Method and device for generating electricity load decomposition data

Abstract

The present invention provides a method and device for generating electricity load decomposition data based on a conditional denoising diffusion model. The method relates to the field of electricity load data generation and includes obtaining raw electricity load data, wherein the raw electricity load data includes date and time data; preprocessing the raw electricity load data to obtain processed data; encoding the date and time data to obtain date-coded data; constructing a conditional denoising diffusion model; training the conditional denoising diffusion model using the processed data and the date-coded data to obtain a trained conditional denoising diffusion model; and randomly resampling the raw electricity load data using the trained conditional denoising diffusion model to obtain new electricity load data. The method aims to ensure that the generated data is both diverse and strongly correlated with date and time information and electricity load data, providing more reliable, accurate, and rich electricity load monitoring data.

Description

Technical Field

[0001] The present invention relates to the field of electricity load data generation, and in particular to a method and device for generating electricity load decomposition data based on a conditional denoising diffusion model. Background Art

[0002] Electricity load decomposition technology uses only a building's main line electricity meter to identify individual electrical devices and their operating status. It has wide application in areas such as building energy conservation, smart cities, and smart grids. Currently, this technology is trained using large amounts of labeled electrical device data. However, collecting such data is challenging, making it a major bottleneck in developing universal, non-intrusive electricity load monitoring solutions.

[0003] Several public datasets exist that cover energy consumption information for electrical appliances. These datasets aim to measure the energy consumption of household appliances using low-frequency or high-frequency intrusive monitoring devices. By continuously monitoring household energy demand, these datasets provide comprehensive and long-term measurements for studying user behavior in residential buildings. These datasets are very useful for performing electricity load decomposition tasks because they provide detailed electricity load monitoring data at the device level. However, the scale of currently available public datasets is relatively limited. To alleviate data limitations, a method for generating residential building electricity load monitoring data is needed. Current data generation algorithms based on generative adversarial models have some problems, such as mode collapse, which leads to a lack of diversity in the generated samples, and they often ignore the strong correlation between date and time information and electricity load data. Summary of the Invention

[0004] In order to solve the above problems, an embodiment of the present invention provides a method for generating electricity load decomposition data based on a conditional denoising diffusion model.

[0005] In a first aspect, an embodiment of the present invention includes a method for generating electricity load decomposition data based on a conditional denoising diffusion model, comprising obtaining original electricity load data, wherein the original electricity load data includes date and time data; S2, preprocessing the original electricity load data to obtain processed data; S3, encoding the date and time data to obtain date-coded data; S4, constructing a conditional denoising diffusion model, and training the conditional denoising diffusion model using the processed data and the date-coded data to obtain a trained conditional denoising diffusion model; S5, randomly resampling the original electricity load data using the trained conditional denoising diffusion model to obtain simulation data.

[0006] Optionally, the S2 includes:

[0007] S21, resampling the original power load data;

[0008] S22. Fill in the missing values of the original power load data;

[0009] S23: Generate a power load curve based on the original power load data, and normalize the power load curve to adjust the data span.

[0010] Optionally, the S3 includes:

[0011] S31. Extracting time features from the time tag based on the date and time data;

[0012] S32. Integrate multi-scale time information using the time features, wherein monthly data is encoded using a one-hot vector format, and daily data is encoded using a cosine periodic feature.

[0013] Optionally, the conditional denoising diffusion model includes a forward process and a backward process, wherein the forward process includes a forward noise addition stage and a forward diffusion stage, wherein in the forward diffusion stage, the completely random noise satisfying the Gaussian distribution obtained by adding Gaussian noise to the processed data can be expressed as:

[0014]

[0015]

[0016]

[0017] α t =1-β t

[0018] Where x0 is the processed data, β t For the predefined schedule [β1,β2,…,β t ] in the unit, ε t is Gaussian noise sampled from a standard normal distribution.

[0019] Optionally, the backward process of the conditional denoising diffusion model includes:

[0020] In the backward denoising stage, the completely random noise is gradually removed and data reconstruction is performed, wherein the conditional probability formula is calculated as:

[0021]

[0022] Where, is the mean of the conditional probability, is the variance of the conditional probability.

[0023] Optionally, using a predicted noise model in the backward denoising stage includes:

[0024] S41, use 4 convolutional layers for downsampling, each layer reduces the feature dimension by half;

[0025] S42, use 4 layers of deconvolution layers to upsample the feature information, each layer amplifies the feature dimension and splices it with the feature map of the corresponding contraction path;

[0026] S43, completing the network transmission and the expansion and reduction of the number of channels through the residual structure;

[0027] S44. Use the last convolutional layer to adjust the output data dimension to be the same as the input to obtain the predicted value of the noise.

[0028] Optionally, the prediction noise model further includes:

[0029] Add a cross attention layer between each connection layer, the formula is:

[0030]

[0031]

[0032]

[0033]

[0034] Where, Represents the learnable weight matrix in the neural network, φ i (x t ) represents the input x t The intermediate vector after passing through the i-th layer of the network structure, ψ(y) defines the encoding of the conditional information, and d is the dimension of the input vector.

[0035] In a second aspect, an embodiment of the present invention further includes a device for generating electricity load decomposition data based on a conditional denoising diffusion model, comprising a data acquisition module for acquiring original electricity load data, wherein the original electricity load data includes date and time data;

[0036] A preprocessing module preprocesses the original power load data to obtain processed data;

[0037] A date encoding module encodes the date and time data to obtain date encoding data;

[0038] A model training module constructs a conditional denoising diffusion model, and trains the conditional denoising diffusion model using the processed data and the date code data to obtain a trained conditional denoising diffusion model;

[0039] The simulation module uses the trained conditional denoising diffusion model to randomly resample the original power load data to obtain simulation data.

[0040] Optionally, the preprocessing module includes:

[0041] A resampling module, resampling the original power load data;

[0042] A data filling module, for filling missing values in the original power load data;

[0043] The normalization module generates an electricity load curve based on the original electricity load data, and normalizes the electricity load curve to adjust the data span.

[0044] Optionally, the date coding module includes:

[0045] A feature extraction module extracts time features from the time tag based on the date and time data;

[0046] The encoding module integrates multi-scale time information using the time features, wherein the monthly data is encoded using a one-hot vector format and the daily data is encoded using cosine periodic features.

[0047] The embodiments of the present invention include the following advantages:

[0048] This paper designs a conditional denoising diffusion model to mine existing residential electricity usage characteristics from existing data, training a generative model to generate electricity load data. Furthermore, because data generation uses date and time codes as conditional information, the generated data exhibits a strong correlation with date, better reflecting users' electricity usage habits. This can aid downstream tasks such as load decomposition, identifying residential electricity usage patterns, and establishing virtual power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0050] Figure 1 This is a flow chart of a method for generating electricity load decomposition data based on a conditional denoising diffusion model provided by an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the structure of a device for generating power load decomposition data based on a conditional denoising diffusion model provided by an embodiment of the present invention;

[0052] Figure 3 Schematic diagram of the conditional denoising diffusion model structure;

[0053] Figure 4 Schematic diagram of the prediction noise model in the conditional denoising diffusion model;

[0054] Figure 5 A schematic diagram of a power load curve generated by an embodiment of the present invention.

[0055] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. DETAILED DESCRIPTION

[0056] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0057] The following describes in detail the method and device for generating power load decomposition data based on the conditional denoising diffusion model provided in the embodiments of the present application through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0058] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0059] Example 1

[0060] The first embodiment of the present invention relates to a method for generating power load decomposition data based on a conditional denoising diffusion model, the process of which is as follows: Figure 1 Shown, incl.

[0061] S1. Obtain original power load data, wherein the original power load data includes date and time data.

[0062] Collect a large amount of representative residential electricity load data, which includes the electricity load curves and date and time information of different equipment in residential buildings.

[0063] S2. Preprocessing the original power load data to obtain processed data;

[0064] The original data are preprocessed to fill the missing values of the original electricity load data, normalize the original curve to adjust the data span, and resample all data to equal time intervals.

[0065] S21, resampling the original power load data;

[0066] S22. Fill in the missing values of the original power load data;

[0067] S23: Generate a power load curve based on the original power load data, and normalize the power load curve to adjust the data span.

[0068] The data is resampled and the data granularity is adjusted to 15 minutes. The missing values of the original power load data are filled and the original curve is normalized to adjust the data span. The normalization formula is:

[0069]

[0070] j is the jth value in the input sequence data, X max ,X min The maximum and minimum values of active power in the input data column x j,nor Refers to the standardized x j , this normalization formula maps the entire time series data to between 0 and 1 to eliminate the impact of different scales between data.

[0071] S3. Encode the date and time data to obtain date coded data;

[0072] Encode date and time data using one-hot vector and cosine periodic encoding.

[0073] S31. Extracting time features from the time tag based on the date and time data;

[0074] S32. Integrate multi-scale time information using the time features, wherein monthly data is encoded using a one-hot vector format, and daily data is encoded using a cosine periodic feature.

[0075] Date and time data is encoded using one-hot vectors and cosine periodic encoding. Multiple features are extracted from each time tag: the day of the month, the day of the week, and the month of the year. These features are used to integrate multi-scale temporal information. The month of the year is encoded in one-hot vector format. The day of the month and the day of the week features are then encoded as sine and cosine features to reflect the temporal periodicity of the user's energy consumption patterns.

[0076] S4. Construct a conditional denoising diffusion model, and train the conditional denoising diffusion model using the processed data and the date code data to obtain a trained conditional denoising diffusion model.

[0077] like Figure 3 As shown, it is the diffusion model of the present invention. Figure 4 The processed data and date-coded data are input into the prediction noise neural network to complete the training.

[0078] The conditional denoising diffusion model includes a forward process and a backward process, and the forward process includes a forward denoising stage and a forward diffusion stage.

[0079] The forward noise addition stage does not require training. In the forward diffusion stage, the data is gradually contaminated by Gaussian noise to become completely random noise that satisfies the Gaussian distribution. In the forward diffusion stage, the completely random noise that satisfies the Gaussian distribution obtained by adding Gaussian noise to the processed data can be expressed as:

[0080]

[0081]

[0082]

[0083] α t =1-β t

[0084] Where x0 is the processed data, β t For the predefined schedule [β1,β2,…,β t ] in the unit, ε t is Gaussian noise sampled from a standard normal distribution.

[0085] The backward process of the conditional denoising diffusion model includes:

[0086] The backward denoising stage is the process of gradually removing noise from the data and reconstructing the data, that is, solving the conditional probability. It is assumed that the conditional probability in the backward denoising stage conforms to the Gaussian distribution.

[0087] In the backward denoising stage, the completely random noise is gradually removed and the data is reconstructed, wherein the conditional probability formula is calculated as:

[0088]

[0089] Where, is the mean of the conditional probability, is the variance of the conditional probability.

[0090] According to the Bayesian formula and the properties of Gaussian distribution, the mean of the conditional probability in the backward stage can be deduced and variance in and It can be expressed as:

[0091]

[0092]

[0093] According to the forward propagation formula, the original input x0 is expressed as x after t steps of noise addition t , we can get

[0094]

[0095] During training, the goal of the algorithm is to minimize the actual noise ε and the predicted noise ε by optimizing the negative log-likelihood using a variational lower bound. θ The loss term can be parameterized as

[0096]

[0097] Where t is obtained by uniform sampling from {1,…,T}, ε is the actual amount of Gaussian noise used in the diffusion noise addition process, and ε θ Predict the amount of noise for the neural network, θ is an abstract representation of all parameters in the neural network, and x0 represents the original input.

[0098] The prediction noise model in the backward denoising stage includes:

[0099] S41, downsampling is performed using 4 convolutional layers, where each layer reduces the feature dimension by half;

[0100] S42, use 4 layers of deconvolution layers to upsample the feature information, each layer amplifies the feature dimension and splices it with the feature map of the corresponding contraction path;

[0101] S43, completing the network transmission and the expansion and reduction of the number of channels through the residual structure;

[0102] S44. Use the last convolutional layer to adjust the output data dimension to be the same as the input to obtain the predicted value of the noise.

[0103] The prediction noise model architecture in the backward denoising stage is composed of a U-Net with the same input and output dimensions. The direct input of the network is the active power information x and time t, and the output of the network is ε θThe network first uses 4 convolutional layers for downsampling, reducing the feature dimension by half each time. Then, 4 deconvolution layers are used to upsample the feature information. Each layer amplifies the feature dimension and splices it with the feature map of the corresponding contraction path. To prevent gradient disappearance and gradient explosion, a residual structure is used to complete the network transmission and the expansion and reduction of the number of channels. The last convolutional layer of the network is responsible for adjusting the output data dimension to be the same as the input, so as to obtain the predicted value ε of the noise ε θ .

[0104] The prediction noise model in the backward denoising stage also includes:

[0105] A cross-attention layer is added between each connection layer of the U-net network to increase the influence of conditional features on prediction noise.

[0106] Add a cross attention layer between each connection layer, the formula is:

[0107]

[0108]

[0109]

[0110]

[0111] Where, Represents the learnable weight matrix in the neural network, φ i (x t ) represents the input x t The intermediate vector after passing through the i-th layer of the network structure, ψ(y) defines the encoding of the conditional information, and d is the dimension of the input vector.

[0112] In this paper, we encode temporal information as ψ(y) to participate in neural network training. Therefore, the final loss function is expanded to:

[0113]

[0114] During training, this paper uses the Classifier-Free Guidance mechanism to jointly train the conditional and unconditional models. Classifier-Free Guidance uses the same neural network to parameterize both models. For the unconditional model, the conditional information encoding ψ(y) is set to zero. The conditional and unconditional models are trained jointly. Therefore, the final loss function is changed to:

[0115]

[0116] Where ω is a predefined parameter that controls the influence ratio between the conditional and unconditional models. The advantage of using this training mechanism is that it can combine the two models simultaneously, achieving a balance between sample quality and diversity.

[0117] S5. Use the trained conditional denoising diffusion model to randomly resample the original power load data to obtain simulation data.

[0118] The present invention obtains raw electricity load data, including date and time data, preprocesses the raw electricity load data to obtain processed data, encodes the date and time data to obtain date-coded data, constructs a conditional denoising diffusion model, trains the conditional denoising diffusion model using the processed data and the date-coded data to obtain a trained conditional denoising diffusion model, performs random resampling from the trained conditional denoising diffusion model to obtain new electricity load data, and inputs the new electricity load data and a preset specified date into the trained conditional denoising diffusion model to obtain simulation data. This method aims to ensure that the generated data is both diverse and strongly correlated with date and time information and electricity load data, thereby providing more reliable, accurate, and rich electricity load monitoring data.

[0119] Example 2

[0120] The second embodiment of the present invention relates to a device 200 for generating power load decomposition data based on a conditional denoising diffusion model, such as Figure 2 Shown, including:

[0121] The data acquisition module 201 acquires original power load data, wherein the original power load data includes date and time data;

[0122] A preprocessing module 202 preprocesses the original power load data to obtain processed data;

[0123] The date encoding module 203 encodes the date and time data to obtain date encoding data;

[0124] A model training module 204 constructs a conditional denoising diffusion model, and trains the conditional denoising diffusion model using the processed data and the date code data to obtain a trained conditional denoising diffusion model;

[0125] The simulation module 205 uses the trained conditional denoising diffusion model to randomly resample the original power load data to obtain simulation data.

[0126] Furthermore, the pre-processing module 202 includes:

[0127] A resampling module 2021 resamples the original power load data;

[0128] A data filling module 2022 fills in missing values of the original power load data;

[0129] The normalization module 2023 generates an electricity load curve based on the original electricity load data, and normalizes the electricity load curve to adjust the data span.

[0130] Furthermore, the date coding module 203 includes:

[0131] A feature extraction module 2031 extracts time features from the time tag based on the date and time data;

[0132] The encoding module 2032 integrates multi-scale time information using the time features, wherein the monthly data is encoded using a one-hot vector format and the daily data is encoded using a cosine periodic feature.

[0133] The present invention constructs an electricity load decomposition data generation device based on a conditional denoising diffusion model, comprising a data acquisition module for acquiring original electricity load data, wherein the original electricity load data includes date and time data; a preprocessing module for preprocessing the original electricity load data to obtain processed data; a date coding module for coding the date and time data to obtain date-coded data; a model training module for constructing a conditional denoising diffusion model, and training the conditional denoising diffusion model using the processed data and the date-coded data to obtain a trained conditional denoising diffusion model; and a simulation module for performing random resampling from the trained conditional denoising diffusion model to obtain simulation data.

[0134] Example 3

[0135] This embodiment performs neural network training based on the Dataport database. Dataport contains device-level active power consumption data (in kilowatts) from 73 households in one year, with a time interval of 1 minute. For active power information, this paper first downsamples the power information, resampling the data set sampling interval from 1 minute to 15 minutes, reducing the number of parameters and making it more consistent with the actual power scheduling interval. The following 6 devices in the data column are selected: air conditioner, furnace, water heater, dishwasher, washing machine, dryer and total power load, a total of 7 columns as input to the neural network for training. For the date and time information in the data, encode it as described in step S3.

[0136] This paper inputs the above data into the neural network for training. After the final training is completed, the model is sampled with the date information encoding as a condition to obtain the final result.

[0137] The results of this paper are compared with the original data. Figure 5shown.

[0138] from Figure 5 It can be seen that the electricity load curve generated by the proposed method has a good correlation with the date. Comparing the proposed method with real-world data, the air conditioning load accounts for a high proportion of the electricity load in July, significantly higher than the data for March. However, a comparison with the results of the commonly used VAE variational autoencoder model shows that the latter does not exhibit a good correlation with the date.

[0139] This example designs a conditional denoising diffusion model to mine existing residential electricity usage characteristics from existing data, training a generative model to generate electricity load data. Furthermore, because data generation uses date and time codes as conditional information, the generated data exhibits a strong correlation with date and is more consistent with user electricity usage habits. This facilitates downstream tasks such as load decomposition, identifying residential electricity usage patterns, and establishing virtual power plants.

[0140] The above describes the technical solutions provided by this application in detail. Specific examples are used herein to illustrate the principles and implementation methods of this application. The descriptions of the above examples are intended only to facilitate understanding of this application, and the contents of this specification should not be construed as limiting this application. Furthermore, those skilled in the art will appreciate that, based on this application, variations in the specific implementation methods and scope of application may occur in various forms. It is not necessary and impossible to enumerate all implementation methods here, but obvious variations or modifications arising therefrom remain within the scope of protection of this application. It should be noted that the order of the embodiments of this application described above is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description of specific embodiments in this specification is limited to specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0141] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0142] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A method for generating electricity load decomposition data based on a conditional denoising diffusion model, characterized in that: include: S1. Obtaining original power load data, wherein the original power load data includes date and time data; S2. Preprocessing the original power load data to obtain processed data; S3. Encode the date and time data to obtain date coded data; S4, constructing a conditional denoising diffusion model, and training the conditional denoising diffusion model using the processed data and the date code data to obtain a trained conditional denoising diffusion model; S5. Randomly resample the original power load data using the trained conditional denoising diffusion model to obtain simulation data; The conditional denoising diffusion model includes a forward process and a backward process, wherein the forward process includes a forward denoising stage and a forward diffusion stage. In the forward diffusion stage, the completely random noise satisfying the Gaussian distribution obtained by adding Gaussian noise to the processed data can be expressed as: a t ＝1-β t Where x0 is the processed data, β t For the predefined schedule [β1,β2,…,β t ] in the unit, ε t is the Gaussian noise sampled from the standard normal distribution; The backward process includes: In the backward denoising stage, the completely random noise is gradually removed and data reconstruction is performed, wherein the conditional probability formula is calculated as: Where, is the mean of the conditional probability, is the variance of the conditional probability; The predicted noise model is used in the backward denoising stage, including: S41, downsampling is performed using 4 convolutional layers, where each layer reduces the feature dimension by half; S42, use 4 layers of deconvolution layers to upsample the feature information, each layer amplifies the feature dimension and splices it with the feature map of the corresponding contraction path; S43, completing the network transmission and the expansion and reduction of the number of channels through the residual structure; S44. Use the last convolutional layer to adjust the output data dimension to be the same as the input, and obtain the predicted value of the noise; The prediction noise model also includes: Add a cross attention layer between each connection layer, the formula is: Where, Represents the learnable weight matrix in the neural network, φ i (x t ) represents the input x t The intermediate vector after passing through the i-th layer of the network structure, ψ(y) defines the encoding of the conditional information, and d is the dimension of the input vector.

2. The method for generating electricity load decomposition data based on a conditional denoising diffusion model according to claim 1, characterized in that: The S2 includes: S21, resampling the original power load data; S22. Fill in the missing values of the original power load data; S23: Generate a power load curve based on the original power load data, and normalize the power load curve to adjust the data span.

3. The method for generating electricity load decomposition data based on a conditional denoising diffusion model according to claim 2, characterized in that: The S3 includes: S31. Extracting time features from the time tag based on the date and time data; S32. Integrate multi-scale time information using the time features, wherein monthly data is encoded using a one-hot vector format, and daily data is encoded using a cosine periodic feature.

4. A device for generating power load decomposition data based on a conditional denoising diffusion model, characterized in that: include: A data acquisition module is used to acquire original power load data, wherein the original power load data includes date and time data; A preprocessing module preprocesses the original power load data to obtain processed data; A date encoding module encodes the date and time data to obtain date encoding data; A model training module constructs a conditional denoising diffusion model, and trains the conditional denoising diffusion model using the processed data and the date code data to obtain a trained conditional denoising diffusion model; A simulation module, which randomly resamples the original power load data using the trained conditional denoising diffusion model to obtain simulation data; The conditional denoising diffusion model includes a forward process and a backward process, wherein the forward process includes a forward denoising stage and a forward diffusion stage. In the forward diffusion stage, the completely random noise satisfying the Gaussian distribution obtained by adding Gaussian noise to the processed data can be expressed as: a t ＝1-β t Where x0 is the processed data, β t For the predefined schedule [β1,β2,…,β t ] in the unit, ε t is the Gaussian noise sampled from the standard normal distribution; The backward process includes: In the backward denoising stage, the completely random noise is gradually removed and data reconstruction is performed, wherein the conditional probability formula is calculated as: Where, is the mean of the conditional probability, is the variance of the conditional probability; The predicted noise model is used in the backward denoising stage, including: S41, use 4 convolutional layers for downsampling, each layer reduces the feature dimension by half; S42, use 4 layers of deconvolution layers to upsample the feature information, each layer amplifies the feature dimension and splices it with the feature map of the corresponding contraction path; S43, completing the network transmission and the expansion and reduction of the number of channels through the residual structure; S44. Use the last convolutional layer to adjust the output data dimension to be the same as the input, and obtain the predicted value of the noise; The prediction noise model also includes: Add a cross attention layer between each connection layer, the formula is: Where, Represents the learnable weight matrix in the neural network, φ i (x t ) represents the input x t The intermediate vector after passing through the i-th layer of the network structure, ψ(y) defines the encoding of the conditional information, and d is the dimension of the input vector.

5. The device for generating power load decomposition data based on a conditional denoising diffusion model according to claim 4, characterized in that: The pre-processing module comprises: A resampling module, resampling the original power load data; A data filling module, for filling missing values in the original power load data; The normalization module generates an electricity load curve based on the original electricity load data, and normalizes the electricity load curve to adjust the data span.

6. The device for generating power load decomposition data based on a conditional denoising diffusion model according to claim 5, characterized in that: The date coding module includes: A feature extraction module extracts time features from the time tag based on the date and time data; The encoding module integrates multi-scale time information using the time features, wherein the monthly data is encoded using a one-hot vector format and the daily data is encoded using cosine periodic features.

Images