Non-intrusive load identification method based on EMD and AlexNet

Abstract

The non-intrusive load identification method based on EMD and AlexNet has the characteristics of high recognition accuracy and small computing space. 1. The resident electricity data information is preprocessed to extract the electrical information of the electrical appliances; 2. The electrical information of the electrical appliances is extracted, and the collected non-stationary signal is processed by using the empirical mode decomposition (EMD) method to extract the time series information of the electrical appliances, including the current, voltage, active power and power factor data at the corresponding time point; 3. The neural network theory is used to train the AlexNet neural network to identify the load of the electrical appliances; 4. The neural network algorithm and non-intrusive load identification technology are used to analyze the voltage, current, active power and power factor information, and the electrical appliance operation information on each branch is obtained according to the electrical appliance operation time series information.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power load identification, and particularly relates to a non-intrusive load identification method based on EMD and AlexNet. BACKGROUND

[0002] Demand side power consumption management and energy efficiency analysis optimization is an important way to solve the problem of energy supply efficiency. Efficient demand side management can not only help the power grid side enhance the operation efficiency of the power grid, but also can alleviate energy pressure and improve energy utilization efficiency. With the promotion of power grid demand side management, residential user domain load identification has become an important factor to realize demand side intelligent management. Through load identification, the actual energy consumption level of various loads in the user domain can be understood, and scientific collection and management of energy efficiency data can be realized.

[0003] At present, the load power consumption information of power customers lacks fine data, which cannot support the realization of demand side intelligent energy management. Traditional load identification adopts an invasive method, which has high hardware cost and is complex in installation, maintenance and management, and is not easy to operate, and the user acceptance is low. At the same time, there are a large number of resident users whose existing meters do not have the condition for transformation. Moreover, the existing load decomposition method needs to install specific identification devices, which has the problems of high cost and large transformation cost, and is difficult to popularize in actual use.

[0004] In the aspect of non-intrusive load identification, traditional methods include simple classification methods such as k-means and svm. With the development of artificial intelligence and deep learning, the rise of AlexNet neural network and twin network, their application in non-intrusive load identification is becoming more and more widespread. In terms of load characteristics, there are active power, reactive power and voltage-current waveform trajectory features of the load. Although the power of the load can quickly obtain the classification result when the power of the load is used as the feature, the classification accuracy is not enough. When the voltage-current waveform trajectory feature is used for classification, most of the loads can be identified, but the accuracy of identification for the voltage-current trajectory feature similar to the load feature still needs to be improved. Moreover, the existing load identification process needs to occupy a large amount of calculation space in calculation, and is not easy to popularize.

[0005] The total active power time sequence signal of the user in the preset time is collected by the common intelligent meter as the signal to be decomposed, and the active power time sequence signal of each electric appliance in the preset time is collected as the prior information; the network structure of the total load time sequence signal is established; the prior information of the neural network signal of each electric appliance is constructed; for one of the electric appliances, the neural network signal of the electric appliance is reconstructed based on the global smoothness function of the neural network signal; the active power time sequence signal of the electric appliance is reconstructed by using the active data corresponding to each working state of the electric appliance and the state duration and combining the normalized neural network signal of the electric appliance; after the time sequence signal of the electric appliance is removed from the total load time sequence signal, the above process is cyclically operated to complete the reconstruction of the active power time sequence signal of the remaining other electric appliances. The method realizes non-intrusive load decomposition.

[0006] However, the network structure of the total load time sequence signal is established, the difference value of the sampling points of the switch state of some electric appliances is similar, the neural network signal of the electric appliance is reconstructed, the value may not meet the previous definition and the actual operation of the electric appliance, and the time sequence signal of the electric appliance cannot be accurately recovered. Moreover, the data is not classified, and the sequence matching also has errors.

[0007] In the prior art, the patent document with publication number CN107546855A and the name "a non-intrusive decomposition method for residential electric load" realizes non-intrusive load decomposition. The technical solution adopted includes: collecting the total load active power time sequence signal of the user within the preset time by the ordinary intelligent electric meter as the signal to be decomposed, and collecting the active power time sequence signal of each electric appliance running alone within the preset time as the prior information; a graph structure is established for the total load time sequence signal; the prior information of the graph signal of each electric appliance is constructed; for one of the electric appliances, based on the global smoothness function of the graph signal, the graph signal of the electric appliance is reconstructed; the active power time sequence signal of the electric appliance is reconstructed by using the active data corresponding to each working state of the electric appliance and the state duration, and combining the regularized graph signal of the electric appliance; after removing the time sequence signal of the electric appliance from the total load time sequence signal, the above process is operated in a loop to complete the reconstruction of the active power time sequence signal of the remaining other electric appliances. The technical solution adopted by the patent document with publication number CN2021110763505 and the name "a non-intrusive load identification method and system based on Alexnet neural network and color coding" includes: collecting the running data of the load and constructing the voltage-current trajectory feature map of the load; the running data of the load is preliminarily identified based on the SVM clustering algorithm to obtain the preliminary load identification result; the voltage-current trajectory feature map and the preliminary load identification result are distinguished by using RGB color to construct the voltage-current trajectory feature map with color distinction; based on the trained Alexnet neural network, the voltage-current trajectory feature map with color distinction is identified to obtain the load identification result; but there are the following shortcomings, first, the feature extraction depends on the resolution of the voltage-current trajectory image, and if the resolution is low, image aliasing is easy to occur under high frequency fluctuation. Secondly, the svm algorithm and the color coding algorithm based on alexnet have little correlation, and in the text, it only serves as a prior correction of load identification data. SUMMARY

[0008] In order to solve the above problems, a non-intrusive load identification method based on EMD and AlexNet is proposed, which reconstructs the neural network signal of the electric appliance, and the numerical value conforms to the previous definition and the actual operation of the electric appliance, and accurately restores the time sequence signal of the electric appliance. This method realizes non-intrusive load decomposition.

[0009] In order to achieve the above purpose, the technical scheme adopted by the present application is:

[0010] The non-intrusive load identification method based on EMD and AlexNet has the following specific steps, and has the characteristics that:

[0011] (1) Data acquisition, collect residential electricity data information;

[0012] (2) data preprocessing, using EMD method to obtain voltage, current, active power, power factor four dimension time series information;

[0013] In step (2), the data preprocessing includes EMD empirical mode decomposition, which is used to eliminate the interference in the sampling data.

[0014] The EMD empirical mode step for nonlinear electrical signal network is as follows:

[0015] Step 1: find all extreme points of the signal, connect the local maximum points into the upper envelope line by cubic spline curve, and connect the local minimum points into the lower envelope line. The upper and lower envelope lines contain all data points;

[0016] Step 2: the average value of the upper envelope and the lower envelope line is obtained, if the IMF condition is met, it is considered to be the first IMF component;

[0017] Step 3: if the IMF condition is not met, take it as the original data, repeat steps 1 and 2 to get the average value of the upper and lower envelopes, and calculate whether it meets the necessary conditions of the IMF component. If it does not meet, repeat the above two steps;

[0018] Step 4: separate the IMF from the signal: take it as the original signal and repeat the above three steps, and get the second IMF component until the nth IMF component;

[0019] Step 5: when it becomes a monotonic function, the remaining becomes a residual component. The sum of all IMF components and residual components is the original signal;

[0020] (3) using AlexNet neural network processing, collecting various types of electrical data including voltage, current, active power, and power factor four-dimensional information, and performing AlexNet neural network training and testing;

[0021] In step (3), the AlexNet neural network processing includes the following steps:

[0022] In the (2) given current electrical appliance combination energy measure X set, containing four dimensions, including voltage, current, active power, and power factor. AlexNet network is the origin of deep convolutional neural network. In this study, the representative AlexNet in recent years is improved;

[0023] AlexNet network contains 5 convolution layers Conv, 3 maximum pooling layers Maxpooling and 3 fully connected layers dense, the convolution layers and the maximum pooling layers are arranged alternately, different convolution layers have different convolution kernels, and the ability to extract low-level features is also different;

[0024] The highlight of AlexNet network is to use double GPU to accelerate network training, compared with using single GPU training learning, the learning speed is greatly improved. The activation function used by AlexNet network is ReLU function, not the traditional Sigmiod function, which also speeds up the learning speed and solves the problem of gradient dispersion. LRN local response normalization is to establish a competition mechanism for local neurons after ReLU, so that the response of larger value becomes relatively larger, and the neurons with smaller response are suppressed, and the generalization ability of the network is strengthened.

[0025] The specific calculation formula of AlexNet network is given as follows:

[0026]

[0027] In the formula, a is the output result after the convolution layer, which is represented as a four-dimensional array; N represents the number of channels; n is the adjacent convolution kernel; k is the bias; and a and β are self-defined values, with a value range of (0, 1), which is determined by neuron parameters and is determined by backtracking verification according to the accuracy result;

[0028] The convolution layer is the most important part of the convolutional neural network, and the feature extraction of the network depends on the convolution layer. The convolution layer extracts objects through the convolution template, and with the change of the convolution template, the convolution layer also operates in different feature channels, so as to realize the extraction of different features and the integration of the same features. According to the change of data, the convolution kernel can be adjusted to the appropriate weight value through optimization algorithm, so as to make the extracted features most effective.

[0029] The operation formula of convolutional neural network is:

[0030] N = (W-F+2P) / S+1 (2-2)

[0031] In the formula, W is the width and height of the input neuron, the width represents the data dimension, and the height represents the numerical value; F is the size of convolution / pooling and; S is the step distance of convolution / pooling; and P is the step number of padding;

[0032] Convolution process description:

[0033]

[0034] In the formula, is the weight of the i-th convolution kernel of the j-th layer; is the j-th convolution local area of the l-th layer; W is the width of the convolution kernel;

[0035] The core idea of the convolutional neural network is to aggregate the information of edges to generate new node representations, which can automatically learn both node features and the correlation between nodes.

[0036] The model proposed in the present application is based on the AlexNet neural network, which consists of 5 convolutional layers and 3 fully connected layers. In the convolutional layers, three pooling layers are inserted to reduce the parameters in the model. The entire network contains 630 million links, 60 million parameters and 650,000 neurons. The structure of AlexNet is shown in network 3. The input layer of the network is a 224*224*4 matrix, corresponding to the size of the input data dimension. The first convolutional layer has 96 convolutional kernels, using a larger convolution with a size of 11*11 and a step of 4. The second layer is an LRN layer, followed by a 3*3 max pooling layer with a step of 4. The convolutional layers after that are relatively small, usually 5*5 with a step of 1. Its purpose is to scan all the matrices. The maximum pooling layer is still 3X3, with a step of 2. It can be seen that in the first few convolutional layers of the network, although the amount of calculation is large, the amount of parameters is small, which is about IM. Most of the parameters of the model are in the fully connected layer. This is determined by the nature of the shared weights of the convolutional layer:

[0037] 1) Input layer. The input layer is the initial layer of the network, which inputs the "power network data" converted from the active power sequence intercepted by the sliding window and subjected to data preprocessing operations such as normalization.

[0038] 2) First convolutional layer. The input dimension is the data feature dimension, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0039] 3) Second convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0040] 4) Third convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0041] 5) Fourth convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0042] 6) The fifth layer of convolution layer. The input dimension is 128, the output dimension is 1, and the ReLu function is used to activate the result.

[0043] 7) The network global average pooling layer. The features obtained by globally average pooling the output results of the convolution layer are used as the features of the target appliance at the corresponding midpoint time of the network.

[0044] 8) The output layer. The features of the target appliance at the corresponding midpoint time of the network output by the network global average pooling layer are used as the power decomposition value of the target appliance at the corresponding midpoint time of the network.

[0045] Based on the EMD and AlexNet, the time series generated by the EMD method is intercepted by a sliding window to obtain new "power, voltage, current, power" factor network data, which is normalized to obtain the input layer of the AlexNet network.

[0046] The four-dimensional electrical characteristic time series data is subjected to five layers of convolution to extract information in the input sequence, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used to extract the feature information in the sequence.

[0047] The matrix extracted by the convolution layer is used as the input of the fully connected layer, which maps the learned "distributed feature representation" to the sample label space, plays the role of "classifier", and preserves the model complexity.

[0048] In the training process, the convergence speed of the AlexNet network is significantly accelerated. When the number of iterations reaches 10, the model accuracy and loss will approach a stable trend due to the use of transfer learning method for pre-training the network, and the model training process is completed. Most of the parameters adjusted in the AlexNet network are trained by collecting 1.4 million data of electrical appliance running data in three months. Therefore, when the network is trained, the network only needs to fine-tune a small number of parameters to make it more suitable for load identification.

[0049] In order to improve the training speed, the samples are sent into the network training in batches using bath size, and dropout and L2 regularization are used to further alleviate the overfitting phenomenon. The Adam optimizer is used to correct the network parameters by back propagation until convergence, and the best training parameters are saved for testing during the training process;

[0050] The absolute mean error Mean Absolute Error, MAE, root mean square error Root Mean Squard Error, RMSE and power decomposition accuracy P acc are selected as the model evaluation indexes, and the specific calculation method is as follows:

[0051]

[0052]

[0053]

[0054] In the formula: y represents the predicted value of the model target appliance power at time t; y t y represents the true value of the target appliance power at time t;

[0055] (4) Load identification, using the trained neural network, the resident power information in the bus is decomposed to obtain the power information of each appliance on each branch;

[0056] (5) Algorithm output, the corresponding is the detected total power rising edge and falling edge, which is most likely caused by a pair of load events, that is, respectively indicating the start and end of the running state of a certain device.

[0057] As a further improvement of the application, in step (1), the data acquisition refers to the acquisition of high-frequency load of the load operation, and the frequency is 1hz.

[0058] As a further improvement of the application, in the (4), if the number of positive clusters is not equal to the number of negative clusters, the part with more cluster numbers is merged, and the two clusters with the smallest mean difference of each cluster are merged during the merging, until the number of positive clusters and the number of negative clusters are equal.

[0059] As a further improvement of the application, in the (4), the establishment of the load characteristic database is induced by artificial measurement of the load characteristics of various devices or automatically performed by using the classification algorithm in machine learning. Since the load characteristics of different electrical devices have obvious differences, it is completely feasible to complete load identification based on the load characteristic database.

[0060] In the load identification process, the accuracy rate of the model and the loss function are used as the judgment basis, each electrical signal obtained by EMD empirical mode decomposition is recognized and decomposed, the running data of each appliance on each branch is obtained, and the time sequence of the obtained load running data is used as the final output result.

[0061] Compared with the prior art, the application has the beneficial effects that:

[0062] 1. This invention employs non-intrusive load identification technology, identifying only the electrical signal at the power inlet. It uses feature extraction and machine learning algorithms to analyze the combined voltage and current, identifying the usage of electrical appliances on each branch circuit. Through analysis, it obtains the types of loads in the user domain and their corresponding operating conditions. This method reduces the economic cost of identification equipment, is easy to install and maintain, and provides power companies and users with a real-time, economical, and effective identification method. It also improves the accuracy of the system when identifying residential electricity users.

[0063] 2. In this invention, the multi-space coupling feature analysis technology for electricity consumption behavior analyzes the user's electricity consumption behavior from multiple spaces, such as electrical space, user social attribute space, environmental space, and user behavior space, and couples them in multiple spaces. Because the information from multiple spaces is combined, the analysis results are more accurate and the efficiency of data analysis is improved. Attached Figure Description

[0064] Figure 1 This is the system frame network of the present invention;

[0065] Figure 2 This is the overall algorithm flow and thought process flow of the AlexNet neural network in this invention;

[0066] Figure 3 This is the step network of EMD empirical mode decomposition in this invention.

[0067] Figure 4 The AlexNet neural network structure is being developed.

[0068] Figure 5 This describes the data processing procedure in the AlexNet neural network of this invention.

[0069] Figure 6 This is the confusion matrix of the AlexNet neural network of this invention.

[0070] Figure 7 To illustrate the effect of the non-invasive load identification method of the present invention, Alexnet represents the identified and decomposed electrical appliance operating data, agg represents bus data, and gt represents the actual value measured by the instrument. The figure uses power information as an example. Detailed Implementation

[0071] Non-invasive load identification methods based on EMD and AlexNet, such as Figure 1 ,

[0072] (1) Data collection: collecting residential electricity consumption data;

[0073] (2) Data preprocessing: Time series information of four dimensions, including voltage, current, active power and power factor, is obtained by using the EMD method.

[0074] (3) Using AlexNet neural network processing, for residential electricity data, collecting various types of electricity data including voltage, current, active power, power factor four-dimensional information, AlexNet neural network training and testing;

[0075] (4) Load identification, using the trained neural network, the residential electricity information in the bus is decomposed to obtain the electricity information of each electric appliance on each branch;

[0076] (5) Algorithm output, the results of the electric appliance electricity data identification and decomposition on each branch are output;

[0077] including data collection (1) data preprocessing (2) AlexNet neural network processing (3) load decomposition (4) algorithm output (5). (1) is used to collect electric appliance data and prepare for data processing; (2) the significance of preprocessing is to ensure the reliability of the identified data, it is necessary to preprocess the data to eliminate the interference and noise in the power data as much as possible, and lay a good foundation for the development of load decomposition. The load decomposition algorithm is a zero-training unsupervised method independent of equipment modeling, different power states of multi-state equipment and different operating modes of equipment.

[0078] Step (1) is internally provided with high-frequency load data collection, the high-frequency load data collection uses a newly added measuring device to record the characteristic data of the electric appliance equipment in three states of steady state, transient state and operation under the states of single electric appliance operation and multiple electric appliances superimposed operation, and to construct a high-frequency load data set in the running process of the electric appliance equipment. The characteristic data of the electric appliance equipment in different states such as starting, running and frequency conversion is recorded, and a high-frequency data set in the running process of the electric appliance equipment is constructed.

[0079] The overall algorithm flow of AlexNet neural network is as follows Figure 2 , by EMD empirical mode decomposition, extracting load running electrical characteristics, combining with Loss loss function calculation, matching load identification model parameters. Through the AlexNet neural network model, the historical running data is trained and tested, and finally each load information is identified, and each electric appliance running condition is decomposed

[0080] Step (2) is provided with EMD empirical mode decomposition method: by finding the maximum and minimum envelope lines, each IMF signal is decomposed from the original signal, and finally the noise is eliminated by signal superposition, effectively avoiding the influence of modal aliasing and noise; as Figure 3 For model training, first use the historical electric appliance running data to pre-train the original AlexNet. Then the training set starts secondary training. Finally, a deep neural network suitable for load identification task is obtained.

[0081] The non-intrusive load identification method based on EMD and AlexNet has the characteristics that specific calculation formulas are given in the AlexNet network:

[0082]

[0083] In the formula: a is the output result after the convolution layer, which is expressed as a four-dimensional array; N represents the number of channels; n is the adjacent convolution kernel; k is the bias; and alpha and beta are self-defined values, the value range is (0, 1), which is determined by the neuron parameters, and the specific value is determined according to the accuracy rate result backtracking verification.

[0084] The convolution layer is the most important part of the convolutional neural network, and the capture of image features depends on the convolution layer. The convolution layer extracts objects through the convolution template, and with the change of the convolution template, the convolution layer also operates in different feature channels, so as to realize the extraction of different features and the integration of the same features. According to the change of data, the convolution kernel can be adjusted to the appropriate weight value through the optimization algorithm, so as to make the extracted features most effective.

[0085] The operation formula of the convolutional neural network is:

[0086] N = (W-F+2P) / S+1 (3-2)

[0087] In the formula: W is the width and height of the input picture; F is the size of convolution / pooling and; S is the convolution / pooling step distance; and P is the padding step number.

[0088] Convolution process description:

[0089]

[0090] In the formula: is the weight of the i-th convolution kernel of the j-th layer; is the j-th convolution local area of the l-th layer; and W is the width of the convolution kernel.

[0091] The root graph data has two important features, namely node features and structure features. The node features describe the inherent properties of the nodes in the network, and the structure features describe the association properties between the nodes. This structure generated by the association not only helps to describe the nodes in the network data, but also plays a key role in describing the whole network. The core idea of the convolutional neural network is to aggregate the node information using the edge information to generate new node representation, which can automatically learn both the node features and the association information between the nodes.

[0092] The model proposed in this study is based on the AlexNet neural network, which consists of 5 convolutional layers and 3 fully connected layers. In the convolutional layers, three pooling layers are inserted to reduce the parameters in the model. The entire network contains 630 million links, 60 million parameters, and 650,000 neurons. The structure of AlexNet is shown in Figure 3 The input layer of the network is a 224X224x3 matrix corresponding to the size of the input network image. The first convolutional layer has 96 convolutional kernels, uses a larger convolution with a size of 11x11 and a step of 4. The second layer is an LRN layer, followed by a 3x3 max pooling layer with a step of 4. The convolution after this is smaller, usually 5x5 or 3x3, with a step of 1. Its purpose is to scan all the pixels. The maximum pooling layer is still 3X3, with a step of 2. It can be seen that in the first few convolutional layers of the network, although the amount of calculation is large, the amount of parameters is small. Most of the parameters of the model are in the fully connected layers. This is determined by the nature of the shared weights of the convolutional layers:

[0093] 1) Input layer. The input layer is the initial layer of the network, which inputs the "power network data" converted from the active power sequence intercepted by the sliding window, the "current network data" converted from the current sequence intercepted by the sliding window, the "voltage network data" converted from the voltage sequence intercepted by the sliding window, and the "power factor network data" converted from the power factor sequence intercepted by the sliding window, and has undergone data preprocessing operations such as normalization.

[0094] 2) First convolutional layer. The input dimension is the data feature dimension, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0095] 3) Second convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0096] 4) Third convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0097] 5) Fourth convolutional layer. The input dimension is 128, the output dimension is 128, the ReLu function is used to activate the result, and then the dropout operation with a probability of 0.5 is used on the activated result to obtain the dropout result.

[0098] 6) The fifth convolutional layer. The input dimension is 128, the output dimension is 1, and the result is activated using the ReLu function.

[0099] 7) The network global average pooling layer. The features obtained by globally average pooling the output of the convolutional layer are used as the features of the target appliance at the corresponding midpoint of the network.

[0100] 8) The output layer. The features of the target appliance at the corresponding midpoint of the network output by the network global average pooling layer are used as the power decomposition value of the target appliance at the corresponding midpoint of the network.

[0101] The convergence speed of the AlexNet network is significantly accelerated during the training process. When the number of iterations reaches 10, the accuracy and loss of the model will approach a stable trend due to the use of the transfer learning method for pre-training the network, and the training process of the model is completed. Most of the parameters in the AlexNet network that have been adjusted are based on the collected three-month appliance operation data for training. Therefore, when the network is trained by the training set in this section, the network only needs to fine-tune a small number of parameters to make it more suitable for load identification.

[0102] In order to improve the training speed, the samples are sent into the network in batches using the bath size, and measures such as dropout early stopping mechanism and adding L2 regularization are used to further alleviate the overfitting phenomenon. The Adam optimizer is used to correct the network parameters through backpropagation until convergence. The best training parameters are saved during the training process for testing.

[0103] The absolute mean error (Mean Absolute Error, MAE), root mean square error (Root Mean Squard Error, RMSE), and power decomposition accuracy P acc are selected as model evaluation indicators. The specific calculation methods are as follows:

[0104]

[0105]

[0106]

[0107] In the formula: y t represents the true value of the target appliance power at time t.

[0108] In (4), for the event detection-based load identification and decomposition model, the decomposition effect is not very good for the electric appliances with low power value and large power fluctuation. A non-intrusive load identification and decomposition method based on AlexNet neural network processing is researched. The running information of each electric appliance under the total electric signal is solved. The identification result is obtained through the convolutional neural network. The over-smoothing problem is reduced in the solving process, and the early stopping mechanism is introduced to prevent overfitting.

[0109] An intelligent load identification system based on power data mining is established. The load identification is visually displayed. The sequence of each electric appliance is obtained by decomposing the total electric appliance sequence input by the system. Combined with the actual application environment, the automatic decomposition of electric appliance load is realized.

[0110] In (5), the absolute mean error (Mean Absolute Error, MAE), root mean square error (Root Mean Squard Error, RMSE) and power decomposition accuracy P acc are selected as model evaluation indexes. Through the index change under different training rounds, the accuracy result of the final model output is obtained. The corresponding electric information time sequence is output under the highest accuracy result

[0111] The non-intrusive load identification technology is adopted to identify only the electric signal at the power inlet. The feature extraction and machine learning algorithm are adopted to analyze the combined voltage and current and identify the use of electric appliances on each branch circuit. The types and corresponding operation of each load in the user domain are obtained through analysis. This method reduces the economic cost of the identification equipment, is easy to install and maintain, and provides a real-time, economical and effective identification means for power companies and users. The accuracy of the system in identifying residential electricity is improved. Since the information in the multi-dimensional space is combined, the analysis result is more accurate, and the data analysis efficiency is improved.

[0112] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a…" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0113] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0114] Confusion matrix is also called possibility table or error matrix. It is a specific matrix used to present the effect of algorithm, such as Figure 6 The matrix effect diagram of hair dryer, microwave oven, computer, air conditioner and table lamp.

[0115] For example Figure 7 The load decomposition diagram of refrigerator and microwave oven: curve AGG is the total power curve of room electricity; curve GT is the real power curve of equipment; AlexNet curve is the power load curve after load decomposition of neural network model.

[0116] By decomposing each running load in the total load power curve, each electrical appliance power load curve is obtained, so as to achieve the purpose of load monitoring.

[0117] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application in any other form, and any modification or equivalent change made according to the technical essence of the present application still falls within the scope of the present application.

Claims

1. A non-invasive load identification method based on EMD and AlexNet, the specific steps of which are as follows, characterized by: (1) Data collection: collecting residential electricity consumption data; (2) Data preprocessing: Time series information of four dimensions, namely voltage, current, active power and power factor, is obtained by using the EMD method. In step (2), data preprocessing includes EMD empirical mode decomposition, which is used to eliminate interference in the sampled data by processing the acquired nonlinear signals. The empirical mode step network steps for EMD of nonlinear electrical signals are as follows: Step 1: Find all extreme points of the signal, connect the local maxima to form the upper envelope and the local minima to form the lower envelope using cubic spline curves. The upper and lower envelopes contain all data points. Step 2: Based on the average of the upper and lower envelopes, if the conditions of IMF are met, then it can be considered as the first IMF component. Step 3: If it does not meet the IMF conditions, it will be used as the raw data. Repeat steps 1 and 2 to obtain the mean of the upper and lower envelopes. Calculate whether it meets the necessary conditions for IMF components. If it does not meet the conditions, repeat the above two steps. Step 4: Separate the IMF from the signal: Repeat the above three steps with the original signal as the original signal, looping times, to obtain the second IMF component until the nth IMF component; Step 5: After becoming a monotonic function, the remaining components become the residual components. The sum of all IMF components and the residual components is the original signal. (3) Using AlexNet neural network processing, for residential electricity consumption data, collect various types of electricity consumption data including four dimensions of information such as voltage, current, active power and power factor, and conduct AlexNet neural network training and testing; In step (3), the AlexNet neural network processing includes the following steps: The AlexNet network contains 5 convolutional layers (Conv), 3 max pooling layers (Maxpooling), and 3 dense fully connected layers (Dense). The convolutional layers and max pooling layers are arranged alternately, and different convolutional layers have different convolutional kernels, which have different abilities to extract low-level features. The specific calculation formula is given in the AlexNet neural network: （1-1） In the formula: a is the output result after the convolutional layer, which is represented as a four-dimensional array; N represents the number of channels; n is the adjacent convolutional kernels; k is the bias; α and β are user-defined values, ranging from (0, 1), determined by the neuron parameters, and specifically determined by backtesting based on the accuracy results. The operational formula for a convolutional neural network is: N = (W - F + 2P) / S + 1 (1-2) In the formula: W represents the width and height of the input neuron, where width represents the data dimension and height represents the numerical value; F represents the size of the convolution / pooling sum; S represents the stride of the convolution / pooling; and P represents the number of padding steps. Description of the convolution process: （1-3） In the formula: Let i be the weight of the i-th convolutional kernel in the j-th layer; Let be the j-th convolutional local region of layer l; W is the width of the convolutional kernel; The sample size is used to feed the samples into the network for training in batches. Dropout and L2 regularization are used to further alleviate overfitting. The Adam optimizer is used to backpropagate and correct the network parameters until convergence. The best training parameters are saved for testing during the training process. Select the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Power Decomposition Accuracy. As a model evaluation metric, the specific calculation method is as follows: (1-4) (1-5) (1-6) In the formula: This represents the predicted power of the target electrical appliance in the model at time t; This represents the actual power of the target electrical appliance at time t; (4) Load identification: using a trained neural network, the residential electricity consumption information in the bus is decomposed to obtain the electricity consumption information of each appliance on each branch. (5) The algorithm output corresponds to the rising and falling edges of the detected total power, which are caused by a pair of load events, namely the start and end of the operating state of a certain device.

2. The non-invasive load identification method based on EMD and AlexNet according to claim 1, characterized in that, In step (1), data acquisition refers to the acquisition of high-frequency loads operating at a frequency of 1 Hz.

3. The non-invasive load identification method based on EMD and AlexNet according to claim 1, characterized in that, In step (4), if the number of positive clusters is not equal to the number of negative clusters, the clusters with more clusters are merged. When merging, the two clusters with the smallest difference in the mean of each cluster are merged until the number of positive clusters and the number of negative clusters are equal.

4. The non-invasive load identification method based on EMD and AlexNet according to claim 1, characterized in that, In (4), a load characteristic database is established to summarize the load characteristics of various equipment through manual measurement or to automatically perform the measurement using classification algorithms in machine learning; In the load identification process, the accuracy and loss function of the model are used as the evaluation criteria. The electrical signals obtained by EMD empirical mode decomposition are identified and decomposed to obtain the operating data of each electrical appliance on each branch. The time series of the obtained load operating data is used as the final output result.

Images