A non-intrusive load identification method and device and a storage medium

Abstract

The application relates to a non-invasive load identification method and device based on an optimized convolutional neural network and a storage medium, wherein the method comprises the following steps: acquiring electric appliance data information; determining a V-I trajectory graph; establishing an optimized convolutional neural network load classification model, which is composed of an optimized convolutional neural network, wherein a double-pooling mixed channel is adopted in the pooling layer of the optimized convolutional neural network, and the double-pooling mixed channel weight is distributed through self-learning weighting mask; initializing hyperparameters; training the optimized convolutional neural network load classification model; optimizing the hyperparameters based on a particle swarm optimization algorithm and a training result; judging whether an error function converges or not, if the error function converges, the optimization is completed; if the error function does not converge, the hyperparameters are updated, and the hyperparameter optimization is performed again; and inputting the V-I trajectory graph into the optimized convolutional neural network load classification model which is optimized, so that electric appliance classification results are obtained. Compared with the prior art, the application has the advantages of good classification and identification performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-intrusive load monitoring, in particular to a non-intrusive load identification method and device based on an optimized convolutional neural network and a storage medium. BACKGROUND

[0002] Non-intrusive load monitoring is cost-effective and increasingly widely used. At present, the important and difficult research on load identification under load monitoring is to install an intelligent electric meter at the power supply inlet and extract the characteristics of electrical appliances to identify the types of electrical appliances. The voltage and current at stable working time are used to draw a V-I trajectory load characteristic. The characteristic parameters extracted from the voltage-current (V-I) trajectory graph are used as the basic data for identifying the load. Fuzzy clustering is used to train the V-I trajectory graph of the load to complete the load identification. Or the steady-state current time domain and frequency domain are used as load characteristics. The good data set is screened through random forest, and the extreme learning machine algorithm is optimized through genetic algorithm to identify and solve the defects of poor traditional separability.

[0003] With the development of modern technology, household appliances are increasingly intelligent and diversified, and the V-I trajectory load characteristics exhibited are increasingly complex and diverse, which is a great test for load identification method research. The existing identification method is difficult to maintain stable classification performance, hindering the development of load monitoring technology and making it difficult to ensure the safety of the residential power environment.

[0004] CN202011443609.8 discloses a non-intrusive load identification method based on V-I trajectory graph and neural network. The method comprises: collecting real-time voltage and current data and active power data; judging whether there is a load switching event and whether the load operating state has reached a stable state according to the change of active power; obtaining the voltage and current data and power data of the load according to the steady-state voltage and current data before and after the event; converting the V-I trajectory into an RGB color image containing voltage and current phase difference, power and other information by using simple image processing technology. After obtaining the RGB color image, normalization processing is performed; load identification is performed through a pre-trained convolutional neural network. The steady-state characteristics of the load are extracted through the convolutional neural network, and the neural network model can be directly run on an embedded device without relying on the operation support of a server. However, the above method still lacks self-learning feature information extraction capability, thereby affecting the overall identification and classification performance. SUMMARY

[0005] The application aims to provide a non-intrusive load identification method and device based on an optimized convolutional neural network and a storage medium.

[0006] The application aims to provide a non-intrusive load identification method and device based on an optimized convolutional neural network and a storage medium.

[0007] A non-intrusive load identification method based on an optimized convolutional neural network comprises the following steps:

[0008] Obtaining electrical appliance data information;

[0009] Determining a V-I trajectory graph based on the electrical appliance data information;

[0010] Establishing an optimized convolutional neural network load classification model, wherein the input of the optimized convolutional neural network load classification model is the V-I trajectory graph, the output is the electrical appliance classification result, and the optimized convolutional neural network is composed of a convolutional layer, a pooling layer, a fully connected layer and a softmax layer connected in sequence, the pooling layer adopts a double-pooling mixed channel, and the double-pooling mixed channel weight is distributed by self-learning weighted mask;

[0011] Initializing the hyperparameters of the optimized convolutional neural network load classification model;

[0012] Training the optimized convolutional neural network load classification model based on the training set data to obtain a training result, wherein the training set data comprises a set of V-I trajectory graphs of different electrical appliances in different working states;

[0013] Optimizing the hyperparameters based on the particle swarm optimization algorithm and the training result, wherein the parameters of the particle swarm optimization algorithm comprise an inertia weight, a learning factor and an error function;

[0014] Determining whether the error function of the particle swarm optimization algorithm converges, if the error function converges, outputting the optimized hyperparameters to obtain the optimized convolutional neural network load classification model, if the error function does not converge, updating the hyperparameters and re-optimizing the hyperparameters;

[0015] Inputting the V-I trajectory graph into the optimized convolutional neural network load classification model to obtain the electrical appliance classification result.

[0016] The convolutional layer utilizes discrete convolution to perform convolutional processing on the V-I trajectory graph to obtain feature points, and performs inner product operation on the pixels of the V-I trajectory graph to generate a feature mapping graph.

[0017] The pooling layer reduces the sampling by a double-pooling mixed channel, and scales the feature map based on the feature points after convolution of the convolution layer, extracts core features and retains background information.

[0018] The output of the double-pooling mixed channel is:

[0019]

[0020] wherein D awp (X) is the output of the double-pooling mixed channel; S(f aw ) is a self-learning weighted mask; max(X i ) is the maximum pooling, is the average pooling.

[0021] The error function is the fitness value.

[0022] The pooling layer and the fully connected layer are further provided with a Dropout layer.

[0023] The hyperparameters include the number and size of the convolution kernel of each convolution layer, the probability of the Dropout layer and the learning rate of the CNN.

[0024] The softmax layer is used to output the appliance classification result, wherein the classification result includes a pre-configured number of intervals, each interval representing an appliance, and the softmax layer determines the appliance classification result based on the maximum probability in the probability of each corresponding interval.

[0025] A non-intrusive load identification device based on an optimized convolutional neural network, comprising a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described above.

[0026] A storage medium having a program stored thereon, wherein the program is executed to implement the method described above.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] (1) The pooling layer of the convolutional neural network is optimized in the present application, a double-pooling mixed channel is adopted, and the double-pooling channel weight is allocated by self-learning through a weighted mask, so that the advantages of traditional maximum pooling and average pooling feature extraction can be combined, and when extracting the appliance feature map, not only the main core feature information can be extracted, but also the overall background feature information can be retained.

[0029] (2) The application adopts a particle swarm optimization algorithm to optimize the hyperparameters of the convolutional neural network, takes the fitness value as the evaluation standard for the advantages and disadvantages of the hyperparameters, automatically finds the optimal hyperparameters, overcomes the uncertainty problem of the classification performance of the traditional artificial network training, makes the overall classification performance tend to be optimized, and improves the overall self-learning feature extraction capability. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A flowchart of the method of the application;

[0031] Figure 2 A schematic diagram of the data acquisition platform of the application;

[0032] Figure 3 An electric appliance working current waveform diagram of an embodiment;

[0033] Figure 4 An electric appliance working voltage waveform diagram of an embodiment;

[0034] Figure 5 V-I trajectory diagrams of common household appliances, wherein (a) is a V-I trajectory diagram of an electric rice cooker, (b) is a V-I trajectory diagram of an electric fan heater, (c) is a V-I trajectory diagram of a vacuum cleaner, (d) is a V-I trajectory diagram of an induction cooker, (e) is a V-I trajectory diagram of a hair dryer working at cold wind two gears, and (f) is a V-I trajectory diagram of a hair dryer working at high temperature two gears;

[0035] Figure 6 A convolution process schematic diagram of a convolutional layer;

[0036] Figure 7 A double-pooling channel schematic diagram;

[0037] Figure 8 A fitness value change curve schematic diagram. DETAILED DESCRIPTION

[0038] The application will be described in detail below in combination with the drawings and specific embodiments. The embodiments are implemented on the premise of the technical solution of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0039] A non-intrusive load identification method based on an optimized convolutional neural network, as shown in Figure 1 includes the following steps:

[0040] 1) Obtain electric appliance data information;

[0041] First, a data acquisition platform is built, as shown in Figure 2As shown, the wall power supply is connected through the power strip to provide power supply for the data collector, household appliances and collection device. In order to highlight the characteristic information of the appliances, the data information of the appliances is collected through the connection of the mutual inductor at a low frequency of 12.5Khz. The collected data information of the appliances includes voltage and current. In an embodiment, taking a rice cooker as an example, the obtained working current waveform diagram of the appliance is as shown in Figure 3 As shown, the working voltage waveform diagram of the appliance is as shown in Figure 4

[0042] 2) Based on the voltage and current data information of the appliance in the steady state, a V-I trajectory diagram is drawn with voltage as the horizontal axis and current as the vertical axis;

[0043] Figure 5 The V-I trajectory diagram of the commonly used household appliance. The determination principle of the commonly used household appliance is the frequency of use of the appliance and whether it is representative, wherein the meaning of being representative can be whether it is universal.

[0044] Modern intelligent household appliances include appliances with a single working state and appliances with multiple working states. Taking a hair dryer as an example, the V-I trajectory diagram of the hair dryer when it works in different working states (such as cold wind mode and high temperature mode) has different characteristics, so for such household appliances, the V-I trajectory diagram of the household appliances in different working states should be drawn respectively.

[0045] 3) Establish an optimized convolutional neural network load classification model

[0046] The input of the optimized convolutional neural network load classification model is the V-I trajectory diagram, and the output is the classification result of the appliance, which is composed of an optimized convolutional neural network, wherein the optimized convolutional neural network includes a convolutional layer, a pooling layer, a fully connected layer and a softmax layer connected in turn.

[0047] The convolutional layer uses discrete convolution to perform convolution processing on the V-I trajectory diagram to obtain feature points, and performs inner product operation on the pixels of the V-I trajectory diagram to generate a feature mapping diagram, and the convolution process of the convolutional layer is as shown in Figure 6

[0048] The pooling layer adopts a double-pooling mixed channel, as shown in Figure 7 The double-pooling mixed channel weights are distributed by self-learning weighted masks, and based on the feature points after convolution of the convolutional layer, the feature mapping diagram is scaled, the core features are extracted and the background information is retained.

[0049] The output of the double-pooling mixed channel is:

[0050]

[0051] Wherein, D awp ​​(X) is a double-pooling mixed channel output; S(f aw ) is a self-learning weighted mask; max(X i ) is max-pooling, is average-pooling.

[0052] The Dropout layer is further arranged between the pooling layer and the fully connected layer, and is used to prevent overfitting.

[0053] The softmax layer is used to output the electrical appliance classification result, wherein the classification result includes a preconfigured number of intervals, each interval represents an electrical appliance, and the softmax layer determines the electrical appliance classification result based on the maximum probability in the probability of each corresponding interval.

[0054] 4) Initialize the hyperparameters of the optimized convolutional neural network load classification model;

[0055] 5) Train the optimized convolutional neural network load classification model based on the training set data, and obtain a training result, wherein the training set data includes a set of V-I trajectory graphs of different electrical appliances in different working states;

[0056] 6) Optimize the hyperparameters based on the particle swarm optimization algorithm and the training result, wherein the parameters of the particle swarm optimization algorithm include an inertia weight, a learning factor, and an error function, and in this embodiment, the error function is a fitness value;

[0057] 7) Determine whether the error function of the particle swarm optimization algorithm converges, if the error function converges, output the optimized hyperparameters, and obtain the optimized convolutional neural network load classification model; if the error function does not converge, update the hyperparameters, and re-optimize the hyperparameters;

[0058] 8) Input the V-I trajectory graph into the optimized optimized convolutional neural network load classification model, and obtain the electrical appliance classification result.

[0059] If the above functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0060] In the present embodiment, an optimized convolutional neural network load classification model is established and trained by Matlab software, and based on the V-I trajectory load characteristics, the load identification research is completed. The to-be-optimized hyperparameters in the optimized convolutional neural network load classification model and the ranges thereof are shown in Table 1.

[0061] Table 1 to-be-optimized hyperparameters and ranges thereof

[0062]

[0063] The inertia weight of the particle swarm optimization algorithm is set to 0.8, the learning factor is 1.45449, and when the fitness value gradually converges to zero, as shown in Figure 8 , the hyperparameter optimization process is ended, and the hyperparameter configuration tending to be optimized is obtained, as shown in Table 2.

[0064] Table 2 hyperparameter optimization results

[0065]

[0066] Finally, through experimental verification and comparison, the recognition accuracy comparison is shown in Table 3, and according to Table 3, it can be concluded that the method of the present application can significantly improve the classification performance.

[0067] Table 3 recognition accuracy comparison

[0068]

[0069] The preferred embodiments of the application have been described above in detail. It should be understood that modifications and variations to the preferred embodiments could be made by those skilled in the art in light of the teachings above. It is therefore contemplated that the application can encompass other variations and modifications that fall within the scope of the claims.

Claims

1. A non-intrusive load identification method based on an optimized convolutional neural network, characterized in that, The method comprises the following steps: obtaining electrical appliance data information; determining a V-I trajectory graph based on the electrical appliance data information; establishing an optimized convolutional neural network load classification model, wherein the input of the optimized convolutional neural network load classification model is the V-I trajectory graph, the output is an electrical appliance classification result, and the optimized convolutional neural network is composed of a convolutional layer, a pooling layer, a fully connected layer and a softmax layer connected in sequence, the pooling layer adopts a double-pooling mixed channel, and the double-pooling mixed channel weight is distributed by self-learning weighting mask; the output of the double-pooling mixed channel is: wherein, is a dual-pooling mixed channel output; is a self-learning weighted mask; is a max-pooling, is an average-pooling; initializing the hyperparameters of the optimized convolutional neural network load classification model; training the optimized convolutional neural network load classification model based on the training set data to obtain a training result, wherein the training set data comprises a set of V-I trajectory graphs of different electrical appliances in different working states; optimizing the hyperparameters based on a particle swarm optimization algorithm and the training result, wherein the parameters of the particle swarm optimization algorithm comprise an inertia weight, a learning factor and an error function; determining whether the error function of the particle swarm optimization algorithm converges, if the error function converges, outputting the optimized hyperparameters to obtain an optimized optimized convolutional neural network load classification model, and if the error function does not converge, updating the hyperparameters and re-optimizing the hyperparameters; inputting the V-I trajectory graph into the optimized optimized convolutional neural network load classification model to obtain the electrical appliance classification result. 2.The non-intrusive load identification method based on an optimized convolutional neural network according to claim 1, characterized in that, The convolutional layer utilizes discrete convolution to perform convolutional processing on the V-I trajectory graph to obtain feature points, and performs inner product operation on the pixels of the V-I trajectory graph to generate a feature mapping graph. 3.The non-intrusive load identification method based on the optimized convolutional neural network according to claim 2, characterized in that, The pooling layer performs down-sampling through a double-pooling mixed channel, scales the feature mapping graph based on the feature points after convolutional processing of the convolutional layer, extracts core features and retains background information. 4.The non-intrusive load identification method based on an optimized convolutional neural network according to claim 1, characterized in that, The error function is an adaptability value.

5. The non-intrusive load identification method based on the optimized convolutional neural network according to claim 1, characterized in that, The pooling layer and the fully connected layer are further provided with a Dropout layer.

6. The non-intrusive load identification method based on an optimized convolutional neural network according to claim 5, characterized in that, The hyperparameters comprise the number and size of the convolutional kernels of each convolutional layer, the probability of the Dropout layer and the CNN learning rate.

7. The non-intrusive load identification method based on the optimized convolutional neural network according to claim 1, characterized in that, The softmax layer is used to output the electrical appliance classification result, wherein the classification result comprises a plurality of pre-configured intervals, each interval represents an electrical appliance, and the softmax layer determines the electrical appliance classification result based on the maximum probability in the probability of each corresponding interval.

8. An apparatus for non-intrusive load identification based on an optimized convolutional neural network, comprising a memory, a processor, and a program stored in the memory, wherein the program comprises the following steps of: The processor implements the method of any one of claims 1-7 when executing the program.

9. A storage medium having stored thereon a program, characterized by The program is executed to implement the method of any one of claims 1-7.

Images