in a new 2N×2N matrix.

[0085]

[0086] In the formula: A = {a | passing through each point of the grid (nI)} j ,nV j )};

[0087] Step 5.2: When the load current is divided into active and reactive components, using only the reactive current to plot the VI trajectory will result in the loss of information regarding the relationship between the active and reactive components. Therefore, the saturation degree S is used to represent the ratio of active power to reactive power over multiple cycles, i.e., the power factor.

[0088]

[0089] In the formula: K is the total number of sampling points, P apparent For apparent power, V rms I rms These are the effective values ​​of the load voltage and load current, respectively.

[0090] Step 5.3: Brightness V, another color-generating attribute of the HSV color space, is also used to represent the repeatability of the VI trajectory; a binary image W with M cycles is generated using voltage and current waveforms. m (m=1,2,...,M), the brightness V is obtained by averaging the pixel values:

[0091]

[0092] The specific method for constructing the ResNet50 neural network in Step 6 above is as follows:

[0093] Step 6.1: Construct a ResNet50 neural network, consisting of 1 + 3 × (3 + 4 + 6 + 3) = 49 convolutional layers, plus the final fully connected layer, for a total of 50 layers. The ResNet50 network structure can be divided into seven parts. The first part does not contain residual blocks and mainly performs convolution, regularization, activation function, and max pooling calculations on the input. The second, third, fourth, and fifth parts all contain residual blocks. The 1×1 convolutional kernels in each part do not change the size of the residual blocks, but only change the dimensions of the residual blocks.

[0094] Step 6.2: To adapt to the actual number of classifications in the load recognition task, the number of neurons in the last fully connected layer was adjusted from 1000 to 11, which is the actual number of device categories.

[0095] Step 6.3: When training the ResNet50 model, set the initial learning rate to 10. -4 During training, if the loss value does not decrease within 10 epochs, the learning rate will be lowered to optimize the model. At the same time, to adapt to the computer processor and memory resources, the batch size during training will be configured to 8, and the number of model iterations will be configured to 300.

[0096] The sampling frequency of the data obtained in Step 1 above needs to be 30kHz, that is, 30,000 points are sampled per second. At the same time, the power frequency is 60Hz, that is, there are 60 cycles in 1 second, and one cycle contains 30,000 / 60 = 500 sampling points.

[0097] In Step 2 above, an event detection algorithm is used to determine whether the electrical appliance has undergone a state change, extract its steady-state current and voltage data, and determine whether the electrical appliance is operating in a steady-state phase by observing whether the voltage and current trajectories in several consecutive cycles overlap.

[0098] Example:

[0099] The present invention will be explained in detail with reference to the PLAID dataset and accompanying figures. The specific implementation steps are as follows: S1: In this embodiment, the PLAID public dataset is selected, with an electrical frequency of 60Hz and a sampling frequency of 30kHz. This dataset records the electricity consumption data of 55 households in Pennsylvania, USA, which includes 1074 sets of voltage and current data for 11 different electrical appliances. The structure of the PLAID dataset is shown in Table 1 below:

[0100] Table 1 PLAID Dataset Composition Information Table

[0101]

[0102] S2: As Figure 2 An event detection algorithm is used to determine whether an appliance has undergone a state change. Once the appliance reaches a steady state, its steady-state current and voltage data are extracted. Based on Fryze theory, the current is decomposed, and the decomposition results are shown below. Figure 3 .

[0103] Specifically, the following steps are included:

[0104] (1) According to Fryze theory, the load current i(t) can be decomposed into active current and reactive current:

[0105] i(t) = i a (t)+i r (t) (1)

[0106] In the formula, i a (t) represents the active current, i r (t) Reactive current; in addition, active current is defined as the orthogonal projection of the load current onto the voltage v(t) in the direction of the voltage, i.e., i a (t) is proportional to the voltage, thus transmitting information about the resistance;

[0107] (2) Calculate the magnitude of the active current based on the active power:

[0108]

[0109]

[0110] In the formula, P active For active power, V rms Where is the effective value of the voltage, and T is the power supply cycle;

[0111] (3) Finally, calculate the magnitude of the reactive current:

[0112]

[0113] Fryze's power theory decomposes the load current into active and reactive components that represent resistive and non-resistive information, respectively, enhancing the uniqueness of the VI trajectory.

[0114] S3: Plot the VI trajectory image using voltage and reactive current and perform binary trajectory mapping on it:

[0115] Taking the preprocessed VI trajectory as an example, the detailed algorithm for mapping the binary trajectory is as follows:

[0116] (1) Obtain the waveforms of voltage v and reactive current i under steady-state load conditions at a high sampling rate;

[0117] (2) Normalize the voltage and reactive current using the maximum absolute values ​​of the voltage and current waveforms:

[0118] i max =max(|i|), v max =max(|v|) (5)

[0119] (3) Divide the two-dimensional plane of voltage and reactive current into a grid of cells with uniform dimensions. Calculate the length and width of each cell:

[0120]

[0121] In the formula: N∈Z + Thus, the 2N×2N elements in the matrix are obtained;

[0122] (4) Interpolate the original trajectory to ensure the continuity of the curve after mapping the binary trajectory; this is because the distance D between adjacent points k Sometimes it is larger than the length or width of a single cell:

[0123]

[0124] In the formula: v k and i k These are the k-th data points of v and i, respectively;

[0125] (5) The interpolated trajectory points (I) j V j ),j=1,2,......J, mapped to 2N×2N grid cells.

[0126] Binary trajectory mapping can not only transfer the trajectory into a matrix without sacrificing computational complexity, but also mitigate small fluctuations in equipment load, facilitating subsequent operations. Therefore, this technique can be used to map the voltage-reactive current trajectory diagram into a binary image for easier subsequent processing.

[0127] S4: Perform HSV color encoding on the binary trajectory image:

[0128] (1) To obtain the motion information of the VI trajectory, we use the hue H to represent the direction of the trajectory, and denote it as:

[0129]

[0130] In the formula: the function arg is the arctangent of the four quadrants, which calculates the phase angle between two consecutive points in the trajectory, and the resulting angle ranges from 0° to 360°; then, the hue of the j-th point is stored in a new 2N×2N matrix;

[0131]

[0132] In the formula: A = {a | passing through each point of the grid (nI)} j ,nV j )} and |.| are the cardinal numbers of the sets;

[0133] (2) When the load current is divided into active and reactive components, only the reactive current is used to plot the VI trajectory, which will result in the loss of information about the relationship between the active and reactive components; therefore, we suggest using the saturation S to represent the ratio of active power to reactive power over multiple cycles, i.e., the power factor.

[0134]

[0135] In the formula: K is the total number of sampling points, P apparent For apparent power, V rms I rms These are the effective values ​​of the load voltage and load current, respectively.

[0136] (3) Another color-creating attribute of the HSV color space, brightness V, is also used to represent the repeatability of the VI trajectory; a binary image W with M cycles is generated using voltage and current waveforms. m (m=1,2,...,M), the brightness V is obtained by averaging the pixel values;

[0137]

[0138] To create a color image that is perceptible to humans, matrices H, S, and V are concatenated along the third dimension, so that hue and saturation values ​​can be converted into equivalent red, green, and blue, such as... Figure 4 As shown;

[0139] S5: Repeat the above S1-S4 operations on 1074 CSV files of 11 types of electrical appliances, and the output will be generated by... Figure 5 The binarized VI trajectory color-coded images of 11 electrical appliances, including air conditioners, energy-saving lamps, fans, refrigerators, hair dryers, heaters, incandescent lamps, laptops, microwave ovens, vacuum cleaners, and washing machines, obtained through color encoding, are shown below. These images are used as input to a ResNet50 neural network, which is then used for feature extraction and classification. The ResNet50 neural network model parameters are shown in Table 2 below.

[0140] Table 2 ResNet50 Neural Network Model Parameters

[0141]

[0142] The specific meanings of each module in Table 2 are as follows:

[0143] (1) Input layer: Accepts input images or features;

[0144] (2) Convolutional and pooling layers: A series of convolutional and pooling layers are used for feature extraction and dimensionality reduction;

[0145] (3) Residual block: The core component in the basic module, which introduces cross-layer connections and residual learning to solve the problems of gradient vanishing and gradient explosion;

[0146] (4) Global average pooling layer: converts the final feature map into a one-dimensional vector;

[0147] (5) Fully connected layer: The output layer used for classification or other tasks.

[0148] In summary, ResNet50 employs a residual network structure, enabling deeper and more efficient training through cross-layer connections and residual learning, avoiding the gradient problem inherent in deep network training. Due to its excellent performance in computer vision tasks, relatively simple structure, and high accuracy and performance, it has been widely used in practical applications. However, when transferring it to non-invasive load recognition tasks, it's important to note that the classic ResNet50 model was originally designed for classification tasks on the ImageNet dataset, which contains images of 1000 different categories. Therefore, the output dimension of the last fully connected layer is 1000, corresponding to each category in the ImageNet dataset. Each neuron corresponds to one category and outputs the probability of that category, making it unsuitable for direct application to load recognition tasks. Therefore, this invention improves the model by reducing the number of neurons in the last fully connected layer from 1000 to 11 to accommodate the actual number of categories in load recognition tasks. Finally, the ResNet50 network is used to classify images, achieving the recognition of 11 different devices.

Claims

1. A non-intrusive load recognition method based on binarized VI trajectory color coding, characterized in that, Includes the following steps: Step 1: Use high-frequency sampling equipment to acquire the voltage and current data of the home bus in real time; Step 2: Assuming that only one device is turned on or off at a time, use an event detection algorithm to determine whether the device has changed its state. Once the device's operating state reaches a steady state, extract its steady-state current and voltage data. Step 3: Preprocess the original current waveform according to Fryze power theory, and decompose the load current into active current and reactive current. Step 4: Normalize the voltage and reactive current data obtained in Step 2 and 3, plot the VI trajectory curve, and perform binary trajectory mapping on it to enhance the uniqueness of the VI trajectory. Step 4.1: Obtain the steady-state voltage of the load at a high sampling rate. and reactive current The waveform; Step 4.2: Normalize the voltage and reactive current using the maximum absolute values ​​of the voltage and current waveforms: ； Step 4.3: Divide the two-dimensional plane of voltage and reactive current into a uniformly sized cell grid, and calculate the length and width of each cell: ； In the formula: Thus, the matrix is ​​obtained element; Step 4.4: Interpolate the original trajectory to ensure the continuity of the curve after mapping the binary trajectory. This is because the distance between adjacent points... Sometimes it is larger than the length or width of a single cell: ； In the formula: and They are and The kth data point; Step 4.5: The interpolated trajectory points Mapped to Grid cells; Step 5: Perform HSV color encoding on the binary VI trajectory obtained in Step 4 to obtain a binary color-encoded VI trajectory image; The binarized VI trajectory is processed using HSV color encoding, and the trajectory direction, power factor, and average pixel value over M periods are integrated into the H, S, and V channels respectively. The specific method is as follows: Step 5.1: To obtain the motion information of the VI trajectory, the hue H is used to represent the direction of the trajectory, and denoted as: ； In the formula: the function arg is the arctangent in the four quadrants, which calculates the phase angle between two consecutive points in the trajectory, with the obtained angle ranging from 0° to 360°. Then, the hue of the j-th point is stored in a new... In the matrix; ； In the formula: ; Step 5.2: When the load current is divided into active and reactive components, using only the reactive current to plot the VI trajectory will result in the loss of information about the relationship between the active and reactive components. Therefore, the saturation degree S is used to represent the ratio of active power to reactive power over multiple cycles, i.e., the power factor. ； In the formula: K is the total number of sampling points, Apparent power, , These are the effective values ​​of the load voltage and load current, respectively. Step 5.3: Brightness (V), another color-generating attribute of the HSV color space, is also used to represent the repeatability of the VI trajectory; a binary image with M cycles is generated using voltage and current waveforms. The brightness V is obtained by averaging its pixel values. ； Step 6: Construct a ResNet50 neural network model, using the binarized color VI trajectory image as input to classify the devices, thus completing non-invasive load identification.

2. The non-intrusive load recognition method based on binary VI trajectory color encoding according to claim 1, characterized in that, In Step 2, when extracting high-frequency data from the device, the voltage waveform is collected starting from the zero-crossing period, and the voltage and current data within one steady-state period are collected.

3. The non-intrusive load recognition method based on binary VI trajectory color encoding according to claim 2, characterized in that, The process of decomposing the load current using Fryze power theory in Step 3 is as follows: ； ； ； ； in, Active current; Reactive current; Active power; The voltage is the effective value; T is the power supply cycle. The voltage at time t is the steady-state high-frequency sampling data. The current at time t represents steady-state high-frequency sampling data.

4. The non-intrusive load recognition method based on binary VI trajectory color encoding according to claim 3, characterized in that, The specific method for constructing the ResNet50 neural network in Step 6 is as follows: Step 6.1: Construct a ResNet50 neural network, consisting of 1 + 3 × (3 + 4 + 6 + 3) = 49 convolutional layers, plus the final fully connected layer, for a total of 50 layers. The ResNet50 network structure can be divided into seven parts. The first part does not contain residual blocks and mainly performs convolution, regularization, activation function, and max pooling calculations on the input. The second, third, fourth, and fifth parts all contain residual blocks. The 1×1 convolutional kernels in each part do not change the size of the residual blocks, but only change the dimensions of the residual blocks. Step 6.2: To adapt to the actual number of classifications in the load recognition task, the number of neurons in the last fully connected layer was adjusted from 1000 to 11, which is the actual number of device categories. Step 6.3: When training the ResNet50 model, set the initial learning rate to 10. -4 During training, if the loss value does not decrease within 10 epochs, the learning rate will be lowered to optimize the model. At the same time, to adapt to the computer processor and memory resources, the batch size during training will be configured to 8, and the number of model iterations will be configured to 300.

5. A non-intrusive load recognition method based on binary VI trajectory color coding according to claim 4, characterized in that, The sampling frequency of the data obtained in Step 1 needs to be 30kHz, that is, 30,000 points are sampled per second, and the power frequency is 60Hz, that is, there are 60 cycles in 1 second, and one cycle contains 30,000 / 60=500 sampling points.

6. The non-intrusive load recognition method based on binary VI trajectory color encoding according to claim 5, characterized in that, In Step 2, an event detection algorithm is used to determine whether the electrical appliance has undergone a state change, and its steady-state current and voltage data are extracted. By observing whether the voltage and current trajectories in several consecutive cycles overlap, it is determined whether the electrical appliance is operating in a steady-state phase.

Images