Non-intrusive load decomposition method based on super-resolution

Abstract

The invention relates to a non-intrusive load decomposition method based on super-resolution. The method comprises the following steps: acquiring low-resolution power sequence data of load equipment; and inputting the low-resolution power sequence data into a conditional noise prediction network CoPNet to obtain prediction noise, and denoising an initial pure noise high-resolution data sequence according to the prediction noise to obtain reconstructed power sequence data, the conditional noise prediction network CoPNet is constructed based on a de-noising diffusion probability model DDPM and is obtained through training of a load sample data set, and the load sample data set comprises low-resolution power sequence training samples; and decomposing the reconstructed power sequence data to obtain power data of the single load equipment. According to the method, the load super-resolution is divided into a high-resolution data mapping process and a super-resolution data reconstruction process, so that accurate modeling of load sequence distribution and effective generation of a high-resolution load sample are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load monitoring, and in particular to a non-invasive load decomposition method based on super-resolution. Background Art

[0002] In the era of smart electricity, in addition to being able to process accurate and real-time fee settlement information, electric energy meters can also further analyze the internal load components of users in combination with artificial intelligence algorithms, that is, to achieve non-intrusive load monitoring or load decomposition (NILM). Non-intrusive load monitoring technology uses artificial intelligence algorithms to analyze the total electricity consumption data on smart meters, and monitor the refined energy consumption data of each load within the monitoring range.

[0003] Since the amount of load information obtained by low-frequency sampling data is limited, it cannot provide sufficient detailed features for load decomposition to ensure the accuracy of load decomposition. Therefore, load super-resolution technology is used to reconstruct high-resolution load data rich in load detail information from low-frequency sampling data. The existing GAN-based load super-resolution method has problems such as over-smoothing and mode collapse, and cannot fully achieve a good reconstruction effect in the scenario of multi-type load data reconstruction. Summary of the invention

[0004] The purpose of the present invention is to provide a non-intrusive load decomposition method based on super-resolution, which can achieve better data reconstruction effect for different types of loads without designing overly complex loss functions.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A non-intrusive load decomposition method based on super-resolution, comprising:

[0007] Obtain low-resolution power series data of load equipment;

[0008] Input the low-resolution power sequence data into a conditional noise prediction network CoPNet to obtain predicted noise, denoise the initial pure noise high-resolution data sequence according to the predicted noise, and obtain reconstructed power sequence data, wherein the conditional noise prediction network CoPNet is constructed based on a denoising diffusion probability model DDPM and is obtained through training with a load sample data set, and the load sample data set includes low-resolution power sequence training samples;

[0009] The reconstructed power sequence data is decomposed to obtain power data of a single load device.

[0010] Optionally, training the conditional noise prediction network CoPNet by using the load sample data set includes: mapping the original high-resolution power sequence data to a Gaussian distribution by forward adding Gaussian noise to obtain pure noise sequence samples;

[0011] Randomly sampling the pure noise sequence samples to obtain noisy high-resolution sequence training data;

[0012] The low-resolution power sequence training samples, the noisy high-resolution sequence training data and the real noise are used to train the conditional noise prediction network CoPNet.

[0013] Optionally, denoising the initial pure noise high-resolution data sequence according to the predicted noise includes:

[0014] The pure noise sequence samples are randomly sampled to obtain the initial pure noise high-resolution data sequence, and the initial pure noise high-resolution data sequence is gradually denoised according to the predicted noise to obtain the reconstructed power sequence data.

[0015] Optionally, the conditional noise prediction network CoPNet includes a data input module, a feature extraction module and a noise output module;

[0016] The data input module is used to upsample the low-resolution power sequence data and splice them in the channel dimension, output splicing features, and encode the time using position coding to obtain the position coding of the time;

[0017] The feature extraction module is used to extract feature information from the splicing feature and the time position code;

[0018] The noise output module is used to integrate the feature information through a convolution layer and a Reshape layer to output predicted noise.

[0019] Optionally, extracting feature information from the splicing feature and the time position encoding includes:

[0020] The preliminary feature information of the splicing feature is extracted through a convolutional layer, and the preliminary feature information and the temporal position code are input into a residual connection structure to extract deep feature information.

[0021] Optionally, the method for extracting deep feature information by the residual connection structure is:

[0022] x i+1 =x i +Swich(Conv1d{[Leaky ReLu(Conv1d(x))+Leaky ReLu(FC((Pos(t)))]})

[0023] Among them, x i+1 is the data input of the i+1th residual block, POS(t) is the position encoding of time step t, Conv1d is the one-dimensional convolutional layer, Leaky Relu is the activation function, and FC is the fully connected layer.

[0024] Optionally, the objective function of the conditional noise prediction network CoPNet is:

[0025]

[0026] Among them, L(θ) is the objective function, is the original high-resolution power sequence data, x L represents low-resolution power sequence data, ∈ is Gaussian noise, ∈ θ To predict the noise, is the noisy high-resolution sequence data at any time step.

[0027] The beneficial effects of the present invention are as follows: the present invention decomposes load super-resolution into two processes: high-resolution data mapping process and super-resolution data reconstruction process, so as to achieve accurate modeling of load sequence distribution and effective generation of high-resolution load samples. Furthermore, a conditional noise prediction network CoPNet is proposed, which can achieve effective reconstruction of high-resolution load sequence data without designing a complex loss function. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 A framework diagram of a non-intrusive load decomposition method based on super-resolution according to an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the structure of a conditional noise prediction network CoPNet according to an embodiment of the present invention;

[0031] Figure 3 Schematic diagrams for comparing indicators of various super-resolution methods according to embodiments of the present invention, wherein (a) is a schematic diagram for comparing numerical deviation indicators, (b) is a schematic diagram for comparing data distribution indicators, and (c) is a schematic diagram for comparing data quality indicators;

[0032] Figure 4 It is a schematic diagram of reconstruction results of different load data under various super-resolution methods according to an embodiment of the present invention;

[0033] Figure 5 A flowchart of a non-intrusive load decomposition method based on super-resolution according to an embodiment of the present invention;

[0034] Figure 6 A schematic diagram of relative errors of load decomposition using power data with different resolutions according to an embodiment of the present invention;

[0035] Figure 7 Schematic diagram of refined total load energy consumption obtained by different load decomposition algorithms under various super-resolution methods in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] DDPM is a generative model using a parameterized Markov chain, which mainly consists of a forward diffusion process of data denoising and a reverse diffusion process of data denoising. Figure 1 The DDPM-based load super-resolution reconstruction architecture shown in the figure mainly consists of two parts: high-resolution data mapping and super-resolution data reconstruction. In the high-resolution data mapping stage, the original high-resolution power sequence data distribution is converted into a Gaussian distribution that is easy to sample by forward adding Gaussian noise, so as to achieve the purpose of implicit learning from the load sequence data distribution; in the super-resolution data reconstruction stage, the Gaussian distribution is first sampled to obtain a high-resolution power sequence containing noise, that is, a pure noise sequence sample. Subsequently, the low-resolution power sequence is used as a conditional input, and the CoPNet network, that is, the DDPM-based load super-resolution model, is used to realize the noise prediction of the high-resolution power sequence. After gradual denoising, a high-resolution power sequence without noise is finally obtained. Using x0 H represents the high-resolution power sequence training sample, x L Represents low-resolution power sequence training samples.

[0039] A non-invasive load decomposition method based on super-resolution, specifically comprising:

[0040] Obtain low-resolution power series data of load equipment;

[0041] The low-resolution power sequence data is input into the conditional noise prediction network CoPNet to obtain the predicted noise, and the initial pure noise high-resolution data sequence is denoised according to the predicted noise to obtain the reconstructed power sequence data, wherein the conditional noise prediction network CoPNet is constructed based on the denoising diffusion probability model DDPM and obtained through training with the load sample data set, and the load sample data set includes low-resolution power sequence training samples;

[0042] The reconstructed power sequence data is decomposed to obtain the power data of a single load device.

[0043] Furthermore, training the conditional noise prediction network CoPNet through the load sample data set includes: mapping the original high-resolution power sequence data to a Gaussian distribution by forward adding Gaussian noise to obtain pure noise sequence samples;

[0044] Randomly sample the pure noise sequence samples to obtain the noisy high-resolution sequence training data;

[0045] The low-resolution power sequence training samples, noisy high-resolution sequence training data and real noise are used to train the conditional noise prediction network CoPNet.

[0046] Furthermore, denoising the initial pure noise high-resolution data sequence according to the predicted noise includes:

[0047] The pure noise sequence samples are randomly sampled to obtain an initial pure noise high-resolution data sequence, and the initial pure noise high-resolution data sequence is gradually denoised according to the predicted noise to obtain reconstructed power sequence data.

[0048] It should be noted that the high-resolution data mapping process includes: in order to more accurately estimate the data distribution of the high-resolution power sequence The mapping process maps the original high-resolution data to a Gaussian distribution through a diffusion process to ensure that the high-resolution data acquired in the reverse direction during the load super-resolution data reconstruction phase conforms to the original target distribution. By comparing the data from the previous step Adding Gaussian noise yields Only the last time step The transition probability in this process It can be expressed by the following formula:

[0049]

[0050] In the formula, β t is the Gaussian noise variance parameter, whose value is between 0 and 1, I is the unit matrix,

[0051] is a Gaussian distribution.

[0052] Based directly on the original high-resolution power sequence Get any time step Setting α t =1-β t , Using formula (1), we can obtain:

[0053]

[0054] in, represents the transition probability from the initial time step to the tth time step.

[0055] By introducing the standard normal random variable ∈ through the reparameterization technique, we can ensure Differentiable, and treat ∈ as a new Gaussian noise:

[0056]

[0057] In the parameter β t If the settings are reasonable, the final It will obey Gaussian distribution, thus realizing the mapping of power series data distribution.

[0058] Super-resolution data reconstruction process: This stage is the inverse process of the high-resolution data mapping process, using the real low-resolution sequence x L As conditional input, complete the generation of conditional probability The modeling of , by denoising the Gaussian distribution sampling data, generates a high-resolution power sequence that obeys the distribution of the original power sequence. It cannot be obtained directly, so the parameterization that follows the Gaussian distribution is adopted Make an estimate:

[0059]

[0060] Where μ θ With Σ θ are the parameterized mean and variance respectively. To simplify the calculation, Σ θ (x t ,t) is set to a fixed value and θ To predict noise, it is used to estimate the high-resolution data mapping process The added Gaussian noise ε is used to obtain the final high-resolution load sequence in the super-resolution data reconstruction process. Using the reparameterization technique To sample:

[0061]

[0062] Among them, z follows Gaussian distribution.

[0063] In order to gradually achieve accurate denoising of power data through formula (7), a CoPNet is proposed to fit ∈ θ Based on variational inference to optimize the negative log-likelihood estimate, the simplified CoPNet objective function L(θ) can be obtained as follows:

[0064]

[0065] Among them, L(θ) is the objective function, is the original high-resolution power sequence data, x L represents low-resolution power sequence data, ∈ is Gaussian noise, ∈ θ To predict the noise, is the noisy high-resolution sequence data at any time step.

[0066] Furthermore, the conditional noise prediction network CoPNet includes a data input module, a feature extraction module and a noise output module;

[0067] A data input module is used to upsample the low-resolution power sequence data and splice it in the channel dimension, output the splicing features, encode the time using the position code, and obtain the position code of the time;

[0068] A feature extraction module is used to extract feature information by encoding the splicing features and time positions;

[0069] The noise output module is used to integrate feature information through the convolution layer and the Reshape layer and output the predicted noise.

[0070] Furthermore, extracting feature information from the splicing feature and the temporal position encoding includes: extracting preliminary feature information of the splicing feature through a convolutional layer, and inputting the preliminary feature information and the temporal position encoding into a residual connection structure to extract deep feature information.

[0071] like Figure 2 As shown in the figure, the network structure of CoPNet consists of three parts: data input, feature extraction and noise output. In the data input part, CoPNet first performs L Upsampling is performed, and each time step is repeatedly sampled so that the data before and after upsampling can reflect the same consumption at the energy consumption level. H Under the premise of keeping consistent, splicing is performed on the channel dimension to obtain the output x of this layer. When processing the position of the power sequence, the position encoding method proposed by Vaswani et al. is used to encode the time t into the network to obtain a unique vector representation Pos t, to enhance the prediction network’s awareness of implicit temporal dependencies in the sequence.

[0072]

[0073] Where Pos t,2i With Pos t,2i+1 Encode the 2ith and 2i+1th vector components at time t respectively, and d is the vector dimension.

[0074] In the feature extraction part, two one-dimensional convolutional layers are first used to extract the preliminary feature information of the sequence data, and the preliminary extracted feature information is sent to the residual connection structure composed of five residual blocks for deep feature information extraction. The residual structure of the deep feature extraction part is used to solve the gradient diffusion problem caused by the increase in network depth. A fully connected layer is used in each residual block to ensure that the dimension of the position encoding is consistent with the current feature dimension. The data expression of each residual block is as follows.

[0075] x i+1 =x i +Swich(Conv1d{[Leaky ReLu(Conv1d(x))+Leaky ReLu(FC((Pos(t)))]})(10)

[0076] Among them, x i+1 is the output of the i-th residual block, and Pos(t) is the position code for t.

[0077] Finally, in the noise output part, the feature extraction output is integrated through a one-dimensional convolution layer and a Reshape layer to achieve the integration of deep high-dimensional features, completing the noise prediction task of super-resolution load data generated in t time steps. In the CoPNet network, in order to avoid the phenomenon of neuron necrosis caused by negative input, except for the output convolution layer of the residual block, which uses the Swish activation function with better smoothness, the other network layers use the Leaky ReLu activation function.

[0078] In terms of network structure parameters, the one-dimensional convolution layer of the upsampling part of the data input is composed of a convolution kernel of size 5 and stride 1; the two one-dimensional convolution layers used to extract preliminary feature information are composed of 16 convolution kernels of size 5 and stride 1. The convolution layer size and stride in the five residual blocks in the residual network are consistent, which are 5 and 1 respectively. The main difference between the five residual blocks is the number of fully connected layers and convolution kernels, which are 32, 64, 128, 64, and 32 respectively; the one-dimensional convolution layer near the output position is composed of a convolution kernel of size 7 and stride 1.

[0079] The training process and model inference process of CoPNet network include:

[0080] The presence of extreme value data in the load sample data set will seriously affect the convergence of the model. In addition, the Gaussian distribution used in the above-mentioned high-resolution data mapping and super-resolution data reconstruction process is a symmetric distribution with a mean of zero. Therefore, normalizing the input data to the range of [-1,1] before network training and model inference can better maintain symmetry while speeding up network training. The symmetric normalization method is shown below.

[0081]

[0082] In the CoPNet network training phase, we first randomly select uniform distributions from t~U(1,T) and The Gaussian distribution of is randomly sampled, and the noisy high-resolution sequence data in the high-resolution data mapping process is obtained by formula (3) Then the gradient descent method is used to optimize the network prediction noise ∈ θ The network parameters of CoPNet are trained with the L2 loss function of real noise. In the model inference stage, we first start with pure noise sequence samples. Random sampling is performed to obtain the initial pure noise high-resolution data sequence. Then CoPNet is used to predict the noisy data at each step in the super-resolution data reconstruction process. After T times of data denoising, the reconstructed high-resolution data x is finally obtained. SR The network training and model inference algorithms are shown in Tables 1 and 2.

[0083] Table 1

[0084]

[0085] Table 2

[0086]

[0087] Evaluation of the performance of the method of this embodiment on the load super-resolution task and the impact of data reconstruction on non-intrusive load decomposition includes:

[0088] (1) Dataset selection: Considering the actual application scenarios of super-resolution methods, this embodiment selects the AMPds dataset with low-frequency sampling and containing multiple load types. This dataset is the electricity consumption data of a household user in Canada from 2012 to 2013. The data collection time interval is 1 minute, and it contains 11 electrical characteristics such as voltage, current, and power of 21 power loads.

[0089] In the selection of load data, this embodiment selects six load devices, such as a dishwasher, a washing machine, and a refrigerator, as shown in Table 3. These six load devices cover different power levels and have greatly different working modes.

[0090] Table 3

[0091]

[0092] (2) To train the CoPNet network model parameters, this embodiment uses the active power from 2012 / 04 / 01 to 2012 / 11 / 27 as training data, and the active power from 2012 / 12 / 01 to 2013 / 03 / 01 as test data. This embodiment uses the original 1-minute time interval data as high-resolution data, and the 5-minute time interval data obtained by downsampling the data as low-resolution data.

[0093] The Adam optimizer was used for network training, and training techniques such as Dropout and Early stopping were not used. The relevant training parameter configurations are shown in Table 4. The experimental platform is based on the Windows 10 operating system and is completed on a computer consisting of an NVIDIA GTX1050 graphics card, an Intel Core i5-10400 CPU, and 32GB of memory. The code is implemented in the Tensorflow 2.8.0 framework in the Python3.9 compilation environment.

[0094] Table 4

[0095]

[0096] (3) In order to comprehensively evaluate the reconstructed data, this embodiment evaluates the three indicators of numerical deviation, data distribution and data quality. In terms of numerical deviation, this embodiment uses the root mean square error (RMSE) indicator for evaluation, which can be used to measure the deviation between the reconstructed value and the true value. In terms of data distribution, this embodiment uses the Hellinger distance (D H ) measures the reconstruction data distribution Q SR With the real data distribution Q H In terms of data quality, this embodiment uses the signal-to-noise ratio (SNR) to measure the proportion of noise contained in the reconstructed data.

[0097] (4) This embodiment selects two algorithms, linear interpolation (LERP) and cubic spline interpolation (CBRP), as comparison algorithms based on interpolation methods, and this embodiment selects SRNet and GAN as comparison algorithms based on deep learning methods.

[0098] From the performance of each algorithm Figure 3From (a) to (c), the DDPM method has achieved the best results in terms of numerical deviation, data distribution and data quality, which are 15.1%, 30.2% and 47.1% higher than the suboptimal methods. Although the DDPM method is not directly trained by the error between the reconstructed power sequence and the true power sequence, the super-resolution results can be closer to the true value by gradually predicting the data noise through CoPNet. In addition, DDPM implicitly learns the mapping relationship from Gaussian distribution to the original power sequence distribution by adding noise, so the reconstructed power sequence of DDPM is more in line with the target distribution.

[0099] The L2 loss function used by SRNet has a low penalty for smaller deviations, which makes it inconsistent with the actual situation in terms of power curve details; although GAN introduces two additional loss functions and another modified network structure, it is still unable to completely find a better solution space based on experimental results. The data interpolation method does not obtain the high-frequency information contained in low-resolution data like through learning, so the reconstructed data is quite different from the actual situation.

[0100] exist Figure 4 In the figure, except for the first column which is the original high-resolution power sequence data of each load, the other columns are the reconstruction results of each super-resolution algorithm for the load. In general, when the super-resolution factor is 5, the quality of the reconstructed data of each super-resolution method for the steady-state stage of the load is the best. The reason is that the power fluctuation of the load is relatively stable during the steady-state stage and the working state remains unchanged for a long time. Therefore, even if the interpolation algorithm is used, a relatively satisfactory reconstruction result can be obtained in this stage. For loads such as CDE, WOE, and CWE that frequently switch working states in the transient stage, since low-resolution data cannot capture the load state of a short duration, the reconstruction effect of each super-resolution algorithm is slightly worse than that of the steady-state stage.

[0101] In order to evaluate the performance of each super-resolution method at different super-resolution multiples, this embodiment is tested under the conditions of super-resolution multiples of 5 (5min reconstruction 1min), 10 (10min reconstruction 1min), and 15 (15min reconstruction 1min). The test results are shown in Table 5.

[0102] Judging from the results, with the increase of super-resolution multiples, all load data reconstruction methods show a deterioration trend in numerical deviation, data distribution and data quality indicators. This shows that using the same low-resolution beam, as the super-resolution multiple increases, the data reconstruction performance of the load data reconstruction method gradually decreases. The reason is that the information contained in the low-resolution data is limited. The larger the super-resolution multiple is set, the less high-resolution detail information corresponding to the low-resolution load data, which makes the reconstruction effect of the high-resolution data worse. In terms of algorithm robustness, the performance of the four types of comparison algorithms decreases significantly with the increase of super-resolution multiples. Even when the super-resolution multiple is 15, the RMSE index of the proposed method still does not exceed 100, and D H The index does not exceed 0.05. In contrast, the RMSE index of other comparison methods exceeds 130 and D H The index exceeds 0.1. In essence, the proposed method reconstructs the Gaussian distribution into the original data distribution under the control of low-resolution data, that is, the generated reconstructed data is closer to the actual situation. Therefore, even when too much high-frequency information is lost, the reconstruction effect is still better than other methods, and it has good stability and reliability.

[0103] Table 5

[0104]

[0105] Through non-invasive load decomposition experiments, the actual performance of the reconstructed data obtained by different super-resolution methods in the decomposition task is analyzed. In the application of load super-resolution methods, the total power consumption of six load devices such as DWE, HPE and FGE is used as the data reconstruction target, and this is used as the input of non-invasive load decomposition. The load decomposition algorithm uses the three most advanced algorithms in the NILM field, including WindowsGRU, RNN and Seq2Seq. In the experiment, the total load data reconstructed by the super-resolution method is used for load decomposition. The active power in the time period from 2012 / 12 / 01 to 2012 / 12 / 14 is selected as the training data, and the test time period is from 2012 / 12 / 15 to 2012 / 12 / 30. The super-resolution multiple is set to 5, such as Figure 5 As shown, the SR Model is a super-resolution model containing a conditional noise prediction network.

[0106] In order to intuitively quantify the refined load sub-item measurement error after load decomposition, the relative decomposition error P is used. error As an indicator to evaluate the impact of load decomposition. First, to intuitively illustrate the impact of high-resolution data on the accuracy of load decomposition, Figure 6The relative error results of energy consumption for load decomposition using low-resolution power data and original high-resolution power data on three different NILM models are shown. The results show that the relative error of energy consumption for load decomposition directly based on low-resolution power data is more than 60%, while the relative error of load decomposition results based on high-resolution power data is relatively small, and even on the RNN method with poor algorithm performance, there is a 25.4% improvement. Experiments show that the temporal resolution of power data has a great influence on the load decomposition results. Therefore, it is necessary to use super-resolution methods to reconstruct high-resolution power data when performing high-precision load decomposition.

[0107] The impact of super-resolution methods on load decomposition results will be analyzed below. Figure 7 The total energy consumption of refined loads using various super-resolution methods under different load decomposition algorithms is shown. Table 6 shows the relative error results of direct load decomposition compared to the original high-resolution data. Figure 7 It can be seen that the effects of various super-resolution methods on different load decomposition methods are different, and the effects on different load decomposition results are quite different. WindowsGRU, RNN and Seq2Seq are all end-to-end learning methods, and data covering more power details are more helpful in mapping from total power data to each load power. From Table 6, the DDPM method improves the performance of WindowsGRU with the best load decomposition performance by 30.6% compared with the second-best method. The reason is that the DDPM method can better restore high-frequency details. Therefore, the load decomposition based on DDPM can obtain the result closest to the decomposition based on real data. Compared with other super-resolution methods, the load decomposition based on DDPM's reconstructed data is closest to the result of directly using high-resolution data in all algorithms, which shows that it is of great help to improve the performance of load decomposition tasks and is more universal than other methods.

[0108] Table 6

[0109]

[0110]

[0111] The above analysis shows the data reconstruction effect of this method on different types of loads and the performance differences under different super-resolution multiples. The results show that this method has better data reconstruction effect than data interpolation, convolutional neural network, generative adversarial network and other methods; in addition, load decomposition based on the reconstructed data obtained by this method can obtain more accurate and refined energy consumption data, which is more conducive to energy consumption management.

[0112] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A non-invasive load decomposition method based on super-resolution, characterized in that: include: Obtain low-resolution power series data of load equipment; Input the low-resolution power sequence data into a conditional noise prediction network CoPNet to obtain predicted noise, denoise the initial pure noise high-resolution data sequence according to the predicted noise, and obtain reconstructed power sequence data, wherein the conditional noise prediction network CoPNet is constructed based on a denoising diffusion probability model DDPM and is obtained through training with a load sample data set, and the load sample data set includes low-resolution power sequence training samples; The reconstructed power sequence data is decomposed to obtain power data of a single load device.

2. The non-intrusive load decomposition method based on super-resolution according to claim 1, characterized in that: Training the conditional noise prediction network CoPNet by using the load sample data set includes: The original high-resolution power sequence data is mapped to a Gaussian distribution by adding Gaussian noise forward to obtain pure noise sequence samples; Randomly sampling the pure noise sequence samples to obtain noisy high-resolution sequence training data; The low-resolution power sequence training samples, the noisy high-resolution sequence training data and the real noise are used to train the conditional noise prediction network CoPNet.

3. The non-invasive load decomposition method based on super-resolution according to claim 2, characterized in that: De-noising the initial pure noise high-resolution data sequence according to the predicted noise comprises: The pure noise sequence samples are randomly sampled to obtain the initial pure noise high-resolution data sequence, and the initial pure noise high-resolution data sequence is gradually denoised according to the predicted noise to obtain the reconstructed power sequence data.

4. The non-invasive load decomposition method based on super-resolution according to claim 2, characterized in that: The conditional noise prediction network CoPNet includes a data input module, a feature extraction module and a noise output module; The data input module is used to upsample the low-resolution power sequence data and splice them in the channel dimension, output splicing features, and encode the time using position coding to obtain the position coding of the time; The feature extraction module is used to extract feature information from the splicing feature and the time position code; The noise output module is used to integrate the feature information through a convolution layer and a Reshape layer to output predicted noise.

5. The non-invasive load decomposition method based on super-resolution according to claim 4, characterized in that: The step of encoding the splicing feature and the time position to extract feature information includes: The preliminary feature information of the splicing feature is extracted through a convolutional layer, and the preliminary feature information and the temporal position code are input into a residual connection structure to extract deep feature information.

6. The non-invasive load decomposition method based on super-resolution according to claim 5, characterized in that: The method for extracting deep feature information using the residual connection structure is: x i+1  i +Switch(Conv1d{[Leaky ReLu(Conv1d(x))+Leaky ReLu(FC((Pos(t)))]}) Among them, x i+1 is the data input of the i+1th residual block, POS(t) is the position encoding of time step t, Conv1d is the one-dimensional convolutional layer, Leaky Relu is the activation function, and FC is the fully connected layer.

7. The non-intrusive load decomposition method based on super-resolution according to claim 2, characterized in that: The objective function of the conditional noise prediction network CoPNet is: Among them, L(θ) is the objective function, is the original high-resolution power sequence data, x L represents low-resolution power series data, ∈ is Gaussian noise, ∈θ To predict the noise, is the noisy high-resolution sequence data at any time step.

Images