Non-intrusive power disaggregation method based on bidirectional long short-term memory network

Abstract

The application relates to the technical field of industrial energy fine management, and discloses a non-intrusive power meter power decomposition method based on a bidirectional long short-term memory network, a physical-time sequence-process three-dimensional equipment classification system is constructed; a "multiple machines one meter" equipment is set with a label according to the classification; an attention mask matrix is generated based on the equipment label; a non-intrusive load decomposition model integrating a bidirectional long short-term memory and an attention mechanism is designed; real-time power data of each independent equipment under a shared power meter is accurately decomposed; the uncertainty of the non-intrusive power decomposition result is quantitatively estimated through a Monte Carlo Dropout estimation, reliability evaluation is provided for equipment-level energy consumption data; and the decomposed energy consumption is associated with specific products through an equipment power-process path mapping matrix. The application solves the problem of missing equipment-level real-time energy consumption data caused by the heterogeneous configuration of the key equipment one machine one meter and the multiple machines one meter of the medium and small power equipment of a whole vehicle manufacturing enterprise.

Description

Technical Field

[0001] This invention relates to the field of industrial energy refinement management technology, and in particular to a non-intrusive meter power decomposition method based on bidirectional long short-term memory networks. Background Technology

[0002] With the advancement of the "dual-carbon" strategy and the development of intelligent manufacturing technology, vehicle manufacturers are gradually establishing energy management systems to achieve real-time monitoring and refined management of energy sources such as electricity, gas, and water during the production process. In the vehicle manufacturing process, welding, painting, and final assembly workshops contain a large number of power equipment, conveying equipment, auxiliary equipment, and process equipment, whose energy consumption is complex, highly volatile, and significantly coupled with load.

[0003] In actual engineering deployments, due to cost control and construction complexity constraints, energy monitoring systems typically adopt the following configuration: critical high-power equipment is metered independently using a "one machine, one meter" approach; a large number of small and medium-power equipment are centrally metered using a "multiple machines, one meter" approach. While this heterogeneous metering structure reduces construction costs, it also introduces the problem of missing equipment-level energy consumption data. Specifically: shared meters can only provide total power data for multiple devices, failing to distinguish the real-time power of individual devices; they cannot accurately assess the energy consumption level and operating efficiency of individual devices; they struggle to identify devices with abnormal energy consumption; and they cannot achieve precise allocation of equipment-level energy consumption to product-level energy consumption. Therefore, in the "multiple machines, one meter" scenario, extracting the power of each individual device from the total power data becomes a key technical issue for refined energy management.

[0004] To address the aforementioned issues, existing technologies typically employ non-intrusive load decomposition (NILM) methods. This method identifies and separates the power consumption behavior of different devices by analyzing the characteristics of the total current or total power signal. NILM offers a potential solution to the energy consumption breakdown problem in "multiple devices per meter" scenarios. This method aims to analyze the energy consumption of each device based solely on the total meter power signal and is currently widely used in residential and commercial building scenarios. Compared to residential electricity consumption scenarios, vehicle manufacturing enterprises have the following characteristics: a large number of devices and frequent load overlap; complex processes and process-dependent equipment operation; obvious periodicity and phased nature of equipment loads; dynamic changes in production cycle time and non-stationary load patterns; and the need to allocate energy consumption from equipment level to product level. Under the energy monitoring structure of "one meter per key device and multiple meters per small and medium power device," the lack of real-time power data at the device level is particularly prominent.

[0005] Therefore, there is an urgent need for a non-intrusive method for decomposing electricity meter power that can integrate industrial semantic information, time series deep modeling capabilities, and uncertainty quantification mechanisms. Summary of the Invention

[0006] This invention provides a non-intrusive meter power decomposition method based on bidirectional long short-term memory networks. By introducing attention mask constraints based on structured device tags and physical conservation losses, it effectively suppresses inter-device feature interference and eliminates power drift under complex industrial loads, achieving high-precision device-level power decoupling and accurate product-level carbon emission traceability in the "multiple machines, one meter" scenario.

[0007] This invention provides a non-intrusive meter power decomposition method based on a bidirectional long short-term memory network, applicable to multiple devices powered by the same shared meter, where the multiple devices jointly process a target product. The method includes the following steps:

[0008] S1. Hierarchically label multiple devices to generate structured device tags for each device; wherein, the structured device tags include the physical characteristics, temporal operation characteristics, process characteristics and coupling relationships between devices for each device;

[0009] S2. Collect the total power raw data of the shared electricity meter and extract multi-dimensional feature sequences from it. Construct an attention mask matrix representing the correlation strength between multiple devices based on the structured device labels of each device. The multi-dimensional feature sequences are used to input into the non-intrusive load decomposition model to extract hidden feature representations. The attention mask matrix is ​​used to perform attention weight calculation in collaboration with the hidden feature representations in the non-intrusive load decomposition model.

[0010] S3. Construct a non-intrusive load decomposition model including a bidirectional long short-term memory network layer and a masked attention layer, and train the non-intrusive load decomposition model using a sample dataset constructed based on historical total power data and historical actual power data of each device.

[0011] S4. Collect the real-time total power data of the shared electricity meter and input it into the trained non-intrusive load decomposition model after preprocessing. Output the real-time initial predicted power value of each device and perform normalization correction to obtain the real-time active power data of each device. Compare the real-time active power data of each device with the preset state judgment threshold to obtain the operating status identifier of each device.

[0012] S5. Obtain the production process information of the target product, construct a device-process mapping matrix representing the correspondence between multiple devices and the manufacturing links of the target product, and calculate the cumulative energy consumption of the target product during the manufacturing period by combining the real-time active power data and operating status identifier of each device, and multiply the cumulative energy consumption by a preset carbon emission factor per unit of electricity to obtain the carbon emission of the target product.

[0013] Furthermore, S1 specifically includes:

[0014] S101. Classify equipment according to energy conversion type, load electrical characteristics and mechanical dynamic characteristics, and extract multi-dimensional physical parameters to construct physical feature vectors for each equipment by combining the electrical characteristics and power levels of the multiple devices.

[0015] S102. Based on the energy consumption fluctuation pattern, multiple devices are divided into steady-state, transient, and periodic devices, and based on the on-site operation records of the devices, the time-series operation mode of each device is constructed.

[0016] S103. Classify and label multiple devices as key devices, auxiliary devices, or infrastructure according to the degree of process coreness to construct the process characteristic classification of each device, and determine the inter-device coupling relationship used to represent the strength of process or physical association between devices. Combine the physical feature vector, time sequence operation mode, process characteristic classification and inter-device coupling relationship of each device to generate a structured device label for each device.

[0017] Further, in S101, the physical feature vector The expression is as follows:

[0018]

[0019] Where i represents the device number; These represent the three-phase voltage levels; These represent the three-phase currents respectively; PF represents total active power; F represents power factor; F represents frequency.

[0020] In S102, the timing operation mode of the i-th device The expression is as follows:

[0021]

[0022] Where t represents the current timestamp; Indicates the device startup time; Indicates the end time of the device; when the device is in operation. When the equipment is in a stopped state .

[0023] Furthermore, S2 specifically includes:

[0024] S201. Collect the raw waveform data and derived data of the total current, voltage, active power and reactive power of the shared meter, wherein each collected data point is timestamped.

[0025] S202. After preprocessing the original waveform data and derived data, extract the time series data, and combine the structured device tags of each device to extract multi-dimensional feature vectors, generating a multi-dimensional feature sequence. Each x represents a vector composed of all features extracted at time t; wherein, the multidimensional feature sequence is used as input to the bidirectional long short-term memory network layer of the non-intrusive load decomposition model to extract forward and backward time-dependent hidden feature representations;

[0026] S203. Based on the structured device labels of each device, construct an attention mask matrix; wherein, the attention mask matrix is ​​used to perform attention weight calculation in conjunction with the hidden feature representation in the mask attention layer of the non-intrusive load decomposition model.

[0027] Furthermore, in S203, the attention mask matrix is ​​a binary mask matrix M_mask, which is constructed as follows:

[0028] Create an N×N matrix and map it to time steps, where N is the number of devices; when device i and device j have no process coupling, specify M_mask[i,j]=-∞; when device i and device j have strong timing linkage, specify M_mask[i,j]=0; or,

[0029] The attention mask matrix is ​​a mask matrix that contains multidimensional correlation strengths. Its specific construction method is as follows:

[0030] Construct a process tag matrix based on the process segments in which the equipment participates. Where, when device i belongs to process segment k, it is specified Otherwise specify The physical feature vector, timing operation mode, and process label matrix of the corresponding equipment are combined to form a three-dimensional label vector for each equipment. ; Calculate the physical feature similarity between devices based on the three-dimensional label vectors. Temporal feature similarity and similarity of process features The mask matrix is ​​constructed according to the following formula. :

[0031]

[0032] in, , , The similarity of the process features is a preset coefficient. The calculation formula is And when equipment i and equipment j belong to different process segments, the following conditions are met. .

[0033] Furthermore, S3 specifically includes:

[0034] S301. Construct a non-intrusive load decomposition model using an encoder-decoder framework, wherein the non-intrusive load decomposition model receives the sample feature sequence as input in the form of a sliding time window data.

[0035] S302. The sample feature sequence is fed into the bidirectional long short-term memory network layer. During the forward and backward feature extraction processes, forward time-dependent features and backward time-dependent features are captured respectively, and the two are concatenated to form a hidden feature representation containing complete contextual information.

[0036] S303. Input the hidden feature representation into the mask attention layer, use the attention mask matrix to constrain similarity calculation to mask the feature information of non-associated devices, calculate the attention weights between multiple devices, and combine the attention weights to generate the initial predicted power value of each device through the fully connected output layer.

[0037] S304. Construct a joint loss function that is a weighted combination of the basic regression loss and the total power conservation constraint loss. Use the sample dataset to optimize the joint loss function through the backpropagation algorithm to update the network parameters. Configure a Monte Carlo Dropout estimation mechanism in the non-intrusive load decomposition model to quantify the uncertainty of the decomposition results during the inference phase.

[0038] Furthermore, step S301 further includes a step of time-frequency feature fusion optimization of the input data:

[0039] The temporal feature representation of the sample feature sequence is extracted using a multi-scale one-dimensional convolutional layer. Simultaneously, the frequency domain features are obtained by performing a Fast Fourier Transform on the sample feature sequence. Frequency domain weights are calculated using the Softmax function. And obtain the frequency domain feature representation through frequency domain weighting. Representing the time-domain features With the frequency domain feature representation The fused feature sequence is obtained by splicing and merging the sequences. The frequency domain weights are input to the subsequent feature extraction layer. The calculation formula is:

[0040]

[0041] In S304, the joint loss function The specific mathematical expression is:

[0042]

[0043] In the formula, These are the preset conservation constraint weighting coefficients; Based on regression loss; The total power conservation constraint loss; the basic regression loss With the total power conservation constraint loss The calculation formulas are as follows:

[0044]

[0045]

[0046] In the formula, T is the total number of time steps for the training samples; N is the number of the multiple devices; This refers to the historical actual power data of the i-th device; This represents the predicted power value of the i-th device output by the model. This refers to the historical total power data.

[0047] Furthermore, in step S303, when the attention mask matrix is ​​a binary mask matrix M_mask, the step of constraining the similarity calculation using the attention mask matrix specifically includes: calculating the similarity score between the decoder query state and the encoder hidden feature representation. The similarity score is then masked and corrected using the binary mask matrix M_mask.

[0048]

[0049] The corrected similarity score is normalized to attention weights using a softmax function. And compute the attention context vector. Using the attention context vector Update the decoder state to generate the initial predicted power value; wherein, The hidden feature representation for the current time step;

[0050] When the attention mask matrix is ​​a mask matrix containing multidimensional correlation strength Attention weight between devices The specific calculation formula is as follows:

[0051]

[0052] In the formula, N is the number of the multiple devices; W is the trainable parameter matrix; and These represent the hidden features of the corresponding devices.

[0053] Furthermore, S4 specifically includes:

[0054] S401. Collect real-time total power time series data of shared electricity meters, construct it into a sliding time window data form as the model input sequence, and perform data standardization mapping preprocessing using the same normalization parameters as in the training phase.

[0055] S402. Input the preprocessed model input sequence into the trained non-intrusive load decomposition model, extract hidden feature representations through forward and backward feature extraction via a bidirectional long short-term memory network layer, and input the hidden feature representations into a mask attention layer. Combine the attention mask matrix to calculate the attention weights between devices to output the initial predicted power values ​​of each device at the corresponding time.

[0056] S403. Normalize and correct the initial predicted power value of each device using the actual total power of the shared meter at the corresponding time to obtain the real-time active power data of each device that meets the physical conservation constraints.

[0057] S404. Based on the real-time active power data of each device, compare it with the status judgment threshold set according to the rated power of the device, identify the operating status identifier of each device, and calculate the cumulative energy consumption data of each device.

[0058] S405. During the inference process of the non-intrusive load decomposition model, the Monte Carlo Dropout mechanism is enabled to perform multiple forward propagations, calculate the mean and variance of the prediction results, and output the confidence interval used to quantify the uncertainty of the power decomposition results.

[0059] Furthermore, in step S403, the specific calculation formula for normalizing and correcting the initial predicted power values ​​of each device is as follows:

[0060]

[0061] In the formula, This refers to the real-time active power data of the i-th device after correction. This represents the initial predicted power value of the i-th device output by the model. The initial predicted power value of the k-th device output by the model; N is the number of the multiple devices; Let be the actual total power of the shared meters at time t; after normalization correction, the real-time active power data of each device satisfies the following physical conservation constraints:

[0062]

[0063] In S404, the operating status identifier of the i-th device The decision logic formula is as follows:

[0064]

[0065] In the formula, This indicates that the equipment is in operation; This indicates that the equipment is in a stopped state; The state determination threshold; the state determination threshold The calculation formula is:

[0066]

[0067] In the formula, Let be the rated power of the i-th device; This is a preset proportionality coefficient, and 0 < <0.2;

[0068] In S405, the specific process of enabling the Monte Carlo Dropout mechanism to perform multiple forward propagations and output confidence intervals includes:

[0069] With the same time series data input, enable Dropout. The forward propagation yields multiple prediction samples from the i-th device at time t. ,in ;

[0070] Calculate the mean of the prediction results With variance :

[0071]

[0072]

[0073] Calculate the uncertainty confidence interval CI based on the mean and the variance: .

[0074] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0075] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0076] The beneficial effects of this invention are as follows:

[0077] This invention addresses the problem of missing equipment-level energy consumption data caused by centralized metering of multiple machines under a single meter in industrial manufacturing. It proposes a non-intrusive meter power decomposition method based on a bidirectional long short-term memory network. By extracting multi-dimensional features such as physical, temporal, and technological characteristics of multiple devices to construct structured equipment labels, and generating an attention mask matrix accordingly, this method leverages the Bi-LSTM network to extract long-range temporal features. This mask matrix then forcibly cuts off feature interference between unrelated devices at the attention layer, enhancing the coupling modeling capability of devices within the same process segment. Simultaneously, this invention introduces a physical conservation constraint loss function during model training to prevent power drift and combines it with a Monte Carlo Dropout mechanism to output confidence intervals. Finally, by combining the equipment-process mapping matrix, it achieves high-precision power decoupling at the equipment level and accurate carbon emission traceability at the product level under complex load conditions, providing reliable and high-quality data support for refined industrial energy management. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the overall process of the non-intrusive meter power decomposition method based on bidirectional long short-term memory network of the present invention.

[0079] Figure 2 This is a schematic diagram of the power decomposition algorithm processing flow in this invention.

[0080] Figure 3 This is a schematic diagram of the BiLSTM-Attention power decomposition model structure based on mask constraints in this invention.

[0081] Figure 4 This is a schematic diagram of the Monte Carlo Dropout uncertainty estimation in this invention.

[0082] Figure 5 This is a schematic diagram of the non-intrusive meter power decomposition device based on a bidirectional long short-term memory network according to the present invention.

[0083] Figure 6 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0084] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0085] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0086] In the vehicle manufacturing process (such as welding, painting, and final assembly workshops), there are numerous power equipment, conveying equipment, auxiliary equipment, and process equipment. Their energy consumption is complex, highly volatile, and significantly coupled with load, and these devices are typically used to process the same target product. In actual energy monitoring system deployments, the system usually adopts a heterogeneous metering configuration. Key high-power equipment is metered independently using a "one machine, one meter" approach, while a large number of small and medium-power equipment are centrally metered using a "multiple machines, one meter" approach.

[0087] This invention applies to scenarios involving multiple devices powered by the same shared electricity meter. In this "multiple devices per meter" structure, the shared meter can only provide the total power data for all devices, and cannot directly obtain the real-time power data of a single independent device within the shared meter. This invention aims to establish a correspondence between device energy consumption and the manufacturing process of a specific product by decomposing the total power of the shared meter.

[0088] like Figure 1 As shown, this invention provides a non-intrusive meter power decomposition method based on a bidirectional long short-term memory network, applicable to multiple devices powered by the same shared meter, where the multiple devices jointly process a target product. The method includes the following steps:

[0089] S1. Classify and label multiple devices to generate structured device tags for each device; wherein, the structured device tags include the physical characteristics, temporal operation characteristics, process characteristics and coupling relationships between devices.

[0090] In one embodiment of the present invention, step S1 specifically includes the following sub-steps:

[0091] S101. Extract the physical characteristics of the equipment and construct a physical feature vector.

[0092] Equipment is categorized based on energy conversion type, load electrical characteristics, and mechanical dynamic characteristics. In vehicle manufacturing companies, the physical characteristics of equipment are specifically classified as follows:

[0093] a. Energy conversion types: electric energy driven type, thermal energy conversion type, fluid power type.

[0094] b. Load electrical characteristics: resistive load, inductive load, capacitive load.

[0095] c. Mechanical dynamic characteristics: rotary motion type, linear motion type, static working type.

[0096] Based on this, multi-dimensional physical parameters are extracted according to the electrical characteristics and power level of the equipment, including: voltage level. Three-phase current Active power Total active power Power factor PF and frequency F. Based on these, the physical characteristic vector of the i-th device is defined. for:

[0097]

[0098] S102. Extract the timing characteristics of the device and construct the timing operation mode.

[0099] Based on energy consumption fluctuation patterns, equipment is classified into steady-state, transient, and periodic equipment. The time-series characteristic classification specifically includes:

[0100] a. Steady-state equipment: Power fluctuation rate ;

[0101] b. Transient devices: Current rise time ;

[0102] c. Periodic equipment: work cycle Duty cycle .

[0103] Furthermore, based on the equipment's on-site operation records (startup time) End time The device's timing operation mode is constructed using the device's operating state (state). The timing operation mode of the i-th device... Represented as:

[0104]

[0105] S103. Extract the process characteristics of the equipment and classify and label them to generate structured equipment labels for each piece of equipment.

[0106] The equipment is classified according to its core technological content, and the specific classification of technological characteristics includes:

[0107] (1) Tier 1 key equipment: such as welding robot clusters, chassis assembly platforms;

[0108] (2) Tier 2 auxiliary equipment: such as pneumatic clamps, tightening guns, etc.;

[0109] (3) Tier 3 infrastructure: such as workshop lighting, air conditioning system, etc.

[0110] Furthermore, the inter-device coupling relationships among the multiple devices are determined. These inter-device coupling relationships represent the strength of the technological or physical association between the devices, specifically including: whether there is a technological coupling relationship between the devices (e.g., belonging to the same process segment), whether there is strong temporal linkage, and the power cross-correlation degree between the devices (e.g., the power cross-correlation degree between the key devices satisfies...). ).

[0111] Finally, the physical feature vectors, timing operation modes, process characteristic classifications, and inter-device coupling relationships of each device are combined to generate structured device tags for each device.

[0112] S2. Collect the total power raw data of the shared electricity meter and extract multi-dimensional feature sequences from it. Construct an attention mask matrix representing the correlation strength between multiple devices based on the structured device labels of each device. The multi-dimensional feature sequences are used to input into the non-intrusive load decomposition model to extract hidden feature representations. The attention mask matrix is ​​used to perform attention weight calculation in collaboration with the hidden feature representations in the non-intrusive load decomposition model.

[0113] In one embodiment of the present invention, such as Figure 2 As shown, step S2 specifically includes the following sub-steps:

[0114] S201. Acquisition and caching of raw data on total power of shared electricity meters

[0115] High-frequency smart meters are installed at the output terminals of the shared meters to be decomposed, collecting raw waveform data and derived data such as total current, voltage, active power, and reactive power from the target shared meters. The collected data devices must have a high sampling rate to capture transient characteristics during device start-up and shutdown; and each collected data point must have a precise second-level timestamp for subsequent time-series alignment and event analysis. The collected raw and derived data are transmitted in real-time to edge computing nodes via a network and buffered and stored in the form of a time-series database for subsequent processing.

[0116] S202, Data Preprocessing and Multidimensional Feature Sequence Extraction

[0117] After preprocessing the buffered and stored data in the time series database, a time period of sequence data is extracted based on the detected event points, and multi-dimensional feature vectors are extracted from it. The extracted features are strongly correlated with the classification dimension of step S1, with the process characteristics features serving more as prior knowledge input. Finally, a multi-dimensional feature sequence is generated. As input to the model, each x represents a vector composed of all features extracted at time t. The multidimensional feature sequence X will be used as input to the non-intrusive load decomposition model as a unified data format, specifically input to the bidirectional long short-term memory network layer of the model to extract hidden feature representations containing complete contextual information and forward and backward time dependencies.

[0118] S203, Construction of Attention Mask Matrix

[0119] Based on the structured device labels and classification labels of S1, the associated and unassociated relationships between multiple devices are defined, and an attention mask matrix representing the strength of the association between multiple devices is constructed. This attention mask matrix is ​​used in the masked attention layer of the non-intrusive load decomposition model for collaborative computation with the hidden feature representation output by the bidirectional long short-term memory network layer. This is to mask the time-step feature information of unassociated devices and suppress feature interference and attention propagation between unrelated devices during attention mechanism computation. The specific construction of the attention mask matrix includes the following two corresponding forms:

[0120] (1) Constructing the binary mask matrix M_mask: Create an N×N matrix (where N is the number of devices) and map it to the time step to generate the binary mask matrix M_mask. The construction rules satisfy: when device i and device j have no process coupling relationship, specify M_mask[i,j]=-∞; when device i and device j have strong timing linkage, specify M_mask[i,j]=0.

[0121] (2) Construct a mask matrix containing multidimensional correlation strength. :

[0122] First, construct a process tag matrix based on the process segments in which the equipment participates. Its specific form is:

[0123]

[0124] Based on this, a three-dimensional tag vector is generated for each device. ,in The device physical feature vector constructed in step S1, This is the device's timing-based operation mode.

[0125] Next, define the mask matrix. as follows:

[0126]

[0127] in, , , These are preset coefficients; Indicates the similarity of physical features; Indicates temporal feature similarity; The similarity of process features is expressed by the following formula: If device i and device j belong to different process segments, then the mask matrix satisfies... .

[0128] S3. Construct a non-intrusive load decomposition model including a bidirectional long short-term memory network layer and a masked attention layer, and train the non-intrusive load decomposition model using a sample dataset constructed based on historical total power data and historical actual power data of each device.

[0129] In one embodiment of the present invention, step S3 specifically includes the following sub-steps:

[0130] S301, Overall Model Architecture Construction

[0131] Constructing a non-intrusive load decomposition model based on deep learning, such as Figure 3 As shown, the model employs an encoder-decoder framework and incorporates an attention mechanism in its overall structure. The model's input layer receives the multidimensional feature sequence extracted in step S2. , where t is the sequence length, and each x is a vector containing various features such as harmonics and power change rate.

[0132] Regarding the input data format, the input layer adopts a sliding time window data format:

[0133]

[0134] In the formula, L is the length of the sliding time window; The total active power of the shared meters at time t is the input data, which has been standardized before input.

[0135] In an optimized network configuration for time-frequency feature fusion, the input samples First, the data is fed into a multi-scale one-dimensional convolutional layer for rapid extraction of temporal features, resulting in a temporal feature representation. Simultaneously, a Fast Fourier Transform is performed on the input signal to obtain its frequency domain characteristics. Subsequently, the Softmax function is used for frequency domain weighted fusion to calculate the frequency domain weights. :

[0136]

[0137] Frequency domain feature representation is obtained by performing frequency domain weighting and aggregation. Finally, the time-domain and frequency-domain features are fused to obtain the fused feature sequence. As input to the subsequent feature extraction layer.

[0138] S302, Bidirectional Long Short-Term Memory Network Layer Hidden Feature Extraction

[0139] The input sequence is fed into a multi-layer bidirectional long short-term memory (Bi-LSTM) network for deep mining of temporal dependencies. During the forward and backward feature extraction processes, the forward LSTM unit processes the sequence sequentially from t=1 to t=T, capturing positive temporal dependency features. The backward LSTM unit processes the sequence in reverse order from t=T to t=1 to capture the inverse time-dependent features. .

[0140] Then, the forward hidden state at each time step is... and backward hidden state The features are concatenated to form a hidden feature representation containing complete contextual information. (i.e., encoded output) ):

[0141]

[0142] By utilizing the bidirectional time-series network structure to simultaneously capture the historical and future trends of equipment load changes, it is possible to accurately capture the long-term time-series dependencies of equipment start-up, shutdown, and operation, thereby enhancing the identification of power linkage characteristics of Tier 1 key equipment.

[0143] S303, Mask Attention Weight Calculation and Decoding Output

[0144] Represent the extracted hidden features The input mask attention layer uses the attention mask matrix constructed in step S2 to perform constraint calculations, dynamically focusing on the feature information that contributes most to the current decomposition at different times. Based on the form of the attention mask matrix constructed in the preceding steps, its internal core calculation mechanism includes the following two implementation methods:

[0145] (1) When using a binary mask matrix M_mask: the decoder generates a query state at time t. ,calculate Similarity score with all hidden feature representations of the encoder The time step scores of non-associated devices are masked and corrected using the binary mask matrix M_mask.

[0146]

[0147] When there is no process coupling between device i and device j, since M_mask[i,j]=-∞, the summation forces their similarity score to be reduced to negative infinity, and then normalized to attention weights through the softmax function. When the attention weight is reduced to almost zero, the interference of non-related features is eliminated; when there is strong temporal linkage between devices, since M_mask[i,j]=0, the original similarity score is retained.

[0148] Next, the attention context vector is calculated. The previous time step's output, the previous time step's decoder state, and the attention context vector are combined. The data is fed into the current LSTM unit of the decoder to update the decoder state. Ultimately, it will and The data is then stitched together again, and the initial predicted power values ​​for each device at the current moment are generated through a fully connected output layer.

[0149] (2) When using a mask matrix that includes multidimensional correlation strength Attention layer: after the hidden feature input mask, according to the mask matrix Calculate the attention weights between devices :

[0150]

[0151] In the formula, N is the number of devices under the shared electricity meter; W is the trainable parameter matrix; This represents the strength of the technological or physical association between device i and device j. The masking mechanism directly suppresses feature interference between unrelated devices during feature propagation, enhances the feature coupling modeling capability between devices in the same process segment, and outputs the initial predicted power value of each device at the corresponding time.

[0152] S304, Model Training Optimization

[0153] Model training is performed after the model network architecture design is completed. Historical total power data (total active power data) from the shared electricity meters is obtained. ) and the corresponding historical actual power data of each device (active power data of individual meters) The sample dataset consists of ( ). The training input sample data is in the format of a sliding window. The output label of supervised training is .

[0154] During supervised or weakly supervised model training, the joint loss function is optimized using the backpropagation algorithm. To update network parameters, the joint loss function is composed of the base regression loss. With respect to the conservation constraints of electrical physics Weighted construction:

[0155]

[0156] In the formula, These are the preset conservation constraint weight coefficients. Wherein, the basic regression loss... The formula used to constrain the mean square error between the predicted power and the historical actual power of each device is as follows:

[0157]

[0158] The conservation constraint loss To constrain the physical conservation of the sum of predicted power from all devices and the total power of the shared meter, thus preventing power drift, the specific formula is as follows:

[0159]

[0160] In the formula, T is the total number of time steps for the training samples; N is the number of devices; This represents the predicted power value of the i-th device at time t, as output by the model.

[0161] Furthermore, to meet the reliability requirements of industrial applications, the model incorporates a Monte Carlo Dropout estimation quantization mechanism during the model building and training phases. During model inference, multiple forward propagations are performed with Dropout enabled to quantify the uncertainty variance of the non-intrusive power decomposition results, providing a reliability confidence interval assessment for the final output device-level energy consumption data.

[0162] S4. Collect the real-time total power data of the shared electricity meter, preprocess it, and input it into the trained non-intrusive load decomposition model. Output the real-time initial predicted power value of each device and perform normalization correction to obtain the real-time active power data of each device. Compare the real-time active power data of each device with the preset state judgment threshold to obtain the operating status identifier of each device.

[0163] In one embodiment of the present invention, step S4 specifically includes the following sub-steps:

[0164] S401, Model Input Data Construction

[0165] During actual operation, real-time total power time-series data of the shared meters to be decomposed is continuously collected. This real-time data is then constructed into a sliding time window format and used as the input sequence for the model. :

[0166]

[0167] In the formula, L is the preset time window length; Let be the total active power actually collected by the shared meter at time t. To ensure that the data distribution in the inference phase is consistent with that in the training phase, the above sliding time window data is preprocessed using the same normalization parameters as in the training phase.

[0168] S402, Model Forward Inference and Initial Power Prediction

[0169] The time series data preprocessed in step S401 is input into the non-intrusive load decomposition model trained in step S3 for inference.

[0170] First, the data is fed into a bidirectional long short-term memory network layer for forward and backward feature extraction to obtain hidden feature representations. :

[0171]

[0172] In the formula: Indicates forward time dependency features; This indicates inverse time-dependent characteristics. This bidirectional structure can simultaneously capture both historical and future trends of device load changes during inference, thereby improving decomposition accuracy under superimposed load conditions.

[0173] Subsequently, the hidden feature representation is input into the mask attention layer, based on the attention mask matrix generated in step S2. Attention weights between computing devices :

[0174]

[0175] In the formula, N is the number of devices under the shared electricity meter; W is the parameter matrix that has been trained and fixed in the model; This represents the strength of the technological or physical association between device i and device j. Through feature decoupling via the masking mechanism described above, the model ultimately outputs the initial predicted power value of each device at time t, denoted as... .

[0176] S403, Normalization Correction and Physical Conservation Constraints

[0177] Since the sum of the predicted power of each device output independently by the model may deviate from the actual power in the total table, in order to ensure that the decomposition results strictly conform to the laws of electrical physics conservation and avoid power drift, the actual total power at time t is used. The initial predicted power values ​​of each device are normalized and corrected to obtain the corrected real-time active power data of each device. :

[0178]

[0179] After this correction step, the sum of the real-time active power of each independent device strictly satisfies the physical conservation constraints:

[0180]

[0181] S404. Equipment Operating Status Determination and Cumulative Energy Consumption Calculation

[0182] Based on the corrected real-time active power data Set the state determination threshold To identify the operating status of each device. Operating status identifier for the i-th device. The judgment logic is as follows:

[0183]

[0184] In the formula, This indicates that the equipment is in operation; This indicates that the equipment is in a stopped state. The state determination threshold is... Set according to the rated power ratio of the equipment:

[0185]

[0186] In the formula, Rated power of the equipment; coefficient The range of values ​​is 0 < <0.2. Simultaneously, based on the decomposed real-time active power, the cumulative energy consumption data of each device is calculated. :

[0187]

[0188] S405. Uncertainty Quantification Assessment Based on Monte Carlo Dropout

[0189] To provide reliability assessments for equipment-level energy consumption data in industrial settings, such as Figure 4 As shown, the Monte Carlo Dropout estimation mechanism is used in the model inference process to quantify the uncertainty of the non-intrusive power decomposition results.

[0190] Specifically, given the same input data, Dropout is enabled multiple times (total... (Number of forward propagations) to obtain multiple prediction samples from the i-th device at time t. (in ).

[0191] Calculate the mean of the prediction results respectively. With variance :

[0192]

[0193]

[0194] Based on this, output the confidence interval CI of the prediction result:

[0195]

[0196] Through the complete S4 process described above, the final output includes real-time active power data at the device level, device operating status indicators, cumulative energy consumption data of the device, and uncertainty confidence intervals, thereby achieving precise decoupling of device-level energy consumption under the "multiple machines, one meter" configuration in real manufacturing scenarios.

[0197] S5. Obtain the production process information of the target product, construct a device-process mapping matrix representing the correspondence between multiple devices and the manufacturing links of the target product, and calculate the cumulative energy consumption of the target product during the manufacturing period by combining the real-time active power data and operating status identifier of each device, and multiply the cumulative energy consumption by a preset carbon emission factor per unit of electricity to obtain the carbon emission of the target product.

[0198] In one embodiment of the present invention, step S5 specifically includes the following sub-steps:

[0199] S501. Obtain the process flow and construct the equipment-process mapping matrix.

[0200] Obtain the production process information for a specific target product and establish a product-process path set:

[0201]

[0202] In the formula, m represents the product number; K represents the number of process sections.

[0203] Based on the production process information and equipment participation, an equipment-process mapping matrix is ​​constructed. :

[0204]

[0205] S502, Time Alignment and Running Status Filtering

[0206] By combining equipment operation records, the time range for the manufacturing of the target product can be obtained:

[0207]

[0208] Based on the device operating status identifier output in step S4, the device operating status data is time-aligned and filtered to extract data intervals that meet the following time constraints for subsequent calculations:

[0209]

[0210] S503, Product-level Energy Consumption Calculation and Allocation

[0211] By combining the equipment-process mapping matrix and the real-time active power data of each device during the manufacturing time interval, the total energy consumption (i.e., cumulative energy consumption) of the target product m during the manufacturing period is calculated. :

[0212]

[0213] In the formula, N represents the number of multiple devices; The real-time active power data of the i-th device obtained by decomposition; For time step.

[0214] If multiple target products are produced simultaneously within a unit of time, the total energy consumption is allocated according to the production quantity to calculate the energy consumption per unit product. :

[0215]

[0216] In the formula, The production quantity of the target product.

[0217] S504, Product Carbon Emission Calculation and System Final Output

[0218] Based on the preset carbon emission factor per unit of electricity Calculate the carbon emissions of the target product :

[0219]

[0220] In the formula, The carbon emissions of the target product; It is the carbon emission factor per unit of electricity, and its unit is kgCO2 / kWh.

[0221] Through the complete processing flow of steps S1 to S5 above, the final output of the system includes: real-time power data at the device level, cumulative energy consumption data at the device level, energy consumption data at the product level, carbon emission data at the product level, and uncertainty confidence intervals used to quantify and decompose reliability.

[0222] To more clearly illustrate the application effect of the method of the present invention in actual industrial scenarios, a complete verification process of an embodiment is given below in combination with specific application scenarios and experimental data.

[0223] Non-intrusive power decomposition verification of electricity meters was carried out in a production unit of a vehicle manufacturing company. The production unit contains 6 independent devices: device D1 (model / number: 722-925), device D2 (model / number: 3-SQL), device D3 (model / number: 471-997), device D4 (model / number: 102-002), device D5 (model / number: CJ-G024), and device D6 (model / number: AP1).

[0224] The above six devices share the same electricity meter, and the corresponding total meter data is marked as D7.ALL. This data represents the total active power of the six devices sharing the meter. It satisfies the following power superposition relationship:

[0225]

[0226] In the formula, This indicates the total active power of the shared electricity meter; This represents the actual active power of the i-th device.

[0227] By extracting and analyzing the electricity meter data and the master meter data of these six independent devices, it was found that each device has significantly different time-series morphological characteristics. The specific characteristics are shown in the table below:

[0228] equipment Power form Feature Description D1 Periodic pulse type There is a clear processing rhythm D2 Step type Start-stop is obvious D3 steady state + short-term fluctuations Auxiliary processes D4 High-power short-time impact Main processing D5 Medium power stability Conveying D6 Low power standby + occasional peak Control Class

[0229] The significant differences in power fluctuation patterns among different devices provide a structural basis for the multidimensional feature extraction and model attention decoupling in this invention. By strictly aligning the timestamps, the errors in the collected data can be verified. Within a reasonable range, it satisfies:

[0230]

[0231] The non-invasive load decomposition model based on bidirectional long short-term memory network and attention mechanism provided by this invention was used to train and decompose the above data in real time. The decomposition accuracy evaluation indicators (mean absolute error MAE, root mean square error RMSE, and coefficient of determination R²) of each device are shown in the table below:

[0232] equipment MAE(kW) RMSE (kW) R² Equipment 1 0.42 0.58 0.972 Equipment 2 0.37 0.51 0.978 Equipment 3 0.48 0.63 0.964 Equipment 4 0.29 0.41 0.983 Equipment 5 0.53 0.70 0.956 Equipment 6 0.33 0.46 0.981

[0233] The verification results show that the average R² of the six devices reached 0.972, indicating extremely high decomposition accuracy.

[0234] Furthermore, to highlight the advantages of the method of the present invention, it was compared with existing traditional decomposition models, and the comparison results are as follows:

[0235] method R² Are there physical constraints? Is carbon emission traceable? Traditional HMM 0.83 no no CNN 0.91 no no Method of the present invention 0.972 yes yes

[0236] Furthermore, during the actual inference phase of the model, this invention introduces a Monte Carlo Dropout mechanism to quantify uncertainty. Taking a specific moment in the system's operation as an example, the device-level predicted power and variance output by the model are shown in the table below:

[0237] equipment Predicted power (kW) variance 722-925 21.4 0.36 3-SQL 12.1 0.21 471-997 24.7 0.42 102-002 10.6 0.19 CJ-G024 14.3 0.28 AP1 8.4 0.17

[0238] The data from the above embodiments fully demonstrate that the method proposed in this invention can not only accurately achieve real-time power decoupling in complex industrial scenarios of "multiple machines and one meter", but also has strict physical conservation constraints and outputs reliable variance evaluation indicators, providing high-quality data support for subsequent accurate calculation of product-level carbon emissions.

[0239] like Figure 5 As shown, the present invention also provides a non-intrusive meter power decomposition device based on a bidirectional long short-term memory network, applicable to multiple devices powered by the same shared meter, wherein the multiple devices jointly process a target product. The device includes:

[0240] The labeling module is used to hierarchically label multiple devices to generate structured device tags for each device; wherein, the structured device tags include the physical characteristics, temporal operation characteristics, process characteristics and coupling relationships between devices.

[0241] A construction module is used to collect the total power raw data of the shared electricity meter and extract multi-dimensional feature sequences from it. Based on the structured device labels of each device, an attention mask matrix representing the correlation strength between multiple devices is constructed. The multi-dimensional feature sequences are used to input into a non-intrusive load decomposition model to extract hidden feature representations. The attention mask matrix is ​​used to perform attention weight calculation in collaboration with the hidden feature representations in the non-intrusive load decomposition model.

[0242] The training module is used to construct a non-intrusive load decomposition model including a bidirectional long short-term memory network layer and a masked attention layer, and to train the non-intrusive load decomposition model using a sample dataset constructed based on historical total power data and historical actual power data of each device.

[0243] The output module is used to collect the real-time total power data of the shared electricity meter, preprocess it and input it into the trained non-intrusive load decomposition model, output the real-time initial predicted power value of each device and perform normalization correction to obtain the real-time active power data of each device, and compare the real-time active power data of each device with the preset state judgment threshold to obtain the operating status identifier of each device.

[0244] The calculation module is used to obtain the production process information of the target product, construct an equipment-process mapping matrix representing the correspondence between multiple devices and the manufacturing links of the target product, and calculate the cumulative energy consumption of the target product during the manufacturing period by combining the real-time active power data and operating status identifiers of each device. The cumulative energy consumption is then multiplied by a preset carbon emission factor per unit of electricity to obtain the carbon emission of the target product.

[0245] Each of the above modules is used to perform the corresponding steps in the non-intrusive meter power decomposition method based on bidirectional long short-term memory network. The specific implementation method is as described in the above method embodiment, and will not be repeated here.

[0246] like Figure 6 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all the data required for the process of the non-intrusive meter power decomposition method based on a bidirectional long short-term memory network. The network interface is used for communication with external terminals via a network connection. The computer program is executed by the processor to implement the non-intrusive meter power decomposition method based on a bidirectional long short-term memory network.

[0247] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0248] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described non-intrusive meter power decomposition methods based on a bidirectional long short-term memory network.

[0249] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0250] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0251] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A non-intrusive meter power decomposition method based on a bidirectional long short-term memory network, applied to multiple devices powered by the same shared meter, wherein the multiple devices jointly process a target product, characterized in that... The method includes the following steps: S1. Hierarchically label multiple devices to generate structured device tags for each device; wherein, the structured device tags include the physical characteristics, temporal operation characteristics, process characteristics and coupling relationships between devices for each device; S2. Collect the total power raw data of the shared electricity meter and extract multi-dimensional feature sequences from it. Construct an attention mask matrix representing the correlation strength between multiple devices based on the structured device labels of each device. The multi-dimensional feature sequences are used to input into the non-intrusive load decomposition model to extract hidden feature representations. The attention mask matrix is ​​used to perform attention weight calculation in collaboration with the hidden feature representations in the non-intrusive load decomposition model. S3. Construct a non-intrusive load decomposition model including a bidirectional long short-term memory network layer and a masked attention layer, and train the non-intrusive load decomposition model using a sample dataset constructed based on historical total power data and historical actual power data of each device. S4. Collect the real-time total power data of the shared electricity meter and input it into the trained non-intrusive load decomposition model after preprocessing. Output the real-time initial predicted power value of each device and perform normalization correction to obtain the real-time active power data of each device. Compare the real-time active power data of each device with the preset state judgment threshold to obtain the operating status identifier of each device. S5. Obtain the production process information of the target product, construct a device-process mapping matrix representing the correspondence between multiple devices and the manufacturing links of the target product, and calculate the cumulative energy consumption of the target product during the manufacturing period by combining the real-time active power data and operating status identifier of each device, and multiply the cumulative energy consumption by a preset carbon emission factor per unit of electricity to obtain the carbon emission of the target product.

2. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 1, characterized in that, S1 specifically includes: S101. Classify equipment according to energy conversion type, load electrical characteristics and mechanical dynamic characteristics, and extract multi-dimensional physical parameters to construct physical feature vectors for each equipment by combining the electrical characteristics and power levels of the multiple devices. S102. Based on the energy consumption fluctuation pattern, multiple devices are divided into steady-state, transient, and periodic devices, and based on the on-site operation records of the devices, the time-series operation mode of each device is constructed. S103. Classify and label multiple devices as key devices, auxiliary devices, or infrastructure according to the degree of process coreness to construct the process characteristic classification of each device, and determine the inter-device coupling relationship used to represent the strength of process or physical association between devices. Combine the physical feature vector, time sequence operation mode, process characteristic classification and inter-device coupling relationship of each device to generate a structured device label for each device.

3. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 2, characterized in that, In S101, the physical feature vector The expression is as follows: Where i represents the device number; These represent the three-phase voltage levels; These represent the three-phase currents respectively; PF represents total active power; F represents power factor; F represents frequency. In S102, the timing operation mode of the i-th device The expression is as follows: Where t represents the current timestamp; Indicates the device startup time; Indicates the end time of the device; when the device is in operation. When the equipment is in a stopped state .

4. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 1, characterized in that, S2 specifically includes: S201. Collect the raw waveform data and derived data of the total current, voltage, active power and reactive power of the shared meter, wherein each collected data point is timestamped. S202. After preprocessing the original waveform data and derived data, extract the time series data, and combine the structured device labels of each device to extract multi-dimensional feature vectors, generating a multi-dimensional feature sequence. Each x represents a vector composed of all features extracted at time t; wherein, the multidimensional feature sequence is used as input to the bidirectional long short-term memory network layer of the non-intrusive load decomposition model to extract forward and backward time-dependent hidden feature representations; S203. Based on the structured device labels of each device, construct an attention mask matrix; wherein, the attention mask matrix is ​​used to perform attention weight calculation in conjunction with the hidden feature representation in the mask attention layer of the non-intrusive load decomposition model.

5. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 4, characterized in that, In S203, the attention mask matrix is ​​a binary mask matrix M_mask, which is constructed as follows: Create an N×N matrix and map it to time steps, where N is the number of the multiple devices; when device i and device j have no process coupling relationship, specify M_mask[i,j]=-∞; when device i and device j have strong timing linkage, specify M_mask[i,j]=0; or, The attention mask matrix is ​​a mask matrix that contains multidimensional correlation strengths. Its specific construction method is as follows: Construct a process tag matrix based on the process segments in which the equipment participates. Where, when device i belongs to process segment k, it is specified Otherwise specify The physical feature vector, timing operation mode, and process label matrix of the corresponding equipment are combined to form a three-dimensional label vector for each equipment. ; The physical feature similarity between devices is calculated based on the three-dimensional label vectors. Temporal feature similarity and similarity of process features The mask matrix is ​​constructed according to the following formula. : in, , , The similarity of the process features is a preset coefficient. The calculation formula is And when equipment i and equipment j belong to different process segments, the following conditions are met. .

6. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 1, characterized in that, S3 specifically includes: S301. Construct a non-intrusive load decomposition model using an encoder-decoder framework, wherein the non-intrusive load decomposition model receives the sample feature sequence as input in the form of a sliding time window data. S302. The sample feature sequence is fed into the bidirectional long short-term memory network layer. During the forward and backward feature extraction processes, forward time-dependent features and backward time-dependent features are captured respectively, and the two are concatenated to form a hidden feature representation containing complete contextual information. S303. Input the hidden feature representation into the mask attention layer, use the attention mask matrix to constrain similarity calculation to mask the feature information of non-associated devices, calculate the attention weights between multiple devices, and combine the attention weights to generate the initial predicted power value of each device through the fully connected output layer. S304. Construct a joint loss function that is a weighted combination of the basic regression loss and the total power conservation constraint loss. Use the sample dataset to optimize the joint loss function through the backpropagation algorithm to update the network parameters. Configure a Monte Carlo Dropout estimation mechanism in the non-intrusive load decomposition model to quantify the uncertainty of the decomposition results during the inference phase.

7. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 6, characterized in that, In step S301, the input data also includes a step of time-frequency feature fusion optimization: The temporal feature representation of the sample feature sequence is extracted using a multi-scale one-dimensional convolutional layer. Simultaneously, the frequency domain features are obtained by performing a Fast Fourier Transform on the sample feature sequence. Frequency domain weights are calculated using the Softmax function. And obtain the frequency domain feature representation through frequency domain weighting. Representing the time-domain features With the frequency domain feature representation The fused feature sequence is obtained by splicing and merging the sequences. The frequency domain weights are input to the subsequent feature extraction layer. The calculation formula is: In S304, the joint loss function The specific mathematical expression is: In the formula, These are the preset conservation constraint weighting coefficients; Based on regression loss; The total power conservation constraint loss; the basic regression loss With the total power conservation constraint loss The calculation formulas are as follows: In the formula, T is the total number of time steps for the training samples; N is the number of the multiple devices; This refers to the historical actual power data of the i-th device; This represents the predicted power value of the i-th device output by the model. This refers to the historical total power data.

8. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 6, characterized in that, In step S303, when the attention mask matrix is ​​a binary mask matrix M_mask, the step of constraining the similarity calculation using the attention mask matrix specifically includes: calculating the similarity score between the decoder query state and the encoder hidden feature representation. The similarity score is then masked and corrected using the binary mask matrix M_mask. The corrected similarity score is normalized to attention weights using a softmax function. And compute the attention context vector. Using the attention context vector Update the decoder state to generate the initial predicted power value; wherein, The hidden feature representation for the current time step; When the attention mask matrix is ​​a mask matrix containing multidimensional correlation strength Attention weight between devices The specific calculation formula is as follows: In the formula, N is the number of the multiple devices; W is the trainable parameter matrix; and These represent the hidden features of the corresponding devices.

9. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 1, characterized in that, S4 specifically includes: S401. Collect real-time total power time series data of shared electricity meters, construct it into a sliding time window data form as the model input sequence, and perform data standardization mapping preprocessing using the same normalization parameters as in the training phase. S402. Input the preprocessed model input sequence into the trained non-intrusive load decomposition model, extract hidden feature representations through forward and backward feature extraction via a bidirectional long short-term memory network layer, and input the hidden feature representations into a mask attention layer. Combine the attention mask matrix to calculate the attention weights between devices to output the initial predicted power values ​​of each device at the corresponding time. S403. Normalize and correct the initial predicted power value of each device using the actual total power of the shared meter at the corresponding time to obtain the real-time active power data of each device that meets the physical conservation constraints. S404. Based on the real-time active power data of each device, compare it with the status judgment threshold set according to the rated power of the device, identify the operating status identifier of each device, and calculate the cumulative energy consumption data of each device. S405. During the inference process of the non-intrusive load decomposition model, the Monte Carlo Dropout mechanism is activated to perform multiple forward propagations, calculate the mean and variance of the prediction results, and output the confidence interval for quantifying the uncertainty of the power decomposition results.

10. The non-intrusive meter power decomposition method based on bidirectional long short-term memory network according to claim 9, characterized in that, In step S403, the specific calculation formula for normalizing and correcting the initial predicted power value of each device is as follows: In the formula, This refers to the real-time active power data of the i-th device after correction. This represents the initial predicted power value of the i-th device output by the model. The initial predicted power value of the k-th device output by the model; N is the number of the multiple devices; Let be the actual total power of the shared meters at time t; after normalization correction, the real-time active power data of each device satisfies the following physical conservation constraints: In S404, the operating status identifier of the i-th device The decision logic formula is as follows: In the formula, This indicates that the equipment is in operation; This indicates that the equipment is in a stopped state; The state determination threshold; The calculation formula is: In the formula, Let be the rated power of the i-th device; This is a preset proportionality coefficient, and 0 < <0.2; In S405, the specific process of enabling the Monte Carlo Dropout mechanism to perform multiple forward propagations and output confidence intervals includes: With the same time series data input, enable Dropout. The forward propagation yields multiple prediction samples from the i-th device at time t. ,in ; Calculate the mean of the prediction results With variance : Calculate the uncertainty confidence interval CI based on the mean and the variance: .

Images