Non-intrusive load identification method based on neural network model

Abstract

The present application relates to the technical field of load identification, and in particular to a non-intrusive power load identification method based on a neural network model. The method comprises: obtaining current signals of a user's electrical appliances in different detection periods; inputting the current signals into a neural network to obtain feature maps; obtaining a parameter vector of each feature channel in each detection period based on the gray scale distribution of different feature channels in the feature maps; obtaining a relative correlation coefficient according to the distribution difference of the parameter vectors of the same feature channel of each detection period and its adjacent previous detection period; evaluating the channel information redundancy performance of each feature channel according to the overall distribution difference of the relative correlation coefficients of the load detection data of a single feature channel in all detection periods and the parameter vectors, introducing an ECA mechanism, determining the feature importance weight of the feature channel, and judging the load operating state. The present application improves the accuracy of load operating state identification.

Description

Technical Field

[0001] This invention relates to the field of load identification technology, and more specifically to a non-intrusive electrical load identification method based on a neural network model. Background Technology

[0002] Non-intrusive load identification refers to the process of acquiring and processing real-time data on different types of electrical loads without altering system operation or adding additional sensors. This data is then used for load decomposition and identification through data analysis or machine learning methods. This allows for monitoring the actual usage and energy consumption of various electrical devices within the line. Load identification is the core of non-intrusive load identification. Common methods include using clustering algorithms to analyze the operating status of different users' appliances, obtaining characteristic curves corresponding to different loads, thereby enabling the decomposition and identification of abnormal situations. This paper presents a method for constructing a load database based on steady-state and transient characteristics, and establishes a corresponding load identification model. Multi-label classification technology is introduced into load identification research, overcoming the limitations of traditional electricity meters for non-intrusive load identification.

[0003] Mixed load usage data includes all power consumption information of the load. Load actions will change the load's own power consumption, which is reflected in the mixed usage data based on the additivity of current. In non-intrusive load identification, when the load's operating state changes, the state of the load characteristics also changes accordingly. Existing non-intrusive load identification methods still have significant limitations in practical household applications. In actual processing, they fail to fully consider the differences in importance between different feature channels, resulting in significant information redundancy in the feature responses between channels, leading to low accuracy in non-intrusive load identification. Summary of the Invention

[0004] To address the issue of low accuracy in existing methods for non-intrusive load identification, this invention aims to provide a non-intrusive electrical load identification method based on a neural network model. The specific technical solution adopted is as follows:

[0005] This invention provides a non-intrusive electrical load identification method based on a neural network model, the method comprising the following steps:

[0006] Acquire current signals from user electrical appliances during different detection cycles;

[0007] The current signal is input into a neural network model to obtain a feature map; the parameter vector of each feature channel in each detection cycle is obtained based on the gray-level distribution of different feature channels in the feature map in each detection cycle; the relative correlation coefficient of the load detection data of each feature channel in each detection cycle is obtained based on the distribution difference of the parameter vector of the same feature channel in each detection cycle and the adjacent previous detection cycle.

[0008] The channel information redundancy performance of each feature channel is evaluated based on the overall distribution difference of the relative correlation coefficients of the load detection data in a single feature channel for all detection cycles and the parameter vector.

[0009] An ECA mechanism is introduced into the neural network model to determine the feature importance weights of feature channels by utilizing the redundancy of channel information; the load operation status is then determined based on these feature importance weights.

[0010] Preferably, obtaining the parameter vector of each feature channel in each detection period based on the grayscale distribution of different feature channels in the feature map in each detection period includes: arranging the grayscale values ​​of all pixels in each feature channel in the feature map in each detection period to obtain the parameter vector of each feature channel in the feature map in each detection period.

[0011] Preferably, the step of obtaining the relative correlation coefficient of the load detection data of each feature channel in each detection cycle based on the distribution difference of the parameter vector of the same feature channel in each detection cycle and its adjacent previous detection cycle includes:

[0012] The range of the elements in the reference vector of each week's expectation analysis channel is denoted as the first eigenvalue of each week's expectation analysis channel;

[0013] Calculate the first difference between the first eigenvalue of the candidate period and the first eigenvalue of the expected analysis channel in the previous week of the candidate period;

[0014] Based on the first difference, the load steady-state coefficient of the candidate cycle expectation analysis channel is obtained;

[0015] Calculate the second difference between the mean of the elements in the reference vector of the candidate week expectation analysis channel and the mean of the elements in the reference vector of the previous week expectation analysis channel.

[0016] Based on the load steady-state coefficient of the candidate week expectation analysis channel and the second difference, the relative correlation coefficient of the load detection data of the candidate week expectation analysis channel is obtained;

[0017] The candidate period is any detection period, and the channel to be analyzed is any feature channel.

[0018] Preferably, obtaining the load steady-state coefficient of the candidate week-expected analysis channel based on the first difference includes:

[0019] Obtain the maximum value of the first eigenvalue of all expected analysis channels for each week;

[0020] Calculate the first ratio between the first difference and the maximum value, and use the negative correlation mapping result of the first ratio as the load steady-state coefficient of the candidate period expectation analysis channel.

[0021] Preferably, obtaining the relative correlation coefficient of the load detection data of the candidate week expectation analysis channel based on the load steady-state coefficient of the candidate week expectation analysis channel and the second difference includes:

[0022] Calculate the second ratio between the load steady-state coefficient of the candidate cycle expectation analysis channel and the second difference;

[0023] The normalized result of the second ratio is determined as the relative correlation coefficient of the load detection data of the candidate weekly expectation analysis channel.

[0024] Preferably, the step of evaluating the channel information redundancy performance of each feature channel based on the overall distribution difference of the relative correlation coefficients of the load detection data in a single feature channel across all detection cycles and the parameter vector includes:

[0025] Based on the relative correlation coefficient of the load detection data of the expected analysis channel in each detection week and the magnitude of the parameter vector of the expected analysis channel in each detection week, the load influence factor of the expected analysis channel in each detection week is obtained.

[0026] The channel information redundancy performance of the channel to be analyzed is obtained by comparing the difference between the load influence factor of the channel to be analyzed in each detection week and the average load influence factor of all channels to be analyzed in all detection weeks.

[0027] Preferably, the step of obtaining the load influence factor of each expected analysis channel based on the relative correlation coefficient of the load detection data of each expected analysis channel in each detection cycle and the magnitude of the parameter vector of each expected analysis channel in each detection cycle includes:

[0028] The ratio between the magnitude of the parameter vector of the expected analysis channel in each detection week and the relative correlation coefficient of the load detection data of the expected analysis channel in the same detection week is used as the load influence factor of the expected analysis channel in each detection week.

[0029] Preferably, the step of obtaining the channel information redundancy performance of the channel to be analyzed based on the difference between the load influence factor of each detection week's expected analysis channel and the average load influence factor of all detection weeks' expected analysis channels includes:

[0030] The negative correlation mapping value between the difference between the load influence factor of each detection week expectation analysis channel and the average load influence factor of all detection week expectation analysis channels is denoted as the second characteristic value of each detection week expectation analysis channel.

[0031] The average value of the second feature value of all channels to be analyzed in the detection cycle is taken as the channel information redundancy of the channel to be analyzed.

[0032] Preferably, the step of determining the feature importance weight of a feature channel using channel information redundancy includes:

[0033] Calculate the cumulative sum of channel information redundancy across all feature channels;

[0034] The ratio of the redundancy of channel information in each feature channel to the sum is determined as the feature importance weight of each feature channel.

[0035] Preferably, determining the load operating status based on the feature importance weights includes:

[0036] The feature importance weights are multiplied channel-by-channel with the original input feature map to obtain a feature map with channel attention;

[0037] The load operating status is classified based on the feature map with channel attention, including single load, dual load, triple load, quadruple load and quintuple load operating status.

[0038] The present invention has at least the following beneficial effects:

[0039] This invention evaluates the relative correlation coefficient of load detection data for each feature channel within a neural network model of current signals from user appliances across different detection cycles. This is achieved by combining the grayscale distribution of different feature channels in the feature map with the distribution differences of parameter vectors for the same feature channel in each detection cycle and its adjacent previous detection cycle. Based on the overall distribution differences of the relative correlation coefficients of load detection data for a single feature channel across all detection cycles, the channel information redundancy performance of each feature channel is evaluated. This provides a more comprehensive analysis of the complex distribution and dynamic changes of electrical parameters. Furthermore, by introducing an ECA mechanism to capture data correlation, the feature importance weights of feature channels are adaptively determined using channel information redundancy performance, enabling accurate identification of load operating states and improving the accuracy and robustness of load operating state identification. This method not only solves the identification confusion problem caused by feature similarity in traditional load identification but also more effectively handles load patterns under complex operating conditions, enhancing the intelligence level of appliance operating status monitoring and energy consumption management. Attached Figure Description

[0040] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1A flowchart of a non-intrusive electrical load identification method based on a neural network model provided in an embodiment of the present invention;

[0042] Figure 2 This is a structural block diagram of a non-intrusive electrical load identification system based on a neural network model, provided in an embodiment of the present invention. Detailed Implementation

[0043] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of a non-intrusive electrical load identification method based on a neural network model, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0045] The following description, in conjunction with the accompanying drawings, details a specific scheme for a non-intrusive electrical load identification method based on a neural network model provided by the present invention.

[0046] Example of a non-intrusive electrical load identification method based on neural network model:

[0047] This embodiment proposes a non-intrusive electrical load identification method based on a neural network model, such as... Figure 1 As shown, the non-intrusive electrical load identification method based on a neural network model in this embodiment includes the following steps:

[0048] Step S1: Obtain the current signal of the user's electrical appliance during different detection cycles.

[0049] First, current data is collected at the user's electricity meter. The user's household appliances include various types such as fans, electric heaters, kettles, refrigerators, and desktop computers. Multiple tests are performed at the user's electricity meter, with each complete test cycle constituting a test period. The current signal within each test period is then acquired. In this embodiment, the current data sampling frequency is 14kHz. In specific applications, the implementer can adjust this setting according to specific circumstances.

[0050] Then, wavelet transform is used to denoise the current signal acquired in each detection cycle. Wavelet transform is an existing technology and will not be elaborated on here. It should be noted that the current signal mentioned below refers to the current signal after wavelet transform processing.

[0051] Step S2: Input the current signal into the neural network model to obtain the feature map; obtain the parameter vector of each feature channel in each detection cycle based on the gray-level distribution of different feature channels in the feature map in each detection cycle; obtain the relative correlation coefficient of the load detection data of each feature channel in each detection cycle according to the distribution difference of the parameter vector of the same feature channel in each detection cycle and the adjacent previous detection cycle.

[0052] Steady-state characteristics of a circuit refer to a relatively stable state maintained during the load's operating time; transient characteristics refer to the instantaneous characteristics exhibited by the load during startup, shutdown, or state switching. Relatively speaking, transient characteristics are more effective in distinguishing the load patterns of different electrical appliances. However, in reality, different electrical appliances exhibit similar variation characteristics, and there is correlation between different appliances. For example, the operation processes of washing machines and dryers are usually sequential, and their electrical characteristics show a certain correlation. Therefore, using the correlation performance of parameters to effectively represent the current mixed electrical appliance load status, and adjusting the analysis process of the improved Swin Transformer model based on it, can more effectively achieve accurate load identification.

[0053] Steady-state characteristic analysis involves monitoring and calculating electrical parameter characteristics during stable load operation. This means the data in the current analysis phase primarily reflects changes in the operating status of the equipment, excluding equipment start-up, shutdown, or the addition or use of other equipment. In other words, data fluctuations depend more on the load characteristics caused by the equipment itself. For example, the load characteristic changes of appliances like hair dryers and washing machines under multiple modes will cause related changes in the electrical parameters of that equipment. The corresponding load changes affect the characteristic performance of the load analysis, which can be categorized as positive or negative. For instance, after a household appliance switches modes, the load condition changes, and its own load characteristics may increase or decrease. However, whether its performance changes within a mixed load depends on the influence of other appliances. Therefore, the steady-state process of the power grid load requires effective control of various parameters and correlation analysis of the load current data.

[0054] For all electrical load steady-state characteristic analysis processes, the actual usage behavior at different stages will exhibit different load characteristic changes on different characteristic channels.

[0055] Next, this embodiment will be described using one detection cycle as an example. The method provided in this embodiment can be used to process other detection cycles.

[0056] Specifically, any detection period is designated as a candidate period, and the current signal of the candidate period is input into a neural network model to obtain a feature map; in this embodiment, the neural network model is the Swing Transformer model. The grayscale values ​​of all pixels in each feature channel of the feature map within each detection period are arranged to obtain the parameter vector of each feature channel in the feature map within each detection period. The grayscale values ​​of pixels in the image can be arranged from left to right and from top to bottom. The neural network processing is based on the assumption that different types of electrical appliances have different load characteristics and exhibit different behaviors in different feature channels. Therefore, steady-state load changes are primarily reflected in the changes of parameter vector values ​​in the corresponding feature channels.

[0057] For a single feature channel in the feature map, the higher the consistency of the variation within the candidate period, the more singular the load characteristics of the electrical appliance type represented by the feature channel are. Inconsistent variation indicates that multiple electrical appliance types are involved, and the overall performance is greatly affected by the correlation between types.

[0058] Any feature channel is designated as the channel to be analyzed. The range of elements in the reference vector of the channel to be analyzed in each detection week is designated as the first eigenvalue of the channel to be analyzed in each detection week. The difference between the first eigenvalue of the channel to be analyzed in the candidate period and the previous week of the candidate period is calculated and designated as the first difference. The maximum value of the first eigenvalue of the channel to be analyzed in all detection weeks is obtained. The first ratio between the first difference and the maximum value is calculated, and the negative correlation mapping result of the first ratio is used as the load steady-state coefficient of the channel to be analyzed in the candidate week.

[0059] As a concrete example, the specific formula for calculating the load steady-state coefficient is given. The load steady-state coefficient of the j-th feature channel in the i-th detection period can be expressed as:

[0060]

[0061] in, This represents the load steady-state coefficient of the i-th feature channel in the j-th detection cycle. This represents the first feature value of the i-th feature channel in the j-th detection period. This represents the first feature value of the i-th feature channel in the (j-1)-th detection period. This represents the maximum value of the first eigenvalue of the i-th feature channel across all periods. This represents an exponential function with the natural constant as its base. Indicates the absolute value sign.

[0062] This represents the first difference between the first feature value of the i-th feature channel in the j-th detection period and the (j-1)-th detection period. This represents the first ratio corresponding to the i-th feature channel in the j-th detection period. The first ratio is equivalent to the normalized result of the first difference. The larger the value, the greater the difference between the first feature value of the i-th feature channel in the j-th detection period and the (j-1)-th detection period, and the smaller the load steady-state coefficient of the j-th feature channel in the i-th detection period.

[0063] After obtaining the steady-state load coefficient of the candidate week's expected analysis channel using the above method, the absolute value of the difference between the mean of the elements in the reference vector of the candidate week's expected analysis channel and the mean of the elements in the reference vector of the previous week's expected analysis channel is calculated, and this absolute value is recorded as the second difference. The ratio between the steady-state load coefficient of the candidate week's expected analysis channel and the second difference is calculated, and this ratio is recorded as the second ratio. The normalized result of the second ratio is determined as the relative correlation coefficient of the load detection data of the candidate week's expected analysis channel. It should be noted that if the second difference is 0, a zero-prevention parameter of 0.001 is first added to the second difference, and then the relative correlation coefficient is calculated using the above method. The relative correlation coefficient is used to reflect the relative correlation performance of the load influence of the electrical appliance type in the candidate week's expected analysis channel. The larger the value, the less obvious the process influence of this characteristic channel is.

[0064] In this embodiment, the maximum and minimum value normalization method is used to normalize the data. The maximum and minimum value normalization method is an existing technology and will not be described in detail here.

[0065] Using the above method, the relative correlation coefficient of the load detection data of each feature channel in each detection cycle can be obtained.

[0066] Step S3: Evaluate the channel information redundancy performance of each feature channel based on the overall distribution difference of the relative correlation coefficients of the load detection data in a single feature channel for all detection cycles and the parameter vector.

[0067] Considering that the relative correlation coefficients of the load detection data of the feature channels obtained in the above process describe more the response of a single feature channel to load changes, the load performance capture capabilities of different feature channels can easily lead to the dilution of important features and the over-amplification of secondary features. For example, in power load identification, different channels may capture the same power change trend, which may be due to certain high-concurrency load processes. In this case, using the traditional MLP that processes all feature channels in the same way will further enhance the performance of this feature, resulting in redundant channel information in the load process. Therefore, it is necessary to further analyze the actual load performance relationship of multiple feature channels and consider the interdependence between channels under the electrical appliance correlation analysis, which can more effectively achieve refined control of the load decomposition process.

[0068] Similar data variation trends across different feature channels represent a homogenization effect. High-concurrency load processes significantly impact the load across different feature channels, reducing their ability to capture load performance. Consequently, the more pronounced the homogenization across different feature channels, the greater the impact of channel information redundancy during the load process, and the less effective the load performance of that feature channel is for load identification.

[0069] The following explanation will still take the channel to be analyzed as an example. Other channels can be processed using the method provided in this embodiment.

[0070] Based on the above characteristics, the load influence factor of each detection week expectation analysis channel is obtained according to the relative correlation coefficient of the load detection data of each detection week expectation analysis channel and the magnitude of the parameter vector of each detection week expectation analysis channel.

[0071] As a specific example, the ratio between the magnitude of the parameter vector of each expected analysis channel in each detection week and the relative correlation coefficient of the load detection data of the expected analysis channel in the same detection week is used as the load influence factor of each expected analysis channel. It should be noted that if the denominator in the calculation formula of the load influence factor is 0, a zero-prevention parameter of 0.001 is added to the denominator before calculating the load influence factor in the above manner.

[0072] Furthermore, based on the difference between the load influence factor of the expected analysis channel in each detection week and the average load influence factor of the expected analysis channels in all detection weeks, the channel information redundancy performance of the channel to be analyzed is obtained.

[0073] As a specific example, the negative correlation mapping value between the load influence factor of each detection week's expected analysis channel and the average load influence factor of all detection weeks' expected analysis channels is denoted as the second characteristic value of each detection week's expected analysis channel. The average of the second characteristic values ​​of all detection weeks' expected analysis channels is taken as the channel information redundancy representation of the channel to be analyzed.

[0074] As a concrete example, a specific formula for calculating channel information redundancy is given. The channel information redundancy of the channel to be analyzed can be expressed as:

[0075]

[0076] in, This indicates the redundancy of channel information in the channel to be analyzed, where J represents the number of detection cycles. This represents the load impact factor of the channel to be analyzed in the j-th detection cycle. This represents the average load impact factor for all detection cycles expected to be analyzed. This represents an exponential function with the natural constant as its base.

[0077] This represents the second eigenvalue of the expected analysis channel in the j-th detection week. It reflects the difference between the load influence factor of the expected analysis channel in the j-th detection week and the average load influence factor of all expected analysis channels in the j-th detection week. The larger the value, the greater the difference between the two. This indicates the negative correlation mapping result of the difference. The greater the difference between the load influence factor of the expected analysis channel in each detection week and the average load influence factor of all expected analysis channels in all detection weeks, the more significant the load process tends to be uniform under the current parameter control, the greater the influence of channel information redundancy in the load process, the less likely the current household power grid related characteristic parameters are to be the optimal parameters, and the less useful the load performance of the corresponding analysis channel is for load identification.

[0078] The above methods are used to evaluate the channel information redundancy performance of each feature channel.

[0079] Step S4: Introduce the ECA mechanism into the neural network model to determine the feature importance weights of feature channels by utilizing the redundancy of channel information; determine the load operation status based on the feature importance weights.

[0080] This embodiment introduces an efficient channel attention (ECA) mechanism into the neural network model. This mechanism dynamically weights the features of each channel by explicitly modeling the interdependencies between feature channels. This mechanism effectively suppresses the responses of unimportant feature channels while enhancing the representational ability of key feature channels, thus enabling the model to learn the feature representations between channels more efficiently. Therefore, it is necessary to perform validation analysis on the overall performance to effectively improve the dynamic response capability of the neural network model in load recognition.

[0081] In the ECA mechanism, the input feature map undergoes spatial dimension compression through Global Average Pooling (GAP), reducing the feature map of size H×W×C to 1×1×C, thereby aggregating the global spatial information of each channel. The compressed feature map is then subjected to 1×1 convolution for cross-channel interactive learning to adaptively capture the non-linear dependencies between different channels and generate channel attention weights.

[0082] The adaptive weighting process for feature channel importance relies on the redundancy of the acquired feature channel information. The weights of the feature channels are adjusted for each detection cycle. This adjustment process can be represented as follows: if the shift window size of the current multi-head self-attention mechanism is r, then the weights are adjusted based on the redundancy of the channel information across all feature channels within that shift window size. Specifically, the sum of the redundancy of the channel information across all feature channels is calculated; the ratio of the redundancy of each feature channel to this sum is determined as the feature importance weight for each feature channel.

[0083] The feature importance weights for all feature channels are obtained, and a channel importance weight matrix of equal size is constructed according to the feature map construction dimension. Channel-wise multiplication is then performed between the feature importance weights and the original input feature map to enhance the feature responses of important channels and suppress redundant information, resulting in a feature map with channel attention. The output is then processed by image patch merging, where adjacent small image patches are merged into larger ones. During this process, due to the different weights assigned by the ECA mechanism, high-importance channels are more strongly represented, while low-importance channels are weakened. At this point, the output size becomes 28×28×192. The input after image patch merging continues to pass through the same Swin Transformer Block and ECA attention mechanism sequentially. This process is repeated twice to convert the output size to 7×7×768. Finally, global average pooling is performed on the resulting 7×7 feature map to generate a 1×1×768 vector, which is then input into the Softmax function for classification.

[0084] In the multi-load classification task, the samples are divided into five categories according to the number of loads running at the same time: single load, dual load, triple load, quadruple load, and five load operation status.

[0085] This embodiment improves the accuracy of steady-state load identification performance results in household electricity scenarios by introducing an efficient channel attention mechanism and optimizing the dynamic weighting between feature channels.

[0086] This embodiment evaluates the relative correlation coefficient of load detection data for each feature channel within a detection cycle by combining the grayscale distribution of different feature channels in the feature map and the distribution differences of parameter vectors of the same feature channel in each detection cycle with its adjacent previous detection cycle in the neural network model of the current signal of the user's electrical appliances. Based on the overall distribution differences of the relative correlation coefficients of load detection data for a single feature channel across all detection cycles, the channel information redundancy performance of each feature channel is evaluated. This provides a more comprehensive analysis of the complex distribution and dynamic changes of electrical parameters. By introducing an ECA mechanism to capture data correlation and utilizing the channel information redundancy performance to adaptively determine the feature importance weights of feature channels, accurate identification of load operating status is achieved, improving the accuracy and robustness of load operating status identification. This method not only solves the identification confusion problem caused by feature similarity in traditional load identification but also more effectively handles load patterns under complex operating conditions, enhancing the intelligence level of electrical appliance operating status monitoring and energy consumption management.

[0087] Example of a non-intrusive power load identification system based on a neural network model:

[0088] See Figure 2 The diagram illustrates a structural block diagram of a non-intrusive electrical load identification system based on a neural network model according to an embodiment of the present invention. The system may include a data acquisition module, a first evaluation module, a second evaluation module, and an identification module.

[0089] The data acquisition module is used to acquire the current signal of the user's electrical appliances during different detection cycles.

[0090] The first evaluation module is used to input the current signal into the neural network model to obtain a feature map; obtain the parameter vector of each feature channel in each detection cycle based on the gray-level distribution of different feature channels in the feature map in each detection cycle; and obtain the relative correlation coefficient of the load detection data of each feature channel in each detection cycle based on the distribution difference of the parameter vector of the same feature channel in each detection cycle and the adjacent previous detection cycle.

[0091] The second evaluation module is used to evaluate the channel information redundancy performance of each feature channel based on the overall distribution difference of the relative correlation coefficients of the load detection data in a single feature channel for all detection cycles and the parameter vector.

[0092] The identification module is used to introduce the ECA mechanism into the neural network model, use the redundancy of channel information to determine the feature importance weight of the feature channel, and determine the load operation status based on the feature importance weight.

[0093] It should be understood that Figure 2The block diagram and modules of the non-intrusive electrical load identification system based on a neural network model shown can be implemented in various ways. For example, in some embodiments, the system and its modules can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by appropriate instructions, such as a microprocessor or dedicated hardware. Those skilled in the art will understand that the above-described methods and apparatus can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The apparatus and modules of this specification can be implemented not only with hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., but also with software, for example, executed by various types of processors, or with a combination of the above-described hardware circuits and software (e.g., firmware).

[0094] For more details about the above modules, please refer to other parts of this manual; they will not be repeated here.

[0095] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-intrusive electrical load identification method based on a neural network model, characterized in that, The method includes the following steps: Acquire current signals from user electrical appliances during different detection cycles; The current signal is input into a neural network model to obtain a feature map; the parameter vector of each feature channel in each detection cycle is obtained based on the gray-level distribution of different feature channels in the feature map in each detection cycle; the relative correlation coefficient of the load detection data of each feature channel in each detection cycle is obtained based on the distribution difference of the parameter vector of the same feature channel in each detection cycle and the adjacent previous detection cycle. The channel information redundancy performance of each feature channel is evaluated based on the overall distribution difference of the relative correlation coefficients of the load detection data in a single feature channel for all detection cycles and the parameter vector. An ECA mechanism is introduced into the neural network model to determine the feature importance weights of feature channels by utilizing the redundancy of channel information; the load operation status is then determined based on these feature importance weights. The relative correlation coefficients for obtaining the load detection data of each feature channel within each detection cycle include: The range of the elements in the reference vector of each week's expectation analysis channel is denoted as the first eigenvalue of each week's expectation analysis channel; Calculate the first difference between the first eigenvalue of the candidate period and the first eigenvalue of the expected analysis channel in the previous week of the candidate period; Based on the first difference, the load steady-state coefficient of the candidate cycle expectation analysis channel is obtained; Calculate the second difference between the mean of the elements in the reference vector of the candidate week expectation analysis channel and the mean of the elements in the reference vector of the previous week expectation analysis channel. Calculate the second ratio between the load steady-state coefficient of the candidate cycle expectation analysis channel and the second difference; The normalized result of the second ratio is determined as the relative correlation coefficient of the load detection data of the candidate week expectation analysis channel; The candidate period is any detection period, and the channel to be analyzed is any feature channel; The evaluation of the channel information redundancy performance of each feature channel includes: Based on the relative correlation coefficient of the load detection data of the expected analysis channel in each detection week and the magnitude of the parameter vector of the expected analysis channel in each detection week, the load influence factor of the expected analysis channel in each detection week is obtained. The channel information redundancy performance of the channel to be analyzed is obtained by comparing the load influence factor of the channel to be analyzed in each detection week with the average load influence factor of the channel to be analyzed in all detection weeks. The step of obtaining the load steady-state coefficient of the candidate cycle expectation analysis channel based on the first difference includes: Obtain the maximum value of the first eigenvalue of all expected analysis channels for each week; Calculate the first ratio between the first difference and the maximum value, and use the negative correlation mapping result of the first ratio as the load steady-state coefficient of the candidate period expectation analysis channel.

2. The non-intrusive electrical load identification method based on a neural network model according to claim 1, characterized in that, The method of obtaining the parameter vector of each feature channel in each detection period based on the grayscale distribution of different feature channels in the feature map in each detection period includes: arranging the grayscale values ​​of all pixels in each feature channel in the feature map in each detection period to obtain the parameter vector of each feature channel in the feature map in each detection period.

3. The non-intrusive electrical load identification method based on a neural network model according to claim 1, characterized in that, The load influence factor for each expected analysis channel is obtained based on the relative correlation coefficient of the load detection data of each expected analysis channel in each detection cycle and the magnitude of the parameter vector of each expected analysis channel in each detection cycle, including: The ratio between the magnitude of the parameter vector of the expected analysis channel in each detection week and the relative correlation coefficient of the load detection data of the expected analysis channel in the same detection week is used as the load influence factor of the expected analysis channel in each detection week.

4. The non-intrusive electrical load identification method based on a neural network model according to claim 1, characterized in that, The method of obtaining the channel information redundancy performance of the channel to be analyzed based on the difference between the load influence factor of each channel to be analyzed in each detection cycle and the average load influence factor of all channels to be analyzed in all detection cycles includes: The negative correlation mapping value between the difference between the load influence factor of each detection week expectation analysis channel and the average load influence factor of all detection week expectation analysis channels is denoted as the second characteristic value of each detection week expectation analysis channel. The average value of the second feature value of all channels to be analyzed in the detection cycle is taken as the channel information redundancy of the channel to be analyzed.

5. The non-intrusive electrical load identification method based on a neural network model according to claim 1, characterized in that, The method of determining the feature importance weights of feature channels using channel information redundancy includes: Calculate the cumulative sum of channel information redundancy across all feature channels; The ratio of the redundancy of channel information in each feature channel to the sum is determined as the feature importance weight of each feature channel.

6. The non-intrusive electrical load identification method based on a neural network model according to claim 1, characterized in that, Determining the load operating status based on the feature importance weights includes: The feature importance weights are multiplied channel-by-channel with the original input feature map to obtain a feature map with channel attention; The load operating status is classified based on the feature map with channel attention, including single load, dual load, triple load, quadruple load and quintuple load operating status.

Images