A load identification method based on convolutional neural network with improved VI trajectory

Abstract

A convolutional neural network load identification method based on an improved VI trajectory relates to the technical field of power load identification. Based on the acquisition of an independent load current waveform, the present invention combines the voltage waveform to improve the traditional single-cycle VI trajectory into a multi-cycle VI trajectory, while introducing the average value characteristics of instantaneous power, current change rate and multi-cycle current effective value, to achieve dual color coding improvement of the VI trajectory and its background, so as to enhance the load feature characterization capability contained in the VI trajectory. The improved VI trajectory is used as the input of the VGGNet network framework in CNN to achieve load identification. In view of the influence of the VGGNet network hyperparameters on the recognition accuracy, the particle swarm optimization PSO algorithm is selected to optimize the hyperparameters during the network training process to improve the load recognition accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of power load identification, and in particular to a load identification method based on a convolutional neural network with an improved VI trajectory. Background Art

[0002] With the development of smart grids, the traditional power industry is shifting toward a highly intensive, knowledge-based, and technologically advanced industry. The power grid has become one of the world's largest and most complex machines. At the same time, non-intrusive load monitoring has garnered widespread attention from both industry and academia, and has become a research hotspot for the application of artificial intelligence technology in related industries such as construction and power. While effective application of non-intrusive load monitoring technology can significantly reduce energy consumption and improve energy efficiency, it still suffers from a low load recognition rate. Summary of the Invention

[0003] The purpose of the present invention is to provide a load identification method based on a convolutional neural network with an improved VI trajectory, so as to solve the problem of low load identification rate.

[0004] The technical solution adopted by the present invention to solve the technical problem is: a load identification method based on a convolutional neural network with an improved VI trajectory, comprising the following steps:

[0005] S1. Improve VI trajectory

[0006] S1.1. Select 25 consecutive cycles of instantaneous voltage and current signals and draw the VI trajectory;

[0007] S1.2. Color-code the VI trajectory: use red to represent the multi-cycle VI trajectory, i.e., the R channel; use green to represent the slope of the straight line segment between adjacent acquisition points of the VI trajectory, i.e., the G channel; and use blue to represent the instantaneous power value, i.e., the B channel, thereby generating a color-coded VI trajectory image;

[0008] S1.3. Design color mapping based on the range of voltage and current values. You can use gradient colors to represent changes in voltage or current.

[0009] S2. Construct a load identification model based on PSO-CNN

[0010] S2.1 CNN-based load identification model

[0011] The VGGNet network, a typical CNN network model, is selected as the load recognition model. The VGGNet network contains 16 convolutional layers, 5 pooling layers, and 3 fully connected layers. The ReLU function is used as the activation function of the convolutional neural network. The VI trajectory feature map is sent to the input layer, that is, the first convolutional layer. The convolutional layer extracts local features and sends them to the pooling layer for feature compression. The fully connected layer integrates the features. The number of units in the last fully connected layer is set according to the number of load types, and the corresponding load category is output. S2.2 Particle Swarm Optimization (PSO) process:

[0012] S2.2.1. For the VGG network, choose an optimization objective that improves validation set accuracy.

[0013] S2.2.2.2. Select the number of convolution kernels, stride, pooling kernel size, and stride as the parameters to be optimized.

[0014] S2.2.3. Initialize the particle swarm and calculate the fitness of each particle, i.e., the hyperparameter combination. Update the position and velocity of the particle based on the fitness. Update the position and velocity of each particle using the following formula:

[0015] v i+1 =v i +c1r1(pbest i -x i )+c2r2(gbest i -x i ) (1)

[0016] x i+1 =x i +v i (2)

[0017] Where i = 1, 2, ..., N, N is the total number of particles in the group; v i is the velocity of the particle; v i The maximum value of v max >0, if v i >v max , then v i =v max ; r1, r2 are random numbers between (0, 1); x i is the current position of the particle; c1 and c2 are learning factors, usually c1=c2=2; pbest i is the best known position of particle i; gbest i is the global optimal position;

[0018] For each particle position x i+1, corresponds to a specific hyperparameter combination, uses this combination to train the VGG network, and updates the particle pbest according to the evaluation index, while updating the global best position gbest;

[0019] S2.3. Particle swarm optimization of CNN hyperparameters:

[0020] First, the fitness function and objective function are determined. In each iterative cycle, a particle is selected and evaluated and optimized in detail using the fitness function. The evaluation point with the greatest potential identified in each iteration is added to the historical dataset. This process continues until the preset termination condition is met. The parameter set to be optimized X, the objective function f, and the fitness function S are used as the input of the particle swarm optimization algorithm to output the optimal VGGNet network model parameters.

[0021] The specific content of step S1.1 is:

[0022] Normalized voltage and current, the formula is as follows:

[0023]

[0024] Where max|v| is the maximum value of instantaneous voltage in a cycle; max|i| is the maximum value of instantaneous current in multiple cycles; v k and i k They represent the voltage and current values ​​of the kth acquisition point in the sequence, Δv k and Δi k are the normalized voltage and current values ​​of the kth acquisition point respectively.

[0025] A VI trajectory is created using the normalized data, and the resulting VI trajectory serves as the R channel of the color-coded VI trajectory.

[0026] Use the arctan function to map the slope of the straight line segment to the range of (0, 1) to complete the creation of the G channel;

[0027]

[0028] Among them, K j is the slope of the jth straight line segment; G j is the G channel depth value of the jth straight line segment; i j Indicates the current value of the current data point; i j+1 Indicates the current value of the next data point; v j Indicates the voltage value of the current data point; v j+1 Indicates the voltage value of the next data point; max|v| is the absolute value of the maximum voltage extracted from the data plotting the VI trajectory; max|i| is the absolute value of the maximum current extracted from the data plotting the VI trajectory;

[0029] G j Mapped to the VI track, the G channel depth value G corresponding to each grid is obtained m,n , and then normalize it to get the value of the G channel of each grid point G m,n ';

[0030] max G=max{G 1,1 ,G 1,2 ,G 1,3 ,…,G n,n} (7)

[0031]

[0032] Using instantaneous power value P j Create the B channel:

[0033] p j =i j ·v j (9)

[0034] The instantaneous power value p of the multi-cycle j Mapped to the VI track, the depth value of each grid is p m,n ,Then the instantaneous power of multiple cycles is normalized;

[0035] max p=max{p 1,1 ,p 1,2 ,p 1,3 ,…,p n,n} (10)

[0036]

[0037] Among them, p m,n is the power superposition value of each grid in the grid; maxp is the maximum power of all grids; p m,n ' is the value of the B channel.

[0038] The background color is determined by first calculating the average of the RMS current energy over multiple adjacent cycles. Second, the current amplitude is mapped to a predefined color range and the color value is calculated using linear interpolation within the specified interval.

[0039] The color coding gradually transitions from gray to blue, matching the color to the current amplitude.

[0040] VGGNet adopts a deep and stacked convolutional layer structure, uses small convolution kernels and pooling layers to improve feature extraction capabilities and position invariance, and applies weight decay and data augmentation techniques to control model complexity and enhance generalization capabilities. It uses images with dual color coding based on background and VI trajectory as model input to complete recognition tasks.

[0041] The convolution layer consists of multiple filters, each of which performs a convolution operation with the input data to produce a series of feature maps, thereby capturing the local characteristics of the input data; the definition of the convolution operation is expressed by Formula 12-Formula 15:

[0042]

[0043]

[0044] Among them, X is the source feature map; K is the convolution kernel; Y is the new feature map; Y=X*K.

[0045] The VGGNet network sets a pooling layer after each convolution layer.

[0046] The beneficial effects of the present invention are as follows: the present invention proposes a convolutional neural network load identification method based on an improved VI trajectory, which first generates VI trajectory data through real-time monitoring of voltage and current signals. After preprocessing, these data extract time domain and frequency domain features, which enhances the network's ability to identify load characteristics. By utilizing the hierarchical structure of the convolutional neural network, the subtle features of load changes can be effectively captured, thereby achieving high-accuracy load identification. The real-time monitoring and feedback of non-invasive load monitoring technology enable users to quickly obtain energy consumption data and optimize electricity consumption behavior; it can accurately identify different types of electrical equipment and their usage status, helping users understand the energy consumption characteristics of each device; it can monitor abnormal energy consumption patterns, detect equipment failures or potential problems in advance, and avoid major losses; after integration with the smart home system, the technology realizes automated control, improves electricity efficiency and convenience; through precise monitoring and analysis, users can better utilize renewable energy such as solar energy and promote the application of green energy; data visualization enhances users' energy-saving awareness and encourages them to take effective energy-saving measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is the active power and reactive power diagram of the electric kettle;

[0048] Figure 2 This is the power change diagram of the inductive load from startup to stable operation;

[0049] Figure 3 Identify the block diagram for the load;

[0050] Figure 4This is the VGGNet network structure diagram;

[0051] Figure 5 It is the convolution process diagram;

[0052] Figure 6 Flowchart of optimizing a VGGNet network for particle swarm optimization. DETAILED DESCRIPTION

[0053] Based on the independent load current waveform, the present invention combines the voltage waveform to improve the traditional single-cycle VI trajectory into a multi-cycle VI trajectory. At the same time, the average value characteristics of instantaneous power, current change rate and multi-cycle current effective value are introduced to achieve dual color coding improvement of the VI trajectory and its background to enhance the load feature characterization capability contained in the VI trajectory. The improved VI trajectory is used as the input of the VGGNet (Visual Geometry Group Net) network framework in the CNN to achieve load recognition. In view of the influence of the VGGNet network hyperparameters on the recognition accuracy, the particle swarm optimization (PSO) algorithm is selected to optimize the hyperparameters during the network training process to improve the load recognition accuracy of the model.

[0054] The load characteristics are analyzed as follows:

[0055] The accuracy of load identification is closely related to the electrical characteristics of the load during operation. Selecting appropriate characteristics can improve identification. By analyzing load types and characteristics, a load characteristic "dictionary" can be constructed. These characteristics are generally divided into two categories: steady-state and transient. Steady-state characteristics include electrical properties such as power, current, and harmonics during stable operation. Transient characteristics involve electrical changes that occur during the transition from startup to stable operation, or from stable operation to shutdown, such as power fluctuations and transient times.

[0056] Active power and reactive power in steady-state load are commonly used features in load identification problems. The calculation formulas for active power P and reactive power Q are as follows:

[0057] P=VIcos(φ) (1)

[0058] Q=VIsin(φ) (2)

[0059] Where V is voltage, I is current, and φ is the phase angle.

[0060] Pure resistive loads mainly show rapid changes in active power during the start and stop process, and can quickly reach stable power when the switch is turned on. At the same time, this type of load consumes almost no reactive power, such as Figure 1 shown.

[0061] When an inductive load starts, there will usually be a synchronous step change in active power and reactive power, and the reactive power will continue to exist after starting. Figure 2 As shown in the figure, the changes in active and reactive power from startup to stable operation of an inductive load indicate that the influence of inductive and capacitive components during startup causes a brief time delay. As the time delay ends, the power experiences a second significant step increase, eventually reaching a steady state. During this steady state, the power exhibits some fluctuation.

[0062] When analyzing steady-state power characteristics, special attention should be paid to the following aspects: whether there is reactive power consumption; the amplitude and rate of change of active power and reactive power during the switching process of the equipment; and the stability of the power characteristics and whether there are obvious fluctuations during the stable operation stage of the load.

[0063] Different types of equipment have significant differences in their current characteristics. Therefore, when analyzing current characteristics, the current amplitude, volatility, range of variation, and whether there are regular periodic changes are important criteria for identifying and distinguishing different devices.

[0064] The characteristic curve drawn with instantaneous voltage u(t) as the horizontal coordinate and instantaneous current i(t) as the vertical coordinate within the same power frequency cycle is called the VI trajectory. Loads with different working principles have obvious differences in the shape of their VI trajectories. Therefore, the VI trajectory is often used for load identification. However, there are also cases where some load types are different but their VI trajectories are relatively similar. In order to reduce the impact of VI trajectory similarity on load identification accuracy, the present invention has made improvements in the following three aspects: the energy loss problem of VI trajectory; the difference in instantaneous power changes between different loads; and the difference in current change rate between different loads.

[0065] During the process from startup to stable operation, various loads exhibit different transient characteristics, such as power fluctuation and duration, due to differences in working principles and internal components.

[0066] 1. Improve VI trajectory

[0067] Compared with other single features, the VI trajectory covers a variety of operating characteristics by combining the changes in voltage and current during the operation of household loads, thereby having a higher degree of discrimination in load identification and outperforming traditional features. In addition, the VI trajectory can be effectively combined with computer vision and non-invasive load monitoring (NILM) tasks. Other features usually require a combination and difference analysis of multiple dimensions, while the VI trajectory transforms the problem into an image recognition task and uses mature image processing technology to simplify the multi-signal processing process. Therefore, the present invention selects the VI trajectory as the key feature for load identification and improves it to enhance the amount of information it contains.

[0068] Traditional VI trajectory analysis methods map VI curves onto a grid and use a binary value of 0 or 1 to represent the presence or absence of VI trajectory information within each cell, thereby transforming the load identification task into an image recognition problem. A limitation of this method is that the binary representation cannot accurately reflect the details of the VI trajectory, resulting in low accuracy in expressing load information and difficulty capturing more subtle changes and features. Therefore, this paper proposes an improved method that aims to improve the feature expressiveness and recognition accuracy of VI trajectories through a richer information representation format, thereby overcoming the shortcomings of traditional methods.

[0069] (1) Traditional VI trajectory analysis methods typically focus only on voltage and current changes within a single cycle and cannot fully capture the load fluctuation characteristics and change patterns over multiple cycles. To address this issue, the present invention uses 25 consecutive cycles of instantaneous voltage and current signals to plot the VI trajectory, aiming to more comprehensively reflect the load changes over different cycles.

[0070] Specifically, the voltage and current signals are first standardized, and then the VI trajectory is drawn based on the standardized data to more accurately express the dynamic characteristics of the load.

[0071] The normalization formula is as follows:

[0072]

[0073] Where max|v| is the maximum value of instantaneous voltage in multiple cycles; max|i| is the maximum value of instantaneous current in multiple cycles; v m and i m Respectively represent the voltage and current values ​​of the mth acquisition point in the sequence, Δv m and Δi m are the normalized voltage and current values ​​of the mth acquisition point respectively.

[0074] (2) The process of color coding the VI trajectory is to more intuitively display the dynamic characteristics and change patterns of the load. In this method, different color channels are used to represent different characteristic information: the red channel (R) is used to represent the changes in the multi-cycle VI trajectory, reflecting the overall trend and characteristics of the load in multiple cycles; the green channel (G) is used to represent the slope of the straight line segment between adjacent sampling points of the VI trajectory, revealing the volatility and rate of the load during the instantaneous change process; the blue channel (B) is used to represent the instantaneous power value, intuitively showing the power level of the load at each moment.

[0075] Specifically, the voltage and current data are first normalized, and then the VI trajectory is plotted using this standardized data. The resulting VI trajectory data is used as the red channel of the image to display the characteristics of multi-cycle variations. Through this color-coding method, the VI trajectory image can more clearly convey the details of load changes in different cycles, instantaneous power fluctuations, and change rates, thereby improving data visualization and feature expression, and contributing to the accuracy and efficiency of load identification tasks.

[0076] Use the arctan function to map the slope of the straight line segment to the range of (0, 1) to complete the creation of the G channel;

[0077]

[0078] Among them, K j is the slope of the jth straight line segment; G j is the G channel depth value of the jth straight line segment; i j Indicates the current value of the current data point; i j+1 Indicates the current value of the next data point; v j Indicates the voltage value of the current data point; v j+1 Indicates the voltage value of the next data point; max|v| is the absolute value of the maximum voltage extracted from the data plotting the VI trajectory; max|i| is the absolute value of the maximum current extracted from the data plotting the VI trajectory.

[0079] G j Mapped to the VI track, the G channel depth value G corresponding to each grid is obtained m,n , and then normalize it to get the value of the G channel of each grid point G m,n ';

[0080] max G=max{G 1,1 ,G 1,2 ,G 1,3 ,…,G n,n} (7)

[0081]

[0082] The following instantaneous power value p is used j Create the B channel:

[0083] p j =i j ·v j (9)

[0084] The multi-cycle instantaneous power value p j Mapped to the VI track, the depth value of each grid is p m,n,Afterwards, the multi-cycle power is normalized.

[0085] max p=max{p 1,1 ,p 1,2 ,p 1,3 ,…,p n,n} (10)

[0086]

[0087] Among them, p m,n is the power superposition value of each grid in the grid; maxp is the maximum power of all grids; p m,n ' is the value of the B channel.

[0088] (3) Considering that different loads may show similar VI traces but have significant differences in current amplitude, which is a key characteristic for distinguishing load types, we color-code the VI trace background according to the load current amplitude to highlight the differences in current amplitude. The specific method is as follows:

[0089] First, the average RMS energy of the current signal is calculated over 25 consecutive cycles. Next, the current amplitude is mapped to a predefined color range, and linear interpolation is used to assign colors to different amplitude intervals. The color coding gradually transitions from gray to blue, covering the range from the lowest to the highest current amplitudes. In this way, the changes in background color reflect the differences in current amplitude, enhancing the correspondence between the VI trace image and the load current characteristics.

[0090] 2. Building a load identification model based on PSO-CNN

[0091] Convolutional neural network (CNN) is used as a load identification model to identify VI trajectories. However, due to the large number of hyperparameters in CNN, the selection of hyperparameters will have a significant impact on the load identification results, and manual adjustment of hyperparameters is cumbersome. Particle swarm optimization (PSO) has high global search capabilities and strong adaptability to complex nonlinear hyperparameter spaces when optimizing VGG networks. PSO searches for the global optimal solution through multiple particles in parallel. Compared with traditional optimization methods, it is easier to implement and adjust. It is suitable for high-dimensional parameter spaces and multi-objective optimization tasks, and can effectively improve the performance and generalization ability of deep learning models. Therefore, using the PSO algorithm to optimize the hyperparameters in CNN can improve the accuracy of load identification. Specifically:

[0092] 2.1 CNN-based load identification model

[0093] The VI trajectory features are converted into images, and then the VI trajectory images are classified using CNN, which is commonly used in image processing, to identify various load types. In the CNN load recognition model, Figure 3 As shown in the figure, the VI trajectory diagrams of different load types are first classified and labeled, and then supervised training is performed on the VI trajectories of various loads. Through the iterative and optimization process, the differences between the VI trajectory of each load and the VI trajectories of other loads are identified.

[0094] The CNN model mainly consists of an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The present invention selects the typical CNN network model - VGGNet network as the load recognition model. The model structure is as follows Figure 4 As shown, the VGGNet network consists of 16 convolutional layers, 5 pooling layers, and 3 fully connected layers. The network uses the ReLU function as the activation function of the convolutional neural network, which is more effective than the Sigmoid function and solves the overfitting problem. In addition, VGGNet adopts a deep and stacked convolutional layer structure, uses small convolution kernels and pooling layers to improve feature extraction capabilities and position invariance, and applies weight decay and data augmentation techniques to control model complexity and enhance generalization capabilities. It uses images with dual color coding based on background and VI trajectory as model input to complete recognition tasks.

[0095] The convolution layer is one of the core layers of VGGNet. It consists of multiple filters. Each filter performs a convolution operation with the input data to generate a series of feature maps, thereby capturing the local features of the input data. Figure 5 This demonstrates how a convolution kernel slides across the original feature map to extract features and generate a new feature map. Depending on specific needs, the desired feature map data can be obtained by adjusting the convolution kernel parameters. The definition of the convolution operation can be expressed as Equations 12-15. The size and stride of the convolution kernel affect the receptive field of pixels in the feature map, which in turn affects the model's feature extraction performance. Furthermore, the number of convolution kernels affects the network's ability to analyze input data from multiple perspectives. Excessive numbers can reduce the network's recognition capabilities and increase computational requirements.

[0096]

[0097] Among them, X is the source feature map; K is the convolution kernel; Y is the new feature map; Y=X*K.

[0098] The VGGNet network has a pooling layer after each convolutional layer. The main function of the pooling layer is to reduce the spatial size of the feature map by downsampling, filtering out unnecessary samples, thereby retaining key information, reducing model parameters, preventing overfitting, and improving computational efficiency. Setting the appropriate pooling kernel size and stride can reduce the computational complexity of the data while preserving data information.

[0099] During network training, randomly set dropout probabilities may affect the recognition performance of the model. The fully connected layer is located at the end of the VGGNet network. The last fully connected layer passes the output to an m-class softmax classifier to perform multi-classification tasks for m different load categories in the dataset.

[0100] In summary, the CNN-based load identification process can be summarized as follows: the VI trajectory feature map is fed into the input layer (i.e., the first convolution layer), the convolution layer extracts local features and feeds them into the pooling layer for feature compression, the fully connected layer integrates the features, and the number of fully connected units in the final layer is set based on the number of load types, and the corresponding load category is output. During the load identification process, a "load not turned on" output state is added. That is, based on the number of load types n, the system sets n+1 output units. If the similarity between the VI trajectory to be identified and the VI trajectory feature library is less than 0.5, the system will output the "load not turned on" state instead of the specific load type.

[0101] 2.2 Particle Swarm Optimization PSO Process:

[0102] First, we need to clearly define the optimization goal. For the VGG network, we choose an optimization goal that improves the accuracy of the validation set.

[0103] Secondly, determine the hyperparameters to be optimized, and select the number of convolution kernels, step size, pooling kernel size and step size as the parameters to be optimized.

[0104] Next, the particle swarm is initialized, and for each particle, that is, the hyperparameter combination, its fitness is calculated. In deep learning, fitness is typically evaluated by training a neural network and calculating the accuracy on a validation set. The position and velocity of the particle are updated based on the fitness. The PSO algorithm updates the position and velocity of each particle using the following formula:

[0105] v i+1 =v i +c1r1(pbest i -x i )+c2r2(gbest i -x i ) (16)

[0106] x i+1 =x i +v i (17)

[0107] Where i = 1, 2, ..., N, N is the total number of particles in the group; v i is the velocity of the particle; v i The maximum value of v max >0, if v i>v max , then v i =v max ; r1, r2 are random numbers between (0, 1); x i is the current position of the particle; c1 and c2 are learning factors, usually c1=c2=2; pbest i is the best known position of particle i; gbest i is the global optimal position;

[0108] For each particle position x i+1 , corresponding to a specific hyperparameter combination, the VGG network is trained using this combination, and the particle's pbest is updated based on the evaluation metric, while the global best position gbest is also updated. PSO iterations can be terminated with a maximum number of iterations or after reaching a satisfactory fitness value. Once the termination condition is met, the hyperparameter combination corresponding to the global best particle is selected as the final optimization result. The obtained optimal hyperparameters are applied to the actual training and use of the VGG network to obtain better performance or efficiency. In this way, PSO can effectively search the hyperparameter space of the VGG network, help optimize the network structure, and improve its performance and generalization ability.

[0109] Through these steps, the PSO algorithm can effectively search and optimize the hyperparameter configuration of the VGG network, improving the performance and generalization ability of the load identification model. The advantage of the PSO algorithm lies in its global search capability, which helps avoid getting stuck in local optimal solutions and quickly converge to a better solution.

[0110] 2.3 Particle Swarm Optimization CNN Hyperparameter Algorithm:

[0111] Regarding the analysis of the influence of hyperparameters on the load recognition effect, the number of convolution kernels, step size, pooling kernel size and step size are selected as the parameters to be optimized.

[0112] In the process of particle swarm optimization of the VGGNet network model to optimize its hyperparameters, such as Figure 6 As shown, the fitness function and objective function are first determined. In each iteration cycle, a particle is selected and evaluated and optimized in detail using the fitness function. The evaluation point with the greatest potential identified in each iteration is added to the historical data set. This process will continue until the preset termination condition is met. The input of the particle swarm optimization algorithm in this invention includes the parameter set to be optimized X, the objective function f, and the fitness function S, and the output is the optimal VGGNet network model parameters.

Claims

1. A load identification method based on a convolutional neural network with improved VI trajectory, characterized in that: The following steps are involved: S1. Improve VI trajectory S1.

1. Select 25 consecutive cycles of instantaneous voltage and current signals and draw the VI trajectory; S1.

2. Color-code the VI trajectory: use red to represent the multi-cycle VI trajectory, i.e., the R channel; use green to represent the slope of the straight line segment between adjacent acquisition points of the VI trajectory, i.e., the G channel; and use blue to represent the instantaneous power value, i.e., the B channel, thereby generating a color-coded VI trajectory image; S1.

3. Design color mapping based on the range of voltage and current values. Gradient colors can be used to represent changes in voltage or current. S2. Construct a load identification model based on PSO-CNN S2.1 CNN-based load identification model The VGGNet network, a typical CNN network model, was selected as the load recognition model. The VGGNet network consists of 16 convolutional layers, 5 pooling layers, and 3 fully connected layers. The ReLU function is used as the activation function of the convolutional neural network. The VI trajectory feature map is sent to the input layer, that is, the first convolutional layer. The convolutional layer extracts local features and sends them to the pooling layer for feature compression. The fully connected layer integrates the features. The number of units in the last fully connected layer is set according to the number of load types, and the corresponding load category is output. S2.2 Particle Swarm Optimization PSO process: S2.2.

1. For the VGG network, choose an optimization objective that improves validation set accuracy. S2.2.2.

2. Select the number of convolution kernels, stride, pooling kernel size, and stride as the parameters to be optimized. S2.2.

3. Initialize the particle swarm and calculate the fitness of each particle, i.e., the hyperparameter combination; Update the position and velocity of the particles according to the fitness, and update the position and velocity of each particle using the following formula: v i+1 =v i +c1 r2 (pbest i -x i )+c2 r2 (gbest i -x i ) (1) x i+1 =x i +v i (2) Where i = 1, 2, ..., N, N is the total number of particles in the group; v i+1 is the updated velocity of the particle; v i is the current velocity of the particle; v i The maximum value of v max >0, if v i >v max , then v i =v max ; r1, r2 are random numbers between (0, 1); x i is the current position of the particle; x i+1 is the updated position of the particle; c1 and c2 are learning factors, usually c1=c2=2; pbest i is the best known position of particle i; gbest i is the global optimal position; For each particle position x i+1 , corresponding to a specific hyperparameter combination, use this combination to train the VGG network, and update the particle pbest according to the evaluation index, and update the global best position gbest at the same time i ; S2.

3. Particle swarm optimization of CNN hyperparameters: First, the fitness function and objective function are determined. In each iterative cycle, a particle is selected and evaluated and optimized in detail using the fitness function. The evaluation point with the greatest potential identified in each iteration is added to the historical dataset. This process continues until the preset termination condition is met. The parameter set to be optimized X, the objective function f, and the fitness function S are used as the input of the particle swarm optimization algorithm to output the optimal VGGNet network model parameters.

2. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 1 is characterized in that: The specific content of step S1.1 is: Normalized voltage and current, the formula is as follows: Where max|v| is the maximum voltage in the data; max|i| is the instantaneous maximum value in the data; v k and i k Respectively represent the voltage and current values ​​of the kth acquisition point in the data, Δv k and Δi k are the normalized voltage and current values ​​of the kth acquisition point respectively.

3. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 2 is characterized in that: A VI trajectory is created using the normalized data, and the resulting VI trajectory serves as the R channel of the color-coded VI trajectory.

4. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 3 is characterized in that: Use the arctan function to map the slope of the straight line segment to the range of (0, 1) to complete the creation of the G channel; Among them, K j is the slope of the jth straight line segment; G j is the G channel depth value of the jth straight line segment; i j Indicates the current value of the current data point; i j+1 Indicates the current value of the next data point; v j Indicates the voltage value of the current data point; v j+1 Indicates the voltage value of the next data point; max|v| is the absolute value of the maximum voltage extracted from the data plotting the VI trajectory; max|i| is the absolute value of the maximum current extracted from the data plotting the VI trajectory; G j Mapped to the VI track, the G channel depth value G corresponding to each grid is obtained m,n , and then normalize it to get the value of the G channel of each grid point G m,n '; maxG=max{G 1,1 ,G 1,2 ,G 1,3 ,…,G n,n } (7) 5. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 4 is characterized in that: Using the instantaneous power value p j Create the B channel: p j =i j v j (9) The instantaneous power value p of the multi-cycle j Mapped to the VI track, the depth value of each grid is p m,n ,Then the instantaneous power of multiple cycles is normalized; maxp=max{p 1,1 ,p 1,2 ,p 1,3 ,…,p n,n } (10) Among them, p m,n is the power superposition value of each grid in the grid; maxp is the maximum power of all grids; p m,n ' is the value of the B channel.

6. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 5 is characterized in that: The background color is determined by first calculating the average of the RMS current energy over 25 consecutive cycles. Second, the current amplitude is mapped to a predefined color range and the color value is calculated using linear interpolation within the specified interval.

7. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 6 is characterized in that: The color coding gradually transitions from gray to blue and is divided into 6 intervals: gray-cyan to green, green to orange, orange to yellow, yellow to red, red to purple and purple to blue, achieving a match between color and current amplitude.

8. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 1 is characterized in that: VGGNet adopts a deep and stacked convolutional layer structure, uses small convolution kernels and pooling layers to improve feature extraction capabilities and position invariance, and applies weight decay and data augmentation techniques to control model complexity and enhance generalization capabilities. It uses images with dual color coding based on background and VI trajectory as model input to complete recognition tasks.

9. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 8 is characterized in that: The convolution layer consists of multiple filters, each of which performs a convolution operation with the input data to produce a series of feature maps, thereby capturing the local characteristics of the input data; the definition of the convolution operation is expressed by Formula 12-Formula 15: Among them, X is the source feature map; K is the convolution kernel; Y is the new feature map; Y=X*K.

10. The load identification method based on a convolutional neural network with improved VI trajectory according to claim 9 is characterized in that: The VGGNet network sets a pooling layer after each convolution layer.

Images