WO2002027616A1

WO2002027616A1 - Energy descriptors using artificial intelligence to maximize learning from data patterns

Info

Publication number: WO2002027616A1
Application number: PCT/US2001/030400
Authority: WO
Inventors: Ellen Jones Maxon; Norman Patrick Maxon
Original assignee: Power Domain, Inc.
Priority date: 2000-09-28
Filing date: 2001-09-28
Publication date: 2002-04-04
Also published as: AU2002211297A1; AU2001294857A1; WO2002027621A1

Abstract

The present invention relates to methods of analyzing utility usage and specifically to a system and a method which develops descriptors to accurately classify and predict an individual user's or group of users (120) usage of a utility resource. Generally, the invention comprises a method for accurately estimating the usage of a utility by a specific customer, or set of customers. Detailed electricaly energy usage records (105), previously collected from a large representative sample of customers, are used to train an intelligent system. The system identifies the key characteristics (135) which lead to a particular customer having a particular usage pattern, and uses these characteristic values as predictors for determining the usage patters of new or prospective customers.

Description

ENERGY DESCRIPTORS USING ARTIFICIAL

INTELLIGENCE TO MAXIMIZE LEARNING

FROM DATA PATTERNS

This application claims priority from U.S. Provisional Patent applications "ENERGY DESCRIPTORS USING ARTIFICIAL INTELLIGENCE TO MAXIMIZE LEARNING FROM DATA PATTERNS", Application No. 60/156,306, filed September 28, 1999 and "GENETIC ALGORITHM METHOD FOR AGGREGATING ELECTRICITY CONSUMPTION AND OPTIMIZING ELECTRIC BUYING GROUPS",

Application No. 60/156,307, filed September 28, 1999, and U.S. Patent Applications "ENERGY DESCRIPTORS USING ARTIFICIAL INTELLIGENCE TO MAXIMIZE LEARNING FROM DATA PATTERNS", Application No. 09/670,895, filed September 28, 2000 and "GENETIC ALGORITHM METHOD FOR AGGREGATING ELECTRICITY

CONSUMPTION AND OPTIMIZING ELECTRIC BUYING GROUPS", Application No. 09/672,589, filed September 28, 2000.

Field of the Invention: The present invention relates to methods of analyzing utility usage and specifically to a system and a method which develops descriptors to accurately classify and predict an individual user's or group of users usage of a utility resource Bαckqround of the Invention:

Utility companies supply the resources essential to the needs of all industries, including manufacturing plants, hospitals, educational institutions, departments stores, office blocks, farms and homes. While traditionally utilities include such metered resources as water, telecommunications, gas, and electricity, utilities may include service-oriented utilities such as postal and courier services, moving services, catering, and the like. A thin line has always existed between supplying the best possible level of utility or service to a customer, while maximizing the profit to the utility supplier. Hence, for many years, regulation of utilities was the norm. However, deregulation of utility industries has placed increasing demand on utilities to become more attuned to profitability. To this end, a utility supplier may choose to segment its total market, identify each segment of the market by some descriptive means, and to tailor product advertising to each segment.

Traditional methods of segmentation and focused marketing rely on simple segmentation of the market by easily identifiable markers, such as geographical region, postal code, or estimated customer size. But this crude methodology lacks the sophistication required to identify users, or types of users, by their actual usage of the utility itself. The inability to readily and reliably characterize usage results in considerable wasted expenditure on marketing to what may amount to an unresponsive or economically unsuitable subset of customers. Pαrticulαrly, to take the example of the power supply industry, and specifically the electrical utility sector, the recent moves towards deregulation of the electric markets in the U.S. will lead to freedom of consumer choice of power supplier in many states by 2002. The players in this deregulated power industry, the utility companies themselves, must ideally have knowledge of their consumers' energy usage patterns in order to sell energy to them. Presently, however, the former monopoly utilities have all the energy data, originally collected from consumers' electric meters for billing purposes. This situation forces other, perhaps much smaller or newer, utility companies to compete at a disadvantage, in effect having to guess the characteristics of their target market before beginning to market their product there. Clearly, this puts them in a weaker position when competing against the larger former monopolies. The smaller utility companies would benefit greatly from a system that supplies them with information that enables them to level the playing field and to survive in an increasingly competitive marketplace.

Additionally, all utility companies, both large and small, need to be able to accurately estimate the periodic (for example monthly, daily, hourly) usage patterns of their customer population so as to best allocate the utility companies available resources, allow for periodic surges in utility usage, and plan for future expansion. To date, the optimum way of doing this has been to a install time-of-use meter in each customers home, and monitor it closely. Since such a process is time-consuming and, in large- scale environments, prohibitively expensive, most utility companies do not have accurate estimates as to their customer usage. Such utility companies would benefit greatly from a system that supplies them with accurately estimated information as to their customer usage on, say, a per-minute or per-hour basis, without the cost normally associated with obtaining such information.

Summary of the Invention:

The invention addresses the need for a system which helps utility companies to accurately estimate accurately estimate the energy usage characteristics of their target population compete with the larger utilities. In one aspect, the invention comprises a method that provides electricity use predictors for commercial, industrial and residential users. Generally, the invention comprises a method for accurately estimating the usage of a utility by a specific customer, or set of customers. Detailed electrical energy usage records, previously collected from a large representative sample of customers, are used to train an intelligent system. The system identifies the key characteristics which lead to a particular customer having a particular usage pattern, and uses these characteristic values as predictors for determining the usage patters of new or prospective customers.

The system described is also capable of accepting as an input a set of values related to the predictors. Such values may be obtained by direct market survey of a potential new geographical marketplace, or by reference to an information database of potential new customers. The system can then determine the usage patterns of the new customers. These usage patterns have α number of applications but are particularly useful for the purposes of segmenting the marketplace for better, more efficient, marketing.

Further, the system described is capable of learning in an intelligent manner. As data is fed through the system on successive analysis runs, a focusing of both the key predictors and the information database is undertaken on a continuous basis, ensuring that each run of new data through the system results in successively more accurate reporting of results. In one embodiment the invention comprises a method of generating a set of suitable consumer utility usage descriptors from a large population of potential utility consumers comprising: developing a utility usage descriptor library; training an artificial intelligence system using the library; and, using the trained artificial intelligence system to assign utility usage descriptors to new customers.

In another embodiment the invention comprises a method of generating a set of suitable consumer utility usage descriptors from a large population of potential utility consumers comprising: developing a utility usage descriptor library; training an artificial intelligence system using the library; and, using the trained artificial intelligence system to assign utility usage descriptors to new customers, wherein the utility usage descriptor library may be updated by reading in actual utility usage data for a customer; comparing said actual usage data for said customer with profile information in the utility usage descriptor library associated with said customer; and, updating said utility usage descriptor library. Brief Description of the Drawings:

These and other features and advantages of the present invention will become more apparent to those skilled in the art from the following detailed description in conjunction with the appended drawings in which:

Figure 1 is a flowchart of a method of an embodiment of the invention showing how key predictors are generated and used to generate and produce a energy library.

Figure 2 is a block diagram of an embodiment of the invention showing how key predictors are created and used to generate and produce a energy library.

Figure 3 is a flowchart of a method of an embodiment of the invention applied to new consumer markets.

Figure 4 is a block diagram of an embodiment of the invention applied to new consumer markets.

Figure 5 is a flowchart of a method of an embodiment of the inverition showing how the energy library is updated to reflect new information.

Figure 6 is a block diagram of a further embodiment of the invention showing how the energy library is updated to reflect new information.

Detailed Description of the Preferred Embodiment:

A system for analyzing and predicting energy usage by utility users is hereinafter described. The system first provides a series of energy descriptors through an analysis of a quantity of usage data, then provides a mechanism for analyzing that data in α manner which allows prediction of energy usage by users with unknown or new usage patterns.

ENERGY DESCRIPTORS A specific embodiment of the invention comprises a method which allows the disclosed system to estimate detailed electric energy consumption for business and residential users. The method is applied initially to a set of metered energy information to develop an energy library and to train an artificial intelligence system. The artificial intelligence system is then used to provide useful information about new sets of customers (i.e. new markets), or to assist a utility company in segmenting the current marketplace so as to best respond to new utility offerings (i.e new products). Figure 1 illustrates the steps in developing an energy library.

As used herein an energy library is a set of energy descriptors, wherein each energy descriptor comprises data pertaining to a customer. The energy descriptor may include both qualitative and quantitative data. A complete energy descriptor comprises a set of key energy descriptors which closely define a customer's energy usage, and a load shape reflecting the customer's usage pattern over an extended period of time.

Raw data (which can be obtained from detailed utility billing records) is used to initially train the analysis process. In this regard, a reference set of energy usage information gathered from a representative sample of homes and businesses is initially input into the system at step 105. It should be understood that a representαtive sample of homes and businesses comprise a statistically well-distributed or valid cross-section of homes and businesses chosen from a much larger population. This reference set of information may be obtained by purchasing a set of metered usage data from, for example, an existing utility company, or group of utility companies, whose customers have time-of-day usage meters. Alternately, it may be recorded directly by installing a set of usage meters into a test sample of customers premises and recording their electric usage over a period of time, or it may even be prepared by extrapolation of a smaller set of data. In either case the reference information reflects both the metered usage parameters (kWh, peak demand, load profile, load factor etc) and the usage patterns of a wide variety of customers over an extended period of time, for example an entire calendar year. In one embodiment the reference data may be precise to the point of knowing the customers energy usage for each interval time period (for example, 5, 15, 30, 60 minute) of a calendar year.

Other detailed information may be known for each customer in the reference set, including for example the customer's field of industry, postal code, residency or business type, geographic location. The more data available in the reference set the more accurate will be the modeling process. Mass selection techniques are used in step 1 10 to classify and separate the customers into primary groups. As used herein, a primary group is one in which all the constituent group members share a common important characteristic. In one embodiment the customers are classified according to Standard Industrial Code (SIC). The Standard Industrial Code (and its counterpart the North American Industry Classification System, NA1CS) is a code system developed by the U.S. Office of Safety and Health Administration (OSHA) to allow classification of each business in the United States. Other forms of business classification are used in the United States, with similar classification systems used in other countries. The SIC code (or simply SIC) describes to a high degree of precision the type of industry or residence a particular customer belongs to. If SIC codes are used to select customers for the primary groupings then each customer in a primary group will either have the same SIC code, or similar SIC codes. The actual number of SIC codes in each primary group can be determined beforehand based on the needs of the operator of the invention. Separating customers into multiple primary groups allows separate analysis models to be customized for each primary group. Other models may operate commonly across all primary groups. For example, in one embodiment wherein the consumer data set is segregated according to SIC code (the primary classification) - a first SIC grouping may represent chemical manufacturers while a second SIC grouping represents mining operation companies. These groupings can be represented as tables, although it will be evident that other data structures could be used. If tables are used, then each line within the table would represent a unique combination of energy parameters. Those entries in the reference set that have common SIC and common energy parameters would thus have only one entry in the appropriate SIC group. Within α particular primary group, individual customers may be further grouped into secondary groups according to their energy usage. In the embodiment using SIC codes, the customers in each secondary group would share both a common SIC code and common usage parameters. Customers not in the same secondary group but in the same primary group would share a common SIC code, but not necessarily common usage parameters or a common usage pattern. In some instances a secondary group may be comprised of a single customer record, reflecting the unique usage pattern of that customer.

The selection methods used to classify the customers into their respective primary and secondary groups may range from simple selection of customers according to common characteristics such as total usage requirement, to more sophisticated techniques including genetic separation techniques. These techniques use statistical methods, including analysis of variance, multiple regression, and discriminant analysis, to classify customers into groups. Such techniques are commonly practiced by geneticists to analyze large biologic populations and are capable of rapidly analyzing and grouping similar clusters of individuals. In one embodiment a mass selection step 1 10 is used to identify, select and classify similar energy users from a heterogeneous population into primary groups. The process of mass selection is well known to one skilled in the art. In this process the parameters or variables attached to each customer are given mathematical values and weighed to ascertain their relative weight or relevance in determining that customers overall usage pattern. The weighing causes customers to be spread out across a spectrum, where they can then be separated.

As an optional step, tags can be used in step 1 15 to further define the selection process. Tags cause some variables to have more weight than others. The use of tags allows an operator to pre-specify those features considered to best determine a customers usage pattern. Common tags may include markers for such electricity-specific variables as kilowatt-hour usage, load factor, or total kilowatt needs, or general non-specific variables such as household type, zip (postal) code, or Standard Industrial Code (SIC code).

Whichever method is used to separate out the initial set of customers, the result of the separation process is a series of groupings. Each primary group contains one or more secondary group, and each secondary group contains one or more customer types. These groups may be represented as tables, although other data structures could be used, including, for example, a relational database, a set of linked lists, or a neural network. If tables are used, then one table might represent a particular primary group, each row in that table would represent a particular secondary group, and each column in the table would represent a parameter or usage pattern associated with the secondary group. Customers that appear have substantially identical usage parameters may be represented by a single combined row in a table. Other customers that have unique usage parameters would have their own row, i.e, their own entry in the table. Stαtisticdl modeling of primary or secondary classifications from the group tables is performed in step 135 to determine which variables in each classification are the most important predictors of energy usage. As mentioned above, the models used may be generic to all primary groups, or may be customized for each primary group to best reflect the particular characteristics of that group. The result of the modeling process is a series of key predictors. As used herein a key predictor is a quantitative or qualitative parameter or variable associated with a particular customer, the value or state of which is a strong determinant of that customer's usage pattern.

In an optional step 125, demographic or marketplace data is used to supplement the data in the primary and secondary groups. For a residential customer this marketplace data may comprise information about each customer's lifestyle, including their assets, income, and hobbies. For a business entity the marketplace may include such information as business size, number of employees, type of heating or air conditioning, and operating hours. Such data can be obtained by a market research survey or predictor, and may be supplied by the same entity supplying the raw reference data, although this need not be the case. Whatever its source the marketplace data should ideally be matched to the reference data so that a high proportion of the reference data entries have a corresponding associated market demographic data entry. The addition of marketplace data in this manner adds to the precision of the model, and allows the model to consider variables such variables when classifying or estimating customer energy usage and usage patterns.

A further optional step 130 adds weather data to supplement the information in the primary and secondary groups. Since weather (both longer-period climate, and shorter-period daily fluctuations) can have a considerable effect on energy usage, information about the weather as it relates to usage is beneficial in increasing the accuracy of the model. As with the marketplace data the weather data should ideally be matched to the reference data so that a high proportion of the reference data entries have a corresponding associated weather data entry. The importance of the weather data and its use in the model varies with the market being analyzed, but is of particular relevance when analyzing markets in which seasonal variations in climate causes fluctuations in energy demand.

The modeling process thus determines the key predictors from the available data (group data, marketplace data and weather data) for each secondary group. The key predictors are associated with the secondary group to which they belong, and are thus also associated with the original customers that are now represented by that group.

In step 140, a load shape is assigned to each secondary group. A load shape is a temporal representation of the energy usage of that particular group or customer over an extended period of time, typically a year although other time periods may be chosen. The load shapes are stored in a load shape library. In one embodiment, each load shape in the load shape library may be initially created using data from the reference set. The demographic, or customer-specific, information is stripped away leaving just the energy usage parameters and energy usage pattern. This remaining information is then stored as a unique entry in the load shape library. Since the load shape is generic to all customers who have common usage parameters, a single load shape can be used to represent the usage characteristics of a wide variety of otherwise non-similar customers.

The load shapes are adjusted (or calibrated) to the correct magnitude so that the energy consumption indicated within each interval sums to the total metered kilowatt hours. The peak metered monthly demand in kilowatts is used to calibrate the load shape's peak magnitude.

The load shape library may be updated with new load shapes on a continuous basis as new data becomes available. When step 140 encounters a secondary grouping or single customer that cannot be represented by a currently stored load shape the process creates a new load shape using averaging, extrapolating, or other estimation means. The load shape library is then updated to reflect the new entry.

In step 145 an energy library is then created. As mentioned above, an embodiment of the energy library comprises a set of energy descriptors, wherein each energy descriptor comprises data pertaining to a customer type or secondary grouping. An embodiment of the energy descriptor comprises a series of entries for each customer type, including their associated usage pαrαmeters, the key predictors that determine those usage parameters, and the load shape.

Data mining techniques are used in step 150 to classify the energy descriptors within the energy library into a set of bins, creating a bin matrix. In the context of this application a bin matrix is defined as a set of bins wherein each bin contains a plurality of user types, culled from a larger population of energy users. The members of a particular bin share some common energy-related characteristic. The characteristic used to allocate members to bins need not be as simple as customer industry, geography, zip or SIC code, but may instead be energy-usage related, for example all members of a bin may have similar load shapes or use substantially the same total amount of energy in a year, even though they are in quite different areas of industry. One embodiment of the bin matrix includes a set of energy bins with, for example similar average load shapes, predictor variables, average monthly kilowatt-hour and kilowatt demand. The bin matrix is as such closely related to the usage patterns most commonly encountered in real life. Using the disclosed method as few as two bins can be used, but any number may be used, for example 5-50 bins. A larger quantity of bins provides better granularity and allows a utility company to better characterize their customer population, but the computational burden imposed upon a system operating the method increases with increased number of bins.

The energy library and the bin matrix created by this method can then be used to develop and train an artificial intelligence system. The artificial intelligent system defined by the invention utilizes a combination of the energy library and the bin matrix to analyze new customers and estimate their usage patterns. Such analysis can be performed by statistical, pattern recognition, or other computational means.

To take a specific example, suppose the initial data comprises the following customer entities:

1. Private house

2. Apartment Complex

3. Apartment complex

4. Small business

5. Department store

6. School

7. School

In the dataset shown above the numbers 1 - 7 may be considered record identifiers. A record identifier uniquely identifies that particular record as it moves through the process. An embodiment of a system which makes use of this data to build an energy library is shown in Figure 2. As shown in Figure 2, the raw data 205 is first entered into the system. Following the separation and grouping of entities into primary and secondary groups the groupings might resemble those of structure 220. Since reference entities 1 ,2, and 3 are all residences they may appear in a first group 221 while the reference entities 4,5,ό,and 7 may appear in a second group 223 for commercial entities. A mass selection module 210 ensures that similar energy users within a primary group are themselves grouped into a common secondary group. Tags 215 can be used to add weight to certain desired pαrαmeters and enhance the selection process. In the illustration of Figure 2 a secondary group can be represented as a single row within a primary group. In this example reference entities 2 and 3 are considered similar entities, as are 6 and 7, because of their typical energy usage, so they are grouped in secondary groups (i.e. along the same row) representing multiple entities. Entities 1 , 4, and 5 have very different energy parameters from their primary group members so they occupy groups of their own. In effect they are represented as unique entities. A modeling processor 235 is then used within each secondary group to determine the key predictors for each entity type. Information from marketing probes or surveys 235, and weather-related information 240, can be used to increase the reliability of the model. The result is an energy library 245, a set of profiles having both energy usage information and associated key predictors. Each of the original entities are associated with one of the profiles and corresponding predictors. If a load shape having the same usage parameters as a profile is available, then it is associated with the corresponding profile in the energy library. If a load shape is not readily available, then a closest matching one will be supplied based on the key predictors. As described herein, the energy library represents a set of 'fingerprints' or unique customer profiles.

A data mining module 250 classifies and assigns the entries in the energy library into a series of predefined bins to create a bin matrix 255. In this example entities 1 ,4 are put into bin A; entities 2,3 into bin B; entity 5 into bin C and entities 6,7 into bin D. Although 1 is α private house and 4 is a small business, for the purposes of distributing electricity 1 and 4 may be considered equivalent. Similarly with 2 and 3, and with 6 and 7.

This information can then be fed to an artificial intelligence system 260 for use in analyzing future data, i.e. data related to new customers, or new energy products.

NEW MARKET ANALYSIS

In one aspect of the present invention, the method described above may be expanded to develop energy descriptors for new business/household entities, i.e. non-reference entities for which there is no metered data available. In one embodiment of the invention an artificial intelligence system is used to classify energy users and estimate their energy usage parameters and usage patterns. To this end a business intelligence library is created. As defined herein a business intelligence library is a collection of heterogenous customer entries, wherein each customer entry corresponds to a customer in a market region or marketplace and contains energy usage information about that particular customer. According to the disclosed method customer entities are selected from the business intelligence library using a marketing probe. A marketing probe is a weighted function of energy and marketing parameters, the exact parameters chosen by the utility company to best reflect their market needs. In one embodiment the selection process is a form of rapid assay, in which each individual energy user in the population is compared against the probe in a step-wise compαrαtive procedure and then slotted into target segments. Rapid assay is a common technique well known to one of skill in the art.

Figure 3 illustrates an embodiment of a process of the invention in which electric energy descriptors are developed for commercial, industrial and residential electricity users, using the artificial intelligence system created earlier.

In step 305 data from the bin matrix is read into the system. The bin matrix comprises a set of bins wherein each bin contains a plurality of user types. The members of a particular bin share a common energy-related characteristic, for example total energy usage per year.

In step 310 data is next read into the system from a business database. In the context of this application a business database is a set of individual business or residence consumers in a geographic area. The exact contents of the business database may vary but typically may include for each consumer their name, address, and other contact information, their SIC code, and their geographical or topographical region. These variables are closely related to the key predictors determined earlier for the reference set of customers. Business databases are readily available from commercial sources, including such consumer market information vendors as Experian, Inc. (Experian, Inc. is a subsidiary of The Great Universal Stores PLC and has headquarters in Nottingham, UK and Orange, California; home page: www.experian.com). A utility company may choose to use a smαll subset of such α database pertaining only to their catchment area or their desired marketing region.

Optionally, weather information from the same geographic area is also read into the system in step 315. The weather information can be used to increase the degree of matching and hence the resolution of the process.

In step 320 the artificial intelligence system merges the data by placing the individual businesses/residences listed in the business library into the bins specified by the bin matrix. The merging process comprises determining for each record in the business database whether there is a corresponding profile in the energy library. Since both the business library and the bin matrix contain information directly related to the key predictors, for example the SIC code, many entries in the business library will have a directly corresponding bin in the bin matrix. Each of these entries in the business database is associated with a bin type and hence with that bin's usage characteristics and usage pattern.

A business intelligence library is then created in step 325. The business intelligence library contains one entry for each entry in the original business database, i.e an entry for each customer within the utility's chosen marketplace. Each entry may include both the information originally presented for that customer together with the usage information calculated for and associated with that customer in step 320. Further classification of the customers into similar groups is based on marketing probes that help to identify and classify users. In developing the marketing probe the utility company may supply certain market criteria to best meet their marketing strategy. For example, they may specify that certain markets comprise high-volume energy users, or interruptible power users, or seasonal users. In step 330, a marketing probe filters the business intelligence library according to the marketing specification. These probes locate desired electricity users by rapid assay, i.e. initial elimination of unwanted segments based on probe parameters, and the weighting of remaining records. The business intelligence library may be tagged with energy fitness weights in step 335 (i.e. with functions of energy descriptors such as load factor and seasonal fluctuations), and then searched for desired energy descriptors, energy fitness, amount of energy, type of business/residence, etc. as defined by probes. In step 340, users are matched by first assigning a score to each record. The classification of records into customer similarity bins in step 345 is based on this score, which is primarily a function of energy descriptor values. A similarity bin is one in which the members of the similarity bin exhibit similar energy requirements.

Figure 4 illustrates an example of an embodiment of a system which employs the method to classify actual records. The bin matrix 410 appears as described above, i.e. it contains four separate bins labeled A,B,C and D. A business database 415 containing a set of new population data is entered into the system, containing for example seven customer entities 1 1 - 17, wherein customer entitiesl 1 - 17 are substantially equivalent to the reference entities 1 -7 discussed earlier:

1 1. Private house

12. Apartment Complex 13. Apartment complex

14. Small business

15. Department store

16. School 17. School

Customer entities 1 1 - 17 represent new customers without associated energy usage or energy pattern information.

The artificial intelligence system 420 compares each entry in the population data with the bin matrix, by comparing the information associated with each entry in the business database with the key predictors associated with each bin of the bin matrix. Weather weights can be used to increase the resolution of the system. When the artificial intelligence system finds the corresponding bin it creates a business intelligence library 430 with the corresponding data. In the example illustrated in Figure 4 the business intelligence library will contain one entry for each member of the incoming population data together with the associated energy bin. The sample data will be given the bin classes 1 1 - A; 12 - B,... 17 - D. This business intelligence library may then be used to create customer similarity bins. A marketing specification function supplied either by the operator or produced by the system itself in accordance with a set of marketing criteria is used to sort the information in the business intelligence library according to energy usage. A fitness function 440 associated with the marketing specification function can be used to tailor the marketing specification and to predict the reliability of the outcome. The result of the process is a series of similarity bins 452-458. The similαrity bins comprise members having similar energy requirements. Bins can be chosen to correspond to one or more matrix bins. In the example similarity binl , 452 corresponds to matrix bins B, (410) and C, (410); similarity bin 2, 454 to matrix bin A, (410); and similarity bin 3, 456 to matrix bin D, (410). The customers 1 1-17 are thus placed into similarity bins: SB1 = 12, 13, 15; SB2=1 1 ,14; SB3=16,17.

These similarity bins are of great use to the utility companies since they allow the company to accurately estimate the usage patterns of a population of potential customers, and to segment their customers according to such usage patterns, easily, rapidly and without the need for extensive statistical analysis.

SELF-LEARNING SYSTEM The self-learning abilities of the invention are illustrated in

Figure 5, where metered electric data obtained from power providers or directly from customers, combined with an energy library and a trained artificial intelligence system, allows the development of new information-rich energy profiles by time-of-day for each new type of business or household. The artificial intelligence system of the invention uses feedback mechanisms to recognize and accommodate any new data pattern different from previous valid patterns. The system will thus learn this new information on a continuous and automatic basis. An embodiment of the method disclosed herein is illustrated in Figure 5. In step 505 a business intelligence library is read into the system. The business intelligence library is a collection of customer entries, wherein each customer entry corresponds to a customer in a market region or marketplace and contains energy usage information about that particular customer. True energy billing information for each customer, as recorded from meters or bills is read into the system in step 510. Those business/residences that appear in both the business intelligence library and the billing information are matched and combined with appropriate energy fitness weights and bin matrix values. These combination records are processed by the artificial intelligence system in step 515, which recognizes familiar energy-use patterns. New patterns or outliers are detected in step 520. An outlier can be created when a business originally estimated to have a certain estimated energy usage pattern (as denoted in the business intelligence library) turns out to have a different actual energy usage pattern (as denoted in the billing information). The new patterns are stored as individual table entries in the business intelligence library. In step 525, the energy library is then updated to include the new information describing the outlier. The business intelligence library may also be updated in this step to reflect the newly acquired data.

The updated energy library may then be processed in step 530 to assign a load shape to each new customer. The method of creating and assigning a new load shape is as described above, and may include such data mining techniques as cluster analysis, tree analysis, and discriminate analysis, all of which are well known to those skilled in the art. The bin matrix is updated in step 535, again as described above. ln this manner the invention comprises an electric energy information and analysis system that continues to learn and improve its accuracy as each geographic electric market develops and additional energy and user information becomes available.

Figure 6 illustrates an example of an embodiment of a system using the above learning mechanism and method of creating new profiles. An example of the business intelligence library 605 may, as discussed earlier, contain seven customer entities wherein the customer entities are in the bins shown 1-A; 2- B; 3-B...7-D. However, the billing information 610 received from the utility company or from the customer meter directly may contradict this information and suggest a different entity usage, for example 1-A; 2-B; 3-A...7-D. In this example, customer entity 3 is predicted to have energy bin matrix value B, as denoted by the business intelligence library, but in reality has the usage characteristics of energy bin matrix A, as denoted by the billing information.

When the system attempts to match the billing data against its information database the artificial intelligence 620 will detect customer entity 3's reported usage pattern as being an outlier 625 , i.e. a new pattern of customer energy usage. The energy library is updated to reflect the new data, adding a new customer type profile for customer entity 3. If customer entity 3 has a time-of-use meter, a load shape may be directly calculable, and the process is complete. Otherwise, if a load shape is not directly calculable from the raw dαtα, then an entirely new load shape can be calculated based on the average of other entities having similar prediction variables and usage characteristics. The energy library 640 is updated to add a new profile 642 for customer entity type 3., together with an associated load shape. The load shape library 630 may be updated with the new load shape for future use. The bin matrix 645 is then regrouped according to energy usage parameters to reflect the new information in the energy library.

DISTRIBUTED PROCESSING

In embodiments of the invention, a portion or all of the disclosed method may be performed by a distributed system. Such a distributed system may be based on a local area network (LAN) operating within a single office location, a wide area network (WAN) encompassing several office locations, or an open systems network such as the Internet or the Web.

In one embodiment an artificial intelligence system operating the method may be located on a central server accessible via the Internet. Other components such as the modeling processor, bin matrix data, energy library, survey information, weather information and load shape library may be located on the same central server, or may be located on other servers in communication with the central server. The central server includes means to allow a user to enter a business database from a remote location. The artificial intelligence system analyzes the information contained within the business database and calculates the desired result, which may include for example, energy estimates for each customer within the business database, or the contents of energy usage groups. The result is then returned to the remote user.

In other embodiments the artificial intelligence system may automatically retrieve data from remote locations using an

Internet protocol such as file transfer protocol (FTP), or another form of dynamic communication protocol. In the context of the application such dynamic communication protocol may include extended markup language (XML), wireless application protocol (WAP), or electronic data interface (EDI). The terms FTP, XML and WAP are common terms known to one skilled in the art. The data thus received may include raw billing information, customer usage information, or new customer population information. The system calculates the desired result and returns it to the user via a similar dynamic communication protocol.

Industrial Applicability:

The foregoing description of preferred embodiments of the present invention has been provided for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Obviously, many modifications and variations will be apparent to practitioners skilled in this art. In particular, it will be obvious that the present invention may be employed in areas other than those related to electrical utilities^ i.e. to other forms of utility and service supply that rely on accurate customer information for optimal resource allocation and profit maximization. The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the following claims and their equivalents.

Claims

Clαims:What is claimed is:

1. A method of generating a set of customer utility usage descriptors from a large population of potential utility customers comprising: developing a utility usage descriptor library; training an artificial intelligence system using the utility usage descriptor library; and, using the trained artificial intelligence system to assign utility usage descriptors to new customers.

2. The method of claim 1 wherein the utility usage descriptor library comprises a plurality of profiles, wherein each of said plurality of profiles describes a unique type of customer.

3. The method of claim 2 wherein each of said plurality of profiles comprises usage information about said unique type of customer.

4. The method of claim 1 wherein said step of developing a utility usage descriptor library comprises the steps of: reading utility billing information pertaining to a plurality of users; grouping similar users by entity type; determining key utility usage predictors related to said entity type; and, creating a utility usage descriptor library wherein each entry in the utility usage descriptor library represents a unique entity type and comprises estimated usage information and key utility usage predictor information related to said unique entity type.

5. The method of claim 4 wherein said step of grouping similar users by entity type comprises: reading variables associated with each user; assigning weights to said variables; weighing each user; grouping users that have similar weights within the same entity type.

6. The method of claim 4 wherein said step of determining said key utility usage predictors related to said entity type comprises: determining by correlation analysis which utility usage parameters associated with said entity type cause the estimated usage pattern of that entity type to most closely match the actual usage pattern of that entity type as denoted by the raw billing information.

7. The method of claim 1 wherein said marketing specification comprises a plurality of usage bins.

8. The method of claim 4 wherein said key utility usage predictors comprise demographic information associated with said unique user type.

9. The method of claim 8 wherein said demographic information is obtained by market survey of the users represented in said raw billing information.

10. The method of claim 1 wherein the utility usage descriptor library may be updated by: reading in actual utility usage data for a customer; comparing said actual usage data for said customer with profile information in the utility usage descriptor library associated with said customer; and, updating said utility usage descriptor library.

1 1. The method of claim 10 wherein said step of comparing said actual usage data with profile information in the utility usage descriptor library comprises: matching usage variables in the actual usage data with key predictors in the utility usage descriptor library to find the profile containing the most correlation between supplied usage variables and key^' predictors.

12. A method of creating a utility usage descriptor library having a plurality of profiles wherein each profile comprises demogrαphic and usage information for a specific type of user, comprising: inputting raw billing information; grouping similar users by utility usage; determine key usage predictors; and, creating a utility usage descriptor library wherein each entry represents a single entity type and comprises estimated usage information and key usage predictor information for that entity type.

13. The method of claim 12 wherein said step of grouping similar users by entity type comprises: reading variables associated with each user; assigning weights to said variables; weighing each user; grouping users that have similar weights within the same entity type.

14. The method of claim 12 wherein said step of determining the key utility usage predictors related to said entity type comprises: determining by correlation analysis which usage parameters associated with said entity type cause the estimated usage pattern of that entity type to most closely match the actual usage pattern of that entity type as denoted by the raw billing information.

15. The method of claim 12 wherein the utility usage descriptor library may be updated by: reading in actual usage data; comparing said actual usage data with profile information in the utility usage library; and, updating said utility usage descriptor library.

16. The method of claim 12 wherein said step of comparing said actual usage data with profile information in the utility usage descriptor library comprises: matching usage variables in the actual usage data with key predictors in the utility usage descriptor library to find the profile containing the most correlation between supplied usage variables and key predictors.

17. A method of estimating the utility usage characteristics of a utility customer comprising the steps of: creating a utility usage descriptor library; training an artificial intelligence system using the utility usage descriptor library; and, using the trained artificial intelligence system to assign utility usage descriptors to new customers.

18. The method of claim 17 wherein said step of creating a utility usage descriptor library comprises the steps of: reading utility billing information pertaining to a plurality of users; grouping similar users by entity type; determining key utility usage predictors related to said entity type; and, creating a utility usage descriptor library wherein each entry in the utility usage descriptor library represents a unique entity type and comprises estimated usage information and key utility usage predictor information related to said unique entity type.

19. The method of claim 18 wherein said step of grouping similar users by entity type comprises: reading variables associated with each user; assigning weights to said variables; weighing each user; grouping users that have similar weights within the same entity type.

20. The method of claim 18 wherein said step of determining said key utility usage predictors related to said entity type comprises: determining by correlation analysis which utility usage parameters associated with said entity type cause the estimated usage pattern of that entity type to most closely match the actual usage pattern of that entity type as denoted by the raw billing information.

21. The method of claim 18 wherein said key utility usage predictors comprise demographic information associated with said unique user type.

22. The method of claim 21 wherein said demographic information is obtained by market survey of the users represented in said raw billing information.

23. The method of claim 17 wherein the utility usage descriptor library may be updated by: reading in actual utility usage data for a customer; comparing said actual usage data for said customer with profile information in the utility usage descriptor library associated with said customer; and, updating said utility usage descriptor library.

24. A system for estimating the utility usage characteristics of a utility customer comprising: a utility usage descriptor library; an artificial intelligence system which uses the utility usage library to assign utility usage descriptors to new customers.

25. The system of claim 24 wherein said utility usage descriptor library comprises a plurality of entries wherein each of said entries represents a unique entity type and comprises estimated usage information and key utility usage predictor information related to said unique entity type.

26. The system of claim 25 wherein the utility usage descriptor library is created by a method comprising: reading utility billing information pertaining to a plurality of users; grouping similar users by entity type by reading variables associated with each user, assigning weights to said variables, weighing each user, and grouping users that have similar weights within the same entity type; determining key utility usage predictors related to said entity type; and, creating a utility usage descriptor library wherein each entry in the utility usage descriptor library represents a unique entity type and comprises estimated usage information and key utility usage predictor information related to said unique entity type.

27. The system of claim 24 further comprising means for updating the utility usage descriptor library.

28. The system of claim 27 wherein said means for updating the utility usage descriptor library comprises: reading in actual utility usage data for a customer; comparing said actual usage data for said customer with profile information in the utility usage descriptor library associated with said customer; and, updαting said utility usage descriptor library.

29. A system for determining a time of use profile for a utility customer, comprising: a utility usage descriptor library having a plurality of load shapes, wherein each of said load shapes represents a temporal indication of a customer's utility usage over an extended period of time; a search engine for receiving a usage predictor associated with said utility customer and determining a load shape associated with said usage predictor; and, a time of use meter for extracting from said load shape a time of use profile indicating the customer's utility usage over a specified time period.

30. The system of claim 29 wherein said utility usage descriptor library is automatically and continuously updated with new data to provide real-time determinations of the time of use profile.

31 . The system of claim 29 wherein the time of use profile is an estimate of a customer's utility usage over a time period represented by the profile.

32. The system of claim 29 wherein the time period is a day, and the time of use profile estimates a time of day usage.

33. The system of claim 29 wherein said usage predictor is any of a customer address, ZIP code, SIC code, or other customer geographic, demographic, or business type identification.

34. A method for determining a time of use profile for a utility customer, comprising the steps of: referencing within a utility usage descriptor library with a usage predictor associated with said utility customer; locating a load shape associated with said usage predictor, wherein said load shape represents a temporal indication of the customer's utility usage over an extended period of time; and, extracting from said load shape a time of use profile indicating the customer's utility usage over a specified time period.

35. The method of claim 34 further comprising the step of automatically and continuously updating the utility usage descriptor library with new data to provide real-time determinations of the time of use profile.

36. The method of claim 34 wherein said step of extracting includes the step of using the time of use profile to estimate a customer's utility usage over a time period represented by the profile.

37. The method of claim 34 wherein the time period is a day, and the time of use profile estimates a time of day usage.

38. The method of claim 34 wherein said usage predictor is any of a customer address, ZIP code, SIC code, or other customer geographic, demographic, or business type identification.