A method for synthesizing electrical appliance fingerprint aggregation data

By normalizing the common data sets and aggregating them, the multi-concurrent generator load signal set is generated, which solves the problem of lack of unified data sets in non-invasive load monitoring, improves the training effect of the model and reduces the data collection cost.

CN113887629BActive Publication Date: 2024-11-22SHANGHAI MENGXIANG INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111166996.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-01
Publication Date
2024-11-22
Estimated Expiration
2041-10-01

AI Technical Summary

Technical Problem

The existing non-invasive load monitoring technology lacks a unified electrical fingerprint aggregation data set, resulting in low model robustness, high cost of load data collection in real scenarios, and useless measurements such as zero load conditions.

Method used

By filtering existing public data sets, normalizing processing, generating unified and standardized basic data sets, and filtering different types of electrical appliances for fingerprint data aggregation, generating multi-concurrent generator load signal sets, and dividing training, verification and test sets , to meet the data needs of non-invasive load monitoring tasks.

Benefits of technology

The generated aggregated data set has the training effect of the model that is almost the same as the real measurement data set, which reduces the cost of load data collection in real scenarios and reduces the occurrence of useless measurements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113887629B_ABST
    Figure CN113887629B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of non-invasive load monitoring, and specifically is a method for synthesizing electrical appliance fingerprint aggregate data. The present invention obtains a unified and standardized basic data set by screening existing public data sets, and then performs normalization processing, and screens different types of electrical appliances to aggregate fingerprint data to generate multiple parallel electrical appliance load signal sets. The method generates different signal representations through different electrical appliance combinations, and finally divides the training, verification and test sets to meet the data requirements of non-invasive load monitoring (non-invasive load monitoring) tasks. The generated aggregated data set has a training effect on the model that is almost the same as the real measurement data set, reducing the cost of load data collection in real scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of non-intrusive load monitoring technology, specifically a method for synthesizing aggregated electrical fingerprint data. Background Art

[0002] Electricity meters can only report the total electricity load data of a household. Therefore, inferring the electricity load of specific devices and decomposing the electricity load are the main research goals at present. How to use deep learning methods to improve the accuracy of load decomposition is a current research hotspot. However, non-intrusive load monitoring uses datasets that sometimes only monitor the load of a single appliance, and sometimes aggregated information with strong distribution (e.g., concentrated electric kettles, refrigerators, etc.). Modeling methods for these datasets are targeted but have low robustness, and do not distinguish appliances into a unified category. For example, the same type of appliance, such as computer 1 and computer 2, will show significant differences in most load conditions.

[0003] Currently, there is a lack of a rich set of appliance fingerprint aggregation datasets to drive the progress of non-intrusive load monitoring tasks. Furthermore, the cost of collecting load data in real-world scenarios is very high, and in many cases, a lot of useless measurements are generated, such as zero-load situations when equipment is not running. Therefore, how to achieve rich aggregation data of various appliance types at a lower cost is one of the main problems at present. Summary of the Invention

[0004] The purpose of this invention is to provide a method for synthesizing aggregated electrical appliance fingerprint data, in order to solve the problems of lack of unified standardization and small maximum concurrent load in existing public datasets in the field of non-intrusive load monitoring.

[0005] The method for aggregating synthetic electrical appliance fingerprint data provided by this invention obtains a unified and standardized basic dataset by screening existing public datasets and performing normalization processing. Then, different types of electrical appliances are selected from this dataset for fingerprint data aggregation, generating a multi-parallel generator load signal set. Different combinations of electrical appliances produce different signal representations. Finally, training, validation, and test sets are divided to meet the data requirements of non-intrusive load monitoring tasks. The generated aggregated dataset provides training performance for the model that is almost identical to that of the real measurement dataset, reducing the cost of collecting load data in real-world scenarios (which sometimes generates many useless measurements, such as in zero-load situations), thus solving the difficulties existing in the aforementioned techniques.

[0006] The method for synthesizing electronic fingerprint aggregation data provided by this invention includes the following specific steps:

[0007] Step 1: Select a public dataset as the initial data source;

[0008] Step 2: Normalize the public dataset to ensure that electrical load signals from different measurement environments can be merged;

[0009] Step 3: Determine the final appliance categories, merge different datasets, and obtain the basic dataset;

[0010] Step 4: Determine the maximum number of concurrent loads, set the limit on the number of different appliance combinations and the limit on the number of combinations represented by a single appliance;

[0011] Step 5: Synthesize the aggregated signal dataset and divide it into training, validation, and test sets.

[0012] The following provides further detailed explanations of each step.

[0013] Step 1: Select a public dataset as the initial data source.

[0014] First, it should be clarified that fingerprint recognition on electronic devices mainly consists of two aspects:

[0015] ①Frequency sampling: mainly manifested in the switching between steady-state and transient states of electrical appliances, as well as the signal waveform under continuous load conditions;

[0016] ②Low-frequency sampling: mainly manifested as root-mean-square (RMS) values, i.e., effective current, voltage and other related information. The electrical fingerprints related to the RMS are related to the current state of the equipment and human usage factors, and are difficult to observe. It is equivalent to a special case under high-frequency sampling.

[0017] The method for synthesizing aggregated appliance fingerprint data proposed in this invention mainly focuses on aggregating appliance fingerprint data from low-frequency sampling. It utilizes a normalized signature synthesizer to synthesize the desired aggregated data, as shown below:

[0018] SNS: P→S w

[0019]

[0020] Where P represents the base dataset, S represents the aggregate dataset, and S... w The aggregated dataset is for a maximum concurrent load of w electrical appliances. Here, the low-frequency sampling mainly represents the root mean square (RMS) value, which is related to effective current and voltage. The electrical fingerprint related to the RMS is related to the current state of the device and human usage factors, and is difficult to observe. It is equivalent to a special case under high-frequency sampling. High-frequency sampling mainly represents the steady-state and transient state switching of electrical appliances, as well as the signal waveform under continuous load.

[0021] Step 2 involves normalizing the public dataset.

[0022] Step 2.1: Voltage control, discarding voltage total harmonic distortion exceeding the threshold t THD fingerprint data;

[0023] Harmonics refer to electrical quantities in a circuit whose frequencies are integer multiples of the fundamental frequency. Generally, they refer to the electrical quantities generated by circuits with frequencies higher than the fundamental frequency after performing Fourier series decomposition on periodic non-sinusoidal electrical quantities. High-order harmonic interference is a major "public nuisance" affecting power quality in current power systems. Aggregated data should be the fingerprint data of electrical appliances under normal operating conditions; therefore, a threshold t needs to be set. THD To ensure that fingerprint data with high total harmonic distortion (THD) (generally greater than 0.1) is not used. The calculation of the harmonic distortion value first obtains the Fourier transform result Sf of the voltage data, then calculates the sum of squares of Sf, subtracts the initial value Sf[0] of Sf and takes the square root, and finally divides by Sf[0] to obtain the distortion value.

[0024] Step 2.2: Extract the required waveform region. Calculate the root mean square (RMS) value of a sliding window of one period across the entire waveform of each data point. The RMS value is set when it exceeds the activation threshold t. on A binary mask is applied to the fingerprint data to facilitate subsequent extraction;

[0025] The activation threshold t on It is a value set manually, determined by examining the relationship between the current and the sampling frequency in relation to the overall data. The root mean square (RMS) is calculated by first dividing the frequency value of the waveform by the reference frequency of 50Hz to obtain the data length, and then dividing the square of the current value by the data length to obtain the RMS.

[0026] Step 2.3: Signal duration control, remove load signal with a duration less than t sec The data is retained for a duration that is rounded to an integer value to avoid spectral leakage in subsequent steps;

[0027] Step 2.4: Frequency normalization. Using the time-stretch method proposed by the librosa library (B. McFee, C. Raffel, D. Liang, D. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in python,” in Proceedings of the 14th Python in Science Conference, vol. 8, 2015.), the waveforms of current and voltage are converted into a reference frequency f. ref And the waveform is f0 / f refThe data is stretched by a factor to avoid fluctuations and matched with fingerprint data from different frequency grids (e.g., 50Hz in China and 60Hz in the United States), where f0 is the actual reference frequency.

[0028] Step 2.5: Downsampling, appropriately setting the frequency f of the final aggregated data. down The value of f is used to downsample the original waveform to frequency f. down To reduce memory usage and ensure that no significant information is lost;

[0029] Step 2.6: Scaling, amplifying the voltage waveform to the reference voltage V. ref The current waveform is also readjusted according to the above steps to maintain the original load power consumption.

[0030] Step 3 determines the final appliance type, merges different datasets, and obtains the basic dataset.

[0031] The public datasets ds, after normalization in step 2, have the same voltage and current waveforms. Since different datasets contain different types of electrical appliances, it is necessary to determine the final types of electrical appliances and set the appliance label set L according to the requirements. n ={l1, l2, ..., l n} indicates that the basic dataset P contains n different electrical appliances, and the merging process can be represented as:

[0032]

[0033] Step 4: Determine the maximum number of concurrent loads, and set limits on the number of different appliance combinations and the number of combinations represented by a single appliance.

[0034] Step 4.1: Determine the maximum number of concurrent load devices w, where w is the number of devices that can be connected during the duration time interval T. obs The number of device tags appearing within, and includes:

[0035]

[0036] Where |*| represents the base of *;

[0037] Step 4.2: S represents the aggregated dataset containing n types of electrical appliances. w Given an aggregated dataset with a maximum concurrent load of w electrical appliances; select w types of electrical appliances from n types for aggregation to obtain S. w The theoretical maximum possible number of data points is num. comb (w); Introduce a constraint The actual number of electrical appliance combinations is:

[0038]

[0039] Limitations under different w Setting basis:

[0040] ① Increase the number of combinations with smaller w (e.g., w < n * 0.3) to achieve a more accurate single electrical appliance fingerprint prediction effect of the model;

[0041] ② Reduce the number of combinations with larger w (e.g., w > n * 0.7) to improve the decomposition ability of the model;

[0042]

[0043] Step 4.3: Denote P l ∈ P(l ∈ L n ) and |P l | ≥ 1 as the basic data set with only electrical appliance l. Therefore, the number of each combination in the aggregated data set is expressed according to a single electrical appliance as:

[0044]

[0045] repr Similarly, for num

[0046]

[0047]

[0048]

[0049] n ), Pr(L n ) = U(1, L w ) to achieve a uniform distribution of different electrical appliances in the aggregated data set; train In summary, the number of unique aggregated signals obtained from a single L

[0050] ​​​​​​​​​​​​​​​​To eliminate the problem of missing fingerprint data unique to different datasets in P, the training set typically occupies more than half of the total dataset, while the test set accounts for a higher percentage than the validation set. S... train For a set S with different concurrent loads w The intervals are taken (e.g., if w = 10, take {2, 4, 6, 8, 10}).

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] This invention obtains a unified and standardized basic dataset by screening existing public datasets and then normalizing them. Different types of electrical appliances are then selected from this dataset to aggregate fingerprint data, generating a multi-parallel generator load signal set. Different combinations of electrical appliances can produce different signal representations. Finally, training, validation, and test sets are created to meet the data requirements of non-intrusive load monitoring tasks. The generated aggregated dataset provides training performance for the model that is nearly identical to that of real-world measurement datasets, reducing the cost of collecting load data in real-world scenarios (which sometimes generates many useless measurements, such as in zero-load situations). Attached Figure Description

[0055] Figure 1 This is a flowchart of the synthetic electrical fingerprint aggregation data method of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0057] Example:

[0058] This invention proposes a method for synthesizing aggregated fingerprint data of electrical appliances, the flowchart of which is shown below. Figure 1 As shown, it can be divided into the following steps:

[0059] Step 1: Select a public dataset as the initial data source;

[0060] Step 2: Normalize the public dataset to ensure that electrical load signals from different measurement environments can be merged;

[0061] Step 3: Determine the final appliance categories and merge different datasets to obtain the basic dataset;

[0062] Step 4: Determine the maximum number of concurrent loads, set the limit on the number of different appliance combinations and the limit on the number of combinations represented by a single appliance;

[0063] Step 5: Synthesize the aggregated signal dataset and divide it into training, validation, and test sets.

[0064] The following provides further detailed explanations of each step.

[0065] 1. Select a public dataset as the initial data source;

[0066] First, it should be clarified that fingerprint recognition on electronic devices mainly consists of two aspects:

[0067] ②Frequency sampling: mainly manifested in the switching between steady-state and transient states of electrical appliances, as well as the signal waveform under continuous load conditions;

[0068] ②Low-frequency sampling: mainly manifested as root-mean-square (RMS) values, i.e., effective current, voltage and other related information. The electrical fingerprints related to the RMS are related to the current state of the equipment and human usage factors, and are difficult to observe. It is equivalent to a special case under high-frequency sampling.

[0069] This paper proposes a method for aggregating electrical appliance fingerprint data, primarily targeting low-frequency sampled fingerprint data. It utilizes a normalized signature synthesizer to synthesize the desired aggregated data, represented as follows:

[0070] SNS: P→S w

[0071]

[0072] Where P represents the base dataset, S represents the aggregate dataset, and S... w The aggregated dataset is for a maximum concurrent load of w electrical appliances. Here, the low-frequency sampling mainly represents the root mean square (RMS) value, which is related to effective current and voltage. The electrical fingerprint related to the RMS is related to the current state of the device and human usage factors, and is difficult to observe. It is a special case under high-frequency sampling. High-frequency sampling mainly represents the steady-state and transient state switching of electrical appliances, as well as the signal waveform under continuous load.

[0073] When constructing P, a public dataset with RMS load signals is selected, and the dataset contains fingerprint data of multiple individual appliances.

[0074] 2. Normalize the public dataset;

[0075] In this example, WHITED(GAO,J.,GIRI,S.,KARA,EC,BERG,M.é 2014.PLAID:a public dataset of high-resoultion electrical appliance measurements for loadidentification research:demo abstract.Proceedings of the 1st ACM Conferenceon Embedded Systems for Energy-Efficient Buildings.Memphis,Tennessee:ACM.) and PLAID(KAHL,M.,HAQ,AU,KRIECHBAUMER,T.&JACOBSEN,H.-A.Whited-a worldwide household and industry transient energy data set.Workshop on Non-IntrusiveLoad Monitoring(NILM),2016Proceedings of the 3rd The datasets from International (2016) were normalized. The two datasets contain 16 and 55 categories of devices, respectively, with 1343 and 1876 available appliance fingerprints (one record for each operating device). Both datasets have 15 common appliance categories: soldering iron, refrigerator, laptop, washing machine, hair dryer, light bulb, heater, fan, refrigerator, coffee maker, kettle, water bottle, flashlight, microwave oven, vacuum cleaner, and air conditioner.

[0076] Step 2.1: Voltage control, total harmonic distortion of voltage exceeds threshold t THD Fingerprint data with a value of 0.1 will be discarded;

[0077] Step 2.2: Extract the required waveform region. Calculate the root mean square (RMS) value of a sliding window of one period across the entire waveform of each data point. The RMS value is set when it exceeds the activation threshold t. on A binary mask is applied to the fingerprint data with a value of 0.1 for subsequent extraction. The required waveform area is masked with 1, and the invalid part is masked with 0.

[0078] Step 2.3: Signal duration control, remove load signal with a duration less than t sec For data with a duration of 1 second (s), the duration of the data is rounded to an integer value to avoid spectral leakage in subsequent steps;

[0079] Step 2.4: Frequency normalization. Using the time-stretch method proposed by the librosa library, the waveforms of current and voltage are converted into a reference frequency f. ref = 50 Hz, and the waveform is at f0 / f ref The data is stretched by a factor to avoid fluctuations and matched with fingerprint data from different frequency grids (e.g., 50Hz in China and 60Hz in the United States), where f0 is the actual reference frequency.

[0080] Step 2.5: Downsampling, setting f appropriately down A value of 4000 Hz ensures that no significant information is lost, and the original waveform is downsampled to frequency f. down To reduce memory usage;

[0081] Step 2.6: Scaling, the voltage waveform is amplified to the reference voltage V. ref = 311 volts (V), and the current waveform is also readjusted according to the above steps to maintain the original load power consumption;

[0082] Four appliance fingerprints were discarded from the WHITD dataset, and 208 appliance fingerprints were discarded from the PLAID dataset.

[0083] 3. Determine the final types of appliances and merge different datasets to obtain the basic dataset;

[0084] The public datasets ds, after normalization in step 2, have the same voltage and current waveforms. Since different datasets contain different types of electrical appliances, it is necessary to determine the final types of electrical appliances and set the appliance label set L according to the requirements. n ={l1, l2, ..., l n} indicates that the basic dataset P contains n different electrical appliances, and the merging process can be represented as:

[0085]

[0086] 4. Determine the maximum number of concurrent loads, and set limits on the number of different appliance combinations and the number of combinations represented by a single appliance;

[0087] Step 4.1: Determine the maximum number of concurrent load devices w, where w is the number of devices that can be connected during the duration time interval T. obs = The number of device tags that appear within 5 seconds (s), and have:

[0088]

[0089] Where |*| represents the base of *.

[0090] Step 4.2: S represents the aggregated dataset containing n types of electrical appliances.w The aggregated dataset with the maximum number of concurrent load appliances being w. Therefore, selecting w appliances from n appliances for aggregation can obtain S w The theoretically maximum possible number of data is num comb (w). However, in many cases, num comb (w) will be very large. Therefore, a limit is introduced The actual number of appliance combinations is:

[0091]

[0092] For the limits under different w The setting basis:

[0093] ① Increase the number of combinations with smaller w (w < n * 0.3) to achieve a more accurate single-appliance fingerprint prediction effect of the model; ② Reduce the number of combinations with larger w (w > n * 0.7) to improve the decomposition ability of the model;

[0094] ② Reduce the number of combinations with larger w (w > n * 0.7) to improve the decomposition ability of the model; ② Reduce the number of combinations with larger w (w > n * 0.7) to improve the decomposition ability of the model;

[0095] Step 4.3: Denote P l ∈ P (l ∈ L n ), |P l | ≥ 1 as the basic dataset with only appliance l. Therefore, each combination number of the aggregated dataset can also be expressed according to a single appliance as:

[0096]

[0097] Similarly, for a limit is introduced In addition, to determine the specific combination number, a weight Ω needs to be introduced, where Pr(L n ) = U(1, L n ) to achieve an even distribution of appliances:

[0098]

[0099] So the number of unique aggregated signals obtained from a single L w is:

[0100]

[0101] For the entire dataset, repr(w) / comb(w) in different sets decreases at a ratio of two. In S train , repr / comb{3, 4, 6, 8, 10} = {12, 12, 6, 3, 3}, S valIn the context of `repr / comb{1~10}={-,2,2,2,1,1,1,1,1,1,1},S` test In the formula `repr / comb{1~10}={-,6,6,6,3,3,3,1,1,1}`, the maximum number of electrical load combinations is 1 when `w=1`, and only one electrical appliance can operate at the same time. Therefore, the number of different electrical appliance combinations is the same as the number of electrical appliances, and there will be no restrictions related to `repr`.

[0102] 5. Synthesize and aggregate the signal dataset, and divide it into training, validation, and test sets;

[0103] The aggregated signal dataset S obtained after processing the data according to the above steps has two statistical methods: one is the number of combinations (comb(w)) based on different appliance combinations, and the other is the number of combinations (repr(w)) based on a single appliance. S is then divided into a training set Si according to a certain percentage. train (60%), validation set S val (10%) and test set S test (30%), the purpose is to eliminate the problem of missing fingerprint data unique to different datasets in P, where S train is a set S of different concurrent loads w The intervals are taken (e.g., if w = 10, take {3, 4, 6, 8, 10}).

Claims

1. A method for synthesizing appliance fingerprint aggregate data, characterized in that: The specific steps are: Step 1: Select a public dataset as the initial data source A public data set with a load signal in the form of RMS is selected, and the data set contains fingerprint data of multiple single electrical appliances; Step 2 normalizes the public data set; the specific process includes: Step 2.1: Voltage control, discard voltage total harmonic distortion exceeding threshold t THD Fingerprint data; Harmonics refer to the amount of electricity in the circuit whose frequency is an integer multiple of the fundamental wave. The threshold value t is set. THD , ensuring that fingerprint data with high voltage total harmonic distortion will not be used; Step 2.2: Extract the required waveform area, calculate the RMS value of a sliding window of the size of a cycle on the entire waveform of each data, and activate it when the RMS value is higher than the activation threshold t on Apply a binary mask to the fingerprint data for subsequent extraction; Step 2.3: Signal duration control, delete the load signal duration less than t sec The duration of the retained data is rounded to an integer value to avoid spectral leakage in subsequent steps; Step 2.4: Frequency normalization: Use the time stretching method proposed by librosa library to convert the current and voltage waveforms to the reference frequency f ref , the waveform is f0 / f ref The multiples are stretched to avoid fluctuations and match the fingerprint data from power grids of different frequencies, where f0 is the actual reference frequency; Step 2.5: Downsampling, reasonably set the frequency f of the final aggregated data down The value of the original waveform is downsampled to the frequency f down To reduce memory usage and ensure no significant loss of information; Step 2.6: Scaling: amplify the voltage waveform to the reference voltage V ref , the current waveform is also readjusted according to the above steps to maintain the original load power consumption; Step 3: Determine the final type of electrical appliances, merge different data sets, and obtain the basic data set; Set the appliance label set L according to your needs n ={l1,l2,…,l n }, indicating that the basic data set P contains n different electrical appliances, and the merging process is expressed as: ds1, ds2, ..., are electrical appliances with labels l1, l2, ..., l n The electrical appliance data corresponding to n different electrical appliances; Step 4 determines the maximum number of concurrent loads, sets the number of different electrical combinations and the number of combinations of a single electrical appliance; the specific process is as follows: Step 4.1: Determine the maximum number of concurrent load appliances w, where w is the number of loads in the continuous time interval T. obs The number of device tags that appear in the , and have: Among them, |*| represents the cardinality of *; Step 4.2: S represents an aggregated dataset containing n types of electrical appliances. w is an aggregated data set with a maximum number of concurrent load appliances of w; w appliances are selected from n appliances for aggregation, and S w The theoretical maximum possible number of data is num comb (w); introduce a restriction The actual number of electrical appliance combinations is: Restrictions under different w The basis for setting: ① Increase the number of smaller w combinations To achieve a more accurate single appliance fingerprint prediction effect of the model; ②Reduce the number of combinations with larger w number to improve the decomposition ability of the model; The number of the smaller w combinations is such that w < n * 0.3; the number of the larger w combinations is such that w > n * 0.7; Step 4.3: Note P l ∈P(l∈L n ),|P l |≥1 is the basic data set with only appliance l, so the number of each combination of the aggregated data set is expressed according to a single appliance as follows: Similarly, for num repr (w), introduces a restriction In addition, in order to determine the specific number of combinations, the weight Ω is introduced: Among them, Pr(L n )=U(1,L n ) to achieve uniform distribution of different appliances in the aggregated dataset; In summary, from a single L w The number of unique aggregate signals obtained is: For the entire dataset, repr(w) / comb(w) decreases by a factor of two across different sets; Step 5: Synthesize the aggregate signal dataset and divide it into training, validation, and test sets The aggregate signal data set S obtained after processing the data according to the above steps has two ways of data statistics: one is the number of combinations comb(w) obtained by different electrical appliances, and the other is the number of combinations represented by a single electrical appliance repr(w); S is divided into training sets S according to a certain percentage train , validation set S val And the test set S test , to eliminate the unique fingerprint data missing problem between different data sets in P. The percentage is determined according to the goal of the network learning task; among them, the training set S train is a set S of different concurrent load numbers w The proceeds are taken at intervals.