Micro-machine learning-based non-intrusive power load decomposition method and system

Abstract

The application discloses a non-intrusive power load decomposition method based on micro machine learning, comprising the following steps: S1, collecting power load data and performing feature extraction, fast Fourier transform and normalization processing, and dividing the processed data into a training data set and a test data set; S2, performing iterative training on the power load data of the training data set based on a residual neural network structure to obtain a power load decomposition model; S3, quantizing and compressing the power load model; S4, inputting the power load data of the test data set into the compressed power load decomposition model and evaluating whether the decomposition model meets the use requirement; and S5, decomposing the real-time power load data of the user by the power load decomposition model to monitor the operating power of each power equipment. The application uses high-frequency sampling data and performs FFT decomposition, can effectively construct load characteristics and improve the model accuracy.

Description

Technical Field

[0001] This invention relates to the field of power system load monitoring technology, and in particular to a non-intrusive load decomposition method and system based on micro-machine learning. Background Technology

[0002] With socio-economic development, residents' electricity demand is constantly increasing, and the diversification of electrical equipment is making the characteristics of residents' electricity load increasingly complex. Residential electricity load is highly flexible and has great demand response potential; therefore, residential electricity load data reflects users' actual electricity consumption behavior and holds immense research value. However, residential electricity data exhibits typical big data characteristics such as large scale, high dimensionality, strong timeliness, high randomness, and low value density, making it difficult to extract features using traditional statistical methods. Therefore, how to utilize next-generation data mining techniques to discover patterns in residents' electricity consumption behavior from smart meter data—i.e., load identification and monitoring technology—and accordingly formulate efficient and energy-saving electricity plans to improve residents' demand response levels has become a research hotspot in recent years.

[0003] Non-intrusive load monitoring (NILM) systems analyze measurement data from monitoring equipment installed at the power inlet to determine the type and operational status of individual loads within a load cluster. For residential users, NILM involves identifying electricity consumption behavior and monitoring the operational status of appliances through household electricity meters. Users can reduce unnecessary energy expenditures through NILM, achieving energy conservation and emission reduction. Power companies can use NILM to understand the composition of user loads, strengthen load-side management, and guide users to consume rationally and schedule load usage time to regulate peak-valley differences and reduce network losses. Furthermore, high-performance NILMs can improve the accuracy of power load forecasting, providing more accurate data for power system simulation analysis and system planning.

[0004] The existing NILM methods are divided into two schemes of low-frequency data decomposition and high-frequency data decomposition. The sampling period of the low-frequency data is 1s, 3s or 5s, the sampling amount is generally active power, the hardware requirement is not high, the cost is relatively low, but the real-time performance is poor; the sampling frequency of the high-frequency data load decomposition is generally above 10kHz, the sampling amount includes voltage, current and other information, the hardware requirement is strict, the cost is relatively high, but the real-time performance is high. The NILM method based on high-frequency data is generally to use high-dimensional feature space machine learning algorithm, and the precision of this method is high, but a large amount of computing resources are needed, and it is more suitable for cloud computing deployment. However, the cloud computing framework has scalability problems in communication delay, bandwidth and privacy, which may endanger the privacy of customers, expose the energy status and daily activities, therefore, the NILM method using the cloud computing framework needs to ensure that the cloud computing does not endanger the privacy of personal data when collecting, storing, transmitting and sharing data, which greatly increases the cost and technical difficulty of the system.

[0005] Therefore, it is necessary to provide a more simple and efficient non-intrusive power load decomposition method and system. SUMMARY

[0006] Based on the above defects in the prior art, the purpose of the present application is to provide a non-intrusive power load decomposition method and system based on micro machine learning, which executes the non-intrusive power load monitoring method in the edge network, so that the end user can well protect the power privacy, and at the same time can take corrective measures faster for abnormal power consumption behavior, reduce energy consumption and avoid damage to power equipment.

[0007] In the first aspect, in order to achieve the above-mentioned purpose of the application, the present application provides a non-intrusive power load decomposition method based on micro machine learning, comprising:

[0008] S1, collecting power load data and extracting, fast Fourier transform and normalizing the power load data features, dividing the power load data after normalization into a training data set and a test data set according to a preset proportion;

[0009] S2, performing iterative training on the power load data of the training data set based on a residual neural network structure to obtain a power load decomposition model;

[0010] S3, quantizing and compressing the power load model by using a TensorFlow Lite framework to obtain a compressed power load decomposition model;

[0011] S4. Input the power load data of the test dataset into the compressed power load decomposition model and calculate the model evaluation parameters. Evaluate whether the compressed power load decomposition model meets the usage requirements through the model evaluation parameters. If it does, output the evaluated power load decomposition model. Otherwise, repeat steps S2 to S4 until the power load model meets the usage requirements.

[0012] S5. Use an evaluated power load decomposition model to decompose the user's real-time power load data in order to monitor the operating power of each electrical device.

[0013] Furthermore, step S1 includes:

[0014] S1-1, Acquire voltage value v at a sampling frequency of 10kHz. t and current value i t , where t = 1, 2, ..., T, and T is the number of times it is used.

[0015] Furthermore, step S1 also includes:

[0016] S1-2, with T s To determine the duration of the data acquisition, the voltage value v is extracted. t and current value i t From the available power load data of each of the T available power loads, the time-domain feature data of the available power load data is extracted, and the time-domain feature data includes apparent power |S t | Active power P t Reactive power Q t Frequency domain feature data of electricity load data is obtained through Fast Fourier Transform, and the frequency domain features include the real component of the current Re{I t} and the imaginary part of the current Im{ t}

[0017] Furthermore, step S1 also includes:

[0018] S1-3. Organize the time-domain and frequency-domain characteristic data of the electricity load data to obtain the time series of the electricity load data:

[0019]

[0020]

[0021] Where t = 1, 2, ..., T;

[0022] S1-4. Obtain the average operating power of each electrical device within a time period Ts:

[0023]

[0024] Wherein, j = 1, 2, …, L, L is the data amount of the data set, and M is the number of types of the electric equipment.

[0025] Further, the step S1 further comprises:

[0026] S1-5, repeatedly performing steps S1-2 to S1-4 to obtain the time series Y of the electric load data and the average running power X of each electric equipment:

[0027] Y = [Y1, Y2, …, Y j , …, Y N ],

[0028] X = [X1, X2, …, X j , …, X N ],

[0029] Wherein, N is the number of times of repeating the steps S1-2 to S1-4.

[0030] Further, the step S1 further comprises:

[0031] S1-6, performing data normalization processing on the time series Y of the electric load data and the average running power X of each electric equipment, and dividing them into a training data set and a test data set according to a preset proportion.

[0032] Further, the step S2 further comprises: verifying the accuracy of the electric load decomposition model by constructing a loss function, wherein the loss function is:

[0033] Wherein, is the model prediction value.

[0034] Further, the evaluation model parameters in the step S4 include the average absolute error, the normalized signal total error, and the standardized decomposition error.

[0035] The second aspect, in order to achieve the above application purposes, the present application further proposes a non-intrusive electric load decomposition system based on micro machine learning, comprising:

[0036] A data collector is configured to collect electric load data of a user.

[0037] A data processor is configured to execute the non-intrusive electric load decomposition method based on micro machine learning according to any one of the above.

[0038] Further, the data processor is configured as an MCU chip based on micro machine learning.

[0039] Compared with the prior art, the present application has the following beneficial effects:

[0040] The load characteristics are extracted by a residual convolutional neural network, the idea of cross-layer connection is used to make up for the defects of gradient dispersion and performance degradation caused by deepening of the network, the original input feature mapping and the output feature mapping of the latter layer are added, and the Selu activation function is used for activation, the feature fusion is completed, the residual connection makes the information transmission of the network more smooth, and the model convergence can be accelerated;

[0041] Tiny machine learning (Tiny ML) is a cross direction of machine learning and embedded Internet of Things (IoT) devices, has the characteristics of low delay, low power consumption, high performance, etc., the residual neural network model is compressed and embedded into the MCU chip, combined with high-frequency sampling, used for NILM technology, has the advantages of strong privacy, no communication bandwidth occupation, high identification performance, etc., the NILM method provided in the application is used, a large amount of installation cost of monitoring and sensing equipment is saved, the manpower and material resources required for maintenance after equipment installation are saved, and the acceptable degree of users for the load monitoring technology application can be increased;

[0042] The high-frequency sampling data is used for FFT decomposition, so that the load characteristics can be effectively constructed, and the model accuracy can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a cloud computing-based NILM system structure schematic diagram of the prior art;

[0044] Figure 2 It is an edge computing-based NILM system structure schematic diagram of the embodiment of the application;

[0045] Figure 3 It is a processing flowchart of the embodiment of the application;

[0046] Figure 4 It is a NILM framework schematic diagram of the embodiment of the application;

[0047] Figure 5 It is a residual neural network structure schematic diagram of the embodiment of the application. DETAILED DESCRIPTION

[0048] The application is described below based on the embodiments, but the application is not limited to only these embodiments. In the following detailed description of the application, some specific details are described in detail. The application can also be completely understood without the description of these details by those skilled in the art. In order to avoid confusion of the essence of the application, the well-known methods, processes, flows, elements and circuits are not described in detail.

[0049] In addition, those skilled in the art should understand that the drawings provided herein are for illustrative purposes only, and the drawings are not necessarily drawn to scale.

[0050] Meanwhile, it should be understood that in the following description, "circuitry" refers to an electrically conductive loop of at least one element or sub-circuitry constructed by electrical or electromagnetic connections. When an element or circuit is said to be "connected to" another element or said to be "connected between" two nodes, it can be directly coupled or connected to another element or there can be an intermediate element, the connection between elements can be physical, logical, or a combination thereof. On the contrary, when an element is said to be "directly coupled to" or "directly connected to" another element, it means that there is no intermediate element between the two.

[0051] Unless the context clearly requires otherwise, throughout the description and the claims, the words "comprise", "comprising", and the like are to be construed in an inclusive sense as opposed to an exclusive or exhaustive sense; that is to say, in the sense of "including, but not limited to".

[0052] In the description of the present application, it should be understood that the terms "first", "second" and the like are only used for descriptive purposes and should not be construed as indicating or implying relative importance. In addition, in the description of the present application, unless otherwise stated, the meaning of "a plurality of" is two or more.

[0053] As Figure 1 shown, the prior art NILM system based on cloud computing collects voltage and current data of each electrical equipment through a data acquisition module, and performs computationally intensive feature extraction and classification algorithms through a cloud server. The data uplink of the data acquisition module and the cloud server requires high bandwidth communication, and the cloud computing framework has scalability problems in communication delay, bandwidth and privacy. They may compromise the privacy of customers, expose energy status and daily activities, therefore, the NILM method using the cloud computing framework needs to ensure that the cloud computing does not compromise the privacy of personal data when collecting, storing, transmitting and sharing data, which greatly increases the cost and technical difficulty of the system.

[0054] As Figure 2 shown, the NILM system based on edge computing of the present embodiment collects voltage and current data of each electrical equipment through a data acquisition module, and performs feature extraction and classification algorithms through an edge device (such as a smart phone, a smart meter) as a data processing module, and the data acquisition module and the data processing module communicate through an edge network, which enables end users to better protect electrical privacy, while enabling faster corrective measures for abnormal electrical behavior, reducing energy consumption and avoiding damage to electrical equipment.

[0055] As Figures 3-5 shown, a non-intrusive electrical load decomposition method based on micro machine learning includes:

[0056] S1, collect electricity load data and extract, fast Fourier transform and normalize the electricity load data features, divide the normalized electricity load data into training data set and test data set according to the preset proportion;

[0057] S2, the electricity load data of the training data set is iteratively trained based on the residual neural network structure to obtain an electricity load decomposition model, the residual neural network structure used in the embodiment is as shown in Figure 5 ;

[0058] S3, quantize and compress the electricity load model using TensorFlow Lite framework to obtain a compressed electricity load decomposition model;

[0059] S4, input the electricity load data of the test data set into the compressed electricity load decomposition model and calculate the model evaluation parameters, evaluate whether the compressed electricity load decomposition model meets the use requirements through the model evaluation parameters, if yes, output the evaluated electricity load decomposition model, otherwise, repeat steps S2-S4 until the electricity load model meets the use requirements;

[0060] S5, use the evaluated electricity load decomposition model to decompose the real-time electricity load data of the user to monitor the running power of each electricity equipment.

[0061] As an implementation manner, as shown in Figure 4 , step S1 includes:

[0062] S1-1, collect voltage value v t and current value i t , where t=1, 2, …, T, T is the number of times of collection;

[0063] S1-2, collect T s length of time, intercept each T available electricity load data of voltage value v t and current value i t , extract the time domain feature data of the available electricity load data, the time domain feature data includes apparent power |S t |, active power P t , reactive power Q t , obtain the frequency domain feature data of the electricity load data by fast Fourier transform, the frequency domain feature includes real part of current Re{I t} and imaginary part of current Im{ t};

[0064] S1-3, arrange the time domain feature data and frequency domain feature data of the electricity load data to obtain the time sequence of the electricity load data:

[0065]

[0066]

[0067] wherein t = 1, 2, … T;

[0068] S1-4, obtaining the average running power of each electrical equipment in the time length Ts:

[0069]

[0070] wherein j = 1, 2, …, L, L is the data amount of the data set, and M is the number of types of electrical equipment;

[0071] S1-5, repeatedly performing steps S1-2 to S1-4 to obtain the time series Y of the electrical load data and the average running power X of each electrical equipment:

[0072] Y = [Y1, Y2, …, Y j , …, Y N ],

[0073] X = [X1, X2, …, X j , …, X N ],

[0074] wherein N is the number of times of repeating the execution of steps S1-2 to S1-4;

[0075] S1-6, performing data normalization processing on the time series Y of the electrical load data and the average running power X of each electrical equipment, and dividing them into a training data set and a test data set according to a preset proportion.

[0076] As an implementation manner, step S2 further comprises: verifying the accuracy of the electrical load decomposition model by constructing a loss function, wherein the loss function is:

[0077] wherein Y is the actual value, and is the model prediction value. The electrical load decomposition model is trained through the residual neural network structure as shown in Figure 5 until the loss function Loss converges.

[0078] As an implementation manner, the evaluation model parameters in step S4 comprise the mean absolute error MAE, the normalized signal total error SAE, and the normalized decomposition error NDE.

[0079] The compressed power consumption load decomposition model is evaluated by using mean absolute error MAE, normalized signal total error SAE, and normalized decomposition error NDE until ideal results are obtained, otherwise the related training parameters in step S2 are adjusted, such as learning rate, optimizer, training times, etc., or the compression parameters in step S3 are adjusted, and steps S2-S4 are repeatedly executed until ideal evaluation results are obtained.

[0080] The evaluated power consumption load decomposition model in step S5 is used to decompose the real-time power consumption load data of the user to monitor the operating power of each power consumption device, including:

[0081] S5-1, collect voltage value v t and current value i t , where t=1, 2, …, T, T is the number of times of collection;

[0082] S5-2, collect T s length of time, intercept each T available power consumption load data of voltage value v t and current value i t , extract the time domain feature data of the available power consumption load data, the time domain feature data includes apparent power |S t |, active power P t , and reactive power Q t , obtain the frequency domain feature data of the power consumption load data by fast Fourier transform, the frequency domain feature includes real part of current Re{I t} and imaginary part of current Im{ t};

[0083] S5-3, arrange the time domain feature data and frequency domain feature data of the power consumption load data to obtain the time series of the power consumption load data:

[0084]

[0085]

[0086] where t=1, 2, …, T;

[0087] S5-4, decompose the load by the non-intrusive power consumption load decomposition model trained and optimized and deployed on the MCU to obtain the average operating power of each device in the time window Ts where i=1, 2, …, M, M is the number of types of power consumption devices.

[0088] A micro machine learning-based non-intrusive power load decomposition system comprises a data collector configured to collect power load data of a user; and a data processor configured to perform the micro machine learning-based non-intrusive power load decomposition method according to any one of the preceding embodiments, wherein the data processor is a micro machine learning-based MCU chip.

[0089] In summary, the beneficial effects of the embodiments of the present application are as follows:

[0090] The load features are extracted by the residual convolutional neural network, the idea of cross-layer connection is used to make up for the defects of gradient dispersion and performance degradation caused by deepening of the network, the original input feature mapping and the output feature mapping of the later layer are added, and the Selu activation function is used for activation to complete the fusion of the features, the residual connection makes the information transmission of the network more smooth, and can accelerate the model convergence;

[0091] Tiny machine learning (Tiny ML) is a cross direction of machine learning and embedded Internet of Things (IoT) devices, which has the characteristics of low delay, low power consumption and high performance, the residual neural network model is compressed and embedded into the MCU chip, combined with high-frequency sampling, used for NILM technology, has the advantages of strong privacy, no communication bandwidth occupation, high recognition performance, etc., using the NILM method proposed in the present application, a large amount of installation cost of monitoring and sensing devices is saved, the manpower and material resources needed for maintenance after installation of the equipment are saved, and the acceptable degree of users for the load monitoring technology application can be increased;

[0092] Using high-frequency sampling data and performing FFT decomposition can effectively construct load features and improve model accuracy.

[0093] The technical solutions of the present application are described above in combination with specific embodiments, but it should be noted that the above description is only for the purpose of explaining the solutions of the present application, and cannot be interpreted as a specific limitation on the protection scope of the invention. Based on the above explanation, other specific embodiments or equivalent replacements of the present application can be conceived by those skilled in the art without creative labor, which will fall within the protection scope of the present application.

Claims

1. A micro machine learning based non-intrusive load disaggregation method, characterized by, The method comprises the following steps: S1, collecting power consumption load data and performing feature extraction, fast Fourier transform and normalization processing on the power consumption load data, and dividing the normalized power consumption load data into a training data set and a test data set according to a preset proportion; S2, performing iterative training on the power consumption load data of the training data set based on a residual neural network structure to obtain a power consumption load decomposition model; S3, quantizing and compressing the power load model using a TensorFlow Lite framework to obtain a compressed power consumption load decomposition model; S4, inputting the power consumption load data of the test data set into the compressed power consumption load decomposition model and calculating model evaluation parameters, and evaluating whether the compressed power consumption load decomposition model meets the use requirements through the model evaluation parameters, if yes, outputting the evaluated power consumption load decomposition model, otherwise, repeating steps S2-S4 until the power consumption load model meets the use requirements; S5, using the evaluated power consumption load decomposition model deployed on the MCU to decompose the real-time power consumption load data of the user to monitor the operating power of each power consumption device.

2. The micro machine learning based non-intrusive load decomposition method according to claim 1, wherein, The step S1 comprises: S1-1, collect the voltage value v with a sampling frequency of 10 kHz t and the current value i t wherein t = 1, 2, … T, T is the number of times of adoption.

3. The micro machine learning based non-intrusive load decomposition method according to claim 2, wherein, The step S1 further comprises: S1-2, with T s as the collection time length, intercepting the voltage value v t and the current value i t of each T available power consumption load data, extracting the time domain feature data of the available power consumption load data, the time domain feature data including apparent power |S t |, active power P t , and reactive power Q t , obtaining the frequency domain feature data of the power consumption load data through fast Fourier transform, and the frequency domain feature including the real part of the current Re{I t} and the imaginary part of the current Im{I t}.

4. The micro machine learning based non-intrusive load decomposition method according to claim 3, wherein, The step S1 further comprises: S1-3, arranging the time domain feature data and the frequency domain feature data of the power consumption load data to obtain a time sequence of the power consumption load data: Wherein, t=1, 2, …T; S1-4, obtaining the average operating power of each power consumption device within a time length Ts: Wherein, j=1, 2, …, L, L is the data amount of the data set, and M is the number of types of power consumption devices.

5. The micro machine learning based non-intrusive load decomposition method according to claim 4, wherein, The step S1 further comprises: S1-5, repeatedly performing steps S1-2-S1-4 to obtain a time sequence Y of the power consumption load data and an average operating power X of each power consumption device: Y = [Y1, Y2,..., Y j ,..., Y N ], X = [X1, X2,..., X j ,..., X N ], Wherein, N is the number of times of repeating steps S1-2-S1-4.

6. The micro machine learning based non-intrusive load decomposition method according to claim 5, wherein, The step S1 further comprises: S1-6, performing data normalization processing on the time sequence Y of the power consumption load data and the average operating power X of each power consumption device, and dividing them into a training data set and a test data set according to a preset proportion.

7. The micro machine learning based non-intrusive load decomposition method according to claim 1, wherein, The step S2 further comprises: verifying the accuracy of the power consumption load decomposition model by constructing a loss function, wherein the loss function is: wherein, is the model predicted value.

8. The micro machine learning based non-intrusive load decomposition method according to claim 1, wherein, The evaluation model parameters in step S4 include mean absolute error, normalized signal total error, and standardized decomposition error.

9. A micro machine learning based non-intrusive electricity load decomposition system, characterized by, The method comprises the following steps: The data acquisition module is configured to collect power consumption load data of a user; The data processing module is configured to perform the micro machine learning-based non-intrusive power consumption load decomposition method according to any one of claims 1-8.

10. The micro machine learning based non-intrusive electricity load decomposition system of claim 9, wherein, The data processing module is configured as a micro machine learning-based MCU chip.

Images