Detecting electricity theft via meter tampering using statistical methods

• US 9,600,773 B2
• Filed: 09/13/2013
• Issued: 03/21/2017
• Est. Priority Date: 06/04/2013
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Dhurandhar, Amit, Kalagnanam, Jayant R., Siegel, Stuart A., Zhu, Yada
• Primary Examiner(s): Shah, Kamini S
• Assistant Examiner(s): COOK, BRIAN S
• Application Number:
  US14/026,695
• Publication Number:
  US 20140358839A1
• Time in Patent Office: 1,285 Days
• Field of Search: 703/6
• US Class Current: 1/1
• CPC Class Codes: G01D 2204/22 Arrangements for detecting or reporting faults, outages or leaks
• CPC Class Codes: G01D 4/00 Tariff metering apparatus in taximeters G07B13/00; apparatus actuated by coins, cards or the like with meter-controlled dispensing of liquid, gas, or electricity G07F15/00
• CPC Class Codes: G01R 22/066 Arrangements for avoiding or indicating fraudulent use
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/04 Inference or reasoning models
• CPC Class Codes: G06Q 50/06 Energy or water supply
• CPC Class Codes: G08B 21/18 Status alarms G08B21/02 takes precedence
• CPC Class Codes: Y04S 20/30 Smart metering, e.g. specially adapted for remote reading

Abstract

A method for detecting anomalous energy usage of building or household entities. The method applies a number of successively stringent anomaly detection techniques to isolate households that are highly suspect for having engaged in electricity theft via meter tampering. The system utilizes historical time series data of electricity usage, weather, and household characteristics (e.g., size, age, value) and provides a list of households that are worthy of a formal theft investigation. Generally, raw utility usage data, weather history data, and household characteristics are cleansed, and loaded into an analytics data mart. The data mart feeds four classes of anomaly detection algorithms developed, with each analytic producing a set of households suspected of having engaged in electricity theft. The system allows a user to select households from each list or a set based on the intersection of all individual sets.

Claims

1. A computer-implemented method for detecting anomalous energy usage amongst building and household entities, said method comprising:

receiving, at a computing system, data comprising energy usage data relating to a building'"'"'s actual energy use over a defined time period (shift), characteristics data of the building, and weather data over one or more defined time periods, said building characteristics data comprising one or more selected from the group consisting of;

a physical size of the building, a number of floors, a number of occupants, building age, a number of bedrooms, a number of bathrooms, a type of heating fuel in use, a type of air conditioning in use, and a latitude and longitude coordinate of the building;

clustering buildings in one or more clusters as determined based on a building'"'"'s energy usage in each time period;

identifying buildings having energy usage that migrate from one cluster to another cluster across time of day shifts, said identifying comprising;

representing a series of data processing operations as a series of data processing nodes, a node representing a usage data source for obtaining said energy usage data for a building;

performing, from said node, concurrent data processing operations as parallel paths according to a respective shift, each parallel path comprising a flow of operations comprising a linking of said node, a node representing a data aggregation operation for that shift and a node representing clustering operations for that shift;

merging data operations at each parallel path to obtain a list of energy use migrations for that building;

generating a model to predict a building'"'"'s energy usage, said model defining expected bounds of energy consumption given a time period (shift) and said weather data and building characteristics data received;

comparing energy usage for each building against an energy use predicted by the model for said building; and

identifying, from said comparison, buildings whose electricity usage is not predicted by the model, said identifying comprising;

performing, from said node representing a usage data source for a building, concurrent data processing operations as parallel paths according to a respective shift, each parallel path comprising a flow of operations comprising a linking of said node, a node for generating non-linear regression models of energy use data aggregation operation for that shift, a node of operations for detecting anomalous buildings for that shift, and a node of operations to rank a building'"'"'s anomalous energy usage based on statistics computed for the number of times a household'"'"'s usage is flagged as anomalous; and

merging data operations at each parallel path to obtain a list of top buildings having anomalous energy usage;

wherein said buildings identified as migrating from one cluster to another cluster between time periods, and said buildings exhibiting electricity usage not predicted by said generated model are flagged as anomalous energy usage entities, and

wherein a processing unit of said computer system is configured to perform said receiving, clustering, identifying of migrating buildings, model generating, comparing, and said identifying of buildings from said comparison.

2. The method of claim 1, wherein said clustering buildings comprises:

for each time period, assigning a building to one of N clusters of buildings determined based on their energy usage;

said clusters including;

low usage buildings, medium usage buildings, and high usage buildings.

3. The method of claim 1, wherein said generated model comprises a non-linear regression model of energy usage for a building and weather type, said energy use comparing comprising:

factoring out variations due to building characteristics and weather conditions from said generated model,

wherein said identified building exhibits an anomalous energy usage given one or more of a weather condition and building characteristic.

4. The method of claim 1, wherein said method further comprises:

sorting the flagged buildings based on a metric for ranking buildings based on a number of times a building'"'"'s actual energy usage falls below the model-defined lower bounds factoring in said weather condition and building characteristic.

5. The method of claim 3, further comprising:

obtaining, from said generated non-linear regression model, a pattern of the residuals of the generated model over time, a residual representing a difference between the energy actual and predicted energy use for a building; and

analyzing residual effects of a building'"'"'s energy usage after said factoring out building characteristics and weather conditions using a randomization test, said randomization test determining, for each building or household, whether its obtained pattern of residuals is random or demonstrates non-randomness,

wherein said buildings identified as demonstrating non-randomness of its obtained pattern of residuals are flagged as anomalous energy usage entities.

6. The method of claim 5, wherein said randomization test detects non-randomness, said test comprising:

obtaining a first statistic value “

Z”

computed over pattern comprising a binary sequence “

s”

of computed residual values;

computing a second statistic value ‘

P’

to detect patterns in the residuals, said ‘

P’

governed according to;

P=Z*f

wherein a value “

f”

is a function of the number of residual values of said binary sequence “

s”

computed as equal to 1, divided by a length of said binary sequence.

7. The method of claim 6, further comprising:

determining whether said computed value P>

1 indicating a residuals pattern, and

flagging said corresponding building as exhibiting anomalous energy usage.

8. The method of claim 3, further comprising:

clustering buildings according to said building characteristics, for each shift, based on energy usage;

building linear or non-linear models of energy usage versus temperature; and

determining, based on a goodness of fit of shift-level models, whether a household is better suited in one cluster over another cluster.

9. The method of claim 8, wherein said building of linear or non-linear models comprises:

invoking a piecewise robust linear regression technique that models a usage pattern of each building, wherein resulting residuals form a random probability field;

{Rn,t}, where t denotes the t-th day, n is the building entity;

estimating a sigma and theft likelihood field {Ln,t} of each building and each day, and obtaining the likelihood field for all the shifts; and

,

identifying said buildings and corresponding defined time periods as anomalous energy usage entities based on rules applied to the theft likelihood field.

10. The method of claim 9, wherein said building of linear or non-linear models comprises:

wherein said identifying buildings as anomalous energy usage entities based on a rule comprises;

determining if a theft likelihood of a shift for a building is inconsistent to a theft likelihood determined for other shifts in a day, and is inconsistent to a theft likelihood determined for other buildings.

11. A computer-implemented method for detecting anomalous energy usage amongst building and household entities, said method comprising:

receiving, at a computing system, data comprising energy usage data relating to a building'"'"'s actual energy use over a defined time period (shift), characteristics data of the building, and weather data over one or more defined time periods, said building characteristics data comprising one or more selected from the group consisting of;

a physical size of the building, a number of floors, a number of occupants, building age, a number of bedrooms, a number of bathrooms, a type of heating fuel in use, a type of air conditioning in use, and a latitude and longitude coordinate of the building;

running a first anomaly energy use detection scheme to cluster buildings in one or more clusters as determined based on a building'"'"'s energy usage in each time period and identify buildings having energy usage that migrate from one cluster to another cluster across time of day shifts, said running a first anomaly energy use detection scheme comprising;

representing a series of data processing operations as a series of data processing nodes, a node representing a usage data source for obtaining said energy usage data for a building;

performing, from said node, concurrent data processing operations as parallel paths according to a respective shift, each parallel path comprising a flow of operations comprising a linking of said node, a node representing a data aggregation operation for that shift and a node representing clustering operations for that shift;

merging data operations at each parallel path to obtain a list of energy use migrations for that building;

running a second anomaly energy use detection scheme to generate a non-linear regression model to predict a building'"'"'s energy usage, said model defining expected bounds of energy consumption given a time period (shift) and weather and building characteristics data received, and identifying buildings whose electricity usage is not predicted by said model, said identifying comprising;

performing, from said node representing a usage data source for a building, concurrent data processing operations as parallel paths according to a respective shift, each parallel path comprising a flow of operations comprising a linking of said node, a node for generating said non-linear regression model to predict energy use data aggregation operation for that shift, a node of operations for detecting anomalous buildings for that shift, and a node of operations to rank a building'"'"'s anomalous energy usage based on statistics computed for the number of times a household'"'"'s usage is flagged as anomalous; and

merging data operations at each parallel path to obtain a list of top buildings having anomalous energy usage; and

running a third anomaly energy use detection scheme to obtain from said generated non-linear regression model, a pattern of residuals of the generated model over time, a residual representing a difference between the energy actual and predicted energy use for a building, and identifying buildings demonstrating non-randomness of its obtained pattern of residuals; and

running a fourth anomaly energy use detection scheme to estimate a sigma and theft likelihood field {Ln,t} of each building and each day, and obtaining a theft likelihood field for all the shifts; and

, invoke a rule applied to said theft likelihood field; and

flagging anomalous energy usage buildings as buildings identified as;

having energy usage that migrate from one cluster to another cluster between time periods;

having electricity usage not predicted by said model;

exhibiting demonstrating non-randomness of its obtained pattern of residuals;

or have a residual energy computed for a building as exceeding a certain distance from an expected residual mean based on said rule;

or combinations thereof.

12. The computer-implemented method of claim 11, wherein running said second anomaly energy use detection scheme comprises:

generating a model to predict a building'"'"'s energy usage, said model defining expected bounds of energy consumption given time of day (shift) and weather and building characteristics data received;

comparing energy usage for each building against an energy use predicted by the model for said building; and

identifying, from said comparison, buildings whose electricity usage is not predicted by model, wherein said generated model comprises a non-linear regression model of energy usage for a building and weather type, said energy use comparing comprising;

factoring out variations due to building characteristics and weather conditions from said generated model.

13. The computer-implemented method of claim 11, wherein running said third anomaly energy use detection scheme comprises:

factoring out variations due to building characteristics and weather conditions from said generated model,

analyzing residual effects of a building'"'"'s energy usage after said factoring out building characteristics and weather conditions using a randomization test, said randomization test determining, for each building or household, whether its obtained pattern of residuals is random or demonstrates non-randomness, wherein said randomization test comprises;

obtaining a first statistic value “

Z”

computed over pattern comprising a binary sequence “

s”

of computed residual values;

computing a second statistic value ‘

P’

to detect patterns in the residuals, said ‘

P’

governed according to;

P=Z*f

wherein a value “

f”

is a function of the number of residual values of said binary sequence “

s”

computed as equal to 1, divided by a length of said binary sequence, and

wherein a computed value P>

1 indicates a residuals pattern resulting in said corresponding building being flagged as exhibiting anomalous energy usage.