CN111382789B - Power load identification method and system based on machine learning - Google Patents
Power load identification method and system based on machine learning Download PDFInfo
- Publication number
- CN111382789B CN111382789B CN202010152529.0A CN202010152529A CN111382789B CN 111382789 B CN111382789 B CN 111382789B CN 202010152529 A CN202010152529 A CN 202010152529A CN 111382789 B CN111382789 B CN 111382789B
- Authority
- CN
- China
- Prior art keywords
- data
- power load
- power
- training
- electrical parameter
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 238000010801 machine learning Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 60
- 238000005070 sampling Methods 0.000 claims description 45
- 238000012360 testing method Methods 0.000 claims description 40
- 238000012795 verification Methods 0.000 claims description 34
- 238000004140 cleaning Methods 0.000 claims description 25
- 239000012634 fragment Substances 0.000 claims description 22
- 230000009467 reduction Effects 0.000 claims description 19
- 238000013527 convolutional neural network Methods 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 14
- 238000010606 normalization Methods 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 6
- 230000000694 effects Effects 0.000 claims description 5
- 230000002159 abnormal effect Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000005265 energy consumption Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000000926 separation method Methods 0.000 claims description 3
- 238000011056 performance test Methods 0.000 claims description 2
- 238000010200 validation analysis Methods 0.000 claims 2
- 238000000605 extraction Methods 0.000 abstract description 6
- 238000010183 spectrum analysis Methods 0.000 abstract description 4
- 230000006698 induction Effects 0.000 abstract description 3
- 238000013528 artificial neural network Methods 0.000 description 6
- 238000012544 monitoring process Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000005259 measurement Methods 0.000 description 4
- 238000003062 neural network model Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000007781 pre-processing Methods 0.000 description 3
- 230000001052 transient effect Effects 0.000 description 3
- 238000005406 washing Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000009434 installation Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000008676 import Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011900 installation process Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
Abstract
The power load identification method and the system based on machine learning provided by the application are characterized in that the basis of actually measured electrical parameter data including current, voltage, power and the like is in a unified format, and on the basis of long-time extraction, collection, analysis, induction and training of power load characteristics, the type of an electrical appliance in use can be accurately identified under the condition that a plurality of pieces of power load overall basis electrical parameter data for a period of time including voltage, current, active power, reactive power and the like are known. Therefore, the machine learning model training method and system for power load identification provided by the application do not need to manually adjust parameters, and compared with the traditional method, such as time domain waveform matching, feature point matching, spectrum analysis and the like, the method and system have high matching accuracy, can automatically learn and automatically acquire the feature parameters required by power load identification, thereby improving the application range of the model and improving the accuracy of power load identification.
Description
Technical Field
The application relates to the technical field of power load detection, in particular to a power load identification method and system based on machine learning.
Background
The power load is characterized by the law that the active power and reactive power extracted by the power load from the power supply of the power system change along with the voltage of the load endpoint and the change of the system frequency; the power load characteristics are an important component of the power system; the identification of the electric equipment through the power load characteristics plays an important role in the development of smart grid technology.
The most common methods of power load identification are invasive and non-invasive identification methods. The method can directly obtain the measurement data of the load, but has the advantages of high installation cost, complex installation process and relatively difficult maintenance; while the non-invasive identification method only needs to install monitoring equipment at the total inlet of the power supply to decompose, monitor and identify each load in the whole system. Specifically, the non-invasive identification method is based on the extraction and identification of the electrical load imprinting characteristics; the electric appliance load imprinting characteristics can reflect unique information of electric equipment in operation, such as voltage, active waveforms, starting current and the like; these load signature features are repeated during operation of the device, and based on this, the consumer can be identified.
The design and extraction of the load imprinting features are the main difficulties of the whole method; feature designs typically employ relatively simple current, voltage, active and reactive power stability/transient features and combinations thereof. However, manually designed signal characteristics need to manually adjust parameters, the problems of low complexity and low dimensionality exist, meanwhile, the traditional matching algorithm such as time domain waveform matching, characteristic point matching, spectrum analysis and other methods are low in matching accuracy, and further the electrical load identification accuracy is low, so that the practical application effect is not ideal; and data modeling is a very necessary and important task. Aiming at electric equipment with stable running state, such as a television, an electric kettle, a computer and the like, the difficulty of load identification is relatively low, and if the working state is more, such as a full-automatic washing machine, the difficulty of load identification is also very high because of more power consumption condition changes in working. Aiming at the problems, the extraction technology based on steady-state characteristics cannot effectively cope with some scenes with higher recognition difficulty.
Disclosure of Invention
The application provides a power load identification method and a system based on machine learning, which are used for solving the technical problem that the accuracy of power load identification is low due to the fact that manually-designed signal characteristics in the existing method are required to manually adjust parameters.
In order to solve the technical problems, the embodiment of the application discloses the following technical scheme:
in a first aspect, an embodiment of the present application discloses a machine learning-based power load identification method, the method including:
acquiring a historical electrical parameter data set of each electrical appliance;
cleaning the historical electrical parameter data sets of the electrical appliances;
dividing the cleaned historical electrical parameter data set of the single electric appliance into an original training set, a verification set and a test set according to a proportion;
intercepting data fragments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data fragments;
respectively establishing a convolutional neural network model based on a noise reduction self-encoder for each target appliance;
training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set formed by data fragments of each target electric appliance to obtain an optimized model of each target electric appliance;
and collecting current data of the power load of the user, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance, and outputting a class result of the power load.
Optionally, the acquiring the historical electrical parameter data set of each electrical appliance includes:
integrating and summarizing the current public data set to obtain a first data set;
installing an electric energy meter at a general incoming line end of a user to obtain electric parameters of the total and single electric loads in one or more spaces to obtain a second data set;
and summarizing the first data set and the second data set to obtain historical electrical parameter data sets of all the electric appliances.
Optionally, the capturing the data segments of the original training set, the verification set and the test set along the time axis, and generating the training set, the verification set and the test set composed of the data segments, includes:
and intercepting the data fragments of the original training set, the verification set and the test set along the time axis direction by using a sliding window with the length of n and the moving step length of 1, and generating the training set, the verification set and the test set which are composed of the data fragments.
Optionally, said cleaning said historical electrical parameter data set of each appliance includes: unification of data formats, downsampling to a specified frequency, voltage normalization.
Optionally, the unifying the data formats includes:
according toConversion of active power to [0,1 ]]Numerical values in between, wherein:
S[i]representing the sampled value, i.e. instantaneous active power, C being the type of electrical load, sa being the sample data, s c Is the active power of the electrical load c.
Optionally, the downsampling to a specified frequency includes:
if the sampling rate is lower than 1Hz, recording according to the original sampling rate;
if the sampling rate is higher than 1Hz, the sampling rate is downsampled to 1Hz;
wherein downsampling the sampling rate to 1Hz comprises:
discarding all other sampling values within 1 second by using the values of the sampling points at intervals of 1 second;
calculating an average value of original sampling points in adjacent 1 second as a 1 second boundary data value;
the median value of the original sampling points within 1 second is calculated as the 1 second boundary data value.
Optionally, the voltage normalization includes:
according toNormalizing the voltages to the same fluctuation range, wherein:
Power normalised representing the normalized Power value, power representing the measured Power value, voltagenal representing the nominal voltage value Voltageobserved representing the measured voltage value.
Optionally, the cleaning the historical electrical parameter data set of each electrical appliance further includes detecting a gap, a normal operation time, and identifying a power load with an energy consumption rank of K, where K is an adjustable parameter.
Optionally, the training the convolutional neural network model based on the noise reduction self-encoder according to the training set, the verification set and the test set formed by the data segments for each target appliance to obtain an optimized model of each target appliance includes:
training parameters of the convolutional neural network model based on the noise reduction self-encoder by using a training set formed by target electrical appliance data fragments;
verifying and testing different models obtained in different training stages on a verification set formed by data fragments until the effect is best used as a corresponding model of a target electric appliance;
and performing performance test on the corresponding model of the target electric appliance by using a test set formed by the data fragments until the performance is optimal to obtain an optimized model of the target electric appliance.
In a second aspect, the present application provides a machine learning-based power load identification method, and the present application further provides a machine learning-based power load identification system, where the system includes:
the data set acquisition module is used for acquiring historical electrical parameter data sets of all the electrical appliances;
the data set cleaning module is used for cleaning the historical electrical parameter data sets of all the electrical appliances;
the data set dividing module is used for dividing the cleaned historical electrical parameter data set of the single electric appliance into an original training set, a verification set and a test set according to the proportion;
the data set segment intercepting module is used for intercepting data segments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data segments;
the model building module is used for building a convolutional neural network model based on the noise reduction self-encoder for each target electric appliance respectively;
the model optimization module is used for training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set which are formed by the data fragments for each target electric appliance to obtain an optimization model of each target electric appliance;
the power load type output module is used for collecting current data of the power load of the user, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance and outputting a type result of the power load.
Compared with the prior art, the application has the beneficial effects that:
as can be seen from the above technical solutions, the machine learning-based power load identification method and system provided in the present embodiment specifically uses measured electrical parameter data including current, voltage, active power, reactive power, and the like as a basis, and unifies basic electrical parameter data, so that on the basis of long-time extraction, collection, analysis, induction and training of power load characteristics, the type of an electrical appliance being used can be correctly identified under the condition that a plurality of pieces of power load overall basic electrical parameter data including voltage, current, active power, reactive power, and the like are known for a period of time. Therefore, the machine learning model training method and system for power load identification provided by the application do not need to manually adjust parameters, and compared with the traditional method, such as time domain waveform matching, feature point matching, spectrum analysis and the like, the method and system have the advantages that the accuracy of matching is high, the method and system can automatically learn and automatically acquire the feature parameters required by power load identification, so that the application range of the model is improved, and the accuracy of power load identification is improved.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application as claimed.
Drawings
In order to more clearly illustrate the technical solution of the present application, the drawings that are needed in the embodiments will be briefly described below, and it will be obvious to those skilled in the art that other drawings can be obtained from these drawings without inventive effort.
FIG. 1 is a schematic flow chart of a power load identification method based on machine learning according to an embodiment of the present application;
FIG. 2 is a schematic structural diagram of a power load identification system based on machine learning according to an embodiment of the present application;
fig. 3 is a schematic diagram of a DSP single-phase multifunctional table for monitoring overall basic electrical parameter data according to an embodiment of the present application.
Detailed Description
In order to make the technical solution of the present application better understood by those skilled in the art, the technical solution of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, but not all embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present application without making any inventive effort, shall fall within the scope of the present application.
The power load imprinting features can reflect unique information reflecting the power utilization state, such as voltage, waveform of active power, current and the like, of one electric equipment in operation, and the load imprinting can repeatedly appear in the operation process of the equipment, so that each electric equipment can be identified.
In a first aspect, an embodiment of the present application discloses a machine learning-based power load identification method, as shown in fig. 1, where the method includes:
s110: a historical electrical parameter dataset of each appliance is obtained.
The acquiring the historical electrical parameter data set of each electrical appliance comprises the following steps:
integrating and summarizing the current public data set to obtain a first data set;
installing an electric energy meter at a general incoming line end of a user to obtain electric parameters of the total and single electric loads in one or more spaces to obtain a second data set;
and summarizing the first data set and the second data set to obtain historical electrical parameter data sets of all the electric appliances.
In the embodiment of the application, besides summarizing the current public data set, the electric energy meter is installed at the general incoming line end of the user to acquire the number of electric parameters.
Since in the transient and steady state signals of the total load obtained by data measurement there are several cases of measurement errors: firstly, the inconsistency of the measuring devices, namely, for the same electric equipment, different measuring devices have different measured values; secondly, the sensor can cause data loss in the process of compressing and transmitting the original data. Because data acquisition and transmission can cause data deviation or loss, the data is necessary to be processed to improve the noise resistance of the load identification method; thirdly, the influence of the period of data sampling on load identification is researched, and the balance point of the data sampling overhead and the system modeling complexity is discussed.
The collected data set is a record of energy consumption, various electric appliances in a room or space are monitored by using a plurality of sets of monitoring instruments and equipment within a certain time range, and a low-frequency data collection mode and a high-frequency data collection mode are adopted. The acquisition frequency of the low frequency acquisition mode is 1Hz, and the acquisition frequency of the high frequency can reach 10kHz to 100kHz. The low-frequency signal is mainly extracted aiming at the load steady-state characteristic, and the high-frequency signal can obtain the load transient characteristic and the high-frequency harmonic characteristic. In general, the high-frequency signal can include more load electricity utilization characteristics, which is beneficial to training of a model and improvement of accuracy, but also puts higher demands on data acquisition, transmission, compression and processing capacity, and improves the complexity of the system. In the whole research process, the complexity and the accuracy of the system need to be selected and changed according to actual conditions.
In addition, we use the single-phase multifunctional electric energy meter of DSP to monitor the overall basic electric parameter data (including voltage, current, active power, reactive power, etc.), as shown in fig. 3, fig. 3 is a schematic diagram of the single-phase multifunctional electric energy meter of DSP provided by the embodiment of the application for monitoring the overall basic electric parameter data in fig. 3, the meter adopts an RS485 remote link to monitor the control panel, the data acquisition is inquired once per minute and the real-time acquisition server of the real-time link data is linked, the device has the advantages that: DIN35MM guide rail installation, has the characteristic of convenient assembly and disassembly; the communication rate can reach 9600bps, and the transmission rate is high; six paths of switching value input and output are adopted, so that the requirements on the input and output of measurement data are met; the DSP chip can be reconstructed and developed according to actual requirements, and the requirements of experimental environments are met.
Such as: we use DSP single phaseThe multifunctional electric energy meter collects overall basic electrical parameter data (including voltage, current, active power, reactive power and the like), and sampling data in the same parameter are arranged in time sequence. Taking active power as an example, p i Representing the degree of voltage detected by the electric energy meter at the ith sampling moment, and arranging the originally acquired data into an active power sequence P according to time sequence active ＝{p 1 ,p 2 ,…p i …, forms the output of the electrical load characteristic data sampling module, and at the same time serves as the input of the electrical load data cleaning module.
S120: the historical electrical parameter data set of each appliance is cleaned.
After the information of the power load data is obtained, the power load data cleaning is a core technology of the system and the method, and mainly comprises automatic screening and abnormal data cleaning, noise identification and separation, downsampling, discarding rate, data normalization, missing data compensation and processing, top-k, error data elimination and the like.
The system needs to create a data CSV file, and the work of cleaning up the data set is completed by deleting incomplete data (partial instrument equipment data have incomplete data or lost data due to different time stamps). After creating the data set and generating the CSV file import, the data will reside in our memory data structure, which can be used throughout the training process (disk save or load data). To address the problem of non-uniformity of data formats from different data sets, several preprocessing tasks are required.
The DSP single-phase multifunctional electric energy meter is used for collecting overall basic electric parameter data (including voltage, current, active power, reactive power and the like), and the sampling data in the same parameter are arranged in time sequence. Taking active power as an example, p i Representing the degree of voltage detected by the electric energy meter at the ith sampling moment, and arranging the originally acquired data into an active power sequence P according to time sequence active ＝{p 1 ,p 2 ,…p i …, forming the output of the power load characteristic data sampling module as power load dataAnd (3) cleaning the input of the module.
In the data cleaning module, the electrical time sequence data is downsampled to the designated frequency f according to the actual application requirement, and the downsampled active power data sequence is P 'under the assumption that the frequency f corresponds to an interval period of 5 sampling points in the original sampled data' active ＝{p 1 ,p 6 ,p 11 ,…p 5i+1 ,…}＝{q 1 ,q 2 ,…q i …. And performing other preprocessing operations such as normalization on the obtained relatively low-frequency time sequence data to generate continuous time sequence data which can be directly used as training by the model. Because the power load characteristic recognition model receives a plurality of electrical parameter inputs at the same time, but the input dimension of a specific electrical parameter is limited and is of a fixed length n, all continuous time sequence data of the electrical parameter cannot be input at one time, so a sliding window (sliding window) mode with the length n is adopted to continuously slide in a step length l, and a data subsequence with the fixed length n in the electrical parameter sequence, such as an active power subsequence P, is intercepted after each sliding i ＝{q in+1 ,q in+2 ,…q in+n And the power load characteristic identification module is used as an input of the power load characteristic identification module.
The method comprises the steps that different electrical appliances are provided with independent corresponding deep neural network models in a load characteristic identification module, data subsequences can be simultaneously input into the deep neural network models corresponding to the different electrical appliances, n sampling values of each electrical parameter in the same time period correspond to n corresponding nodes of a deep neural network input layer, and an mth target electrical appliance working state sequence S in the data subsequences is output through forward propagation operation of the network m,i ＝{s m,in+1 ,s m,in+2 ,…s m,in+n }. The state sequence length is n, but not limited to n, and is determined by the model architecture. All the kinds of electric appliances in the sequence can be judged based on the working state of each electric appliance.
The method specifically comprises the following steps:
(1) Unified format
Since the original data set formats are not uniform, the method needs to extract the characteristics of each data set for evaluation, so as to avoid the consumption of power due to different electric appliancesThe rate difference is too large, so that great interference is generated to judgment, data are required to be cleaned, and the operation is normalized, namely conversion into [0,1 ]]Values between S [ i ]]Representing the sampled value, i.e. instantaneous active power, C being the type of electrical load, sa being the sample data, s c For the active power of the electrical load c, the formula is as follows:
(2) Downsampling
The sampling rate of the device monitor is between 0.008Hz and 16kHz in the data set so the system will downsample the data set to a specified frequency using an aggregate function such as average, mode and median.
If the sampling rate is lower than 1Hz, recording is carried out according to the original sampling rate.
If the sampling rate is higher than 1Hz, the sampling rate is downsampled to 1Hz. The specific downsampling method comprises the following steps:
1. discarding all other sampling values within 1 second by using the values of the sampling points at intervals of 1 second;
2. calculating an average value of original sampling points in adjacent 1 second as a 1 second boundary data value;
3. calculating the median value of original sampling points in adjacent 1 second as a 1 second boundary data value;
(3) Voltage normalization
Due to fluctuations in the acquired voltages, for example, the same dataset shows voltages varying from 180-250V, while the other dataset shows voltages varying in the range of 118-123V. The present system must take into account the effects of these voltage fluctuations, as they can significantly affect power consumption.
According toNormalizing the voltages to the same fluctuation range, wherein:
Power normalised representing normalized Power values, power representing measured Power values, voltakenal representing nominal voltageThe value Voltageobserved represents the measured voltage value.
(4)Top-k
In general, our identification system targets the top k-bit (where k is an adjustable parameter) high-energy devices instead of all devices, because there are three points where, first, the top k-bit power consuming devices can already provide most of the reference information for the overall power consumption situation; second, these devices have the most prominent features, and the rest of the devices can be considered to generate noise only; third, modeling and identification for larger duty cycle power consuming devices can greatly improve the reliability of the data.
During the data cleansing process, the present system also solves other common problems of the data set, such as: the device sensor does not report readings, small data loss, removal of outliers such as observed voltages exceeding twice the rated voltage, loss of mains data, etc.
(5) Detecting gaps
Many algorithms today assume that the communication of each data acquisition device is continuous, however, in practice, it may happen that the data acquisition device is disconnected or fails, and if we set a parameter value, then a gap may be considered to exist in one continuous power data sample when the time of disconnection or failure is greater than the set parameter value. For example, we calculate the difference between the time stamps of adjacent samples and consider that there is a gap in the dataset if it is greater than a certain parameter, e.g., 10 seconds. For data sequences with gaps, the data sequences can not be directly used for a system training set and a test set, and all training data and test data sequences must be selected from data sequences with gaps in the middle.
(6) Normal run time
The uptime is the total time recorded by the sensor. It is the last timestamp, minus the first timestamp, minus the duration resulting from all existing gaps.
S130: the cleaned historical electrical parameter data set of the single electrical appliance is divided into an original training set, a verification set and a test set according to proportion.
S140: and intercepting the data fragments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data fragments.
For example, in the data cleaning module, we downsampled the electrical timing data to the specified frequency f according to the actual application requirement, and supposing that the frequency f corresponds to an interval period of 5 sampling points in the original sampled data, the downsampled active power data sequence is P' active ＝{p 1 ,p 6 ,p 11 ,…p 5i+1 ,…}＝{q 1 ,q 2 ,…q i …. And performing other preprocessing operations such as normalization on the obtained relatively low-frequency time sequence data to generate continuous time sequence data which can be directly used as training by the model. Because the power load characteristic recognition model receives a plurality of electrical parameter inputs at the same time, but the input dimension of a specific electrical parameter is limited and is of a fixed length n, all continuous time sequence data of the electrical parameter cannot be input at one time, so a sliding window (sliding window) mode with the length n is adopted to continuously slide in a step length l, and a data subsequence with the fixed length n in the electrical parameter sequence, such as an active power subsequence P, is intercepted after each sliding i ＝{q in+1 ,q in+2 ,…q in+n And the power load characteristic identification module is used as an input of the power load characteristic identification module.
S150: and respectively establishing a convolutional neural network model based on the noise reduction self-encoder for each target appliance.
In the embodiment of the application, through the designed power load characteristic identification module, the electric appliance in use can be identified by monitoring the power load data of the bus under the condition of unknown power load characteristics of the branch lines.
The design of the power load characteristic recognition module is based on a deep neural network, the characteristic parameters required by recognizing the power load can be automatically obtained without complex characteristic parameters artificially designed, and in the patent, data acquisition and data cleaning are the preparation work for intelligent power distribution and utilization power load characteristic recognition, and the power load characteristic recognition module is the key technology of the patent.
The power load decomposition is considered as a "noise reduction" task, wherein the power of the target appliance is equivalent to a "clean" main signal, while the power generated by other appliances is background "noise", and the task is to separate the power of the target appliance from the background "noise" power.
Deep neural networks are built based on the de-noising AutoEncoder, DAE concept, with convolutional layers for the first and last layer networks, with their position, scaling and warp invariance, to extract useful low-level features, such as the 1000 watt step change feature, which is likely to be a useful feature wherever the power segment occurs.
The architecture and detailed parameters of the convolutional neural network model based on the noise reduction self-encoder are shown in the following table:
architecture and parameters of convolutional neural network model based on noise reduction self-encoder
In the power load characteristic recognition module, for example, learning rate, batch size, a counter-propagation optimization model, a dropout function, selection of an activation function and the like, different parameter combinations bring different machine learning effects, and the system supports different parameter combination optimization, so that the type of electric equipment can be accurately recognized in the power load characteristic recognition module without artificially designing characteristic parameters of the power load.
S160: and training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set formed by the data fragments for each target electric appliance to obtain an optimized model of each target electric appliance.
S170: and collecting current data of the power load of the user, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance, and outputting a class result of the power load.
The method comprises the steps that different electrical appliances are provided with independent corresponding deep neural network models in a load characteristic identification module, data subsequences can be simultaneously input into the deep neural network models corresponding to the different electrical appliances, n sampling values of each electrical parameter in the same time period correspond to n corresponding nodes of a deep neural network input layer, and an mth target electrical appliance working state sequence S in the data subsequences is output through forward propagation operation of the network m,i ＝{s m,in+1 ,s m,in+2 ,…s m,in+n }. The state sequence length is n, but not limited to n, and is determined by the model architecture. All the kinds of electric appliances in the sequence can be judged based on the working state of each electric appliance.
We input into the power load identification module the basic electrical parameter data (including current, voltage, active power, reactive power) processed by the aforementioned power load acquisition and cleaning module as input data, with the objective of deducing the device class of each consumer from the electrical data.
The data and identification method are specifically as follows:
selecting data base electrical parameter data (including current, voltage, active power and reactive power) of a total circuit;
selecting the data from 1 minute to 5 minutes, taking a sampling rate of 1Hz as an example, and 60 to 300 sampling points are provided;
processing the data into a plurality of sequences with 16 data points as one sequence, wherein sequences with gaps are eliminated;
loading a machine learning model program for power load identification trained in advance, which can be realized by any existing machine learning or neural network programming framework, such as tensor flow, keras and the like;
the data sequence is input into a machine learning model, and the identification module outputs current device class data (expressed in integers, such as 1=air conditioner, 2=water kettle, 3=washing machine, 4=electric lamp, etc.) corresponding to each time sampling point.
The basic electrical parameter data processed by the power load acquisition and cleaning module comprise current, voltage, active power and reactive power as input data, and the aim is to infer the equipment category of each electric equipment through the electrical data.
The data and identification method are specifically as follows:
selecting data base electrical parameter data of the total circuit, such as current, voltage, active power and reactive power;
selecting the data from 1 minute to 5 minutes, taking a sampling rate of 1Hz as an example, and 60 to 300 sampling points are provided;
processing the data into a plurality of sequences with 16 data points as one sequence, wherein sequences with gaps are eliminated;
loading a machine learning model program which is trained in advance and used for power load identification can be realized by any existing machine learning or neural network programming framework, such as tensor flow, keras and the like;
the data sequence is input into a machine learning model, and the identification module outputs current device class data (expressed in integers, such as 1=air conditioner, 2=water kettle, 3=washing machine, 4=electric lamp, etc.) corresponding to each time sampling point.
The power load identification method and the system based on machine learning provided by the embodiment specifically take measured electrical parameter data including current, voltage, active power, reactive power and the like as the basis, unify the basic electrical parameter data into a format, and can accurately identify the type of an electrical appliance in use under the condition that a plurality of pieces of power load overall basic electrical parameter data for a period of time including voltage, current, active power, reactive power and the like are known on the basis of long-time power load feature extraction, collection, analysis, induction and training. Therefore, the machine learning model training method and system for power load identification provided by the application do not need to manually adjust parameters, and compared with the traditional method, such as time domain waveform matching, feature point matching, spectrum analysis and the like, the method and system have the advantages that the accuracy of matching is high, the method and system can automatically learn and automatically acquire the feature parameters required by power load identification, so that the application range of the model is improved, and the accuracy of power load identification is improved.
In a second aspect, the present application further provides a power load identification system based on machine learning, as shown in fig. 2, where the system includes:
the data set acquisition module is used for acquiring historical electrical parameter data sets of all the electrical appliances;
the data set cleaning module is used for cleaning the historical electrical parameter data sets of all the electrical appliances;
the data set dividing module is used for dividing the cleaned historical electrical parameter data set of the single electric appliance into an original training set, a verification set and a test set according to the proportion;
the data set segment intercepting module is used for intercepting data segments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data segments;
the model building module is used for building a convolutional neural network model based on the noise reduction self-encoder for each target electric appliance respectively;
the model optimization module is used for training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set which are formed by the data fragments for each target electric appliance to obtain an optimization model of each target electric appliance;
the power load type output module is used for collecting current data of the power load of the user, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance and outputting a type result of the power load.
Since the foregoing embodiments are all described in other modes by reference to the above, the same parts are provided between different embodiments, and the same and similar parts are provided between the embodiments in the present specification. And will not be described in detail herein.
Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure of the application herein. This application is intended to cover any variations, uses, or adaptations of the application following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the application pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.
The embodiments of the present application described above do not limit the scope of the present application.
Claims (8)
1. A machine learning-based power load identification method, the method comprising:
acquiring a historical electrical parameter data set of each electrical appliance;
cleaning the historical electrical parameter data sets of the electrical appliances;
dividing the cleaned historical electrical parameter data set of the single electric appliance into an original training set, a verification set and a test set according to a proportion;
intercepting data fragments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data fragments;
respectively establishing a convolutional neural network model based on a noise reduction self-encoder for each target appliance;
training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set formed by data fragments of each target electric appliance to obtain an optimized model of each target electric appliance;
collecting current data of a user power load, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance, and outputting a class result of the power load;
said cleaning said historical electrical parameter data set of individual appliances includes: automatic screening and abnormal data cleaning, noise identification and separation, downsampling, discarding rate, data normalization, missing data compensation and processing, top-k and error data rejection are realized;
the acquiring the historical electrical parameter data set of each electrical appliance comprises the following steps:
integrating and summarizing the current public data set to obtain a first data set;
installing an electric energy meter at a general incoming line end of a user to obtain electric parameters of the total and single electric loads in one or more spaces to obtain a second data set;
and summarizing the first data set and the second data set to obtain historical electrical parameter data sets of all the electric appliances.
2. The machine learning based power load identification method of claim 1, wherein said intercepting data segments along a time axis for the original training set, validation set and test set generates a training set, validation set and test set comprised of data segments, comprising:
and intercepting the data fragments of the original training set, the verification set and the test set along the time axis direction by using a sliding window with the length of n and the moving step length of 1, and generating the training set, the verification set and the test set which are composed of the data fragments.
3. The machine learning based power load identification method of claim 1, further comprising unification of data formats:
according toConversion of active power to [0,1 ]]Numerical values in between, wherein:
S[i]representing the sampled value, i.e., instantaneous active power, C is the electrical load type,in order to obtain the sample data,s c is the active power of the electrical load c.
4. The machine learning based power load identification method of claim 1, further comprising downsampling to a specified frequency:
if the sampling rate is lower than 1Hz, recording according to the original sampling rate;
if the sampling rate is higher than 1Hz, the sampling rate is downsampled to 1Hz;
wherein downsampling the sampling rate to 1Hz comprises:
discarding all other sampling values within 1 second by using the values of the sampling points at intervals of 1 second;
calculating an average value of original sampling points in adjacent 1 second as a 1 second boundary data value;
the median value of the original sampling points within 1 second is calculated as the 1 second boundary data value.
5. The machine learning based power load identification method of claim 1, further comprising voltage normalization:
according to Power=/>Power normalizes the voltage to the same fluctuation range, wherein:
Powerrepresenting normalized Power value, power representing measured Power value, voltage nominal Representing a nominal Voltage value Voltage observed Representing the measured voltage value.
6. The machine learning based power load identification method of claim 1, wherein said cleaning said historical electrical parameter data set of each appliance further comprises detecting gaps, normal run times, and identifying power loads with energy consumption ranked top K bits, where K is an adjustable parameter.
7. The machine learning-based power load identification method according to claim 1, wherein the training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set formed by data segments for each target appliance to obtain an optimized model of each target appliance comprises:
training parameters of a convolutional neural network model based on the noise reduction self-encoder by using a training set consisting of target electrical appliance data fragments;
verifying and testing different models obtained in different training stages on a verification set formed by data fragments until the effect is best used as a corresponding model of a target electric appliance;
and performing performance test on the corresponding model of the target electric appliance by using a test set formed by the data fragments until the performance is optimal to obtain an optimized model of the target electric appliance.
8. A machine learning based power load identification system, the system comprising:
the data set acquisition module is used for acquiring historical electrical parameter data sets of all the electrical appliances; the acquiring the historical electrical parameter data set of each electrical appliance comprises the following steps: integrating and summarizing the current public data set to obtain a first data set; installing an electric energy meter at a general incoming line end of a user to obtain electric parameters of the total and single electric loads in one or more spaces to obtain a second data set; summarizing according to the first data set and the second data set to obtain historical electrical parameter data sets of all the electric appliances;
the data set cleaning module is used for cleaning the historical electrical parameter data sets of all the electrical appliances; said cleaning said historical electrical parameter data set of individual appliances includes: automatic screening and abnormal data cleaning, noise identification and separation, downsampling, discarding rate, data normalization, missing data compensation and processing, top-k and error data rejection are realized;
the data set dividing module is used for dividing the cleaned historical electrical parameter data set of the single electric appliance into an original training set, a verification set and a test set according to the proportion;
the data set segment intercepting module is used for intercepting data segments of the original training set, the verification set and the test set along a time axis to generate the training set, the verification set and the test set which are composed of the data segments;
the model building module is used for building a convolutional neural network model based on the noise reduction self-encoder for each target electric appliance respectively;
the model optimization module is used for training the convolutional neural network model based on the noise reduction self-encoder according to a training set, a verification set and a test set which are formed by the data fragments for each target electric appliance to obtain an optimization model of each target electric appliance;
the power load type output module is used for collecting current data of the power load of the user, inputting the current data into an optimization model of each target electric appliance, separating out the working state of the electric appliance and outputting a type result of the power load.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010152529.0A CN111382789B (en) | 2020-03-06 | 2020-03-06 | Power load identification method and system based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010152529.0A CN111382789B (en) | 2020-03-06 | 2020-03-06 | Power load identification method and system based on machine learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111382789A CN111382789A (en) | 2020-07-07 |
| CN111382789B true CN111382789B (en) | 2023-11-14 |
Family
ID=71217025
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010152529.0A Active CN111382789B (en) | 2020-03-06 | 2020-03-06 | Power load identification method and system based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111382789B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111917114B (en) * | 2020-08-20 | 2022-02-08 | 国网湖南省电力有限公司 | Power load event detection method suitable for embedded platform |
| WO2022141330A1 (en) * | 2020-12-31 | 2022-07-07 | Guizhou Power Grid Company Limited | A non-intrusive load identification method based on fingerprint characteristics of load power |
| CN112801115B (en) * | 2021-01-26 | 2022-09-06 | 广西电网有限责任公司电力科学研究院 | Power supply fluctuation control method based on microgrid source-load comprehensive characteristic image and application |
| SE2150281A1 (en) * | 2021-03-11 | 2022-09-12 | Ctek Sweden Ab | Method for detecting performance deterioration of components |
| CN113033633B (en) * | 2021-03-12 | 2022-12-09 | 贵州电网有限责任公司 | Equipment type identification method combining power fingerprint knowledge and neural network |
| CN113191600A (en) * | 2021-04-13 | 2021-07-30 | 清科优能(深圳)技术有限公司 | Intelligent house non-invasive load identification method based on data mining |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108416695A (en) * | 2018-02-24 | 2018-08-17 | 合肥工业大学 | Electric load probability density prediction technique based on deep learning and system, medium |
| CN108616120A (en) * | 2018-04-28 | 2018-10-02 | 西安理工大学 | A kind of non-intrusive electrical load decomposition method based on RBF neural |
| CN109116100A (en) * | 2018-07-09 | 2019-01-01 | 清华大学 | It is a kind of based on coding-decoding structure electric load electricity consumption decomposition method |
| CN109145949A (en) * | 2018-07-19 | 2019-01-04 | 山东师范大学 | Non-intrusive electrical load monitoring and decomposition method and system based on integrated study |
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN110536237A (en) * | 2019-09-04 | 2019-12-03 | 国网四川省电力公司电力科学研究院 | Location information acquisition method based on UWB |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100305889A1 (en) * | 2009-05-27 | 2010-12-02 | General Electric Company | Non-intrusive appliance load identification using cascaded cognitive learning |
-
2020
- 2020-03-06 CN CN202010152529.0A patent/CN111382789B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108416695A (en) * | 2018-02-24 | 2018-08-17 | 合肥工业大学 | Electric load probability density prediction technique based on deep learning and system, medium |
| CN108616120A (en) * | 2018-04-28 | 2018-10-02 | 西安理工大学 | A kind of non-intrusive electrical load decomposition method based on RBF neural |
| CN109116100A (en) * | 2018-07-09 | 2019-01-01 | 清华大学 | It is a kind of based on coding-decoding structure electric load electricity consumption decomposition method |
| CN109145949A (en) * | 2018-07-19 | 2019-01-04 | 山东师范大学 | Non-intrusive electrical load monitoring and decomposition method and system based on integrated study |
| CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
| CN110536237A (en) * | 2019-09-04 | 2019-12-03 | 国网四川省电力公司电力科学研究院 | Location information acquisition method based on UWB |
Non-Patent Citations (2)
| Title |
|---|
| 基于大数据Bayes分类的家电设备识别算法;晋远等;《建筑科学》;20170415;第33卷(第04期);全文 * |
| 胡忠义著."实验结果与分析".《基于计算智能技术的电力负荷预测理论及应用》.2019,58-60. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111382789A (en) | 2020-07-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111382789B (en) | Power load identification method and system based on machine learning | |
| CN111242391B (en) | Machine learning model training method and system for power load identification | |
| CN110376457B (en) | Non-invasive load monitoring method and device based on semi-supervised learning algorithm | |
| Rueda et al. | Wavelet-based analysis of power system low-frequency electromechanical oscillations | |
| Garcia-Perez et al. | Fully-convolutional denoising auto-encoders for NILM in large non-residential buildings | |
| JP6455431B2 (en) | Monitoring device, monitoring method and program | |
| CN108021736A (en) | A kind of load switching action monitoring method based on sliding window Remanent Model | |
| Yu et al. | Nonintrusive appliance load monitoring for smart homes: Recent advances and future issues | |
| US20130338948A1 (en) | Methods and system for nonintrusive load monitoring | |
| Luo et al. | Monitoring HVAC equipment electrical loads from a centralized location-methods and field test results | |
| Andrean et al. | A hybrid method of cascade-filtering and committee decision mechanism for non-intrusive load monitoring | |
| CN105143892A (en) | Power entry point energy measuring apparatus and energy measurement information labeling system using same | |
| GB2476246A (en) | Diagnosing an operation mode of a machine | |
| CN114580666A (en) | Multi-mode fusion motor intelligent maintenance system | |
| CN112215406A (en) | Non-invasive type residential electricity load decomposition method based on time convolution neural network | |
| CN103018611B (en) | Non-invasive load monitoring method and system based on current decomposition | |
| CN112765149A (en) | System and method for calculating capacity of energy storage system | |
| Matthews et al. | Automatically disaggregating the total electrical load in residential buildings: a profile of the required solution | |
| CN102799108B (en) | Energy consumption depolymerization method based on low frequency energy consumption information and home energy management system | |
| CN112365090A (en) | Deep learning-based non-invasive electrical load identification method and device | |
| CN112615428A (en) | Line loss analysis and treatment system and method | |
| Azizi et al. | Quantification of disaggregation difficulty with respect to the number of smart meters | |
| CN115825602A (en) | Load identification method and system for comprehensive multivariate information similarity analysis | |
| CN111199014B (en) | Time sequence based seq2point NILM method and device | |
| CN115133659A (en) | On-line monitoring method and system for low-voltage line loss of low-voltage distribution network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |