Apparatus and method for detecting multi-type energy consumption anomaly, and recording medium

Abstract

The present invention relates to an apparatus for detecting a multi-type energy consumption anomaly, comprising: a data input unit for receiving multi-type energy consumption data about at least one energy source; a first multi-type instance learning unit for detecting whether an anomaly exists in consideration of the correlation between energy sources of the multi-type energy consumption data, and generating and aggregating description information about the energy sources in which the anomaly is detected; a second multi-type instance learning unit for detecting whether an anomaly exists in consideration of the correlation between the result of aggregating the description information in the first multi-type instance learning unit and time points in which the multi-type energy consumption data is generated, and generating and aggregating description information about the time point at which the anomaly is detected; and an output unit for outputting a final determination of whether an anomaly exists in the multi-type energy consumption data on the basis of the result of aggregating the description information in the second multi-type instance learning unit.

Description

Device, method and recording medium for detecting abnormal consumption of various types of energy

[0001] The present invention relates to a device, method and recording medium for detecting abnormal consumption of various types of energy.

[0002] Outlier detection in energy consumption plays a crucial role in reducing energy consumption and carbon emissions by detecting wasteful energy consumption and quickly identifying incorrect consumption patterns.

[0003] Smart meters are advanced meters equipped with communication capabilities that measure energy usage, such as electricity, gas, and water. Unlike traditional meters, they offer innovative features such as real-time data collection, remote control and management, energy usage information provision, and enhanced customer service. Furthermore, the proliferation of Advanced Metering Infrastructure (AMI), which automatically analyzes power usage by remotely reading data from smart meters, has made real-time data collection for a wide range of energy sources possible.

[0004] Meanwhile, multi-energy data or submetering data is being utilized because it can additionally detect anomalies that are only observed through correlations between energy consumption amounts that are not observed in single energy consumption amounts.

[0005] However, current methodologies for detecting abnormal consumption of multiple types of energy focus on learning and detection, focusing on whether the input time series data for multiple types of energy consumption contains anomalies. This limits the ability to easily explain the energy or submetering source that caused the anomaly, at what point in the input time series data.

[0006] Additionally, in order to detect anomalies in energy consumption, each energy consumption must be collected in real time, but the communication process for collecting energy consumption raises various security risks.

[0007] [Prior Art Literature]

[0008] [Patent Document]

[0009] (Patent Document 1) Korean Patent Publication No. 10-2531291

[0010] The present invention has been devised to solve the above problems, and an object of the present invention is to provide a device, method, and recording medium for detecting abnormal consumption by considering correlations between energy sources of various types of energy and correlations between time points.

[0011] In order to achieve the above object, according to one embodiment of the present invention, a device for detecting abnormal consumption of multiple types of energy comprises: a data input unit for receiving multiple types of energy consumption data for at least one energy source; a first multiple instance learning unit for detecting anomalies by considering correlations between energy sources of the multiple types of energy consumption data, and generating and aggregating explanatory information for energy sources in which anomalies are detected; a second multiple instance learning unit for detecting anomalies by considering correlations between time points at which the multiple types of energy consumption data are generated and a result of aggregating the explanatory information in the first multiple instance learning unit, and generating and aggregating explanatory information for the time points at which the anomalies are detected; and an output unit for outputting a final abnormality result for the multiple types of energy consumption data based on a result of aggregating the explanatory information in the second multiple instance learning unit.

[0012] In order to achieve the above object, according to one embodiment of the present invention, a method for detecting abnormal consumption of a multi-type energy abnormal consumption detection device includes: a step of a data input unit receiving multi-type energy consumption data for at least one energy source; a step of a first multi-type instance learning unit detecting whether an anomaly exists by considering correlations between energy sources of the multi-type energy consumption data, and generating and aggregating explanatory information for the energy source in which the anomaly is detected; a step of a second multi-type instance learning unit detecting whether an anomaly exists by considering correlations between time points at which the multi-type energy consumption data is generated and generating and aggregating explanatory information for the time points at which the anomaly is detected; and a step of an output unit outputting whether a final abnormality exists for the multi-type energy consumption data based on a result of aggregating the explanatory information in the second multi-type instance learning unit.

[0013] According to one aspect of the present invention described above, by providing a method for detecting abnormal consumption that considers correlations between energy sources and time points of multiple energy sources, abnormal consumption of multiple energy sources can be more effectively detected. Furthermore, by simultaneously detecting abnormal consumption and providing an explanation of the time point at which the abnormality occurred and the energy source responsible for the abnormality, the method reduces the need for additional expert analysis and allows for an immediate explanation of the abnormality.

[0014] Additionally, unlike existing methodologies that require additional re-access to the original data to determine the cause of an anomaly, this method ensures safety from a security perspective by finding the cause of an anomaly through a pre-trained model.

[0015] Figure 1 is an overall framework of a multi-energy abnormal consumption detection device according to an embodiment of the present invention;

[0016] Figure 2 is a device diagram showing the internal blocks of a multi-energy abnormal consumption detection device according to an embodiment of the present invention.

[0017] FIG. 3 is a diagram showing an abnormal consumption detection procedure of a multi-energy abnormal consumption detection device according to an embodiment of the present invention;

[0018] FIG. 4 is a diagram showing a process of detecting multiple types of energy abnormal consumption and generating related explanatory information by applying the CAL-MIL model according to an embodiment of the present invention.

[0019] And, Fig. 5 is a flowchart showing the abnormal consumption detection operation of the multi-energy abnormal consumption detection device according to an embodiment of the present invention.

[0020] The following detailed description of the present invention refers to the accompanying drawings, which illustrate specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention, while different from each other, are not necessarily mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the present invention. Furthermore, it should be understood that the positions or arrangements of individual components within each disclosed embodiment may be modified without departing from the spirit and scope of the present invention. Accordingly, the following detailed description is not intended to be limiting, and the scope of the present invention is defined only by the appended claims, along with the full scope of equivalents to which such claims are entitled, if properly described. Like reference numerals in the drawings designate the same or similar functionality throughout the several aspects.

[0021] The components according to the present invention are defined by functional distinctions rather than physical distinctions, and can be defined by the functions each component performs. Each component may be implemented as hardware or program code and processing units that perform each function, and the functions of two or more components may be implemented by including them in a single component. Therefore, the names given to the components in the following embodiments are not intended to physically distinguish each component, but rather to suggest the representative functions performed by each component, and it should be noted that the technical spirit of the present invention is not limited by the names of the components.

[0022] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

[0023] Figure 1 is an overall framework of a multi-energy abnormal consumption detection device according to an embodiment of the present invention.

[0024] Referring to FIG. 1, the illustrated abnormal consumption detection device (100) inputs multiple energy consumption data for at least one energy source. The energy source may be, for example, electricity, gas, water, hot water, etc. In addition, submetering data may be input into the abnormal consumption detection device (100) instead of the multiple energy consumption data.

[0025] An abnormal consumption detection device (100) into which multiple energy consumption data is input outputs abnormality detection results of multiple energy consumption and explanatory information about the energy source and time point at which the abnormality was detected. That is, the abnormal consumption detection device (100) applies a Correlation Aware Lightweight-Multiple Instance Learning (CAL-MIL) model that considers correlation between instances while reducing computational complexity based on a self-attention mechanism to detect abnormalities in multiple energy consumption in two stages and generates and outputs explanatory information about the detected energy source and time point.

[0026] Hereinafter, the operation of detecting abnormal consumption of various types of energy by the above abnormal consumption detection device (100) will be described in more detail through FIGS. 2 and 3.

[0027] FIG. 2 is a device diagram showing the internal blocks of a multi-energy abnormal consumption detection device according to an embodiment of the present invention.

[0028] Referring to FIG. 2, the illustrated abnormal consumption detection device (100) includes a data input unit (110), a first multi-instance learning unit (120), a second multi-instance learning unit (130), and an output unit (150). In addition, the abnormal consumption detection device (100) may include a feature extraction unit (not shown).

[0029] The data input unit (110) receives multiple energy consumption data for at least one energy source.

[0030] The first multi-instance learning unit (120) detects whether there is an anomaly in the multi-instance energy consumption data by considering the correlation between energy sources in the multi-instance energy consumption data, and generates and aggregates explanatory information about the energy sources in which an anomaly is detected.

[0031] The second multi-instance learning unit (130) considers the correlation between the time points at which normal / abnormal and multi-instance energy consumption data are generated, i.e., the result of aggregating the explanatory information from the first multi-instance learning unit (120), to detect whether multi-instance energy consumption data is abnormal, and generates and aggregates explanatory information for the time points at which the abnormality is detected.

[0032] The output unit (140) outputs whether there is a final abnormality for the multi-type energy consumption data based on the result of aggregating the description information in the second multi-type instance learning unit (130). At this time, if the final abnormality indicates "abnormality," the output unit (140) outputs the final detection result (abnormality) and a description of the energy source and time point where the abnormality was detected. In other words, the output unit (140) outputs a description of the energy source and time point where the abnormality was detected, generated in the first and second description units (120, 130).

[0033] FIG. 3 is a diagram showing an abnormal consumption detection procedure of a multi-energy abnormal consumption detection device according to an embodiment of the present invention, and FIG. 4 is a diagram showing a process of detecting multi-energy abnormal consumption and generating related explanatory information by applying a CAL-MIL model according to an embodiment of the present invention.

[0034] The data input unit (110) of the multi-energy abnormal consumption detection device (100) receives N multi-energy consumption data (energy 1 consumption data, energy 2 consumption data, ..., energy N-1 consumption data, energy N consumption data). Here, the multi-energy consumption data includes energy instance information and time instance information.

[0035] The first multi-instance learning unit (120) detects whether there is an anomaly by considering the correlation between energy instance information generated at each time point, and generates and aggregates explanatory information about the energy instance in which an anomaly is detected. That is, the first multi-instance learning unit (120) applies an attention weight calculated through the CAL-MIL model to the energy instance information generated at each of the first to T-th time points (time-point 1 to time-point T), aggregates the anomaly detection results (normal / abnormal) and explanatory information about the energy source in which an anomaly is detected, and transmits the aggregated information to the second multi-instance learning unit (130).

[0036] The second multi-instance learning unit (130) detects anomalies by considering the correlation between the aggregated results from the first multi-instance learning unit (120) and the time instance information, and generates and aggregates explanatory information for the time instances where anomalies are detected. That is, the second multi-instance learning unit (130) applies attention weights to the aggregated results from the first multi-instance learning unit (120) and outputs the final anomaly detection result (anomaly) through the output unit (140).

[0037] In particular, the first and second multi-instance learning units (120, 130) apply the CAL-MIL model proposed in the present invention to detect abnormalities in multi-energy consumption in two steps and generate and output explanatory information about the energy source and time point where the abnormality was detected.

[0038] As described in more detail through Fig. 4, the feature extraction unit extracts a feature vector from energy instance information and transfers it to the first multi-instance learning unit (120), and extracts a feature vector from time instance information and transfers it to the second multi-instance learning unit (130).

[0039] The first multi-instance learning unit (120) calculates an anomaly score by energy instance from the value feature vector among the feature vectors of value, key, and query received from the feature extraction unit. Here, the value feature vector represents the anomaly detection result of individual energy instances without considering correlation and is derived by applying the LogSoftmax activation function, and the key and query feature vectors are derived by applying the sigmoid activation function to prevent gradient explosion. In addition, the element v constituting the anomaly score matrix by energy instance ij represents the ideal score for the ith energy instance information occurring at point j.

[0040] The first multi-instance learning unit (120) calculates the attention weight by energy instance using the key feature vector and query feature vector as shown in the following mathematical expression 1.

[0041] [Mathematical Formula 1]

[0042]

[0043] Here, the key feature vector K is a one-dimensional matrix of length N, the query feature vector Q is an N-dimensional square matrix, and the attention weight matrix is ​​normalized by dividing the product of the key feature vector matrix and the query feature vector matrix by N. Therefore, each element of the attention weight matrix of length N represents the attention weight to be multiplied by the anomaly detection result of an individual energy instance, and the element a ij represents the attention weight for the jth energy instance information occurring at point i.

[0044] Thereafter, the first multi-instance learning unit (120) multiplies the anomaly score by the energy instance, i.e., the anomaly detection result of each energy instance, by the attention weight by the energy instance, and adds them to generate explanatory information for the energy instance in which an anomaly is detected. Then, the explanatory information is aggregated according to a preset method and the anomaly detection result (normal / abnormal) is transmitted to the second multi-instance learning unit (130). Here, the anomaly detection result represents the anomaly detection result for the entire set of energy instances occurring at each point in time.

[0045] The second multi-instance learning unit (130) sets the aggregated result received from the first multi-instance learning unit (120) as an ideal score by the time instance, and calculates the attention weight by the time instance using the feature vectors received from the feature extraction unit, i.e., the key feature vector and the query feature vector. Here, the element aT constituting the attention weight matrix by the time instance i represents the attention weight for the i-th time instance information.

[0046] The second multi-instance learning unit (130) generates explanatory information for the time instance in which an anomaly is detected by multiplying the anomaly score by the attention weight of the time instance and adding them together. Then, the explanatory information is aggregated according to a preset method and the anomaly detection result (normal / abnormal) is transmitted to the output unit (140).

[0047] The CAL-MIL model employed in the present invention is trained to describe anomalies, i.e., the occurrence of abnormal consumption of multiple types of energy and the individual energy sources causing the anomalies, without the need for labels for specific time points and individual energy sources. Therefore, the CAL-MIL model detects anomalies in the consumption of multiple types of energy while simultaneously providing descriptions of the time points and energy sources at which the anomalies were detected.

[0048] To apply the CAL-MIL model to each energy source, a feature extraction unit is utilized, as described above. The feature extraction unit utilized in the CAL-MIL model is designed with a simple structure to reduce computational overhead and incorporates a self-attention mechanism to account for correlations between extracted feature vectors.

[0049] In addition, the feature extraction unit uses a 1D (1Dimension) convolution layer and a 1D max pooling layer to extract feature vectors of the same size from individual energy instance information, and additionally uses a fully connected (FC) layer to extract value, key, and query feature vectors. Here, the FC layer is designed to detect anomalies that fit the characteristics of each energy source.

[0050] Meanwhile, the layer for extracting query and key feature vectors is designed to reflect global characteristics using a common layer and enable extraction using an attention mechanism, thereby extracting specific query and key feature vectors for each air support. The query and key feature vectors are then used to calculate attention weights, and by multiplying the value feature vector by the attention weight, the anomaly detection results for each time point and an explanation of the energy source that caused the anomaly at that time point can be obtained.

[0051] To account for energy instances that cause anomalous consumption of multiple energies, the first CAL-MIL is performed using individual energy instances from different energy sources within the same energy window as the initial energy instance. A second CAL-MIL is then performed using the anomaly detection results for each energy window and the feature vectors extracted from the energy sources within the corresponding energy window. In this process, information about each energy window is used as a viewpoint instance to perform CAL-MIL based on the presence or absence of anomalies in each energy window. The feature vectors of individual energy instances extracted through the encoder corresponding to each stage are concatenated to obtain the feature vector of the viewpoint instance. The query and key feature vectors are extracted using the FC layer of the viewpoint instance. The value feature vector of the viewpoint instance is obtained by performing CAL-MIL on the energy instance.

[0052] FIG. 5 is a flowchart showing an abnormal consumption detection operation of a multi-energy abnormal consumption detection device according to an embodiment of the present invention.

[0053] The data input unit of the multi-energy abnormal consumption detection device receives multi-energy consumption data for at least one energy source. (S501)

[0054] The first multi-instance learning unit detects whether there is an anomaly in the multi-instance energy consumption data by considering the correlation between energy sources of the multi-instance energy consumption data input in S501. (S503) Then, the first multi-instance learning unit generates and aggregates explanatory information about the energy sources for which an anomaly is detected. (S505)

[0055] The second multi-instance learning unit detects whether there is an anomaly in the multi-instance energy consumption data by considering the correlation between the results (normal / abnormal) aggregated in S505 and the points in time at which the multi-instance energy consumption data is generated. (S507) Then, the second multi-instance learning unit generates and aggregates explanatory information about the points in time at which the anomaly is detected. (S509)

[0056] The output section outputs whether the final abnormality is present for the various energy consumption data based on the results (normal / abnormal) aggregated in S509 (S511). At this time, if the results aggregated in S509 are “abnormal,” the output section outputs the explanatory information generated in S505 and S509 together with the aggregated results.

[0057] The method for detecting abnormal energy consumption of various types of the present invention may be implemented in the form of program commands that can be executed by various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program commands, data files, data structures, etc., either singly or in combination.

[0058] The program commands recorded on the above computer-readable recording medium may be specially designed and configured for the present invention or may be known and available to those skilled in the art of computer software.

[0059] Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, and flash memory.

[0060] Examples of program instructions include not only machine language codes, such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.

[0061] Although various embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and various modifications can be made by a person having ordinary skill in the art to which the present invention pertains without departing from the gist of the present invention as claimed in the claims, and such modifications should not be understood individually from the technical idea or prospect of the present invention.

Claims

1. A data input unit for receiving multiple energy consumption data for at least one energy source; A first multi-instance learning unit that detects anomalies by considering correlations between energy sources of the above multi-instance energy consumption data, and generates and aggregates explanatory information about energy sources in which anomalies are detected; A second multi-instance learning unit that detects anomalies by considering the correlation between the results of aggregating the explanatory information in the first multi-instance learning unit and the time points at which the multi-instance energy consumption data is generated, and generates and aggregates explanatory information for the time points at which the abnormality is detected; and A multi-type energy abnormality consumption detection device, comprising an output unit that outputs whether or not there is a final abnormality in the multi-type energy consumption data based on the result of aggregating the description information in the second multi-type instance learning unit.

2. In paragraph 1, A multi-energy abnormal consumption detection device, wherein the above multi-energy consumption data includes energy instance information and time instance information.

3. In paragraph 2, A multi-type energy abnormal consumption detection device further comprising a feature extraction unit that extracts a feature vector from the energy instance information and transfers the feature vector to the first multi-type instance learning unit, and extracts a feature vector from the time instance information and transfers the feature vector to the second multi-type instance learning unit.

4. In paragraph 3, The above feature extraction unit, Extracting feature vectors of values, keys, and queries from the energy instance information and transferring them to the first multi-instance learning unit, A multi-type energy abnormality consumption detection device that extracts key and query feature vectors from the above-mentioned point-in-time instance information and transmits them to the second multi-type instance learning unit.

5. In paragraph 4, The above first multi-instance learning unit, Calculate the anomaly score by energy instance from the above value feature vector, Using the above key feature vector and query feature vector, the attention weight by energy instance is calculated, Generate explanatory information for an energy instance in which an anomaly is detected by applying an attention weight by the energy instance to the anomaly score by the energy instance, A multi-energy abnormal consumption detection device that aggregates the above-described information according to a preset method and transmits the result to the second multi-instance learning unit.

6. In paragraph 5, The second multi-instance learning unit is, The aggregated result received from the first multi-instance learning unit is set as the ideal score by the point instance, Using the above key feature vector and query feature vector, the attention weight by viewpoint instance is calculated, Generate explanatory information for the time instance where an anomaly is detected by applying the attention weight by the time instance to the anomaly score by the time instance, A multi-energy abnormal consumption detection device that aggregates the above-described information according to a preset method and transmits the result to the output unit.

7. A step in which the data input unit receives multiple energy consumption data for at least one energy source; A step in which a first multi-instance learning unit detects anomalies by considering correlations between energy sources of the multi-instance energy consumption data, and generates and aggregates explanatory information about energy sources in which anomalies are detected; A step of detecting anomalies by considering the correlation between the results of the second multi-instance learning unit aggregating the explanatory information from the first multi-instance learning unit and the time points at which the multi-instance energy consumption data is generated, and generating and aggregating explanatory information for the time points at which the anomalies are detected; and A method for detecting abnormal consumption of a multi-energy abnormal consumption detection device, comprising: a step of outputting whether or not there is a final abnormality in the multi-energy consumption data based on the result of aggregating the explanatory information in the second multi-instance learning unit; 8. In paragraph 7, A method for detecting abnormal consumption of multiple types of energy, wherein the above-mentioned multiple types of energy consumption data includes energy instance information and time instance information.

9. In paragraph 8, A method for detecting abnormal multi-type energy consumption, further comprising a step of a feature extraction unit extracting a feature vector from the energy instance information and transferring it to the first multi-type instance learning unit, and a step of extracting a feature vector from the time instance information and transferring it to the second multi-type instance learning unit.

10. In paragraph 9, The steps that the above feature extraction unit transmits are: Extracting feature vectors of values, keys, and queries from the energy instance information and transferring them to the first multi-instance learning unit, A method for detecting abnormal energy consumption of multiple types, wherein feature vectors of keys and queries are extracted from the above-mentioned point-in-time instance information and transmitted to the second multiple-instance learning unit.

11. In paragraph 10, The step of aggregating the above first multi-instance learning unit is as follows: Calculate the anomaly score by energy instance from the above value feature vector, Using the above key feature vector and query feature vector, the attention weight by energy instance is calculated, Generate explanatory information for an energy instance in which an anomaly is detected by applying an attention weight by the energy instance to the anomaly score by the energy instance, A method for detecting abnormal multi-energy consumption, which aggregates the above-mentioned description information according to a preset method and transmits the result to the second multi-instance learning unit.

12. In paragraph 11, The step of aggregating the second multi-instance learning unit is as follows: The aggregated result received from the first multi-instance learning unit is set as the ideal score by the point instance, Using the above key feature vector and query feature vector, the attention weight by viewpoint instance is calculated, Generate explanatory information for the time instance where an anomaly is detected by applying the attention weight by the time instance to the anomaly score by the time instance, A method for detecting abnormal consumption of multiple types of energy, which aggregates the above-mentioned explanatory information according to a preset method and transmits the result to the above-mentioned output unit.

13. A recording medium having recorded thereon a computer program for performing an abnormal consumption detection method of a multi-energy abnormal consumption detection device of claim 7.

Images