A non-intrusive industrial load decomposition method and apparatus

Abstract

This invention discloses a non-invasive industrial load decomposition method and apparatus. It collects and preprocesses industrial aggregated active power sequences using smart meters, and constructs a periodic timestamp embedding module to complete feature mapping and fusion. An improved Transformer module is constructed, replacing traditional self-attention with a parallel multi-scale branch enhanced linear self-attention module to process the sequences. Local patterns are extracted through multi-scale lexical generation, and an amplitude-aware linear self-attention mechanism is used to avoid the attention dispersion problem of linear attention while maintaining linear computational complexity. Multi-scale time-domain features are formed through deep fusion. Frequency-domain features are extracted using a frequency-domain feature enhancement module using discrete cosine transform, and a weight matrix is ​​generated for feature recalibration. Finally, the output projection head maps the output to achieve power decomposition of the target device. This invention effectively solves the problems of computational bottlenecks, insufficient frequency-domain feature mining, and limited multi-scale feature fusion in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of power energy management technology, specifically to a non-intrusive industrial load decomposition method and apparatus. Background Technology

[0002] Currently, building a new power system based on new energy sources has become the core direction of energy industry transformation. Industrial loads, as a core component of electricity consumption, are characterized by their large proportion, high energy consumption, and complex characteristics, making their refined management crucial for grid supply and demand balance, peak shaving and valley filling, and energy conservation and emission reduction. Non-intrusive load monitoring (NILM) technology, with its advantages of not requiring modification of internal industrial wiring, low cost, and easy maintenance, collects power signals from the power input bus and uses algorithms to decompose the operating status and energy consumption details of each electrical device, providing support for enterprise energy efficiency analysis, fault early warning, and carbon emission accounting. It has already received widespread attention and initial application in the field of industrial energy management. Early combinatorial optimization algorithms (CO), Hidden Markov Models (HMM), and more recently developed deep learning methods such as CNNs and RNNs have shown some practicality in residential loads or simple industrial scenarios, but their adaptability to complex industrial environments still needs improvement.

[0003] However, existing non-intrusive load decomposition techniques still face significant bottlenecks in complex industrial scenarios. On the one hand, industrial load data exhibits long-term dependencies, and the computational complexity of traditional full-attention mechanisms increases quadratically when processing long sequences, leading to computational bottlenecks. Window-based processing methods, on the other hand, sever global connections, failing to capture long-term dependencies in equipment operation. On the other hand, existing methods primarily focus on time-domain feature analysis, lacking in-depth mining of frequency-domain discriminative information. This makes it difficult to distinguish between industrial equipment with similar power but different spectral characteristics. Furthermore, traditional time coding suffers from boundary discontinuities, failing to accurately capture industrial load patterns spanning shifts, weeks, and other cycles. In addition, industrial equipment experiences frequent start-ups and shutdowns, significant power fluctuations, and complex background noise. Existing models employ simplistic multi-scale feature fusion strategies with insufficient generalization capabilities, further limiting the improvement of load decomposition accuracy and failing to meet the practical needs of refined industrial energy management.

[0004] Therefore, there is an urgent need for a non-intrusive industrial load decomposition method to address the problems of existing technologies, such as computational bottlenecks, insufficient frequency feature mining, discontinuous time coding boundaries, and limited multi-scale feature fusion. Summary of the Invention

[0005] To address these issues, the present invention provides a non-intrusive industrial load decomposition method and apparatus, which solves the problems of computational bottlenecks, insufficient frequency feature mining, discontinuous time coding boundaries, and single multi-scale feature fusion in existing non-intrusive industrial load decomposition technologies when processing long-sequence data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a non-invasive industrial load decomposition method, comprising: Aggregated active power sequence data in industrial scenarios is obtained by acquiring data from smart meters and preprocessing it to obtain processed aggregated active power sequence data. Based on the processed aggregated active power sequence data, a periodic timestamp embedding module is constructed; the periodic timestamp embedding module is used for mapping and fusion processing to obtain a fused feature sequence. An improved Transformer module is constructed, replacing the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module; the fused feature sequence is input into the improved Transformer module for temporal feature extraction; the parallel multi-scale branch enhanced linear attention module consists of several parallel scale branches, each scale branch integrating a word generation module and an amplitude-aware linear attention module at a specified scale; Each branch extracts multi-scale local features using parallel deep convolutions with different kernel sizes, outputting multi-scale feature maps. The extracted multi-scale feature maps are then fed into the amplitude-aware linear self-attention module of the corresponding scale branch, achieving parallel processing with linear computational complexity to model global context dependencies and outputting single-scale context features for each branch. The single-scale context features output by each branch are concatenated and fused using linear mapping, and then combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, resulting in the temporal global context aggregation features output by the improved Transformer module. The time-domain global context aggregation feature is input into the frequency domain feature enhancement module. The frequency domain feature enhancement module performs a 1D discrete cosine transform on the time-domain global context aggregation feature channel by channel to extract the comprehensive frequency domain feature and perform a nonlinear mapping to obtain the frequency domain feature enhancement matrix. The matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation feature to achieve frequency domain enhancement of the time-domain global context aggregation feature and output the enhanced feature map. The enhanced feature map is mapped from the feature space to the power value space by the output projection head, and the active power decomposition result of the target device is output.

[0007] As a preferred method for non-intrusive industrial load decomposition, during the preprocessing of the aggregated active power sequence data, the aggregated active power sequence is divided into several subsequences by calculating the mean and variance of the total power sequence; the subsequences are then converted into sequences with a mean of 0 and a standard deviation of approximately 1 through standardization processing to obtain the processed aggregated active power sequence data.

[0008] As a preferred embodiment of the non-intrusive industrial load decomposition method, in the process of mapping and fusing through the periodic timestamp embedding module to obtain the fused feature sequence, the time scalar is decomposed into three scalar components: minutes, hours, and days of the week. Each scalar component is mapped into a two-dimensional time embedding vector in the geometric space of a unit circle through trigonometric transformation. The three two-dimensional time embedding vectors are concatenated and then fused through a learnable linear projection layer to obtain a global time-series embedding vector. The global time-series embedding vector is fused with the processed aggregated active power sequence data to obtain the fused feature sequence.

[0009] The formula for calculating the two-dimensional time embedding vector is: ; In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The expression for the global temporal embedding vector is: ; In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

[0010] As a preferred method for non-intrusive industrial load decomposition, in the improved Transformer module, feature mapping and attention calculation are performed within each scale branch of the parallel multi-scale branch enhanced linear attention module in the following manner: This branch uses depthwise convolution to extract scale feature maps from the fused feature sequence, and then performs linear projection on these scale feature maps to generate a query matrix specific to this scale branch. Key matrix Sum matrix The amplitude-aware linear self-attention module within this branch is configured to address the feature divergence problem of linear attention by introducing an amplitude-aware mechanism; the amplitude-aware linear self-attention module employs a nonlinear kernel function. query matrix AND key matrix Mapped to a nonnegative feature space to satisfy the positive definiteness requirement of the kernel method; the amplitude-aware linear self-attention module introduces a dynamic scaling factor that depends on the magnitude of the query feature. With dynamic offset terms The linear attention calculation process is reconstructed; For the A query vector, the amplitude-aware mechanism introduces a dynamic scaling factor. and dynamic offset terms Attention score is defined as: ; Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ; ; in, Given the sequence length; from this, we can deduce: ; The amplitude-aware linear self-attention mechanism adaptively adjusts the attention distribution based on changes in the magnitude of the query vector. After obtaining the attention score, apply the attention score to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ; In the actual calculation process, pre-calculation and As a context vector, and shared among all query vectors, the overall attention calculation only requires a single accumulation operation on the sequence dimension to obtain the single-scale context features corresponding to each scale branch; After all scale branches have completed parallel computation, the single-scale context features output by each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features are combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network. After nonlinear activation and mapping through a fully connected layer, the temporal global context aggregation features are output.

[0011] As a preferred embodiment of the non-invasive industrial load decomposition method, during the generation of the enhanced feature map through the frequency domain feature enhancement module, the temporal global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain a frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the frequency domain feature enhancement matrix; the matrix is ​​then used to perform a Hadamard product with the temporal global context aggregation feature to obtain the enhanced feature map. The specific processing flow of the frequency domain feature enhancement module in generating frequency feature enhanced feature maps is as follows: The feature decoupling process uses the time-domain global context aggregation features as the input feature map. Decoupled along the feature dimension A single independent feature vector: ; in, Indicates the first Feature vectors of each channel; The frequency domain feature extraction process based on one-dimensional discrete cosine transform (DCT) utilizes the fundamental functions of 1D DCT to convert the time-domain signal into a frequency-domain signal. The definition of the fundamental functions of 1D DCT is as follows: ; in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ; This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ; Then, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the input feature map: ; The frequency domain feature enhancement process utilizes multi-layer projection to perform a nonlinear mapping on the frequency domain representation: ; in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is selected to capture the nonlinear dependencies of the load frequency domain representation; the mapped frequency domain representation... As the frequency domain feature enhancement matrix, and the input feature map Perform the Hadamard product to enhance the frequency domain of the input feature map: ; in, This represents the obtained enhanced feature map.

[0012] The present invention also provides a non-invasive industrial load decomposition device, employing the above-described non-invasive industrial load decomposition method, comprising: The aggregated active power sequence data acquisition and preprocessing unit is used to acquire aggregated active power sequence data in industrial scenarios through smart meters and perform preprocessing to obtain processed aggregated active power sequence data. The fusion feature sequence acquisition unit is used to construct a periodic timestamp embedding module based on the processed aggregated active power sequence data; and to obtain a fusion feature sequence by performing mapping and fusion processing through the periodic timestamp embedding module. A parallel multi-scale global context aggregation unit is used to construct an improved Transformer module, which replaces the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module. The fused feature sequence is input into the improved Transformer module for temporal feature extraction. The parallel multi-scale branch enhanced linear attention module consists of several scale branches of parallel operations, and each scale branch integrates a word generation module and an amplitude-aware linear attention module at a specified scale. Each branch extracts multi-scale local features using parallel deep convolutions with different kernel sizes, outputting multi-scale feature maps. The extracted multi-scale feature maps are then fed into the amplitude-aware linear self-attention module of the corresponding scale branch, achieving parallel processing with linear computational complexity to model global context dependencies and outputting single-scale context features for each branch. The single-scale context features output by each branch are concatenated and fused using linear mapping, and then combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, resulting in the temporal global context aggregation features output by the improved Transformer module. The enhanced feature map generation unit is used to input the time-domain global context aggregation features into the frequency-domain feature enhancement module. The frequency-domain feature enhancement module performs a one-dimensional discrete cosine transform on the time-domain global context aggregation features channel by channel to extract the comprehensive frequency-domain features and performs a nonlinear mapping on them to obtain the frequency-domain feature enhancement matrix. The matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation features to achieve frequency-domain enhancement of the time-domain global context aggregation features and output the enhanced feature map. The active power decomposition result output unit maps the enhanced feature map from the feature space to the power value space through the output projection head, and outputs the active power decomposition result of the target device.

[0013] As a preferred embodiment of the non-invasive industrial load decomposition device, in the aggregated active power sequence data acquisition and preprocessing unit, during the preprocessing of the aggregated active power sequence data, the aggregated active power sequence is divided into several subsequences by calculating the mean and variance of the total power sequence; the subsequences are then converted into sequences with a mean of 0 and a standard deviation of approximately 1 through standardization processing to obtain the processed aggregated active power sequence data.

[0014] As a preferred embodiment of a non-invasive industrial load decomposition device, in the fusion feature sequence acquisition unit, during the mapping and fusion processing through the periodic timestamp embedding module to obtain the fusion feature sequence, the time scalar is decomposed into three scalar components: minutes, hours, and days of the week; each scalar component is mapped into a two-dimensional time embedding vector in the unit circle geometric space through trigonometric transformation; the three two-dimensional time embedding vectors are concatenated and then fused through a learnable linear projection layer to obtain a global time-series embedding vector; the global time-series embedding vector is fused with the processed aggregated active power sequence data to obtain the fusion feature sequence. The formula for calculating the two-dimensional time embedding vector is: ; In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The expression for the global temporal embedding vector is: ; In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

[0015] As a preferred solution for a non-invasive industrial load decomposition device, within each scale branch of the improved Transformer module in the parallel multi-scale global context aggregation unit, feature mapping and attention calculation are performed in the following manner: This branch uses deep convolution to extract scale feature maps from the fused feature sequence, and performs linear projection on these scale feature maps to generate query matrix, key matrix, and value matrix specific to this scale branch. The amplitude-aware linear self-attention module within this branch is configured to solve the feature divergence problem of linear attention by introducing an amplitude-aware mechanism. The amplitude-aware linear self-attention module uses a nonlinear kernel function to map the query matrix and key matrix to a non-negative feature space to meet the positive definiteness requirement of the kernel method. The amplitude-aware linear self-attention module introduces a dynamic scaling factor and a dynamic offset term that depend on the magnitude of the query feature to reconstruct the linear attention calculation process. For the A query vector, the amplitude-aware mechanism introduces a dynamic scaling factor. and dynamic offset terms Attention score is defined as: ; Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ; ; in, Given the sequence length; from this, we can deduce: ; The amplitude-aware linear self-attention mechanism adaptively adjusts the attention distribution based on changes in the magnitude of the query vector. After obtaining the attention score, apply the attention score to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ; In the actual calculation process, pre-calculation and As a context vector, and shared among all query vectors, the overall attention computation requires only one accumulation operation on the sequence dimension to obtain the single-scale context features corresponding to each scale branch. After all scale branches have completed parallel computation, the single-scale context features output by each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features are combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network. After nonlinear activation and mapping through a fully connected layer, the temporal global context aggregation features are output.

[0016] As a preferred embodiment of a non-invasive industrial load decomposition device, in the enhanced feature map generation unit, during the process of generating the enhanced feature map through the frequency domain feature enhancement module, the time-domain global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain a frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the frequency domain feature enhancement matrix; the matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation feature to obtain the enhanced feature map. The specific processing flow of the frequency domain feature enhancement module in generating frequency feature enhanced feature maps is as follows: The feature decoupling process uses the aggregated features of the temporal global context as the input feature map. Decoupled along the feature dimension A single independent feature vector: ; in, Indicates the first Feature vectors of each channel; The frequency domain feature extraction process based on one-dimensional discrete cosine transform (DCT) utilizes the fundamental functions of 1D DCT to convert the time-domain signal into a frequency-domain signal. The definition of the fundamental functions of 1D DCT is as follows: ; in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ; This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ; Then, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the input feature map: ; The frequency domain feature enhancement process utilizes multi-layer projection to perform a nonlinear mapping on the frequency domain representation: ; in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is selected to capture the nonlinear dependencies of the load frequency domain representation; the mapped frequency domain representation... As the frequency domain feature enhancement matrix, and the input feature map Perform the Hadamard product to enhance the frequency domain of the input feature map: ; in, This represents the obtained enhanced feature map.

[0017] The present invention has the following advantages: First, it balances computing power and long-term dependence: By using an amplitude-aware linear attention mechanism, the complexity of long sequence processing is reduced to linear, and the attention divergence problem of traditional linear attention mechanisms is avoided. This not only breaks through the computing power bottleneck, but also fully captures the long-term dependence of industrial loads. Second, deep fusion of time and frequency features: combining multi-scale time-domain feature extraction with dynamic weighted enhancement in the frequency domain, the frequency domain discrimination information specific to each device is strengthened, effectively distinguishing industrial devices with similar power but different spectra; Third, the physical timestamp features are continuous and accurate: periodic timestamp embedding is used to eliminate discontinuities at time boundaries, adapting to the cross-cycle operation mode of industrial loads; Fourth, improved accuracy and generalization: By suppressing noise interference in industrial scenarios through dynamic weight calibration, the accuracy of load decomposition and the model's generalization ability under complex working conditions are improved, adapting to the actual needs of refined industrial energy management. Attached Figure Description

[0018] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0019] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0020] Figure 1 This is a schematic flowchart of a non-invasive industrial load decomposition method provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the model structure of a non-invasive industrial load decomposition method provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the multi-scale amplitude sensing linear attention module processing flow in a non-invasive industrial load decomposition method provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the frequency domain feature enhancement module processing flow in a non-intrusive industrial load decomposition method provided in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram comparing the load decomposition results in one possible embodiment provided in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the architecture of a non-invasive industrial load decomposition device provided in Embodiment 2 of the present invention. Detailed Implementation

[0021] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a non-invasive industrial load decomposition method, comprising the following steps: S1. Obtain aggregated active power sequence data in industrial scenarios through smart meters and perform preprocessing to obtain processed aggregated active power sequence data; S2. Based on the processed aggregated active power sequence data, construct a periodic timestamp embedding module; perform mapping and fusion processing through the periodic timestamp embedding module to obtain a fused feature sequence; S3. Construct an improved Transformer module, replacing the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module; input the fused feature sequence into the improved Transformer module for temporal feature extraction; the parallel multi-scale branch enhanced linear attention module consists of several parallel scale branches, each scale branch integrating a word generation module of a specified scale and an amplitude-aware linear attention module; S4. Each branch uses parallel deep convolutions with different kernel sizes to extract multi-scale local features and output multi-scale feature maps. The extracted multi-scale feature maps are fed into the amplitude-aware linear self-attention module of the corresponding scale branch to model the global context dependency in parallel with linear computational complexity, and output the single-scale context features of each branch. The single-scale context features output by each branch are concatenated and linearly mapped and fused. Then, combined with the pre-normalization and residual connection mechanism, they are input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, to obtain the temporal global context aggregation features output by the improved Transformer module. S5. Input the time-domain global context aggregation feature into the frequency domain feature enhancement module. Perform a 1D discrete cosine transform on the time-domain global context aggregation feature channel by channel through the frequency domain feature enhancement module to extract the comprehensive frequency domain feature and perform a nonlinear mapping on it to obtain the frequency domain feature enhancement matrix. Then, use this matrix to perform a Hadamard product with the time-domain global context aggregation feature to realize the frequency domain enhancement of the time-domain global context aggregation feature and output the enhanced feature map. S6. The enhanced feature map is mapped from the feature space to the power value space through the output projection head, and the active power decomposition result of the target device is output.

[0023] In this embodiment, in step S1, aggregated active power sequence data in an industrial scenario is obtained through a smart meter and preprocessed to obtain processed aggregated active power sequence data.

[0024] Specifically, the HIPE public industrial electricity dataset was used for case analysis. This dataset has a sampling frequency of 1 / 5Hz and includes three months of master meter data and 10 types of industrial load data. A portion of the industrial load data from this dataset was selected as the aggregated active power sequence data for subsequent processing. Simultaneously, 80% of the data in this dataset was selected as the training set for load decomposition, and 20% as the test set for the load decomposition algorithm. During the testing phase, the factory master meter data was input into the trained model, and the model output results were compared and analyzed with the actual active power data of the equipment.

[0025] In this embodiment, the preprocessing process for the aggregated active power sequence data is as follows: First, calculate the mean and variance of the power sequence data in the training set table. and The training set data of the total power sequence and the equipment operating power sequence are divided into overlapping subsequences. and Window overlap rate Each subsequence contains 512 sampling points. The total power sequence is then standardized using the following formula, transforming it into a sequence with a mean of 0 and a standard deviation of approximately 1. This eliminates the impact of power magnitude differences between different devices on subsequent feature extraction, providing a high-quality input data foundation for subsequent modules.

[0026] The conversion formula is: ; In the formula, The mean of the power of the training set; To train the variance of the power of the settable; This is the standardized total power subsequence.

[0027] In this embodiment, in step S2, a periodic timestamp embedding module is constructed based on the processed aggregated active power sequence data; the periodic timestamp embedding module is used for mapping and fusion processing to obtain a fused feature sequence.

[0028] Specifically, firstly, based on the processed aggregated active power sequence data, a periodic timestamp embedding module is constructed; then, a... The timestamp of the current sampling point is decomposed into three scalar components: minutes (MOH), hours (HOD), and days of the week (DOW), denoted as set. Their corresponding physical periods are respectively .

[0029] For any periodic Time components We can map it to a two-dimensional time embedding vector plane using trigonometric transformations, as shown in the following formula: ; In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The three periodic vectors are then concatenated and fused through a learnable linear projection layer to obtain the final global temporal embedding vector. : ; In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

[0030] Finally, the global time-series embedding vector is concatenated element by element with the processed aggregated active power sequence data to obtain the fused feature sequence. This ensures that the Euclidean distance between the end of a complete time cycle (e.g., 23:59) and the beginning of a new cycle (e.g., 00:00) in the embedding space is minimized, guaranteeing the continuity of time and enabling the model to accurately capture load operation modes that span physical time cycles.

[0031] In this embodiment, in step S3, parallel deep convolution is performed through a multi-scale lexical generation module to extract local multi-granular features of the fused feature sequence.

[0032] Specifically, such as Figure 3 As shown, the multi-scale lexical generation module contains three sets of parallel deep convolution branches, which use 3×1, 5×1 and 7×1 one-dimensional convolution kernels respectively to perform convolution operations on the fused feature sequence and extract local multi-granularity features of the fused feature sequence.

[0033] In this embodiment, in step S4, within each scale branch of the parallel multi-scale branch enhanced linear attention module in the improved Transformer module, feature mapping and attention calculation are specifically performed as follows: This branch uses depthwise convolution to extract scale feature maps from the fused feature sequence, and then performs linear projection on these scale feature maps to generate a query matrix specific to this scale branch. Key matrix Sum matrix The amplitude-aware linear self-attention module within this branch is configured to address the feature divergence problem of linear attention by introducing an amplitude-aware mechanism; the amplitude-aware linear self-attention module employs a nonlinear kernel function. query matrix AND key matrix Mapped to a nonnegative feature space to satisfy the positive definiteness requirement of the kernel method; the amplitude-aware linear self-attention module introduces a dynamic scaling factor that depends on the magnitude of the query feature. With dynamic offset terms The linear attention calculation process is reconstructed; For the A query vector, the amplitude-aware mechanism introduces a dynamic scaling factor. and dynamic offset terms Attention score is defined as: ; Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ; ; in, Given the sequence length; from this, we can deduce: ; Through the above design, the amplitude-aware linear self-attention mechanism can adaptively adjust the attention distribution according to the changes in the magnitude of the query vector. When the magnitude of the query vector is large, the attention weights are more concentrated, exhibiting a sharp distribution similar to the softmax attention mechanism; when the magnitude of the query vector is small, the attention distribution tends to be flatter, thus significantly alleviating the attention dispersion problem existing in the traditional linear attention mechanism.

[0034] After obtaining the attention score, apply it to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ; In the actual calculation process, pre-calculation and As a context vector, shared across all query vectors, the overall attention computation requires only a single accumulation operation along the sequence dimension to obtain the single-scale context features corresponding to each scale branch, ensuring linear time complexity. This enables the amplitude-aware linear attention mechanism to efficiently simulate the amplitude-aware characteristics of softmax attention, thereby accurately capturing the global context in industrial load power signals. This improvement allows the model to efficiently process long aggregated power sequences in computationally limited equipment environments, while maintaining good modeling capabilities for local power change patterns and long-range dependencies. After parallel computation is completed in all scale branches, the single-scale context features output from each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features, combined with pre-normalization and residual connection mechanisms, are input into the feedforward neural network, mapped through nonlinear activation and fully connected layers, to output the temporal global context aggregation features.

[0035] In this embodiment, in step S5, during the process of generating the enhanced feature map through the frequency domain feature enhancement module, the temporal global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain a frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the frequency domain feature enhancement matrix; the matrix is ​​then used to perform a Hadamard product with the temporal global context aggregation feature to obtain the enhanced feature map. Specifically, such as Figure 4 As shown, firstly, the feature decoupling process uses the temporal global context aggregation features as the input feature map. Decoupled along the feature dimension A single independent feature vector: ; in, Indicates the first The feature vectors of each channel.

[0036] Then, a one-dimensional discrete cosine transform (DCT) is performed on each eigenvector. The core idea is to convert the time-domain signal into a frequency-domain signal using the fundamental functions of the one-dimensional DCT. The definition of the fundamental functions of the one-dimensional DCT is as follows: ; in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ; This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ; in, This indicates that the frequency domain representation of the time-domain sequence is calculated using the Discrete Cosine Transform (DCT). This embodiment utilizes the even-symmetric extension property of the signal through the DCT, ensuring the continuity of the extended signal at the boundaries and effectively eliminating high-frequency artifact noise caused by the Gibbs phenomenon in traditional Fast Fourier Transform (FFT) for time-domain signal frequency feature extraction. Subsequently, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the original input feature map. ; In order to adaptively mine key frequency features, the frequency domain feature enhancement process utilizes multi-layer projection to perform nonlinear mapping on the frequency domain representation: ; in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is chosen here to capture the nonlinear dependencies of the load frequency domain representation. Finally, the mapped frequency domain representation... Compared with the original input feature map Perform the Hadamard product to enhance the frequency domain of the original features: ; in, This represents the original input feature map after frequency domain feature enhancement.

[0037] The above steps are equivalent to a learnable soft filter that can dynamically enhance characteristic channels containing significant operating frequencies of the device, while suppressing channels containing background noise or invalid frequencies, thereby significantly improving the accuracy of load decomposition.

[0038] In this embodiment, in step S6, the enhanced feature map is mapped from the feature space to the power value space by the output projection head, and the active power decomposition result of the target device is output.

[0039] Specifically, the output projection head consists of two linear transformation layers and one activation function: First, the high-dimensional enhanced feature map is mapped to a feature dimension matching the number of target devices through a first-layer linear transformation. After ReLU activation, effective features are retained. Then, a second-layer linear transformation maps the vectors in the feature space to the power value space, outputting the active power sequence of the target devices with the same sampling frequency as the original, thus obtaining the power decomposition prediction result for the target devices. This process achieves a precise conversion from abstract time-frequency characteristics to specific equipment power values. The final output decomposition results can be directly used for energy consumption monitoring, fault early warning, and energy efficiency analysis of industrial equipment, meeting the practical needs of refined industrial energy management.

[0040] In this embodiment, the device power prediction result and the actual operating power of the device are used. Calculate the MSE loss value, update the model parameters through gradient backpropagation, until the model converges.

[0041] To evaluate the performance of this invention, mean absolute error (MAE), root mean square error (RMSE), signal aggregation error (SAE), and normalized decomposition error (NDE) were selected as performance indicators for the proposed non-intrusive load decomposition algorithm. The specific formulas are as follows: ; ; ; ; In the formula, This represents the actual power of the equipment. This is the predicted power value for the equipment; This represents the total number of sample points in the sequence.

[0042] In one possible embodiment, a simulation verification example is provided as follows: This example uses the HIPE public industrial power dataset for case analysis. This dataset has a sampling frequency of 1 / 5Hz and includes three months of summary data and data on 10 types of industrial loads. This example selects five major loads with relatively high power: PCB laminator, high-temperature sintering furnace, reflow oven, vacuum pump, and PCB cleaning machine. The PCB laminator's active power is characterized by a stable operating power of approximately 25kW, with a power fluctuation of 4kW during operation. The high-temperature sintering furnace's active power is characterized by an initial power of approximately 7kW, with peak and ramp-up effects during operation. The reflow oven has two different operating modes: in the first mode, there is a power peak of approximately 8.5kW upon startup, and a stable operating power of approximately 7.5kW; in the second mode, there is rapid power fluctuation with an amplitude of approximately 4kW, occurring approximately once per minute. The vacuum pump's active power is characterized by a 1kW peak upon startup, and a stable operating power of approximately 0.9kW. The PCB cleaning machine's active power is characterized by multiple operating states: approximately 1kW in standby mode and approximately 11kW under full load.

[0043] Table 1 Comparison of Evaluation Indicators for Load Decomposition Results

[0044] Table 1 shows a performance comparison between the present invention and commonly used comparison algorithms such as Seq2Point, Seq2Seq, and UNetNiLM. Figure 5 As shown, the load decomposition visualization results of this invention are presented compared with two other algorithms, Seq2Point and UNetNiLM.

[0045] Overall, this invention demonstrates superior power decomposition accuracy across all device categories, with a significant advantage when handling diverse load characteristics. Specifically, it excels in handling highly complex and volatile loads such as laminators. Comparative algorithms exhibit high Normalized Decomposition Error (NDE) on laminators (0.423 and 0.469 respectively) and Root Mean Square Error (RMSE) exceeding 670, indicating that traditional models struggle to accurately capture the details of intense power oscillations. In contrast, this invention significantly reduces the NDE to 0.059 and the RMSE to 250.452 for laminators, demonstrating that the introduced multi-scale feature extraction and frequency enhancement modules effectively reconstruct complex waveform profiles. Furthermore, this invention demonstrates excellent performance in Signal Convergence Error (SAE) on devices such as reflow ovens and cleaning machines. For example, the SAEs for reflow ovens and cleaning machines are only 0.033 and 0.066 respectively, far lower than comparative algorithms (such as UNetNiLM's 0.252 on cleaning machines). This indicates that the algorithm is not only accurate in fitting instantaneous power, but also has extremely high consistency in estimating total energy consumption over long periods.

[0046] It is worth noting that although the proposed method and S2P have their respective advantages and disadvantages in terms of mean absolute error (MAE) for sintering furnaces and vacuum pumps, this invention consistently maintains the overall best performance in terms of RMSE (more sensitive to outliers) and SAE (lowest SAE of 0.057 for vacuum pumps). Furthermore, the proposed model demonstrates robustness in industrial high-noise environments, effectively suppressing extreme errors and ensuring the reliability of energy efficiency analysis. In addition to the aforementioned accuracy advantages, this invention also exhibits significant superiority in model lightweighting. The S2P algorithm has as many as 30.17M parameters, while the proposed model has only 1.10M parameters. This means that this invention achieves a high degree of model volume compression while maintaining high accuracy. This extremely low space occupation not only significantly reduces storage costs but also significantly reduces memory overhead, making it extremely easy to deploy in resource-constrained industrial edge computing gateways or embedded terminals.

[0047] The analysis results show that the present invention has achieved the desired effect. With its low model size and superior load decomposition performance, it provides important support for optimizing energy allocation in industry, reducing operating costs, enhancing grid stability, promoting the integration of renewable energy, supporting refined demand-side management, risk assessment, and improving user satisfaction by accurately identifying equipment operating status and predicting changes in electricity demand at resource-constrained edge. This, in turn, promotes the development of smart grid edge sensing capabilities and the efficient operation of the electricity market.

[0048] The application scenarios of this invention are as follows: In the context of refined energy efficiency management in high-energy-consuming industrial enterprises, this invention can decompose the total incoming power to each production equipment in real time, accurately calculate the energy consumption of a single piece of equipment, help enterprises identify inefficient operating equipment, optimize production scheduling, and achieve energy saving and cost reduction.

[0049] In the scenario of dynamic load control in industrial parks, this invention can quickly decompose the equipment composition of the total load of the park, provide accurate equipment-level load data for the demand-side response of the power grid, support the park to dynamically adjust equipment operation strategies during peak electricity consumption periods, and alleviate the pressure on the power grid.

[0050] In industrial equipment fault early warning scenarios, this invention can continuously decompose equipment power, capture abnormal fluctuation characteristics of equipment, and combine equipment operating thresholds to achieve early fault warning and reduce the risk of unplanned downtime.

[0051] In industrial carbon emission accounting scenarios, this invention can accurately decompose the energy consumption of equipment in each production process, and combine it with the carbon emission coefficient per unit energy consumption to realize real-time calculation of equipment-level carbon emissions, providing data support for enterprises to track their carbon footprint and achieve carbon reduction targets.

[0052] In the scenario of industrial load modeling in new power systems, this invention can provide fine-grained equipment-level load time-series data, helping the power grid to build a more accurate industrial load prediction model and improve the power grid's adaptability to the consumption of new energy sources.

[0053] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0054] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0055] Example 2

[0056] See Figure 6 Embodiment 2 of the present invention also provides a non-invasive industrial load decomposition device, comprising: The aggregated active power sequence data acquisition and preprocessing unit 001 is used to acquire aggregated active power sequence data in industrial scenarios through smart meters and perform preprocessing to obtain processed aggregated active power sequence data. The fusion feature sequence acquisition unit 002 is used to construct a periodic timestamp embedding module based on the processed aggregated active power sequence data; and to obtain a fusion feature sequence by performing mapping and fusion processing through the periodic timestamp embedding module. A parallel multi-scale global context aggregation unit 003 is used to construct an improved Transformer module, replacing the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module. The fused feature sequence is input into the improved Transformer module for temporal feature extraction. The parallel multi-scale branch enhanced linear attention module consists of several parallel scale branches, each scale branch integrating a word generation module and an amplitude-aware linear attention module at a specified scale. Each branch uses parallel depthwise convolutions with different kernel sizes to extract multi-scale local features and output multi-scale feature maps. The extracted multi-scale feature maps are respectively input into the amplitude-aware linear self-attention modules of the corresponding scale branches to achieve modeling of global context dependencies with linear computational complexity in parallel processing, outputting single-scale context features of each branch. The single-scale context features output by each branch are concatenated and linearly mapped and fused, and then combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, obtaining the temporal global context aggregation features output by the improved Transformer module. The enhanced feature map generation unit 004 is used to input the time-domain global context aggregation feature into the frequency-domain feature enhancement module. The frequency-domain feature enhancement module performs a one-dimensional discrete cosine transform on the time-domain global context aggregation feature channel by channel to extract the comprehensive frequency-domain feature and perform a nonlinear mapping to obtain the frequency-domain feature enhancement matrix. The matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation feature to achieve frequency-domain enhancement of the time-domain global context aggregation feature and output the enhanced feature map. The active power decomposition result output unit 005 maps the enhanced feature map from the feature space to the power value space through the output projection head, and outputs the active power decomposition result of the target device.

[0057] In this embodiment, in the aggregated active power sequence data acquisition and preprocessing unit 001, during the preprocessing of the aggregated active power sequence data, the aggregated active power sequence is divided into several subsequences by calculating the mean and variance of the total power sequence; the subsequences are converted into sequences with a mean of 0 and a standard deviation of approximately 1 through standardization processing to obtain the processed aggregated active power sequence data.

[0058] In this embodiment, in the fusion feature sequence acquisition unit 002, during the mapping and fusion processing by the periodic timestamp embedding module to obtain the fusion feature sequence, the time scalar is decomposed into three scalar components: minutes, hours, and days of the week; each scalar component is mapped to a two-dimensional time embedding vector in the unit circle geometric space through trigonometric transformation; the three two-dimensional time embedding vectors are concatenated and then fused through a learnable linear projection layer to obtain a global time-series embedding vector; the global time-series embedding vector is fused with the processed aggregated active power sequence data to obtain the fusion feature sequence.

[0059] The formula for calculating the two-dimensional time embedding vector is: ; In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The expression for the global temporal embedding vector is: ; In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

[0060] In this embodiment, within each scale branch of the improved Transformer module in the parallel multi-scale global context aggregation unit 003, feature mapping and attention calculation are performed in the following manner:

[0061] This branch uses depthwise convolution to extract scale feature maps from the fused feature sequence and performs linear projection on these scale feature maps to generate query, key, and value matrices specific to this branch. The amplitude-aware linear self-attention module within this branch is configured to address the feature divergence problem of linear attention by introducing an amplitude-aware mechanism. This module uses a non-linear kernel function to map the query and key matrices to a non-negative feature space to satisfy the positive definiteness requirement of the kernel method. Furthermore, the amplitude-aware linear self-attention module introduces a dynamic scaling factor and a dynamic offset term that depend on the query feature amplitude to reconstruct the linear attention calculation process, thereby recovering the query vector amplitude information lost during the normalization calculation process. The attention score is defined as follows: ; Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ; ; in, Given the sequence length; from this, we can deduce: ; Through the above design, the amplitude-aware linear self-attention mechanism can adaptively adjust the attention distribution according to the changes in the magnitude of the query vector. When the magnitude of the query vector is large, the attention weights are more concentrated, exhibiting a sharp distribution similar to the softmax attention mechanism; when the magnitude of the query vector is small, the attention distribution tends to be flatter, thus significantly alleviating the attention dispersion problem existing in the traditional linear attention mechanism.

[0062] After obtaining the attention score, apply it to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ; In the actual calculation process, pre-calculation and As a context vector, shared across all query vectors, the overall attention computation requires only a single accumulation operation along the sequence dimension to obtain the single-scale context features corresponding to each scale branch, ensuring linear time complexity. This enables the amplitude-aware linear attention mechanism to efficiently simulate the amplitude-aware characteristics of softmax attention, thereby accurately capturing the global context in industrial load power signals. This improvement allows the model to efficiently process long aggregated power sequences in computationally limited equipment environments, while maintaining good modeling capabilities for local power change patterns and long-range dependencies. After parallel computation is completed in all scale branches, the single-scale context features output from each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features, combined with pre-normalization and residual connection mechanisms, are input into the feedforward neural network, mapped through nonlinear activation and fully connected layers, to output the temporal global context aggregation features.

[0063] In this embodiment, in the enhanced feature map generation unit 004, during the process of generating the enhanced feature map through the frequency domain feature enhancement module, the global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain a frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the spectral weight matrix; the original input feature map is dynamically weighted element-wise using the spectral weight matrix to obtain the enhanced feature map; The specific processing flow of the frequency domain feature enhancement module in generating the frequency domain feature enhanced feature map is as follows: The feature decoupling process uses the time-domain global context aggregation features as the input feature map. Decoupled along the feature dimension A single independent feature vector: ; in, Indicates the first Feature vectors of each channel; The core of the frequency domain feature extraction process based on one-dimensional discrete cosine transform (DCT) is to convert the time-domain signal into a frequency-domain signal using the fundamental functions of 1D DCT. The definition of the fundamental functions of 1D DCT is as follows: ; in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ; This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ; Then, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the input feature map: ; In order to adaptively mine key frequency features, the frequency domain feature enhancement process utilizes multi-layer projection to perform nonlinear mapping on the frequency domain representation: ; in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is chosen here to capture the nonlinear dependencies of the load frequency domain representation. Finally, the mapped frequency domain representation... As the frequency domain feature enhancement matrix, and the input feature map Perform the Hadamard product to enhance the frequency domain of the input feature map: ; in, This represents the original input feature map after frequency domain feature enhancement.

[0064] It should be noted that the information interaction and execution process between the modules of the above system are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0065] Example 3

[0066] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a non-intrusive industrial load decomposition method. The program code includes instructions for executing a non-intrusive industrial load decomposition method according to Embodiment 1 or any possible implementation thereof.

[0067] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0068] Example 4

[0069] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor; The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute a non-intrusive industrial load decomposition method according to Embodiment 1 or any possible implementation thereof by calling the program instructions.

[0070] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0071] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0072] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0073] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A non-invasive industrial load decomposition method, characterized in that, include: Aggregated active power sequence data in industrial scenarios is obtained by acquiring data from smart meters and preprocessing it to obtain processed aggregated active power sequence data. Based on the processed aggregated active power sequence data, a periodic timestamp embedding module is constructed. The periodic timestamp embedding module is used for mapping and fusion processing to obtain a fused feature sequence; An improved Transformer module is constructed, replacing the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module; the fused feature sequence is input into the improved Transformer module for temporal feature extraction; the parallel multi-scale branch enhanced linear attention module consists of several parallel scale branches, each scale branch integrating a word generation module and an amplitude-aware linear attention module at a specified scale; Each branch uses parallel deep convolutions with different kernel sizes to extract multi-scale local features and output multi-scale feature maps. The extracted multi-scale feature maps are then fed into the amplitude-aware linear self-attention module of the corresponding scale branch. Parallel processing with linear computational complexity is used to model the global context dependency and output the single-scale context features of each branch. The single-scale context features output from each branch are concatenated and linearly mapped and fused. Then, combined with pre-normalization and residual connection mechanisms, they are input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, resulting in the temporal global context aggregation features output by the improved Transformer module. The time-domain global context aggregation feature is input into the frequency domain feature enhancement module. The frequency domain feature enhancement module performs a 1D discrete cosine transform on the time-domain global context aggregation feature channel by channel to extract the comprehensive frequency domain feature and perform a nonlinear mapping to obtain the frequency domain feature enhancement matrix. The matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation feature to achieve frequency domain enhancement of the time-domain global context aggregation feature and output the enhanced feature map. The enhanced feature map is mapped from the feature space to the power value space by the output projection head, and the active power decomposition result of the target device is output.

2. The non-invasive industrial load decomposition method according to claim 1, characterized in that, In the preprocessing of the aggregated active power sequence data, the aggregated active power sequence is divided into several subsequences by calculating the mean and variance of the total power sequence; the subsequences are then converted into sequences with a mean of 0 and a standard deviation of approximately 1 through standardization processing to obtain the processed aggregated active power sequence data.

3. The non-invasive industrial load decomposition method according to claim 2, characterized in that, In the process of mapping and fusing through the periodic timestamp embedding module to obtain the fused feature sequence, the time scalar is decomposed into three scalar components: minutes, hours, and days of the week. Each scalar component is mapped into a two-dimensional time embedding vector in the geometric space of a unit circle through trigonometric transformation. The three two-dimensional time embedding vectors are concatenated and then fused through a learnable linear projection layer to obtain a global time-series embedding vector. The global time-series embedding vector is fused with the processed aggregated active power sequence data to obtain the fused feature sequence. The formula for calculating the two-dimensional time embedding vector is: ； In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The expression for the global temporal embedding vector is: ； In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

4. The non-invasive industrial load decomposition method according to claim 3, characterized in that, In the improved Transformer module, within each scale branch of the parallel multi-scale branch enhanced linear attention module, feature mapping and attention calculation are performed in the following manner: This branch uses depthwise convolution to extract scale feature maps from the fused feature sequence, and then performs linear projection on these scale feature maps to generate a query matrix specific to this scale branch. Key matrix Sum matrix The amplitude-aware linear self-attention module within this branch is configured to address the feature divergence problem of linear attention by introducing an amplitude-aware mechanism. The amplitude-sensing linear self-attention module employs a nonlinear kernel function. query matrix AND key matrix Mapped to a nonnegative feature space; the amplitude-aware linear self-attention module introduces a dynamic scaling factor that depends on the magnitude of the query feature. With dynamic offset terms The linear attention calculation process is reconstructed; For the A query vector, the amplitude-aware mechanism introduces a dynamic scaling factor. and dynamic offset terms ; Attention score is defined as: ； Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ； ； in, Given the sequence length; from this, we can deduce: ； The amplitude-aware linear self-attention mechanism adaptively adjusts the attention distribution based on changes in the magnitude of the query vector. After obtaining the attention score, apply the attention score to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ； In the actual calculation process, pre-calculation and As a context vector, and shared among all query vectors, the overall attention calculation only requires a single accumulation operation on the sequence dimension to obtain the single-scale context features corresponding to each scale branch; After all scale branches have completed parallel computation, the single-scale context features output by each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features are combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network. After nonlinear activation and mapping through a fully connected layer, the temporal global context aggregation features are output.

5. A non-invasive industrial load decomposition method according to claim 4, characterized in that, In the process of generating the enhanced feature map through the frequency domain feature enhancement module, the time domain global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain the frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the frequency domain feature enhancement matrix. The enhanced feature map is obtained by performing a Hadamard product between the matrix and the aforementioned temporal global context aggregated features. The specific processing flow of the frequency domain feature enhancement module in generating frequency feature enhanced feature maps is as follows: The feature decoupling process uses the time-domain global context aggregation features as the input feature map. Decoupled along the feature dimension A single independent feature vector: ； in, Indicates the first Feature vectors of each channel; The frequency domain feature extraction process based on one-dimensional discrete cosine transform (DCT) utilizes the fundamental functions of 1D DCT to convert the time-domain signal into a frequency-domain signal. The definition of the fundamental functions of 1D DCT is as follows: ； in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ； This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ； Then, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the input feature map: ； The frequency domain feature enhancement process utilizes multi-layer projection to perform a nonlinear mapping on the frequency domain representation: ； in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is selected to capture the nonlinear dependencies of the load frequency domain representation; the mapped frequency domain representation... As the frequency domain feature enhancement matrix, and the input feature map Perform the Hadamard product to enhance the frequency domain of the input feature map: ； in, This represents the obtained enhanced feature map.

6. A non-invasive industrial load decomposition device, employing the non-invasive industrial load decomposition method according to any one of claims 1-5, characterized in that, include: The aggregated active power sequence data acquisition and preprocessing unit is used to acquire aggregated active power sequence data in industrial scenarios through smart meters and perform preprocessing to obtain processed aggregated active power sequence data. The feature sequence acquisition unit is used to construct a periodic timestamp embedding module based on the processed aggregated active power sequence data. The periodic timestamp embedding module is used for mapping and fusion processing to obtain a fused feature sequence; A parallel multi-scale global context aggregation unit is used to construct an improved Transformer module, which replaces the traditional self-attention mechanism with a parallel multi-scale branch enhanced linear attention module. The fused feature sequence is input into the improved Transformer module for temporal feature extraction. The parallel multi-scale branch enhanced linear attention module consists of several scale branches of parallel operations, and each scale branch integrates a word generation module and an amplitude-aware linear attention module at a specified scale. Each branch uses parallel deep convolutions with different kernel sizes to extract multi-scale local features and output multi-scale feature maps. The extracted multi-scale feature maps are then fed into the amplitude-aware linear self-attention module of the corresponding scale branch. Parallel processing with linear computational complexity is used to model the global context dependency and output the single-scale context features of each branch. The single-scale context features output from each branch are concatenated and linearly mapped and fused. Then, combined with pre-normalization and residual connection mechanisms, they are input into the feedforward neural network of the improved Transformer module for nonlinear feature transformation, resulting in the temporal global context aggregation features output by the improved Transformer module. The enhanced feature map generation unit is used to input the time-domain global context aggregation features into the frequency-domain feature enhancement module. The frequency-domain feature enhancement module performs a one-dimensional discrete cosine transform on the time-domain global context aggregation features channel by channel to extract the comprehensive frequency-domain features and performs a nonlinear mapping on them to obtain the frequency-domain feature enhancement matrix. The matrix is ​​then used to perform a Hadamard product with the time-domain global context aggregation features to achieve frequency-domain enhancement of the time-domain global context aggregation features and output the enhanced feature map. The active power decomposition result output unit maps the enhanced feature map from the feature space to the power value space through the output projection head, and outputs the active power decomposition result of the target device.

7. A non-invasive industrial load decomposition device according to claim 6, characterized in that, In the aggregated active power sequence data acquisition and preprocessing unit, during the preprocessing of the aggregated active power sequence data, the aggregated active power sequence is divided into several subsequences by calculating the mean and variance of the total power sequence; the subsequences are then converted into sequences with a mean of 0 and a standard deviation of approximately 1 through standardization processing to obtain the processed aggregated active power sequence data.

8. A non-invasive industrial load decomposition device according to claim 7, characterized in that, In the fusion feature sequence acquisition unit, during the mapping and fusion processing by the periodic timestamp embedding module to obtain the fusion feature sequence, the time scalar is decomposed into three scalar components: minutes, hours, and days of the week. Each scalar component is mapped to a two-dimensional time embedding vector in the unit circle geometric space through trigonometric transformation. The three two-dimensional time embedding vectors are concatenated and then fused through a learnable linear projection layer to obtain a global time-series embedding vector. The global time-series embedding vector is fused with the processed aggregated active power sequence data to obtain the fusion feature sequence. The formula for calculating the two-dimensional time embedding vector is: ； In the formula, For the first The time component The two-dimensional temporal embedding vector corresponding to each sampling point; For the first The time component The time scalar value of each sampling point; For the first The physical period corresponding to each time component; The expression for the global temporal embedding vector is: ； In the formula, For the first The global temporal embedding vector corresponding to each timestamp; Linear(.) is a learnable linear projection layer; Concat(.) is a vector concatenation operation; These are two-dimensional time embedding vectors obtained by triangulating the "minute, hour, and weekday" components of the corresponding time scalars.

9. A non-invasive industrial load decomposition device according to claim 8, characterized in that, Within the parallel multi-scale global context aggregation unit, feature mapping and attention calculation are performed within each scale branch of the improved Transformer module as follows: This branch uses deep convolution to extract scale feature maps from the fused feature sequence, and performs linear projection on the scale feature maps to generate query matrix, key matrix and value matrix specific to this scale branch; the amplitude-aware linear self-attention module in this branch is configured to solve the feature divergence problem of linear attention by introducing an amplitude-aware mechanism; The amplitude-aware linear self-attention module uses a nonlinear kernel function to map the query matrix and the key matrix to a non-negative feature space. The amplitude-aware linear self-attention module introduces a dynamic scaling factor and a dynamic offset term that depend on the amplitude of the query feature to reconstruct the linear attention calculation process; For the A query vector, the amplitude-aware mechanism introduces a dynamic scaling factor. and dynamic offset terms Attention score is defined as: ； Among them, dynamic scaling factor With dynamic offset terms All are based on the dynamic calculation of the magnitude of the query vector, specifically defined as: ； ； in, Given the sequence length; from this, we can deduce: ； The amplitude-aware linear self-attention mechanism adaptively adjusts the attention distribution based on changes in the magnitude of the query vector. After obtaining the attention score, apply the attention score to the value vector. The complete computation of the amplitude-aware linear self-attention mechanism is shown below: ； In the actual calculation process, pre-calculation and As a context vector, and shared among all query vectors, the overall attention computation requires only one accumulation operation on the sequence dimension to obtain the single-scale context features corresponding to each scale branch. After all scale branches have completed parallel computation, the single-scale context features output by each branch are concatenated along the channel dimension and fused into multi-scale attention features through a linear projection layer. Finally, the multi-scale attention features are combined with pre-normalization and residual connection mechanisms and input into the feedforward neural network. After nonlinear activation and mapping through a fully connected layer, the temporal global context aggregation features are output.

10. A non-invasive industrial load decomposition device according to claim 9, characterized in that, In the enhanced feature map generation unit, during the process of generating the enhanced feature map through the frequency domain feature enhancement module, the time domain global context aggregation feature is decoupled into several independent single feature vectors along the feature dimension; a one-dimensional discrete cosine transform is performed on each single feature vector to obtain the frequency domain representation; the frequency domain representations of all channels are stacked into a tensor and input into a multilayer perceptron to generate the frequency domain feature enhancement matrix. The enhanced feature map is obtained by performing a Hadamard product between the matrix and the aforementioned temporal global context aggregated features. The specific processing flow of the frequency domain feature enhancement module in generating frequency feature enhanced feature maps is as follows: The feature decoupling process uses the aggregated features of the temporal global context as the input feature map. Decoupled along the feature dimension A single independent feature vector: ； in, Indicates the first Feature vectors of each channel; The frequency domain feature extraction process based on one-dimensional discrete cosine transform (DCT) utilizes the fundamental functions of 1D DCT to convert the time-domain signal into a frequency-domain signal. The definition of the fundamental functions of 1D DCT is as follows: ； in, This is the time position index corresponding to the feature vector. This represents the index of the frequency components of the transformed sequence; each feature vector... Positionally with all basic functions Multiplying these components yields the energy intensity representation of each eigenvector at different frequency components, as shown in the following formula: ； This process is viewed as an element-wise weighted summation of the input feature vector and its frequency components; by calculating all... Given several frequency components, the frequency domain representation of the eigenvector is obtained. : ； Then, the frequency domain representations of the 1D feature vectors of all channels are re-stacked along the feature dimension to obtain the comprehensive frequency domain representation of the input feature map: ； The frequency domain feature enhancement process utilizes multi-layer projection to perform a nonlinear mapping on the frequency domain representation: ； in, and All are linear weight matrices that can be learned by the projection layer. For the sigmoid function, For nonlinear activation functions, the ReLU activation function is selected to capture the nonlinear dependencies of the load frequency domain representation; the mapped frequency domain representation... As the frequency domain feature enhancement matrix, and the input feature map Perform the Hadamard product to enhance the frequency domain of the input feature map: ； in, This represents the obtained enhanced feature map.

Images