CN114169435A

CN114169435A - Non-invasive load identification method based on feature visualization

Info

Publication number: CN114169435A
Application number: CN202111487529.7A
Authority: CN
Inventors: 王澍; 徐艺文; 董秀青; 何念; 郑旭丹
Original assignee: FUZHOU UNIVERSITY ZHICHENG COLLEGE
Current assignee: FUZHOU UNIVERSITY ZHICHENG COLLEGE
Priority date: 2021-12-07
Filing date: 2021-12-07
Publication date: 2022-03-11

Abstract

The invention relates to a non-invasive load identification method based on feature visualization. Aiming at the problems that the network model parameters of the traditional load identification algorithm based on one-dimensional sequence characteristic quantity are huge, and the identification accuracy is low due to high-power load and harmonic abundant load when multiple loads work simultaneously, the harmonic analysis is carried out on the original current data, and then the one-dimensional harmonic characteristic sequence is converted into a two-dimensional image by methods of Gramian Angular Field (GAF) and recursion graph (RP) and is used as the input of an image classification model based on CNN. After the feature visualization method is adopted, the load identification accuracy rate of 98.272% on the PLAID data set and 96.573% on the self-collected data set can be achieved by using a network and calculation cost which are smaller than the traditional algorithm scale.

Description

Non-invasive load identification method based on feature visualization

Technical Field

The invention relates to a non-invasive load identification method based on feature visualization.

Background

According to the traditional load identification algorithm based on the one-dimensional sequence characteristic quantity, the network model parameters are huge, and the accuracy is low due to the fact that high-power loads and harmonic abundant loads work simultaneously under the condition of multiple loads. The current research focuses on the analysis of one-dimensional sequence characteristics of the load, does not consider the time correlation of signals, and wastes information characteristics on a time scale. With the increase of the number of household electric equipment, loads with similar power are easy to be mistaken; meanwhile, the low-power load is easily submerged by the high-power load; the defects of huge network model parameters, high model training overhead, fuzzy recognition and the like of the algorithm generally exist.

Disclosure of Invention

The invention aims to solve the defects in the prior art and provides a non-invasive load identification method based on feature visualization.

In order to achieve the purpose, the technical scheme of the invention is as follows: a non-intrusive load identification method based on feature visualization is realized as follows:

and converting the harmonic content sequence corresponding to the one-dimensional current amplitude sequence into a two-dimensional image by using a method of a gram angular field and a recursion graph, and then completing the task of load type identification by using an image classification model based on CNN.

In an embodiment of the present invention, the fundamental wave and each subharmonic component are obtained by performing fourier series decomposition on a periodic non-sinusoidal wave, and the following formula is given:

wherein, c₀Representing a direct current component, c_mRepresenting the amplitude of the m-th harmonic, m omega representing the angular frequency of the m-th harmonic,

represents the initial phase of the m-th harmonic;

the harmonic content is the ratio of the root mean square value of the harmonic component to the root mean square value of the fundamental component, and is expressed by percentage; the h-order current harmonic content formula is as follows:

in the formula I_{h ms}Representing the effective value of the h harmonic current, i_k(k) Representing the instantaneous value of the h-th harmonic current at time k, simplified to equal the peak value I of the fundamental current_1mAnd the peak value I of the harmonic component_hmThe ratio of.

In one embodiment of the invention, the normalized one-dimensional sequence data is converted from a Cartesian coordinate system to a polar coordinate system, and then the intrinsic correlations at different points in time are identified by considering the angles and/or differences between the different points in time; the angle making and corresponding realizing method is GASF, and the angle making and corresponding realizing method is GADF.

In an embodiment of the invention, the gram-sum angular field GASF and the gram-difference angular field GADF are defined as follows:

wherein, X is a current waveform signal.

In an embodiment of the present invention, the phase space is reconstructed, and the embedding dimension m and the delay time τ are selected in the following manner: calculating an embedding dimension m by a false nearest neighbor method or a C-C algorithm, calculating a delay time tau by an average mutual information method and an average displacement analysis method, and then performing phase space reconstruction to obtain the following matrix:

in phase space, i time vector X_iAnd j time vector X_jProximity D_ij＝||X_i-X_jExpressed in | l, the recursion graph is drawn by computing the following equation:

R_ij＝Θ(ε-D_ij)

where ε is the threshold and Θ (·) is the Heaviside function.

In an embodiment of the invention, the CNN has good performance in the field of image classification, and performs supervised training on a sample by extracting features of different layers, wherein the CNN model has an input layer, four convolutional layers, two pooling layers, three full-connection layers and an output layer.

Compared with the prior art, the invention has the following beneficial effects: the invention provides a novel load identification algorithm based on feature visualization, which converts a harmonic content rate sequence corresponding to a one-dimensional current amplitude sequence into a two-dimensional image by a method of a gram angular field and a recursion graph, and then efficiently completes the task of load type identification by utilizing an image classification model based on CNN. The recognition rate of the method provided by the invention on the PLAID data set and the self-collected data set respectively reaches 98.272% and 96.573%, and the effectiveness and reliability of the method provided by the invention are fully verified.

Drawings

FIG. 1 is an overall flow chart of the method of the present invention.

Fig. 2 is a graph of current waveform and harmonic content.

Fig. 3 is a cartesian to polar coordinate system conversion.

Fig. 4 is a graph of the effect of the gram angular field (left GASF, right GADF).

Fig. 5 is a recursive graph effect diagram.

FIG. 6 is a gram angular field plot generated from 10 current harmonic sequences in an acquired dataset.

Fig. 7 is a recursive diagram generated from 10 current harmonic sequences in an acquired data set.

Fig. 8 is a gram angle field plot generated by 10 current harmonic sequences in the PLAID data set.

Fig. 9 is a recursive diagram of generation of 10 current harmonic sequences of the PLAID data set.

Fig. 10 is a diagram of a CNN network structure used in the present invention.

Fig. 11 is a comparison of the classification recognition results of three load feature images.

Fig. 12 is a training process for CNN-based image classification model input GASF maps (self-acquired dataset).

Fig. 13 is a training process of inputting a GASF map (PLAID dataset) based on the CNN image classification model.

Fig. 14 is a one-dimensional signature sequence recognition confusion matrix based on the self-collected data set and CNN.

Fig. 15 is a confusion matrix based on self-collected data sets and the identification of the results by the GASF method.

Fig. 16 is a confusion matrix based on the identification results of the PLAID dataset and the GASF method.

Detailed Description

The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.

The invention relates to a non-invasive load identification method based on feature visualization, which is realized as follows:

The following is a specific embodiment of the present invention.

The specific flow of the algorithm provided by the invention is shown in figure 1: carrying out high-frequency sampling on electric power data by using a non-invasive method, and firstly carrying out preprocessing such as filtering, error detection and the like on original current amplitude data; then, harmonic analysis is carried out, and fundamental wave and higher harmonic features are extracted; next, converting the one-dimensional harmonic feature sequence into a two-dimensional image based on a method of a Gram Angular Field (GAF) and a Recurrence Plot (RP); secondly, the two-dimensional image is used as a characteristic input model of the load to be trained so as to improve the accuracy of load identification; and finally, the effectiveness and reliability of the algorithm provided by the invention are verified on a public data set PLAID and a self-collected data set.

1. Feature visualization method based on current harmonics

A. Current harmonic extraction

Various household appliances contain a plurality of different types of electric elements, resistors, capacitors, inductors and the like have different properties, and the complex circuit causes the current waveform actually consumed by the load to be non-sinusoidal. The fundamental wave and each subharmonic component can be obtained by performing Fourier series decomposition on the periodic non-sinusoidal wave, and the formula is as follows:

indicating the initial phase of the m harmonic.

The harmonic content is the ratio of the root mean square value of the harmonic component to the root mean square value of the fundamental component, expressed as a percentage. The h-order current harmonic content formula is as follows:

in the formula I_{h ms}Representing the effective value of the h harmonic current, i_k(k) Representing the instantaneous value of the h harmonic current at time k. After simplification, equal to current fundamental wavePeak value I_1mAnd the peak value I of the harmonic component_hmThe ratio of.

Due to the difference in the electric element and the circuit structure, the harmonic contents of different loads are clearly distinguished, and can be roughly classified into a pure resistance type, a pump driving type, a motor driving type, and an electronic power supply type.

The harmonic condition of common household appliances is shown in fig. 2, the left graph of each appliance is the original current waveform, the middle is the fundamental wave and each harmonic amplitude, and the right graph is the harmonic content. The hot water kettle and the electric hair drier belong to pure resistance type, because the work of the hot water kettle and the electric hair drier mainly depends on resistance heating, almost no higher harmonic wave exists, and the state switching is almost completed instantly; the washing machine and the refrigerator belong to pump driving type, more odd harmonic waves can be generated in the operation process, and obvious and longer transient processes can occur in state switching; motor driving equipment such as a fan and the like is similar to pump driving equipment, but the state switching is not obvious and short; electronic power supply equipment such as mobile phone chargers and computers can generate rich higher harmonics due to the fact that the high-frequency switch is used for adjusting output voltage.

B. Feature visualization method based on Gelam angular field GAFs

The gram angular field has good effects in the fields of rolling bearing fault diagnosis, electrocardiosignal analysis, arrhythmia classification and the like. The idea is to convert the normalized one-dimensional sequence data from a cartesian coordinate system to a polar coordinate system and then identify the intrinsic correlations at different points in time by taking into account the angles and/or differences between the different points in time. The angle making and corresponding realizing method is GASF, and the angle making and corresponding realizing method is GADF.

In a cartesian coordinate system, the current waveform signal X ═ { X ═ X₁，x₂，…，x_nBelongs to a typical one-dimensional signal, x_nRepresenting the current amplitude sampled at time n, first normalized:

the current amplitude after normalization with the sampling time n as the radius

For angle phi, the processed current waveform sequence

From a cartesian coordinate system to a polar coordinate system (as shown in fig. 3).

Due to the fact that

Therefore φ ∈ [0, π ∈]And cos (phi) is at [0, pi ]]The above is monotonous, and it is known that the encoding method of expression (2) has double mapping properties, and can prevent information loss at the time of one-dimensional sequence conversion and ensure uniqueness between mapping and inverse mapping. At the same time, radius maintains the time dependence of the sequence.

Define the GAF matrix as:

G_{(i，j||i-j|＝k)}indicating the sum of the direction values phi, the principal diagonal G, in the time interval k_i，iThe special case is that k is 0, which contains the original information of the sequence point.

In a polar coordinate system, the time dependencies in different time intervals can be identified by trigonometric functions and/or differences between points. The gram-sum angular field (GASF) and the gram-difference angular field (GADF) are defined as follows:

the length of the original current waveform sequence is N, the size of the obtained gram matrix GAF is N multiplied by N, when N is large, the GAF is caused to be too large, and the calculation amount is increased sharply. Therefore, before the sequence is input into the gram angle field, the sequence is processed by a piecewise aggregation approximation method, that is, the sequence is divided into small intervals with fixed length, and the average value of data in each interval is calculated to represent the whole interval. The resulting G is a matrix representation of the two-dimensional image. Fig. 4 is a graph of the effect of the gram angular field (left GASF, right GADF).

C. Characteristic visualization method based on recursive graph RP

The Recursion Plot (RP) is an important method for analyzing the periodicity, chaos and non-stationarity of a time sequence, and can reveal the internal structure of the time sequence and give a priori knowledge about similarity, information quantity and predictability. The recursive graph is particularly suitable for short-time sequence data, and can check the stationarity and the internal similarity of the time sequence. The recursion diagram is widely applied to analysis of no-load closing vibration signals of the transformer and identification of surface defects of materials^[18]And structural internal damage detection.

First, the most important step, phase space reconstruction, directly affects the quality of the recursive graph. The key to reconstructing the phase space is the choice of the embedding dimension m and the delay time τ. The embedding dimension m is calculated by a false nearest neighbor method or a C-C algorithm, and the delay time tau can be calculated by an average mutual information method and an average displacement analysis method. Then, phase space reconstruction is carried out to obtain the following matrix:

in phase space, i time vector X_iAnd j time vector X_jProximity D_ij＝||X_i-X_jExpressed by | |, by calculation

R_ij＝Θ(ε-D_ij) (9)

To draw a recursive graph as shown in figure 5.

2. Network model training based on feature visualization

A. Load data set

In order to verify the accuracy and reliability of the algorithm proposed by the present invention, the present invention uses a self-sampled data set and a PLAID data set. The data of sampling by oneself is established based on STM 32's non-intrusive electric power data acquisition terminal, has used open type current sensor and AD7606 module, samples five kinds of common household electrical appliances of thermos, desk lamp, refrigerator, fan, TV set with 12 KHz's sampling rate, totally 32 kinds of combination service conditions, 1000 groups of data of each kind, totally 32000 group current amplitude sequence. And calculating the corresponding harmonic content of each group of sequences, and converting the harmonic content into a gram angular field graph and a recurrence graph with the image size of 40 multiplied by 40.

The Plug-Level application Identification Dataset contains current and voltage data collected at a 30KHz sampling rate from 11 different types of devices in 60 households of pittsburgh, pa. The various loads in the data set, such as air conditioners, refrigerators, and washing machines, etc., have different instances of different brands/models. The PLAID dataset has been widely used in recent NILM studies. The invention also selects 32 different types of data with single load and multi-load consumption.

Fig. 6-9 show feature visualization images of partial load types in self-acquired datasets and PLAID datasets.

Each gram angular field map contains the original current harmonic information and the corresponding time relationship. The [0, 1] interval is mapped by a purple-to-red gradient process, in the graph, the harmonic amplitudes are different, the color display is different, the difference of different subharmonics can be obviously enhanced, and more characteristic information than the original one-dimensional current harmonic sequence is extracted. The recursion graph strengthens the internal structure information of the original one-dimensional current harmonic sequence through the distribution and the variation trend of black and white pixels, and can better judge the internal similarity of different sequences.

B. Image classification model structure

The CNN has good performance in the field of image classification, and supervised training is carried out on samples by extracting features of different layers. CNNs are generally composed of convolutional layers, pooling layers, and fully-connected layers. A convolution layer (Conv) performs a convolution operation on an input image to extract an image feature. Pooling layers (Poolling layers) are used to reduce the number of connections between convolutional layers to reduce the computational burden, while reducing feature dimensions. A Fully connected layer (FC) connects two-dimensional images into one-dimensional vector for output of results.

The CNN model used in the present invention is composed of an input layer, four convolutional layers, two pooling layers, three full-link layers, and an output layer, as shown in fig. 10.

Defining a volume block: firstly, extracting the characteristics of a two-dimensional image by using a convolution operator; then, using the RELU function to execute activation operation, and increasing the learning ability of the network to the nonlinear characteristics; and then adding a pooling layer to perform dimension reduction and abstraction on the features to prevent an overfitting phenomenon. And two rolling blocks are used in the linear superposition, three full-connection layers are connected, the classification is carried out by using a Softmax function, and finally the result is output by an output layer. The model has 82792 parameters in total and the size of the model is 1.01 Mb. The evaluation of the models can compare their classification performance from Accuracy (Accuracy), Precision (Precision), Recall (Recall), and F1 score (F1-score).

3. Results and analysis of the experiments

Comparison of different feature visualization methods with conventional methods

All models of the invention are trained on a TX9-CU5DK notebook platform, and a CPU is

Core^TMi5-10400, Yingwei RTX20708G GDDR6 independent display card, 16GB operation memory.

1) Results of conventional methods of identification

For a one-dimensional current amplitude Sequence, the conventional method widely uses a CNN network, a Long short-term memory (LSTM) network combined with an attention mechanism, and a Sequence-to-Sequence (seq 2seq) network, and performs experiments on self-collected data sets and pladd data, and the recognition results are shown in table 1 below:

TABLE 1 comparison of results of conventional one-dimensional signature sequence identification methods

As can be seen from table 1, the performance of the three algorithms on the plidi data set is better than that of the self-collected data set, because the self-collected data contains a table lamp with power of only 7w, when the load works simultaneously, the table lamp is easily covered by other high-power loads and is difficult to identify; the power of each electric appliance in the PLAID is larger, and certain difference exists among the electric appliances, and when various electric appliances work simultaneously, the identification effect is superior to that of a self-sampling data set. The Seq2Seq algorithm has the largest parameter scale, 2183632 parameters exist, the evaluation indexes such as the accuracy and the F1 score are also the highest, the accuracy on the PLAID data set reaches 86.271%, and the accuracy on the self-collected data also has 84.317%. Secondly, the LSTM algorithm achieved 84.127% and 81.345% accuracy on the plid dataset and on the self-collected dataset, respectively. And the CNN network only considers the local characteristics of the sequence and does not consider the relevance before and after time, so that the CNN network cannot effectively identify the overall weak difference of the current sequence, and the accuracy rate on the two data sets is not more than 60%. Therefore, the problems of large scale of model parameters and low recognition degree generally exist in the traditional recognition algorithm based on the one-dimensional characteristic sequence.

2) Results of the feature visualization method

In the current research, the main method for visualizing the load characteristics is a V-I trace graph, which integrates the information of current and voltage at a certain time and generates the trace graph as the training input of a network model. Compared with the method of the GASF graph and the recursion graph provided by the invention, the performance comparison is carried out on the CNN-based image classification network, and the result is shown in fig. 11.

As can be seen from FIG. 11, both methods proposed by the present invention performed better on both datasets than the V-I trace plot. The GASF graph method has 96.573% accuracy on the self-acquisition data set, and even reaches 98.272% accuracy on the PLAID; the recursion plot achieved 95.031% and 96.317% accuracy on the self-collected data set and the PLAID, respectively. The performance of the feature visualization method on the two data sets is consistent with the performance of the one-dimensional feature sequence method, the recognition rate of the PLAID data set is slightly higher than that of the self-acquisition data set, and the effectiveness of the data sets is explained.

The recursion graph divides the current sequence according to time, and then the corresponding harmonic characteristic information is mapped into position information in the two-dimensional image, so that the information density of the internal structure of the sequence is obviously improved, and more detailed information can be obtained compared with a V-I track graph; the GASF graph overcomes the limitation that a recursion graph and a V-I track graph cannot reflect more interharmonic differences only by using gray values, the interharmonic differences are increased through color components, and the load characteristics are strengthened. Therefore, the GASF graph-based approach has the highest recognition rate, followed by the recursive graph-based approach.

As can be seen from fig. 12 and 13, the accuracy of the model is still low in the initial stage, and after 10 rounds of training and learning, the Loss is rapidly reduced to below 1.0, and the accuracy is improved to above 50%; after about 30 rounds of training and learning, the accuracy rate is continuously improved and reaches more than 90 percent; after 50 rounds, the network tends to stabilize, indicating that the network has converged. Under the action of Early Stopping mechanism, the training of self-collected data sets is stopped at the 60 th round, the training of PLAID data sets is stopped at the 63 th round, the accuracy rates of 96.931% and 98.704% on the test set are respectively reached, and the accuracy rates on the verification set are 96.573% and 98.272% respectively. As can be seen from the loss curves of the training set and the test set, the variation trends of the training set and the test set are basically consistent, which indicates that the network runs well, and the problems of under-fitting and over-fitting do not occur.

B. Type analysis for payload recognition

Fig. 14 shows the worst identification effect in the experiment, which is based on the one-dimensional feature sequence identification method of the self-collected data set and CNN. The single-load operation recognition effect is good, but when multiple loads operate simultaneously, only the desk lamp and the hot water kettle are mistakenly considered to be working, and the desk lamp with rich harmonic waves is generated due to the fact that the high-power hot water kettle and the LEDs are stroboscopic. The power difference between the hot water kettle and the fan and the power difference between the hot water kettle and the desk lamp are two orders of magnitude, and the hot water kettle is easily ignored as the current noise of the mains supply when working at the same time. The second harmonic wave to the twelfth harmonic wave of the desk lamp are distributed in a hump shape, the content of the odd harmonic wave is obviously larger than that of the even harmonic wave, and the odd harmonic wave and the even harmonic wave are sequentially decreased and are similar to the harmonic wave characteristic of a television; on the other hand, the pump drive and motor drive loads of refrigerators, fans and the like only have high odd harmonic content of three to seven times, and are easily covered by the harmonic characteristics of table lamps and televisions during working. The condition that the desk lamp and the hot water kettle do not participate in simultaneous working is avoided, and the condition of recognition confusion is obviously improved. The method based on LSTM and the method based on Seq2Seq considers the time correlation of sequence characteristics, thereby relieving the problem of fuzzy recognition caused by high power and harmonic rich load to a certain extent, but the recognition accuracy is still low.

Fig. 15 shows the recognition effect of the GASF graph method and the classification model based on CNN images on self-collected data, which fully illustrates that the method provided herein can effectively recognize various types of loads under the condition of simultaneous operation of multiple loads. Fig. 16 is a representation of the method based on the GASF diagram on the PLAID data, which is the best recognition effect in all experiments, and compared with the CNN, LSTM and Seq2Seq algorithms using one-dimensional feature vectors, the method not only has a smaller parameter scale and relatively simple iterative computation, but also greatly reduces training overhead, shortens training time, and successfully solves the problem that the characteristics of high-power load and harmonic rich load are too prominent.

4. Conclusion

Aiming at the problems that the traditional load identification algorithm based on one-dimensional sequence characteristic quantity has huge network model parameters and the accuracy is low due to high-power load and harmonic abundant load when multiple loads work simultaneously, the invention provides a novel load identification algorithm based on characteristic visualization, which converts a harmonic content rate sequence corresponding to a one-dimensional current amplitude sequence into a two-dimensional image by a method of a gram angle field and a recursion diagram, and then utilizes an image classification model based on CNN to efficiently complete the task of load type identification. The identification rates of the method on the PLAID data set and the self-collected data set respectively reach 98.272% and 96.573%, and the effectiveness and the reliability of the method are fully verified. In future work, other characteristic information such as power and voltage is combined in the method, the generalization performance of the model is further improved, and more complex multi-state load types under the multi-load combined working condition are identified.

The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.

Claims

1. A non-intrusive load identification method based on feature visualization is characterized by comprising the following steps:

2. The method of claim 1, wherein the fundamental and each harmonic component is obtained by fourier series decomposition of a periodic non-sinusoidal wave, and the formula is as follows:

represents the initial phase of the m-th harmonic;

in the formula I_hrmsRepresenting the effective value of the h harmonic current, i_k(k) Representing the instantaneous value of the h-th harmonic current at time k, simplified to equal the peak value I of the fundamental current_1mAnd the peak value I of the harmonic component_hmThe ratio of.

3. The non-invasive load identification method based on feature visualization as claimed in claim 1, wherein the normalized one-dimensional sequence data is converted from Cartesian coordinate system to polar coordinate system, and then the intrinsic correlation of different time points is identified by considering the angle and/or difference between different points; the angle making and corresponding realizing method is GASF, and the angle making and corresponding realizing method is GADF.

4. The non-invasive load identification method based on feature visualization as claimed in claim 3, wherein the gram sum angular field GASF and the gram difference angular field GADF are defined as follows:

wherein, X is a current waveform signal.

5. The method of claim 1, wherein the phase space is reconstructed, and the embedding dimension m and the delay time τ are selected by: calculating an embedding dimension m by a false nearest neighbor method or a C-C algorithm, calculating a delay time tau by an average mutual information method and an average displacement analysis method, and then performing phase space reconstruction to obtain the following matrix:

R_ij＝Θ(ε-D_ij)

where ε is the threshold and Θ (·) is the Heaviside function.

6. The method as claimed in claim 1, wherein the CNN model has an input layer, four convolutional layers, two pooling layers, three full-link layers and an output layer, and performs supervised training on the sample by extracting features of different levels.