A non-invasive load identification method, device, equipment and storage medium

Abstract

The present invention relates to a non-invasive load identification method, device, equipment, and storage medium; it belongs to the technical field of electricity load identification; the method includes: processing the collected current and voltage data to generate a Markov transition field; each element in the Markov transition field is the state transition probability filled corresponding to the collected current and voltage data arranged in the time domain; the state transition probability is the state transition probability obtained by normalizing and quantile processing the collected current and voltage data and then mapping the obtained normalized current and voltage data to the corresponding quantile unit; performing downsampling and pseudo-color coding on the Markov transition field to obtain an RGB color image; inputting the RGB color image into a constructed neural network model for non-invasive load identification. The present invention is used for load monitoring and management in smart grids, improving the efficiency and reliability of energy use.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical load identification, and in particular to a non-intrusive load identification method, device, equipment and storage medium. Background Art

[0002] Smart grids provide strong technical support for the efficient management and effective utilization of electric energy. As a key link in the research of smart grid technology, load identification provides a complete data basis and technical means for smart grids. Non-Intrusive Load Monitoring (NILM), as an emerging identification technology, only needs to install a monitoring device at the bus, and then it can analyze the types, operating states and energy consumption information of residential loads, reduce the installation and maintenance costs, and minimize the interference with user privacy.

[0003] There are problems of low identification accuracy and large network parameter quantity in non-intrusive load identification, which will lead to the situation that it is easy to confuse the identification of electrical appliances with multiple states, and it is easy to misjudge the identification between resistive loads. Summary of the Invention

[0004] In view of the above analysis, the present invention aims to provide a non-intrusive load identification method, device, equipment and storage medium to solve the problems of low identification accuracy and large network parameter quantity.

[0005] The object of the present invention is mainly achieved through the following technical solutions:

[0006] On the one hand, the present invention discloses a non-intrusive load identification method, including:

[0007] A data preprocessing step; processing the collected current and voltage data to generate a Markov transition field;

[0008] Each element in the Markov transition field is the state transition probability filled corresponding to the collected current and voltage data arranged in the time domain; the state transition probability is the state transition probability that maps the obtained current and voltage normalization data to the corresponding quantile unit after normalizing and quantiling the collected current and voltage data;

[0009] A feature image generation step; performing downsampling and pseudo-color coding on the Markov transition field to obtain an RGB color image;

[0010] A load identification step; inputting the RGB color image into the constructed neural network model for non-intrusive load identification.

[0011] On the other hand, the present invention also discloses a non-intrusive load identification device based on a Markov transfer field, including: a data preprocessing module, a feature image generation module, and a load identification module; wherein,

[0012] The data preprocessing module is used to process the collected current and voltage data into a Markov transfer field;

[0013] Each element in the Markov transfer field is the state transition probability filled for the collected current and voltage data arranged in the time domain; the state transition probability is the state transition probability obtained by normalizing and quantiling the collected current and voltage data and mapping the normalized current and voltage data to the corresponding quantile unit;

[0014] The feature image generation module is used to perform downsampling and pseudo-color coding on the Markov transfer field to obtain an RGB color image;

[0015] The load identification module is used to input the RGB color image into the constructed neural network model for non-intrusive load identification.

[0016] On the other hand, the present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the above-mentioned non-intrusive load identification method is implemented to perform non-intrusive load identification.

[0017] On the other hand, the present invention also discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the above-mentioned non-intrusive load identification method is implemented to perform non-intrusive load identification.

[0018] The beneficial effects of the present invention are as follows:

[0019] Aiming at the problems of the existing non-intrusive load identification method with weak learning ability and a large network model, considering the correlation between voltage and current data, the difference between different electrical appliance feature images is increased; it is beneficial to perform load identification; a SimAM parameter-free attention module is added to the network of the lightweight network ShuffleNetV2 as a feature extraction network, and the convolution kernel in the depthwise separable convolution is expanded to obtain a larger receptive field, so as to realize load classification and identification with fewer parameters. The present invention can be applied to the load monitoring and management of smart grids, improving the efficiency and reliability of energy use. Description of the Drawings

[0020] The drawings are only for the purpose of showing specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference signs represent the same components;

[0021] Figure 1 Flow chart of the non-invasive load identification method in the embodiment of the present invention;

[0022] Figure 2 Schematic diagram of the process of obtaining an RGB color image using voltage and voltage-normalized data in the embodiment of the present invention;

[0023] Figure 3 Load Markov transition field image generated by 11 electrical appliances in the PLAID dataset in the embodiment of the present invention;

[0024] Figure 4 Schematic diagram of the network model structure in the embodiment of the present invention where the SimAM parameter-free attention module is placed in front of the classification layer of the lightweight network ShuffleNetV2;

[0025] Figure 5 Network structure diagram of ShuffleNetV2-0.5 as a lightweight network for load identification in the embodiment of the present invention;

[0026] Figure 6 Schematic diagram of the training process of the PLAID dataset in the embodiment of the present invention;

[0027] Figure 7 Confusion matrix diagram of the experimental results of the PLAID test set in the embodiment of the present invention;

[0028] Figure 8 Schematic diagram of the training process of the WHITED dataset in the embodiment of the present invention;

[0029] Figure 9 Confusion matrix diagram of the experimental results of the WHITED test set in the embodiment of the present invention;

[0030] Figure 10 Markov transition field images of the microwave oven in the LAID dataset with bin values equal to 5, 10, 20, and 40 in the embodiment of the present invention;

[0031] Figure 11 Comparison of the accuracies of the test set under two strategies with different bin values in the embodiment of the present invention is shown in the figure;

[0032] Figure 12 Schematic diagram of the composition and connection of the non-invasive load identification device based on the Markov transition field in the embodiment of the present invention. Detailed implementation manners

[0033] The preferred embodiments of the present invention will be specifically described below with reference to the accompanying drawings, where the accompanying drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention.

[0034] Example 1

[0035] This example discloses a non-intrusive load identification method, as Figure 1 shown, including:

[0036] Step S1, data preprocessing; processing the collected current and voltage data into a Markov transition field;

[0037] Each element in the Markov transition field is the state transition probability filled corresponding to the collected current and voltage data arranged in the time domain; the state transition probability is the state transition probability obtained by normalizing and quantile processing the collected current and voltage data and mapping the obtained normalized current and voltage data to the corresponding quantile unit;

[0038] Step S2, feature image generation; performing downsampling and pseudo-color coding on the Markov transition field to obtain an RGB color image;

[0039] Step S3, load identification; inputting the RGB color image into the constructed neural network model for non-intrusive load identification;

[0040] Preferably, the neural network model is a network model that adds a SimAM parameter-free attention module to the lightweight network ShuffleNetV2 as a feature extraction network.

[0041] The neural network model in this example adds a SimAM parameter-free attention module to the lightweight network ShuffleNetV2 as a feature extraction network, enabling the neural network to focus on specific parts of the input data and thus improving the model performance.

[0042] Specifically, the step S1 includes:

[0043] 1) Performing normalization processing on the collected discrete current and voltage signals arranged in the time domain to obtain normalized current and voltage data;

[0044] Collect N discrete current and voltage signals arranged in the time domain to obtain a discrete voltage signal sequence u = {u1, u2,... u N} and a discrete current signal sequence i = {i1, i2,... i N};

[0045] Respectively perform normalization processing to obtain normalized voltage data u d and normalized current data i d :

[0046]

[0047]

[0048] In the formula, u min and u max represent the minimum and maximum values of the voltage respectively; i min and i max represent the minimum and maximum values of the current respectively.

[0049] 2) Perform quantile processing on the normalized data of current and voltage;

[0050] Divide the normalized voltage data u d into Q quantile units (bin values) according to the amplitude, and each quantile unit has the same number of sampling points; map each normalized voltage data u d and the normalized current data i d into the Q quantile units respectively;

[0051] That is, map each u d , i d to the corresponding bin value q j respectively, where j ∈ [1, Q].

[0052] Among them, the bin value is taken as 5 - 40, and 20 is more optimal.

[0053] 3) Generate a Markov state transition matrix;

[0054] Use the transition probability of the Markov chain to represent the relationship between the two sequences of the normalized data of current and voltage mapped to the corresponding quantile units; arrange all the transition probabilities of the Markov chain along the transition rule to obtain the Markov state transition matrix;

[0055] Use the transition probability of the Markov chain to represent the relationship between the two sequences of the normalized data of current and voltage mapped to the corresponding quantile units, as shown in the following formula:

[0056] P j,k = p j,k (i d ∈ q j | u d ∈ q k ) ;

[0057] P j,k represents the conditional probability that at the same moment, in u d and i d , it transfers from the quantile unit q k to q j , where k, j ∈ [1, Q];

[0058] Arrange all the transition probabilities of the Markov chain along the transition rule, and the obtained Q×Q Markov state transition matrix T is:

[0059]

[0060] 4) Fill the state transition probability according to the time-domain arrangement of the collected current and voltage to obtain the Markov transition field;

[0061] The Markov transition field M is:

[0062]

[0063] In the formula, M jk = p j,k (i j ∈ q j | u k ∈ q k ) represents that at the moment when the current normalized data i d = i j and the voltage normalized data is u d = u k , the normalized data i j and u k are transferred to the transition probability corresponding to the quantile units q j and q k on the state transition matrix T.

[0064] The size of the Markov transition field finally obtained in step S1 is N×N. When the amount of discrete data N collected is too large, the size of the subsequent generated image is large, which is not conducive to the calculation and data storage of the neural network, and there is a lot of redundant information.

[0065] Based on this, in step S2, it includes:

[0066] 1) Downsampling processing;

[0067] Use the kernel blur method to downsample the Markov transition field to obtain a two-dimensional matrix of m×m;

[0068] 2) Generate a grayscale image;

[0069] Plot the data in the two-dimensional matrix of m×m corresponding to the grayscale values encoded inside the computer to generate a grayscale image;

[0070] 3) Convert the grayscale image to an RGB color image;

[0071] Map different grayscale values to different colors for pseudo-color coding to obtain an RGB color image.

[0072] The RGB color image generated by the solution of this embodiment is constructed according to the two time series of voltage and current. The image contains both the time-domain information of the sequence and the correlation between current and voltage, thus increasing the difference between different electrical feature images; it is conducive to load identification.

[0073] The schematic flow chart of obtaining an RGB color image using voltage and voltage-normalized data is as Figure 2 shown.

[0074] The load Markov transition field images generated by 11 electrical appliances in the PLAID dataset are as shown in the appendix Figure 3 shown.

[0075] Specifically, in the neural network model in step S3, the lightweight network ShuffleNetV2 is adopted;

[0076] As one of the most representative lightweight networks at present, ShuffleNetV2 has the advantages of few network parameters, low computational complexity, fast operation speed, high recognition accuracy, etc. Since lightweight networks usually have fewer parameters and layers, they are usually more efficient and easier in the training and tuning processes. This can save time and computing resources and make the model development and optimization processes smoother.

[0077] The ShuffleNetV2 model optimizes the residual structure using depthwise separable convolutions and channel shuffle. Among them, the depthwise separable convolution operation is divided into depth convolution and pointwise convolution. Depth convolution first performs independent convolution operations on each channel of the input, and then uses pointwise convolution to fuse the results of the previous step; channel shuffle means that after the convolution operation is completed, the intra-group features are scattered into different groups and then the next group convolution is performed.

[0078] The lightweight network ShuffleNetV2 model optimizes the ResNet residual structure using channel transformation and depthwise separable convolutions, which not only improves the running efficiency of the model but also ensures the accuracy of the network.

[0079] However, there is no effective attention mechanism in ShuffleNetV2 to help the model learn features in terms of structure.

[0080] Preferably, in this embodiment, the ShuffleNetV2 model is improved by adding a SimAM parameter-free attention module to the lightweight network ShuffleNetV2 as the feature extraction network.

[0081] The attention mechanism can enable the neural network to focus on specific parts of the input data, thereby improving the model performance. The SimAM parameter-free attention module adopted in this embodiment is different from the existing channel attention or spatial attention, and no additional parameters are required when obtaining the three-dimensional attention weights of the feature map.

[0082] In the SimAM parameter-free attention module, in order to successfully implement attention, the importance of a single neuron needs to be estimated. The weight of the neuron is calculated through the minimum energy function. The formula for the minimum energy function is:

[0083]

[0084] where λ is the regularization term, t l is the l-th neuron on a single channel of the input feature map, is the mean value of all neurons on a single channel, is the variance of all neurons on a single channel.

[0085] The smaller the value, the lower the energy, the greater the difference between neuron l and its surrounding neurons, and the more important it is for visual processing. Therefore, the importance of each neuron can be obtained by . The output feature map is shown as follows:

[0086]

[0087] where E is the set of all neurons , the purpose of the sigmoid function is to limit the value of E, and X is the input feature map.

[0088] As Figure 4 shown, in order to avoid the attention module increasing the complexity of the model, in the preferred solution of this embodiment, the SimAM parameter-free attention module is placed before the classification layer of the lightweight network ShuffleNetV2. After the input feature map passes through H×W SimAM attention modules, the calculation of formula is performed to obtain the final output feature map; H and W are the two-dimensional widths of the output data of the classification layer respectively.

[0089] In this embodiment, by placing the SimAM parameter-free attention module before the classification layer, the model's perception ability of channel information is improved, and further the model can better extract spatial features.

[0090] Furthermore, the convolution kernel in the depthwise separable convolution of the ShuffleNetV2 model of the lightweight network is expanded to obtain a larger receptive field.

[0091] Specifically, the 3×3 convolution in the depthwise separable convolution is replaced with a 5×5 convolution, thereby obtaining a larger receptive field.

[0092] Preferably, in this embodiment of the research, ShuffleNetV2-0.5 with the fewest parameters is selected as the lightweight network for load recognition, and the network structure of the present invention is as shown in the appendix Figure 5 .

[0093] In a specific solution of this embodiment, in order to verify the effectiveness of the solution of the present invention well, the PLAID dataset and the WHITED dataset are used to train and test the neural network model respectively;

[0094] The environmental parameters for the specific training and testing are shown in Table 1, and the specific experimental settings are as follows:

[0095] Table 1

[0096]

[0097] During the training and testing processes, the accuracy rate, confusion matrix, and F1 score are used as evaluation indicators for load identification. The accuracy rate is used to evaluate the overall identification effect of the dataset, and the calculation method is as follows:

[0098]

[0099] In the formula, A represents the total number of samples, and a represents the number of samples correctly identified.

[0100] The F1 score is used to evaluate the identification effect of each type of load and is the harmonic mean of the precision rate and the recall rate. The calculation formula is as follows:

[0101]

[0102]

[0103]

[0104] In the formula, T P represents the number of samples that are actually positive classes and are simultaneously identified as positive classes; F N represents the number of samples that are actually positive classes but are identified as negative classes; F P represents the number of samples that are actually negative classes but are identified as positive classes. P re is the precision rate, and R e is the recall rate.

[0105] The main experimental results in this embodiment include the experimental results of the PLAID dataset, the experimental results of the WHITED dataset, the comparison results of different strategies of the Markov transfer field, and the comparison results with existing methods.

[0106] Experimental results of the PLAID dataset: The PLAID dataset contains 1,074 instances of 15 types of electrical appliances, and detailed information from startup to stable operation is recorded at a sampling frequency of 30 kHz. According to the present invention, the voltage and current in the PLAID dataset are processed, and one data point is taken every 20 cycles to generate a load image set with fused features, totaling 7,953 samples. These samples are divided into a training set, a validation set, and a test set in a ratio of 6:2:2. Adam optimizer is used for training, and an adaptive adjustment mechanism is set for the learning rate. The initial learning rate is 0.0001. If the loss function does not decrease within 7 times, the learning rate is updated to 1 / 10 of the original learning rate to optimize the model. The batch size set for model training is 64, and the training includes 100 iterations.

[0107] Figure 6 The training process for the PLAID dataset is shown. It can be seen that the model converges completely around 80 times. After convergence, the highest accuracy rate of the validation set reaches 99.31%, and the loss value of the validation set converges to 0.018. The test set is input into the trained model, and the recognition accuracy rate on the PLAID test set reaches 98.99%.

[0108] Figure 7 The confusion matrix is drawn based on the results. The recognition effect of the refrigerator is lower than that of other loads because there are more unknown states of the refrigerator in the PLAID dataset, which is prone to misjudgment as other loads; the fluorescent lamp and the laptop have similar current trajectories, resulting in a small difference in their Markov transition matrices and being prone to confusion; the heater and the hair dryer belong to the same type of heating equipment and have similar working modes, making it difficult to distinguish them. The average F1 score of the recognition reaches 98.94%.

[0109] Experimental results of the WHITED dataset: The WHITED dataset contains 54 types of electrical appliances, with a total of 1,339 instances, and the sampling frequency is 44.1 kHz. A similar experimental process as that of the PLAID dataset is adopted for the WHITED dataset, and a total of 6,450 samples are generated. The accuracy rate and loss function during the training process are as Figure 7 shown. The inter-class differences in the WHITED dataset are relatively large, so the convergence speed is higher than that of the PLAID dataset, and it achieves complete convergence around 70 times. The highest accuracy rate of the validation set is 99.53%, and the loss value is as low as 0.016. The test set is input into the trained model, and the overall accuracy rate reaches 99.22%, and the average F1 score is 99.21%. The confusion matrix is as Figure 8 shown.

[0110] There are many heating-type loads in the WHITED dataset, such as kettles, irons, rice cookers, sandwich makers, etc. The current signals of these loads are almost sinusoidal signals, and the phase angles are also close to 0, which are extremely prone to confusion. Therefore, further research is carried out on the accurate identification of heating-type loads. The current waveforms and phase differences of game consoles, treadmills, televisions, and fluorescent lamps are similar, and the generated Markov transition field images have little difference after normalization, resulting in misjudgments among these loads.

[0111] Comparison results of different strategies for Markov transition fields: Different bin values have a great impact on image generation. When the bin value becomes larger, the probabilities of the Markov transition matrix are more diverse, and the Markov transition field images are more complex.

[0112] Figure 10 Shows the Markov transition field images of the microwave oven in the PLAID dataset when the bin values are equal to 5, 10, 20, and 40. It can be seen that as the bin value increases, the segmentation of the image increases and the texture becomes more complex. Using the Markov transition field generated from two sequences of current and voltage as the feature image for load identification, compared with the image generated from the traditional single current sequence, the method of the present invention has higher discrimination. The accuracy rates of the test set under two strategies at different bin values are compared as Figure 11 shown. It can be seen that the accuracy rate of the method of the present invention for the two datasets is significantly higher than that of the Markov transition field method of the single current sequence at different bin values.

[0113] Under the method of the embodiment of the present invention, the PLAID dataset reaches the highest accuracy rate of 98.99% at 20 bins, and the WHITED dataset reaches the highest accuracy rate of 99.30% at 40 bins. Since the accuracy rates of the two datasets at 20 bins and 40 bins are not very different, considering the computational amount, the present invention selects 20 as the final bin value for experiments.

[0114] Comparison results with existing methods: To comprehensively verify the effectiveness of the method of the present invention, the load identification effects of the method of the present invention in the PLAID and WHITED datasets are compared with those of current advanced methods. The recognition accuracies, network parameter quantities, and computational amounts of different methods are shown in Table 2. It can be seen from the results that the accuracy of the method of the present invention is better than that of the other four load identification methods. Method 1 uses a feature combination grayscale image as the feature map and a lightweight improved network ZFNet-Inception as the recognition network; Method 2 uses a color image encoded with U-I trajectories as the feature image and the classic network VGG16 for load identification; Method 3 first extracts the reactive current and then uses the Gram angle field of the reactive current as the feature image and uses Inception-v3 to achieve load identification. It can be seen from the results that, compared with the existing methods, the present invention has a higher recognition accuracy than the three methods, and the parameter quantity of the feature image is reduced by 75.5%-91.8%; the network parameter quantity is much lower than that of the VGG16 and Inception-v3 models of Method 2 and Method 3, and the parameter quantity is reduced by 78.3% compared with the lightweight network ZFNet of Method 1; in terms of computational amount, the method of the present invention is only about 1 / 1000 of the three methods. It can be seen that the method of the present invention achieves a higher recognition accuracy with fewer parameter quantities and computational amounts, reflecting the superiority of the method of the present invention.

[0115] Table 2

[0116]

[0117] In summary, the non-intrusive load identification method in this embodiment considers the correlation between voltage and current data, increases the difference between different electrical appliance feature images; is conducive to load identification; adds a SimAM parameter-free attention module to the network of the lightweight network ShuffleNetV2 as the feature extraction network, expands the convolution kernel in the depthwise separable convolution, and obtains a larger receptive field, so as to achieve load classification and recognition with fewer parameter quantities. The solution of this embodiment can be applied to the load monitoring and management of smart grids, improving the efficiency and reliability of energy use.

[0118] Embodiment Two

[0119] Another embodiment of the present invention discloses a non-intrusive load identification device based on a Markov transfer field, as Figure 12 shown, including: a data preprocessing module, a feature image generation module, and a load identification module; wherein,

[0120] The data preprocessing module is used to process the collected current and voltage data into a Markov transfer field;

[0121] Each element in the Markov transition field is the state transition probability filled for the collected current and voltage data arranged in the time domain; the state transition probability is the state transition probability obtained by mapping the normalized current and voltage data to the corresponding quantile unit after normalizing and quantiling the collected current and voltage data;

[0122] A feature image generation module, configured to perform downsampling and pseudo-color coding on the Markov transition field to obtain an RGB color image;

[0123] A load identification module, configured to input the RGB color image into a constructed neural network model for non-intrusive load identification.

[0124] The specific technical details and beneficial effects in this embodiment are the same as those in Embodiment 1. Please refer to them specifically and will not be elaborated here one by one.

[0125] Embodiment 3

[0126] An embodiment of the present invention discloses an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor.

[0127] In one example, the above-mentioned processor may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0128] The memory may include a read-only memory (ROM), a random access memory (RAM), a disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform non-intrusive load identification by referring to the non-intrusive load identification method in Embodiment 1.

[0129] The processor runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory, so as to implement the non-intrusive load identification method in Embodiment 1 for non-intrusive load identification.

[0130] In one example, a communication interface and a bus may also be included.

[0131] Among them, the memory, the processor, and the communication interface are connected through a bus and complete communication with each other.

[0132] The communication interface is mainly used to implement communication between various modules, devices, units, and / or devices in the embodiments of the present application. It can also access input devices and / or output devices through the communication interface.

[0133] The bus includes hardware, software, or both, and the components of the electronic device are coupled together. By way of example and not limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses or a combination of two or more of these. In a suitable case, the bus may include one or more buses. Although the embodiments of the present application describe and illustrate specific buses, the present application contemplates any suitable bus or interconnect.

[0134] Embodiment 4

[0135] An embodiment of the present invention discloses a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the non-intrusive load identification method in Embodiment 1 can be implemented to perform non-intrusive load identification and achieve the same technical effects. To avoid repetition, it will not be elaborated here. Among them, the above computer-readable storage medium may include a non-transitory computer-readable storage medium, such as a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disc, etc., which is not limited herein.

[0136] As described above, it is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A non-intrusive load identification method, characterized in that, Including: Data preprocessing step; Processing the collected current and voltage data to generate a Markov transition field, including: 1) Normalizing the discrete current and voltage signals arranged in the time domain to obtain normalized current and voltage data; 2) Performing quantile processing on the normalized current and voltage data; Normalize the voltage data u d Divide it into Q quantile units according to the amplitude, and each quantile unit has the same number of sampling points; for each normalized voltage data u d and the normalized current data i d Map and divide them into Q quantile units respectively; 3) Generating a Markov state transition matrix; Using the transition probability of the Markov chain to represent the two-sequence relationship between the normalized current and voltage data mapped to the corresponding quantile units, as shown in the following formula: P j,k = p j,k (i d ∈ q j | u d ∈ q k ); P j,k represents the conditional probability of transferring from the quantile unit q d and i d to q k at the same moment, where k, j ∈ [1, Q]; j ​ Arranging all the transition probabilities of the Markov chain along the transition rule, the Q×Q Markov state transition matrix T obtained is: 4) Filling the state transition probabilities according to the time-domain arrangement of the collected current and voltage to obtain a Markov transition field; The Markov transition field M is: where, M jk = p j,k (i j ∈ q j | u k ∈ q k ) represents the transfer probability corresponding to the quantile units q d = i j and u d = u k at the moment when the normalized data of current is i j and u k are transferred to the quantile units q j and q k on the Markov state transition matrix T; Feature image generation step; Downsampling and pseudo-color encoding the Markov transition field to obtain an RGB color image, including: 1) Downsampling processing; Using the kernel blur method to downsample the Markov transition field to obtain a two-dimensional matrix of m×m; 2) Generating a grayscale image; Mapping the data in the two-dimensional matrix of m×m to the grayscale values encoded inside the computer to draw a grayscale image; 3) Converting the grayscale image to an RGB color image; Mapping different grayscale values to different colors for pseudo-color encoding to obtain an RGB color image; The RGB color image is constructed based on the two time series of voltage and current, and the image contains both the time-domain information of the series and the correlation between current and voltage; Load identification step; Inputting the RGB color image into the constructed neural network model for non-intrusive load identification.

2. The non-intrusive load identification method according to claim 1, wherein The neural network model is a network model that adds a SimAM parameter-free attention module to the lightweight network ShuffleNetV2 as a feature extraction network.

3. The non-intrusive load identification method according to any one of claims 1-2, wherein Place the SimAM parameter-free attention module in front of the classification layer of the lightweight network ShuffleNetV2. After the input feature map passes through the SimAM attention module, perform the calculation of formula to obtain the final output feature map; X is the input feature map, is the output feature map, and E is the set of weights of all neurons; the sigmoid function restricts the value of E.

4. The non-intrusive load identification method according to claim 3, wherein Calculating the weights of neurons through a minimum energy function; The minimum energy function formula is: where λ is the regularization term, and t l is the l-th neuron of the input feature map on a single channel, is the mean of all neurons on a single channel, is the variance of all neurons on a single channel.

5. The non-intrusive load identification method according to claim 3, wherein Expanding the convolution kernels in the depthwise separable convolution of the lightweight network ShuffleNetV2 model to obtain a larger receptive field.

6. A non-intrusive load identification device for implementing the non-intrusive load identification method according to any one of claims 1-5, characterized in that, Including: A data preprocessing module, a feature image generation module, and a load identification module; wherein, The data preprocessing module is used to process the collected current and voltage data to generate a Markov transition field; Each element in the Markov transition field is the state transition probability filled for the collected current and voltage data arranged in the time domain; The state transition probability is the state transition probability obtained by normalizing and performing quantile processing on the collected current and voltage data and mapping the normalized current and voltage data to the corresponding quantile units; The feature image generation module is used to downsample and pseudo-color encode the Markov transition field to obtain an RGB color image; A load identification module, configured to input an RGB color image into a constructed neural network model for non-intrusive load identification.

7. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the non-intrusive load identification method according to any one of claims 1 to 5 to perform non-intrusive load identification.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the non-intrusive load identification method according to any one of claims 1 to 5 to perform non-intrusive load identification.

Images