KR20160141032A

KR20160141032A - Non-Intrusive Appliance Load Monitoring Method using a Switching Factorial Hidden Markov Model and System applying the same

Info

Publication number: KR20160141032A
Application number: KR1020150073706A
Authority: KR
Inventors: 윤정미
Original assignee: 전자부품연구원
Priority date: 2015-05-27
Filing date: 2015-05-27
Publication date: 2016-12-08

Abstract

A non-intrusive appliance load monitoring method using S-FHMM is provided. A non-intrusive appliance load monitoring method according to an embodiment of the present invention includes measuring supply power, detecting power change, and accurately estimating an appliance which causes a change in power detected based on the power usage features of the appliance. Thereby, the appliances which are in use can be estimated without measuring power usage for each device.

Description

TECHNICAL FIELD The present invention relates to a non-contact type load identification method and system using a S-FHMM,

The present invention relates to device load identification, and more particularly, to a non-contact device load identification method and system capable of estimating devices in use without measuring power usage data for each of the devices.

Conventional home appliance load identification is largely divided into non-intrusive appliance load monitoring and intrusive appliance load monitoring according to the data extraction method.

The contact type extracts the power usage data directly by connecting it to the household appliance, and extracts the total power amount data by using the smart meter located outside the house.

That is, as shown in FIG. 1, the non-contact type has one measurement point such as AMI (Advanced Metering Infrastructure) and the data is extracted from various measurement points, such as a meter embedded in a sub meter, a smart plug, .

The goal of individual household appliances load identification is to infer the state of each appliance by analyzing the extracted data and manage it efficiently. Particularly, there is a need to find a way to increase the accuracy in the identification of non-contact type home appliances.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a non-contact load identification method and system for estimating a device that caused a power change with reference to power usage characteristics of the devices.

According to an aspect of the present invention, there is provided a method for identifying a non-contact type load, comprising: measuring supply power; Detecting a power change; And estimating the device that caused the detected power change, with reference to power usage features of the devices.

And, the power usage features may include at least three power related components.

In addition, the at least three power-related components may be three power-related components having the largest eigenvalue of the selected covariance using SVD (Singular Value Decomposition) among the power-related components.

The estimating step may estimate the device that caused the power change through the stochastic k-nearest neighbor classification algorithm.

Further, the non-contact load identification method according to an embodiment of the present invention may further include the step of sensing the surrounding environment, and the estimating step may further include using sensing data generated in the sensing step, The device can be estimated.

According to another aspect of the present invention, there is provided a non-contact type load identification system comprising: a measurement unit for measuring supply power; A detection unit for detecting a power change; And an analyzer for estimating a device that caused the detected power change, with reference to the power usage characteristics of the devices.

As described above, according to embodiments of the present invention, it is possible to estimate a device that caused a power change with reference to power usage characteristics of devices, so that it is possible to avoid the cost burden of measuring the power use per device do.

In addition, according to the embodiments of the present invention, the accuracy of the non-contact type device load identification can be further improved by utilizing the sensing data.

Figure 1 shows the classification of IALM and NIALM,
FIG. 2 is a flowchart illustrating a non-contact type household appliances load identification process,
FIG. 3 is a drawing for deriving an individual usage amount in the total usage amount,
FIG. 4 illustrates a K-Nearest Neighbor (KNN)
FIG. 5 illustrates probabilistic K-Nearest Neighbor (PKNN)
Figure 6 is an example of an FHMM structure,
7 shows an example of a switching state-space model,
FIG. 8 is a diagram illustrating a Switching Factorial Hidden Markov Model (S-FHMM)
9 shows the components extracted from the AMI,
10 is a graph showing a result of three-dimensional reduction in principal component analysis,
Fig. 11 is a diagram showing a k-NN classification (k = 10)
FIG. 12 is a graph showing the relationship between training data and test data,
13 shows results of inference of three kinds of equipment using only active power,
Fig. 14 shows the result of inference using HMM only for effective power (HMM technology for five devices)
FIG. 15 is a graph showing the relationship between training data and test data using active power and sensor data,
16 shows results of inference of three kinds of equipment using active power and sensor data,
17 shows the result of inference using active power and sensor data (HMM technology for five devices)
18 shows only the comparison when using active power,
19 is a graph showing a comparison between the active power and the ambient light sensor data,
20 is a block diagram of an apparatus load identification system according to an embodiment of the present invention.

Hereinafter, the present invention will be described in detail with reference to the drawings.

1. Appliance Load monitoring

Non-contact individual household appliances By analyzing the changes in power supplied to the home (non-intrusive load monitoring, NILM or non-intrusive appliance load monitoring, NIALM), it is possible to deduce the energy consumption of household appliances or products. The basic processing procedure of the non-contact type individual household appliance load identification is shown in FIG.

As shown in FIG. 2, the digital AC monitor is connected to the home. The amount of power and current change is measured through the Admittance Measurement Unit and recorded in a Net Change Detector Unit after normalization by a Scaler. Finally, cluster analysis is used to classify each appliance in operation.

Non-contact individual appliances The load identification system senses the power usage of these individual appliances as a continuous change of individual appliances. As shown in FIG. 3, each of the devices has an ON / OFF state characteristic, and identifies the state of the device by searching for the characteristics of each individual home appliances hidden in the total power value using the feature.

2. Principal Component Analysis

Principal component analysis is a method of reducing the dimension by resetting the new main axis while maintaining the information of the feature with the great dimension. At this time, the feature vectors are relocated on the basis of the principal axis without correlation in order to represent each variable efficiently. In this method, the variable having the largest covariance between feature vectors is set as the main axis, and the next major axis is set as the next main axis. In order to perform such a process, the symmetric covariance matrix of each feature vector is decomposed into eigen-analysis or SVD (Singular Value Decomposition) to extract principal components. Specifically, SVD is used to extract d principal components in order of greatest eigenvalue of covariance (assuming that the number of principal components is d), and a low dimensional value projected as a new d dimensional space Can be obtained. This projected variation can improve the performance of data analysis by eliminating unnecessary dimension or eliminating noise signal.

3. k-Nearest Neighbor (k-Nearest Neighbor)

The k-Nearest Neighbor algorithm is a classification method that determines the target value of the variance based on the similarity with the training set. The classifier learned by the training set calculates the similarity between the new incoming variable and the existing training set. In this case, similarity between variables can be measured by various methods from Euclidean distance to mutual information. Based on this similarity, the closest k neighbor variables are selected and the target value is determined by a number of labels.

The existing kNN technique is a deterministic technique and it is difficult to consider the uncertainty existing in the data. For example, if k = 3, the number of TV neighbors is large. Therefore, it can be determined that the new data is a signal for the TV. However, as shown in FIG. 5, the probabilistic k- neighbor classifier calculates the probabilities of the neighbors, When it is high, it is a technology that makes judgment by light.

4. NIALM Related to modelling technique

HMM (Hidden Markov Model) is NIALM based on graphical model. The HMM technique is a technique for solving the problem by considering the state of the device as a hidden variable and considering the composite signals extracted from the AMI as Observation. One of these techniques, Factorial Hidden Markov Model (FHMM), directly affects Observation according to the state of each device, and each state has an efficient computation amount based on the Markovian Model, so real time is possible .

4.1 Factorial Hidden Markov Model (FHMM)

FHMM is a variant of HMM, which is one of the well known state space model inference methods, and it deduces Hidden value S, which is the state value of the discrete device through observation value Y which is continuous value. In particular, assume that the FHMM has a multiplex structure in which the number of states is extended in a basic HMM model, and the states of the individual entities are independent of each other. The transition probability (see equation (1) 1st order Markov chain), which makes efficient computation and modeling possible.

(One)

4.2 Switching state - space model

Another model is a model that combines HMM and state-space model and utilizes discrete state variable S. This implies that the transition probability of S follows the equation (2) using the transition matrix.

(Φ is transition matrix) (2)

This model has the problem that it provides more free model than FHMM, but it also increases the amount of computation obtained by increasing the hidden variable to be measured.

5. HMM-based appliance load identification

5.1 Separated Unit Identification (IALM)

First, to extract only three components from the 169 components extracted from the AMI, a process of eliminating unnecessary components is performed through principal component analysis, which is a dimension reduction algorithm. We then use the labeling of the training set to perform instrument identification with the KNN classifier.

5.2 Identification of Electric Appliances Based on Total Contactless Power NIALM )

The technique used in identification of household electric appliances based on total contactless electric power uses the same model as the multiple sensor based home appliance identification technology, so the details are described in Section 5.3.

5.3 Composite (sensor) based appliance identification (NIALM + sensors)

The performance of the multi-label sorter is evaluated by using the pre-information such as the data extracted from the sensor and the NIALM data of the first power, which is a model in which the FHMM and the switching state-space model are mixed. . 8 is a graphical model of the model, where O is a hidden value to be estimated,? Is an observation value to be used for inference,

Represents a variable that can be omitted.

The definition of each variable in FIG. 8 is as follows.

- W: Existing NIALM identification data

- C: Feature space is constructed by sensor value (temperature, humidity, etc.), voltage and frequency, and the result of applying PKNN classification technique

- S: Time series data indicating the state of each device. It can be a binary value of on / off or a value of 3 or more discrete values

- X: Parameter value obtained from training phase as power value for each device and state, and this variable itself may be omitted in mathematical calculation process

- Y: Actual power data observed

The above model has the following mathematical modeling. As can be seen from the equation below and FIG. 8, it is assumed that Y, W, and C, which affect the S value to be inferred, are observed values and are independent of each other. This proposed Switching Factorial HMM (S-FHMM) meets the requirements to derive the state of each device with several incomplete data.

(3)

(4)

(5)

Each of the functions inside this relation can be defined in detail as follows.

-

: Since it is not assumed to be a specific distribution, it is decided at the training stage

-

: Assuming a normal distribution with mean and device-specific variance for each state value (deriving the variance for each state value is assumed to be one variance because it greatly increases the complexity of the problem)

-

: power variable according to x value

The probability of the parameter is derived from the preprocessing step

-

: Enter the first inferred state and the accuracy value for the information

Looking at Eqs. (3), (4), and (5), we can see that there are some parameters that need to be resolved before the algorithm is implemented. First, the average value per state of the power consumption of K devices

And the standard deviation value

And a calculation of the standard deviation r taking into consideration the noise signal generated in the composite signal should be preceded. To this end, the characteristics of each device are learned in a training step, and each parameter is measured in advance, so that the device can be stably identified in a test step.

(6)

After the calculation of the model parameters through this preprocessing process, the algorithm performs the direct operation. To do this, we calculate the marginalized likelihood as follows to construct a new system that provides an efficient computation process for the model used in Figure 8 and (3) (4) (5). This marginal likelihood makes it possible to obtain the hidden value directly from the observed value without having to separately obtain the power value for each device. In other words, the expression of the state space based HMM is derived by FHMM.

(7)

Basically, the HMM is a Bayesian model and needs a Prior distribution. The model for this is as follows.

(8)

The W value used here is a value for utilizing the performance of the incomplete classifier obtained from the first-identified power information, and the C value is a parameter having a classifier result value using an ancillary sensor, which is a probabilistic K nearest neighbor (PKNN) . The formula using PKNN is as follows.

(9)

From PKNN

The process of extracting and tracing a feature is a method of inferring likelihood through comparison with trained data. The sensor can use roughness, temperature, and humidity sensor. When the illuminance sensor value is high, the Likelihood value is obtained for the nearest state of the devices operated such as a fluorescent light, and the electric heater and the like obtain the corresponding information by using the temperature and humidity sensor, which is used incidentally. For example, assuming that the power usage characteristics of the stand and the humidifier are similar, when the power change is detected and the illuminance is increased, the power change is caused by the stand, and when the power change is detected, It is assumed that it was caused by a humidifier.

5.4 Posterior value derivation method

In order to derive the posterior which represents the state value for each device, the filtering process and the predication process are repeated to derive the result.

The process of obtaining posterior is repeated from Filtering → Prediction → Filtering → Prediction. That is, assuming the initial value, Prediction value is obtained, and on the basis of this, Filtering and the next Prediction value are derived based on that.

In order to obtain the filtering and predication, we assume that the initial state value at t = 0 is off or uniform distribution. In this case, the filtering equation is decomposed on the assumption that W, C and Y are independent.

Is the Prediction value at t-1, and the Filtering step and the Prediction step are proceeded by the following equations.

5.4.1 Filtering step:

(10)

5.4.2 Prediction step :

(11)

6. Used data features and extraction

6.1 Data Description

When extracting data from a power signal, the higher the sampling frequency is, the more advantageous it is to identify the change in power consumption of each device. The data used in the above algorithm is a power signal extracted at 20 Hz, and the extracted feature is composed of 169 components in total. 169 components are calculated based on the basic active power (P), reactive power (Q), voltage (Urms), current (Irms)

6.2 Data used

6.2.1 Separate unit-to-unit identification (IALM)

In order to eliminate this problem, we use Principal Component Analysis, which is a dimensional reduction technique, to apply 169 dimensions to three-dimensional data To perform device identification.

6.2.2 Consumer identification based on the total amount of contactless power NIALM )

To make the same technology as the existing NIALM, the device identification for the composite signal utilizes only the active signal, which is a representative signal component, not 169 components.

6.2.3 Complex (sensor) based appliance power consumption identification accuracy ( NIALM + sensors )

Finally, in order to improve the performance of the existing NIALM by additionally using additional sensor data in the composite signal, the result of performing PKNN by adding sensor data to the active power component is synthesized.

7. Implementation Results

7.1 Separate Unit Identification (IALM)

The unit identification among the separated devices first extracted 169 components from the AMI. Then, we set up a new main axis that can express all feature information more efficiently by using Principal Component Analysis which is a dimension reduction algorithm. After the variables were projected on the newly set main axis, classification into KNN classifier was performed. For this purpose, data corresponding to audio, TV, and set-top box was sampled. In the case of set-top box, the data size was small and extracted twice. Fig. 10 shows the result of reducing the size of the 169 components to three main components. In the case of the set-top box, it can be confirmed that clustering is well performed even though it is not the data extracted at once. What this means is that the device of the same model can be effectively identified through the KNN classifier as in FIG. Currently, unit identification between single isolated devices has a very high probability accuracy of 99.20%.

7.2 Identification of Home Electric Appliances Based on Total Contactless Power NIALM )

7.2.1 On three devices PKNN apply

12 shows the effective power value with respect to time according to the change of the equipment state with time. In order to classify the state of the device using only the active power using the probabilistic K-neighborhood classifier selected in the previous section, assuming that all devices have On and Off states for all three devices, There are a number of situations 2 ³ , so we gather information in advance on this situation in the training step.

We use stochastic k-nearest neighbor classifiers to know the state of each device in the test data. This is inferred to be the largest number of adjacent k when the active power, which is a characteristic of 8 (= 2 ³ ) cases of training data, is compared with the effective power of the time to be searched through a stochastic comparison method. Here, k is 5, and other values can be used.

The inference results for each device are shown in FIG. First, if the state value of each equipment consists only of On and Off, the result of the algorithm will be a probability distribution for each state. In other words, P (TV) = [0.8, 0.2] shows the result as 0.8, and the probability that the TV is turned on is 0.2. That is, it can be said that P (TV = Off) = 0.8 and P (TV = On) = 0.2.

When measuring the accuracy, the result is replaced with a large probability value. For example, if P (TV) = [0.8, 0.2] is given as the result, the TV is off (Off) and the value is compared with the hidden real state value. If it is wrong, it is called. If you look at the results of the experiment, you will notice that the humidifier is not actually turned on because there is no additional sensor data, but the stand is turned on and the humidifier is mistaken for being turned on. However, the identification values for the three instruments showed fairly high accuracy. Exactly 200% of the time except 8 times, the accuracy is 96%. The results of adding sensor data (temperature, humidity) to improve accuracy can be found in 6.3.1.

7.2.2 In five different devices HMM apply

The HMM technique proposed in the embodiment of the present invention is presented as a model for linking the result obtained from the first-identified power identification algorithm with the information obtained from incidentally obtained sensor values. This problem is used to identify five devices. The input value is the active power signal value extracted from the AMI, and it distinguishes ON / OFF of 5 devices only with this signal value. The input value is the active power signal and the result is the probability distribution for each state of the five instruments. For example, P (audio) = [0.4, 0.6] gives the result and produces a result in a vector form. If the result is [0.4, 0.6], P (Audio = Off) = 0.4 and P (Audio = On) = 0.6. When the accuracy is measured later, the probability value of the two states is determined to be a measured value having a large value. In this case, P (Audio = On)> P (Audio = Off), so it can be said that the audio is on. These results are compared with the actual state values hidden and measured for accuracy.

As the number of devices increases, the performance of accuracy decreases. According to the result of analyzing virtual data having 10,000 samples randomly generated over 10 times using virtual data having similar values to real data, it is confirmed that the accuracy is about 70% for 5 devices , While the heaters were almost 100% accurate, while audio and light data with weak information showed frequent errors and the overall accuracy was estimated to be about 70% due to errors in the instrument.

Fig. 14 is a result of identification of real data for five devices by HMM technology. a) the active power signal, b) a separate set-top box, c) a separate lamp signal, d) a separate heater signal, e) a separate audio signal, f) a separate TV signal, If you look at a set-top box or a light bulb, you can see a value that is frequently shaken off.

7.3 Composite (Sensor) Based Consumer Power Consumption Identification Accuracy ( NIALM + sensors )

7.3.1 For three devices PKNN apply

FIG. 15 shows the values of the actual power, the humidity (blue line) and the temperature (red line) with respect to time according to the change of the equipment state with time. As in the previous experiment, we use probabilistic k-nearest neighbor classifiers to train the situation of all situations of the whole equipment in order to classify the state of equipment by using active power and sensor data in combination.

In the test data, we use stochastic k-nearest neighbor classification method to know the state of each equipment. This is inferred to be the largest number of adjacent k when comparing the active power, which is the characteristic of the eight cases of the training data, and the sensor data with the active power of the time to be searched with the sensor data through the stochastic comparison method. k was 5.

The inference results for each device can be seen to infer the real data closer to the real data than the inference results using the stochastic K-nearest neighbor classification technique using only the active power previously (see FIG. 16). In the entire 200 hour experiment, the accuracy is 99.5% except for one. Earlier, 3.5% accuracy was upgraded from the probability of 96% using only active power.

7.3.2 In five different devices HMM Technology

This experiment is the same except for the inference result of FIG. 16 using only the active power and one environment. One difference is that in this environment, ancillary sensor values are used. Humidity and temperature mainly affect the heaters. The heaters have good results without using the sensor data, so it is difficult to distinguish them by using the ambient light sensor rather than the humidity and temperature sensors. Attempt to identify a device with a small amount of power signal, such as a lamp. To do this, we learn what changes occur in each equipment when sensor data is applied. In the training phase, we obtain the values and change the input values for the state of each equipment whenever sensor data generated in event format is received. This leads to a change in the function.

FIG. 17 is a view showing each of devices by HMM technology with real data obtained from five devices, signals obtained from temperature, illuminance and humidity sensor. It is interesting to note that frequent on / off changes that occur when there is no sensor data in the light value are eliminated, providing a more stable result. Since there is not enough real data experiment, it is possible to measure the virtual power value with 10,000 samples randomly generated 10 times of virtual data similar to the real data and the result of measurement through the sensor data. As a result, accuracy of about 75% , Which showed a 5% improvement in identification of devices using small amounts of power such as light fixtures and set-top boxes that were difficult to identify using sensor data.

In addition, a comparison of the light signal is shown in Figures 18 and 19 below.

8. Home Appliance Load Identification System

10 is a block diagram of an apparatus load identification system according to an embodiment of the present invention. The apparatus load identification system according to an embodiment of the present invention includes a measurement unit 110, a detection unit 120, an analysis unit 130, and a sensor 140, as shown in FIG.

The measuring unit 110 measures the supplied power, the detecting unit 120 detects a change in electric power, and the sensor 140 senses the surrounding environment.

The analysis unit 130 learns and holds the power usage characteristics of the home appliances and estimates the home appliances that caused the detected power change in the detection unit 120. [ The device estimation uses the above-described S-FHMM, and the sensing data generated by the sensor 140 is also used.

So far, we have designed and implemented various environment setting and algorithm for non - contact device identification. In order to do this more effectively, we divided into three steps.

First, unit identification between the separated devices was performed through the PCA and KNN classifiers.

Second, a basic model is proposed for multiple identification of contactless data (Fig. 8) with Switching Factorial Hidden Markov Model with Multi-Observed data.

Third, to increase overall accuracy, sensor data values were added to the base model as new observations.

We have designed the Switching Factorial HMM to create a model that is more efficient than the existing NIALM techniques. We have also used sensor data to identify devices such as light bulbs and set-top boxes that were difficult to separate in the existing NIALM environment with a small amount of power consumption.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, It will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the present invention.

110:
120:
130:
140: sensor

Claims

Measuring supply power;
Detecting a power change; And
And estimating a device that caused the detected power change with reference to power usage characteristics of the devices.

The method according to claim 1,
The power usage features include:
Comprising at least three power related components.

The method of claim 2,
Said at least three power related components comprising:
Wherein the power-related components are three power-related components having the largest eigenvalue of the selected covariance using SVD (Singular Value Decomposition) among the power-related components.

The method according to claim 1,
Wherein,
And estimating a device that caused the power change through a stochastic k-nearest neighbor classification algorithm.

The method of claim 4,
Sensing the surrounding environment,
Wherein,
Further comprising sensing data generated in the sensing step to estimate a device that caused the power change.

A measuring unit for measuring supply power;
A detection unit for detecting a power change; And
And an analyzer for estimating a device that caused the detected power change, with reference to power usage characteristics of the devices.