Non-intrusive load disaggregation method based on wavelet decomposition and improved neural network

Abstract

The application discloses a non-intrusive load decomposition method based on wavelet decomposition and an improved neural network, mainly uses a non-intrusive load identification device to collect power consumption information, carries out pretreatment operation on the collected data information, extracts load characteristics through four-layer wavelet decomposition, and then inputs the load characteristics into an improved neural network for training and learning; the improvement mainly lies in that the network is multi-scale and multi-input, the training process is multi-resolution, and the network is divided into four parts according to the wavelet decomposition layers and is trained respectively, is spliced again finally, then a large amount of data is used to adjust network parameters, a non-intrusive load decomposition model is perfected, a load decomposition task is completed, and the non-intrusive load decomposition result is analyzed. Through the application, the time-frequency domain can be combined, the problem that important information can be lost, data characteristic utilization is low and a multi-state process electric appliance is difficult to identify when load data is too much can be solved.

Description

Technical Field

[0001] This invention relates to the technical field of power load decomposition, and in particular to a non-intrusive load decomposition method based on wavelet decomposition and improved neural networks. Background Technology

[0002] With the continuous development of technology, electricity accounts for an increasingly larger proportion of final energy consumption. Therefore, the efficient use of electricity will bring significant energy savings, which has spurred the implementation and innovation of smart grid technology. A key feature of smart grid technology is the integration of smart meters, which provides energy consumers and suppliers with comprehensive information on appliance power consumption. In fact, some studies have shown that providing comprehensive appliance power consumption information can promote potential energy savings of over 20%. Furthermore, the availability of power consumption information offers significant benefits for power system operation and dispatch, as well as grid planning. Therefore, it can improve asset utilization and energy efficiency. An important method for acquiring and analyzing power consumption information is Non-Intrusive Load Monitoring (NILM), which infers the power consumption of individual appliances based on the overall power consumption measured at a single point, rather than installing multiple meters. Numerous existing projects have demonstrated that personalized energy efficiency services based on NILM technology can reduce household electricity consumption by approximately 14%. Therefore, NILM is considered a novel and cost-effective technology for monitoring detailed load power consumption and has been rapidly implemented in recent years.

[0003] The non-intrusive load identification device consists of a smart energy measurement device located at the entrance of the home. For example... Figure 1 As shown, this device can monitor the power consumption of various loads under it, and in practice, its functions can also be integrated into smart meter modules. The non-intrusive load identification device can obtain rich electrical information, such as total load power, voltage, and current, and includes information on load components with different characteristics. By extracting the characteristic information of these electrical quantities, the non-intrusive load identification device can decompose the load and estimate the switching status and power consumption of each electrical device.

[0004] Based on the above discussion, the invention of a non-invasive load decomposition method that satisfies both high precision and high efficiency has high practical application value. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a non-intrusive load decomposition method based on wavelet decomposition and improved neural network. This method can be applied to non-intrusive online monitoring of residential electricity loads and has the characteristics of low cost, easy implementation, and high decomposition accuracy. By combining the time and frequency domains, this invention can solve the problems that important information may be lost, data feature utilization is low, and multi-state process electrical appliances are difficult to identify when there is too much load data.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a non-intrusive load decomposition method based on wavelet decomposition and improved neural networks, comprising the following steps:

[0007] 1) Use a non-intrusive load identification device to collect power consumption information on the load side, perform data preprocessing to obtain two-dimensional image sequence load data, and divide the data into training set and test set according to the proportion;

[0008] 2) The preprocessed load data is decomposed into load features using a four-level wavelet decomposition. Each level of wavelet decomposition can obtain low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features.

[0009] 3) The load feature information obtained through four-layer wavelet decomposition is input into the constructed non-intrusive load decomposition model for decomposition training. The model parameters are continuously adjusted using data from the training set to improve the non-intrusive load decomposition model. Then, the model is validated using data from the test set, and the non-intrusive load decomposition results are analyzed. The constructed non-intrusive load decomposition model is an improved neural network. The improvements of this neural network include: First, the network is multi-scale and multi-input, and the training process is multi-resolution; Second, the input images of four wavelet input layer channels are obtained from four-layer wavelet decomposition, and a training network composed of CNN, BiLSTM, and Attention is connected to each of the four wavelet input layer channels for training; Third, the results of the four inputs at different scales after training are concatenated through a concatenation layer to achieve wavelet reconstruction.

[0010] Furthermore, in step 1), power data is collected at the household power inlet by power sensors deployed in a specific area to provide a data source for non-intrusive load research. The collected data consists of the total power time series and the power time series of individual electrical appliances. In order to obtain more noise-resistant, higher quality data and better training results, the following preprocessing operations are required.

[0011] The first step is to clear outliers in the data. Outliers may be caused by errors in the measuring device or environmental factors. The formula for clearing outliers is as follows:

[0012] P t -P t±1 >P th or P t >P max

[0013] In the formula, P t P is the power consumption value of a single electrical device at time t. t±1P is the power consumption value of a single electrical device at time t±1. th and P max These are the switching power value and maximum power value of a single electrical device, respectively.

[0014] Secondly, data standardization is performed to make the data easier for neural networks to train, which is beneficial for obtaining better results. The standardization formula is expressed as follows:

[0015]

[0016] In the formula, x t This represents the power value at time t; σ and σ represent the average power and standard deviation of the electrical equipment, respectively, and x is the power data after standardization.

[0017] Secondly, sliding windowing is applied to the data to improve the training speed of the neural network. Sliding windowing divides the data according to the length of the window, as shown in the following formula:

[0018] N = N W -W+1

[0019] In the formula, N is the number of processed load time series. W W represents the number of load time series before processing, and W is the window length that needs to be segmented.

[0020] Next, the data dimension is transformed, converting the one-dimensional time series load data into a two-dimensional image sequence load data. Since the load data is now divided into N data points of window length W, the dimension transformation only requires copying these N load data points W times, so that the load data changes from a one-dimensional W time series to a W*W image sequence.

[0021] Finally, the transformed data is divided into training and testing sets in a 7:3 ratio to complete the database construction.

[0022] Furthermore, in step 2), the image sequence payload data is decomposed using Discrete Wavelet Transform (DWT). Each p×q image undergoes one DWT transformation to obtain four sub-images of size (p / 2)×(q / 2), and the decomposition formula is as follows:

[0023]

[0024] In the formula, j represents the decomposition level; k and m are the row and column indices, respectively; l and n are the translation factors; h and g represent the low-pass and high-pass filters, respectively; and c represents the image at level j. j After one wavelet decomposition, four sub-images c are obtained. j+1 d j+1,hd j +1,v d j+1,d These include the low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features of the image, respectively.

[0025] Furthermore, in step 3), firstly, in order to fully utilize the features in each direction and improve the feature utilization rate, a multi-scale and multi-resolution method is adopted to improve the neural network. The idea of ​​multi-scale is reflected in the fact that the input consists of four data of different lengths, which also happens to conform to the fact that each wavelet decomposition layer halves the input length. The four-layer wavelet decomposition can construct different input lengths and construct input images for four wavelet input layer channels. The idea of ​​multi-resolution is reflected in the fact that the network training is performed with two-dimensional images, and the resolution of the images changes with the modules of the neural network.

[0026] Secondly, the four-layer wavelet decomposition yields four wavelet input layer channels. These four wavelet input layer channels are paired with four training networks consisting of CNN, BiLSTM, and Attention. That is, a training network is connected after each wavelet input layer channel for training. The purpose of using CNN is to further extract features from the data, achieving data compression and ensuring that the total input and output lengths are consistent. BiLSTM is chosen due to its advantages in time series problems. Attention is used to assign spatial attention weights, giving higher weights when the device is on or off, allowing the network to learn according to the weight allocation, making it easier to recognize the multiple states of the device and improving network accuracy.

[0027] Next, the four trained networks are concatenated by a concatenate connection layer, and CNN is used again to ensure that the lengths of the input and output are consistent, thereby ensuring the single-branch reconstruction of the wavelet;

[0028] Finally, the four trained networks are concatenated through a concatenate layer and then fed into a fully connected layer to achieve a non-linear mapping from deep features to the target electrical power. Combined with the seq2point learning model, the decomposition result is obtained. There are three fully connected layers with dimensions of 258, 128, and 1, respectively. The last fully connected layer and the output constitute the seq2point model.

[0029] Furthermore, the improved neural network includes the following modules:

[0030] Module 1, the repetition layer, can transform the load data from a one-dimensional W time series into a two-dimensional W×W image sequence, where W is the length of the input;

[0031] Module 2, Wavelet Input Layer 1, represents performing one wavelet decomposition;

[0032] Module 3, wavelet input layer 2, represents performing two wavelet decompositions;

[0033] Module 4, wavelet input layer 3, represents performing three wavelet decompositions;

[0034] Module 5, wavelet input layer 4, represents performing four wavelet decompositions;

[0035] Module 6, CNN, stride=2, kernel size=3, activation function is ReLU;

[0036] Module 7, BiLSTM, units = 64, activation function is Tanh;

[0037] Module 8 is an attention mechanism module;

[0038] Module 9 is the Concatenate connection layer;

[0039] Module 10 is a three-layer fully connected layer with dimensions of 258, 128, and 1 respectively;

[0040] Module 11, training layer 1, is composed of modules 2, 6, 7, 8, and 9 connected in sequence;

[0041] Module 12, training layer 2, is composed of modules 3, 6, 7, 8, and 9 connected in sequence;

[0042] Module 13, training layer 3, is composed of modules 4, 6, 7, 8, and 9 connected in sequence;

[0043] Module 14, training layer 4, is composed of modules 5, 6, 7, 8, and 9 connected in sequence;

[0044] The improved neural network's total loss function is composed of the loss functions of four training layers concatenated. Each training layer's loss function includes the loss functions of CNN, BiLSTM, and Attention. Finally, a fully connected layer is used to obtain the final loss function. The formula for each training layer is as follows:

[0045] loss train =loss(Attention(BiLSTM(CNN(X))))

[0046] In the formula, X represents the input load power value, and loss train The loss function representing the training layer;

[0047] Therefore, the total loss function is expressed as:

[0048] loss total =loss(Dense(Cat(Cat(Cat(losstrain1 loss train2 loss train3 loss train4 )))

[0049] In the formula, loss total Represents the total loss function, Dense represents a fully connected layer, Cat represents a connected layer, and loss train1 loss train2 loss train3 loss train4 This represents the loss function for training layers 1, 2, 3, and 4.

[0050] Furthermore, in step 3), the load feature information obtained through four-layer wavelet decomposition is input into the non-intrusive load decomposition model for decomposition training. The parameters of the training neural network are set as follows: Batch size is set to 512, training iterations are 250, the Adam optimizer is used, the initial learning rate is 0.0001, the decay rate is 0.1, the decay step size is 1000, and the loss function is the minimum mean square error (MSE). An early stopping mechanism is also added, that is, when the mean square error stops decreasing after 10 iterations, the network will automatically stop training. By continuously optimizing and adjusting the model parameters, the model accuracy can be improved. After training, the model parameters with the smallest loss on the validation set are selected from the saved training parameters as the model parameters for the detection model. Then, the data in the test set is input into the trained model to obtain the user's decomposition results.

[0051] Furthermore, in step 3), to more intuitively demonstrate the performance of the model in non-invasive load decomposition, regression metrics, including Mean Absolute Error (MAE), Normalized Total Error (SAE), and classification metrics F1, Accuracy, Precision, and Recall, are used for evaluation.

[0052]

[0053] In the formula, and x t Let T represent the predicted and actual power values ​​of a certain electrical appliance at time t, where T is the sample size.

[0054]

[0055] In the formula, and This represents the actual total power and the predicted total output power of an electrical appliance over a time period t. and

[0056]

[0057]

[0058]

[0059] In the formula, precision(c j ) represents accuracy; recall(c j F1-Score(c) represents recall rate; j TP(c) is a combined metric of accuracy and recall. j 0 indicates category c j Number of identified samples; FN9c j 0 indicates category c j The number of unidentified samples; FP(c j ) is identified as c j Category but not belonging to c j The number of samples.

[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0061] 1. The data preprocessing method of the present invention, especially the method for removing data outliers, can remove measurement outliers caused by non-intrusive load identification devices or environmental factors, making the data more noise-resistant and the load characteristics more obvious.

[0062] 2. The wavelet decomposition feature extraction method proposed in this invention can fully combine the features of the time and frequency domains, and solve the problems that important information may be lost, data feature utilization is low, and multi-state process electrical appliances are difficult to identify when there is too much load data.

[0063] 3. The data dimension conversion method and multi-layer wavelet decomposition method proposed in this invention can convert one-dimensional time series into two-dimensional image sequences, and can further obtain more component feature information of the images at different scales in the horizontal, vertical and diagonal directions.

[0064] 4. The non-intrusive load decomposition model proposed in this invention can fully utilize the feature information of multi-layer wavelet decomposition, improve the feature utilization rate, and further extract features and ensure the consistency of input and output scales by using convolution in training the network. It leverages the advantages of bidirectional long short-term memory networks in solving sequence problems, which is conducive to improving the accuracy of load decomposition. The attention mechanism is used to filter the network features, multiply them by different weights, and increase the weight of important features, which is conducive to using features to detect multi-state process equipment when the equipment is switched on and off and to improve the network training speed. Attached Figure Description

[0065] Figure 1 This is a hardware diagram of a non-intrusive load identification device.

[0066] Figure 2 This is a flowchart of the method of the present invention.

[0067] Figure 3 Comparison images showing the removal of abnormal points in the refrigerator.

[0068] Figure 4 This is a feature information map obtained from two-level wavelet decomposition.

[0069] Figure 5 This is a structural diagram of a non-intrusive load decomposition model.

[0070] Figure 6 This is a detailed exploded view of a section of the dishwasher.

[0071] Figure 7 This is a detailed exploded view of a portion of the refrigerator.

[0072] Figure 8 This is a detailed exploded view of a part of the washing machine.

[0073] Figure 9 This is a detailed exploded view of a portion of the microwave oven. Detailed Implementation

[0074] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0075] like Figure 2 As shown, this embodiment discloses a non-intrusive load decomposition method based on wavelet decomposition and improved neural networks, including the following steps:

[0076] 1) Power data is collected at the household power inlet using power sensors deployed in specific areas, providing a data source for non-invasive load studies. The collected data consists of total power time series and power time series of individual electrical appliances. To obtain more robust, higher-quality data and better training results, the following preprocessing operations are required:

[0077] The first step is to clear outliers from the data. Outliers may be caused by errors in the measuring device or environmental factors. Taking the collected refrigerator load data as an example, the comparison before and after clearing outliers is as follows: Figure 3 As shown, the formula for clearing outliers can be expressed as follows:

[0078] P t -P t±1 >P th or P t >P max

[0079] Among them, P t P is the power consumption value of a single electrical device at time t. t±1 P is the power consumption value of a single electrical device at time t±1. th and P max These are the switching power value and maximum power value of a single electrical device, respectively.

[0080] Secondly, data standardization is performed to make the data easier for neural networks to train, which is beneficial for obtaining better results. The standardization formula is expressed as follows:

[0081]

[0082] Where, x t This represents the power value at time t; σ and σ represent the average power and standard deviation of the electrical equipment, respectively, and x is the power data after standardization. The average power, standard deviation, maximum value and switching value of typical equipment are shown in the table below.

[0083] Apparatus Switch value Average power (w) Maximum power (w) Standard deviation Kettle 200 700 3948 1000 Microwave oven 200 500 3138 800 Refrigerator 50 200 2572 400 Dishwasher 10 700 3230 1000 Washing machine 20 400 3962 700

[0084] Secondly, sliding windowing is applied to the data to improve the training speed of the neural network. Sliding windowing divides the data according to the length of the window, as shown in the following formula:

[0085] N = N W -W+1

[0086] Where N is the number of processed load time series, N W W represents the number of load time series before processing, and W is the window length that needs to be segmented.

[0087] Next, the data dimension is transformed, converting the one-dimensional time series load data into a two-dimensional image sequence load data. Since the load data is now divided into N data points of window length W, the dimension transformation only requires copying these N load data points W times, so that the load data changes from a one-dimensional W time series to a W*W image sequence.

[0088] Finally, the transformed data was divided into training and testing sets in a 7:3 ratio. To make the load decomposition results more obvious and demonstrate the advantages of the proposed method, we used several typical load devices, such as refrigerators, washing machines, dishwashers, and microwave ovens, for decomposition and used this to complete the database construction.

[0089] 2) Regarding feature extraction, considering the strong fluctuations in load power data, along with its periodicity and time correlation, most existing methods only involve the time domain. When the sequence is too long, important information may be lost, and it is difficult to identify the multi-state processes of electrical appliances. Wavelet decomposition can extract both time and frequency domain information from a given time series, while also providing data compression. This solves the problems of lost time information and excessive training time due to excessive data. In image analysis, abrupt change points are one of the most important pieces of information, often located at the edges of important structures. Wavelet decomposed sub-images contain rich texture information. Therefore, we can utilize wavelet decomposition to extract feature information from the transformed two-dimensional image sequence. Discrete Wavelet Transform (DWT) is commonly used to perform wavelet decomposition on two-dimensional load data image sequences. Each p×q image undergoes one DWT transformation, resulting in four (p / 2)×(q / 2) sub-images. The decomposition formula is as follows:

[0090]

[0091] In the formula, j represents the decomposition level; k and m are the row and column subscripts, respectively; l and n are translation factors; h and g represent the low-pass and high-pass filters, respectively; and c represents the image at level j. j After one wavelet decomposition, four sub-images c are obtained. j+1 ,d j+1,h ,d j +1,v ,d j+1,d These features include the low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features of the image. Taking the collected refrigerator data as an example, the feature information obtained after two layers of wavelet decomposition is as follows: Figure 4 As shown, "image" refers to image window data that has been preprocessed but not decomposed into two dimensions; "1-level" refers to the feature information obtained after one level of wavelet decomposition; and "2-level" refers to the feature information obtained after two levels of wavelet decomposition.

[0092] 3) After obtaining horizontal, vertical and diagonal information through wavelet decomposition, a non-intrusive load decomposition model needs to be established for load decomposition training. This non-intrusive load decomposition model is an improved neural network.

[0093] First, to fully utilize features from all directions and improve feature utilization, we adopt a multi-scale, multi-resolution approach to improve the neural network. The multi-scale idea is reflected in the fact that the input consists of four data points of different lengths, which coincides with the fact that each wavelet decomposition layer halves the input length. The four-layer wavelet decomposition can construct input images of different lengths, creating four wavelet input layer channels. The multi-resolution idea is reflected in the fact that training in the network is performed using two-dimensional images, and the resolution of the images changes with the modules of the neural network.

[0094] Secondly, the four-layer wavelet decomposition yields four wavelet input layer channels. These four wavelet input layer channels are paired with four training networks consisting of CNN (Convolutional Neural Network), BiLSTM (Bidirectional Long Short-Term Memory Network), and Attention. That is, a training network is connected after each wavelet input layer channel for training. The purpose of using CNN is to further extract data features, achieving data compression, and ensuring that the total input and total output lengths are consistent. BiLSTM is chosen due to its advantages in time series problems. Attention is used to assign spatial attention weights, giving higher weights when the device is on or off, allowing the network to learn according to the weight allocation, making it easier to recognize the multiple states of the device and improving network accuracy.

[0095] Next, the four trained networks are concatenated by a concatenate connection layer, and CNN is used again to ensure that the lengths of the input and output are consistent, thereby ensuring the single-branch reconstruction of the wavelet;

[0096] Finally, the four trained networks are concatenated through a concatenate layer and then fed into a fully connected layer to achieve a non-linear mapping from deep features to the target electrical power. Combined with the seq2point learning model, the decomposition result is obtained. There are three fully connected layers with dimensions of 258, 128, and 1 respectively. The last fully connected layer and the output constitute the seq2point model.

[0097] In summary, the improved neural network includes the following modules:

[0098] Module 1, the repetition layer, can transform the load data from a one-dimensional W time series into a two-dimensional W*W image sequence, where W is the length of the input;

[0099] Module 2, Wavelet Input Layer 1, represents performing one wavelet decomposition;

[0100] Module 3, wavelet input layer 2, represents performing two wavelet decompositions;

[0101] Module 4, wavelet input layer 3, represents performing three wavelet decompositions;

[0102] Module 5, wavelet input layer 4, represents performing four wavelet decompositions;

[0103] Module 6, CNN, stride=2, kernel size=3, activation function is ReLU;

[0104] Module 7, BiLSTM, units = 64, activation function is Tanh;

[0105] Module 8 is an attention mechanism module;

[0106] Module 9 is the Concatenate connection layer;

[0107] Module 10 is a three-layer fully connected layer with dimensions of 258, 128, and 1 respectively;

[0108] Module 11, training layer 1, is composed of modules 2, 6, 7, 8, and 9 connected in sequence;

[0109] Module 12, training layer 2, is composed of modules 3, 6, 7, 8, and 9 connected in sequence;

[0110] Module 13, training layer 3, is composed of modules 4, 6, 7, 8, and 9 connected in sequence;

[0111] Module 14, training layer 4, is composed of modules 5, 6, 7, 8, and 9 connected in sequence;

[0112] The improved neural network's total loss function is composed of the concatenation of the loss functions of four training layers. Each training layer's loss function includes the loss functions of CNN, BiLSTM, and Attention. Finally, a fully connected layer is used to obtain the final loss function. The formula for each training layer can be expressed as follows:

[0113] loss train =loss(Attention(BiLSTM(CNN(X))))

[0114] Where X represents the input load device power value, loss train The loss function representing the training layer;

[0115] Therefore, the total loss function can be expressed as:

[0116] loss total =loss(Dense(Cat(Cat(Cat(loss train1 loss train2 loss train3 loss train4 )))

[0117] Where, loss total Represents the total loss function, Dense represents a fully connected layer, Cat represents a connected layer, and loss train1 loss train2 loss train3 loss train4 The loss function represents the training layers 1, 2, 3, and 4;

[0118] Therefore, the improved neural network structure is as follows: Figure 5 As shown, the Input layer refers to the input layer, Slidingwindow=48 means the sliding window length is 48, Repeat layer means transforming from a one-dimensional window of 48 to a two-dimensional image of 48*48, Reshape means dimensional transformation, 1-level wavelet transform, 2-level wavelet transform, 3-level wavelet transform, and 4-level wavelet transform refer to wavelet decomposition at levels 1, 2, 3, and 4, respectively, LL, HL, LH, and HH refer to the low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features obtained after wavelet decomposition, respectively, Conv.layer refers to the convolutional neural network, whose kernel size, filters, and stride are the kernel size, number of kernels, and stride, respectively, BiLSTM is a bidirectional long short-term memory network, whose units refer to units, Attention is the attention mechanism, Concat refers to the connection layer, Dense layer refers to the fully connected layer, and Output is the output layer.

[0119] 4) The load feature information obtained through four-layer wavelet decomposition is input into the non-intrusive load decomposition model for decomposition training. The training neural network parameters are set as follows: Batch size is set to 512, training iterations are 250, the Adam optimizer is used, the initial learning rate is 0.0001, the decay rate is 0.1, the decay step size is 1000, and the loss function is the minimum mean square error (MSE). An early stopping mechanism is also added, that is, when the mean square error stops decreasing after 10 iterations, the network will automatically stop training. By continuously optimizing and adjusting the model parameters, the model accuracy can be improved. After training, the model parameters with the minimum loss on the validation set are selected from the saved training parameters as the model parameters for the detection model. Then, the data in the test set is input into the trained model to obtain the user's decomposition results, such as the power sequence of individual loads such as dishwashers, washing machines, refrigerators, microwave ovens, and kettles.

[0120] To more intuitively demonstrate the performance of the model in non-invasive load decomposition, this paper will use regression metrics including Mean Absolute Error (MAE), Normalized Total Error (SAE), and classification metrics including F1, Accuracy, Precision, and Recall for evaluation. These metrics are as follows:

[0121]

[0122] in, and x t Let T represent the predicted and actual power values ​​of a certain electrical appliance at time t, where T is the sample size.

[0123]

[0124] in, and This represents the actual total power and the predicted total output power of an electrical appliance over a time period t. and

[0125]

[0126]

[0127]

[0128] Among them, precision(c j 0 represents accuracy; recall(c j F1-Score(c) represents recall rate; j TP(c) is a combined metric of accuracy and recall. j ) represents category c j The number of identified samples; FN(c j ) represents category c j The number of unidentified samples; FP(c j ) is identified as c j Category but not belonging to c j The sample size; the above formula verifies that the non-intrusive load decomposition model is feasible. To more intuitively display the decomposition results of this model, the results of the load regression index and decomposition index are shown in the table below. Detailed decomposition diagrams of the load localities for dishwashers, refrigerators, washing machines, and microwave ovens are shown below. Figure 6-9 As shown, the total power refers to the total household power value directly measured from the non-intrusive load identification device, the actual device power refers to the power value of a single device directly measured by the identification device, and the text decomposition value refers to the power value obtained by decomposition using the method described in this paper.

[0129]

[0130]

[0131] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A non-intrusive load decomposition method based on wavelet decomposition and improved neural networks, characterized in that, Includes the following steps: 1) Use a non-intrusive load identification device to collect power consumption information on the load side, perform data preprocessing to obtain two-dimensional image sequence load data, and divide the data into training set and test set according to the proportion; 2) The preprocessed load data is decomposed into load features using a four-level wavelet decomposition. Each level of wavelet decomposition can obtain low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features. 3) The load feature information obtained through four-layer wavelet decomposition is input into the constructed non-intrusive load decomposition model for decomposition training. The model parameters are continuously adjusted using data from the training set to improve the non-intrusive load decomposition model. Then, the model is validated using data from the test set, and the non-intrusive load decomposition results are analyzed. The constructed non-intrusive load decomposition model is an improved neural network. The improvements of this neural network include: First, the network is multi-scale and multi-input, and the training process is multi-resolution; Second, the input images of four wavelet input layer channels are obtained from four-layer wavelet decomposition, and a training network composed of CNN, BiLSTM, and Attention is connected to each of the four wavelet input layer channels for training; Third, the results of the four inputs at different scales after training are concatenated through a concatenation layer to achieve wavelet reconstruction.

2. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 1, characterized in that, In step 1), power data is collected at the household power inlet by power sensors deployed in a specific area to provide a data source for non-intrusive load research. The collected data consists of the total power time series and the power time series of individual electrical appliances. In order to obtain more noise-resistant, higher quality data and better training results, the following preprocessing operations are required. The first step is to clear outliers in the data. Outliers may be caused by errors in the measuring device or environmental factors. The formula for clearing outliers is as follows: or ； In the formula, It is the power consumption value of a single electrical device at time t. It is the power acquisition value of a single electrical device at time t±1. and These are the switching power value and maximum power value of a single electrical device, respectively. Secondly, data standardization is performed to make the data easier for neural networks to train, which is beneficial for obtaining better results. The standardization formula is expressed as follows: ； In the formula, This represents the power value at time t; and These represent the average power and standard deviation of the electrical equipment, respectively. The power data is after standardization. Secondly, sliding windowing is applied to the data to improve the training speed of the neural network. Sliding windowing divides the data according to the length of the window, as shown in the following formula: ； In the formula, N is the number of processed load time series. W represents the number of load time series before processing, and W is the window length that needs to be segmented. Next, the data dimension is transformed, converting the one-dimensional time series load data into a two-dimensional image sequence load data. Since the load data is now divided into N data points of window length W, the dimension transformation only requires copying these N load data points W times, so that the load data changes from a one-dimensional W time series to a W*W image sequence. Finally, the transformed data is divided into training and testing sets in a 7:3 ratio to complete the database construction.

3. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 2, characterized in that, In step 2), the Discrete Wavelet Transform (DWT) is used to perform wavelet decomposition on the two-dimensional image sequence load data. Each p×q image undergoes one DWT transformation to obtain four sub-images of size (p / 2)×(q / 2). The decomposition formula is as follows: ； In the formula, j represents the decomposition level; k and m are the row and column subscripts, respectively; l and n are the translation factors; h and g represent the low-pass and high-pass filters, respectively; and j represents the image at level j. After one wavelet decomposition, four sub-images are obtained. , , , These include the low-frequency approximate component features, horizontal component features, vertical component features, and diagonal component features of the image, respectively.

4. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 3, characterized in that, In step 3), firstly, in order to fully utilize the features in each direction and improve the feature utilization rate, a multi-scale and multi-resolution method is adopted to improve the neural network. The idea of ​​multi-scale is reflected in the fact that the input consists of four data of different lengths, which also coincides with the fact that each wavelet decomposition layer halves the input length. The four-layer wavelet decomposition can construct different input lengths and construct input images for four wavelet input layer channels. The idea of ​​multi-resolution is reflected in the fact that the network training is performed on two-dimensional images, and the resolution of the images changes with the modules of the neural network. Secondly, the four-layer wavelet decomposition yields the input images of four wavelet input layer channels. These four wavelet input layer channels are paired with four training networks consisting of CNN, BiLSTM, and Attention. A training network is connected after each wavelet input layer channel for training. The purpose of using CNN is to further extract the features of the data, play the role of data compression, and ensure that the total input and total output lengths are consistent. BiLSTM is used because of its advantages in time series problems; Attention is used to assign spatial attention weights, giving higher weights when the electrical device is turned on or off, so that the network can learn according to the weight allocation, making it easier to identify the multiple process states of the device and improve the network accuracy. Next, the four trained networks are concatenated by a concatenate connection layer, and CNN is used again to ensure that the lengths of the input and output are consistent, thereby ensuring the single-branch reconstruction of the wavelet; Finally, the four trained networks are concatenated through a concatenate layer and then fed into a fully connected layer to achieve a non-linear mapping from deep features to the target electrical power. Combined with the seq2point learning model, the decomposition result is obtained. There are three fully connected layers with dimensions of 258, 128, and 1, respectively. The last fully connected layer and the output constitute the seq2point model.

5. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 4, characterized in that, The improved neural network includes the following modules: Module 1, the repetition layer, can transform the load data from a one-dimensional W time series into a two-dimensional W×W image sequence, where W is the length of the input; Module 2, Wavelet Input Layer 1, represents performing one wavelet decomposition; Module 3, wavelet input layer 2, represents performing two wavelet decompositions; Module 4, wavelet input layer 3, represents performing three wavelet decompositions; Module 5, wavelet input layer 4, represents performing four wavelet decompositions; Module 6, CNN, stride=2, kernel size=3, activation function is ReLU; Module 7, BiLSTM, units=64, activation function is Tanh; Module 8 is an attention mechanism module; Module 9 is the Concatenate connection layer; Module 10 is a three-layer fully connected layer with dimensions of 258, 128, and 1 respectively; Module 11, training layer 1, is composed of modules 2, 6, 7, 8, and 9 connected in sequence; Module 12, training layer 2, is composed of modules 3, 6, 7, 8, and 9 connected in sequence; Module 13, training layer 3, is composed of modules 4, 6, 7, 8, and 9 connected in sequence; Module 14, training layer 4, is composed of modules 5, 6, 7, 8, and 9 connected in sequence; The improved neural network's total loss function is composed of the loss functions of four training layers concatenated. Each training layer's loss function includes the loss functions of CNN, BiLSTM, and Attention. Finally, a fully connected layer is used to obtain the final loss function. The formula for each training layer is as follows: ； In the formula, X represents the input load power value. The loss function representing the training layer; Therefore, the total loss function is expressed as: ； In the formula, Represents the total loss function. Represents a fully connected floor. Represents the connection layer. This represents the loss function for training layers 1, 2, 3, and 4.

6. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 5, characterized in that, In step 3), the load feature information obtained through four-layer wavelet decomposition is input into the non-intrusive load decomposition model for decomposition training. The training neural network parameters are set as follows: Batch size is set to 512, training iterations are 250, the Adam optimizer is used, the initial learning rate is 0.0001, the decay rate is 0.1, the decay step size is 1000, and the loss function is the minimum mean square error (MSE). An early stopping mechanism is also added. When the mean square error stops decreasing after 10 iterations, the network will automatically stop training. By continuously optimizing and adjusting the model parameters, the model accuracy can be improved. After training, the model parameters with the minimum loss on the validation set are selected from the saved training parameters as the model parameters for the detection model. Then, the data in the test set is input into the trained model to obtain the user's decomposition results.

7. The non-intrusive load decomposition method based on wavelet decomposition and improved neural network according to claim 6, characterized in that, In step 3), to more intuitively demonstrate the performance of the model in non-invasive load decomposition, regression metrics, including Mean Absolute Error (MAE), Normalized Total Error (SAE), and classification metrics F1, Accuracy, Precision, and Recall, are used for evaluation. ； In the formula, and Let T be the predicted and actual power values ​​of a certain electrical appliance at time t, where T is the sample size. ； In the formula, and This represents the actual total power and the predicted total output power of an electrical appliance over a time period t. and ; ； ； ； In the formula, For accuracy; Recall rate; It is a comprehensive indicator of accuracy and recall. Indicates category The number of samples identified; Indicate category The number of unidentified samples; To be identified as Category but not belonging to The number of samples.

Images