Non-intrusive load monitoring method and device based on multi-chain decomposition method

In non-invasive load monitoring through multi-chain decomposition method, power consumption prediction is used to predict multiple decomposition chains and sub-models, which solves the problem of inaccurate power consumption prediction in the prior art and improves the accuracy and efficiency of monitoring.

CN113033076BActive Publication Date: 2024-11-19LUMLUX CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110229250.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-03-02
Publication Date
2024-11-19
Estimated Expiration
2041-03-02

AI Technical Summary

Technical Problem

The existing non-invasive load monitoring methods cannot accurately predict the power consumption of each power-consuming device, and there are problems such as high data acquisition cost and large errors.

Method used

Using a non-invasive load monitoring method based on multi-chain decomposition method, by obtaining the power consumption data in the bus line, inputting it to the power consumption monitoring model, using multiple decomposition chains and sub-models for prediction, each decomposition chain corresponds to one The power consumption equipment to be monitored is connected to output the power consumption prediction value to improve the overall decomposition performance.

Benefits of technology

It improves the accuracy of power consumption prediction of power consumption of power consumption equipment, reduces data acquisition costs, and reduces errors, and makes independent models connect to improve overall performance through multi-chain decomposition method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113033076B_ABST
    Figure CN113033076B_ABST
Patent Text Reader

Abstract

The present application relates to a non-intrusive load monitoring method and device based on a multi-chain decomposition method, belonging to the field of electronic technology. The method comprises: inputting power consumption data in a total line into a total power consumption monitoring model to obtain power consumption prediction values ​​of various power consuming devices; the total power consumption monitoring model comprises a plurality of decomposition chains, and different power consuming devices correspond to different decomposition chains; each decomposition chain comprises one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power consuming device, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power consuming device corresponding to the decomposition chain, so as to output the power consumption prediction value of the power consuming device corresponding to the decomposition chain; the problem that the power consumption prediction values ​​of various power consuming devices obtained by the existing non-intrusive load monitoring method are not accurate enough can be solved, and the overall model is constructed by using independent models as sub-models, so that the various sub-models are connected and the overall performance of the decomposition is improved.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present application relates to a non-invasive load monitoring method and device based on a multi-chain decomposition method, and belongs to the field of electronic technology. [Background technology]

[0002] Energy is the material basis for the progress of human society. Since the Second Industrial Revolution, electricity as an energy source has been gradually applied to all aspects of life, and electricity is gradually becoming an increasingly important resource. Nowadays, global electricity consumption is increasing year by year, and society is facing an increasingly serious energy shortage problem.

[0003] At the same time, with the continuous changes in the earth's climate, countries are becoming increasingly aware of the importance of environmental protection. At present, many countries have reached a consensus on reducing greenhouse gas emissions and are committed to reducing the emission of greenhouse gases such as carbon dioxide, which has also stimulated the demand for reducing energy consumption. In order to reasonably reduce the consumption of electricity, it is necessary to properly manage and optimize the use of electricity. A clear understanding of the specific use of electricity can effectively help people optimize electricity management. Analyzing household electricity consumption has gradually become a research field that attracts attention.

[0004] In terms of home energy management, load monitoring is an application that can help households save energy. The definition and purpose of load monitoring is to obtain the usage of various devices in the entire house of the resident. This helps the power company understand the composition of the user's consumption load, strengthen load-side management, and achieve the purpose of adjusting peak-to-valley differences and reducing network losses by guiding users to consume reasonably and reasonably arranging the use time of high-power non-emergency loads.

[0005] At present, load monitoring can be divided into two categories: intrusive and non-intrusive:

[0006] (1) Traditional intrusive load monitoring collects data and monitors the operation and power consumption of each load by installing sensors on each line and each load. However, a large number of hardware devices need to be installed, the system structure is complex, and the installation efficiency is low. At the same time, this solution requires entering the user's home to install data collection equipment, that is, "intrusive" load monitoring, which affects the user's living safety.

[0007] (2) Non-intrusive load monitoring (NILM), also known as load decomposition, analyzes the total meter data in a specific area to obtain relevant information about each power load in the area, such as the number of loads, load types, working status, and corresponding energy consumption. Non-intrusive load monitoring is easy to install, but there is a certain deviation between the inferred power consumption data and the actual value, and the error is large.

[0008] For non-intrusive load monitoring algorithms, they can be divided into transient signal-based and steady-state signal-based methods according to the different sampling periods of their input signals.

[0009] Transient signals are generated through high-frequency acquisition, and their acquisition frequency is usually as high as several thousand hertz to tens of thousands of hertz. The higher the acquisition frequency of the data acquisition device, the more comprehensive and accurate it can capture the power consumption characteristics and detailed information of the equipment, making it easier to infer the type of equipment and power consumption based on this information. However, there are several defects in using transient signals for load monitoring. First, in order to obtain transient signals, the sampling rate of the data acquisition device is required to be high enough, which leads to higher data acquisition costs; secondly, when the data acquisition device collects at a high frequency, a large amount of data will be generated in a short period of time, and these data need to be transmitted to the data processing module to execute the load monitoring algorithm, which puts higher requirements on data transmission and storage.

[0010] In contrast, the method of non-intrusive load monitoring based on steady-state signals has lower requirements for data acquisition equipment, and steady-state signals can be obtained through low-frequency acquisition. Low-frequency acquisition means that its acquisition frequency is less than one hertz, and the sampling interval of low-frequency acquisition ranges from a few seconds to even a few minutes. Smart meters that are now widely used among home users can fully meet such low sampling requirements. Although the non-intrusive load monitoring method based on low-frequency signals is low-cost, it cannot obtain high-frequency information of the line. At the same time, the low-frequency characteristics of different devices have certain similarities, and for devices whose usage time is less than the sampling interval, there may be situations where they cannot be captured. [Summary of the invention]

[0011] The present application provides a non-intrusive load monitoring method and device based on a multi-chain decomposition method, which can solve the problem that the power consumption prediction value of each power-consuming device obtained by the existing non-intrusive load monitoring method is not accurate enough. The present application provides the following technical solutions:

[0012] In a first aspect, a non-intrusive load monitoring method based on a multi-chain decomposition method is provided, the method comprising:

[0013] Get the power consumption data in the total circuit;

[0014] Inputting the power consumption data into the power consumption monitoring overall model to obtain the power consumption prediction value of each power consuming device;

[0015] The power consumption monitoring overall model includes a plurality of decomposition chains, each of which corresponds to a power-consuming device to be monitored, and different power-consuming devices correspond to different decomposition chains;

[0016] Each decomposition chain includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power-consuming equipment, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power-consuming equipment corresponding to the decomposition chain, so as to output the predicted value of the power consumption of the power-consuming equipment corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power-consuming equipment.

[0017] Optionally, for a decomposition chain including n sub-models, where n is an integer greater than 1, the calculation process of the power consumption prediction value output by the decomposition chain includes:

[0018] Inputting the power consumption data into the first sub-model to obtain an output result of the first sub-model;

[0019] Input the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1;

[0020] When k+1 is less than n, the value of k is equal to k+1, and the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model is performed again;

[0021] When k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.

[0022] Optionally, the method further comprises:

[0023] Using the training data set, the sub-model corresponding to each power-consuming device is trained respectively to obtain multiple trained sub-models;

[0024] For each trained sub-model, other trained sub-models are used as the preceding nodes of the trained sub-model to obtain a temporary decomposition chain; and the temporary decomposition chain is trained to obtain a trained temporary decomposition chain;

[0025] Using a test data set to test the trained sub-model and the trained temporary decomposition chain respectively;

[0026] When the test result indicates that the performance of the trained sub-model is lower than the performance of the trained temporary decomposition chain, determining the preceding node to be an associated node of the trained sub-model;

[0027] When the test result indicates that the performance of the trained sub-model is higher than the performance of the trained temporary decomposition chain, determining that the preceding node is not an associated node of the trained sub-model;

[0028] Based on each trained sub-model and the associated nodes of each trained sub-model, a graph structure is constructed; each trained sub-model in the graph structure is used as a vertex of the graph, and the weight w(j, i) of the directed edge e(j, i) connecting vertex i and vertex j is the optimization degree of the temporary decomposition chain compared with the trained sub-model, and a negative weight represents performance improvement; i represents the trained sub-model, and j represents the associated node of the trained sub-model;

[0029] Add a special vertex in the graph structure, where the special vertex has outgoing edges pointing to all other vertices and has a weight of 0;

[0030] The shortest path from the special vertex to all other vertices without loops is calculated to obtain a multi-chain decomposition model structure of the overall power consumption monitoring model.

[0031] Optionally, the method further comprises:

[0032] Generate a plurality of model structures of the overall model based on each sub-model, wherein the model structure of each overall model includes at least a plurality of decomposition chains;

[0033] Use the training data set to train each sub-model in the model structure of each overall model to obtain a trained model structure;

[0034] Use the model search set to search for the model structure with the best performance among the trained model structures to obtain the overall power consumption monitoring model;

[0035] The power consumption monitoring overall model is tested using a test data set.

[0036] Optionally, the generating of a plurality of model structures of the overall model based on the respective sub-models includes:

[0037] Setting a maximum length of each decomposition chain in the model structure, wherein the maximum length is less than or equal to the total number of the power-consuming devices to be monitored;

[0038] Model structures of multiple overall models are generated according to the maximum length, and each decomposition chain in the model structure of each overall model is less than or equal to the maximum length.

[0039] Optionally, each set of training data includes sample values ​​and labels;

[0040] For a sub-model without a preceding node, the sample value is power consumption data, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model;

[0041] For a sub-model with a preceding node, the sample value is the difference between the power consumption data and the actual power consumption value corresponding to each preceding node, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model.

[0042] Optionally, the power consumption data includes: the power consumption of each of the power-consuming devices to be monitored, the power consumption of the virtual device, and the noise value of the power consumption;

[0043] The virtual device refers to a device in the bus line that does not need to be monitored.

[0044] Optionally, the number of decomposition chains in the overall power consumption monitoring model is the same as the number of power-consuming devices to be monitored.

[0045] In a second aspect, a non-intrusive load monitoring device based on a multi-chain decomposition method is provided, the device comprising:

[0046] A total amount acquisition module is used to obtain the power consumption data in the total circuit;

[0047] A component prediction module, used for inputting the power consumption data into the power consumption monitoring overall model to obtain the power consumption prediction value of each power consuming device;

[0048] The power consumption monitoring overall model includes a plurality of decomposition chains, each of which corresponds to a power-consuming device to be monitored, and different power-consuming devices correspond to different decomposition chains;

[0049] Each decomposition chain includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power-consuming equipment, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power-consuming equipment corresponding to the decomposition chain, so as to output the predicted value of the power consumption of the power-consuming equipment corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power-consuming equipment.

[0050] Optionally, for a decomposition chain including n sub-models, where n is an integer greater than 1, the component prediction module is used to:

[0051] Inputting the power consumption data into the first sub-model to obtain an output result of the first sub-model;

[0052] Input the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1;

[0053] When k+1 is less than n, the value of k is equal to k+1, and the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model is performed again;

[0054] When k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.

[0055] The beneficial effects of the present application are: by acquiring power consumption data in the total line; inputting the power consumption data into the total power consumption monitoring model to obtain the power consumption prediction value of each power consuming device; wherein the total power consumption monitoring model includes multiple decomposition chains, each decomposition chain corresponds to a power consuming device to be monitored, and different power consuming devices correspond to different decomposition chains; for each decomposition chain, one or more sub-models connected in sequence are included; each sub-model is used to predict the power consumption of the corresponding power consuming device, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power consuming device corresponding to the decomposition chain, so as to output the power consumption prediction value of the power consuming device corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power consuming devices; the problem that the power consumption prediction values ​​of each power consuming device obtained by the existing non-invasive load monitoring method are not accurate enough can be solved, and an overall model is constructed by designing a multi-chain decomposition method to use independent models as sub-models, so that each sub-model is connected and the overall performance of the decomposition is improved.

[0056] In addition, when solving a multi-chain model, the present application also provides two simplified solutions to address the situation where the complexity of the exhaustive method for obtaining a multi-chain model with the best performance is too high: (1) by limiting the maximum length of the model decomposition chain, the total number of multi-chain decomposition models that can be generated is reduced, thereby reducing the search range when using an exhaustive method to search for a multi-chain model with the best performance; (2) based on the relationship between the sub-models that constitute the multi-chain decomposition model, the type and position relationship of each sub-model in the multi-chain model with better performance is solved, and the sub-model is trained based on this, thereby reducing some unnecessary sub-model training processes, thereby speeding up the solution process of the multi-chain decomposition model.

[0057] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present application in conjunction with the accompanying drawings.

Brief Description of the Drawings

[0058] Figure 1 It is a schematic diagram of an independent model predicting device power consumption provided by an embodiment of the present application;

[0059] Figure 2 It is a schematic diagram of a single-chain prediction device power consumption provided by an embodiment of the present application;

[0060] Figure 3 is a flow chart of a non-intrusive load monitoring method based on a multi-chain decomposition method provided by an embodiment of the present application;

[0061] Figure 4It is a schematic diagram of predicting device power consumption based on multi-chain decomposition provided by an embodiment of the present application;

[0062] Figure 5 It is a block diagram of a non-intrusive load monitoring device based on a multi-chain decomposition method provided by an embodiment of the present application. [Specific implementation method]

[0063] The specific implementation methods of the present application are further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.

[0064] Non-intrusive load monitoring, sometimes also called load decomposition, is a typical time series analysis problem. The non-intrusive load monitoring task is to infer the energy consumption of a single device from the observed total energy consumption data. Assuming that Y represents the total energy consumption data of a household observed from time 1 to time T, then Y = {y(t)|y(t)∈R + , 1≤t≤T, t∈N + The set of N household electrical appliances whose power consumption needs to be inferred in the problem is defined as A, that is, A = {A i |1≤i≤N,i∈N + Device A i The energy consumption data can be expressed as X i ={x i (t)|x i (t)∈R + , 1≤t≤T, t∈N + In addition to the several devices that need to be paid attention to, the total household power consumption also includes some devices whose power consumption does not need to be inferred. The power consumption of these devices is added together as a whole and regarded as a virtual device. The power consumption of this virtual device can be expressed as V = {v(t)│v(t)∈R + , 1≤t≤T, t∈N +}. In addition, the total amount of household electricity consumption also contains a certain amount of noise, and ε(t) represents the noise value in the total amount of electricity consumption data at time t. Then, the following formula can be obtained:

[0065]

[0066] The purpose of non-intrusive load monitoring is to restore the energy consumption data x of a single device from the measured total line energy consumption data y(t). i (t). Taking power as an example, the total power consumption at a certain moment is determined by the working state of each device and the corresponding power at that moment. The total power consumption is the sum of the power of each device and the noise.

[0067] When performing load decomposition, the traditional method is to input the total line data into a specific model, and the model outputs the estimated value of the power consumption of a specific device, that is, N independent models are required to decompose N devices. N is an integer greater than 1.

[0068] For example, three deep neural networks are used to complete non-intrusive load monitoring tasks based on low-frequency signals. The three schemes used are long short-term memory network (LSTM), denoising autoencoder (DAE), and a network that can regress the start time, end time and average power demand of each activated device.

[0069] For example, convolutional neural network (CNN) and recurrent neural network (RNN) are used to perform energy decomposition and compared.

[0070] For example, based on the characteristics of the gated recurrent unit (GRU), the GRU is applied to low-frequency non-intrusive load monitoring tasks.

[0071] For another example: a sequence-to-point (Seq2point) decomposition method is used to generate the input of the model using a sliding window from the data of the total line, and the model output is the inferred value of the power consumption data of a specific device at the middle moment of the time period corresponding to the input data.

[0072] For example, the generative adversarial network (GAN) was applied to the non-intrusive load monitoring task, and a sequence-to-subsequence (Seq2subseq) load monitoring method based on the conditional generative adversarial network (cGAN) was proposed.

[0073] All of the above methods treat the models inferred from different devices as independent individuals, and there is no correlation between the models, for example: Figure 1 The load monitoring model shown in the figure, each device in the load monitoring model corresponds to a sub-model. Figure 1 In the figure, an air conditioner corresponds to a sub-model 11, a refrigerator corresponds to a sub-model 12, a washing and drying machine corresponds to a sub-model 13, a dishwasher corresponds to a sub-model 14, and a microwave oven corresponds to a sub-model 15. The sub-models of each device are independent of each other, and the input of each sub-model is the power consumption data of the total line, and the output is the predicted power consumption value of the corresponding device.

[0074] In addition, the traditional tree structure model called treeCNN can also be used for load monitoring. This method breaks the situation that the models used to decompose various devices are independent of each other. The method adopted by treeCNN is to place different models as sub-models in the nodes of the tree structure to generate a total model. This not only maintains the relative independence of each sub-model, but also creates a connection between them. Tree structure, if N devices need to be decomposed, N! trees can be generated. The performance of these trees is different, so determining the structure of the tree is a key step in creating a treeCNN model. The existing method is a greedy method based on the idea of ​​determining the tree order and reducing the complexity of the solution.

[0075] The sub-models used to infer the power consumption of each device in each treeCNN model are in the same chain, and the length of the chain is the number of devices that need to be inferred N. This is also referred to as the single-chain decomposition scheme below. For example: Figure 2 In the load monitoring model shown, the models corresponding to all devices in the load monitoring model are connected in sequence to obtain a single chain. Figure 2 In the example, the nodes on the single chain are: sub-model 21 corresponding to the microwave oven, sub-model 22 corresponding to the refrigerator, sub-model 23 corresponding to the air conditioner, sub-model 24 corresponding to the washing machine and sub-model 25 corresponding to the dishwasher. The input of the first node on the single chain is the power consumption data of the total line, and the output is the device corresponding to the first node (such as Figure 2 The power consumption prediction value of the microwave oven in the figure); the input of each node after the first node is the power consumption data minus the difference between the power consumption prediction values ​​of each preceding node of the node, and the output is the power consumption prediction value of the node.

[0076] In this application, the preceding node refers to each sub-model located before the current node.

[0077] However, each electrical device has its own characteristics and different power consumption patterns. Since the types and order of nodes before the end node in the decomposition chain vary greatly, these will affect the final decomposition effect of the decomposition chain, so the optimal decomposition chain corresponding to each device is also quite different. It is not guaranteed that all devices in a decomposition chain can achieve their best decomposition performance. When inferring different devices, decomposition chains with different structures need to be used.

[0078] Based on the above technical problems, in this application, an independent model for inferring the power consumption of a specific device is used as a sub-model, and multiple decomposition chains are composed of multiple sub-models. The output of the end node of each decomposition chain is the inferred value of the power consumption of the specific device. Multiple decomposition chains constitute an overall model to obtain a multi-chain decomposition model.

[0079] In addition, since the time cost of finding the multi-chain decomposition model with the best performance is too high, the present application also provides two simplified solutions: (1) by limiting the maximum length of the model decomposition chain, the total number of multi-chain decomposition models that can be generated is reduced, thereby reducing the search range when using an exhaustive method to search for the multi-chain model with the best performance; (2) based on the relationship between the sub-models that constitute the multi-chain decomposition model, the type and position relationship of each sub-model in the multi-chain model with better performance is solved, and the sub-model is trained based on this, reducing some unnecessary sub-model training processes, thereby speeding up the solution process of the multi-chain decomposition model.

[0080] The non-invasive load monitoring method based on the multi-chain decomposition method provided by the present application is introduced in detail below.

[0081] Optionally, the present application uses the execution subject of each embodiment as an electronic device with computing capability as an example for illustration. The electronic device may be a terminal or a server. The terminal may be a computer, a laptop computer, a tablet computer, etc. This embodiment does not limit the type of terminal and the type of electronic device.

[0082] Figure 3 This is a flow chart of a non-intrusive load monitoring method based on a multi-chain decomposition method provided by an embodiment of the present application. The method includes at least the following steps:

[0083] Step 301, obtaining power consumption data in the main circuit.

[0084] The power consumption data includes: the power consumption of each power-consuming device to be monitored, the power consumption of a virtual device, and the noise value of the power consumption; wherein a virtual device refers to a device in the main line that does not need to be monitored.

[0085] Step 302, inputting the power consumption data into the overall power consumption monitoring model to obtain the power consumption prediction value of each power consuming device; wherein the overall power consumption monitoring model includes multiple decomposition chains, each decomposition chain corresponds to a power consuming device to be monitored, and different power consuming devices correspond to different decomposition chains; for each decomposition chain, one or more sub-models connected in sequence are included; each sub-model is used to predict the power consumption of the corresponding power consuming device, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power consuming device corresponding to the decomposition chain, so as to output the power consumption prediction value of the power consuming device corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power consuming devices.

[0086] The number of decomposition chains in the overall power consumption monitoring model is the same as the number of power consuming devices to be monitored.

[0087] Assume that the overall model for power consumption monitoring is as follows Figure 4 As shown, Figure 4In the example, there are five power-consuming devices to be monitored, namely, air conditioner, refrigerator, washing machine, dishwasher and microwave oven. Accordingly, the model includes five decomposition chains. The first decomposition chain is used to output the power consumption prediction value of the air conditioner, including sub-models corresponding to three power-consuming devices, namely, sub-model corresponding to the refrigerator, sub-model corresponding to the microwave oven and sub-model corresponding to the air conditioner. The second decomposition chain is used to output the power consumption prediction value of the refrigerator, including sub-models corresponding to two power-consuming devices, namely, sub-model corresponding to the washing machine and sub-model corresponding to the refrigerator. The third decomposition chain is used to output the power consumption prediction value of the washing machine, including a sub-model corresponding to one power-consuming device, namely, the sub-model corresponding to the washing machine. The fourth decomposition chain is used to output the power consumption prediction value of the dishwasher, including sub-models corresponding to five power-consuming devices, namely, sub-model corresponding to the microwave oven, sub-model corresponding to the refrigerator, sub-model corresponding to the air conditioner, sub-model corresponding to the washing machine and sub-model corresponding to the dishwasher. The fifth decomposition chain is used to output the predicted power consumption value of the microwave oven, including sub-models corresponding to three power-consuming devices, namely: a sub-model corresponding to the air conditioner, a sub-model corresponding to the refrigerator, and a sub-model corresponding to the microwave oven.

[0088] For a decomposition chain including n sub-models, where n is an integer greater than 1, the calculation process of the power consumption prediction value output by the decomposition chain includes: inputting the power consumption data into the first sub-model to obtain the output result of the first sub-model; inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1 in sequence; when k+1 is less than n, the value of k is equal to k+1, and the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model is executed again; when k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.

[0089] In this application, it is assumed that the framework for inferring the power consumption of all devices is called the overall model, which is represented by M. The decomposition chain that constitutes the overall model is represented by M chain The sub-models that make up the decomposition chain are represented by M sub (Z\A', i), where Z is the set of all components of the main line, including electrical equipment, virtual equipment and noise components, and A' is a subset of A, that is, Z\A' represents the set Z minus the device set A', where i is the index of the device. sub (Z\A', i) represents the power consumption data of the corresponding components of the elements in the set Z\A' and the inference of A i In particular, when A' is an empty set, Z\A' can be abbreviated to Z, Z\A j Removing device A from set Z jIn addition, P(Z, i) represents M sub The performance of (Z, i).

[0090] The overall process of using the multi-chain decomposition framework to infer device power consumption is as follows: First, the original dataset is divided into three subsets, called sub-model training datasets D train , total model search dataset D search And the total model test dataset D test , which are used to train sub-models, determine the structure of the overall model, and test the performance of the overall model. When it is necessary to infer the power consumption of multiple devices, first use D train Train multiple sub-models to form multiple total models, using D search Select the best performing model M from these models opt , and finally D test Enter M opt Infer the power consumption of the equipment and evaluate M with relevant evaluation indicators opt performance.

[0091] In the present application, the multi-chain decomposition model is composed of sub-models, and the sub-models can be created by any method suitable for non-intrusive load monitoring tasks. The first step in building a multi-chain decomposition model is to train multiple sub-models. Each set of training data used by each sub-model during training includes sample values ​​and labels. For a sub-model without a predecessor node, the sample value is the power consumption data, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model; for a sub-model with a predecessor node, the sample value is the difference between the power consumption data and the actual power consumption value corresponding to each predecessor node, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model.

[0092] During the model testing phase, the input of the sub-model is the total power consumption minus the output value of the previous node of the sub-model, and the output is the predicted power consumption value of the sub-model.

[0093] In one example, a method for generating a power consumption monitoring overall model includes: generating a plurality of overall model model structures based on respective sub-models, wherein the model structure of each overall model includes at least a plurality of decomposition chains; using a training data set to train each sub-model in the model structure of each overall model to obtain a trained model structure; using a model search set to search for the model structure with the best performance among the trained model structures to obtain a power consumption monitoring overall model; and using a test data set to test the power consumption monitoring overall model.

[0094] When there are N devices that need to analyze the power consumption, the length of the decomposition chain for one of the devices can vary from 1 to N. When the decomposition chain length is L (1≤L≤N), L-1 devices need to be selected from N-1 devices to be combined as the front nodes of the terminal node, and a total of When training a sub-model, the way the input data of the sub-model is processed has nothing to do with the order of the preceding nodes, so the sub-models used in the terminal nodes of these decomposition chains have Since 1≤L≤N, there are Different lengths of decomposition chains are required When inferring the power consumption of N devices, there are Total models are needed to construct these total models Seed model. Taking N=5 as an example, the total number of sub-models that need to be trained can reach 80. The more sub-models there are, the more time and space are required to create them. Since the multi-chain decomposition method can use multiple methods to create sub-models, when the deep learning method is used to build sub-models, model training is extremely time-consuming. Therefore, the need to train so many sub-models greatly limits the application of the multi-chain decomposition method.

[0095] Based on this, in this embodiment, model structures of multiple overall models are generated based on each sub-model, including: setting the maximum length of each decomposition chain in the model structure, the maximum length is less than or equal to the total number of power-consuming devices to be monitored; generating model structures of multiple overall models according to the maximum length, each decomposition chain in the model structure of each overall model is less than or equal to the maximum length.

[0096] Assume that when the maximum length of the decomposition chain is k (k∈N + , k≤N), a total of Different total models require different sub-models. Obviously, reducing k can reduce the scale of the number of multi-chain decomposition models and reduce the number of sub-models required. Of course, limiting k may affect the performance of the overall model. In extreme cases, when k=1, the multi-chain decomposition model degenerates into an independent model. At this time, the number of sub-models required to build the overall model is the least, and the time cost is the lowest, but at the same time the performance of the model will also be greatly affected. In this embodiment, the purpose of setting the maximum length is mainly to strike a balance between the time cost of building the model and the model performance, and to exchange a tolerable model performance loss for a faster model building process.

[0097] In another example, a method for generating a total power consumption monitoring model includes: using a training data set to train the sub-model corresponding to each power consuming device respectively to obtain multiple trained sub-models; for each trained sub-model, using other trained sub-models as predecessor nodes of the trained sub-model respectively to obtain a temporary decomposition chain; and training the temporary decomposition chain to obtain a trained temporary decomposition chain; using a test data set to test the trained sub-models and the trained temporary decomposition chain respectively; when the test result indicates that the performance of the trained sub-model is lower than the performance of the trained temporary decomposition chain, determining the predecessor node to be an associated node of the trained sub-model; when the test result indicates that the performance of the trained sub-model is higher than the performance of the trained temporary decomposition chain When it is possible, determine that the preceding node is not an associated node of the trained sub-model; construct a graph structure based on each trained sub-model and the associated nodes of each trained sub-model; in the graph structure, each trained sub-model is taken as a vertex of the graph, and the weight w(j, i) of the directed edge e(j, i) connecting vertex i and vertex j is the optimization degree of the temporary decomposition chain compared with the trained sub-model, and a negative weight represents performance improvement; wherein i represents the trained sub-model, and j represents the associated node of the trained sub-model; add special vertices to the graph structure, and the special vertices have outgoing edges pointing to all other vertices, and the weight is 0; calculate the shortest path from the special vertex to all other vertices without a loop, and obtain the multi-chain decomposition model structure of the total power consumption monitoring model.

[0098] In this embodiment, a graph-based algorithm (GBA) is provided to generate a model based on the relationship between two devices. Specifically, first, N sub-models M corresponding to all power-consuming devices to be decomposed are trained respectively. sub (Z,i) and obtain its evaluation index P(Z,i). Then, for each device A i , respectively, the remaining equipment A j As its front-end device, train the sub-model M sub (Z\A j ,i), the sub-model Ms ub (Z,j) and M sub (Z\A j ,i) form a decomposition chain, input the total power consumption data for model testing to obtain P(Z\A j ,i). Compare P(Z,i) and P(Z\A j ,i), if P(Z\A j ,i) is better than P(Z,i), which means that when we need to infer A i When the power consumption is j In A i It was separated before than directly separating Ai In order to solve the final multi-chain model result, in this embodiment, the relationship between two devices is converted into a graph structure, each device is a vertex of the graph, and the weight w(j,i) of the directed edge e(j,i) connecting vertex j and vertex i is P(Z\A j ,i) Compared with the optimization degree of P(Z,i), a negative weight represents an improvement in performance. The specific calculation method is related to the selection of evaluation indicators. Add a special vertex in the graph. This node has outgoing edges pointing to all other vertices and the weight is 0. The multi-chain decomposition model structure is obtained by finding the shortest path from the special vertex to all other vertices without loops. The requirement that there is no loop in the path is because the process of inferring and subtracting the power consumption of a device as an intermediate value in a decomposition chain should only exist once. The reason for seeking the shortest path is that reaching the end point through this path can make its decomposition performance improvement effect most obvious.

[0099] In the process of using GBA to obtain a multi-chain decomposition model, if it is necessary to infer the power consumption of N devices, first, it is necessary to train N sub-models for inferring the power consumption of specific devices from the total line data. Then, for each device A_i, the remaining devices A_j are used as their front-end devices, and (N-1)×N sub-models need to be trained. Then, the multi-chain model structure can be solved, and finally, the remaining sub-models that are missing from the multi-chain model structure are trained. Depending on the different model structures obtained, the number of remaining required sub-models is between 0 and N×(N-2). Compared with the exhaustive method, this method can greatly reduce the number of sub-models that need to be trained.

[0100] Optionally, when evaluating the performance of the model, the model performance evaluation indicators used include but are not limited to at least one of the following: mean absolute error (MAE), signal aggregate error (SAE) and F1 score.

[0101] MAE is the error between the actual value and the inferred value of the device power consumption. We use x i (t) represents device A i The actual value of power consumption at time t is represented by x' i (t) represents the corresponding estimated value. i MAE of estimated power consumption during time period T i The calculation formula is:

[0102]

[0103] SAE represents the total error of energy over a period of time, expressed as r i and r' i Respectively represent A iThe actual and estimated total power consumption over a period of time, SAE i The calculation formula is:

[0104]

[0105] Non-intrusive load monitoring can also be viewed as a classification task, where the state of the device is used as the value to be inferred, where the state is divided into two categories: on and off. We set a power-on threshold for each device and determine its state based on the relationship between the device's power consumption and the power-on threshold. Because the F1 score takes into account both accuracy and recall, and is more comprehensive, this article uses the F1 score as an evaluation indicator for the model classification performance. The calculation formula is as follows:

[0106]

[0107] In summary, the non-intrusive load monitoring method based on the multi-chain decomposition method provided in this embodiment obtains the power consumption data in the total line; inputs the power consumption data into the total power consumption monitoring model to obtain the power consumption prediction value of each power consuming device; wherein the total power consumption monitoring model includes multiple decomposition chains, each decomposition chain corresponds to a power consuming device to be monitored, and different power consuming devices correspond to different decomposition chains; for each decomposition chain, it includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power consuming device, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power consuming device corresponding to the decomposition chain, so as to output the power consumption prediction value of the power consuming device corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power consuming devices; the problem that the power consumption prediction value of each power consuming device obtained by the existing non-intrusive load monitoring method is not accurate enough can be solved, and an overall model is constructed by designing a multi-chain decomposition method to use independent models as sub-models, so that each sub-model is connected and the overall performance of the decomposition is improved.

[0108] In addition, when solving a multi-chain model, the present application also provides two simplified solutions to address the situation where the complexity of the exhaustive method for obtaining a multi-chain model with the best performance is too high: (1) by limiting the maximum length of the model decomposition chain, the total number of multi-chain decomposition models that can be generated is reduced, thereby reducing the search range when using an exhaustive method to search for a multi-chain model with the best performance; (2) based on the relationship between the sub-models that constitute the multi-chain decomposition model, the type and position relationship of each sub-model in the multi-chain model with better performance is solved, and the sub-model is trained based on this, thereby reducing some unnecessary sub-model training processes, thereby speeding up the solution process of the multi-chain decomposition model.

[0109] Figure 51 is a block diagram of a non-intrusive load monitoring device based on a multi-chain decomposition method provided by an embodiment of the present application. The device comprises at least the following modules: a total amount acquisition module 510 and a component prediction module 520 .

[0110] A total amount acquisition module 510 is used to acquire power consumption data in the total circuit;

[0111] The component prediction module 520 is used to input the power consumption data into the power consumption monitoring overall model to obtain the power consumption prediction value of each power consuming device;

[0112] The power consumption monitoring overall model includes a plurality of decomposition chains, each of which corresponds to a power-consuming device to be monitored, and different power-consuming devices correspond to different decomposition chains;

[0113] Each decomposition chain includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power-consuming equipment, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power-consuming equipment corresponding to the decomposition chain, so as to output the predicted value of the power consumption of the power-consuming equipment corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power-consuming equipment.

[0114] Optionally, for a decomposition chain including n sub-models, where n is an integer greater than 1, the component prediction module 520 is configured to:

[0115] Inputting the power consumption data into the first sub-model to obtain an output result of the first sub-model;

[0116] Input the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1;

[0117] When k+1 is less than n, the value of k is equal to k+1, and the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model is performed again;

[0118] When k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.

[0119] For relevant details, refer to the above method embodiment.

[0120] It should be noted that: the non-invasive load monitoring device based on the multi-chain decomposition method provided in the above embodiment only uses the division of the above-mentioned functional modules as an example when performing non-invasive load monitoring based on the multi-chain decomposition method. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the non-invasive load monitoring device based on the multi-chain decomposition method is divided into different functional modules to complete all or part of the functions described above. In addition, the non-invasive load monitoring device based on the multi-chain decomposition method provided in the above embodiment belongs to the same concept as the non-invasive load monitoring method embodiment based on the multi-chain decomposition method. The specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0121] Optionally, the present application also provides a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the non-intrusive load monitoring method based on the multi-chain decomposition method of the above method embodiment.

[0122] Optionally, the present application also provides a computer product, which includes a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the non-intrusive load monitoring method based on the multi-chain decomposition method of the above method embodiment.

[0123] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described 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.

[0124] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

[0125] The above is only a specific implementation of the present application, and any other improvements made based on the concept of the present application are deemed to be within the protection scope of the present application.

Claims

1. A non-intrusive load monitoring method based on a multi-chain decomposition method, characterized in that: The non-intrusive load monitoring method comprises: Get the power consumption data in the total circuit; Inputting the power consumption data into the power consumption monitoring overall model to obtain the power consumption prediction value of each power consuming device; The power consumption monitoring overall model includes a plurality of decomposition chains, each of which corresponds to a power-consuming device to be monitored, and different power-consuming devices correspond to different decomposition chains; Each decomposition chain includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power-consuming device, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power-consuming device corresponding to the decomposition chain, so as to output the power consumption prediction value of the power-consuming device corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power-consuming devices; Using the training data set, the sub-model corresponding to each power-consuming device is trained respectively to obtain multiple trained sub-models; For each trained sub-model, other trained sub-models are used as the preceding nodes of the trained sub-model to obtain a temporary decomposition chain; and the temporary decomposition chain is trained to obtain a trained temporary decomposition chain; Using a test data set to test the trained sub-model and the trained temporary decomposition chain respectively; When the test result indicates that the performance of the trained sub-model is lower than the performance of the trained temporary decomposition chain, determining the preceding node to be an associated node of the trained sub-model; When the test result indicates that the performance of the trained sub-model is higher than the performance of the trained temporary decomposition chain, determining that the preceding node is not an associated node of the trained sub-model; Based on each trained sub-model and the associated nodes of each trained sub-model, a graph structure is constructed; each trained sub-model in the graph structure is used as a vertex of the graph, and the weight w(j, i) of the directed edge e(j, i) connecting vertex i and vertex j is the optimization degree of the temporary decomposition chain compared with the trained sub-model, and a negative weight represents performance improvement; i represents the trained sub-model, and j represents the associated node of the trained sub-model; Add a special vertex in the graph structure, where the special vertex has outgoing edges pointing to all other vertices and has a weight of 0; The shortest path from the special vertex to all other vertices without loops is calculated to obtain a multi-chain decomposition model structure of the overall power consumption monitoring model.

2. The non-intrusive load monitoring method according to claim 1, characterized in that: For a decomposition chain including n sub-models, where n is an integer greater than 1, the calculation process of the power consumption prediction value output by the decomposition chain includes: Inputting the power consumption data into the first sub-model to obtain an output result of the first sub-model; Input the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1; When k+1 is less than n, set the value of k to be equal to k+1, and execute again the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; When k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.

3. The non-intrusive load monitoring method according to claim 1, characterized in that: The non-intrusive load monitoring method further comprises: Generate a plurality of model structures of the overall model based on each sub-model, wherein the model structure of each overall model includes at least a plurality of decomposition chains; Use the training data set to train each sub-model in the model structure of each overall model to obtain a trained model structure; Use the model search set to search for the model structure with the best performance among the trained model structures to obtain the overall power consumption monitoring model; The power consumption monitoring overall model is tested using a test data set.

4. The non-intrusive load monitoring method according to claim 3, characterized in that: The model structure of generating multiple overall models based on various sub-models includes: Setting a maximum length of each decomposition chain in the model structure, wherein the maximum length is less than or equal to the total number of the power-consuming devices to be monitored; Model structures of multiple overall models are generated according to the maximum length, and each decomposition chain in the model structure of each overall model is less than or equal to the maximum length.

5. The non-intrusive load monitoring method according to claim 1 or 3, characterized in that: Each set of training data includes sample values ​​and labels; For a sub-model without a preceding node, the sample value is power consumption data, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model; For a sub-model with a preceding node, the sample value is the difference between the power consumption data and the actual power consumption value corresponding to each preceding node, and the label is the actual power consumption value of the power-consuming device corresponding to the sub-model.

6. The non-intrusive load monitoring method according to any one of claims 1 to 4, characterized in that: The power consumption data includes: the power consumption of each of the power-consuming devices to be monitored, the power consumption of the virtual device, and the noise value of the power consumption; The virtual device refers to a device in the bus line that does not need to be monitored.

7. The non-intrusive load monitoring method according to any one of claims 1 to 4, characterized in that: The number of decomposition chains in the overall power consumption monitoring model is the same as the number of the power consuming devices to be monitored.

8. A non-intrusive load monitoring device based on a multi-chain decomposition device using the non-intrusive load monitoring method according to any one of claims 1 to 7, characterized in that: The non-intrusive load monitoring device comprises: A total amount acquisition module is used to obtain the power consumption data in the total circuit; A component prediction module, used for inputting the power consumption data into the power consumption monitoring overall model to obtain the power consumption prediction value of each power consuming device; The power consumption monitoring overall model includes a plurality of decomposition chains, each of which corresponds to a power-consuming device to be monitored, and different power-consuming devices correspond to different decomposition chains; Each decomposition chain includes one or more sub-models connected in sequence; each sub-model is used to predict the power consumption of the corresponding power-consuming equipment, and the sub-model at the end of the decomposition chain is used to predict the power consumption of the power-consuming equipment corresponding to the decomposition chain, so as to output the predicted value of the power consumption of the power-consuming equipment corresponding to the decomposition chain; different sub-models on the decomposition chain correspond to different power-consuming equipment.

9. The non-intrusive load monitoring device according to claim 8, characterized in that: For a decomposition chain including n sub-models, where n is an integer greater than 1, the component prediction module is used to: Inputting the power consumption data into the first sub-model to obtain an output result of the first sub-model; Input the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; k is an integer from 1 to n-1; When k+1 is less than n, set the value of k to be equal to k+1, and execute again the step of inputting the output result of the kth sub-model into the k+1th sub-model to obtain the output result of the k+1th sub-model; When k+1 is equal to n, the output result of the k+1th sub-model is used as the power consumption prediction value output by the decomposition chain.