A non-intrusive load decomposition method based on adaptive graph convolutional network

Through the adaptive graph convolution network method, graph modeling and feature extraction are carried out in combination with temperature and user behavior characteristics, the problem of insufficient accuracy of non-invasive load decomposition technology is solved, and higher load decomposition accuracy and feature correlation are achieved extract.

CN114970313BActive Publication Date: 2024-11-12SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210343618.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-02
Publication Date
2024-11-12
Estimated Expiration
2042-04-02

AI Technical Summary

Technical Problem

The existing non-invasive load decomposition technology has limitations in accuracy, and does not fully utilize the temperature and user behavior characteristics, and the data input is limited to European space, so the correlation characteristics between data cannot be effectively mined.

Method used

Adaptive graph convolution network (AGCN) method is used to form a sequence of total load feature, combine temperature and user behavior characteristics for graph modeling, and use sliding window segmentation and Chebishev graph convolution to extract features to capture node features. The dynamic correlation of the two breaks the limitations of traditional European space.

Benefits of technology

It improves the accuracy of non-invasive load decomposition, can decompose electrical power more accurately, fully explore the correlation of load characteristics, and breaks through the limitations of single consideration of traditional load characteristics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114970313B_ABST
    Figure CN114970313B_ABST
Patent Text Reader

Abstract

The present invention is a non-intrusive load decomposition method based on an adaptive graph convolutional network, comprising the following steps: (1) forming a total load feature sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence; (2) performing sliding window segmentation on the formed total load feature sequence; (3) coupling user behavior characteristics, and performing graph modeling on each segmented total load feature sequence; (4) dividing the graph data and inputting it into an AGCN model for training; (5) performing state power threshold judgment on the output target electrical appliance power curve to obtain the electrical appliance state. The present invention forms a new sequence based on the collected traditional load feature sequence and temperature, performs graph modeling on the formed new sequence coupled with user behavior characteristics, and fully extracts the steady-state characteristics of the load through a Chebyshev graph convolutional network based on a spatial attention mechanism, thereby achieving a more accurate decomposition of the non-intrusive load.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of non-intrusive load decomposition, and in particular to a non-intrusive load decomposition method based on an adaptive graph convolutional network. Background Art

[0002] With the development of smart grids, user-side electricity management and energy efficiency analysis are becoming increasingly important. Load decomposition is a key link in realizing intelligent management on the user side. Compared with the intrusive type, the non-intrusive load decomposition only needs to analyze the data collected at the user's total meter end to obtain the electricity consumption of each appliance of the user, which has strong practical application value.

[0003] At present, the update of non-intrusive load decomposition model is mostly focused on the selection and extraction of load features and the optimization of load decomposition algorithm. Load features have evolved from power and voltage to VI trajectory diagrams and harmonics; feature extraction methods have evolved from statistical analysis and regression analysis to wavelet analysis and high-frequency filtering; load decomposition algorithms have evolved from simple mathematical models at the beginning to traditional machine learning methods and then to the recent hot deep learning. The accuracy of non-intrusive load decomposition is also constantly improving.

[0004] However, the selection of load characteristics still only focuses on the mining of traditional load electrical characteristics, lacking attention to multi-faceted characteristic data such as temperature and user behavior characteristics. The use of electrical appliances changes with different temperatures. For example, fans are used more in summer and are basically idle in winter; electric heaters are just the opposite, and their use time is mainly concentrated in winter. In addition, electrical appliances are used by users, who are the users and operators of electrical appliances. The user's behavioral habits also have a crucial impact on the operation of electrical appliances. For example, the use time of electrical appliances by working people is concentrated in the evening stage of weekdays and non-working days, while the use time of electrical appliances by families with infants and young children and the elderly is more uniform every day. Moreover, the data input of the current non-invasive load decomposition model is almost all sequential input of data in Euclidean space, which results in the use of convolution and other methods that can only map and extract the attributes of the data itself, lacking the full mining of the correlation characteristics between data. The existence of these problems has led to the limitation of the accuracy of current non-invasive load decomposition, and it is urgent to consider seeking new ways to perform non-invasive load decomposition to further break through its decomposition accuracy. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a non-intrusive load decomposition method based on an adaptive graph convolutional network, which can improve the accuracy of non-intrusive load decomposition and break the limitation of only considering traditional load characteristics and sequential input of data in Euclidean space.

[0006] The present invention provides a non-intrusive load decomposition method based on an adaptive graph convolutional network, comprising the following steps, characterized in that:

[0007] 1) Forming a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence: The steps of forming a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence are as follows:

[0008] Step 1: Use non-intrusive methods to collect the total load active power sequence, denoted as P = [p 1 ,p 2 ,...,p T ] T ; Collect the total load current sequence, recorded as I = [i 1 ,i 2 ,...,i T ] T . Collect the temperature sequence, denoted as T c =[t 1 ,t 2 ,...,t T ] T . Among them, 1, 2, …, T represents the duration of the acquisition sequence;

[0009] Step 2: The collected total load active power sequence P = [p 1 ,p 2 ,...,p T ] T 、Total load current sequence I=[i 1 ,i 2 ,...,i T ] T and temperature series T c =[t 1 ,t 2 ,...,t T ] T Splicing to form the total load characteristic sequence, recorded as

[0010] L=[l 1 , l 2 ,…, where l i =[p i ,i i ,t i ];

[0011] 2) Segment the formed total load characteristic sequence by sliding window;

[0012] The steps of performing sliding window segmentation on the formed total load characteristic sequence are as follows:

[0013] Step 1: Construct a sliding time window with a length of M;

[0014] Step 2: The total load characteristic sequence L = [l1 , l 2 ,…,l T ] T The sliding window moves forward one sampling point each time, and the first load sequence formed is recorded as L 1 =[l 1 , l 2 , …, 1 M ] T , sliding in sequence, and finally forming T-M+1 load sequence vectors L=[L 1 , L 2 , …, L T-M+1 ];

[0015] 3) Coupled with user behavior characteristics, graph modeling is performed on the characteristic sequence of each segmented total load;

[0016] The steps of graph modeling the segmented total load feature sequence for the coupled user behavior features are as follows:

[0017] Step 1: The total load characteristic sequence vector L = [L 1 , L 2 , …, L T-M+1 ] in each section of the total load characteristic sequence L i Each is constructed into an undirected graph G(X, A), where X is the feature matrix of the nodes in the graph and A is the adjacency matrix;

[0018] Step 2: Define the node characteristics, and express the node characteristics as the total load characteristic sequence formed at that moment, that is, the combination of the total load active power value, the total load current value and the temperature at that moment. The expression is:

[0019]

[0020] Where: M is the number of nodes, i.e., the total load characteristic sequence L of each segment i Length;

[0021] Step 3: Define the adjacency matrix. Time can best describe the characteristics of user behavior. Considering the characteristics of coupled user electricity consumption, each edge of the adjacency matrix of the input graph data is given a weight based on the time interval. The higher the weight, the shorter the time interval between the two nodes, and the closer the relationship between the two nodes.

[0022]

[0023] Where: t i represents the time when node i is located; t j represents the time when node j is located; σ is the scaling factor;

[0024] 4) Divide the graph data and input it into the AGCN model for training;

[0025] 5) Perform state power threshold judgment on the output target electrical appliance power curve to obtain the electrical appliance state.

[0026] As a further improvement of the present invention,

[0027] In step 4), the graph data is divided and input into the AGCN model for training steps:

[0028] Step 1: Divide the graph data into training set, validation set, and test set in a ratio of 6:2:2;

[0029] Step 2: Define the spatial attention module of the AGCN submodule to capture the dynamic correlation between node features. The specific expression is:

[0030] S=V s .σ(W s (X (r-1) ) T +b s )

[0031]

[0032] Where N is the number of nodes. is the input of the rth spatial attention module, C r-1 is the number of channels of the input data in the rth layer. When r = 1, C 0 =F, where F is the node feature dimension. is a learnable parameter, using sigmoidσ as the activation function, and using the adjacency matrix A and the spatial attention matrix when performing graph convolution Dynamically adjust the influence weights between nodes;

[0033] Step 3: Define the Chebyshev graph convolution of the AGCN submodule for feature extraction. The specific expression is:

[0034]

[0035] in, L is the Laplace matrix, A is the adjacency matrix, is the degree matrix, D ii =∑ j A ij , I N is the identity matrix, and the parameters is the vector of polynomial coefficients, is the largest eigenvalue of the Laplace matrix, and the recursive definition of the Chebyshev polynomial is T k (x) = 2xT k-1(x)-T k-2 (x), where T 0 (x) = 1, T 1 (x) = x;

[0036] Step 4: Perform layer normalization on the output of the Chebyshev graph convolution layer. Therefore, the training set graph data is input in parallel to the AGCN submodule for feature extraction. The specific expression is:

[0037]

[0038]

[0039] Among them, ⊙ is the Hadamard product, gθ*Gx′ represents the result of layer normalization of gθ*Gx, E[gθ*Gx] and Var[gθ*Gx] represent the mean and variance of gθ*Gx respectively, ε takes the default value 1e-5, and γ and β are trainable parameters;

[0040] Step 5: Input the final output of the graph convolution module into the pooling layer, use graph average pooling to compress the data features and focus on effective information. The specific calculation formula is:

[0041]

[0042] Step 6: Use the fully connected layer to realize the nonlinear mapping of deep features to target electrical power using the pooling result. The specific calculation formula is:

[0043]

[0044] Among them, W (k-1) is the weight matrix of the kth fully connected layer; b (k-1) is the bias matrix of the kth fully connected layer;

[0045] Step 7: Through the above network model, the decomposition sequence output in represents the decomposition result of the i-th graph data, that is, the target electrical appliance power value at the i-th graph point. Compared with the true value, the Adam optimizer is used to back-propagate and correct the network parameters. At the same time, an early stopping mechanism is added to the validation set to train the model.

[0046] As a further improvement of the present invention, the step of determining the state power threshold of the output target electrical appliance power curve to obtain the electrical appliance state in step 5) is as follows:

[0047] Step 1: For the optimized final network model, use the test set to test the model and finally apply it in the actual non-intrusive load decomposition work;

[0048] Step 2: The output target electrical appliance power curve is used to set the state threshold by using the actual state power value of the electrical appliance for state judgment, and the state of the electrical appliance at the corresponding moment can be obtained.

[0049] The innovation and application prospects of this method are as follows: (1) When forming the total load feature sequence, in addition to considering the traditional load features of active power and current, temperature features are also added, making the feature sequence more comprehensive and diverse; (2) The total load feature sequence formed is graph-modeled, breaking the limitation of sequential input of data in Euclidean space, considering the connection relationship between nodes, and fully mining the correlation features between the total load feature sequence; (3) When graph modeling, the node features include traditional load features and temperature, the adjacency matrix considers user behavior characteristics, and couples traditional load characteristics, temperature and user behavior characteristics. Compared with considering traditional load characteristics alone, feature extraction is more accurate and comprehensive; (4) The dynamic correlation between node features is captured through the spatial attention mechanism, and the Chebyshev graph convolutional network is used to automatically learn the attribute characteristics of the node itself and the correlation information between node features, fully extracting the steady-state characteristics of the load and achieving a more accurate decomposition of non-intrusive loads. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Load decomposition process;

[0051] Figure 2 Load decomposition model based on AGCN. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0053] The present invention provides a non-intrusive load decomposition method based on an adaptive graph convolutional network, which can improve the accuracy of non-intrusive load decomposition and break the limitation of only considering the traditional load characteristics and the sequential input of data in the Euclidean space. As an embodiment of the present invention, the present invention provides a non-intrusive load decomposition method based on AGCN, wherein the load decomposition flow chart is as follows Figure 1 As shown in Figure 2, the load decomposition model based on AGCN is as follows: Figure 2 As shown, and achieved by the following steps:

[0054] 1) Form a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence:

[0055] The steps of forming a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence are as follows:

[0056] Step 1: Use non-intrusive methods to collect the total load active power sequence, denoted as P = [p1 ,p 2 ,...,p T ] T ; Collect the total load current sequence, recorded as I = [i 1 ,i 2 ,...,i T ] T . Collect the temperature sequence, denoted as T c =[t 1 ,t 2 ,...,t T ] T . Among them, 1, 2, …, T represents the duration of the acquisition sequence;

[0057] Step 2: The collected total load active power sequence P = [p 1 ,p 2 ,...,p T ] T 、Total load current sequence I=[i 1 ,i 2 ,...,i T ] T and temperature series T c =[t 1 ,t 2 ,...,t T ] T Splicing to form the total load characteristic sequence, recorded as

[0058] L=[l 1 , l 2 ,…, where l i =[p i ,i i ,t i ];

[0059] 2) Segment the formed total load characteristic sequence by sliding window;

[0060] The steps of performing sliding window segmentation on the formed total load characteristic sequence are as follows:

[0061] Step 1: Construct a sliding time window with a length of M;

[0062] Step 2: The total load characteristic sequence L = [l 1 , l 2 ,…,l T ] T The sliding window moves forward one sampling point each time, and the first load sequence formed is recorded as L 1 =[l 1 , l 2 , …, 1 M ] T, sliding in sequence, and finally forming T-M+1 load sequence vectors L=[L 1 , L 2 , …, L T-M+1 ];

[0063] 3) Coupled with user behavior characteristics, graph modeling is performed on the characteristic sequence of each segmented total load;

[0064] The steps of graph modeling the segmented total load feature sequence for the coupled user behavior features are as follows:

[0065] Step 1: The total load characteristic sequence vector L = [L 1 , L 2 , …, L T-M+1 ] in each section of the total load characteristic sequence L i Each is constructed into an undirected graph G(X, A), where X is the feature matrix of the nodes in the graph and A is the adjacency matrix;

[0066] Step 2: Define the node characteristics, and express the node characteristics as the total load characteristic sequence formed at that moment, that is, the combination of the total load active power value, the total load current value and the temperature at that moment. The expression is:

[0067]

[0068] Where: M is the number of nodes, i.e., the total load characteristic sequence L of each segment i Length;

[0069] Step 3: Define the adjacency matrix. Time can best describe the characteristics of user behavior. Considering the characteristics of coupled user electricity consumption, each edge of the adjacency matrix of the input graph data is given a weight based on the time interval. The higher the weight, the shorter the time interval between the two nodes, and the closer the relationship between the two nodes.

[0070]

[0071] Where: t i represents the time when node i is located; t j represents the time when node j is located; σ is the scaling factor;

[0072] 4) Divide the graph data and input it into the AGCN model for training;

[0073] The steps of dividing the graph data and inputting it into the AGCN model for training are as follows:

[0074] Step 1: Divide the graph data into training set, validation set, and test set in a ratio of 6:2:2;

[0075] Step 2: Define the spatial attention module of the AGCN submodule to capture the dynamic correlation between node features. The specific expression is:

[0076] S=V s .σ(W s (X (r-1) ) T +b s )

[0077]

[0078] Where N is the number of nodes. is the input of the rth spatial attention module, C r-1 is the number of channels of the input data in the rth layer. When r = 1, C 0 =F, where F is the node feature dimension. is a learnable parameter, using sigmoidσ as the activation function, and using the adjacency matrix A and the spatial attention matrix when performing graph convolution Dynamically adjust the influence weights between nodes;

[0079] Step 3: Define the Chebyshev graph convolution of the AGCN submodule for feature extraction. The specific expression is:

[0080]

[0081] in, L is the Laplace matrix, A is the adjacency matrix, is the degree matrix, D ii =∑ j A ij , I N is the identity matrix, and the parameters is the vector of polynomial coefficients, is the largest eigenvalue of the Laplace matrix, and the recursive definition of the Chebyshev polynomial is T k (x) = 2xT k-1 (x)-T k-2 (x), where T 0 (x) = 1, T 1 (x) = x;

[0082] Step 4: Perform layer normalization on the output of the Chebyshev graph convolution layer. Therefore, the training set graph data is input in parallel to the AGCN submodule for feature extraction. The specific expression is:

[0083]

[0084]

[0085] Among them, ⊙ is the Hadamard product, gθ*Gx′ represents the result of layer normalization of gθ*Gx, E[gθ*Gx] and Var[gθ*Gx] represent the mean and variance of gθ*Gx respectively, ε takes the default value 1e-5, and γ and β are trainable parameters;

[0086] Step 5: Input the final output of the graph convolution module into the pooling layer, use graph average pooling to compress the data features and focus on effective information. The specific calculation formula is:

[0087]

[0088] Step 6: Use the fully connected layer to realize the nonlinear mapping of deep features to target electrical power using the pooling result. The specific calculation formula is:

[0089]

[0090] Among them, W (k-1) is the weight matrix of the kth fully connected layer; b (k-1) is the bias matrix of the kth fully connected layer;

[0091] Step 7: Through the above network model, the decomposition sequence output in represents the decomposition result of the i-th graph data, that is, the target electrical appliance power value at the i-th graph point. Compared with the true value, the Adam optimizer is used to back-propagate and correct the network parameters. At the same time, an early stopping mechanism is added to the validation set to train the model.

[0092] 5) Perform state power threshold judgment on the output target electrical appliance power curve to obtain the electrical appliance state.

[0093] The steps of determining the state power threshold of the output target electrical appliance power curve to obtain the electrical appliance state are as follows:

[0094] Step 1: For the optimized final network model, use the test set to test the model and finally apply it in the actual non-intrusive load decomposition work;

[0095] Step 2: The output target electrical appliance power curve is used to set the state threshold by using the actual state power value of the electrical appliance for state judgment, and the state of the electrical appliance at the corresponding moment can be obtained.

[0096] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A non-intrusive load decomposition method based on an adaptive graph convolutional network comprises the following steps, characterized in that: 1) Form a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence: The steps of forming a total load characteristic sequence consisting of a total load active power sequence, a total load current sequence and a temperature sequence are as follows: Step 1: Use a non-intrusive method to collect the total load active power sequence, denoted as P = [p1, p2, ..., p T ] T ; Collect the total load current sequence, recorded as I = [i1, i2, ..., i T ] T ; Collect temperature sequence, denoted as T c =[t1, t2, ..., t T ] T ; Wherein, 1, 2, ..., T represents the duration of the acquisition sequence; Step 2: The collected total load active power sequence P = [p1, p2, ..., p T ] T 、Total load current sequence I=[i1,i2,...,i T ] T and temperature series T c =[t1, t2, ..., t T ] T The total load characteristic sequence is formed by splicing, which is recorded as L = [l1, l2, ..., where l i =[p i ,i i , t i ]; 2) Segment the formed total load characteristic sequence by sliding window; The steps of performing sliding window segmentation on the formed total load characteristic sequence are as follows: Step 1: Construct a sliding time window with a length of M; Step 2: The total load characteristic sequence L = [l1, l2, ..., l T ] T The sliding window moves forward one sampling point each time, and the first load sequence formed is recorded as L1 = [l1, l2, ..., I M ] T , sliding in sequence, and finally forming T-M+1 load sequence vectors L=[L1, L2, …, L T-M+1 ]; 3) Coupled with user behavior characteristics, graph modeling is performed on the characteristic sequence of each segmented total load; The steps of graph modeling the segmented total load feature sequence for the coupled user behavior features are as follows: Step 1: The total load characteristic sequence vector L = [L1, L2, ..., L T-M+1 ] in each section of the total load characteristic sequence L i Each is constructed into an undirected graph G(X, A), where X is the feature matrix of the nodes in the graph and A is the adjacency matrix; Step 2: Define the node characteristics, and express the node characteristics as the total load characteristic sequence formed at that moment, that is, the combination of the total load active power value, the total load current value and the temperature at that moment. The expression is: Where: M is the number of nodes, i.e., the total load characteristic sequence L of each segment i Length; Step 3: Define the adjacency matrix. Time can best describe the characteristics of user behavior. Considering the characteristics of coupled user electricity consumption, each edge of the adjacency matrix of the input graph data is given a weight based on the time interval. The higher the weight, the shorter the time interval between the two nodes, and the closer the relationship between the two nodes. Where: t i represents the time when node i is located; t j represents the time when node j is located; σ is the scaling factor; 4) Divide the graph data and input it into the AGCN model for training; 5) Perform state power threshold judgment on the output target electrical appliance power curve to obtain the electrical appliance state.

2. According to claim 1, a non-intrusive load decomposition method based on an adaptive graph convolutional network is characterized in that: In step 4), the graph data is divided and input into the AGCN model for training steps: Step 1: Divide the graph data into training set, validation set, and test set in a ratio of 6:2:2; Step 2: Define the spatial attention module of the AGCN submodule to capture the dynamic correlation between node features. The specific expression is: S=V s .σ(W s (X (r-1) ) T +b s ) Where N is the number of nodes; is the input of the rth spatial attention module, C r-1 is the number of channels of the input data in the rth layer; when r = 1, C0 = F, F is the node feature dimension; V s , is a learnable parameter, using sigmoidσ as the activation function, and using the adjacency matrix A and the spatial attention matrix when performing graph convolution Dynamically adjust the influence weights between nodes; Step 3: Define the Chebyshev graph convolution of the AGCN submodule for feature extraction. The specific expression is: in, L is the Laplace matrix, A is the adjacency matrix, is the degree matrix, D ii =∑ j A ij , I N is the identity matrix, and the parameters is the vector of polynomial coefficients, is the largest eigenvalue of the Laplace matrix, and the recursive definition of the Chebyshev polynomial is T k (x) = 2xT k-1 (x)-T k-2 (x), where T0(x) = 1, T1(x) = x; Step 4: Normalize the output of the Chebyshev graph convolution layer; therefore, the training set graph data is input in parallel into the AGCN submodule for feature extraction. The specific expression is: Among them, ⊙ is the Hadamard product, gθ*Gx′ represents the result of layer normalization of gθ*Gx, E[gθ*Gx] and Var[gθ*Gx] represent the mean and variance of gθ*Gx respectively, ε takes the default value 1e-5, and γ and β are trainable parameters; Step 5: Input the final output of the graph convolution module into the pooling layer, use graph average pooling to compress the data features and focus on effective information. The specific calculation formula is: Step 6: Use the fully connected layer to realize the nonlinear mapping of deep features to target electrical power using the pooling result. The specific calculation formula is: Among them, W (k-1) is the weight matrix of the kth fully connected layer; b (k-1) is the bias matrix of the kth fully connected layer; Step 7: Through the above network model, the decomposition sequence output in represents the decomposition result of the i-th graph data, that is, the target electrical appliance power value at the i-th graph point; then Compared with the true value, the Adam optimizer is used for back propagation to correct the network parameters; at the same time, an early stopping mechanism is added to the validation set to train the model.

3. The non-intrusive load decomposition method based on an adaptive graph convolutional network according to claim 1, characterized in that: In step 5), the steps of determining the state power threshold of the output target electrical appliance power curve to obtain the electrical appliance state are as follows: Step 1: For the optimized final network model, use the test set to test the model and finally apply it in the actual non-intrusive load decomposition work; Step 2: The output target electrical appliance power curve is used to set the state threshold by using the actual state power value of the electrical appliance for state judgment, and the state of the electrical appliance at the corresponding moment can be obtained.