Non-intrusive Load Monitoring Method Based on Multi-Physical Quantity Fusion

Abstract

The present application relates to a non-invasive load monitoring method based on multi-physical quantity fusion, belonging to the technical field of load monitoring. The method includes: acquiring N physical quantities of electrical energy data in a main line; inputting the N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of a target device; wherein, the fully convolutional autoencoder network includes N feature extraction modules, a feature fusion module and an output layer; different feature extraction modules correspond to different physical quantities; the feature extraction modules are established based on autoencoders; the feature fusion module is used to fuse the feature values output by the N feature extraction modules to obtain a feature sequence matching the data size of the output layer; the feature fusion module is established based on N successively connected transposed convolutional layers; no fully connected layer is adopted in the network structure, and convolutional or transposed convolutional operations are used to replace the fully connected operation to process data, which can reduce the number of model parameters, save computing resources and improve training efficiency.

Description

Technical Field

[0001] The present application relates to a non-intrusive load monitoring method based on multi-physical quantity fusion, and belongs to the technical field of load monitoring. Background Art

[0002] Non-intrusive load monitoring (NILM) refers to installing a monitoring device at the entrance of the user's main line (or bus), and decomposing the collected total load electrical quantity into independent operation data of each electrical appliance through intelligent algorithms, and identifying the working states of each electrical appliance and analyzing the user's electricity consumption behavior. In most of the existing work on non-intrusive load monitoring problems, the monitoring purpose is usually to understand the power consumption of various different devices in the residence, and use non-intrusive load monitoring technology to assist in home energy management. For residential users, the most direct benefit of the home energy management system is that it can help users reduce the home power consumption and the user's electricity bill by real-time monitoring and scheduling of various household appliances while maintaining a comfortable living condition. Currently, when conducting research on non-intrusive load decomposition algorithms, since it is expected to analyze the active power of the device from the main line data, only the active power information in the main line is used for power analysis.

[0003] However, in addition to active power, various other physical quantity information in the main line can also be used as input information for energy decomposition. This includes reactive power, apparent power, etc. Compared with the method of only using active power for energy decomposition, introducing reactive power or other physical quantity information in the process of inferring the active power of the device can bring an overall performance improvement. For example, the invention patent "Non-intrusive Load Monitoring Smart Meter and Power Decomposition Method" with the application number 202010643323.8 discloses a non-intrusive load monitoring smart meter, in which the power decomposition model is a sequence-to-point model with an attention mechanism, and the attention mechanism is used to weight values for sliding windows of various different sizes and various physical quantities including active power, reactive power, and voltage; the input of the sequence-to-point model with an attention mechanism is the sequence data corresponding to the total electricity consumption, and the output is the electricity consumption of each device.

[0004] In the above method, although it is proposed to calculate the power consumption of each device using multiple physical quantities, the power decomposition model uses a Sequence-to-point (Seq2point) model, and the number of parameters of the Seq2point model is more than three million two hundred thousand. The reason for the large number of parameters in the Seq2point model is the existence of fully connected layers in the network structure. In the Seq2point network structure, two fully connected layers are placed before the output layer, and the number of parameters in these two fully connected layers accounts for 93% of the entire network. Based on this, the large number of model parameters will lead to problems such as the model occupying a large amount of computing resources and the low training efficiency of the model. Summary of the Invention

[0005] The present application provides a non-intrusive load monitoring method based on multi-physical quantity fusion, which can not use fully connected layers in the network structure, and uses convolutional or transposed convolutional operations to replace fully connected operations to process data, which can greatly reduce the number of model parameters, thereby saving the computing resources of the model and improving the training efficiency of the model. The present application provides the following technical solutions:

[0006] Obtain N physical quantities of electrical energy data in the main line, where N is an integer greater than 1;

[0007] Input the N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device;

[0008] Among them, the fully convolutional autoencoder network includes N feature extraction modules, a feature fusion module connected to the N feature extraction modules, and an output layer connected to the feature fusion module;

[0009] Each feature extraction module is used to extract features from a corresponding physical quantity, and different feature extraction modules correspond to different physical quantities; the feature extraction module is established based on an autoencoder;

[0010] The feature fusion module is used to fuse the feature values output by the N feature extraction modules to obtain a feature sequence matching the data size of the output layer; the feature fusion module is established based on N sequentially connected transposed convolutional layers;

[0011] The output layer is used to output the active power value based on the feature sequence.

[0012] Optionally, the autoencoder includes an encoder and a decoder connected to the encoder;

[0013] The encoder includes 3 sequentially connected one-dimensional convolutional layers, and after the data input to the encoder passes through the 3 one-dimensional convolutional layers, the size of the features decreases sequentially;

[0014] The decoder includes 4 sequentially connected one-dimensional transposed convolutional layers. After the data input to the decoder passes through the 4 one-dimensional transposed convolutional layers, the size of the features is sequentially restored to the original size of the data input to the autoencoder.

[0015] Optionally, the convolutional kernel sizes of the 3 sequentially connected one-dimensional convolutional layers are all 8, and the number of convolutional kernels in each one-dimensional convolutional layer is 1 / 2 of the number of convolutional kernels in the next one-dimensional convolutional layer; each one-dimensional convolutional layer includes a rectified linear unit (relu) function and has a stride of 1.

[0016] The convolutional kernel sizes of the 4 sequentially connected one-dimensional transposed convolutional layers are all 8, and the number of convolutional kernels in the first one-dimensional transposed convolutional layer is 2 times the number of convolutional kernels in the second one-dimensional transposed convolutional layer; each one-dimensional transposed convolutional layer includes a Sigmoid function and has a stride of 1.

[0017] Optionally, in the N sequentially connected transposed convolutional layers, the size of the convolutional kernel in the time direction of each transposed convolutional layer is not 1, and each transposed convolutional layer does not use an activation function.

[0018] The number of convolutional kernels in each transposed convolutional layer is 1 more than the number of convolutional kernels in the next transposed convolutional layer.

[0019] Optionally, the convolutional kernel size in each transposed convolutional layer is 8 and the stride is 1.

[0020] Optionally, before obtaining the N physical quantities of the electric energy data in the trunk line, it further includes:

[0021] Determining the types of the N physical quantities.

[0022] Optionally, the determining the types of the N physical quantities includes:

[0023] Determining the initial types of the physical quantities, where the initial types include at least two of active power, active component of current, reactive component of current, admittance, and power factor.

[0024] Training the fully convolutional autoencoder network using the physical quantities of the initial types to obtain the first result value of the model evaluation index.

[0025] Training the fully convolutional autoencoder network again using the physical quantities of the initial types and other types of physical quantities to obtain the second result value of the model evaluation index; where the other types are different from the initial types, and the other types are data types that can be collected from the trunk line or data types that can be calculated based on the electric energy data collected from the trunk line.

[0026] In the case where the second result value is greater than the first result value, add the other type to the initial type, and execute again the step of using the physical quantity of the initial type and the physical quantity of the other type to train the fully convolutional autoencoder network again to obtain the second result value of the model evaluation index;

[0027] In the case where the second result value is less than or equal to the first result value, determine the initial type as the type of the N physical quantities.

[0028] Optionally, the determining the type of the N physical quantities includes:

[0029] Collect the electrical energy data to obtain the first physical quantity;

[0030] Use the electrical energy data to calculate the second physical quantity of each type;

[0031] Use the Pearson correlation coefficient, chi-square verification, mutual information, maximum information coefficient, and distance correlation coefficient to score each first physical quantity and second physical quantity;

[0032] Determine the types of the first physical quantity and / or the second physical quantity greater than the preset scoring threshold as the types of the N physical quantities.

[0033] Optionally, the determining the type of the N physical quantities includes:

[0034] Collect the electrical energy data to obtain the first physical quantity;

[0035] Use the electrical energy data to calculate the second physical quantity of each type;

[0036] Input the first physical quantity and the second physical quantity into a pre-trained weight calculation model to obtain the weight coefficients of each physical quantity;

[0037] Select the types of the N physical quantities in the order of the weight coefficients from large to small.

[0038] Optionally, the inputting the N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device includes:

[0039] Process the abnormal data in the N physical quantities, where the abnormal data refers to the physical quantity exceeding the normal range, and the normal ranges corresponding to different types of physical quantities are different;

[0040] Perform standardization processing on the processed physical quantities to obtain N standardized physical quantities;

[0041] Input the N standardized physical quantities into N feature extraction modules respectively for the fully convolutional autoencoder network to perform load monitoring calculations, and obtain the active power value of the target device.

[0042] The beneficial effects of this application at least include: designing a fully convolutional autoencoder network built by convolutional layers and transposed convolutional layers. This network consists of multiple parallel feature extraction modules and a feature fusion module, and can infer the active power data of the target device corresponding to a certain period through various different physical quantity data in the trunk line; it can solve the problems that the traditional power decomposition model has a large number of model parameters, resulting in a large amount of computing resources occupied by the model and a low training efficiency of the model; since the fully connected layer is not adopted in the network structure, convolutional or transposed convolutional operations are used to replace the fully connected operation to process data, which can greatly reduce the number of model parameters, thereby saving the computing resources of the model and improving the training efficiency of the model.

[0043] In addition, by selecting five physical quantities from more than ten physical quantities that can greatly improve the model performance, it can be ensured that the physical quantities input into the model are all physical quantities that contribute to predicting the active power value, which can not only improve the accuracy of model calculation, but also avoid wasting the computing resources of the model.

[0044] The above description is only an overview of the technical solution of this application. In order to understand the technical means of this application more clearly and implement it according to the content of the specification, the following uses the preferred embodiments of this application and combines the accompanying drawings to describe in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flowchart of a non-intrusive load monitoring method based on multi-physical quantity fusion provided by an embodiment of this application;

[0046] Figure 2 is a schematic structural diagram of a fully convolutional autoencoder network provided by an embodiment of this application;

[0047] Figure 3 is a schematic structural diagram of a feature extraction module provided by an embodiment of this application;

[0048] Figure 4 is a schematic structural diagram of a feature fusion module provided by an embodiment of this application;

[0049] Figure 5 is a schematic diagram showing the influence of inputting different types of physical quantities on different performance indicators of the model provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0051] The present application proposes a method for non-intrusive load monitoring that can comprehensively utilize various physical quantity information. It mainly uses a full convolution auto encoder network (Full Convolution Auto Encoder Network Based on Multi Physical Quantity Fusion, FC-AE-MPQF) to solve the non-intrusive load monitoring problem. The main idea of this method includes: first, using a convolutional neural network to build an auto encoder feature extraction module to extract features of different physical quantities, then using a convolutional neural network to build a feature fusion module to fuse different features, and finally using the fused features to infer the power consumption of the target device.

[0052] Optionally, the present application takes the non-intrusive load monitoring method based on multi-physical quantity fusion provided in each embodiment as an example for illustration in an electronic device. The electronic device is a terminal or server with computing capabilities. The terminal can be an intelligent electric meter connected to the main line, or a computer, tablet, etc. that is communicatively connected to the intelligent electric meter. The type of the electronic device is not limited in this embodiment.

[0053] Figure 1 It is a flowchart of a non-intrusive load monitoring method based on multi-physical quantity fusion provided by an embodiment of the present application. The method at least includes the following steps:

[0054] Step 101, obtain N physical quantities of the electrical energy data in the main line.

[0055] Wherein, N is an integer greater than 1.

[0056] Generally speaking, increasing the number of types of input data of the model can improve the performance of the model. However, it should be noted that the newly added data needs to be data related to the task and the original data. Otherwise, adding data with little relevance to the task in the model is not conducive to the training and convergence of the model, and will cause losses to the model performance. In addition, when the number of types of input information of the model is relatively large, it is necessary to design a suitable network structure to comprehensively integrate various feature information, so that the neural network can combine different types of data to obtain its target data.

[0057] In the load monitoring scenario, most of the household electricity of residents is alternating current. Based on this, the main physical quantities existing in the circuit system are as follows:

[0058] 1. Active power (P). When there are energy storage elements such as inductors and capacitors in an AC circuit, the energy storage elements may cause the direction of energy flow in the circuit system to change periodically. In a complete cycle, the net flow of energy in one direction is called active power.

[0059] 2. Reactive power (Q). In a complete cycle, the part of the energy that reciprocates between the energy storage element and the power source is called reactive power.

[0060] 3. Phase difference Inductive or capacitive elements present in the circuit will cause a phase difference between the voltage and current with the same frequency. This difference is called the phase difference.

[0061] 4. Apparent power (S). The magnitude of the vector sum of active power and reactive power is called apparent power. Apparent power can reflect the total capacity of the system and is a specific manifestation of both active and reactive powers.

[0062] 5. Power factor (PF). The power factor is a specific physical quantity in an AC power system, which refers to the ratio of active power to apparent power, and its value ranges from [0, 1].

[0063] 6. Admittance (Y). The essence of non-intrusive load monitoring tasks is to analyze the internal information of the circuit system through main line data, including the types and operating states of devices in the circuit, etc. Admittance is a physical quantity that can describe the internal information of a device. Its physical meaning is to characterize the difficulty of alternating current passing through a circuit or system. Admittance consists of two parts: conductance (G) and susceptance (B). Conductance represents the strength of the conductor's ability to transmit current; susceptance is the reciprocal of reactance, and reactance is used to describe the magnitude of the hindrance of capacitance and inductance to current.

[0064] The above physical quantities can all represent the operating state of the entire circuit system from different aspects. This is related to the operating state and power of the devices in the circuit. Comprehensively using the information of these physical quantities can extract more useful physical quantity information from limited data.

[0065] When the above physical quantities all exist in the same power grid system, the instantaneous voltage of the alternating current is denoted as u(t), the instantaneous current is denoted as i(t), and the instantaneous power is denoted as p(t). The effective values of voltage and current in sinusoidal alternating current are denoted as V rms and I rms , and the imaginary unit is denoted as j. There are the following constraint relationships among the physical quantities in the circuit:

[0066] p(t) = u(t)i(t)

[0067]

[0068]

[0069]

[0070]

[0071] Y = G + jB

[0072]

[0073] As can be seen from the above calculation formulas, there are certain mathematical relationships between different physical quantities. There are far more physical quantities in the actual circuit than those mentioned above, and there are more than a dozen meaningful physical quantities that can be directly collected or indirectly calculated in the circuit. There may be information redundancy among these physical quantities, and some physical quantities may not be helpful for the non-intrusive load monitoring task. The number of physical quantities input into the model affects the number of feature extraction modules and the number of layers of the feature fusion module. In order to avoid overfitting of the network model caused by information redundancy and also to avoid too many parameters in the model, it is necessary to select a small number of representative key physical quantities (i.e., the N physical quantities in this step) from more than a dozen physical quantities as the input of the model, that is, feature selection is required. The purpose of feature selection is to find the optimal feature subset. By feature selection, redundant or irrelevant features in the feature set can be removed, which can reduce the number of features and achieve the purpose of improving the model performance and reducing the running time.

[0074] Based on this, the acquisition methods of the N physical quantities in this embodiment include but are not limited to the following several types:

[0075] The first type: Acquire N physical quantities preset by the user related to the active power of the predicted target device.

[0076] The second type: Before acquiring N physical quantities of the electric energy data in the main line, determine the types of the N physical quantities.

[0077] Among them, the methods for determining the types of the N physical quantities are: Filter method, Wrapper method or Embedded method. The following will introduce each determination method separately.

[0078] 1. Filtering method. That is, the features are scored according to Pearson correlation coefficient, chi-square test, mutual information, maximum information coefficient, distance correlation coefficient, etc., and then a threshold or the number of candidate features to be selected is set for screening. Specifically, determine N types of physical quantities, including: collecting electrical energy data of the main line to obtain the first physical quantity; calculating the second physical quantity of each type using the electrical energy data; scoring each first physical quantity and second physical quantity using Pearson correlation coefficient, chi-square test, mutual information, maximum information coefficient, and distance correlation coefficient; determining the types of the first physical quantity and / or the second physical quantity greater than the preset scoring threshold as the N types of physical quantities.

[0079] Among them, the second physical quantity is directly calculated using the first physical quantity, or is further calculated using the result calculated from the first physical quantity. This embodiment does not limit the implementation manner of the second physical quantity.

[0080] 2. Wrapper method. That is, according to the result value of the model evaluation index, several features are selected or removed from the feature set each time. Specifically, determine N types of physical quantities, including: determining the initial type of physical quantity; training a fully convolutional autoencoder network using the physical quantity of the initial type to obtain the first result value of the model evaluation index; training the fully convolutional autoencoder network again using the physical quantity of the initial type and other types of physical quantities to obtain the second result value of the model evaluation index; where the other type is different from the initial type, and the other type is a data type that can be collected from the main line or a data type that can be calculated based on the electrical energy data collected from the main line; in the case where the second result value is greater than the first result value, add the other type to the initial type, and execute again the step of training the fully convolutional autoencoder network again using the physical quantity of the initial type and other types of physical quantities to obtain the second result value of the model evaluation index; in the case where the second result value is less than or equal to the first result value, determine the initial type as the N types of physical quantities.

[0081] Among them, the initial type includes at least two of active power, active component of current, reactive component of current, admittance, and power factor.

[0082] 3. Embedding method. That is, first obtain the weight coefficient of each feature through a specific machine learning model, and then select features according to the coefficient value from large to small. This method is similar to the filtering method, and the difference is that the feature weight coefficient in this method is obtained through the model training method. Specifically, determine N types of physical quantities, including: collecting electrical energy data to obtain the first physical quantity; calculating the second physical quantity of each type using the electrical energy data; inputting the first physical quantity and the second physical quantity into a pre-trained weight calculation model to obtain the weight coefficient of each physical quantity; selecting the N types of physical quantities in descending order of the weight coefficient.

[0083] Among them, the weight calculation model is established based on a neural network and trained using sample physical quantities and the weight coefficient labels of the sample physical quantities.

[0084] Step 102: Input N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device.

[0085] Assume that when there are N physical quantity data in the main line as the input of the model, the set of physical quantities in the main line is denoted as Y, and each physical quantity data is denoted as Y j , then the Y from time point 1 to T j data can be expressed as:

[0086]

[0087] The active power data of the target device is denoted as X p , there is:

[0088]

[0089] The problem to be solved by the technical solution provided in this embodiment is to obtain X through Y p , that is, to infer the active power value of the target device through the set of multiple physical quantity data in the main line. Based on this, a neural network structure based on multi-physical quantity fusion. At the same time, another problem existing in the existing deep learning methods for non-intrusive load monitoring is that the number of model parameters is too large. In this embodiment, a fully connected layer is not used in the network structure, and convolution or transposed convolution operations are used to replace the fully connected operation to process the data, which can greatly reduce the number of model parameters. Based on the above principle, this embodiment proposes a fully convolutional autoencoder network based on multi-physical quantity fusion to solve the non-intrusive load monitoring problem. The idea of the autoencoder and depthwise separable convolution is borrowed in the design process of the model, and the model can infer the active power data value of the target device through various different physical quantity information in the main line.

[0090] Since some abnormal data will inevitably be generated due to problems with the sensors themselves during the data acquisition process of the main line, if the original data is not processed, it will affect the model's analysis of the data and increase the difficulty and calculation accuracy of model training. Therefore, before the data is input into the model, it is first necessary to process the outlier abnormal value points of the data.

[0091] Specifically, input N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device, including: processing the abnormal data among the N physical quantities, where the abnormal data refers to the physical quantities exceeding the normal range, and the normal ranges corresponding to different physical quantities are different; performing standardization processing on the processed physical quantities to obtain N standardized physical quantities; correspondingly inputting the N standardized physical quantities into N feature extraction modules for the fully convolutional autoencoder network to perform load monitoring calculations and obtain the active power value of the target device.

[0092] Among them, processing the abnormal data among the N physical quantities includes: collecting the data on the main line and the branch lines of the device to obtain the original data of the N physical quantities; analyzing the normal operating state of the device and the value ranges of various physical quantities, and then replacing the data exceeding the maximum or minimum value ranges in the original data with the maximum or minimum values respectively.

[0093] After performing the operation of removing data outliers, the multi-physical quantity fusion method needs to fuse the data of different physical quantities into one data through a convolutional neural network. The numerical ranges of different physical quantities are different, with different dimensions and dimension units. If the data belonging to different physical quantities are directly fused through convolution, the fused data will be more biased towards being affected by the physical quantity data with larger numerical values. To eliminate the influence of dimensions, it is necessary to perform standardization processing on the data.

[0094] In this embodiment, performing standardization processing on the processed physical quantities to obtain N standardized physical quantities includes: performing min-max standardization and Z-score standardization processing on the processed physical quantities to obtain N standardized physical quantities.

[0095] Among them, min-max standardization is also called deviation standardization, and its purpose is to perform a linear transformation on the data to map it to a number between 0 and 1. Its conversion function is:

[0096]

[0097] Among them, for each processed physical quantity, min is the minimum value of the processed physical quantity, max is the maximum value of the processed physical quantity, and x represents each processed physical quantity.

[0098] The purpose of Z-score standardization is to convert the data into data conforming to a normal distribution. Its conversion function is:

[0099]

[0100] Among them, for each processed physical quantity, μ is the average value of the processed physical quantity, and σ is the standard deviation of the processed physical quantity.

[0101] After min-max normalization and Z-score normalization, the data of different physical quantities are transformed into data that conforms to the normal distribution with a minimum value of 0 and a maximum value of 1. Each data is at the same order of magnitude. Data normalization can improve the accuracy of the model and accelerate the convergence speed of the model.

[0102] In this embodiment, referring to Figure 2 , the overall fully convolutional autoencoder network has a multi-input single-output structure. Its input is various different physical quantity information in the main line, and the output is the active power value of the target device.

[0103] Referring to Figure 2 , the execution process of the non-intrusive load monitoring method provided in this embodiment specifically includes: First, preprocess the different physical quantity data 21 in the main line, and then input them into different input layers 22 respectively. Immediately after the input layer are multiple parallel feature extraction modules 23. The number of feature extraction modules is the same as the number of types of physical quantities input to the model. The design idea of the feature extraction module comes from the autoencoder. The purpose of autoencoding is to learn a representation for a set of data, and the representation is also called a characterization or encoding. The specific network design in the feature extraction module will be described in detail below. The feature fusion module 24 is connected after the feature extraction module 23. The role of the feature fusion module 24 is to fuse the features extracted from different physical quantities together. The feature fusion module will be described in more detail below. The last part of the network is the output layer 25. The output layer 25 is responsible for processing the fused features into the output of the model, and finally performing denormalization processing on the output of the model to obtain the active power value of the target device.

[0104] Compared with other methods that only use the active power data in the main line to infer the power consumption of the target device, the non-intrusive load monitoring method based on multi-physical quantity fusion proposed in this embodiment can make full use of various physical quantity information in the main line to infer the power consumption of the target device, and has more comprehensive input feature information than other methods. In addition, this method improves the existing network structure and proposes a fully convolutional neural network for solving the non-intrusive load monitoring problem. The following subsections will describe several details of the non-intrusive load monitoring method based on multi-physical quantity fusion in detail.

[0105] According to Figure 2It can be seen that the fully convolutional autoencoder network provided in this embodiment includes N feature extraction modules, a feature fusion module connected to the N feature extraction modules, and an output layer connected to the feature fusion module. Each feature extraction module is used to extract features of a corresponding physical quantity, and different feature extraction modules correspond to different physical quantities; the feature extraction module is established based on an autoencoder; the feature fusion module is used to fuse the feature values output by the N feature extraction modules to obtain a feature sequence matching the data size of the output layer; the feature fusion module is established based on N sequentially connected transposed convolutional layers; the output layer is used to output the active power value based on the feature sequence.

[0106] Among them, the design idea of the feature extraction module comes from the denoising autoencoder. To some extent, non-intrusive load decomposition can be regarded as a denoising task. Typical denoising tasks include removing grain from old photos or removing reverberation from recordings. As for non-intrusive load decomposition, it can be regarded as recovering the "clean" power data of the target device from the "mixed" data containing "noise" in the main line. The autoencoder is a commonly used artificial neural network for denoising tasks at present. The autoencoder can use the input information as the target to be learned and thus learn a kind of encoding according to the input information. The autoencoder mainly consists of two parts: an encoder and a decoder. Among them, the encoder is used to compress and encode the input into a latent space representation, and the role of the decoder is to reconstruct the input signal through the latent space representation. The autoencoder is a typical Encoder-Decoder structure, and Encoder-Decoder, as a general framework, can be used for processing tasks of various types of data such as text, language, image, and video. The encoder and the decoder can be built using different network structures respectively. The specific network structure of the feature extraction module used in the multi-physical quantity fusion non-intrusive load monitoring method is as Figure 3 shown. The autoencoder includes an encoder 31 and a decoder 32 connected to the encoder.

[0107] The encoder 31 includes 3 sequentially connected one-dimensional convolutional layers, and after the data input to the encoder passes through the 3 one-dimensional convolutional layers, the size of the feature decreases sequentially. The decoder 32 includes 4 sequentially connected one-dimensional transposed convolutional layers, and after the data input to the decoder passes through the 4 one-dimensional transposed convolutional layers, the size of the feature is restored to the original size of the data input to the autoencoder sequentially.

[0108] Figure 3Take, for example, that the kernel sizes of three consecutive one-dimensional convolutional layers are all 8. The number of kernels in each one-dimensional convolutional layer is 1 / 2 of the number of kernels in the next one-dimensional convolutional layer. For example, the number of kernels in the first one-dimensional convolutional layer is 8, the number of kernels in the second one-dimensional convolutional layer is 16, and the number of kernels in the third one-dimensional convolutional layer is 32. Each one-dimensional convolutional layer includes a Rectified Linear Unit (relu) function and has a stride of 1.

[0109] Figure 3 In [description], the kernel sizes of four consecutive one-dimensional transposed convolutional layers are also 8. The number of kernels in the first one-dimensional transposed convolutional layer is 2 times the number of kernels in the second one-dimensional transposed convolutional layer. For example, the number of kernels in the first one-dimensional transposed convolutional layer is 16, and the number of kernels in the second one-dimensional transposed convolutional layer is 8. The number of kernels in the third one-dimensional transposed convolutional layer is 1, and the number of kernels in the fourth one-dimensional transposed convolutional layer is 1; each one-dimensional transposed convolutional layer includes a Sigmoid function and has a stride of 1. The number of kernels in the third one-dimensional convolutional layer is 2 times the number of kernels in the first one-dimensional transposed convolutional layer. In this way, the number of kernels in the 3 one-dimensional convolutional layers and the first 2 one-dimensional transposed convolutional layers can be set as: 8, 16, 32, 16, 8, which is overall symmetric.

[0110] Through Figure 3 It can be seen that the parameter settings of the convolutional layer and the transposed convolutional layer are to make the encoder and decoder network structures symmetric. The data of the non-intrusive load monitoring task is time-series data, so it is more suitable to use one-dimensional convolution to operate on it.

[0111] Optionally, in this embodiment, take the activation function of the 3 one-dimensional convolutional layers as relu and the activation function of the 4 one-dimensional transposed convolutional layers as the Sigmoid function. In actual implementation, the activation functions of the two can also be set to the same activation function. This embodiment does not limit the setting method of the activation function.

[0112] Refer to Figure 4 , in N consecutive transposed convolutional layers, the size of the kernel in the time direction of each transposed convolutional layer is not 1, and each transposed convolutional layer does not use an activation function; the number of kernels in each transposed convolutional layer is 1 more than the number of kernels in the next transposed convolutional layer.

[0113] Schematically, the kernel size in each transposed convolutional layer is 8 and the stride is 1.

[0114] According to Figure 4It can be seen that the difference between N transposed convolutional layers lies in their different numbers of convolutional kernels, which decrease from the number of features N to 1 one by one. Through such parameter settings, it can be achieved that the number of features decreases by one every time the features pass through a transposed convolutional layer. After multiple one-dimensional transposed convolutional operations, different features can be gradually fused into a single feature.

[0115] After multiple parallel feature extraction modules process data of different physical quantities through the encoders and decoders therein, eigenvalue features of different physical quantities are extracted. However, at this time, the eigenvalue features are still independent of each other and have no association. The feature fusion method can achieve the complementary advantages of multiple features and obtain more robust and accurate results. Traditional methods use a merging function (concat) or an adding function (add) to operate on the features and then perform feature fusion through a fully connected layer. Different from traditional methods, in this embodiment, transposed convolutional operations are used for feature fusion. The advantage of using transposed convolution compared to using a fully connected layer for feature fusion is that it can greatly reduce the amount of computation and the number of parameters. Depthwise separable convolution is mainly divided into two processes: depthwise convolution and pointwise convolution. In depthwise convolution, each convolutional kernel is only responsible for one channel, and each channel is only convolved by a single convolutional kernel. The depthwise convolution operation performs independent convolution operations on each channel and cannot effectively utilize the feature information of the same channel in the spatial position. Therefore, it is necessary to combine the features through pointwise convolution operations to generate new features. Pointwise convolution is very similar to the conventional convolution operation. The only difference is that the convolutional kernel size of pointwise convolution is 1×1×M, where M is the number of channels in the previous layer. The pointwise convolution operation will perform a weighted combination of the features generated by the depthwise convolution operation to produce new features.

[0116] Each feature extraction module in the non-intrusive load monitoring method provided in this embodiment also only performs feature extraction operations on individual physical quantities, and cannot effectively utilize the information of features in the time dimension. The characteristics of the feature extraction module are very similar to those of pointwise convolution. Therefore, the features that have been extracted can be fused through the idea of pointwise convolution. However, different from the pointwise convolution operation, while fusing the features from different physical quantities, the feature fusion module needs to restore the features whose size has been reduced due to the convolution operation to the original input size, so as to form a corresponding relationship at the same moment between the data of the main line and the device branch lines, meeting the requirements of sequence-to-sequence. In view of this, transposed convolution is selected for feature fusion operation. In addition, the convolution kernel size in the feature fusion module is also different from that in the pointwise convolution. The first dimension of the convolution kernel size of the transposed convolution, that is, the size in the time direction, cannot be 1, because performing transposed convolution on physical quantity data at a single time point is meaningless. Incorporating the data in a time period into the receptive field of the transposed convolution can extract features such as the degree of change of the data. Therefore, in the method of this chapter, the size of the transposed convolution kernel is set to 8. The purpose of the feature fusion module is to combine the features extracted by the feature extraction module, replacing the fully connected layer in Seq2Point. Therefore, no activation function is used in the transposed convolution layer of the feature fusion module.

[0117] Among them, the fully convolutional autoencoder network is trained using the training data for Figures 2 to 4 the network structure shown. The training process at least includes the following steps:

[0118] Step 1: Obtain training data. Among them, each piece of training data includes N types of sample physical quantities and the active power label values of the target devices corresponding to the N types of sample physical quantities.

[0119] Schematically, the training data is obtained from a public dataset. For example: The commonly used public datasets are shown in Table 1 below. As can be seen from Table 1, there are more physical quantities in the main line data than in the branch line data in most datasets. The datasets with more physical quantities in the main line data are AMPds2, BLUED, and iAWE. Among the three, the AMPds2 dataset has the longest acquisition duration and the largest amount of data. Since deep learning requires a large amount of data for model training and verification, in this embodiment, AMPds2 is used as an example of training data.

[0120] Table 1:

[0121]

[0122] The AMPds2 dataset contains one-year measurement data of 21 electricity meters with a sampling interval of one minute. The data acquisition source is a house built in the Vancouver area in 1955. AMPds2 contains two years of data starting from January 1, 2012 to April 1, 2014. Among them, the sub-circuit data has a total of 15 types of equipment data. If all the data is used for model training and testing, it will lead to too long model training time and too large workload. Therefore, in this embodiment, only five typical devices, namely Clothes Washer (CWE), Kitchen Fridge (FGE), Clothes Dryer (CDE), Dish Washer (DWE), and Heat Pump (HPE), are selected as target devices. In actual implementation, the target devices can be fewer or more, or can also include other types of electrical equipment. This embodiment does not limit the type of target devices. For example: The data of 10 months starting from May 2012 is used as the dataset, and at the same time, the dataset of 10 months is divided. The data of the first 6 months is used as the training set (training data), the data of the middle 2 months is used as the validation set, and the data of the last two months is used as the test set. In addition to the physical quantities directly collected in the dataset, during the experiment, some additional physical quantity data can also be obtained through indirect calculation by combining the mathematical relationships between physical quantities.

[0123] Step 2: Use the training data to iteratively train the fully convolutional neural network to obtain the initial parameters of the fully convolutional autoencoder network.

[0124] Step 3: Use the validation set to optimize and adjust the initial parameters to obtain the trained fully convolutional autoencoder network.

[0125] Optionally, after obtaining the trained fully convolutional autoencoder network, the performance of the fully convolutional autoencoder network can also be tested using the test set.

[0126] In this embodiment, the performance evaluation indicators of the fully convolutional autoencoder network include but are not limited to: Mean Absolute Error (MAE), signal aggregate error (SAE), and F1-Score.

[0127] When calculating the F1-Score, it is necessary to set the power-on threshold of the device according to the device characteristics of the dataset selected in this embodiment. For example, the power-on thresholds of each target device in this embodiment are set as shown in Table 2.

[0128] Table 2:

[0129]

[0130] In step 101, it is mentioned that there are many types of physical quantities that can be directly or indirectly obtained in the circuit, and feature selection is required. Assume that five physical quantities are used as inputs of the model through manual feature selection, and the performance of the model is tested. These five physical quantities and the reasons for their selection are as follows.

[0131] Active power (P): Since the physical quantity output by the model is active power, it is most intuitive to use active power as input information.

[0132] The active component of current (Ia), that is, the ratio of active power to voltage: In an AC power system, the voltage will fluctuate to a certain extent due to the influence of the power grid. Dividing the active power by the voltage to obtain the ratio can minimize the impact of voltage fluctuations.

[0133] The reactive component of current (Ir), that is, the ratio of reactive power to voltage: the reason is the same as that of the active component of current.

[0134] Admittance (Y): Admittance is a voltage-independent characteristic quantity whose value depends only on the properties of the device itself.

[0135] Power Factor (PF): The size of the power factor is closely related to the nature of the load in the circuit. Different electrical appliances have different inductive, capacitive and resistive components. The value of the power factor will change due to the different devices in the circuit. It is more suitable for distinguishing the types of devices contained in the circuit system and is more suitable as input information for non-intrusive load decomposition.

[0136] In addition to the above physical quantities, voltage data was also selected as the input information of the model to explore whether voltage (U) data can help improve the performance of the model. In the experiment, the washing machine was used as the target device, and different physical quantity combinations were used as the input information of the model for model training and testing. MAE, SAE and F1-Score were selected as performance evaluation indicators. The results are as follows: Figure 5 As shown. Figure 5 It can be seen that, except for voltage, the other physical quantities can help improve the model performance to varying degrees. The more types of physical quantities input, the better the model performance. Different physical quantity information improves the model performance to different degrees. From the experimental results, it can be seen that adding voltage data to the model input leads to a decrease in model performance, indicating that voltage data is not helpful for non-intrusive load monitoring tasks and is not suitable for using voltage data as model input information.

[0137] As can be seen from the above, for the non-intrusive load monitoring method provided in this embodiment, on the one hand, the input information adopted is data of multiple physical quantities such as active power and admittance, rather than only taking the single active power data as the input; on the other hand, the network structure adopts a fully convolutional neural network, which does not contain a fully connected layer. Based on this, the fully convolutional autoencoder network provided in this embodiment is compared with two traditional non-intrusive load monitoring methods using the above performance evaluation indicators. The two traditional non-intrusive load monitoring methods selected in this embodiment are: Autoregressive model (AR)-NILM and Fully Convolutional Networks (FCN)-Denoising autoencoders (dAE). Among them, the reason for selecting AR-NILM is that this traditional method provides the reactive power in the main line as an additional feature of the active power to the model, and it is convenient to change the input of this model to the same physical quantity input as the method in this embodiment. Using this method as a comparison method can verify whether the fully convolutional autoencoder network has better performance when using the same physical quantity information. The reason for selecting FCN-dAE is that the neural networks proposed in this method are all fully convolutional neural networks, but the input information of the model in this method is only the single active power information in the main line. Using this method as a comparison method for the method in this application can verify whether the method with multi-physical quantity input is more effective than the method with single-physical quantity input when using the fully convolutional neural network.

[0138] The comparison of the experimental results of the three methods on the AMPds2 dataset is shown in Table III below. It can be seen from Table III that for different devices and evaluation indicators, the method of this application shows the best performance among the three methods.

[0139] Table III:

[0140]

[0141] The performance optimization results of the method of this application compared with the other two methods are shown in Table IV below. In terms of a single device, the method of this application also has different degrees of performance improvement compared with FCN-dAE and AR NILM. The comparison results between this application and FCN-dAE can show that when using the fully convolutional neural network, the multi-physical quantity feature fusion method has better decomposition performance than the method using only a single physical quantity. The comparison results between this application and AR-NILM can show that when the input information of the model is the same, both are multi-physical quantity information, the network structure proposed in the method of this application has better performance than AR-NILM.

[0142] Table IV:

[0143]

[0144]

[0145] In summary, the non-intrusive load monitoring method based on multi-physical quantity fusion provided by this embodiment designs a fully convolutional autoencoder network constructed by convolutional layers and transposed convolutional layers. The network consists of multiple parallel feature extraction modules and a feature fusion module, and can infer the active power data of the target device in the corresponding time period through various different physical quantity data in the main line; it can solve the problems that the model parameters of the traditional power decomposition model are relatively large, resulting in a large amount of computing resources occupied by the model and a low training efficiency of the model; since the fully connected layer is not used in the network structure and convolutional or transposed convolutional operations are used to replace the fully connected operation to process the data, the number of model parameters can be greatly reduced, thereby saving the computing resources of the model and improving the training efficiency of the model.

[0146] In addition, by selecting five physical quantities from more than ten physical quantities that can greatly improve the model performance, it can be ensured that the physical quantities input into the model are all physical quantities that contribute to predicting the active power value, which can not only improve the accuracy of model calculation, but also avoid wasting the computing resources of the model.

[0147] Optionally, the present application also provides a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the non-intrusive load monitoring method based on multi-physical quantity fusion in the above method embodiment.

[0148] Optionally, the present application also provides a computer product, which includes a computer-readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the non-intrusive load monitoring method based on multi-physical quantity fusion in the above method embodiment.

[0149] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0150] The above embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A non-invasive load monitoring method based on multi-physical quantity fusion, characterized in that The method includes: Obtaining N physical quantities of electrical energy data in the main line, where N is an integer greater than 1; Inputting the N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device; Wherein, the fully convolutional autoencoder network includes N feature extraction modules, a feature fusion module connected to the N feature extraction modules, and an output layer connected to the feature fusion module; Each feature extraction module is used to extract features of a corresponding physical quantity, and different feature extraction modules correspond to different physical quantities; the feature extraction module is established based on an autoencoder; The feature fusion module is used to fuse the feature values output by the N feature extraction modules to obtain a feature sequence matching the data size of the output layer; the feature fusion module is established based on N sequentially connected transposed convolutional layers; The output layer is used to output the active power value based on the feature sequence; The autoencoder includes an encoder and a decoder connected to the encoder; The encoder includes 3 sequentially connected one-dimensional convolutional layers, and after the data input to the encoder passes through the 3 one-dimensional convolutional layers, the size of the feature decreases sequentially; The decoder includes 4 sequentially connected one-dimensional transposed convolutional layers, and after the data input to the decoder passes through the 4 one-dimensional transposed convolutional layers, the size of the feature is sequentially restored to the original size of the data input to the autoencoder; In the N sequentially connected transposed convolutional layers, the size of the convolutional kernel in the time direction of each transposed convolutional layer is not 1, and each transposed convolutional layer does not use an activation function; The number of convolutional kernels in each transposed convolutional layer is 1 more than the number of convolutional kernels in the next transposed convolutional layer.

2. The method according to claim 1, wherein The convolutional kernel sizes of the 3 sequentially connected one-dimensional convolutional layers are all 8, and the number of convolutional kernels in each one-dimensional convolutional layer is 1 / 2 of the number of convolutional kernels in the next one-dimensional convolutional layer; each one-dimensional convolutional layer includes a rectified linear unit (relu) function and has a stride of 1; The convolutional kernel sizes of the 4 sequentially connected one-dimensional transposed convolutional layers are all 8, and the number of convolutional kernels in the first one-dimensional transposed convolutional layer is 2 times the number of convolutional kernels in the second one-dimensional transposed convolutional layer; each one-dimensional transposed convolutional layer includes a sigmoid function and has a stride of 1.

3. The method according to claim 1, wherein The convolutional kernel size in each transposed convolutional layer is 8 and the stride is 1.

4. The method according to any one of claims 1 to 3, characterized in that Before obtaining the N physical quantities of electrical energy data in the main line, it further includes: Determining the types of the N physical quantities.

5. The method according to claim 4, wherein The determining the types of the N physical quantities includes: Determining the initial types of the physical quantities, where the initial types include at least two of active power, active component of current, reactive component of current, admittance, and power factor; Training the fully convolutional autoencoder network with the physical quantities of the initial types to obtain the first result value of the model evaluation index; Retrain the fully convolutional autoencoder network using the physical quantity of the initial type and other types of physical quantities to obtain a second result value of the model evaluation index; wherein, the other types are different from the initial type, and the other types are data types that can be collected from the main line or data types that can be calculated based on the electrical energy data collected from the main line; When the second result value is greater than the first result value, add the other type to the initial type and execute again the step of retraining the fully convolutional autoencoder network using the physical quantity of the initial type and other types of physical quantities to obtain a second result value of the model evaluation index; When the second result value is less than or equal to the first result value, determine the initial type as the type of the N physical quantities.

6. The method according to claim 4, characterized in that, The determining the type of the N physical quantities includes: Collect the electrical energy data to obtain a first physical quantity; Calculate second physical quantities of each type using the electrical energy data; Score each first physical quantity and second physical quantity using Pearson correlation coefficient, chi-square verification, mutual information, maximum information coefficient, and distance correlation coefficient; Determine the types of the first physical quantity and / or second physical quantity greater than a preset scoring threshold as the types of the N physical quantities.

7. The method according to claim 4, characterized in that The determining the type of the N physical quantities includes: Collect the electrical energy data to obtain a first physical quantity; Calculate second physical quantities of each type using the electrical energy data; Input the first physical quantity and the second physical quantity into a pre-trained weight calculation model to obtain weight coefficients of each physical quantity; Select the types of the N physical quantities in descending order of the weight coefficients.

8. The method according to any one of claims 1 to 3, characterized in that The inputting the N physical quantities into a pre-trained fully convolutional autoencoder network to obtain the active power value of the target device includes: Process the abnormal data in the N physical quantities, where the abnormal data refers to physical quantities exceeding the normal range, and the normal ranges corresponding to different types of physical quantities are different; Perform standardization processing on the processed physical quantities to obtain N standardized physical quantities; Correspondingly input the N standardized physical quantities into N feature extraction modules for the fully convolutional autoencoder network to perform load monitoring calculation to obtain the active power value of the target device.

Images