Non-intrusive electric energy load decomposition method based on spatial-temporal feature reconstruction technology

Abstract

The invention relates to the field of power system load analysis, and discloses a non-intrusive electric energy load decomposition method based on a spatial-temporal feature reconstruction technology. According to the method, fast Fourier transform (FFT) is adopted to find a periodic mode of power data, then a one-dimensional time sequence is remodeled into a two-dimensional space-time tensor according to a period, data feature information of the two-dimensional space-time tensor is deeply mined by using a convolutional neural network, and non-intrusive decomposition of the power data is realized. According to the method, the spatial-temporal characteristics of the power data are reconstructed, so that the method can learn the change relation between the period and the period, the power utilization habit of a user is deeply mined, and the load decomposition precision is improved. Meanwhile, the identified power load periodicity corresponds to the power equipment, so that the influence of similar power electric appliances can be avoided.

Description

Technical Field

[0001] The present invention relates to the field of non-invasive load decomposition, and in particular to a non-invasive electric energy load decomposition method based on spatiotemporal feature reconstruction technology. Background Art

[0002] In the current power system, accurate decomposition of user-side electric energy load is crucial for the optimization of power system scheduling, energy efficiency management and demand response strategy formulation. Traditional methods of electric energy load decomposition mostly rely on invasive installation of sensors or dedicated measurement equipment, which are not only costly, but also complicated to install and maintain, causing inconvenience to users. Therefore, non-invasive load decomposition technology has gradually become a research hotspot.

[0003] Existing non-intrusive load decomposition technologies are mainly based on statistical methods or machine learning algorithms. Although these methods can achieve load decomposition to a certain extent, they often have the following limitations: First, the extraction of periodic features in power data is insufficient, resulting in low accuracy of load decomposition; second, there is a lack of deep-level feature mining of power loads in time and space dimensions, making it difficult for the decomposition results to meet the needs of practical applications. Summary of the invention

[0004] The present invention aims to provide a non-invasive electric energy load decomposition method based on spatiotemporal feature reconstruction technology. The method of the present invention uses fast Fourier transform to find the periodicity of load data, reconstructs the one-dimensional time series into a two-dimensional spatiotemporal tensor according to the period, and uses a two-dimensional convolutional network to deeply mine the data features of the two-dimensional spatiotemporal tensor, thereby capturing the change information within the period and between cycles, and achieving the purpose of equipment identification and load decomposition.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A non-intrusive electric energy load decomposition method based on spatiotemporal feature reconstruction technology, the method is implemented based on a non-intrusive load decomposition model, and the construction method of the non-intrusive load decomposition model includes the following steps:

[0007] Step S1, collecting load data, including high-frequency active power data of the total electric meter and each electric device, and constructing an original data set; the high-frequency active power data is specifically the electric load active power data collected at time intervals of seconds or even shorter;

[0008] Step S2, data preprocessing, deletion, filling, data resampling and normalization of missing values ​​in the original data set, and division into training data set and test data set; according to the required decomposition accuracy, the high-frequency data set can be resampled to convert it to the required resolution;

[0009] Step S3, training the non-intrusive load decomposition model, taking batches of the training data set, training each batch of training data, setting the MSE index as the loss function, measuring the loss of the decomposition results and the high-frequency active power data of n electrical equipment, and then adjusting the parameters; and using the test data set for verification, selecting the model with the best verification result as the trained model;

[0010] The training of each batch of training data is performed based on the time module. The specific training process is as follows:

[0011] 1) Feature enhancement: reshape the one-dimensional time series of the total meter load data in the training dataset into a two-dimensional space-time tensor;

[0012] 2) Feature extraction, using a convolutional neural network to deeply mine the data features of the two-dimensional spatiotemporal tensor to obtain n one-dimensional aggregation sequences;

[0013] 3) Vertically splice n one-dimensional aggregation sequences to obtain the decomposition results of each electrical equipment

[0014] In the above technical solution, further, in step S1: the original data set is D = {X, y 1 ,…,y i ,…y n}; X represents the high-frequency active power data of the total electric meter, y i (i=1,2,…,n) represents the high-frequency active power data of the i-th electrical equipment.

[0015] Furthermore, in step S2, data preprocessing is performed to delete, fill, resample and normalize the missing values ​​in the original data set to obtain a training data set D train And the test dataset D test .

[0016] Furthermore, in the step 1), the total electric meter load data is input into the time module, and firstly a fast Fourier transform is performed to obtain its periodicity.

[0017] A=Avg(Amp(FFT(X)))

[0018]

[0019] Among them, A represents the intensity of each frequency component in the total meter load data X, and the most significant k frequencies are selected (and arranged in order as f 1 ,…,f j ,…,f k ), that is, we get k cycles p 1 ,…,pj ,…,p k FFT stands for Fast Fourier Transform, Amp stands for the amplitude value of the signal after Fast Fourier Transform, Avg stands for the average value of these amplitude values; T stands for the total time length.

[0020] According to the obtained cycle, the load data is stacked to obtain k two-dimensional space-time tensors. The process can be described as:

[0021]

[0022] in, Represents the period p j The corresponding two-dimensional space-time tensor, Padding means expanding the time series along the time dimension, and Reshape means stacking the load data based on period and frequency.

[0023] Furthermore, in step 2), the method for obtaining the one-dimensional polymerization sequence is specifically:

[0024] 21), in the convolutional neural network, the convolution kernel is used to extract features from the two-dimensional spatiotemporal tensor, and k one-dimensional sequences can be obtained after residual connection and normalization.

[0025] 22) Adaptively aggregate the k one-dimensional sequences obtained in step 21) to obtain a one-dimensional aggregated sequence. The adaptive aggregation is specifically weighted summing of the k one-dimensional sequences.

[0026] Furthermore, in step S3, the MSE indicator is used as the loss function, the model is repeatedly trained until a specified number of times, and the model with the best verification result is selected. The loss function is specifically:

[0027]

[0028] Among them, y i Indicates the power value of electrical equipment i, It represents the power value of the electrical equipment i obtained by decomposing the power value of the total electric meter through the non-intrusive load decomposition model.

[0029] Furthermore, the non-intrusive load decomposition model is used to decompose the electric energy load, and the total load data of the electric meter is directly input into the trained non-intrusive load decomposition model to obtain the load decomposition value of each electrical equipment.

[0030] The beneficial effects of the present invention are:

[0031] The present invention realizes a high-precision non-invasive electric energy load decomposition method by combining fast Fourier transform and convolutional neural network technology, captures the relationship between the changes within the cycle and between cycles of load data, effectively improves the load feature extraction capability, reduces deployment costs, is applicable to a variety of power usage scenarios, helps optimize power system scheduling and energy efficiency management, and has significant economic and social benefits. The method of the present invention can adaptively aggregate periodic sequences, effectively improve the accuracy and adaptability of load decomposition, and provide a new technical approach for the refined management of power systems. The method of the present invention reconstructs the spatiotemporal characteristics of power data, so that the method can learn the relationship between the changes within the cycle and between cycles, deeply explore the user's power usage habits, and improve the accuracy of load decomposition. At the same time, since the identified power load periodicity corresponds to the power equipment, the influence of similar power appliances can be avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0033] Figure 1 It is a flow chart of the method for constructing the non-invasive load decomposition model of the present invention.

[0034] Figure 2 Schematic diagram of the training process of the time module of the non-intrusive load decomposition model.

[0035] Figure 3 It is a flow chart of electric energy load decomposition using a non-intrusive load decomposition model.

[0036] Figure 4 It is a schematic diagram of the process of reconstructing spatiotemporal features based on cycles. DETAILED DESCRIPTION

[0037] The present invention provides a non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology, which is implemented based on a non-intrusive load decomposition model. Figure 1 As shown, the method for constructing the non-intrusive load decomposition model includes the following steps:

[0038] Step S1, collecting load data, including high-frequency active power data of a total electric meter and n electric devices, and constructing an original data set, wherein the high-frequency active power data in the original data set is a one-dimensional time series;

[0039] Step S2, data preprocessing, deleting, filling, resampling and normalizing the missing values ​​in the original data set, and dividing it into a training data set and a test data set;

[0040] Step S3, training the non-intrusive load decomposition model, taking batches of the training data set, training each batch of training data, setting the MSE index as the loss function, measuring the loss of the decomposition results and the high-frequency active power data of n electrical equipment, and then adjusting the parameters; and using the test data set for verification, selecting the model with the best verification result as the trained model;

[0041] The training data of each batch is trained, and the training process is implemented based on the time module (such as Figure 2 ), the specific training process is:

[0042] 1) Feature enhancement: reshape the one-dimensional time series of the total meter load data in the training dataset into a two-dimensional space-time tensor;

[0043] 2) Feature extraction, using a convolutional neural network to deeply mine the data features of the two-dimensional spatiotemporal tensor to obtain n one-dimensional aggregation sequences;

[0044] 3) Vertically splice n one-dimensional aggregation sequences to obtain the decomposition results of each electrical equipment

[0045] The original data set in step S1 is D = {X, y 1 ,…,y i ,…y n}; X represents the total meter load active power data, y i (i=1,2,…,n) represents the decomposed active power data of the electrical equipment load.

[0046] After data preprocessing in step S2, the training data set D is obtained. train And the test dataset D test .

[0047] In step 1), the total meter load data is input into the time module, and firstly a fast Fourier transform is performed to obtain its periodicity, as shown in the following example: Figure 4 .

[0048] A=Avg(Amp(FFT(X)))

[0049]

[0050] Among them, A represents the intensity of each frequency component in the total meter load data X, and the most significant k frequencies are selected (and arranged in order as f 1 ,…,f j ,…,f k ), that is, we get k cycles p 1 ,…,p j ,…,p kFFT stands for Fast Fourier Transform, Amp stands for the amplitude value of the signal after Fast Fourier Transform, Avg stands for the average value of these amplitude values; T stands for the total time length.

[0051] According to the obtained cycle, the load data is stacked to obtain k two-dimensional space-time tensors. The process can be described as:

[0052]

[0053] in, Represents the period p j The corresponding two-dimensional space-time tensor, Padding means expanding the time series along the time dimension, and Reshape means stacking the load data based on period and frequency.

[0054] In step 2), the method for obtaining the one-dimensional polymerization sequence is specifically as follows:

[0055] 21), in the convolutional neural network, the convolution kernel is used to extract features from the two-dimensional spatiotemporal tensor, and k one-dimensional sequences can be obtained after residual connection and normalization.

[0056] 22) Adaptively aggregate the k one-dimensional sequences obtained in step 21) to obtain a one-dimensional aggregated sequence. The adaptive aggregation is specifically weighted summing of the k one-dimensional sequences.

[0057] In step S3: using the MSE indicator as the loss function, repeatedly train the model until the specified number of times, and select the model with the best verification result. The loss function is specifically:

[0058]

[0059] Among them, y i Indicates the power value of electrical equipment i, It represents the power value of the electrical equipment i obtained by decomposing the power value of the total electric meter through the non-intrusive load decomposition model.

[0060] By inputting the meter load data into the trained non-intrusive load decomposition model, the load decomposition value of each electrical equipment (including power and operating status) can be obtained.

[0061] like Figure 3 The flowchart of using the non-intrusive load decomposition model to decompose the electric energy load is shown, and the specific decomposition process is as follows:

[0062] The meter load data is input into the non-invasive load decomposition model. The entire model consists of multiple layers of stacked time modules. The time modules are used to perform feature enhancement and feature extraction on the meter load data. The deep features are extracted through multiple layers of time modules. The features extracted by the last layer of time modules are integrated and compressed by the fully connected layer to obtain the decomposition results of all electrical equipment, which can be output.

[0063] In each time module, the input sequence is processed in two ways: in one way, the input sequence is processed by fast Fourier transform (FFT) to obtain n periods, and the two-dimensional space-time vectors stacked by each period are convolved to mine the space-time features, and the obtained one-dimensional sequence is adaptively aggregated; in the other way, the input sequence is processed by the activation layer; finally, the two output results are connected with residuals to solve the gradient vanishing and gradient exploding problems in deep neural network training.

Claims

1. A non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology, characterized in that: Based on a non-intrusive load decomposition model, the construction method of the non-intrusive load decomposition model is: Step S1, collecting load data, including high-frequency active power data of a total electric meter and n electric devices, and constructing an original data set, wherein the high-frequency active power data in the original data set is a one-dimensional time series; Step S2, data preprocessing, deleting, filling, resampling and normalizing the missing values ​​in the original data set, and dividing it into a training data set and a test data set; Step S3, training the non-intrusive load decomposition model, taking batches of the training data set, training each batch of training data, setting the MSE index as the loss function, measuring the loss of the decomposition result and the high-frequency active power data of n electrical equipment, and then adjusting the parameters; The test data set is used for verification, and the model with the best verification result is selected as the trained model; The training of each batch of training data is performed based on the time module. The specific training process is as follows: 1) Feature enhancement: reshape the one-dimensional time series of the total meter load data in the training dataset into a two-dimensional space-time tensor; 2) Feature extraction, using a convolutional neural network to deeply mine the data features of the two-dimensional spatiotemporal tensor to obtain n one-dimensional aggregation sequences; 3) Vertically splice n one-dimensional aggregation sequences to obtain the decomposition results of each electrical equipment 2. According to the non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology of claim 1, it is characterized in that: In step S1, the original data set constructed is D = {X, y1, ..., y i ,…y n }; X represents the high-frequency active power data of the total electric meter, y i Represents the high-frequency active power data of the i-th electrical equipment, where i=1,2,…,n.

3. According to the non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology of claim 1, it is characterized in that: The step 1) comprises the following steps: First, the total meter load data is subjected to fast Fourier transform to calculate the frequency components in the load data; then, the most significant k frequencies are selected to obtain k periods; finally, the load data are stacked according to the obtained periods to obtain k two-dimensional space-time tensors.

4. The non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology according to claim 3 is characterized in that: In step 2), the acquisition of the one-dimensional polymerization sequence specifically includes the following steps: 21), in the convolutional neural network, the convolution kernel is used to extract features from the two-dimensional spatiotemporal tensor, and k one-dimensional sequences are obtained after residual connection and normalization; 22), adaptively aggregating the k one-dimensional sequences obtained in step 21) to obtain a one-dimensional aggregated sequence; the adaptive aggregation is specifically weighted summing of the k one-dimensional sequences.

5. The non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology according to claim 1 is characterized in that: In the step S3: during the training process of the non-invasive load decomposition model, the training is repeated until a specified number of times, and the model with the best verification result is selected.

6. The non-intrusive power load decomposition method based on spatiotemporal feature reconstruction technology according to claim 1 is characterized in that: The total load data of the electric meter is passed as input to the trained non-intrusive load decomposition model to obtain the load decomposition value of each electrical equipment.

Images