A non-intrusive load decomposition method based on progressive learning structure

Abstract

This invention discloses a non-intrusive load decomposition method based on a progressive learning structure. The method comprises the following steps: collecting data on the total active power of a house and the active power of each target appliance; preprocessing the data to construct a dataset; using the constructed dataset to train a non-intrusive load decomposition algorithm model based on the progressive learning structure; and inputting the total active power of an unknown house into the non-intrusive load decomposition algorithm model based on the progressive learning structure to decompose and obtain the power consumption sequence of the target appliance. By employing a progressive learning structure and a timestamp embedding method, the present invention enhances the model's feature extraction capabilities and enriches the model's input, significantly improving the accuracy and generalization capabilities of the non-intrusive load decomposition algorithm.

Description

Technical Field

[0001] The present invention relates to a non-invasive load decomposition method based on a progressive learning structure, which belongs to the field of power technology and artificial intelligence technology. Background Art

[0002] The development of modern power systems has driven a demand for fine-grained data at the appliance level. Appliance consumption data can be used to calculate demand response resources, identify equipment failures, generate detailed electricity bills, and build power profiling systems. The most straightforward way to obtain appliance power consumption is to install meters directly on the target appliances. While this invasive load monitoring method can produce accurate results, the large-scale deployment of smart metering equipment is prohibitively expensive. Non-intrusive load monitoring (NILM) has become a mainstream research topic because it leverages existing infrastructure and can significantly reduce hardware costs through software algorithms. The entire process of non-intrusive load decomposition relies solely on single-point measurements, using the power, voltage, and current recorded by the master meter as features to identify the operating status of individual appliances and estimate their power consumption. Compared to invasive methods, non-intrusive load decomposition offers significantly lower costs and higher software algorithm flexibility, demonstrating its significant potential for application.

[0003] The problem of non-intrusive load decomposition can be broken down into two subtasks: determining the on / off state of a device and estimating its specific power consumption during its on-state. Existing deep learning-based non-intrusive load decomposition algorithms typically assign these two complex tasks to a single network, limiting the accuracy of model predictions. Furthermore, most current non-intrusive load decomposition methods rely solely on a mapping between total power and target appliance power to achieve decomposition, ignoring the cyclical nature of appliance operation and failing to learn the distinct patterns of electricity usage habits. Summary of the Invention

[0004] The present invention provides a non-invasive load decomposition method based on a progressive learning structure. The two subtasks decomposed from the load decomposition problem are progressively completed through a two-part network, and time information is used to assist the model in making power predictions, aiming to improve the accuracy and generalization ability of load decomposition.

[0005] A non-intrusive load decomposition method based on a progressive learning structure, comprising:

[0006] Collect the total active power of the house and the active power data of each target electrical appliance;

[0007] Preprocess the data and construct the data set;

[0008] The constructed dataset is used to train a non-intrusive load decomposition algorithm model based on a progressive learning structure;

[0009] The total active power of the unknown house is input into a non-intrusive load decomposition algorithm model based on a progressive learning structure, and the power consumption sequence of the target electrical appliances is decomposed.

[0010] Furthermore, the data preprocessing step includes:

[0011] Resample the power data at intervals of T;

[0012] For data missing less than t time, fill in the value by forward filling, and for data missing more than t time, fill in the value with 0;

[0013] Delete the total power below P thres data;

[0014] Add a status mark to the data set. If the power of the appliance is greater than the turn-on threshold, the device is turned on and the status value is marked as 1. If the power of the appliance is less than the turn-on threshold, it is marked as 0.

[0015] Furthermore, the non-intrusive load decomposition algorithm model based on the progressive learning structure includes network I and network II. Network I is responsible for determining the on / off state of the target appliance, and network II accurately restores the power curve of the target appliance based on the decomposition result of network I.

[0016] The input of the network I is the total power signal and timestamp data, and the output is the preliminary predicted power sequence;

[0017] The input of the network II is the total power signal, timestamp data and the target electrical appliance state sequence initially predicted by the network I, and the output is the final predicted power sequence.

[0018] Furthermore, both network I and network II include an information embedding layer, a core decomposition layer and an output layer; the core decomposition layer of network I has the same structure as the core decomposition layer of network II, the output layer of network I has the same structure as the output layer of network II, and the information embedding layer of network I and the information embedding layer of network II differ due to different input information.

[0019] Furthermore, the information embedding layer includes embedding of aggregated power data, embedding of state data and embedding of time information;

[0020] For each type of input sequence data, a convolutional layer is used to extract features from the sequence while increasing the dimension, and then a pooling layer is used to shorten the sequence length. Finally, a summation is performed to fuse multiple embedded information.

[0021] Furthermore, the time information encoding step in the time information embedding includes:

[0022] Get timestamp data;

[0023] Extract the three features of hour of day, day of week, and month of year for each timestamp;

[0024] These three features are linearly encoded into values ​​in the range [-0.5, 0.5].

[0025] Furthermore, the core decomposition layer includes two layers of encoders, and each layer of encoder network is composed of a multi-head self-attention network and a feedforward neural network.

[0026] Furthermore, the output layer includes a deconvolution layer and two fully connected layers.

[0027] Furthermore, the loss functions of the network I and the network II include:

[0028] Mean square error

[0029] Bulldozing distance error

[0030] Dice loss

[0031] Mean absolute error

[0032] in, and y t They represent the appliance power predicted by the model and the actual appliance power at time t respectively; and s t They represent the appliance status predicted by the model and the actual appliance status at time t respectively; T is the total number of sequence points; and y represent the distribution of the model-predicted electrical power and the actual electrical power series, respectively. is a set of distributions on ×, whose marginals are and y, u and v are integral variables; O is the timing point at which the appliance is actually turned on or the state prediction error occurs;

[0033] According to the different tasks focused on by the two networks, the loss function calculation formulas of Network I and Network II are as follows:

[0034]

[0035] Among them, λ is the weight of the mean absolute error loss term.

[0036] Furthermore, after the power consumption sequence of the target electrical appliance is obtained by the decomposition, four indicators are used to evaluate the effect of the non-intrusive load decomposition algorithm model based on the progressive learning structure, including accuracy, F1 score, mean absolute error and mean relative error.

[0037] The non-invasive load decomposition method provided by the present invention has at least the following beneficial effects:

[0038] The load decomposition algorithm of the present invention utilizes a two-part network to progressively solve the two sub-problems of judging the switch state of electrical appliances and predicting the power value, so that a single network can focus on a single task, thus improving the completion effect of both tasks. In addition, the present invention introduces time information in the information embedding layer to improve the model's ability to capture the characteristics of electricity usage habits. In the REDD dataset, microwave ovens, washing machines, dishwashers, and refrigerators were selected as research objects, and houses 2, 3, 4, 5, and 6 were used as training sets, and house 1 was used as a test set. The Pytorch deep learning framework was used in the Python programming platform to train and test the model. The model of the present invention outperformed the other four non-invasive load decomposition algorithms based on deep learning in terms of classification index accuracy, F1 index, and regression index mean absolute error and mean relative error, and was particularly able to significantly reduce the mean absolute error. The load decomposition algorithm of the present invention has a certain generalization ability for fresh samples and performs well in cross-family tests. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of an embodiment of a non-invasive load decomposition method based on a progressive learning structure provided by the present invention;

[0040] Figure 2 This is a diagram of the progressive learning structure proposed by the present invention;

[0041] Figure 3 A single network structure diagram in the present invention;

[0042] Figure 4 This is a schematic diagram of the timestamp description method provided by the present invention. DETAILED DESCRIPTION

[0043] In order to illustrate the technical solution more clearly, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0044] The present invention provides a non-invasive load decomposition method based on a progressive learning structure, which progressively completes the two subtasks decomposed from the load decomposition problem through a two-part network, and uses time information to assist the model in making power predictions, thereby improving the accuracy of load decomposition.

[0045] refer to Figure 1 In some embodiments, a non-intrusive load decomposition method based on a progressive learning structure is provided, comprising:

[0046] S1. Use a sampling device to collect the total active power of the house and the active power data of each target electrical appliance;

[0047] S2, preprocess the data and construct the data set;

[0048] S3. Using the constructed data set to train a non-intrusive load decomposition algorithm model based on a progressive learning structure;

[0049] S4. Input the total active power of the unknown house into the model and decompose it to obtain the power consumption sequence of the target electrical appliances.

[0050] The power sampling is low-frequency sampling, which ensures the low cost of the sampling equipment.

[0051] The data preprocessing step includes:

[0052] S2.1. Resample the power data at intervals of T.

[0053] S2.2. For missing data less than t time, fill in the value by forward filling, and for missing data more than t time, fill in the value with 0;

[0054] S2.3, delete the total power below P thres The data indicates that no electrical appliances are turned on and no disassembly is necessary;

[0055] S2.4. Add a status flag to the data set. If the power of the appliance is greater than the on threshold, the device is considered on and the status value is marked as 1; otherwise, it is marked as 0.

[0056] The non-intrusive load decomposition algorithm model based on progressive learning structure includes two networks, which feed the output features of the previous network together with the original features into the next network, thereby improving the feature extraction and fitting capabilities of the system. Figure 2 Network I is primarily responsible for determining the on / off state of the target appliance, while Network II accurately restores the target appliance's power curve based on the decomposition results of Network I. Network I takes the total power signal and timestamp data as input, and outputs a preliminary predicted power sequence. Network II takes the total power signal, timestamp data, and the target appliance state sequence initially predicted by Network I as input, and outputs the final predicted power sequence.

[0057] Further, refer to Figure 3 ,Network I and Network II both consist of three parts: information embedding layer, core decomposition layer and output layer.,The core decomposition layer and output layer of the two networks have the same structure,,while the information embedding layer varies due to the different input,information.

[0058] Furthermore, the information embedding layer includes embedding of aggregated power data, embedding of state data and embedding of time information;

[0059] For each type of input sequence data, a convolutional layer is used to extract features from a sequence of length L, while increasing the dimension to H. The sequence length is then shortened to L / 2 through a pooling layer, and finally multiple embedded information is fused by summation calculation.

[0060] refer to Figure 4 In some embodiments, the time information encoding step includes:

[0061] Get timestamp data;

[0062] Extract the three features of hour of day, day of week, and month of year for each timestamp;

[0063] These three features are linearly encoded into values ​​in the range of [-0.5, 0.5].

[0064] Furthermore, the core decomposition layer includes two encoder layers, each consisting of a multi-head self-attention network and a feedforward neural network. The multi-head self-attention network has two heads; the feedforward neural network uses two layers of full connection and GELU activation to process input elements, then uses residual connections to prevent network degradation and gradient vanishing, and then performs layer normalization to stabilize the forward input distribution.

[0065] Furthermore, the output layer maps the features extracted by the core decomposition layer into a single appliance power sequence of the same length as the total power sequence. The deconvolution operation expands the data length to L, and the two fully connected layers convert the feature matrix into one-dimensional data of length L.

[0066] The loss functions of Network I and Network II include:

[0067] Mean square error

[0068] Bulldozing distance error

[0069] Dice loss

[0070] Mean absolute error

[0071] in, and y t They represent the appliance power predicted by the model and the actual appliance power respectively; and s represent the appliance status predicted by the model and the actual appliance status, respectively; T is the total number of sequence points; is a set of distributions on ×, whose marginals are and y t ; O is the timing point when the appliance is actually turned on or the state prediction error occurs.

[0072] Network I's contribution to the final output lies primarily in providing relatively accurate classification information to Network II. Therefore, its loss function consists of three terms: mean square error, bulldozing distance error, and dice loss. The dice loss plays a major role in guiding the network to make accurate judgments about the on / off state of the target appliance. Based on the appliance's state information, Network II needs to fit the power curve of the appliance's operation. Therefore, its loss function consists of three terms: mean square error, bulldozing distance error, and mean absolute error. The mean absolute error loss term corrects the power value when the device is turned on. The loss function calculation formulas for Network I and Network II can be expressed as:

[0073]

[0074] Among them, λ is the weight of the mean absolute error loss term.

[0075] The training process of the non-intrusive load decomposition model provided in this embodiment uses the Adam optimizer to accelerate gradient descent.

[0076] Furthermore, after decomposing the power consumption sequence of the target appliance, four indicators are used to evaluate the model effect, including accuracy, F1 score, mean absolute error (MAE), and mean relative error (MRE). The calculation formula is as follows:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] Among them, TP, FP, and FN represent the number of appliances that are actually turned on and the decomposition result is also on, the number of appliances that are actually turned off and the decomposition result is on, and the number of appliances that are turned off and the decomposition result is also off. P and N represent the number of appliances that are actually turned on and off, respectively. and y t They represent the appliance power predicted by the model and the actual appliance power respectively; T is the total number of sequence points.

[0084] The proposed algorithm was tested on low-frequency data from the REDD dataset. The average performance of each evaluation metric outperformed four baseline models. When tested individually on a wide range of appliances, the proposed method also demonstrated superior overall performance. In particular, the proposed method outperformed other algorithms on every appliance using the mean absolute error metric.

[0085] The present invention introduces time information into the information embedding layer to improve the model's ability to capture electricity usage habit characteristics.

[0086] The load decomposition algorithm of the present invention has a certain generalization ability for fresh samples and performs well in cross-family tests.

[0087] The load decomposition algorithm of the present invention utilizes a two-part network to progressively solve the two sub-problems of judging the switch state of electrical appliances and predicting the power value, so that a single network can focus on a single task, thus improving the completion effect of both tasks. In addition, the present invention introduces time information in the information embedding layer to improve the model's ability to capture the characteristics of electricity usage habits. In the REDD dataset, microwave ovens, washing machines, dishwashers, and refrigerators were selected as research objects, and houses 2, 3, 4, 5, and 6 were used as training sets, and house 1 was used as a test set. The Pytorch deep learning framework was used in the Python programming platform to train and test the model. The model of the present invention outperformed the other four non-invasive load decomposition algorithms based on deep learning in terms of classification index accuracy, F1 index, and regression index mean absolute error and mean relative error, and was particularly able to significantly reduce the mean absolute error. The load decomposition algorithm of the present invention has a certain generalization ability for fresh samples and performs well in cross-family tests.

[0088] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.

Claims

1. A non-intrusive load decomposition method based on a progressive learning structure, characterized in that: include: S1. Collect the total active power of the house and the active power data of each target electrical appliance; S2. Preprocess the data to construct a data set. The data preprocessing step includes: Resample the power data at intervals of T; For data missing less than t time, fill in the value by forward filling, and for data missing more than t time, fill in the value with 0; Delete the total power below P thres data; Add a status mark to the data set. If the power of the appliance is greater than the threshold, the device is turned on and the status value is marked as 1. If the power of the appliance is less than the threshold, the status value is marked as 0. S3. Using the constructed data set to train a non-intrusive load decomposition algorithm model based on a progressive learning structure, the non-intrusive load decomposition algorithm model based on a progressive learning structure includes a network I and a network II. The network I is responsible for determining the on / off state of the target electrical appliance, and the network II accurately restores the power curve of the target electrical appliance based on the decomposition result of the network I. The input of the network I is the total power signal and timestamp data, and the output is the preliminary predicted power sequence; The input of the network II is the total power signal, timestamp data and the target electrical appliance state sequence initially predicted by the network I, and the output is the final predicted power sequence; The loss functions of the network I and network II include: Mean square error Bulldozing distance error Dice loss Mean absolute error in, and y t They represent the appliance power predicted by the model and the actual appliance power at time t respectively; and s t They represent the appliance status predicted by the model and the actual appliance status at time t respectively; T is the total number of sequence points; and y represent the distribution of the model-predicted electrical power and the actual electrical power series, respectively. yes A set of distributions on , whose marginals are and y, u and v are integral variables; O is the timing point at which the appliance is actually turned on or the state prediction error occurs; According to the different tasks focused on by the two networks, the loss function calculation formulas of Network I and Network II are as follows: Among them, λ is the weight of the mean absolute error loss term; S4. The total active power of the unknown house is input into a non-intrusive load decomposition algorithm model based on a progressive learning structure to decompose the power consumption sequence of the target electrical appliances.

2. The non-invasive load decomposition method based on progressive learning structure according to claim 1 is characterized in that: Both network I and network II include an information embedding layer, a core decomposition layer and an output layer; the core decomposition layer of network I has the same structure as the core decomposition layer of network II, the output layer of network I has the same structure as the output layer of network II, and the information embedding layer of network I and network II differs due to different input information.

3. The non-invasive load decomposition method based on progressive learning structure according to claim 2 is characterized in that: The information embedding layer includes embedding of aggregated power data, embedding of state data and embedding of time information; For each type of input sequence data, a convolutional layer is used to extract features from the sequence while increasing the dimension, and then a pooling layer is used to shorten the sequence length. Finally, a summation is performed to fuse multiple embedded information.

4. The non-invasive load decomposition method based on progressive learning structure according to claim 3 is characterized in that: The time information encoding step in the time information embedding includes: Get timestamp data; Extract the three features of hour of day, day of week, and month of year for each timestamp; These three features are linearly encoded into values ​​in the range [-0.5, 0.5].

5. The method according to claim 2, characterized in that The core decomposition layer includes two layers of encoders, and each layer of encoder network is composed of a multi-head self-attention network and a feedforward neural network.

6. The non-invasive load decomposition method based on progressive learning structure according to claim 2, characterized in that: The output layer includes a deconvolution layer and two fully connected layers.

7. The non-invasive load decomposition method based on progressive learning structure according to claim 1, characterized in that: After the power consumption sequence of the target electrical appliance is obtained by the decomposition, four indicators are used to evaluate the effect of the non-intrusive load decomposition algorithm model based on the progressive learning structure, including accuracy, F1 score, mean absolute error and mean relative error.

Images