CN113687182A - Household load identification method, program and system based on noise reduction automatic encoder - Google Patents
Household load identification method, program and system based on noise reduction automatic encoder Download PDFInfo
- Publication number
- CN113687182A CN113687182A CN202110927020.3A CN202110927020A CN113687182A CN 113687182 A CN113687182 A CN 113687182A CN 202110927020 A CN202110927020 A CN 202110927020A CN 113687182 A CN113687182 A CN 113687182A
- Authority
- CN
- China
- Prior art keywords
- load
- noise reduction
- load identification
- power
- data
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/081—Locating faults in cables, transmission lines, or networks according to type of conductors
- G01R31/086—Locating faults in cables, transmission lines, or networks according to type of conductors in power transmission or distribution networks, i.e. with interconnected conductors
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R21/00—Arrangements for measuring electric power or power factor
- G01R21/001—Measuring real or reactive component; Measuring apparent energy
- G01R21/002—Measuring real component
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R22/00—Arrangements for measuring time integral of electric power or current, e.g. electricity meters
- G01R22/06—Arrangements for measuring time integral of electric power or current, e.g. electricity meters by electronic methods
- G01R22/061—Details of electronic electricity meters
- G01R22/063—Details of electronic electricity meters related to remote communication
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Power Engineering (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Remote Monitoring And Control Of Power-Distribution Networks (AREA)
Abstract
The invention provides a household load identification method, a household load identification program and a household load identification system based on a noise reduction automatic encoder, which relate to the technical field of power grid load identification. Through example tests, the household load identification method has high accuracy in identifying the real equipment power consumption and judging the state of the equipment, has high universality, and can effectively identify the same equipment with different models and brands.
Description
Technical Field
The invention relates to the technical field of power grid load identification, in particular to a household load identification method, a household load identification program and a household load identification system based on a noise reduction automatic encoder.
Background
At present, the household electric energy meter can only count the total household electricity consumption and cannot perform classified metering on various load consumed electric energy. The load identification enables family users and electric power companies to have better knowledge on power utilization behaviors and equipment energy consumption, and supports upper-layer application of the intelligent power distribution network. In addition, with the increase of new energy, the power distribution network is required to have a faster and more accurate demand side response function, and the realization of the function also needs to depend on a load identification technology.
The existing non-intrusive load identification method has high requirement on measurement data, and the measurement data, namely high-frequency load current data or transient waveform at the moment of load starting, can be obtained only after a common electric energy meter is modified, so that extra cost is increased. In recent years, load identification only by means of low-frequency single measurement has been proposed, such as load judgment by means of current effective value or on/off judgment of electrical equipment by means of steady-state time domain active and reactive power, but these methods have a common disadvantage that identification effect is poor when a plurality of loads with similar steady-state waveforms are simultaneously turned on.
Disclosure of Invention
One of the objectives of the present invention is to provide a non-invasive method for identifying a home load with low measurement requirements and better resolution of similar loads of a steady-state power waveform.
In order to achieve the purpose, the invention adopts the following technical scheme: in the household load identification method based on the noise reduction automatic encoder, the power of other equipment outside the target equipment is taken as noise or noise, the power characteristic of the target equipment is extracted through machine learning training of the noise reduction automatic encoder, in a load separation stage, an input mixed power signal is analyzed by adopting a sliding window, a median filter is used for processing an overlapping part of the sliding window, the output value of the overlapping window is replaced by the statistical median of all values in a neighborhood, and the load separation is carried out according to the following steps:
firstly, encoding a network:
1. processing the raw total input power consumption data by one or more one-dimensional convolution layers to generate a set of feature maps;
2. each convolution layer sequentially passes through a linear activation function, a maximum pooling layer, an additional convolution layer and a pooling layer to form a fully-connected multilayer sensor;
3. the full connection layer is processed by the modified linear unit activation function, and the whole encoding process is finished;
secondly, decoding the network:
4. performing up-sampling on the fully-connected multilayer perceptron through deconvolution;
5. performing pooling on the result in the step 4, wherein the pooling is an inverse process of maximum pooling;
6. performing up-sampling on the result in the step 5 through deconvolution;
7. and obtaining a noise reduction signal reconstructed by decoding, and obtaining power consumption information of the target equipment according to the noise reduction signal reconstructed by decoding.
In step 2, the neighboring maxima are obtained by maximum pooling, making the activation function positions within the analysis window more independent and reducing the length of the feature map and the number of full link layer elements.
In step 3, the modified linear unit activation function compares the input with zero and outputs a larger value.
Specifically, when the noise reduction automatic encoder is trained, the mean square error between the minimized output and the activation function of the target device is taken as a target, and a random gradient descent method is adopted to optimize the training parameters.
In addition, the invention also relates to a household load identification program based on the noise reduction automatic encoder, which is stored in a computer or a server connected with a data input device and executes the steps 1 to 7 in the household load identification method, and when the program runs, the data input device is used for acquiring the active power measurement data of a target household within a set time period range.
The invention further provides a household load identification system based on the noise reduction automatic encoder, which comprises a data input device and a server running the household load identification program, wherein the data input device is connected with the server and transmits the obtained active power measurement data to the server running the household load identification program.
Further, the household load identification system further comprises a user side, the user side is connected with the server through a communication network, and the server sends the identified power consumption information of the electrical equipment to the corresponding user side through the communication network.
The invention provides a brand-new non-intrusive load identification method, which is based on the principle of a noise automatic encoder, only depends on the conventional sampling rate and single active power measurement data, treats the total mixed power as pictures or sound records needing to be processed, treats the power generated by other equipment (equipment without concern) except a target equipment as 'noise' or 'noise', and performs target change to identify the load power of the target equipment (single equipment of interest) from the total mixed power. Through example tests, the method has higher accuracy in identifying the real power consumption of the equipment and judging the state of the equipment, and particularly has the advantage of good universality, and can effectively identify the same equipment with different models and brands.
Drawings
FIG. 1 is a graph showing power waveforms of various loads in a day in a home;
FIG. 2 is a diagram of a noise reduction autoencoder architecture;
FIG. 3 shows the identification results of three devices in the home 1;
FIG. 4 is a graph of the actual energy consumption of the dishwasher in the household 1 for one day;
FIG. 5 is a graph comparing the load recognition accuracy of the method of the present invention and FHMM-based recognition method;
FIG. 6 is a comparison graph of the identification results of the home 2 devices from the network obtained by the home 1 data training;
fig. 7 is a diagram of recognition results of three devices in TraceBase;
FIG. 8 is a graph comparing the load recognition results of the desktop computers within the time period of 15000-25000 seconds.
Detailed Description
In order to facilitate a better understanding of the improvements of the present invention over the prior art, those skilled in the art will now make a further description of the present invention with reference to the accompanying drawings and examples.
In general, considering that a household has a plurality of electric devices turned on simultaneously in most of the time, and therefore the total active power of the household is formed by combining the sub-powers of the electric devices, what is needed in this embodiment is to extract the power characteristics of the electric devices and separate them from the mixed total power, and this separation process can be regarded as a noise reduction process in image processing or speech recognition. Typical noise reduction processes include removing noise from old pictures, or removing noise from a piece of sound, or even filling in an unclear portion of a picture. The essence of load identification is load separation, which can be regarded as the picture or sound recording that needs to be processed, and the power generated by other equipment without concern is regarded as "noise". The purpose of training a noise reduction auto encoder (DAE) is to separate a "clean" target signal from a noisy input. Input signal containing noise by DAE
Mapping to hidden proxies
Thereby constructing a decoded output vector z-gθ′(y) is carried out. The structure of the noise reduction auto-encoder is shown in fig. 2, in which,
signal obtained by adding random noise to the original input x, fθFor the encoder, y is the intermediate proxy after the encoding mapping, gθ′For the decoder, z is the reconstruction input, LH(x, z) is the reconstruction loss, which is used to measure the reconstruction error, and the parameters θ and θ' are trained to minimize the average reconstruction error during the training process, i.e., to make the output z as close as possible to the original input vector x, z, which is not "contaminated"
A deterministic function.
In the load separation stage, the sliding window is used to analyze the input mixed power signal y (t), and the length of the sliding window is determined by the usage time of the corresponding electrical equipment. For a mixed power obtained by simultaneously turning on a plurality of devices, the sliding windows must overlap. A common solution is to reconstruct this overlapping window using the average of the overlapping parts. This presents a problem in that when only a small portion of the on-time of a device is contained in the overlapping window, the result of load recognition is significantly higher than the actual power usage value, and the error of recognition increases further as the window slides. Here, the present embodiment uses a median filter to process the overlapped part, that is, the output signal of the overlapped part is the result of median filtering y (t). Specifically, because the power change of the overlapping window is relatively smooth, the output value of the overlapping window may be replaced by a statistical median of all values in a neighborhood, which is called a window (window), and the size of the window is determined according to the actual mixed power characteristic.
The following describes each step of load identification in this embodiment in detail.
First, load identification based on a noise reduction auto-encoder.
The problem of non-intrusive load identification in this embodiment can be represented by the following equation (1):
wherein, yiAnd (t) represents the electrical quantity value of a single electric device, and the electrical quantity value can be power, voltage or current, and can be regarded as an active power value without loss of generality. y (t) represents the total power consumption of the household, e (t) represents the total measurement error, wherein the measurement error is considered to be 0, and N represents the number of the electric equipment of the household. Thus, it can be seen from equation (1) that non-invasive load discriminationThe problem is identified that the power consumption value y of the single electrical equipment is obtained by using an algorithm under the condition that only the total load power y (t) is knowni(t) of (d). Further, load decomposition is translated into noise reduction problems as shown in the following equation:
y(t)＝yk(t)+ck(t),k＝1,2,...,N (2)；
wherein, ck(t) represents the sum of the powers of all the devices except device k, yk(t) represents the load k to be separated. Therefore, it is necessary to obtain the value of the active power consumed by the load (target load) k of interest, only c needs to be addedk(t) separating from the total load y (t) by the following steps:
1. coding the network:
1.1 the raw total input power values are processed by one or more one-dimensional convolutional layers to produce a set of signature maps.
1.2 each convolution layer sequentially passes through a linear activation function, a maximum pooling layer, an additional convolution layer and a pooling layer to finally form the fully-connected multilayer sensor.
1.3 the full link layer is processed by the modified linear unit (ReLU) activation function to end the whole encoding process.
2. Decoding the network:
2.1 upsampling the fully connected multi-layered perceptron by deconvolution.
2.2 the results from 2.1 were pooled (reverse of maximal pooling).
2.3 continue to up-sample the results in 2.2 by deconvolution.
And 2.4, obtaining the noise reduction signal reconstructed by decoding, and obtaining the power consumption information of the target equipment according to the noise reduction signal reconstructed by decoding.
Specifically, in step 1.2 above, neighboring maxima are obtained by a max pooling operation, thereby making the activation function position within the analysis window more independent, and in addition, the length of the feature map and the number of full link layer elements can be reduced. And comparing the input with zero by a modified linear unit (ReLU) activation function, and outputting a larger value, thereby avoiding negative values of the load power after decomposition. The objective of the above-mentioned noise-reducing autoencoder training network is to minimize the Mean Square Error (MSE) between the output and the activation function of the device to be separated, and to perform training parameter optimization using a random gradient descent (SGD) method. It is noted that unlike the conventional DAE, which requires artificial addition of noise data to input data, only the device power of a non-target object needs to be taken as noise in the present embodiment. Therefore, the non-intrusive load identification method adopted in the embodiment is not equivalent to the traditional picture or sound noise reduction, but takes the noise reduction as a training standard and better learns how to extract useful features, so as to better construct a high-level proxy.
And II, testing an example.
1. Selection of a data set.
At present, there are a plurality of opening source data sets for non-invasive load identification research at home and abroad, and the following are commonly used:
(1) a REDD data set. It is known as Reference Energy differentiation Dataset developed by massachusetts university j.kolter and m.johnson, the first Dataset for NILM studies. The REDD data set provides high frequency data with a sampling frequency of 15kHz and low frequency data with sampling frequencies of 0.5Hz and 1 Hz. In this case, electricity consumption data of 10 households, 119 days, 268 devices, and 1T were recorded. Fig. 1 is an example of a REDD data set, which shows the power consumption of each device in a day in a home. The REDD data set can be processed by excel, the operation is convenient, and the data downloading website is as follows:http：//redd.csail.mit,edu.
(2) TraceBase dataset. The TraceBase data set was developed by university of Darmstadt a.reinhardt et al, germany, monitoring and recording over ten homes and offices, 31 different types of equipment, 122 equipment, 1270 load electricity usage data. Table 1 shows the power consumption of a certain microwave oven during a period of time, the left entry is time, and the two rightmost numbers respectively represent the average active power consumption in1 second and 8 second periods. The TraceBase data set is also stored in an excel table form, the format of data items is as follows, and the data downloading website is as follows: http: // www.TraceBase.org.
05/01/201220：47：36；1373；478
05/01/201220：47：37；1378；648
05/01/201220：47：38；1382；994
05/01/201220：47：40；1378；1149
05/01/201220：47：41；1378；1301
The noise reduction automatic encoder network mentioned above is trained through actual measurement data of two data sets of REDD and TraceBase, and a test result is compared with an identification method of a hidden Markov model (FHMM) based on factors. All codes were in Python language and data was analyzed using NILMTK and Pandas tools. The neural network training environment is Win10 family edition, Intel i5-10210U processor, 8G memory and NVIDIA GeForce MX110 display card.
2. And (6) testing indexes.
The evaluation of the test results is divided into two aspects: the accuracy of energy decomposition and the correctness of equipment state judgment. The evaluation indicators in terms of energy decomposition are: degree of truth, accuracy and F1Indices, respectively using
Pi (E)And
and (4) showing. The specific calculation formulas of the first two indexes are shown in formulas (4) and (5):
wherein the content of the first and second substances,
representing the separated energy signal, yi(T) represents the true energy consumption of the device and T represents the total number of samples. To analyze the overall performance of the two methods, a comparison can be made by calculating the average degree of realism and accuracy of all the devices, as follows:
R(E)and P(E)Respectively, the averages found by considering the trueness and the accuracy of the load resolution of all the devices are represented, and the overall performance of the non-intrusive load identification scheme is reflected. While
The measurement index is a geometric mean of the trueness and accuracy, calculated as follows:
in addition, in this embodiment, a standard error NEP of load identification is further defined, which is used to represent a sum of deviations between the equipment energy consumption obtained after identification and the standard energy consumption, and the sum of the deviations is normalized by the total real equipment energy consumption, and a calculation formula of the sum is:
the determination of the device state refers to the determination of the device on/off state, and may be specifically divided into four indicators, True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN). The specific definitions of the four indices are as follows:
in formulae (10) to (13), si(t) and
respectively representing the real and recognized states of the device i at time t, and on and off representing the two states of the device "on" and "off". The degree of truth and accuracy of recognition based on the state of the device is defined as:
similarly, the truth and accuracy of all equipment state judgment and identification are considered, and indexes are obtained:
to obtain a decision based on the state of the device
Index:
in addition, the embodiment further utilizes the mausus correlation coefficient as an identification accuracy index, which is defined as:
the overall mazis correlation coefficient is:
the MCC has a value between-1, the identification is more accurate when the value is larger, and the random prediction identification is indicated when the value is 0.
3. And (6) testing results.
3.1REDD dataset.
In the REDD data set, family 1 and family 2 data are selected as test objects. The data is updated every 3 seconds, thus containing a total of 28800 pieces of data during the day. In order to verify the effectiveness of the load identification method in the present embodiment, the load decomposition effects of 10 electrical devices in the family 1 and 8 electrical devices in the family 2 are tested and compared. Wherein, the 10 kinds of electrical equipment of the family 1 are respectively an oven, a refrigerator, a dish washer, a disinfection cabinet, a lamp, a dryer, a microwave oven, a bath heater, an electric heater and a stove; the 8 kinds of electric appliances of the family 2 are a kitchen appliance 1, a kitchen appliance 2, a lamp, a stove, a microwave oven, a dryer, a refrigerator and a dish washer respectively. In the data training process, considering that the devices may show different power waveforms in different time periods, 10 days of data are selected for training for each device, and the 10 days of data are tested and verified, so that 576000 pieces of data are needed in total. In the REDD data set, the power consumption data of all 10 electrical devices exceeds 60 ten thousand.
For simplicity of presentation, only the power split results of three appliances in the household 1, respectively a dishwasher, a refrigerator and a lamp, are shown. As shown in fig. 3, the abscissa is time in seconds, and since it is desired to better observe the load identification effect of the method involved in this embodiment and the FHMM identification method, only the power waveform of the device on time period is intercepted, the abscissa time only lasts 6000 seconds, i.e. 2000 data points. In fig. 3, the top waveform represents the actual power curve of the load, the middle waveform represents the load identification result of the method involved in this embodiment, and the bottom waveform represents the load identification result of the FHMM identification method.
It should be noted that, common household electrical appliances can be divided into three categories from the operation state: single state class, continuous change class and multi-state class. The single state type means that the equipment has only one stable state after being started, and the power is generally kept unchanged, such as a lamp, a kettle, a microwave oven lamp and the like; the continuous change class refers to a process that the power of the equipment is continuously increased/decreased in the process of turning on/off, such as a television (power change is 50W-75W), a computer (80W-100W) and the like; the multi-state class refers to a device having multiple power states during operation, such as a refrigerator, a washing machine, a dishwasher, a dryer, and the like. In these three types of electrical devices, the identification of the single-state class and the continuously-changing class is relatively simple, while the multi-state class is easily confused with other devices due to the fact that the power of the multi-state class is greatly different in different state stages. As can be seen from fig. 3, for the lamps belonging to the single-state class, the identification effect of the two methods is good, the on and off states of the device can be well reflected, and the power consumption value is more accurately determined. For dishwashers and refrigerators belonging to the multi-state category, the load identification effect of the method is better, and the method is specifically represented in two aspects: 1. the real power consumption value of the equipment can be accurately resolved; 2. the different state stages of the equipment can be more accurately judged, so that the probability of misjudgment is reduced.
FIG. 4 is a diagram of the dishwasher in the household 1 during a day, the usage time is in the interval of 10000-12000 seconds, the interval is enlarged and the identification results of the two algorithms are compared, as shown in FIG. 5. It can be clearly seen from the figure that the method related by the invention only has a little jitter in the high-power operation state, the jitter error is not more than 5%, and the switching process between the fitting states can be well realized. The FHMM-based identification method is generally seen to be high in power identification, with a magnitude of approximately 20%, and cannot accurately represent the load handover process.
Table 1 below compares the 4 indicators of the two methods, the four indicators being defined and illustrated in 3.1, which are respectively indicative of the accuracy of the energy consumption identification
(the larger the better), the accuracy of the judgment of the state of the apparatus is shown
(the larger the better), NEP indicating the deviation of the power recognition result from the actual value (the smaller the better), and the mazis correlation coefficient MCC indicating the accuracy of the state judgment (the closer to 1 the better). In order to simplify the expression, the table only lists the index comparison conditions of 5 devices, and as can be seen from the table, all indexes obtained by the method disclosed by the invention are superior to those obtained by the FHMM method, the rightmost side of the table lists the overall expression value, and the black bold shows that the indexes have better expression.
TABLE 1REDD data set several equipment identification effect index comparisons
Due to the fact that household appliances are various in models, power utilization behaviors of different models of appliances can be different. In order to test the universality of the method, the data of the family 1, the family 3 and the family 4 are trained, and the network obtained by training decomposes the aggregate power of the family 2. Fig. 6 shows the result of identifying each device in the family 2 after the network training using the data of the family 1, and only the comparison results of three devices, respectively, a stove, a microwave oven and a sterilizer, are shown for the sake of simplicity. In the figure, the top waveform represents the actual power curve of the load, the middle waveform represents the load recognition result of the present invention, and the bottom waveform represents the load recognition result based on the FHMM method. It can be seen from the figure that both methods provide better equipment identification for single-state microwave ovens and cabinets, while the present invention is clearly superior to the FHMM method for fires with multiple states.
3.2TraceBase dataset.
The TraceBase data set contains 31 different types of equipment, 122 equipment, 1270 load electricity consumption data, and the data acquisition interval is 1-2 seconds. In this embodiment, 20 devices are identified by two methods, and the identification results of the television, the desktop computer and the electric iron are selected for display, as shown in fig. 7. Similarly, the top waveform in the figure represents the actual power curve of the load, the middle waveform represents the load recognition result of the present invention, and the bottom waveform represents the load recognition result of the FHMM method. It can be seen from the figure that the present invention has significant advantages in both identifying the real device power consumption and determining the different stages at which the device is located.
Fig. 8 compares the recognition effect of the two methods in the time period of 15000 seconds to 25000 seconds for the desktop computer. It can be seen from the figure that the jitter error of the method of the present invention is not more than 4%, and the switching process between the bonding states can be well performed, but the FHMM identification method is inaccurate in judgment at the moment of starting and stopping the load, and the total identified load power is higher. Table 2 compares 4 indices of the two methods, and it can be seen from the table that all the indices obtained by the present invention are superior to FHMM method, and the right-most side of the table lists the total expression values.
TABLE 2 comparison of several equipment identification performance indicators for TraceBase data sets
In summary, the present embodiment provides a non-invasive load identification method relying on only conventional sampling rate single active power measurement, which is based on the principle of a noise automatic encoder, and considers the total mixed power as a picture or a sound recording to be processed, and the power generated by other devices without concern as "noise" or "noise", and the execution target is changed to identify the load power of the interested single device from the total mixed power.
In an example test, the REDD and TraceBase data sets are used for comparing load identification effects of the provided method and the FHMM method (the power identification and the state judgment are two aspects, namely four specific indexes), and a test result shows that the method has obvious advantages in the aspects of identifying the power consumption of real equipment and judging the state of the equipment, and the method has the advantage of better universality and can effectively identify the same equipment with different models and brands.
It should be noted that, in the above embodiment, the above-mentioned method for identifying a home load based on a noise reduction automatic encoder may be implemented by a computer program, where the computer program is stored in a computer or a server connected with a data input device, and when the program runs, the data input device obtains active power measurement data of a target home within a set time period, and executes the foregoing steps 1.1 to 2.4. Furthermore, a household load identification system based on the noise reduction automatic encoder can be further designed, the system comprises a data input device, a server and a user side, wherein the server runs the household load identification program, the data input device is connected with the server and transmits the obtained active power measurement data to the server running the household load identification program, the user side is connected with the server through a communication network, and the server sends the identified power consumption information of the electrical equipment to the corresponding user side through the communication network. It should be understood by those skilled in the art that the number of the user terminals is not limited, and the system may configure the user terminals with different network address codes/device codes for different households, so that the server may accurately send the obtained power consumption information of the electrical equipment to the user terminal of the corresponding household.
The above embodiments are preferred implementations of the present invention, and the present invention can be implemented in other ways without departing from the spirit of the present invention.
Some of the drawings and descriptions of the present invention have been simplified to facilitate the understanding of the improvements over the prior art by those skilled in the art, and some other elements have been omitted from this document for the sake of clarity, and it should be appreciated by those skilled in the art that such omitted elements may also constitute the subject matter of the present invention.
Claims (7)
1. The household load identification method based on the noise reduction automatic encoder is characterized by comprising the following steps: taking the power of other equipment outside the target equipment as noise points or noises, extracting the power characteristics of the target equipment through a machine learning training noise reduction automatic encoder, analyzing an input mixed power signal by adopting a sliding window in a load separation stage, processing an overlapped part of the sliding window by using a median filter, replacing an output value of the overlapped window by a statistical median of all values in a neighborhood, and carrying out load separation according to the following steps:
firstly, encoding a network:
1. processing the raw total input power consumption data by one or more one-dimensional convolution layers to generate a set of feature maps;
2. each convolution layer sequentially passes through a linear activation function, a maximum pooling layer, an additional convolution layer and a pooling layer to form a fully-connected multilayer sensor;
3. the full connection layer is processed by the modified linear unit activation function, and the whole encoding process is finished;
secondly, decoding the network:
4. performing up-sampling on the fully-connected multilayer perceptron through deconvolution;
5. performing pooling on the result in the step 4, wherein the pooling is an inverse process of maximum pooling;
6. performing up-sampling on the result in the step 5 through deconvolution;
7. and obtaining a noise reduction signal reconstructed by decoding, and obtaining power consumption information of the target equipment according to the noise reduction signal reconstructed by decoding.
2. A home load recognition method according to claim 1, wherein: in step 2, neighboring maxima are obtained by maximum pooling, making the activation function positions within the analysis window more independent and reducing the length of the feature map and the number of full link layer elements.
3. A home load recognition method according to claim 1, wherein: in step 3, the modified linear unit activation function compares the input with zero and outputs a larger value.
4. A home load recognition method according to claim 1, wherein: when the noise reduction automatic encoder is trained, the mean square error between the minimized output and the activation function of the target equipment is taken as a target, and a random gradient descent method is adopted to optimize the training parameters.
5. Household load identification program based on noise reduction automatic encoder, its characterized in that: storing the data in a computer or a server connected with a data input device, and executing the steps 1 to 7 of the household load identification method according to any one of claims 1 to 4, wherein when the program runs, the data input device is used for acquiring active power measurement data of a target household within a set time period.
6. Domestic load identification system based on automatic encoder of making an uproar falls, its characterized in that: a server running the home load identification program of claim 7, and a data input device, the data input device being connected to the server and transmitting the obtained active power measurement data to the server running the home load identification program.
7. A home load recognition system according to claim 6, wherein: the power consumption information of the electrical equipment is identified by the server, and the power consumption information of the electrical equipment is transmitted to the corresponding user side through the communication network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110927020.3A CN113687182A (en) | 2021-08-12 | 2021-08-12 | Household load identification method, program and system based on noise reduction automatic encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110927020.3A CN113687182A (en) | 2021-08-12 | 2021-08-12 | Household load identification method, program and system based on noise reduction automatic encoder |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113687182A true CN113687182A (en) | 2021-11-23 |
Family
ID=78579628
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110927020.3A Pending CN113687182A (en) | 2021-08-12 | 2021-08-12 | Household load identification method, program and system based on noise reduction automatic encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113687182A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114511058A (en) * | 2022-01-27 | 2022-05-17 | 国网江苏省电力有限公司泰州供电分公司 | Load element construction method and device for power consumer portrait |
-
2021
- 2021-08-12 CN CN202110927020.3A patent/CN113687182A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114511058A (en) * | 2022-01-27 | 2022-05-17 | 国网江苏省电力有限公司泰州供电分公司 | Load element construction method and device for power consumer portrait |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110376457B (en) | Non-invasive load monitoring method and device based on semi-supervised learning algorithm | |
| CN109387712B (en) | Non-invasive load detection and decomposition method based on state matrix decision tree | |
| US11067482B2 (en) | Appliance operation signal processing system and method | |
| US9104189B2 (en) | Methods and apparatuses for monitoring energy consumption and related operations | |
| US20150377935A1 (en) | Signal identification methods and systems | |
| US20180096439A1 (en) | Utility consumption signal processing system and a method of processing a utility consumption signal | |
| CN111382789B (en) | Power load identification method and system based on machine learning | |
| CN112215406B (en) | Non-invasive resident electricity load decomposition method based on time convolution neural network | |
| CN110991263B (en) | Non-invasive load identification method and system for resisting background load interference | |
| WO2010129414A1 (en) | System and method for power utility usage, monitoring and management | |
| Zhang et al. | Power transformer fault diagnosis considering data imbalance and data set fusion | |
| CN110416995B (en) | Non-invasive load decomposition method and device | |
| CN111639586A (en) | Non-invasive load identification model construction method, load identification method and system | |
| US20220120586A1 (en) | System and method for identifying appliances under recall | |
| Anderson | Non-intrusive load monitoring: Disaggregation of energy by unsupervised power consumption clustering | |
| CN113687182A (en) | Household load identification method, program and system based on noise reduction automatic encoder | |
| Sanabria-Villamizar et al. | Characterization of household-consumption load profiles in the time and frequency domain | |
| Cimen et al. | Smart-building applications: Deep learning-based, real-time load monitoring | |
| CN114037178A (en) | Non-invasive load decomposition method based on unsupervised pre-training neural network | |
| Khazaei et al. | Evaluation of low-complexity supervised and unsupervised NILM methods and pre-processing for detection of multistate white goods | |
| Devlin et al. | Load identification and classification of activities of daily living using residential smart meter data | |
| Bucci et al. | A non-intrusive load identification system based on frequency response analysis | |
| Dinesh et al. | Non-intrusive load monitoring based on low frequency active power measurements | |
| CN117039840A (en) | Feature automatic enhancement-based household load identification method and system | |
| Massidda et al. | Non-intrusive load disaggregation via a fully convolutional neural network: improving the accuracy on unseen household |
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 |