A non-intrusive load decomposition method and device based on transfer learning and a medium

Abstract

The application discloses a non-invasive load decomposition method and device based on transfer learning and a medium, and relates to the field of load decomposition; the method comprises the following steps: acquiring total power consumption data of a target power system; determining a TimesNet network model; the TimesNet network model is stacked by multiple TimesBlock layers; the TimesNet network model is obtained by training according to total power consumption data of a known source domain and target domain, and based on the minimum of a weighted and overall loss function; the total power consumption data of the target power system is input into the TimesNet network model, time series data processing is performed by using a fast Fourier transform method, mapping and correlation processing of migration is performed according to the processed total power consumption data of the target power system, and then the power of equipment is determined; and the application can improve the load decomposition precision.

Description

Technical Field

[0001] This invention relates to the field of load decomposition, and in particular to a non-invasive load decomposition method, apparatus and medium based on transfer learning. Background Technology

[0002] Given the current context of dual-carbon and highly uncertain wind and solar power integration into the power system, the construction of new power systems has significant practical implications. Load decomposition technology can provide information on the operating status of various user-side equipment, user electricity consumption habits, and the dispatchability of various flexible loads, which is crucial for the construction of new power systems. Through detailed user-side electricity consumption information and technologies such as demand response and user-side management, it can provide strong support for the power system, such as providing peak shaving and valley filling, and new energy consumption services, which is conducive to the safe, stable, efficient, and economical operation of the power grid.

[0003] Currently, methods for obtaining detailed electricity consumption data on the user side include traditional invasive load monitoring, which involves installing monitoring devices on all target devices of the user to obtain detailed electricity consumption information. However, this method is costly, complex to install, and raises issues related to user privacy. Non-invasive load decomposition, on the other hand, can obtain detailed electricity consumption data for each device of the user simply from the user's total electricity consumption data. It is low-cost and easy to install. The obtained load data can provide a basis for power system dispatching and can also help achieve energy conservation and emission reduction by providing electricity consumption guidance to users.

[0004] However, in existing technologies, similar solutions mainly utilize deep learning methods. By training models on large amounts of data, they establish a connection between total load data and data from individual devices, achieving relatively accurate load decomposition. This typically models the load decomposition problem as a regression model, with the neural network's input and output being the building's total electricity consumption data and the electricity consumption data of a single target device, respectively. Training the neural network model requires a large amount of data—that is, data on the building's total electricity consumption and the target device's consumption over a period of time. However, a problem with this approach is the difficulty in obtaining training data. Generally, obtaining total building electricity consumption data is relatively easy, but obtaining the target device's electricity consumption data is more difficult, requiring the installation of smart meters on individual devices, thus making large-scale data collection challenging.

[0005] Conventional deep learning methods require a large amount of data for training. Current technologies do not make full use of existing data, and the neural network models they employ are poorly adapted to time-series data. Therefore, improving the accuracy of load decomposition is crucial. Summary of the Invention

[0006] The purpose of this invention is to provide a non-invasive load decomposition method, apparatus, and medium based on transfer learning, which can improve the accuracy of load decomposition.

[0007] To achieve the above objectives, the present invention provides the following solution: a non-intrusive load decomposition method based on transfer learning.

[0008] The method includes: acquiring total electricity consumption data of the target power system; determining a TimesNet network model; the TimesNet network model is composed of multiple stacked TimesBlock layers; the TimesNet network model is trained based on the total electricity consumption data of known source domain data and target domain data, with the goal of minimizing the weighted sum and overall loss function; the source domain data is the total electricity consumption data corresponding to a known target area or target device; the target domain data is the total electricity consumption data corresponding to an unknown target area or target device.

[0009] The total electricity consumption data of the target power system is input into the TimesNet network model, and the time series data is processed using the Fast Fourier Transform method. Then, the migration mapping association processing is performed based on the processed total electricity consumption data of the target power system to determine the equipment power. The equipment power is the electricity consumption data of each sub-region or sub-equipment corresponding to the target power system.

[0010] Optionally, the method for determining the TimesNet network model specifically includes: acquiring training data; the training data includes: total electricity consumption data of known source domain data and target domain data; constructing a TimesNet network; the TimesNet network includes multiple TimesBlock layers; inputting the source domain data and target domain data from the training data into the TimesBlock layers, and performing time series data processing using the Fast Fourier Transform method to obtain processed time series data; the processed time series data includes: source domain time series data and target domain time series data.

[0011] The mean squared error method is used to determine the regression loss function based on the source domain time series data; the kernel two-sample detection method is used to perform mapping and correlation processing on the target domain time series data and the source domain time series data to obtain the domain adaptive loss function; the regression loss function and the domain adaptive loss function are weighted and summed to obtain the weighted sum overall loss function.

[0012] Based on the target domain time series data and the source domain time series data, with the goal of minimizing the weighted sum and overall loss function, the parameters of the TimesNet network are adjusted and trained under a set number of training rounds to obtain the trained TimesNet network; the parameters include: weights and learning rate; the trained TimesNet network is determined as the TimesNet network model.

[0013] Optionally, the time series data is processed using the Fast Fourier Transform (FFT) method to obtain the processed time series data. Specifically, this includes: using the FFT method to obtain periodic components from the training data according to the length of each period to obtain harmonic information data; the harmonic information data includes amplitude and frequency.

[0014] For any period length, the harmonic information data is subjected to dimensionality transformation to obtain two-dimensional data; two-dimensional convolution is used to extract features from each of the two-dimensional data, and adaptive aggregation is performed to obtain multiple aggregated data; all the aggregated data are weighted and summed to obtain the processed time series data.

[0015] Optionally, the expression for the regression loss function is:

[0016]

[0017] Among them, L R Here, n is the regression loss function; n is the number of source domain time series data. This represents the output data of the source domain time series data in the TimesNet network; y i This refers to the data corresponding to the known target region or target device in the source domain time series data.

[0018] Optionally, the expression for the domain adaptive loss function is:

[0019]

[0020] Among them, L D For the domain adaptive loss function; MMD(X) s X t X is the source domain time series data. s and target domain time series data X t The unbiased estimator corresponding to the two-sample nuclear test; n s φ is the number of source domain time series data; i is the index of the source domain time series data; φ(·) represents the direction of the Hillbert space to the regeneration kernel. Mapping association; n t is the number of time series data in the target domain; j is the sequence number of the time series data in the target domain. For the i-th source domain time series data To the Hillbert space of regenerated nuclei Mapping associations; For the j-th target domain time series data The mapping association to the regenerating kernel Hillbert space H; Here is the mapping formula for the regenerated kernel Hillbert space H.

[0021] Optionally, the expression for the weighted summation overall loss function is:

[0022]

[0023] Where L is the weighted sum overall loss function; L R L is the regression loss function; D X is the domain adaptive loss function; s Source domain time series data; X t The target domain time series data; Θ is the parameter; μ is the weight corresponding to the regression loss function; λ is the weight corresponding to the domain adaptive loss function; Z s This refers to the output data of the TimesBlock layer for the source domain time-series data; Z t This is the output data of the TimeBlock layer for the target domain time series data.

[0024] Optionally, for any period length, the harmonic information data is subjected to dimensional transformation processing to obtain two-dimensional data. Specifically, this includes: for any period length, the harmonic information data is folded and arranged to obtain harmonic information processed data; and the harmonic information processed data is subjected to sequence zero-padding processing to obtain two-dimensional data.

[0025] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the steps of the above-described non-intrusive load decomposition method based on transfer learning.

[0026] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described non-intrusive load decomposition method based on transfer learning.

[0027] According to specific embodiments provided by the present invention, the following technical effects are disclosed.

[0028] This invention discloses a non-intrusive load decomposition method, apparatus, and medium based on transfer learning. The method involves acquiring total electricity consumption data of a target power system; determining a TimesNet network model; the TimesNet network model being composed of multiple stacked TimesBlock layers; and training the TimesNet network model based on the total electricity consumption data of known source and target domain data, with the goal of minimizing the weighted sum and overall loss function. The total electricity consumption data of the target power system is input into the TimesNet network model, and the Fast Fourier Transform (FFT) method is used for time series data processing. Transfer mapping and correlation processing are then performed based on the processed total electricity consumption data of the target power system to determine equipment power. TimesBlock is used to capture correlation relationships, and the TimesNet network model is transferred, decomposing the time series data into data with different frequency periods using FFT, facilitating the model's capture of time series data features and improving load decomposition accuracy. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the non-intrusive load decomposition method based on transfer learning provided in Embodiment 1 of the present invention.

[0031] Figure 2 This is a schematic diagram illustrating the overall principle of an embodiment of the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the capture relationship between intra-cycle and inter-week data provided in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the folding and arranging process provided in an embodiment of the present invention.

[0034] Figure 5 This is a schematic diagram of the TimesBlock layer provided in an embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram of the transfer learning process provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0037] Different households or electricity users have certain similarities in their electrical equipment and electricity consumption data, and existing data can be used to train non-intrusive load decomposition models for other users.

[0038] To address the issue of insufficient data during model training in existing technologies, this invention proposes a non-intrusive load decomposition method based on transfer learning. This method utilizes existing user training data combined with the total electricity consumption data of new users to train the model, resulting in a model that can obtain detailed electricity consumption data for new users. To achieve higher decomposition accuracy, this invention employs the TimesNet model, which decomposes time-series data into data with different frequency periods using Fast Fourier Transform, facilitating the model's capture of the characteristics of the time-series data.

[0039] The purpose of this invention is to provide a non-invasive load decomposition method, apparatus, and medium based on transfer learning, which aims to improve the accuracy of load decomposition.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1: As Figure 1 As shown in this embodiment, a non-intrusive load decomposition method based on transfer learning is presented.

[0042] The method includes: Step 100: Obtaining the total power consumption data of the target power system.

[0043] Step 200: Determine the TimesNet network model. The TimesNet network model is composed of multiple stacked TimesBlock layers. The TimesNet network model is trained based on the total power consumption data of known source domain data and target domain data, with the goal of minimizing the weighted sum and overall loss function. The source domain data is the total power consumption data of the known target area or target device. The target domain data is the total power consumption data of the unknown target area or target device.

[0044] Step 300: Input the total electricity consumption data of the target power system into the TimesNet network model, perform time series data processing using the Fast Fourier Transform method, and perform migration mapping and association processing based on the processed total electricity consumption data of the target power system to determine the equipment power. The equipment power is the electricity consumption data of each sub-region or sub-device corresponding to the target power system.

[0045] In one embodiment, the method for determining the TimesNet network model specifically includes: acquiring training data; the training data includes: total power consumption data of known source domain data and target domain data.

[0046] Construct a TimesNet network; the TimesNet network consists of multiple TimesBlock layers.

[0047] The source domain data and target domain data in the training data are input into the TimesBlock layer, and the time series data are processed using the Fast Fourier Transform method to obtain the processed time series data. The processed time series data includes: source domain time series data and target domain time series data.

[0048] Specifically, the Fast Fourier Transform (FFT) method is used to process time series data to obtain processed time series data. This includes: using the FFT method to obtain periodic components from the training data according to the length of each period to obtain harmonic information data; the harmonic information data includes amplitude and frequency.

[0049] For any period length, the harmonic information data is transformed to obtain two-dimensional data.

[0050] Specifically, for any period length, the harmonic information data is subjected to dimensional transformation processing to obtain two-dimensional data. This includes: folding and arranging the harmonic information data for any period length to obtain processed harmonic information data; and padding the processed harmonic information data with zeros to obtain two-dimensional data.

[0051] Two-dimensional convolution is used to extract features from each two-dimensional data point, and adaptive aggregation is performed to obtain multiple aggregated data points.

[0052] We perform a weighted summation on all the aggregated data to obtain the processed time series data.

[0053] The mean squared error method is used to determine the regression loss function based on the source domain time series data; the kernel two-sample detection method is used to perform mapping and correlation processing on the target domain time series data and the source domain time series data to obtain the domain adaptive loss function.

[0054] Based on the regression loss function and the domain adaptive loss function, a weighted sum is performed to obtain the weighted overall loss function.

[0055] Based on the time series data of the target domain and the time series data of the source domain, with the goal of minimizing the weighted sum and the overall loss function, the parameters of the TimesNet network are adjusted and trained under a set number of training epochs to obtain the trained TimesNet network; the parameters include: weights and learning rate.

[0056] The trained TimesNet network is identified as the TimesNet network model.

[0057] Specifically, the expression for the regression loss function is shown below.

[0058]

[0059] Among them, L R Here, n is the regression loss function; n is the number of source domain time series data. This represents the output data of the source domain time series data in the TimesNet network; y i This refers to the data corresponding to the known target region or target device in the source domain time series data.

[0060] The expression for the domain adaptive loss function is shown below.

[0061]

[0062] Among them, L D For the domain adaptive loss function; MMD(X) s M t X is the source domain time series data. s and target domain time series data X t The unbiased estimator corresponding to the two-sample nuclear test; n s φ is the number of source domain time series data; i is the index of the source domain time series data; φ(·) represents the direction of the Hillbert space to the regeneration kernel. Mapping association; n t is the number of time series data in the target domain; j is the sequence number of the time series data in the target domain. For the i-th source domain time series data To the Hillbert space of regenerated nuclei Mapping associations; For the j-th target domain time series data To the Hillbert space of regenerated nuclei Mapping associations; Hillbert space for regenerating nuclei The mapping formula.

[0063] The expression for the weighted summation loss function is as follows.

[0064]

[0065] Where L is the weighted sum overall loss function; L R L is the regression loss function; D X is the domain adaptive loss function; s Source domain time series data; X t The target domain time series data; Θ is the parameter; μ is the weight corresponding to the regression loss function; λ is the weight corresponding to the domain adaptive loss function; Z s This refers to the output data of the TimesBlock layer for the source domain time-series data; Z t This is the output data of the TimeBlock layer for the target domain time series data.

[0066] Specifically, the electricity consumption data of a target area or target device is identified from the total electricity consumption data of a certain building or area. The total electricity consumption data is generally the meter reading at a user's household. If there are m devices under a certain meter, and i is the serial number, then the operating power corresponding to the i-th device is P. i If (t), then the expression for the total power P(t) is as follows.

[0067]

[0068] Where σ(t) represents the measurement noise. Non-intrusive load decomposition aims to identify P from the total load, i.e., the total power P(t). i (t). Furthermore, the formula for total power can also be expressed as follows.

[0069]

[0070] Where P i Let s be the rated power of the i-th device. i (t) represents the on / off state of the device at time t.

[0071] The overall principle is as follows Figure 2 As shown, House1 contains detailed total power consumption data and power consumption data for each individual device, which is called source domain data. House2 only contains total power consumption data, which is called target domain data.

[0072] In reality, time-series data is generally influenced by various periodic factors, and these factors often interact with each other, making the changes in time-series data highly complex and difficult to accurately model using conventional neural networks. Load electricity consumption data is a typical type of time-series data, influenced by temporal factors such as season, date, and time of day, as well as non-temporal factors such as different user types and user electricity consumption habits. While the power fluctuation at a single node may be small, the load fluctuation over time can be significant in smaller areas. Therefore, TimesNet networks can be used for analysis.

[0073] Traditional time series analysis methods primarily focus on the continuity and recursion of time series, neglecting their potential multiple periodicities. In contrast, the TimesNet network treats individual points in a time series as relatively independent. For data at a given point in time, it not only focuses on data before and after that point but also on data across different periodic components at the same point in time. Therefore, it can simultaneously capture the relationships between data within a period and between periods, such as... Figure 3 As shown.

[0074] To obtain the multi-period components in one-dimensional time series data, the one-dimensional data can be processed using a Fast Fourier Transform (FFT) to obtain the processed one-dimensional data X. 1D This yields the period and amplitude information of each component. Its general expression can be represented as: X 1D ∈R T×d .

[0075] Where T represents the period length of each component, and d represents the period dimension of the data. In a general sense, the information for each period can be calculated using the following formula.

[0076] A = Avg(Amp(FFT(X)) 1D )))

[0077]

[0078]

[0079] Where A∈R T R represents the amplitude of each component in the time series data. T This represents a sequence with a period length of T. The transformation here is to split a one-dimensional data into several short sequences with a period length of T. k is also a sequence number, here representing the sequence number of the harmonic. {f1, ..., f k} represents the frequency of the component with the largest amplitude, and the corresponding number of periods is {p1, ..., p2}. k The process of acquiring multi-period information from time-series data can also be represented as follows.

[0080] A,{f1,…,f k},{p1,…,p k}=MultiP(X 1D ).

[0081] In order to capture the relationships during the week, such as Figure 4 As shown, the data from multiple periods is folded and arranged using the following formula.

[0082]

[0083] Padding involves adding zeros to the end of the sequence to ensure that the sequence data can be appended to the period length p, which has already been folded and arranged. i Divisibility. After performing the above operation on each periodic component, a two-dimensional tensor can be obtained. Two-dimensional data

[0084] Figure 5 This is a schematic diagram of the structure of the TimesBlock layer, the basic unit of the TimesNet network model. The entire model is composed of multiple stacked TimesBlock layers. The process of capturing the periodic and intra-period relationships of time series data through the TimesBlock layer consists of the following steps.

[0085] FFT and Reshape.

[0086] In this step, harmonic information, including amplitude, frequency, and period length, is obtained from the input data using FFT. The one-dimensional time series data is then converted into two-dimensional data through padding and reshaping operations. The formula for this process is as follows: l represents the number of TimesBlock layers.

[0087]

[0088]

[0089] Two-dimensional feature extraction from one-dimensional time series data.

[0090] The two-dimensional data obtained in step 1 This paper uses two-dimensional convolution to extract features, and implements this two-dimensional convolution using the Inception module. Inception is a module name in the field of deep learning that uses convolutional kernels of different sizes to reduce the number of parameters in a neural network and greatly improve computational efficiency. The determination of a two-dimensional convolution is shown in the following equation.

[0091]

[0092] Reshape back and adaptive aggregation.

[0093] After extracting the features from the two-dimensional data, the two-dimensional data is transformed back into one-dimensional time-series data. Its process is the reverse of the process described above.

[0094]

[0095] The input data has the same dimension, ensuring that the input and output have the same shape. Dep() means removing the values ​​that are padded by padding the initial one-dimensional data.

[0096] Finally, the outputs of the Inception module for each periodic component are weighted and summed to obtain the final output.

[0097]

[0098]

[0099] After performing the above operations on the total power P(t), assuming the source domain dataset... Labeled data, target domain data Unlabeled. Source domain data. As input to the model, These are labels for the source domain data, representing the power data of the target device. The target domain data is also used as input to the model, with the aim of learning as much knowledge as possible from the source domain data. Let X be the feature space and label space of the source domain. s and Y s The feature space of the target domain is s t The marginal probability distribution of a feature space is typically defined as the marginal probability distribution of that space. In practical applications, the marginal probability distributions of the source and target domains differ, which can lead to poor performance of a model trained in the source domain in the target domain, resulting in poor generalization ability and reduced transferability. This invention proposes a method that measures the gap between the source and target domains using a loss function, comprehensively considering both labeled and unlabeled data, to achieve the transfer of knowledge from the source domain to the target domain.

[0100] 1) Regression loss.

[0101] In domain adaptation, the source domain labels are known, and the power consumption of the target device in the source domain is also known. A regression model is used, and the difference L between the model output of the source domain data and the actual labels is calculated. R The mean squared error (MSELoss) is used as a metric.

[0102]

[0103] Among them, L R Here, n is the regression loss function; n is the number of source domain time series data. This represents the output data of the source domain time series data in the TimesNet network; y i This refers to the data corresponding to the known target region or target device in the source domain time series data.

[0104] 2) Domain adaptive loss.

[0105] In cross-domain learning, the key is to enable the target domain to learn as much knowledge as possible from the source domain. However, the data distributions of the source and target domains often differ. Therefore, to achieve the goal of transfer learning, it is necessary to reduce the data distribution differences between the domains. Currently, the main metrics for measuring the data distribution differences between two domains are MMD and deepCORALLoss. This paper chooses MMD as the domain adaptive loss. MMD is a kernel two-sample test that rejects or accepts the null hypothesis p=q based on the observed samples. The basic idea behind MMD is that if the generated distributions are the same, then all statistics are the same. MMD can be represented as follows.

[0106]

[0107] in, Let the reproducing kernel Hillbert space (RKHS) have a characteristic kernel k and source domain time series data X. s and target domain time series data X t The distributions are p and q, respectively, E p and E q Each represents the expectation placed upon them.

[0108] The kernel k is represented as follows.

[0109] k(x s x t )=(φ(x s ), φ(x t )).

[0110] That is, the vector inner product of the mappings of the source domain data and the target domain data in RKHS. Therefore, it is true if and only if the difference in the data distributions of the two domains is 0, that is, p = q. It is 0.

[0111] In practical applications, the empirical formula for calculating MMD for the source domain dataset Ds and the target domain dataset Dt is shown below.

[0112]

[0113] Among them, L D For the domain adaptive loss function; MMD(X) s X t X is the source domain time series data. s and target domain time series data X t The unbiased estimator corresponding to the two-sample nuclear test; n s φ is the number of source domain time series data; i is the index of the source domain time series data; φ(·) represents the direction of the Hillbert space to the regeneration kernel. Mapping association; n t is the number of time series data in the target domain; j is the sequence number of the time series data in the target domain. For the i-th source domain time series data To the Hillbert space of regenerated nuclei Mapping associations; For the j-th target domain time series data To the Hillbert space of regenerated nuclei Mapping associations; Hillbert space for regenerating nuclei The mapping formula.

[0114] To better evaluate the model's performance in load decomposition, two loss functions are combined using source domain time series data X. s and its label Y s To calculate the regression loss, the TimeNet layer is used to process the Z-axis of the source and target domain data. s and Z t The domain adaptive loss is calculated, and the weighted sum of the two is used as the overall loss to supervise model training, as shown in the following equation.

[0115]

[0116] Where L is the weighted sum overall loss function; L R L is the regression loss function; D X is the domain adaptive loss function; s Source domain time series data; X t The target domain time series data; Θ is the parameter; μ is the weight corresponding to the regression loss function; λ is the weight corresponding to the domain adaptive loss function; Z s This refers to the output data of the TimesBlock layer for the source domain time-series data; Z t This is the output data of the TimeBlock layer for the target domain time series data.

[0117] μ represents the influence of the regression loss on the model. A larger μ indicates a greater impact from the source domain data and labels, but excessively large μ can easily lead to overfitting. Conversely, a small μ may prevent the model from learning enough from the source domain, resulting in insufficient model performance. λ represents the influence of the domain adaptation loss on the model. A higher λ indicates a greater influence from the target domain data, but it may reduce regression accuracy. Too small a λ will result in insufficient learning. Reflected in the load decomposition model, both excessively large and excessively small weightings of the regression loss will lead to insufficient decomposition accuracy.

[0118] To address the aforementioned issues, a method for dynamically adjusting loss weights and the learning rate is proposed. The knowledge learned by the model on source domain data may not be applicable to the target domain; for example, the power of the target device obtained from decomposition is often smaller than the true value. Generally, a larger learning rate and a larger weight assigned to the regression loss should be used in the early stages of model training to enable the model to converge quickly to near the optimal value and obtain the basic power consumption characteristics of the target device, such as its operating hours. In the later stages of model training, a smaller learning rate and a larger weight assigned to the domain adaptive loss should be used to adjust the model's performance in the target domain, especially the device's operating power, to achieve the best possible results.

[0119] The specific weights of the loss function are shown in the following formula:

[0120]

[0121] The learning rate is dynamically adjusted as shown in the following formula:

[0122] Lr=L initial ×γ [epoch / 5] .

[0123] In the above formula, L initial γ represents the learning rate at the beginning of model training, γ represents the decay coefficient of the learning rate, and [·] is for rounding down.

[0124] In the above formula, epoch represents the number of training iterations of the model, and ρ represents the number of weight transformation iterations.

[0125] Non-intrusive load decomposition was achieved based on transfer learning, solving the data dependency problem of conventional deep learning methods. Using the TimesNet network model, it is more suitable for time-series data such as electricity data, improving the accuracy of the decomposition.

[0126] This invention proposes a non-intrusive load decomposition method based on transfer learning. Compared to existing conventional deep learning methods, this method fully utilizes existing data, using data from existing scenarios to help train a load decomposition model for new scenarios. It can achieve high-accuracy load decomposition even when the power consumption data of the target equipment in the new scenario is unknown. See [link / reference] Figure 6The TimesNet network model was used, which improved the decomposition accuracy compared to existing models. Figure 6 The triangles in the image are different colors to represent that they come from different datasets, and the different shapes are only for differentiation.

[0127] Example 2: A computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the non-intrusive load decomposition method based on transfer learning in Example 1.

[0128] Example 3: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the non-intrusive load decomposition method based on transfer learning in Example 1.

[0129] Example 4: A computer device, which may be a database. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores pending transactions. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements the non-intrusive load decomposition method based on transfer learning in Example 1.

[0130] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0131] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A non-intrusive load decomposition method based on transfer learning, characterized in that, The method includes: Obtain the total electricity consumption data of the target power system; A TimesNet network model is determined; the TimesNet network model is composed of multiple stacked TimesBlock layers; the TimesNet network model is trained based on the total electricity consumption data of known source domain data and target domain data, with the goal of minimizing the weighted sum and overall loss function; the source domain data is the total electricity consumption data corresponding to the known target area or target device; the target domain data is the total electricity consumption data corresponding to the unknown target area or target device. The total electricity consumption data of the target power system is input into the TimesNet network model, and the time series data is processed using the Fast Fourier Transform method. Then, the migration mapping and association processing is performed based on the processed total electricity consumption data of the target power system to determine the equipment power. The equipment power is the electricity consumption data of each sub-region or sub-equipment corresponding to the target power system. The method for determining the TimesNet network model specifically includes: Acquire training data; the training data includes: total electricity consumption data of known source domain data and target domain data; Construct a TimesNet network; the TimesNet network includes multiple TimesBlock layers; The source domain data and target domain data in the training data are input into the TimesBlock layer, and the time series data are processed using the Fast Fourier Transform method to obtain the processed time series data; the processed time series data includes: source domain time series data and target domain time series data; The mean squared error method is used to determine the regression loss function based on the source domain time series data; A kernel two-sample detection method is adopted, and a mapping and association process is performed on the target domain time series data and the source domain time series data to obtain a domain adaptive loss function; Based on the regression loss function and the domain adaptive loss function, a weighted sum is performed to obtain the weighted overall loss function; Based on the target domain time series data and the source domain time series data, with the goal of minimizing the weighted sum and overall loss function, the parameters of the TimesNet network are adjusted and trained under a set number of training epochs to obtain the trained TimesNet network; the parameters include: weights and learning rate; The trained TimesNet network is identified as the TimesNet network model; The expression for the domain adaptive loss function is: ； in, For domain-adaptive loss function; Source domain time series data and target domain time series data The unbiased estimator corresponding to the two-sample nuclear test; The number of time-series data in the source domain; The sequence number of the source domain time series data; To the Hillbert space of regenerated nuclei Mapping associations; The number of time-series data in the target domain; The sequence number of the target domain time series data; For the first Time series data from each source domain To the Hillbert space of regenerated nuclei Mapping associations; For the first Time series data of each target domain To the Hillbert space of regenerated nuclei Mapping associations; Hillbert space for regenerating nuclei The mapping formula.

2. The non-intrusive load decomposition method based on transfer learning according to claim 1, characterized in that, The Fast Fourier Transform (FFT) method is used to process time series data, resulting in processed time series data, specifically including: The training data is processed using a Fast Fourier Transform (FFT) method to obtain periodic components according to each period length, thereby obtaining harmonic information data; the harmonic information data includes amplitude and frequency. For any period length, the harmonic information data is subjected to dimensional transformation processing to obtain two-dimensional data; Two-dimensional convolution is used to extract features from each of the two-dimensional data, and adaptive aggregation processing is performed to obtain multiple aggregated data. The aggregated data is weighted and summed to obtain the processed time-series data.

3. The non-intrusive load decomposition method based on transfer learning according to claim 1, characterized in that, The expression for the regression loss function is: ； in, For regression loss function; The number of time-series data in the source domain; This represents the output data of the source domain time series data in the TimesNet network. This refers to the data corresponding to the known target region or target device in the source domain time series data.

4. The non-intrusive load decomposition method based on transfer learning according to claim 1, characterized in that, The expression for the weighted summation loss function is: ； in, For the weighted sum overall loss function; For regression loss function; For domain-adaptive loss function; This is source domain time series data; For target domain time series data; For parameters; These are the weights corresponding to the regression loss function; The weights corresponding to the domain adaptive loss function; This is the output data of the TimesBlock layer for the source domain time series data; This is the output data of the TimeBlock layer for the target domain time series data.

5. The non-intrusive load decomposition method based on transfer learning according to claim 2, characterized in that, For any period length, the harmonic information data undergoes a dimensionality transformation to obtain two-dimensional data, specifically including: For any period length, the harmonic information data is folded and arranged to obtain harmonic information processing data; The harmonic information processing data is subjected to zero-padding after the sequence to obtain two-dimensional data.

6. A computer device, comprising: The memory and processor contain a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the non-intrusive load decomposition method based on transfer learning as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the non-intrusive load decomposition method based on transfer learning as described in any one of claims 1-5.

Images