A non-invasive load identification method and storage medium

Abstract

This application discloses a non-intrusive load identification method and storage medium. The method includes: slicing the collected total current data of the power load through a preset time window to obtain long-cycle data segments; performing dense overlapping sampling within the long-cycle data segments to generate two or more local sub-windows; calculating the saliency score of each local sub-window and arranging them in descending order to obtain N keyframes; extracting the long-range logical features of the long-cycle data segments and the local morphological fingerprint features of the keyframes through a multi-scale feature extraction model; calculating the mean of the saliency scores of the N keyframes to generate independent attention weights for the two features; weighting and concatenating the two features and inputting them into a classifier to output the operating status of each electrical appliance. This application effectively balances macroscopic control logic and microscopic transient fingerprints by using multi-scale heterogeneous feature extraction and non-mutually exclusive gating fusion, significantly improving the identification accuracy of complex mixed loads.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, specifically to a non-intrusive load identification method and storage medium. Background Technology

[0002] Non-intrusive load monitoring (NILM) aims to analyze the total voltage and current data at the user's power inlet to decompose the operating status and energy consumption information of each internal appliance. It is a key technology for achieving refined demand-side management and improving the intelligence level of the power grid. Compared with invasive monitoring, which requires installing sensors on each appliance, non-invasive methods have advantages such as low cost, high user acceptance, and ease of promotion.

[0003] Existing non-intrusive load identification methods have inherent contradictions in window settings: short time windows are difficult to fully capture the periodic duty cycle control logic of temperature-controlled appliances, easily misjudging the heating phase as a continuous heating appliance or the disconnection phase as a shutdown; while simply extending the window can cover the macro cycle, the transient load features that last only a few seconds are diluted by a large amount of background data in the long sequence, resulting in a decrease in the model's sensitivity to transient events; at the same time, the traditional slicing method with a fixed step size is prone to segmenting and destroying the start-up transient waveform that crosses the window boundary, causing feature truncation; in addition, the Softmax normalization mechanism widely used in multi-scale feature fusion forces the sum of feature weights to be 1, forming a zero-sum game, making it impossible for the model to take into account the effective synergy of strong start-up transients and regular duty cycles, thus limiting the identification accuracy of complex mixed loads.

[0004] Therefore, how to design a non-intrusive load identification technology that can retain long-term macro-control logic, focus on micro-transient morphological characteristics, and avoid the mutual exclusion constraints of feature truncation and fusion weights has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a non-intrusive load identification method and storage medium to solve the problem that existing non-intrusive load identification methods cannot simultaneously capture long-cycle duty cycle logic and micro-transient fingerprints, resulting in feature truncation and mutual exclusion fusion, which reduces the identification accuracy of complex loads.

[0006] This application provides a non-intrusive load identification method, which specifically includes a data acquisition step, a data slicing step, a dense overlapping sampling step, a keyframe acquisition step, a model building step, a feature extraction step, and a running status acquisition step.

[0007] The data acquisition step involves collecting total current data of the power load at the user's electricity meter at a sampling rate of no less than 6400Hz; the data slicing step uses a preset time window to slice the acquired total current data of the power load, obtaining long-period data segments; the dense overlapping sampling step uses a preset duration and a preset sliding step size to perform dense sampling within the long-period data segments, generating two or more local sub-windows; the keyframe acquisition step calculates the saliency score of the local sub-windows, and based on the obtained saliency score, arranges the local sub-windows in descending order of saliency score, obtaining N keyframes; the model construction step uses an asymmetric heterogeneous modeling strategy to construct a multi-model... The multi-scale feature extraction model includes a parallel bidirectional state-space model and a one-dimensional convolutional neural network. The feature extraction step, based on the multi-scale feature extraction model, extracts long-range logical features of the long-period data segment and local morphological fingerprint features of the N keyframes. The operating status acquisition step calculates the mean saliency score of the N keyframes, generates independent attention weights for the long-range logical features and the local morphological fingerprint features, concatenates the weighted long-range logical features and the local morphological fingerprint features, and inputs them into the classifier of the multi-scale feature extraction model to output the operating status of each electrical appliance corresponding to the total power load current data.

[0008] Furthermore, the keyframe acquisition step specifically includes a saliency score calculation step and a keyframe filtering step.

[0009] The saliency score calculation step is used to extract the temporal fingerprint index of the local sub-window, and calculate the saliency score of the local sub-window based on the obtained temporal fingerprint index, the formula of which is:

[0010] ;

[0011] ;

[0012] ;

[0013] Among them, S i This represents the saliency score of the i-th local sub-window. This represents the coefficient of the first normalization constant. V represents the set second normalization constant coefficient. i E represents the energy volatility of the i-th local sub-window. i Let p represent the information entropy of the i-th local sub-window, exp represent the exponential function, K represent the set number of intervals for dividing the current signal amplitude within the local sub-window, and p j This represents the probability that the current signal falls into the j-th interval. This represents the standard deviation of the current or power signal within the i-th local sub-window. This represents the average current signal within the i-th local sub-window. This represents a very small constant.

[0014] The keyframe filtering step is based on the saliency score of the obtained local sub-windows. The local sub-windows are sorted in descending order of saliency score, and the top N local sub-windows are the N keyframes obtained.

[0015] Furthermore, the feature extraction step specifically includes a global feature acquisition step and a local feature acquisition step.

[0016] The global feature acquisition step involves inputting the long-period data segment into the multi-scale feature extraction model and extracting the long-range logical features of the long-period data segment through the bidirectional state space model; the local feature acquisition step involves inputting the key frame into the multi-scale feature extraction model and extracting the local morphological fingerprint features of the key frame through the one-dimensional convolutional neural network.

[0017] Furthermore, the running state acquisition step includes a transient intensity characterization calculation step. This step calculates the mean of the saliency scores of the N keyframes as the transient intensity characterization value, using the following formula:

[0018] ;

[0019] in, This represents the transient intensity characterization value, where N represents the number of keyframes. This represents the saliency score corresponding to the k-th keyframe.

[0020] Furthermore, the operation status acquisition step specifically includes a local original bias calculation step and a global original bias calculation step.

[0021] The local original bias calculation step is based on the mean of the obtained saliency scores. Using an independent nonlinear gating mechanism, the local original bias of the local morphological fingerprint features is calculated through the linear mapping layer of the gating network of the multi-scale feature extraction model. The formula is as follows:

[0022] ;

[0023] in, Indicates the local original bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network.

[0024] The global original bias calculation step is based on the mean of the obtained saliency scores. Using an independent nonlinear gating mechanism, the global original bias of the long-range logistic features is calculated through the linear mapping layer of the gated network of the multi-scale feature extraction model. The formula is as follows:

[0025] ;

[0026] in, This represents the global primitive bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network.

[0027] Furthermore, the running state acquisition step also includes an attention weight generation step, which generates non-mutually exclusive global attention weights and local attention weights based on the obtained global original bias and local original bias, using the Sigmoid activation function.

[0028] Furthermore, the attention weight generation step specifically includes a local attention weight generation step and a global attention weight generation step.

[0029] The local attention weight generation step is based on the obtained local original bias, and generates local attention weights through the Sigmoid activation function, with the following formula:

[0030] ;

[0031] in, This represents the local attention weight.

[0032] The global attention weight generation step is based on the obtained global original bias, and generates global attention weights through the Sigmoid activation function, with the formula as follows:

[0033] ;

[0034] in, This represents the global attention weight.

[0035] Furthermore, the operation status acquisition step also includes a weighting step, a splicing step, and an operation status output step.

[0036] The weighting step involves multiplying the local morphological fingerprint features by the local attention weight and then multiplying the long-range logical features by the global attention weight. The concatenation step involves concatenating the weighted local morphological fingerprint features and the weighted long-range logical features along the channel dimension to obtain a concatenated feature vector. The operating status output step involves inputting the concatenated feature vector into the classifier of the multi-scale feature extraction model and outputting the operating status of each electrical appliance corresponding to the total power load current data.

[0037] Furthermore, the duration of the time window at least covers the minimum complete start-stop control cycle of temperature-controlled electrical appliances in the total power load current data.

[0038] This application also provides a storage medium storing computer-readable instructions that, when read by at least one processor, cause at least one processor to perform at least one step in the non-intrusive load identification method.

[0039] This application provides a non-intrusive load identification method and storage medium. By constructing a long-cycle time window covering the complete start-stop cycle of temperature-controlled appliances and performing dense overlapping sampling and explicit saliency calculation within it, the method effectively solves the problems of transient signal dilution and feature truncation caused by fixed slices under long windows. At the same time, it adopts an asymmetric heterogeneous dual-branch architecture to extract long-range duty cycle logic and local transient morphological fingerprints respectively. The method achieves non-exclusive weighted fusion of the two features through an independent Sigmoid gating mechanism, breaking the zero-sum game constraint of traditional Softmax fusion. This allows the model to simultaneously take into account macroscopic control logic and microscopic physical fingerprints, significantly improving the identification accuracy and robustness of complex mixed loads including temperature-controlled appliances, transient appliances, and multi-state appliances. Attached Figure Description

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

[0041] Figure 1 This is a flowchart of the non-intrusive load identification method described in the embodiments of this application;

[0042] Figure 2 This is a flowchart of the keyframe acquisition steps described in the embodiments of this application;

[0043] Figure 3 This is a flowchart of the feature extraction steps described in the embodiments of this application;

[0044] Figure 4 This is a flowchart of the steps for obtaining the running status as described in the embodiments of this application;

[0045] Figure 5 This is a flowchart of the attention weight generation steps described in the embodiments of this application;

[0046] Figure 6 This is a schematic diagram of the storage medium and processor described in the embodiments of this application.

[0047] Explanation of reference numerals in the attached figures:

[0048] 100 storage media, 200 processors. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] like Figure 1 As shown, this application provides a non-intrusive load identification method, which specifically includes steps S1 (data acquisition), S2 (data slicing), S3 (dense overlapping sampling), S4 (keyframe acquisition), S5 (model construction), S6 (feature extraction), and S7 (running status acquisition).

[0051] Step S1, the data acquisition step, involves collecting the total current data of the power load at the user's electricity meter at a sampling rate of not less than 6400Hz.

[0052] Step S2 is a data slicing step. A time window is preset to slice the acquired total power load current data to obtain long-cycle data segments. The time window is a long-cycle time window, and the duration of the time window at least covers the minimum complete start-stop control cycle of temperature control appliances in the total power load current data.

[0053] Its technical advantage lies in ensuring the complete capture of macroscopic duty cycle characteristics by setting a long-cycle time window that covers the complete start-stop cycle of the temperature control appliance.

[0054] Step S3, the dense overlapping sampling step, uses a preset duration and a preset sliding step size to perform dense sampling within the long periodic data segment, generating two or more local sub-windows;

[0055] The technical advantage lies in the fact that dense overlapping sampling is performed within a long-period window to generate multiple local sub-windows. This ensures that any millisecond-level start-up transient signal can be completely covered by at least one sub-window. This avoids the feature truncation problem caused by traditional fixed-step slicing and effectively solves the dilution effect of long-period windows on transient signals by densely focusing on local regions in a long-period background. This preserves a complete and comparable transient candidate set for subsequent keyframe selection.

[0056] In one feasible embodiment of this application, the duration of the time window is set to 180 seconds, the preset duration of the local sub-window is set to 10 seconds, and the preset sliding step size of the local sub-window is set to 5 seconds. A total of 35 sub-windows with a 50% overlap rate are generated within the long period time window of 180 seconds.

[0057] Step S4 is the keyframe acquisition step. Calculate the saliency score of the local sub-window. Based on the obtained saliency score, arrange the local sub-windows in descending order of saliency score to obtain N keyframes.

[0058] like Figure 2 As shown, step S4, the keyframe acquisition step, specifically includes step S41, the saliency score calculation step, and step S42, the keyframe filtering step.

[0059] Step S41, the saliency score calculation step, involves extracting the temporal fingerprint index of the local sub-window and calculating the saliency score of the local sub-window based on the obtained temporal fingerprint index. The formula is as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] Among them, S i This represents the saliency score of the i-th local sub-window. This represents the coefficient of the first normalization constant. V represents the set second normalization constant coefficient. i E represents the energy volatility of the i-th local sub-window. i Let p represent the information entropy of the i-th local sub-window, exp represent the exponential function, K represent the set number of intervals for dividing the current signal amplitude within the local sub-window, and p j This represents the probability that the current signal falls into the j-th interval. This represents the standard deviation of the current or power signal within the i-th local sub-window. This represents the average current signal within the i-th local sub-window. This represents a very small constant.

[0064] Step S42, the keyframe filtering step, involves sorting the local sub-windows in descending order of their salience scores based on the obtained scores, with the top N local sub-windows being the N keyframes obtained.

[0065] Its technical effectiveness lies in constructing an explicit saliency calculation formula based on energy fluctuation rate and information entropy, and performing unsupervised transient intensity quantification on all local sub-windows generated by densely overlapping sampling. It can accurately select N keyframes containing high-fluctuation, low-entropy transient spike signals without manual annotation. This mechanism directly achieves the screening of effective transient information within long-period windows from a mathematical operation level, which not only avoids the computational redundancy and poor interpretability problems caused by implicit search in neural networks, but also fundamentally solves the dilution effect of long-term windows on transient signals by eliminating a large number of silent background sub-windows, providing high-quality input data for accurate modeling of subsequent local branches.

[0066] Step S5 is the model construction step, in which an asymmetric heterogeneous modeling strategy is used to construct a multi-scale feature extraction model. The multi-scale feature extraction model includes a parallel bidirectional state space model and a one-dimensional convolutional neural network. The bidirectional state space model is a Bi-Mamba model. The Bi-Mamba model branch can effectively capture the global dependencies in long-period data segments and the periodic duty cycle logic of temperature control appliances through its bidirectional state space structure.

[0067] The technical advantage lies in achieving multi-scale decoupling and differentiated modeling of load features through the parallel design of a bidirectional state-space model and a one-dimensional convolutional neural network. The bidirectional state-space model, with its long-sequence modeling capability, fully captures the global contextual information and the periodic duty cycle logic of the temperature control appliance within long-period data segments. The one-dimensional convolutional neural network focuses on extracting microscopic transient morphological fingerprints from the selected N key frames. This asymmetric heterogeneous architecture enables the multi-scale feature extraction model to simultaneously consider load features with different time scales and physical meanings, laying a solid feature foundation for subsequent accurate fusion and identification.

[0068] Step S6, feature extraction step, involves extracting the long-range logical features of the long-period data segment and the local morphological fingerprint features of the N key frames based on the multi-scale feature extraction model.

[0069] like Figure 3 As shown, step S6, feature extraction, specifically includes step S61, global feature acquisition, and step S62, local feature acquisition.

[0070] Step S61, the global feature acquisition step, involves inputting the long-period data segment into the multi-scale feature extraction model, and extracting the long-range logistic features of the long-period data segment through the bidirectional state-space model. .

[0071] Step S62, the local feature acquisition step, involves inputting the key frame into the multi-scale feature extraction model and extracting the local morphological fingerprint features of the key frame through the one-dimensional convolutional neural network.

[0072] Specifically, the selected N keyframes are input into a one-dimensional convolutional neural network to extract N independent local feature vectors V. k The saliency scores are normalized and weighted, then fused into a unique fixed-length local morphological feature vector. This fully preserves multiple valid transient fingerprints that may exist within the long window. The formula is as follows:

[0073] ;

[0074] ;

[0075] in, This represents the saliency score corresponding to the k-th keyframe. This represents the original saliency score corresponding to the nth keyframe. This represents a fixed-length local morphological feature vector, i.e., the local morphological fingerprint feature, where N represents the number of keyframes. This represents the normalized aggregate weight corresponding to the k-th keyframe.

[0076] Its technical advantage lies in the fact that by inputting long-period data segments into the Bi-Mamba model to extract long-range logical features, and simultaneously inputting the selected N key frames into a one-dimensional convolutional neural network to extract local morphological fingerprint features, it achieves the decoupling and parallel extraction of macroscopic duty cycle information and microscopic transient waveform information. This heterogeneous feature extraction strategy enables the model to obtain two complementary load representations at two scales from the same raw data—the global branch captures the periodic control logic of the temperature control appliance, while the local branch focuses on the refined morphology of transient events, providing a feature foundation with complete information and controllable dimensions for subsequent dynamic fusion.

[0077] Step S7 is the running status acquisition step. The mean of the saliency scores of the N keyframes is calculated, the independent attention weights of the long-range logical features and the local morphological fingerprint features are generated, the long-range logical features and the local morphological fingerprint features are weighted and concatenated, and then input into the classifier of the multi-scale feature extraction model to output the running status of each electrical appliance corresponding to the total power load current data.

[0078] like Figure 4As shown, step S7, the running state acquisition step, includes step S71, transient intensity characterization calculation step, step S72, local original bias calculation step, step S73, global original bias calculation step, step S74, attention weight generation step, step S75, weighting step, step S76, splicing step, and step S77, running state output step.

[0079] Step S71, the transient intensity characterization calculation step, involves calculating the mean of the saliency scores of the N keyframes as the transient intensity characterization value, using the following formula:

[0080] ;

[0081] in, This represents the transient intensity characterization value, where N represents the number of keyframes. This represents the saliency score corresponding to the k-th keyframe.

[0082] The technical effect lies in the fact that by aggregating the saliency scores of the selected N keyframes, the transient intensity information of multiple discrete sub-windows is compressed into a scalar value that can characterize the overall transient activity level of the current long-period window. The transient intensity characterization value serves as the input for subsequent gating mechanisms, enabling the model to dynamically adjust the fusion weights of global and local features based on the richness of transient events within the entire window. This achieves adaptive gating adjustment based on the characteristics of the data itself, providing a data-driven decision-making basis for breaking away from traditional fixed weights or zero-sum game fusion.

[0083] Step S72, the local original bias calculation step, involves calculating the local original bias of the local morphological fingerprint features based on the mean of the obtained saliency scores using an independent nonlinear gating mechanism through the linear mapping layer of the gating network of the multi-scale feature extraction model. The formula is as follows:

[0084] ;

[0085] in, Indicates the local original bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network.

[0086] Step S73, the global original bias calculation step, is based on the mean of the obtained saliency scores. Using an independent nonlinear gating mechanism, the global original bias of the long-range logistic features is calculated through the linear mapping layer of the gating network of the multi-scale feature extraction model. The formula is as follows:

[0087] ;

[0088] in, This represents the global primitive bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network.

[0089] Its technical advantage lies in introducing two independent sets of learnable linear mapping layers to represent transient intensity values. Original biases mapped to local features respectively Original bias with global features This design enables the gating mechanism to autonomously learn the baseline activation levels of local and global branches through a data-driven approach based on the overall transient activity level of the current window. This provides a parameterized adaptive control basis for the subsequent generation of non-mutually exclusive attention weights. At the same time, the setting of two independent sets of parameters ensures that the generation process of local and global weights does not interfere with each other, laying a structural foundation for breaking the zero-sum game constraint of traditional Softmax fusion.

[0090] Step S74, the attention weight generation step, generates non-mutually exclusive global attention weights and local attention weights based on the obtained global original bias and local original bias, using the Sigmoid activation function.

[0091] like Figure 5 As shown, step S74, the attention weight generation step, specifically includes step S78, the local attention weight generation step, and step S79, the global attention weight generation step.

[0092] Step S78, the local attention weight generation step, involves generating local attention weights based on the obtained local original bias using the Sigmoid activation function. The formula is as follows:

[0093] ;

[0094] in, This represents the local attention weight.

[0095] Step S79, the global attention weight generation step, involves generating global attention weights based on the obtained original global bias using the sigmoid activation function. The formula is as follows:

[0096] ;

[0097] in, This represents the global attention weight.

[0098] Its technical effect lies in achieving the effect of locally applying the original bias. Compared with the global original bias The Sigmoid activation function was input independently, generating local attention weights α and global attention weights β, which are uncoupled and have values ​​ranging from (0,1). This independent nonlinear mapping mechanism breaks the zero-sum game constraint of "the sum of weights is 1" in the traditional Softmax gating, allowing α and β to take values ​​independently according to the original bias of their respective branches. They can be increased simultaneously to strengthen the contribution of dual-path features, or adaptively increased and decreased to highlight the dominant information. This fundamentally solves the mutual exclusion problem in multi-scale feature fusion and provides a non-mutually exclusive and dynamically balanced weight foundation for subsequent flexible weighted concatenation.

[0099] Step S75 is a weighting step, in which the local morphological fingerprint features are multiplied by the local attention weight for weighting, and the long-range logical features are multiplied by the global attention weight for weighting.

[0100] Step S76, the concatenation step, concatenates the weighted local morphological fingerprint features and the weighted long-range logical features along the channel dimension to obtain the concatenated feature vector. .

[0101] Step S77 is the running status output step, in which the spliced ​​feature vector is input into the classifier of the multi-scale feature extraction model, and the running status of each electrical appliance corresponding to the total power load current data is output.

[0102] The technical effect lies in achieving adaptive dynamic weighting of two heterogeneous features by multiplying local morphological fingerprint features and long-range logical features by their independently generated attention weights α and β, respectively. The weighted features are then concatenated along the channel dimension to form a joint feature vector that integrates macroscopic duty cycle information and microscopic transient morphological information. This vector is then input into a classifier to complete end-to-end identification of the appliance's operating status. This fusion mechanism allows the model to flexibly adjust the contribution ratio of the two types of features based on the transient activity level of the current window. It strengthens the dominant role of local morphological features during strong transient events and highlights the discriminative value of global logical features under steady-state conditions. Furthermore, due to the non-mutually exclusive nature of the weights, the two types of effective information are fully preserved rather than suppressed, significantly improving the identification accuracy and generalization ability in complex mixed load scenarios.

[0103] like Figure 6 As shown, this application also provides a storage medium 100 storing computer-readable instructions that, when read by at least one processor 200, cause at least one processor 200 to perform at least one step in the non-intrusive load identification.

[0104] The advantages of this application are that it provides a non-intrusive load identification method and storage medium. By constructing a long-cycle time window covering the complete start-stop cycle of the temperature control appliance and performing dense overlapping sampling and explicit saliency calculation within it, it effectively solves the problems of transient signal dilution and feature truncation caused by fixed slices under long windows. At the same time, it adopts an asymmetric heterogeneous dual-branch architecture to extract long-range duty cycle logic and local transient morphological fingerprints respectively. It also achieves non-mutually exclusive weighted fusion of the two features through an independent Sigmoid gating mechanism, breaking the zero-sum game constraint of traditional Softmax fusion. This allows the model to simultaneously take into account macroscopic control logic and microscopic physical fingerprints, significantly improving the identification accuracy and robustness of complex mixed loads including temperature control appliances, transient appliances, and multi-state appliances.

[0105] The above provides a detailed description of the non-intrusive load identification method and storage medium provided by this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A non-intrusive load disaggregation method, characterized in that, Specifically, the steps include the following: The data acquisition step involves collecting the total current data of the power load at the user's electricity meter at a sampling rate of no less than 6400Hz. The data slicing step involves setting a time window and slicing the acquired total power load current data to obtain long-cycle data segments. The densely overlapping sampling step uses a preset duration and a preset sliding step to perform dense sampling within the long-period data segment, generating two or more local sub-windows. The keyframe acquisition step involves calculating the saliency score of the local sub-window, arranging the local sub-windows in descending order of saliency score based on the obtained saliency score, and acquiring N keyframes. The N keyframes represent the top N local sub-windows in descending order of saliency score among the local sub-windows. The model building steps involve constructing a multi-scale feature extraction model using an asymmetric heterogeneous modeling strategy. The multi-scale feature extraction model includes a parallel bidirectional state space model and a one-dimensional convolutional neural network. The asymmetric heterogeneous modeling strategy means that the parallel design of the bidirectional state space model and the one-dimensional convolutional neural network forms an asymmetric heterogeneous architecture multi-scale feature extraction model, which enables the multi-scale feature extraction model to simultaneously take into account load features with different time scales and different physical meanings. The feature extraction step involves extracting the long-range logical features of the long-period data segment and the local morphological fingerprint features of the N key frames based on the multi-scale feature extraction model. The long-range logic feature represents the global context information and the periodic duty cycle logic of the temperature control appliance in the long-period data segment fully captured by the bidirectional state-space model. The local morphological fingerprint feature representation of the key frame is obtained by inputting N key frames into the one-dimensional convolutional neural network, extracting local feature vectors of N key frames, and using the saliency score to normalize and weight the N local feature vectors to form a unique fixed-length local morphological feature vector. The operation status acquisition step involves calculating the mean of the saliency scores of the N keyframes, generating independent attention weights for the long-range logical features and the local morphological fingerprint features, concatenating the weighted long-range logical features and the local morphological fingerprint features, and inputting them into the classifier of the multi-scale feature extraction model to output the operation status of each electrical appliance corresponding to the total power load current data.

2. The non-invasive load identification method as described in claim 1, characterized in that, The keyframe acquisition step specifically includes the following steps: The saliency score calculation steps involve extracting the temporal fingerprint index of the local sub-window, and calculating the saliency score of the local sub-window based on the obtained temporal fingerprint index. The formula is as follows: ; ; ; Among them, S i This represents the saliency score of the i-th local sub-window. This represents the set normalization constant coefficient. V represents the set normalization constant coefficient. i E represents the energy fluctuation rate of the i-th local sub-window. i Let p represent the information entropy of the i-th local sub-window, K represent the set number of intervals for dividing the current signal amplitude within the local sub-window, and p j This represents the probability that the current signal falls into the j-th interval. This represents the standard deviation of the current or power signal within the i-th local sub-window. This represents the average current signal within the i-th local sub-window. Denotes a minimal constant; and In the keyframe filtering step, based on the saliency scores of the obtained local sub-windows, the local sub-windows are sorted in descending order of saliency scores, and the top N local sub-windows are the N keyframes obtained.

3. The non-invasive load identification method as described in claim 1, characterized in that, The feature extraction step specifically includes the following steps: The global feature acquisition step involves inputting the long-period data segment into the multi-scale feature extraction model, and extracting the long-range logistic features of the long-period data segment through the bidirectional state-space model; and The local feature acquisition step involves inputting the keyframe into the multi-scale feature extraction model and extracting the local morphological fingerprint features of the keyframe through the one-dimensional convolutional neural network.

4. The non-invasive load identification method as described in claim 1, characterized in that, The steps for obtaining the operating status include: The transient intensity characterization calculation step involves calculating the mean of the saliency scores of the N keyframes as the transient intensity characterization value, using the following formula: in, This represents the transient intensity characterization value, where N represents the number of keyframes. This represents the saliency score corresponding to the k-th keyframe.

5. The non-invasive load identification method as described in claim 4, characterized in that, The step of obtaining the operating status also includes: The local original bias calculation step involves calculating the local original bias of the local morphological fingerprint features based on the mean of the obtained saliency scores, using an independent nonlinear gating mechanism through the linear mapping layer of the gating network of the multi-scale feature extraction model. The formula is as follows: in, Indicates the local original bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network; The global original bias calculation step involves calculating the global original bias of the long-range logistic features based on the mean of the obtained saliency scores, using an independent nonlinear gating mechanism through the linear mapping layer of the gating network of the multi-scale feature extraction model. The formula is as follows: in, This represents the global primitive bias. This represents the learnable weight parameters in the linear mapping layer of the gated network. This represents the learnable bias parameters in the linear mapping layer of the gated network.

6. The non-invasive load identification method as described in claim 5, characterized in that, The step of obtaining the operating status also includes: The attention weight generation step, based on the obtained global and local original biases, generates non-mutually exclusive global and local attention weights through the Sigmoid activation function.

7. The non-invasive load identification method as described in claim 6, characterized in that, The attention weight generation step specifically includes the following steps: The local attention weight generation step involves generating local attention weights based on the obtained local initial biases using the sigmoid activation function. The formula is as follows: in, Represents local attention weights; The global attention weight generation step involves generating global attention weights based on the obtained initial global bias using the sigmoid activation function. The formula is as follows: in, This represents the global attention weight.

8. The non-invasive load identification method as described in claim 6, characterized in that, The step of obtaining the operating status also includes: The weighting step involves multiplying the local morphological fingerprint features by the local attention weight and then multiplying the long-range logical features by the global attention weight. The splicing step involves splicing the weighted local morphological fingerprint features and the weighted long-range logical features along the channel dimension to obtain the spliced ​​feature vector. The operation status output step involves inputting the spliced ​​feature vector into the classifier of the multi-scale feature extraction model and outputting the operation status of each electrical appliance corresponding to the total power load current data.

9. The non-invasive load identification method as described in claim 1, characterized in that, The duration of the time window must at least cover the minimum complete start-stop control cycle of temperature-controlled appliances in the total power load current data.

10. A storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are read by at least one processor, the at least one processor performs at least one step in the non-intrusive load identification method as described in any one of claims 1 to 9.

Images