CN115758246A - Non-invasive load identification method based on EMD and AlexNet - Google Patents
Non-invasive load identification method based on EMD and AlexNet Download PDFInfo
- Publication number
- CN115758246A CN115758246A CN202211451964.9A CN202211451964A CN115758246A CN 115758246 A CN115758246 A CN 115758246A CN 202211451964 A CN202211451964 A CN 202211451964A CN 115758246 A CN115758246 A CN 115758246A
- Authority
- CN
- China
- Prior art keywords
- data
- alexnet
- load
- information
- emd
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The non-invasive load identification method based on the EMD and the AlexNet has the characteristics of high identification accuracy, small occupied calculation space and the like. 1. Preprocessing resident electricity consumption data information and extracting electrical appliance operation electrical information; 2. extracting the characteristics of electrical information of the running of the electrical appliance, processing the acquired non-stationary signals by using an Empirical Mode Decomposition (EMD) method, and extracting time series information of the running of the electrical appliance, wherein the time series information comprises four items of data of current, voltage, active power and power factor at corresponding time points to form the time series information of the running of the electrical appliance; 3. adopting a neural network theory and using a trained AlexNet neural network to carry out load identification on the time sequence information of the electric appliance; 4. and analyzing information such as the converged voltage, current, active power, power factor and the like by adopting a neural network algorithm and a non-invasive load identification technology, and acquiring the operation information of the electric appliances on each branch circuit according to the operation time sequence information of the electric appliances.
Description
Technical Field
The invention 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
The demand side power utilization management and energy efficiency analysis optimization are important ways for solving the problem of energy supply efficiency, and the efficient demand side management can not only help the power grid side to enhance the operation efficiency of the power grid, but also relieve the energy pressure and improve the energy utilization efficiency. With the advance of management work of a power grid demand side, load identification of a residential user domain becomes an important factor for realizing intelligent management of the demand side. The actual energy consumption levels of various loads of the user domain can be known through load identification, and scientific collection and management of energy efficiency data are achieved.
At present, load electricity utilization information of an electric power customer lacks fine data and cannot support and achieve intelligent energy utilization management of a demand side. The traditional load identification adopts an intrusive method, the hardware cost is high, the installation, maintenance, management and other aspects are complex and difficult to operate, and the user acceptance degree is low. Meanwhile, a large amount of resident user stored electric meters do not have transformation conditions. In addition, the existing load decomposition method needs to install a specific identification device, so that the problems of high cost, high modification cost and the like exist, and the load decomposition method is difficult to popularize in practical application.
In terms of non-intrusive load identification, the conventional methods are: by adopting simple classification methods such as k-means, svm and the like, along with the development of artificial intelligence and deep learning, the AlexNet neural network and the twin network and the like grow up, the application of the AlexNet neural network and the twin network is more and more extensive in the aspect of non-invasive load identification, and the load characteristics comprise active power, reactive power and voltage-current waveform track characteristics of the load. When the power of the load is used as the characteristic, although the classification result can be obtained quickly, the classification accuracy is not enough, when the voltage-current waveform track characteristic is used for classification, most of the loads can be identified, but the identification accuracy of the voltage-current track characteristic with similar load characteristics is still to be improved, and the existing load identification process needs to occupy a large amount of calculation space during calculation and is not easy to popularize and use.
Acquiring a total load active power time sequence signal within a preset time of a user as a signal to be decomposed by a common intelligent ammeter, and acquiring an active power time sequence signal during independent operation of each electrical appliance within the preset time as prior information; establishing a network structure for the total load time sequence signal; constructing prior information of a neural network signal of each electrical appliance; for one of the electrical appliances, reconstructing the neural network signal of the electrical appliance based on the global smoothness function of the neural network signal; active power time sequence signals of the electrical appliance are reconstructed by utilizing active data and state duration time corresponding to each working state of the electrical appliance and combining the regulated neural network signals of the electrical appliance; and after the electrical appliance time sequence signal is removed from the total load time sequence signal, performing cycle operation on the process to complete reconstruction of the active power time sequence signals of the rest other electrical appliances. The method achieves non-intrusive load splitting.
However, the network structure is established for the total load time sequence signal, and the difference values of the sampling points of the on-off states of some electrical appliances are similar, so that after the neural network signal of the electrical appliance is reconstructed, the numerical value may not conform to the previous definition and the actual operation of the electrical appliance, and the time sequence signal of the electrical appliance cannot be accurately recovered. And the data is not classified, and the sequence matching has errors.
In the prior art, publication no: CN107546855A, name: the patent document "a non-invasive decomposition method of residential electrical load" realizes non-invasive load decomposition. The technical scheme adopted comprises the following steps: acquiring a total load active power time sequence signal within a preset time of a user as a signal to be decomposed through a common intelligent electric meter, and acquiring an active power time sequence signal when each electric appliance operates independently within the preset time as prior information; establishing a graph structure for the total load time sequence signal; constructing prior information of a graph signal of each electrical appliance; for one of the electrical consumers, reconstructing a graph signal of the electrical consumer based on the global smoothness function of the graph signal; active power time sequence signals of the electrical appliance are reconstructed by utilizing active data and state duration time corresponding to each working state of the electrical appliance and combining the graph signals normalized by the electrical appliance; and after the electrical appliance time sequence signal is removed from the total load time sequence signal, performing cycle operation on the process to complete reconstruction of active power time sequence signals of the rest other electrical appliances. Publication No.: CN2021110763505, name: the technical scheme adopted by the 'non-invasive load identification method and system based on Alexnet neural network and color coding' comprises the steps of collecting operation data of a load and constructing a voltage-current track characteristic diagram of the load; performing preliminary identification on the load operation data based on an SVM clustering algorithm to obtain a preliminary load identification result; distinguishing the voltage-current track characteristic diagram and the initial load identification result by adopting RGB colors, and constructing a voltage-current track characteristic diagram with color distinction; identifying a voltage-current track characteristic diagram with color distinction based on a trained Alexnet neural network to obtain a load identification result; however, there is a disadvantage that the feature extraction depends on the resolution of the voltage-current trajectory image in the first place, and image aliasing easily occurs at high-frequency fluctuations if the resolution is low. Secondly, the svm algorithm is not greatly related to the color coding algorithm based on alexnet, and only serves as the prior correction function of the load identification data in the text.
Disclosure of Invention
In order to solve the problems, a non-invasive load identification method based on EMD and AlexNet is provided, after a neural network signal of an electrical appliance is reconstructed, the numerical value accords with the previous definition and the actual operation of the electrical appliance, and the time sequence signal of the electrical appliance is accurately recovered. The method realizes non-invasive load decomposition.
In order to achieve the purpose, the invention adopts the technical scheme that:
the non-invasive load identification method based on the EMD and the AlexNet comprises the following specific steps and is characterized in that:
(1) Data acquisition, namely acquiring resident electricity consumption data information;
(2) Data preprocessing, namely acquiring time sequence information of four dimensions of voltage, current, active power and power factor by using an EMD method;
in the step (2), data preprocessing comprises EMD empirical mode decomposition, and is performed on the acquired nonlinear signals to eliminate interference in the sampling data;
the empirical mode step network for the nonlinear electrical signal EMD comprises the following steps:
step 1: and searching all extreme points of the signal, connecting the local maximum points into an upper envelope line through a cubic spline curve, and connecting the local minimum points into a lower envelope line. The upper and lower envelopes contain all data points;
step 2: obtaining a first IMF component if the IMF condition is met according to the average value of the upper envelope and the lower envelope;
and step 3: if the IMF condition is not met, the obtained data is used as original data, the step 1 and the step 2 are repeated to obtain the mean value of the upper envelope and the lower envelope, whether the required condition of the IMF component is met or not is calculated, and if the required condition is not met, the steps are repeated;
and 4, step 4: separating the IMF from the signal: repeating the three steps as an original signal for a plurality of times to obtain a second IMF component till an nth IMF component;
and 5: when becoming a monotonic function, the remainder becomes a residual component. The sum of all IMF components and residual components is an original signal;
(3) Utilizing AlexNet neural network processing to acquire various electricity data including four-dimensional information of voltage, current, active power and power factor aiming at resident electricity data, and carrying out AlexNet neural network training and testing;
in the step (3), the AlexNet neural network processing comprises the following steps:
in the step (2), the set of the current electric appliance combined energy measures X comprises four dimensions including voltage, current, active power and power factor. The AlexNet network is the origin of the deep convolutional neural network. The research improves AlexNet which is relatively representative in recent years;
the AlexNet network comprises 5 convolutional layers Conv,3 maximum pooling layers Maxpooling and 3 fully-connected layers dense, wherein the convolutional layers and the maximum pooling layers are alternately arranged, different convolutional layers have different convolutional cores, and the low-level feature extraction capability is different;
the AlexNet network is characterized in that double GPUs are utilized for network accelerated training, and compared with single GPU training learning, the learning speed is greatly improved. The activation function used by the AlexNet network is a ReLU function, rather than a traditional signiod function, which also speeds up learning and solves the problem of gradient dispersion well. The LRN local response normalization is to establish a competition mechanism for local neurons after ReLU, so that the value with larger response becomes relatively larger, the neurons with smaller response are inhibited, and the generalization capability of the network is enhanced.
A specific calculation formula is given in the AlexNet network:
in the formula: a is the output result after convolution layer, which is expressed as a four-dimensional array; n represents the number of channels; n is an adjacent convolution kernel; k is a deviation; alpha and beta are self-defined values, the value range (0, 1) is determined by neuron parameters, and the values are determined by backtracking verification according to accuracy results;
convolutional layers are the most important parts of convolutional neural networks, and capture of network image features depends on convolutional layers. The convolutional layer is subjected to object extraction through the process convolution template, and the convolutional layer can also be operated in different characteristic channels along with the transformation of the convolution template, so that the extraction of different characteristics and the integration of the same characteristics are realized. According to the change of data, the convolution kernel can be adjusted to a proper weight value through an optimization algorithm, and therefore the extracted features are most effective.
The operation formula of the convolutional neural network is as follows:
N＝(W-F+2P)/S+1 (2-2)
in the formula: w is the width and height of the input neuron, wherein the width represents the data dimension, and the height represents the numerical value; f is the size of the convolution/pooling sum; s is the step pitch of convolution/pooling; p is the step number of padding;
description of the convolution procedure:
in the formula:
the weight of the i convolution kernel for the j layer;
the jth convolution local region for l layers; w is the width of the convolution kernel;
the core idea of the convolutional neural network is to aggregate node information by using edge information to generate a new node representation, which can automatically learn not only node characteristics but also association information between nodes.
The model proposed in this application is based on an AlexNet neural network, which consists of 5 convolutional layers and 3 fully connected layers. In the convolutional layer, three pooling layers are inserted to reduce the parameters in the model. The entire network contains 6.3 million links, 6000 million parameters and 65 million neurons. The structure of AlexNet is shown in network 3. The input layer of the network is a matrix of 224x 4, corresponding to the size of the input data dimension. The first convolutional layer has 96 convolutional kernels in total, and uses a larger convolution with a size of 11 × 11 and a step size of 4. The second layer is the LRN layer, followed by a 3 × 3 max pooling layer, with a step size of 4. The convolutional layer after this is relatively small, typically 5 × 5, with a step size of 1. Its purpose is to scan all the matrices. While the maximum pooling layer is still 3X3, step size is 2. It can be seen that in several convolutional layers in front of the network, although the amount of computation is large, the amount of parameters is small, which is about IM. Most parameters of the model are in the fully connected layer. This is determined by the nature of the convolutional layer sharing weights:
1) And (6) an input layer. The input layer is an initial layer of the network, and the input layer is 'power network data' converted from the active power sequence intercepted by the sliding window and subjected to data preprocessing operations such as normalization and the like.
2) A first layer of convolutional layers. And the input dimension is a data characteristic dimension, the output dimension is 128, the result is activated by using a ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
3) A second layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result, so that the dropout result is obtained.
4) A third layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
5) A fourth layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
6) A fifth layer of convolution layers. The input dimension is 128 and the output dimension is 1, and the result is activated using the ReLu function.
7) And network global average pooling layer. And taking the features obtained by performing global average pooling on the convolutional layer output results as the features of the network corresponding to the midpoint time.
8) And (6) an output layer. And outputting the characteristics of the corresponding midpoint moment of the network, which are output by the network global average pooling layer, as the power decomposition value of the target electrical appliance at the corresponding midpoint moment of the network.
According to the non-invasive load identification method based on the EMD and the AlexNet, a time sequence generated by the EMD method is intercepted through a sliding window to obtain new power, voltage, current and power factor network data, and the input layer of the AlexNet network is obtained through normalization processing.
And performing five-layer convolution on the four-dimensional electrical characteristic time sequence data, extracting information in an input sequence, activating the result by using a ReLu function, and extracting characteristic information in the sequence by using a dropout operation with the probability of 0.5 on the activation result.
The matrix extracted from the convolutional layer is used as the input of the fully connected layer, the learned distributed feature expression is mapped to the sample mark space, the function of a classifier is achieved, and the model complexity is reserved.
The convergence speed of the AlexNet network is obviously accelerated in the training process. When the iteration times reach 10 times, the network is pretrained by using a transfer learning method, the precision and the loss of the model approach to a steady trend, and the training process of the model is finished. Most parameters of the AlexNet network, which are adjusted by the network, are trained by collecting operation data of the electric appliance within three months, and the total number of the parameters is 140 ten thousand. Thus, when the network is trained, the network only needs to fine-tune a small number of parameters, making it more suitable for load recognition.
In order to improve the training speed, the bath size is used for sending samples into a network for training in batches, measures such as dropout and addition of L2 regularization are used for further relieving the overfitting phenomenon, an Adam optimizer is used for reversely propagating and correcting network parameters until convergence is achieved, and the best training parameters are stored in the training process for testing;
selecting Absolute average Error Mean Absolute Error (Mean Absolute Error, MAE), root Mean square Error (Root Mean square Error, RMSE) and power decomposition accuracy rate P acc As the model evaluation index, the specific calculation method is as follows:
in the formula:
representing a predicted value of the model target electrical appliance power at the time t; y is t Representing the true value of the target appliance power at time t;
(4) Load identification, namely decomposing the electricity utilization information of residents in the bus by using the trained neural network to obtain the electricity utilization information of each electric appliance on each branch;
(5) The algorithm outputs, corresponding to the detected rising and falling edges of the total power, are most likely caused by a pair of load events, i.e. indicating the beginning and end of a certain plant operating state, respectively.
As a further improvement of the invention, in the step (1), the data acquisition refers to the acquisition of the high-frequency load of the load operation, and the frequency is 1hz.
As a further improvement of the present invention, in the step (4), if the number of the positive clusters is not equal to the number of the negative clusters, the obtained clusters are merged, and during merging, the two clusters with the smallest difference value between the mean values of the clusters are merged until the number of the positive clusters is equal to the number of the negative clusters.
As a further improvement of the present invention, in the step (4), the load characteristics of various devices are summarized by manual measurement through establishing the load characteristic database or automatically performed by using a classification algorithm in machine learning, and since the load characteristics of different electrical devices have obvious differences, it is technically completely feasible to complete load identification based on the load characteristic database.
And in the load identification process, the accuracy and the loss function of the model are used as judgment basis, the EMD empirical mode decomposition is carried out to obtain each electrical signal for identification and decomposition, the operation data of each electrical appliance on each branch is obtained, and the time sequence of the obtained load operation data is used as a final output result.
Compared with the prior art, the invention has the beneficial effects that:
1. in the invention, a non-invasive load identification technology is adopted, only the electric signal at the electric power inlet is identified, the converged voltage and current are analyzed by adopting a characteristic extraction and machine learning algorithm, the service condition of the electric appliances on each branch circuit is identified, and the type and the corresponding operation condition of each load in a user domain are obtained by analysis. The method reduces the economic cost of the identification equipment, is easy to install and maintain, provides a real-time, economic and effective identification means for power companies and users, and also improves the accuracy of the system in identifying the electricity utilization rows of residents.
2. According to the invention, the electricity consumption behavior multi-space coupling characteristic analysis technology analyzes the electricity consumption behavior of the user from a plurality of spaces such as an electric space, a user social attribute space, an environment space and a user behavior space, and performs coupling in the plurality of spaces.
Drawings
FIG. 1 is a system block network of the present invention;
FIG. 2 is an overall algorithm flow and idea flow network of the AlexNet neural network of the present invention;
FIG. 3 is a network of steps for EMD decomposition in accordance with the present invention
FIG. 4 is the AlexNet neural network structure of the invention;
FIG. 5 is a data processing process in the AlexNet neural network of the present invention;
FIG. 6 is a confusion matrix of the AlexNet neural network of the present invention.
Fig. 7 shows the effect of the non-intrusive load identification method of the present invention, where Alexnet is the operation data of the electrical appliance for identification and decomposition, agg is the bus data, gt is the actual value measured by the instrument, and the power information is taken as an example in the figure.
Detailed Description
The non-intrusive load identification method based on EMD and AlexNet, as shown in figure 1,
(1) Data acquisition, namely acquiring resident electricity consumption data information;
(2) Data preprocessing, namely acquiring time sequence information of four dimensions of voltage, current, active power and power factor by using an EMD method;
(3) Processing by utilizing an AlexNet neural network, collecting various electricity utilization data comprising four-dimensional information of voltage, current, active power and power factor aiming at resident electricity utilization data, and training and testing the AlexNet neural network;
(4) Load identification, namely decomposing the electricity utilization information of residents in the bus by using the trained neural network to obtain the electricity utilization information of each electric appliance on each branch;
(5) Algorithm output, namely outputting the result after identifying and decomposing the electricity utilization data of the electric appliances on each branch circuit;
the method comprises the steps of data acquisition (1), data preprocessing (2), alexNet neural network processing (3), load decomposition (4) and algorithm output (5). (1) The device is used for acquiring the data of the electric appliance and preparing for data processing; (2) The significance of the preprocessing is that in order to ensure the reliability of the identification data, data preprocessing is necessary to eliminate the interference and noise in the power data as much as possible, and a good foundation is laid for the development of load decomposition. The load decomposition algorithm is a zero-training unsupervised method that does not rely on device modeling, different power states of a multi-state device, and different operating modes of the device.
The method comprises the following steps that (1) high-frequency load data collection is arranged inside the electric appliance, a newly-added measuring device is used for collecting the high-frequency load data, the states of single electric appliance operation, superposition operation of a plurality of electric appliances and the like are recorded, the characteristic data of the electric appliance under the states of steady state, transient state and operation 3 types are recorded, the high-frequency load data set in the operation process of the electric appliance is constructed, the characteristic data of the electric appliance under the different states of starting, operation, frequency conversion and the like are recorded, and the high-frequency data set in the operation process of the electric appliance is constructed.
The overall algorithm flow of the AlexNet neural network is shown in figure 2, load operation electrical characteristics are extracted through EMD empirical mode decomposition, and load Loss function calculation is combined to match load identification model parameters. Training and testing historical operating data through an AlexNet neural network model, finally completing identification of each load information, and decomposing the operating condition of each electric appliance
The EMD empirical mode decomposition method is arranged in the step (2): by searching a maximum value envelope line and a minimum value envelope line, each IMF signal is decomposed from the original signal, and finally, the signals are superposed to eliminate noise, so that the influence of mode aliasing and noise is effectively avoided; for model training, raw AlexNet is first pre-trained using historical appliance operating data, as in fig. 3. The training set then begins a second training. And finally obtaining a deep neural network suitable for the load identification task.
The non-intrusive load identification method based on the EMD and the AlexNet is characterized in that: a specific calculation formula is given in the AlexNet network:
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 an adjacent convolution kernel; k is a deviation; alpha and beta are self-defined values, the value range (0, 1) is determined by neuron parameters, and the determination is specifically determined according to the result backtracking verification of the accuracy rate.
The convolutional layer is the most important part of the convolutional neural network, and the capture of the image characteristics depends on the convolutional layer. The convolutional layer is subjected to object extraction through the process convolution template, and the convolutional layer can also be operated in different characteristic channels along with the transformation of the convolution template, so that the extraction of different characteristics and the integration of the same characteristics are realized. According to the change of data, the convolution kernel can be adjusted to a proper weight value through an optimization algorithm, and therefore the extracted features are most effective.
The operation formula of the convolutional neural network is as follows:
N＝(W-F+2P)/S+1 (3-2)
in the formula: w is the width and height of the input picture; f is the size of the convolution/pooling sum; s is the step pitch of convolution/pooling; p is the number of padding steps.
Description of the convolution procedure:
in the formula:
the weight of the i convolution kernel of the j layer;
the jth convolution local region for l layers; w is the width of the convolution kernel.
Root graph data has two important features, namely node features and structural features. The node characteristics describe the inherent properties of the nodes in the network and the structural characteristics describe the nature of the associations between the nodes. The structure generated by association not only helps much in the characterization of nodes in the network data, but also plays a key role in the characterization of the whole network. The core idea of the convolutional neural network is to aggregate node information by using edge information to generate a new node representation, which can automatically learn not only node characteristics but also association information between nodes.
The model proposed in this study is based on an AlexNet neural network, which consists of 5 convolutional layers and 3 fully connected layers. In the convolutional layer, three pooling layers are inserted to reduce the parameters in the model. The entire network contains 6.3 million links, 6000 million parameters and 65 million neurons. The structure of AlexNet is shown in fig. 3. The input layer of the network is a 224X 3 matrix, corresponding to the size of the input network image. The first convolutional layer has 96 convolutional kernels in total, and uses a larger convolution with a size of 11 × 11 and a step size of 4. The second layer is the LRN layer, followed by a 3 × 3 max pooling layer, with a step size of 4. The resulting convolution laver is relatively small, typically 5 x 5 or 3x3, with a step size of 1. The purpose of which is to scan all pixels. While the maximum pooling layer is still 3X3, step size is 2. It can be seen that in several convolutional layers in front of the network, although the amount of calculation is large, the amount of parameters is small. Most parameters of the model are in the fully connected layer. This is determined by the nature of the convolutional layer sharing weights:
1) And (6) an input layer. The input layer is an initial layer of the network, the input layer inputs 'power network data' converted from an active power sequence intercepted by a sliding window, inputs 'current network data' converted from a current sequence intercepted by the sliding window, inputs 'voltage network data' converted from a voltage sequence intercepted by the sliding window, inputs 'power factor network data' converted from a power factor sequence intercepted by the sliding window, and is subjected to data preprocessing operations such as normalization.
2) A first layer of convolutional layers. And the input dimension is a data characteristic dimension, the output dimension is 128, the result is activated by using a ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
3) A second layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
4) A third layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result to obtain the dropout result.
5) A fourth layer of convolutional layers. And the input dimension is 128, the output dimension is 128, the result is activated by using the ReLu function, and then dropout operation with the probability of 0.5 is used for the activation result, so that the dropout result is obtained.
6) A fifth layer of convolution layers. The input dimension is 128 and the output dimension is 1, and the result is activated using the ReLu function.
7) And network global average pooling layer. And taking the features obtained by performing global average pooling on the convolutional layer output results as the features of the network corresponding to the midpoint time.
8) And (5) outputting the layer. And outputting the characteristics of the corresponding midpoint moment of the network, which are output by the network global average pooling layer, as the power decomposition value of the target electrical appliance at the corresponding midpoint moment of the network.
The convergence rate of the AlexNet network is obviously accelerated in the training process. When the iteration times reach 10 times, the precision and the loss of the model approach to a steady trend due to the fact that the network is pre-trained by the transfer learning method, and the training process of the model is completed. Most parameters which are adjusted by the network in the AlexNet network are trained by the collected electric appliance operation data within three months. Thus, when the network is trained by this section of the training set, the network only needs to fine-tune a small number of parameters, making it more suitable for load recognition.
In order to improve the training speed, the samples are fed into the network training in batches by using the bath size, meanwhile, the overfitting phenomenon is further relieved by using the dropout early-stopping mechanism, adding L2 regularization and other measures, and network parameters are corrected by back propagation of an Adam optimizer until convergence. The best training parameters were saved for testing during training.
Selecting an absolute average Error (MAE), a Root Mean Square Error (RMSE) and a power decomposition accuracy rate P acc As a model evaluation index. The specific calculation method is as follows:
in the formula:
showing the predicted value of the model target electrical appliance power at the moment t; y is t Representing the true value of the target appliance power at time t.
(4) In the method, for the load identification and decomposition model based on event detection, the electric appliance decomposition effect on the electric appliances with lower power values and particularly large power fluctuation is not good, a non-invasive load identification and decomposition method based on AlexNet neural network processing is researched, the problem of operation information of each electric appliance under the total electric signals is solved in a combined manner, the identification result is obtained through a convolutional neural network, the excessive smoothness problem is reduced in the solving process, an early-stop mechanism is introduced, and the over-fitting phenomenon is prevented.
And establishing an intelligent load identification system for power utilization data mining, visually displaying the load identification, and decomposing a total electrical appliance sequence input by the system to obtain the sequence of each electrical appliance. And the automatic decomposition of the load of the electric appliance is realized by combining the practical application environment.
(5) In the method, an Absolute average Error (MAE), a Root Mean Square Error (RMSE) and a power decomposition accuracy rate P are selected acc As a model evaluation index. Obtaining the accuracy result of the final model output through index changes under different training rounds, and outputting the corresponding electrical information time sequence under the highest accuracy result
The method is characterized in that a non-invasive load identification technology is adopted, only electric signals at an electric power inlet are identified, feature extraction and machine learning algorithms are adopted, the converged voltage and current are analyzed, the service conditions of electric appliances on each branch circuit are identified, and the types and corresponding operation conditions of each load in a user domain are obtained through analysis. The method reduces the economic cost of the identification equipment, is easy to install and maintain, provides a real-time, economic and effective identification means for power companies and users, and also improves the accuracy of the system in identifying the residential electricity utilization rows.
It should be noted that, in this document, relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrases "comprising a component of' 8230; \8230;" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.
The above examples are only intended to illustrate the technical solution of the present invention, and not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
The confusion matrix is also called a probability table or an error matrix. It is a specific matrix used to present the effect of the algorithm, for example, fig. 6 is a matrix effect diagram of a blower, a microwave oven, a computer, an air conditioner, and a desk lamp.
Fig. 7 is a load-exploded view of a refrigerator and a microwave oven: the curve AGG is a total power utilization curve of a room; curve GT is the device real power curve; the AlexNet curve is a power load curve after model load decomposition of the neural network.
By decomposing each operation load in the total load power curve, each electric appliance power load curve is obtained, thereby achieving the purpose of load monitoring.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and any modifications or equivalent variations made in accordance with the technical spirit of the present invention may fall within the scope of the present invention as claimed.
Claims (4)
1. The non-invasive load identification method based on the EMD and the AlexNet comprises the following specific steps and is characterized in that:
(1) Data acquisition, namely acquiring resident electricity consumption data information;
(2) Data preprocessing, namely acquiring time sequence information of four dimensions of voltage, current, active power and power factor by using an EMD method;
in the step (2), data preprocessing comprises EMD empirical mode decomposition, and is performed on the acquired nonlinear signals to eliminate interference in the sampling data;
the empirical mode step network for the nonlinear electrical signal EMD comprises the following steps:
step 1: and searching all extreme points of the signal, connecting the local maximum points into an upper envelope line through a cubic spline curve, and connecting the local minimum points into a lower envelope line. The upper and lower envelopes contain all data points;
and 2, step: obtaining a first IMF component if the IMF condition is met according to the average value of the upper envelope and the lower envelope;
and step 3: if the IMF condition is not met, the obtained data is used as original data, the step 1 and the step 2 are repeated to obtain the mean value of the upper envelope and the lower envelope, whether the required condition of the IMF component is met or not is calculated, and if the required condition is not met, the steps are repeated;
and 4, step 4: separating the IMF from the signal: repeating the three steps as an original signal for a plurality of times to obtain a second IMF component till an nth IMF component;
and 5: when becoming a monotonic function, the remainder becomes a residual component. The sum of all IMF components and residual components is an original signal;
(3) Utilizing AlexNet neural network processing to acquire various electricity data including four-dimensional information of voltage, current, active power and power factor aiming at resident electricity data, and carrying out AlexNet neural network training and testing;
in the step (3), the AlexNet neural network processing comprises the following steps:
the AlexNet network comprises 5 convolutional layers Conv,3 maximum pooling layers Maxpooling and 3 fully-connected layers dense, wherein the convolutional layers and the maximum pooling layers are alternately arranged, different convolutional layers have different convolutional cores, and the low-level feature extraction capability is different;
a specific calculation formula is given in the AlexNet neural network:
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 an adjacent convolution kernel; k is a deviation; alpha and beta are self-defined values, the value range (0, 1) is determined by neuron parameters, and the values are determined by backtracking verification according to accuracy results;
the operation formula of the convolutional neural network is as follows:
N＝(W-F+2P)/S+1 (1-2)
in the formula: w is the width and height of the input neuron, wherein the width represents the data dimension, and the height represents the numerical value; f is the size of the convolution/pooling sum; s is the step pitch of convolution/pooling; p is the step number of padding;
description of the convolution procedure:
in the formula:
the weight of the i convolution kernel for the j layer;
the j-th convolution local region for l layers; w is the width of the convolution kernel;
using bath size to send samples into network training in batches, simultaneously using dropout and adding L2 regularization to further relieve overfitting phenomena, using an Adam optimizer to reversely propagate and correct network parameters until convergence, and storing the best training parameters in the training process for testing;
selecting Absolute average Error Mean Absolute Error (Mean Absolute Error, MAE), root Mean square Error (Root Mean square Error, RMSE) and power decomposition accuracy rate P acc As the model evaluation index, a specific calculation method is as follows:
in the formula:
representing a predicted value of the model target electrical appliance power at the time t; y is t The actual value of the target electrical appliance power at the time t is represented;
(4) Load identification, namely decomposing the electricity utilization information of residents in the bus by using the trained neural network to obtain the electricity utilization information of each electrical appliance on each branch;
(5) The algorithm outputs, corresponding to the detected rising and falling edges of the total power, are most likely caused by a pair of load events, i.e. indicating the beginning and end of a certain plant operating state, respectively.
2. The non-invasive load identification method based on EMD and AlexNet according to claim 1, characterized in that in step (1), the data acquisition means acquiring the high frequency load of the load operation, and the frequency is 1hz.
3. The non-intrusive load identification method based on EMD and AlexNet of claim 1, wherein in the step (4), if the number of positive clusters is not equal to the number of negative clusters, the obtained clusters are merged with the part having the larger number of clusters, and during merging, the two clusters having the smallest difference value between the mean values of the clusters are merged until the number of positive clusters is equal to the number of negative clusters.
4. The non-invasive load identification method based on EMD and AlexNet according to claim 1, characterized in that in (4), the load characteristic database is established, and the load characteristics of various devices are summarized through manual measurement or automatically performed by using a classification algorithm in machine learning;
and in the load identification process, the accuracy and the loss function of the model are used as judgment bases, the EMD empirical mode decomposition is carried out to obtain each electrical signal for identification and decomposition, the operation data of each electrical appliance on each branch is obtained, and the time sequence of the obtained load operation data is used as a final output result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211451964.9A CN115758246A (en) | 2022-11-21 | 2022-11-21 | Non-invasive load identification method based on EMD and AlexNet |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211451964.9A CN115758246A (en) | 2022-11-21 | 2022-11-21 | Non-invasive load identification method based on EMD and AlexNet |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115758246A true CN115758246A (en) | 2023-03-07 |
Family
ID=85333041
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211451964.9A Pending CN115758246A (en) | 2022-11-21 | 2022-11-21 | Non-invasive load identification method based on EMD and AlexNet |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115758246A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116051910A (en) * | 2023-03-10 | 2023-05-02 | 深圳曼顿科技有限公司 | Non-invasive load identification method, device, terminal equipment and storage medium |
| CN116089800A (en) * | 2023-04-10 | 2023-05-09 | 武汉工程大学 | Method and system for extracting and correcting ringing component of dynamic pressure measurement signal of shock wave flow field |
| CN116595488A (en) * | 2023-07-19 | 2023-08-15 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
| CN117330816A (en) * | 2023-12-01 | 2024-01-02 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
| CN118013259A (en) * | 2024-04-09 | 2024-05-10 | 中国人民解放军海军工程大学 | Data analysis method based on non-contact measurement and related equipment thereof |
-
2022
- 2022-11-21 CN CN202211451964.9A patent/CN115758246A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116051910A (en) * | 2023-03-10 | 2023-05-02 | 深圳曼顿科技有限公司 | Non-invasive load identification method, device, terminal equipment and storage medium |
| CN116089800A (en) * | 2023-04-10 | 2023-05-09 | 武汉工程大学 | Method and system for extracting and correcting ringing component of dynamic pressure measurement signal of shock wave flow field |
| CN116595488A (en) * | 2023-07-19 | 2023-08-15 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
| CN116595488B (en) * | 2023-07-19 | 2023-11-14 | 青岛鼎信通讯股份有限公司 | Non-invasive load identification method based on intelligent ammeter |
| CN117330816A (en) * | 2023-12-01 | 2024-01-02 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
| CN117330816B (en) * | 2023-12-01 | 2024-01-26 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
| CN118013259A (en) * | 2024-04-09 | 2024-05-10 | 中国人民解放军海军工程大学 | Data analysis method based on non-contact measurement and related equipment thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115758246A (en) | Non-invasive load identification method based on EMD and AlexNet | |
| Xuan et al. | Multi-model fusion short-term load forecasting based on random forest feature selection and hybrid neural network | |
| CN113962364B (en) | Multi-factor power load prediction method based on deep learning | |
| CN111080032A (en) | Load prediction method based on Transformer structure | |
| CN111880044B (en) | Online fault positioning method for distribution network containing distributed power supply | |
| CN109685314B (en) | Non-intrusive load decomposition method and system based on long-term and short-term memory network | |
| CN111178616B (en) | Wind speed prediction method based on negative correlation learning and regularization extreme learning machine integration | |
| CN109063911A (en) | A kind of Load aggregation body regrouping prediction method based on gating cycle unit networks | |
| CN104523264B (en) | Electrocardiosignal processing method | |
| CN110414788A (en) | A kind of power quality prediction technique based on similar day and improvement LSTM | |
| Liu et al. | Heating load forecasting for combined heat and power plants via strand-based LSTM | |
| CN106779219A (en) | A kind of electricity demand forecasting method and system | |
| CN109255726A (en) | A kind of ultra-short term wind power prediction method of Hybrid Intelligent Technology | |
| CN114462718A (en) | CNN-GRU wind power prediction method based on time sliding window | |
| CN114022311A (en) | Comprehensive energy system data compensation method for generating countermeasure network based on time sequence condition | |
| CN111242353A (en) | Wind power combined prediction modeling and prediction method | |
| Akpinar et al. | Forecasting natural gas consumption with hybrid neural networks—Artificial bee colony | |
| CN115659254A (en) | Power quality disturbance analysis method for power distribution network with bimodal feature fusion | |
| CN115828755A (en) | Method and device for evaluating participation of micro-grid group in power grid service and readable storage medium | |
| CN115759389A (en) | Day-ahead photovoltaic power prediction method based on weather type similar day combination strategy | |
| CN115409369A (en) | Comprehensive energy system reliability evaluation method based on mechanism and data hybrid driving | |
| CN110059737B (en) | Distribution transformer connection relation identification method based on integrated deep neural network | |
| CN114488069A (en) | Radar high-resolution range profile identification method based on graph neural network | |
| CN113449920A (en) | Wind power prediction method, system and computer readable medium | |
| CN117151770A (en) | Attention mechanism-based LSTM carbon price prediction method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |