A non-intrusive load identification method based on multi-dimensional features

Abstract

The present invention discloses a non-intrusive load identification method based on multi-dimensional features, which includes: obtaining total electrical appliance data within a target time period and performing preprocessing; inputting the preprocessed data into a pre-built non-intrusive load identification model to perform load identification and obtain identification results; due to the low frequency of minute-level meter data collection, it increases the difficulty of capturing the characteristics of residential electricity loads, which seriously affects the final load identification results. By introducing the physical characteristics of household appliances and the social characteristics of residential users, combined with the gating mechanism and the attention mechanism, the ability to analyze and mine residential electricity load characteristics from minute-level meter data is enhanced, and the identification accuracy of the non-intrusive load identification method on minute-level data is improved.

Description

Technical Field

[0001] The invention relates to a non-intrusive load identification method based on multi-dimensional features, and belongs to the technical field. Background Art

[0002] Under the background of vigorously developing the global energy Internet, my country is promoting the construction of smart grids in an orderly manner. Realizing two-way interactive smart electricity consumption has gradually become the main development goal of the power system. The key to developing smart electricity consumption lies in obtaining power users' load monitoring information and intelligent analysis of users' electricity consumption behavior. Residential users have the characteristics of diverse electricity consumption behaviors and a large number of users, and are the main service group of the two-way interactive service system for smart electricity consumption. Although my country's smart grid construction is becoming increasingly perfect, due to the limitations of technology and other factors, my country has not carried out large-scale research on load identification technology on the residential user side. Conventional electricity analysis technology can no longer meet the needs of flexible and interactive smart electricity consumption. Residential electricity load identification and intelligent analysis technology urgently need to be improved and innovated.

[0003] The load identification of residential electrical equipment is mainly achieved through two methods: invasive load monitoring and non-invasive load monitoring. Intrusive load monitoring refers to installing hardware monitoring equipment on the power load to obtain the operating status and power consumption of each equipment load in real time. Although this method accurately realizes load monitoring, it consumes a lot of manpower and financial resources in the early stage and affects the lives of users, and is not easy to maintain. Non-invasive load monitoring refers to installing a non-invasive load decomposition device directly on the main circuit to obtain the current, voltage and power information of the total power load, and identify the load operating status of each electrical equipment in the residential user based on this information. This method does not need to go deep into the user's residence, but only installs the collection equipment outside. It is relatively simple and convenient, and has become the mainstream direction of load monitoring research in recent years. However, in actual applications, the existing conventional non-invasive load monitoring method is achieved by modifying or adding monitoring equipment, which faces the difficulties of high investment, high cost, wide scope, and difficulty in promotion. It also faces the problems that the accuracy and timeliness of intelligent analysis cannot meet actual needs. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a non-intrusive load identification method based on multi-dimensional features to solve the problems that the accuracy and timeliness of existing intelligent analysis cannot meet actual needs.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] The present invention provides a non-intrusive load identification method based on multi-dimensional features, comprising:

[0007] Obtain the total electrical appliance data within the target time period and perform preprocessing;

[0008] The preprocessed data is input into a pre-built non-intrusive load identification model to perform load identification and obtain identification results;

[0009] The construction process of the non-intrusive load identification model includes:

[0010] Construct inputs for non-intrusive load identification models based on the physical characteristics of electrical equipment;

[0011] Construct a subnet model for classifying the switch status of electrical equipment based on electrical characteristics and gating mechanisms;

[0012] Construct an electrical equipment load regression subnetwork model based on electrical features and attention mechanism;

[0013] Construct a user social feature regression subnetwork model based on social features and sliding window mechanism;

[0014] Combining the outputs of the electrical equipment switch state classification subnet model, the electrical equipment load regression subnet model, and the user social characteristics regression subnet model, a non-intrusive load identification model for input is constructed;

[0015] Obtain electrical characteristic data, physical characteristic data of electrical equipment and social characteristic data of users;

[0016] Label the electrical characteristic data, physical characteristic data and social characteristic data respectively and generate training sets;

[0017] The non-intrusive load identification model is trained and parameter optimized through the training set to obtain the final non-intrusive load identification model.

[0018] Optionally, the physical characteristics include:

[0019] The power consumption equipment whose operation status change is lower than the preset value, whose single operation cycle is higher than the preset value and whose power adjustment amplitude is less than the preset value is recorded as 0;

[0020] Electrical equipment whose operating status change is greater than or equal to the preset value, whose single operating cycle is less than or equal to the preset value, and whose power switching amplitude is greater than or equal to the preset value is recorded as 1.

[0021] Optionally, the input for constructing the non-intrusive load identification model based on the physical characteristics of the electrical equipment includes:

[0022] If the physical feature is 0, a sequence-to-sequence model is used, which predicts the power of the electrical equipment at a single moment multiple times in a sliding window manner and calculates the average value, and uses the predicted average value as the input of the non-intrusive load identification model;

[0023] If the acquired label information is 1, a sequence-to-point model is used, and the sequence-to-point model predicts the power of the electrical equipment at a single moment only once in a sliding window manner, and uses the predicted value as the input of the non-intrusive load identification model.

[0024] Optionally, the electrical characteristics include:

[0025] Use a standardized method to convert the power of the electrical equipment at the corresponding moment into a distribution with a mean of 0 and a standard deviation of 1:

[0026]

[0027] Among them, y t is the electrical characteristic of power at time t, x t is the value of power at time t, μ and σ are the expected value and standard deviation of the power distribution respectively.

[0028] Optionally, the construction of the electrical equipment switch state classification subnet model based on electrical characteristics and gating mechanism includes:

[0029] Determine the switch status of electrical equipment based on electrical characteristics:

[0030]

[0031] in, is the switch state of the i-th electrical equipment at time t, is the electrical characteristic of the i-th electrical equipment at time t, and threshold is the preset threshold value of the switch of the electrical equipment;

[0032] The neural network mapping of the switch state is:

[0033]

[0034] in, is the mapping from the power of the i-th electrical device to the switch state;

[0035] Generate switch state classification subnetwork model based on neural network mapping:

[0036]

[0037] in, is the power of the i-th electrical equipment at time t, is the mapping of the switch state of each electrical device at time t, ⊙ is the characteristic product;

[0038] The encoder network of the electrical equipment switch state classification subnet model includes a six-layer one-dimensional convolutional neural network, which uses elu as an activation function to extract the mapping relationship between the device switch state and the total power; the decoder network mainly includes a three-layer fully connected neural network, and the upper layer uses a simoid activation function to output the decomposition result of the switch state.

[0039] Optionally, the construction of an electrical equipment load regression subnet model based on electrical features and an attention mechanism includes:

[0040] Get the attention weight of the power device at each moment:

[0041]

[0042] e tj =f(σ t-1 ,h j )

[0043] e tk =f(σ t-1 ,h k )

[0044] The attention weights are taken as the weighted combination of all hidden states of the encoder:

[0045]

[0046] Among them, h j is the conversion function of the encoding stage, σ t-1 is the conversion function in the decoding stage, and f(·) is the function for calculating h j The learning function of the scaling importance, α tj is the weight corresponding to the output of the switch state classification subnetwork model at time t, and T is the number of electrical devices.

[0047] Optionally, the physical characteristics include:

[0048] Use minimum and maximum scaling for data type data and normalize it to the [0,1] interval:

[0049] z′=(z-min) / (max-min)

[0050] Wherein, z is the data value of the current data type data, z′ is the physical characteristic of the current data type data, max and min are the maximum and minimum values ​​of the current data type data; the data type data includes the user's family population, the user's house area, and the user's house temperature;

[0051] Physical features are manually determined for non-digital data that protects the structure of the user's property.

[0052] Optionally, the construction of the user social feature regression subnet model based on social features and a sliding window mechanism includes: adopting a sliding window mechanism with the same window size as the electrical equipment load regression subnet model, the user social feature regression subnet model extracts features and refines information on social features through five layers of one-dimensional convolutional neural network layers, and its output feature map is input into a multi-layer fully connected neural network for processing, and the final output is a hidden state sequence of a certain length.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention provides a non-intrusive load identification method based on multidimensional features. On the basis of the electrical features of minute-level electrical equipment collected from residential electricity meters, the physical features of electrical equipment and the social features of residential users are introduced to perform minute-level non-intrusive load identification. The input and output types of the model are determined according to the physical features of the electrical equipment to improve the accuracy of the load identification results. For electrical equipment with relatively stable operating status, long single operating cycle, and small power fluctuation amplitude, a sequence-to-sequence model is used to predict multiple times at a single time point in a sliding window manner to calculate the average. For electrical equipment with frequent operating status changes, short single operating cycle, and large power fluctuation amplitude, a sequence-to-point model is used to predict only once at a single time point in a sliding window manner. For families with a large population, the frequency of use of large appliances such as washing machines and water heaters may be relatively high; in periods of high temperature, the frequency of use of air conditioners may be high; for families in the north, the frequency of use of appliances such as air humidifiers may be relatively high. Such social characteristics can have a positive impact on the final load identification results. Each type of characteristic data will be preprocessed separately, such as scaling and standardization, and finally the data will be output to the joint layer to integrate the social characteristics with the electrical characteristics, and the load identification results will be output through the fully connected layer, which improves the accuracy of load identification of minute-level data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of a non-intrusive load identification method based on multi-dimensional features provided by an embodiment of the present invention;

[0056] Figure 2 is a flow chart of a construction process of a non-intrusive load identification model provided by an embodiment of the present invention;

[0057] Figure 3 It is a schematic diagram of the working principle of the non-intrusive load identification model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0059] Embodiment 1:

[0060] like Figure 1 As shown, the present invention provides a non-intrusive load identification method based on multi-dimensional features, comprising the following steps:

[0061] S1. Obtaining total electrical appliance data within a target time period and preprocessing it;

[0062] S2, inputting the preprocessed data into a pre-built non-intrusive load identification model to perform load identification and obtain identification results;

[0063] like Figure 2 As shown in Figure 2, the construction process of the non-intrusive load identification model includes:

[0064] S101, constructing an input of a non-intrusive load identification model based on physical characteristics of electrical equipment;

[0065] S102, constructing a subnet model for classifying switch states of electrical equipment based on electrical characteristics and gating mechanisms;

[0066] S103, constructing an electrical equipment load regression subnet model based on electrical characteristics and attention mechanism;

[0067] S104, constructing a user social feature regression subnet model based on social features and a sliding window mechanism;

[0068] S105, constructing an input non-intrusive load identification model by combining the outputs of the electrical equipment switch state classification subnet model, the electrical equipment load regression subnet model, and the user social feature regression subnet model;

[0069] S106, obtaining electrical characteristic data, physical characteristic data of the electrical equipment and social characteristic data of the user;

[0070] S107, labeling the electrical characteristic data, the physical characteristic data, and the social characteristic data respectively and generating a training set;

[0071] S108. Train the non-intrusive load identification model and optimize parameters using the training set to obtain a final non-intrusive load identification model.

[0072] (1) Specific:

[0073] Physical characteristics include:

[0074] The power consumption equipment whose operation status change is lower than the preset value, whose single operation cycle is higher than the preset value and whose power adjustment amplitude is less than the preset value is recorded as 0;

[0075] Electrical equipment whose operating status change is greater than or equal to the preset value, whose single operating cycle is less than or equal to the preset value, and whose power switching amplitude is greater than or equal to the preset value is recorded as 1.

[0076] The inputs for building a non-intrusive load identification model based on the physical characteristics of electrical equipment include:

[0077] If the physical feature is 0, the sequence-to-sequence model is used. The sequence-to-sequence model predicts the power of the electrical equipment at a single moment multiple times in a sliding window manner and calculates the average value, and uses the predicted average value as the input of the non-intrusive load identification model;

[0078] If the obtained label information is 1, the sequence-to-point model is used. The sequence-to-point model predicts the power of the electrical equipment at a single moment only once in a sliding window manner, and uses the predicted value as the input of the non-intrusive load identification model.

[0079] (2) Specific:

[0080] Electrical characteristics include:

[0081] Use a standardized method to convert the power of the electrical equipment at the corresponding moment into a distribution with a mean of 0 and a standard deviation of 1:

[0082]

[0083] Among them, y t is the electrical characteristic of power at time t, x t is the value of power at time t, μ and σ are the expected value and standard deviation of the power distribution respectively.

[0084] Optionally, a subnet model for classifying the switch status of electrical equipment based on electrical characteristics and gating mechanisms includes:

[0085] Determine the switch status of electrical equipment based on electrical characteristics:

[0086]

[0087] in, is the switch state of the i-th electrical equipment at time t, is the electrical characteristic of the i-th electrical equipment at time t, and threshold is the preset threshold value of the switch of the electrical equipment;

[0088] The neural network mapping of the switch state is:

[0089]

[0090] in, is the mapping from the power of the i-th electrical device to the switch state;

[0091] Generate switch state classification subnetwork model based on neural network mapping:

[0092]

[0093] in, is the power of the i-th electrical equipment at time t, is the mapping of the switch state of each electrical device at time t, ⊙ is the characteristic product;

[0094] The encoder network of the electrical equipment switch state classification subnet model includes a six-layer one-dimensional convolutional neural network, which uses elu as the activation function to extract the mapping relationship between the equipment switch state and the total power; the decoder network mainly includes a three-layer fully connected neural network, and the upper layer uses the simoid activation function to output the decomposition result of the switch state.

[0095] The electrical equipment load regression subnetwork model based on electrical features and attention mechanism includes:

[0096] Get the attention weight of the power device at each moment:

[0097]

[0098] e tj =f(σ t-1 ,h j )

[0099] e tk =f(σ t-1 ,h k )

[0100] The attention weights are taken as the weighted combination of all hidden states of the encoder:

[0101]

[0102] Among them, h j is the conversion function of the encoding stage, σ t-1 is the conversion function in the decoding stage, and f(·) is the function for calculating h j The learning function of the scaling importance, α tj is the weight corresponding to the output of the switch state classification subnetwork model at time t, and T is the number of electrical devices.

[0103] (3) Specific:

[0104] Physical characteristics include:

[0105] Use minimum and maximum scaling for data type data and normalize it to the [0,1] interval:

[0106] z′=(z-min) / (max-min)

[0107] Among them, z is the data value of the current data type data, z′ is the physical characteristics of the current data type data, max and min are the maximum and minimum values ​​of the current data type data; the data type data includes the user's family population, the user's house area, and the user's house temperature;

[0108] Non-data data protects the structure of the user's house by manually determining physical characteristics.

[0109] The user social feature regression subnet model is constructed based on social features and a sliding window mechanism, including: adopting a sliding window mechanism with the same window size as the electrical equipment load regression subnet model; the user social feature regression subnet model extracts features and refines information on social features through five layers of one-dimensional convolutional neural network layers; the output feature map is input into a multi-layer fully connected neural network for processing, and the final output is a hidden state sequence of a certain length.

[0110] like Figure 3 As shown, the working principle of this embodiment is as follows:

[0111] (1) Use open source machine learning frameworks such as TensorFlow, SKlean, and Numpy to build Figure 3 The model structure shown is used to construct a minute-level non-intrusive load identification model.

[0112] (2) Determine the operating mode of the electrical equipment based on its physical characteristics. For electrical equipment with relatively stable operating status, long single operating cycle, and small power fluctuation, use the sequence-to-sequence model; for electrical equipment with frequent operating status changes, short single operating cycle, and large power fluctuation, use the sequence-to-point model.

[0113] (3) Combine the gating mechanism to build a subnet model for classifying the switch status of electrical equipment.

[0114] (4) Combine the attention mechanism to construct an electrical equipment load regression subnetwork model.

[0115] (5) Introduce social characteristics and construct a user social characteristics regression subnetwork model.

[0116] (6) Integrate electrical characteristics and social characteristics, connect the hidden state sequences output by the three types of subnetworks, and complete the overall model construction.

[0117] (7) For the load identification models constructed for different types of electrical equipment, the parameters are set and adjusted separately, such as the length of the time window. For electrical equipment with a shorter operating cycle, a smaller time window is set; for electrical equipment with a longer operating cycle, the time window is appropriately increased. The same applies to other parameters.

[0118] (8) Deploy a minute-level non-intrusive load identification system to complete the integrated data reception, processing, and import processes, and display the identification results in real time on a visual interface.

[0119] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0120] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A non-intrusive load identification method based on multi-dimensional features, characterized in that: include: Obtain the total electrical appliance data within the target time period and perform preprocessing; The preprocessed data is input into a pre-built non-intrusive load identification model to perform load identification and obtain identification results; The construction process of the non-intrusive load identification model includes: Construct inputs for non-intrusive load identification models based on the physical characteristics of electrical equipment; Construct a subnet model for classifying the switch status of electrical equipment based on electrical characteristics and gating mechanisms; Construct an electrical equipment load regression subnetwork model based on electrical features and attention mechanism; Construct a user social feature regression subnetwork model based on social features and sliding window mechanism; Combining the outputs of the electrical equipment switch state classification subnet model, the electrical equipment load regression subnet model, and the user social characteristics regression subnet model, a non-intrusive load identification model for input is constructed; Obtain electrical characteristic data, physical characteristic data of electrical equipment and social characteristic data of users; Label the electrical characteristic data, physical characteristic data and social characteristic data respectively and generate training sets; The non-intrusive load identification model is trained and parameter optimized through the training set to obtain the final non-intrusive load identification model.

2. The non-intrusive load identification method based on multi-dimensional features according to claim 1 is characterized in that: The physical characteristics include: The power consumption equipment whose operation status change is lower than the preset value, whose single operation cycle is higher than the preset value and whose power fluctuation amplitude is less than the preset value is recorded as 0; Electrical equipment whose operating status change is greater than or equal to the preset value, whose single operating cycle is less than or equal to the preset value, and whose power fluctuation amplitude is greater than or equal to the preset value is recorded as 1.

3. The non-intrusive load identification method based on multi-dimensional features according to claim 2 is characterized in that: The input of the non-intrusive load identification model based on the physical characteristics of the electrical equipment includes: If the physical feature is 0, a sequence-to-sequence model is used, which predicts the power of the electrical equipment at a single moment multiple times in a sliding window manner and calculates the average value, and uses the predicted average value as the input of the non-intrusive load identification model; If the acquired label information is 1, a sequence-to-point model is used, and the sequence-to-point model predicts the power of the electrical equipment at a single moment only once in a sliding window manner, and uses the predicted value as the input of the non-intrusive load identification model.

4. The non-intrusive load identification method based on multi-dimensional features according to claim 1 is characterized in that: The electrical characteristics include: Use a standardized method to convert the power of the electrical equipment at the corresponding moment into a distribution with a mean of 0 and a standard deviation of 1: Among them, y t is the electrical characteristic of power at time t, x t is the value of power at time t, μ and σ are the expected value and standard deviation of the power distribution respectively.

5. The non-intrusive load identification method based on multi-dimensional features according to claim 4 is characterized in that: The construction of the electrical equipment switch state classification subnet model based on electrical characteristics and gating mechanism includes: Determine the switch status of electrical equipment based on electrical characteristics: in, is the switch state of the i-th electrical equipment at time t, is the electrical characteristic of the i-th electrical equipment at time t, and threshold is the preset threshold value of the switch of the electrical equipment; The neural network mapping of the switch state is: in, is the mapping from the power of the i-th electrical device to the switch state; Generate switch state classification subnetwork model based on neural network mapping: in, is the power of the i-th electrical equipment at time t, is the mapping of the switch state of each electrical device at time t, ⊙ is the characteristic product; The encoder network of the electrical equipment switch state classification subnet model includes a six-layer one-dimensional convolutional neural network, which uses elu as an activation function to extract the mapping relationship between the device switch state and the total power; the decoder network mainly includes a three-layer fully connected neural network, and the upper layer uses a simoid activation function to output the decomposition result of the switch state.

6. The non-intrusive load identification method based on multi-dimensional features according to claim 4 is characterized in that: The electrical equipment load regression subnet model based on electrical characteristics and attention mechanism includes: Get the attention weight of the power device at each moment: e tj =f(σ t-1 ,h j ) e tk =f(σ t-1 ,h k ) The attention weights are taken as the weighted combination of all hidden states of the encoder: Among them, h j is the conversion function of the encoding stage, σ t-1 is the conversion function in the decoding stage, and f(·) is the function for calculating h j The learning function of the scaling importance, α tj is the weight corresponding to the output of the switch state classification subnetwork model at time t, and T is the number of electrical devices.

7. The non-intrusive load identification method based on multi-dimensional features according to claim 1 is characterized in that: The physical characteristics include: Use minimum and maximum scaling for data type data and normalize it to the [0,1] interval: z′=(z-min) / (max-min) Wherein, z is the data value of the current data type data, z′ is the physical characteristic of the current data type data, max and min are the maximum and minimum values ​​of the current data type data; the data type data includes the user's family population, the user's house area, and the user's house temperature; Physical characteristics are manually determined for non-digital data, including the structure of the user's house.

8. The non-intrusive load identification method based on multi-dimensional features according to claim 7 is characterized in that: The method for constructing a user social feature regression subnet model based on social features and a sliding window mechanism includes: adopting a sliding window mechanism with the same window size as the electrical equipment load regression subnet model; the user social feature regression subnet model extracts features and refines information on social features through five layers of one-dimensional convolutional neural network layers; the feature map outputted by the model is inputted into a multi-layer fully connected neural network for processing; and the final output is a hidden state sequence of a certain length.

Images