A non-intrusive load decomposition method combining external attention mechanism and Seq2Point

Abstract

The present invention discloses a non-invasive load decomposition method that combines an external attention mechanism with Seq2Point. The present invention proposes a Seq2Point non-invasive load decomposition method that combines an external attention mechanism, which achieves an improvement in decomposition accuracy without increasing the training time cost. First, the collected power sequence is preprocessed and window data is constructed; then, the window data is processed by an encoder to obtain the original feature code; then, the external attention mechanism is used to correct the original feature code to obtain the corrected feature code; the original feature code and the corrected feature code are residually connected to obtain a comprehensive feature code; finally, the comprehensive feature code is decoded by a decoder to achieve non-invasive load decomposition. This method can utilize low-frequency power sequence information, spend a small amount of training time, and achieve accurate non-invasive load decomposition. This method can provide assistance for further mining user electricity consumption information and provide a reference for grid demand-side response.

Description

Technical Field

[0001] The present invention belongs to the field of non-intrusive load decomposition detection and relates to a non-intrusive load decomposition method combining an external attention mechanism with Seq2Point. Background Art

[0002] Non-Intrusive Load Disaggregation (NILD) obtains total electricity usage information from smart meters at the entrance of a specific area and then decomposes it using certain technical means to obtain user electricity usage and energy consumption data. This decomposed information is of great significance for equipment fault monitoring, demand-side response, and data analysis. Compared with traditional intrusive load disaggregation (ILD), NILD not only reduces the number of sensor devices but also eliminates the need for modification and maintenance of existing power lines, making it a low-cost and easy-to-implement decomposition method. NILD was first proposed by Professor Hart in the 1980s, but due to limitations in computing power and insufficient model accuracy, it did not attract widespread attention. However, in the past decade, with the rapid advancement of smart grid construction worldwide and the rise of artificial intelligence technologies represented by companies such as NVIDIA and Google, power systems have begun to shift towards cyber-physical convergence systems powered by big data and the Internet of Things. The large number of smart meters deployed at the user end not only performs traditional electricity billing functions but also provides a vast amount of electricity data. Summary of the Invention

[0003] NILD is a typical time series analysis problem and can be divided into two categories: decomposition based on transient characteristic signals and decomposition based on steady-state characteristic signals. Decomposition based on transient characteristic signals usually uses high-frequency sampling data. Methods based on steady-state characteristic signals usually use low-frequency sampling data, usually power, voltage, current and other data collected at low frequencies. Some scholars have applied deep learning to NILD and achieved good results. Some models use the Seq2Seq method for non-invasive load decomposition, while some models use the attention mechanism for load decomposition. However, there is a contradiction between the training time cost and the model decomposition accuracy. The inventors proposed a Seq2Point non-invasive load decomposition method combined with an external attention mechanism to improve the accuracy of non-invasive load decomposition without increasing the training time cost.

[0004] The Seq2Point non-invasive load decomposition method proposed in the present invention in combination with an external attention mechanism includes the following steps:

[0005] Step 1: Preprocessing of input data and preparation of window data;

[0006] Step 1-1: Data collation;

[0007] Downsample all data to reduce the data frequency and align the timestamps of device power data with total load data;

[0008] Step 1-2: Data standardization;

[0009] use All load data are normalized, where x' t Represents the result after preprocessing, x t Indicates the power reading of the acquisition device at time t, represents the mean value of the power reading sequence collected, and σ is the standard deviation of the power reading sequence collected;

[0010] Step 1-3: Construct window data;

[0011] The data sequence from time t to time t+s-1 after normalization is taken as a window data, which is expressed as z={x' t ,x' t+1 ,x' t+2 ,...,x' t+s-1}, s represents the window length.

[0012] Step 2: Build a non-intrusive payload decomposition model that combines external attention mechanism and Seq2Point for non-intrusive payload decomposition;

[0013] The window data is feature-encoded using an encoder based on a one-dimensional convolutional layer to obtain original feature coding information; the original feature coding information is corrected using an external attention mechanism to obtain corrected feature coding information; the original feature coding information and the corrected feature coding information are residually connected to obtain comprehensive feature coding information; finally, the comprehensive feature coding information is decoded by a decoder based on a linear layer to obtain the equipment load decomposition value O, thereby realizing non-intrusive load decomposition from window data z to load decomposition value O.

[0014] Step 2-1: Use an encoder based on a one-dimensional convolutional layer to encode the window data, expressed as C = Encode(z);

[0015] Encode(z)=[Conv1D(z)] k , k=1,2,3,...,K, which means that the window data z undergoes k one-dimensional convolution layers for continuous change, rather than the window data z undergoing k powers after passing through one convolution layer;

[0016] Conv1D(z)=(c1,c2,...,c i ,...,c s-l-1 ), ci The eigenvector of the ith position, c i =f(kernal e ·x' i:i+l-1 +b), b is the bias vector, kernal e represents the e-th convolution kernel of size 1×l, Represents the convolution kernel kernal e The eigenvalue of the ith position, x' i:i+l-1 Represents a vector (x' i ,x' i+1 ,x' i+2 ,...,x' i+l-1 ), f represents the activation function;

[0017] Step 2-2: Use the external attention mechanism to correct the original feature encoding information to obtain the corrected feature encoding information, expressed as E = EA(C);

[0018] First, perform a Query operation on C, expressed as Q = Query (C); then use the external memory unit M k Process Q, that is, R = M k (Q), then normalize R to get S, that is, S = Noraml (R), and finally use the external memory unit M v Process S to obtain the modified feature coding information E, that is, E = M v (S);

[0019] The query operation is implemented through the convolution layer;

[0020] M k (·) and M v () is implemented through the linear layer; the linear layer is represented as Y = f(WX + b), which is the input of the X linear layer and W is the weight matrix;

[0021] The calculation process of Noraml(·) normalization operation is: in is the data at the jth column position in the i-th row of the input initial data matrix, For the matrix The result of normalization operation on the rows; i,j For the matrix The column is standardized;

[0022] Step 2-3: Perform residual connection on the original feature coding information and the modified feature coding information to obtain comprehensive feature coding information, expressed as F=E+C;

[0023] Step 2-4: Decode the comprehensive feature encoding information using a linear layer-based decoder, expressed as O = Decode(F), to achieve non-intrusive load decomposition from the window data z to the load decomposition value O;

[0024] Decode(F)=[Y=f(WF+b)] u , u=1,2,3,...,U, which means that the comprehensive feature coding information F is continuously changed through u linear layers to obtain the decoded information, rather than F being raised to the uth power after passing through one linear layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flowchart for implementing non-intrusive load decomposition;

[0026] Figure 2 The specific structure of the Seq2Point non-intrusive load decomposition model combined with the external attention mechanism;

[0027] Figure 3 Schematic diagram of the structure of the external attention mechanism;

[0028] Figure 4 Schematic diagram of the decomposition effects of the Seq2Point method, CBAM method and this method. DETAILED DESCRIPTION

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Reference Figure 1 , a non-intrusive load decomposition method combining external attention mechanism with Seq2Point, including the following steps:

[0031] This example is based on the dataset REDD, which is a public dataset used to verify load decomposition;

[0032] Step 1: Data collation;

[0033] The data is downsampled to 0.125 Hz, i.e. one data point every 8 seconds, and the timestamp is aligned with the total load data;

[0034] Step 2: Data normalization;

[0035] use All load data are normalized, where x' t Represents the result after preprocessing, x t Indicates the power reading of the acquisition device at time t, represents the mean value of the power reading sequence collected, and σ is the standard deviation of the power reading sequence collected;

[0036] Step 3: Construct window data;

[0037] The window length s is set to 599 to generate window data, and the window data z={x' t ,x' t+1 ,x' t+2 ,...,x' t+s-1};

[0038] Step 4: Set up the encoder;

[0039] The encoder consists of 5 one-dimensional convolutional layers. The first layer has 30 convolution kernels with a convolution kernel size of 1×10; the second layer has 40 convolution kernels with a convolution kernel size of 1×8; the third layer has 40 convolution kernels with a convolution kernel size of 1×6; the fourth layer has 40 convolution kernels with a convolution kernel size of 1×5; the fifth layer has 50 convolution kernels with a convolution kernel size of 1×5, and the activation function is ReLU.

[0040] Step 5: Set the Query operation;

[0041] Query is a one-dimensional convolution layer with a convolution kernel size of 1×1 and 50 convolution kernels.

[0042] Step 6: Set up the external memory unit;

[0043] M k The weight matrix size of () is 570×32, M v The weight matrix size of () is 32×570, and the activation function is ReLU;

[0044] Step 7: Set up the decoder;

[0045] The decoder consists of two linear layers. The W size of the first linear layer is 570×1024, and the activation function of the first layer is ReLU. The W size of the second linear layer is 1024×1, and no activation function is set for the second layer.

[0046] Step 8: Train the model;

[0047] A model that can be used for non-invasive load decomposition is obtained through training;

[0048] Step 9: Input the window data into the model to perform non-intrusive load decomposition. The evaluation indicators of the decomposition effect are the normalized signal aggregation error (SAE), mean absolute error (MAE), and normalized disaggregation error (NDE). The expressions of each evaluation indicator are as follows:

[0049]

[0050] Where: Output decomposition value at time t; x t Input real data at time t; r = ∑ t x t is the total energy consumption of electrical equipment; is the total energy consumption output by the model;

[0051] The decomposition results of this paper are compared with the current mainstream non-intrusive load decomposition methods (Seq2Point's non-intrusive load decomposition method, T2V's non-intrusive load decomposition method, and CBAM's non-intrusive load decomposition method). The comparison results are shown in Tables 1, 2, and 3.

[0052] Table 1 Comparison of SAE indicators of various methods

[0053]

[0054] Table 2 Comparison of MAE indicators of various methods

[0055]

[0056] Table 3 Comparison of NDE indicators of various methods

[0057]

[0058] The smaller the value of the SAE index, the higher the decomposition accuracy. In terms of SAE accuracy, this method is significantly superior to the non-invasive load decomposition methods of Seq2Point, T2V, and CBAM.

[0059] The smaller the value of the MAE indicator, the better the decomposition effect. This method has better decomposition effects on refrigerators and dishwashers than other methods, and the average decomposition effect is also better than other methods.

[0060] The smaller the value of the NDE index, the better the decomposition effect. The decomposition effect of this method on refrigerators and dishwashers is better than other methods, and the average decomposition effect is also better than other methods.

[0061] Taking the refrigerator as an example, the Seq2Point, CBAM and decomposition results of this method under the same window are plotted as follows: Figure 4 As shown, this method can better decompose the power curve of the refrigerator.

Claims

1. A non-intrusive load decomposition method combining an external attention mechanism with Seq2Point, comprising the following steps: Step 1: Preprocessing of input data and preparation of window data; In order to standardize the input data and expand the scope of application of the decomposition method, the data is processed through three operations: data normalization, data downsampling, and window data construction, considering the inconsistency of the power data sequence numerical range and the inconsistency of the power data sampling frequency. Step 2: Build a non-intrusive payload decomposition model that combines external attention mechanism and Seq2Point for non-intrusive payload decomposition; The window data is feature-encoded using an encoder based on a one-dimensional convolutional layer to obtain the original feature encoding information; The original feature coding information is corrected using the external attention mechanism to obtain the corrected feature coding information; the original feature coding information and the corrected feature coding information are residually connected to obtain the comprehensive feature coding information; Finally, the comprehensive feature encoding information is decoded by a linear layer-based decoder to obtain the equipment load decomposition value O, realizing non-intrusive load decomposition from window data z to load decomposition value O.

2. The non-intrusive load decomposition method combining external attention mechanism and Seq2Point according to claim 1 is characterized in that In step 1, the steps of data standardization, data downsampling and window data construction are as follows: Step 1-1: Data collation; To unify the data frequency and ensure consistency between the device power data and the total load data in subsequent processing, all data are downsampled, the data frequency is reduced, and the timestamps of the device power data and the total load data are aligned; Step 1-2: Data standardization; In order to analyze the power data and total load data of different devices at the same scale and improve the accuracy and stability of non-intrusive load decomposition, the formula All load data are normalized, where x' t Represents the result after preprocessing, x t Indicates the power reading of the acquisition device at time t, represents the mean value of the power reading sequence collected, and σ is the standard deviation of the power reading sequence collected; Step 1-3: Construct window data; In order to meet the model input requirements, the data sequence from time t to time t+s-1 after normalization is taken as a window data, which is expressed as z = {x' t ,x' t+1 ,x' t+2 ,...,x' t+s-1 }, s represents the window length.

3. The non-intrusive load decomposition method combining external attention mechanism and Seq2Point according to claim 1 is characterized in that In step 2, the steps of constructing a non-intrusive load decomposition model combining an external attention mechanism and Seq2Point for non-intrusive load decomposition are as follows: Step 2-1: Use an encoder based on a one-dimensional convolutional layer to perform feature encoding on the window data to obtain the original feature encoding information, which is expressed as C = Encode(z); Encode(z)=[Conv1D(z)] k , k=1,2,3,...,K, which means that the window data z undergoes k one-dimensional convolution layers for continuous change, rather than the window data z undergoing k powers after passing through one convolution layer; Conv1D(z)=(c1,c2,...,c i ,...,c s-l-1 ), c i The eigenvector of the ith position, c i =f(kernal e ·x' i:i+l-1 +b), b is the bias vector, kernal e represents the e-th convolution kernel of size 1×l, Represents the convolution kernel kernal e The eigenvalue of the ith position, x' i:i+l-1 Represents a vector (x' i ,x' i+1 ,x' i+2 ,...,x' i+l-1 ), f represents the activation function; Step 2-2: Use the external attention mechanism to correct the original feature encoding information to obtain the corrected feature encoding information, expressed as E = EA(C); First, perform a Query operation on C, expressed as Q = Query (C); then use the external memory unit M k Process Q, that is, R = M k (Q), then normalize R to get S, that is, S = Noraml (R), and finally use the external memory unit M v Process S to obtain the modified feature coding information E, that is, E = M v (S); The query operation is implemented through the convolution layer; M k () and M v () is implemented through the linear layer; the linear layer is represented as Y = f(WX + b), which is the input of the X linear layer and W is the weight matrix; The calculation process of Noraml(·) normalization operation is: in is the data at the jth column position in the i-th row of the input initial data matrix, For the matrix The result of normalization operation on the rows; i,j For the matrix The column is standardized; Step 2-3: Perform residual connection on the original feature coding information and the modified feature coding information to obtain comprehensive feature coding information, expressed as F=E+C; Step 2-4: Decode the comprehensive feature encoding information using a linear layer-based decoder, expressed as O = Decode(F), to achieve non-intrusive load decomposition from the window data z to the load decomposition value O; Decode(F)=[Y=f(WF+b)] u , u=1,2,3,...,U, which means that the comprehensive feature coding information F is continuously changed through u linear layers to obtain the decoded information, rather than F being raised to the uth power after passing through one linear layer.

Images