IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 7, NO. …

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 7, NO. 3, AUGUST 2011 399

A Data Mining Framework for ElectricityConsumption Analysis From Meter Data

Daswin De Silva, Xinghuo Yu, Fellow, IEEE, Damminda Alahakoon, and Grahame Holmes, Senior Member, IEEE

Abstract—This paper presents a novel data mining frameworkfor the exploration and extraction of actionable knowledge fromdata generated by electricity meters. Although a rich source ofinformation for energy consumption analysis, electricity metersproduce a voluminous, fast-paced, transient stream of data thatconventional approaches are unable to address entirely. In orderto overcome these issues, it is important for a data mining frame-work to incorporate functionality for interim summarization andincremental analysis using intelligent techniques. The proposedIncremental Summarization and Pattern Characterization (ISPC)framework demonstrates this capability. Stream data is struc-tured in a data warehouse based on key dimensions enablingrapid interim summarization. Independently, the IPCL algorithmincrementally characterizes patterns in stream data and corre-lates these across time. Eventually, characterized patterns areconsolidated with interim summarization to facilitate an overallanalysis and prediction of energy consumption trends. Resultsof experiments conducted using the actual data from electricitymeters confirm applicability of the ISPC framework.

Index Terms—Data mining, electricity meters, energy consump-tion analysis, incremental learning, interim summarization.

I. INTRODUCTION

E NERGY consumption analysis is a primary research areain power systems planning and management. Recent de-

velopments in energy market deregulation and provision of sus-tainable energy have contributed to increased interest in thisarea. Electricity meters record energy consumption and powerquality at a preset interval, usually an hour or less. Readingsfrom meters in a specific geographical area are transmitted inreal-time to a central location where it can be analyzed to pro-duce immediate outcomes that support decision-making. Theavailability of accurate and updated information on energy con-sumption presents the potential to transform demand forecastingand energy conservation from passive historical data-based ac-tivities to active real-time data driven operations [1].

Research in energy consumption analysis is traditionally con-ducted in three distinct areas based on the time scale of inves-tigation; short, medium and long term. A short-term forecastattempts to predict several hours or days in advance, medium-

Manuscript received April 06, 2011; accepted May 15, 2011. Date of publi-cation June 20, 2011; date of current version August 10, 2011. Paper no. TII-11-207.

D. De Silva, X. Yu, and G. Holmes are with the Platform TechnologiesResearch Institute, RMIT University, Melbourne VIC 3001, Australia (e-mail:[email protected]).

D. Alahakoon is with the Cognitive and Connectionist Systems Laboratory,Monash University, Victoria 3800, Australia.

Digital Object Identifier 10.1109/TII.2011.2158844

term predictions focus on weeks to months and long-term fore-casting looks at years. Many research efforts have focused onshort-term demand forecasting [2]–[4], compared to mediumand long-term prediction [5], [6].

Techniques adopted for load forecasting can be further clas-sified into model-based techniques, rule-based systems (expertsystems) and nonlinear neural learning systems. Model-basedtechniques present the classical approach, where curve-fittingprocedures [7], linear regression, multiplicative autoregressivemodels [8], and state-space models have been used mainlydue to transparency as the end-user is able to examine itsoperational behavior. Rule-based systems have primarily fo-cused on short-term forecasting as explicated in [9] and [10].Several research endeavors have treated consumption analysisas a knowledge discovery problem using intelligent techniques[11]. Both forms of learning, supervised and unsupervised,have been adopted in these studies [12], [13]. In [14], theauthors proposed the use of unsupervised learning based onthe SOM algorithm for three tasks, classification, filtering andidentification, of customer load patterns. Clustering has alsobeen successful in other industry applications of data streammining, such as [15] and [16].

In an electricity meter environment, the first constraint to ac-curate analysis is the voluminous, transient and randomly or-dered meter readings continuously transmitted by multiple me-ters to a central location. Second, given the nature of data thereexists a high likelihood of repeating patterns; such as daily andweekly cycles. Therefore, it is appropriate for the pattern ex-traction technique to be incremental and continue to accumu-late new information based on past patterns. These two key con-straints introduce new dynamics to the problem domain and theyare not adequately addressed by conventional data mining andsummarization methods.

The proposed framework overcomes these constraints usingtwo novel features. They are an interim data summarizationcomponent to handle the dynamics of the data stream and anincremental learning and knowledge accumulating technique toextract patterns from large volumes of continuously receiveddata. Outcomes from the two components are also integratedusing fuzzy pattern matching for analysis and prediction.

This paper is organized as follows. Section II delineates theISPC framework. It elaborates on the main functions; interimsummarization, incremental learning, fuzzy fusion of inter-mediate results, and prediction from final results. Section IIIpresents the results of experiments conducted using meterdata collected over an year from RMIT University premises.Section IV outlines future directions and concluding remarks.

1551-3203/$26.00 © 2011 IEEE

400 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 7, NO. 3, AUGUST 2011

Fig. 1. Interim summarization and pattern characterization (ISPC) framework.

II. THE ISPC FRAMEWORK

As stated in Section I, the ISPC framework is composed oftwo functions; interim summarization and incremental learning.An illustration of the two functions with association of corre-sponding outcomes is shown in Fig. 1.

Interim summarization has been used in the past work fordata analysis [17]. In ISPC, this feature generates online di-gests of selected segments of the data stream based on accu-mulation in a data warehouse. This also ensures critical digestsare noticed and further analyzed with incremental learning out-comes. Contrary to its general use, data is not accumulated in thewarehouse but discarded after a predetermined time period thatdoes not affect its storage capacity. In the incremental learningfunction, the IPCL algorithm characterizes patterns existing instream data. This actuates offline in an incremental manner. Thecharacterized patterns are accumulated in a columnar structureso that past learning outcomes are not lost and new learning oc-curs in the context of past outcomes. A formal description of anincremental learning environment is useful to justify the needfor two separate functions and the consolidation of outcomesfrom each.

An attribute represents a qualitative or quantitative domainof interest in the learning environment. A collection of attributes

are generally used to identify the learningenvironment. Attribute can take values from a finite set

predefined for each .A learning environment can be represented by a collection

of environment variables; attributes , where is a finiteset; . Instances of the learning environment,

, are defined using values taken by attributes . Thereby,becomes a finite set of such instances; .

Environment instances introduce spatial complexity to thelearning environment.

The data space of an instance is generated by the Carte-sian product of the collection of attribute-value pairs ex-isting in the corresponding instance .

.Data space vectors are the individual points in , represented

by the attribute-value pairs . Each individual state is rep-resented by a data space vector. Thereby, .

An incremental learning environment denotes a collec-tion of learning environments; . Fluctua-tions and variations of environment instances generate dy-namism in the learning environment. This occurs over a pe-riod of time, . Thereby, uniquely identifies learning en-vironment in . Dynamism over time introduces tem-poral complexity to the learning environment. An incremental

learning environment can also be represented in terms ofdata space vectors as

(1)

A. Incremental Learning

The incremental learning feature of the ISPC frameworkis implemented by the IPCL algorithm. It is developed basedon the Incremental Knowledge Acquiring and Self-Learning(IKASL) algorithm.

The IKASL algorithm was designed to address limitations toincremental learning; the issues of catastrophic forgetting, sta-bility-plasticity dilemma and lack of knowledge representation.It extends dynamic self-organization with the novel topologypreserving guidance method to incrementally learn new patternswithin the context of past learning outcomes. It also featuresthe accumulation of learning outcomes in an evolving columnarstructure. This columnar structure contributes to incrementallearning and scales down the combinatorial increase of knowl-edge uncovered by the continuous learning process. In [18], theIKASL algorithm was applied to large volume text mining andin [19], it was coupled with a supplementary real-time onlineclustering operation and used for demand forecasting. The sameadvantages of IKASL are maintained in the IPCL algorithm asit incrementally learns and accumulates characterized patterns.In the ISPC framework, The IPCL algorithm does not require anonline clustering operation as the interim summarization func-tion addresses stream constraints.

Dynamic topology preservation in IPCL facilitates un-constrained learning using a structure-adapting feature mapintroduced in [20]. This is in contrast to fixed-structure featuremap algorithms (e.g., SOM), which use a fixed grid for statictopology preservation. Structure adaptation ensures naturalgroupings in the data are exposed. Structure adaptation iscontrolled by total quantization error and a predefinedvalue, the spread factor (SF) [20]

(2)

When exceeds SF the network structure is adapted to suitthis change, either by growing new nodes or by distributing theerror to neighboring nodes. Despite structural adaptation, dy-namic topology preservation is unable to maintain continuity oflearning.

The topology preserving guidance technique actuates incre-mental learning in the IPCL algorithm. Taking weight vectoras the representative vector, learning outcomes from dynamictopology preservation can be shown as an integration

(3)

where is the input data space with records,the probability distribution of input vectors and as

the learning rate.The topology preserving guidance technique maintains an in-

cremental learning behavior for period by sustaining learning

DE SILVA et al.: A DATA MINING FRAMEWORK FOR ELECTRICITY CONSUMPTION ANALYSIS FROM METER DATA 401

Fig. 2. Columnar structure generated by the IPCL algorithm.

outcomes from the previous period as the gener-alization layer for learning in period . The followingsimilarity measurement is used to determine the which isclosest to input vector :

(4)

where is an input vector from period is a node fromthe generalization layer and is the set of node indexesin .

In this manner, guides new inputs based on the learningoutcomes from . In learning, inputs assigned to each gen-eralization node will grow from into individual topolog-ical maps. Unassigned inputs signify previously unseen vectors,these have to be learned anew. A separate map is used to deter-mine patterns in this set of new inputs.

The generalization layers produced from learning at eachphase are accumulated in a columnar structure. The columnarstructure is an evolving hetero-hierarchy of learning outcomesgenerated from each phase. It is able to accumulate patternscharacterized by the incremental learning process of the IPCLalgorithm. A structural view of the columnar organization isshown in Fig. 2.

The complete algorithm is given below. Notations are asfollows.

th learning phase;

th generalization phase;

th output node in th learning phase, outputnodes are created and grown during learning;

th hit node in th learning phase, hit nodes arecluster centers in the feature map which developinto aggregated representations;

th node in the th neighborhood of hit node;

th aggregate node in th generalization phase,aggregate nodes form the generalization layer,contributing to incremental learning;

learning rate for weight adaptations;

spread factor, controls growth of the dynamicfeature map;

total quantization error of node ;

proximity matrix; holds proximity values ofneighborhood nodes to the corresponding hitnode.

• I. Initialization.1) Determine SF, learning rate , neighborhood size for

weight adaptations and neighborhood size for hit nodegeneralization.

2) Initialize weight vectors of the four starting nodes ofthe feature map with random values.

• II. Incremental learning.1) Feed input vector, into the feature map.2) The node closest to is identified using Euclidean ge-

ometry, find node such that,, where are the input and weight vectors, respec-

tively, is the position vector, and is the set of naturalnumbers.


3) The network adapts to the input vector by modifyingweights in and nodes in the defined neighborhood

(5)

where the learning rate is a sequenceof positive parameters converging to zero as

are the weight vectors of node be-fore and after the adaptation and is the neighbor-hood of the winning neuron at the th iteration.The decreasing value of depends on the numberof nodes existing in the map at time .

4) Increase the error value of winner (the difference be-tween the input vector and the weight vectors).

5) When , grow nodes if is a boundary node.Distribute weights to neighbors if not.

6) Initialize new node weights to match neighborhoodnode weights.

7) Identify hit nodes with respective neighborhood nodesfor fuzzy aggregation (3).

8) Calculate the proximity matrix , where containsthe proximity of , the th node of the thneighborhood to the corresponding hit node, . Thisis calculated using, , where areweight vectors of and , respectively, and

the position vector.9) will represent fuzzy measure of , the

th node of the th neighborhood. Using these valuesthe fuzzy integral

(6)

where the finite set of inputvalues, is calculated for all positions of the weightvector of the corresponding aggregate node, . Allsuch nodes form the th generalization layer, .

• Learning from a Generalization layer.1) All input vectors are fed, in sequence, to the nodes of

the generalization layer. The aggregate node closest tothe input vector is identified using (4).

2) Winning node is assigned the input vector. For alllearning epochs to follow, node will develop a fea-ture map representing the set of assigned input vectors.

An aggregate node will start with a single node con-taining a weight vector identical to that of the aggregatenode and node growth will start from this single node. Allother parameters have the same use as with learning in

B. Interim Summarization

The main purpose of the interim summarization function isto generate online digests of stream data. The digests contributeto analysis in several forms, such as ad hoc querying, explo-ration of granularity and trend analysis. The basis for interimsummarization is the dimensional model representing the meterenvironment. Dimensional model design is a denormalizationtechnique that provides an intuitive view of data correspondingto the main areas of interest of the problem space [21]. It

Fig. 3. Dimensional model of meter environment.

is a subject-oriented structure composed of fact tables anddimensional tables. The fact table contains keys and numericalmeasures that can be summarized in terms of dimensions.Dimensions are composed of attributes and hierarchies thatidentify areas of interest in the problem domain. The facttable is linked to all dimensional tables resembling a star orsnowflake structure (Fig. 3). The data warehouse organizesattributes of each input vector into the star or snowflake schemaof the dimensional model. As data is accumulated in the ware-house, the analyst is able to conduct interim summarization.Apart from the computing cost of summarizing measures inthe fact table, interim summarization is not a process-intensiveoperation. Furthermore, the dimensional model facilitates sub-ject-oriented querying of accumulated data. Most conventionalapproaches carry a high overhead for online summarizationand interim querying of this nature. The main contributionsof interim summarization are further discussed in Section IValongside results.

C. Fuzzy Fusion

It is necessary to fuse together the outcomes from the two fea-tures. Digests extracted from interim summarization can thusbe compared with IPCL outcomes to determine similar trendsand consumption patterns that are likely to follow. A fuzzy pat-tern matching technique is used for fusion of outcomes. Fuzzypattern matching [22], [23] has been developed to address theimprecision and uncertainty of variables which need to be com-pared during a matching process. It takes into account the se-mantic representation of the input component and the patternset and provides an evaluation of the similarity between the twoas a value in [0,1]. The imprecise nature of the key aspects ofinterest in the problem space (time, day and location) furtherjustify the use of fuzzy pattern matching.

In [22], Dubois and Prade introduced two scalar measuresto estimate the compatibility between a pattern and an input

. They are a degree of possibility of matching and adegree of necessity of matching

(7)

(8)

where is the universe of discourse, is the grade ofpossibility of the input and the grade of possibility of the


Fig. 4. Multidimensional querying. (a) Mean consumption of all meters atmidday across the week. (b) Total consumption recorded by all meters inbuilding 10 on Thursdays during peak hours.

pattern. These two measures are used to match digests of interestwith IPCL outcomes that bear similarity. IPCL outcomes withstrong proximity to the selected digest can be further analyzedand used to predict the consumption record of the immediatetime period.

III. EXPERIMENTS

Electricity usage at RMIT University premises was recordedusing six meters over a period of one year from 9/1/2008 to9/1/2009. The recorded information consists of power factor(PF), the ratio of the real power flowing to the load to the ap-parent power and energy consumed measured in Kilowatt hours(Kwh). Both PF and Kwh were recorded at half hour intervalsevery day of the said year.

A electricity meter data stream can be defined as an incre-mental learning environment

(9)

where is the set of indexes of all time periods under scrutinyand is the set of indexes of all meter readings in the th timeperiod. The readings were fed into the main functions of theISPC framework and the outcomes are discussed below.

Interim summarization generates three key outcomes; ad hocquerying, exploration of granularity and trend analysis. The di-mensional model enables subject-oriented multidimensional adhoc querying. As illustrated in Fig. 4, we are able to determineintricate details with multiple dimensions such as the averageconsumption during a selected hour of the day over a week andthe total consumption of a selected building during the peakhours of a selected day.

The dimensional model is able to organize dimensional dataon granularity. In addition to the time dimension which is a nat-ural hierarchy, geography and meter dimensions pose a furtherinteresting hierarchy. Information regarding the highest energyconsuming week, month, quarter, and meter can be extractedfrom the granular structure.

The presence of trends across time at different levels of granu-larity can also be identified using interim summarization. Fig. 5

Fig. 5. Trend analysis. (a) Mean consumption recorded by meter by hour ofday. (b) Mean consumption recorded by meter by month across the year.

Fig. 6. IPCL outcomes from week by week learning. Node [9 3] learning out-come splits in week 2 and merges in week 5. Node [13 5] learning outcomecharacterizes new patterns occurring in week 2.

captures two such trends, in (a), average consumption across24 h recorded by each individual meter is shown. Three metersshow consumption pattern analogous to daily university activity,while the rest of the meters are consistently low throughout theday. In (b), average consumption is plotted across the available


Fig. 7. Histograms visualizing three IPCL outcomes. (a) Low consumption during peak hours. (b) High consumption during peak hours. (c) Very low consumptionduring off-peak hours.

months with yearly separation. It is interesting to note that al-though daily trends are similar, the yearly consumption trendof the three meters varies heavily. Meters andregister contrasting peaks and dips, possibly indicative of a sea-sonal trend.

Incremental learning occurs over time, therefore it is neces-sary to associate patterns with a predetermined time interval.The IPCL algorithm maintains flexibility of this time interval,allowing hourly, daily, weekly, and monthly associations. It gen-erates a columnar structure that characterizes patterns and main-tains continuity of learning across the time periods. Fig. 6 illus-trates a segment of this structure depicting learning outcomesfrom week 1–5 generalization layers. The learning outcomes arenamed after the coordinates of the nodes from the feature mapat each phase. This figure illustrates splits, merges from weekto week learning as well as new learning that occurs in week 2.These outcomes are distinct to the learning in week 1 thus ap-pear later in the structure.

Each learning outcome can be visualized in terms of fivehistograms. The five histograms display information about thepattern characterized by the node. The Daily graph depicts thenumber of readings at each hour of day. Meter ID graph depictsthe number of readings by meter ID (and thus location). TheWeekly graph depicts the number of readings by day of week.The Kwh and PF graphs depict the number of readings by mag-nitude of Kwh or PF measure. Fig. 7 illustrates three such cases.In learning outcome (a), meter records consumption. Itis characterized by low Kwh within a short range (35–38 Kwh)that occurs during peak hours on all week days. In (b), meter

records high Kwh (165–185 Kwh) during peak hourson all week days. In (c), very low consumption (7.5–20 Kwh)is recorded by meter during off-peak hours across thewhole week.

Repeating patterns are a common occurrence across allweeks. Fig. 8 depicts a consistent cyclic pattern existing acrossall weeks 1–46. It characterizes the general peak period con-sumption in the range of 160–185 Kwh on weekdays. This

Fig. 8. Characterization of cyclic patterns in electricity consumption.

characterization is consistent throughout the weeks as shown inDay and Kwh histograms for weeks 1 and 46.

The capacity to incrementally identify splits and merging pat-terns is also a novel feature of the IPCL algorithm. Fig. 9 illus-trates a high energy consumption pattern. Due to space restric-tions only the Kwh histograms are shown. A distinct separationis visible in week 4 learning, where 140–165 Kwh readings aresplit from 170–210 Kwh readings. This split is maintained untilweek 16, which is a vacation week. In week 16, the two pat-terns merge into one signifying low consumption levels duringthat week.

The fuzzy fusion functionality determines learning outcomesthat strongly compare with digests extracted through interimsummarization. Each digest is translated into a numericalrepresentation within [0,1] to enable a discrete fuzzy patternmatching process with similar representations of IPCL out-comes. The following discrete versions of (7) and (8) are usedhere

(10)

(11)


Fig. 9. Characterization of split and merge patterns.

Fig. 10. Proximity of IPCL outcomes from weeks 41–37 to digest � fromweek 42.

Both the digest, and IPCL outcomes were selected fromin-semester weeks at RMIT to demonstrate fuzzy fusion of out-comes. An approximation of late morning peak activity, con-sumption recorded at 11:00 am by meter on Tuesday ofweek 42 (10/14/2008), was selected as the digest along withIPCL outcomes from the most recent history (weeks 41–37).Measures and were calculated for each combination. Prox-imity of each learning outcome to the digest was derived andplotted, as shown in Fig. 10. Proximity values greater than 0.75were obtained mainly from those IPCL outcomes that identifyreadings from meter , the same meter from which wasobtained. The capacity to distinguish recordings from the samemeter significantly reduces the processing cost and time takenfor analysis.

Only consumption recorded on working days in weeks 41–37are plotted here as the digest under scrutiny is also from a weekday - Tuesday of week 42. The straightforward observation ishigh proximity from consumption recorded on past Tuesdays.However, interestingly, consumption recorded on past Thurs-days also shows high proximity. The ISPC framework is thusable to accurately determine similar trends to an ad hoc digestfrom a collection of past recordings.

The ability to predict the likelihood of the subsequent trend ofconsumption of the selected day stems from this plot.After identifying significantly proximate days from IPCL out-comes, it is now possible to compare the consumption record of

Fig. 11. Consumption records of high proximity patterns (from weeks 41–39)compared with � and �� (from weeks 42). (a) Records from Tuesdays.(b) Records from Thursdays.

those selected days with the consumption record of . Fig. 11plots the consumption record for the rest of the day for thosehigh proximity days from weeks 41–39 from IPCL outcomes.Using available data the plot for can be completed forthis experiment. The dotted line graph shows the consumptionrecord for . In Fig. 11(a), the dissimilarity of the con-sumption record on past Tuesdays can be clearly noticed andcompared with the similarity of the consumption record on pastThursdays in Fig. 11(b). The closest IPCL outcome to con-tains recordings from Thursday week 41, with a proximity of0.9963. This is clearly reflected in Fig. 11(a), where the twoconsumption records show high similarity throughout the 24 h.

IV. DISCUSSION AND CONCLUSION

This paper proposed the ISPC framework for data mining,intelligent analysis and prediction of energy consumption basedon electricity meter readings. The framework addresses thedata stream nature of such an environment with the interimsummarization feature. This feature temporarily accumulatesreadings, structured on dimensions of interest for quick extrac-tion of subject-oriented digests of consumption information.The IPCL algorithm is the second feature which empowersincremental learning and knowledge accumulation from meterreadings. It characterizes patterns in energy consumptionthrough autonomous learning and uses past learning outcomesto incrementally acquire new knowledge. The IPCL algorithmgenerates an evolving columnar structure of characterizedpatterns (IPCL outcomes). IPCL outcomes are able to identifycyclic patterns, splits and merging patterns. Fuzzy patternmatching is used to consolidate interim digests with IPCL out-comes. Numerical representation of the dimensions of interestof a selected digest are compared with those of IPCL outcomes


from past time periods. Outcomes with high proximity to thedigest are further scrutinized to aid the prediction of the energyconsumption.

Overall, the ISPC framework addresses limitations to datamining in an electricity meter environment with functionalityfor interim querying, autonomous incremental learning, fuzzyfusion and data-driven prediction.

ACKNOWLEDGMENT

The authors are grateful to Dr. W. Peng for preliminary ob-servations and support with data during the initial stages of thisresearch.

REFERENCES

[1] F. Benzi, N. Anglani, E. Bassi, and L. Frosini, “Electricity smart metersinterfacing the households,” IEEE Trans. Ind. Electron., 2011, to bepublished.

[2] D. Srinivasan, S. S. Tan, C. S. Chang, and E. K. Chan, “Parallel neuralnetwork-fuzzy expert system strategy for short-term load forecasting:System implementation and performance evaluation,” IEEE Trans/Power Syst., vol. 14, pp. 1100–1106, 1999.

[3] S. Tzafestas and E. Tzafestas, “Computational intelligence techniquesfor short-term electric load forecasting,” J. Intell. Robot. Syst., vol. 31,pp. 7–68, 2001.

[4] J. W. Taylor and R. Buizza, “Neural network load forecasting withweather ensemble predictions,” IEEE Trans. Power Syst., vol. 17, no.3, pp. 626–632, Aug. 2002.

[5] E. González-Romera, M. A. Jaramillo-Morán, and C.-F. Diego,“Monthly electric energy demand forecasting based on trend extrac-tion,” IEEE Trans. Power Syst., vol. 21, no. 4, pp. 1946–1953, Nov.2006.

[6] N. X. Jia, R. Yokoyama, Y. C. Zhou, and Z. Y. Gao, “A flexiblelongterm load forecasting approach based on new dynamic simulationtheory,” Elec. Power Energy Syst., vol. 23, pp. 549–556, 2001.

[7] J. W. Taylor and S. Majithia, “Using combined forecasts with changingweights for electricity demand profiling,” J. Oper. Res. Soc., vol. 51,pp. 72–82, 2000.

[8] G. A. N. Mbamalu and M. E. El-Hawary, “Load forecasting viasuboptimal seasonal autoregressive models and iteratively reweightedleast squares estimation,” IEEE Trans. Power Syst., vol. 8, no. 1, pp.343–348, Feb. 1993.

[9] S. Rahman and O. Hazin, “Generalized knowledge-based short-termload-forecasting technique,” IEEE Trans. Power Syst., vol. 8, no. 2, pp.508–514, May 1993.

[10] M. Tamimi and R. Egbert, “Short term electric load forecasting viafuzzy neural collaboration,” Elect. Power Syst. Res., vol. 56, pp.243–248, 2000.

[11] X. Li, C. Bowers, and T. Schnier, “Classification of energy consump-tion in buildings with outlier detection,” IEEE Trans. Ind. Electron.,vol. 57, no. 11, pp. 3639–3644, 2010.

[12] K. Metaxiotis, A. Kagiannas, D. Askounis, and J. Psarras, “Artificialintelligence in short term electric load forecasting: a state-of-the-artsurvey for the researcher,” Energy Conversion Manage., vol. 44, no. 9,pp. 1525–1534, 2003.

[13] H. S. Hippert, C. E. Pedreira, and R. C. Souza, “Neural networks forshort-term load forecasting: a review and evaluation,” IEEE Trans.Power Syst., vol. 16, no. 1, pp. 44–55, Feb. 2001.

[14] S. Verdu, M. Garcia, C. Senabre, A. Marin, and F. Franco, “Classifi-cation, filtering, and identification of electrical customer load patternsthrough the use of self-organizing maps,” IEEE Trans. Power Syst., vol.21, no. 4, pp. 1672–1682, Nov. 2006.

[15] Z. Song and A. Kusiak, “Constraint-based control of boiler efficiency:A data-mining approach,” IEEE Trans. Ind. Informat., vol. 3, no. 1, pp.73–83, Feb. 2007.

[16] J. Iglesias, P. Angelov, A. Ledezma, and A. Sanchis, “Creatingevolving user behavior profiles automatically,” IEEE Trans. Knowl.Data Eng., 2011, to be published.

[17] R. Gong, S. Huang, and T. Chen, “Robust and efficient rule extractionthrough data summarization and its application in welding fault diag-nosis,” IEEETrans. Ind. Informat., vol. 4, no. 3, pp. 198–206, Aug. 2008.

[18] D. De Silva and L. D. Alahakoon, “Incremental knowledge acquisitionand self learning from text,” in Proc. 2010 Int. Joint Conf. Neural Net-works (IJCNN), Jul. 2010, pp. 1–8.

[19] D. De Silva, X. Yu, L. D. Alahakoon, and G. Holmes, “Incremental pat-tern characterization learning and forecasting for electricity consump-tion using smart meters,” in Proc. 20th IEEE Int. Symp. Ind. Electron.,Jun. 2011, accepted.

[20] L. D. Alahakoon, S. K. Halgamuge, and B. Srinivasan, “Dynamicself-organizing maps with controlled growth for knowledge discovery,”IEEE Trans. Neural Networks, vol. 11, no. 3, pp. 601–614, 2000.

[21] R. Agrawal, A. Gupta, and S. Sarawagi, “Modeling multidimensionaldatabases,” in Proc. 13th Int. Conf. Data Eng., 1997, pp. 232–243.

[22] D. Dubois, H. Prade, and C. Testemale, “Weighted fuzzy patternmatching,” Fuzzy Sets Syst., vol. 28, pp. 313–331, Dec. 1988.

[23] L. A. Zadeh, “Fuzzy sets as a basis for a theory of possibility,” FuzzySets Syst., vol. 100, no. 1, pp. 9–34, 1999.

Daswin De Silva received the B.Sc. with first classhonours from Manchester Metropolitan University,Manchester, U.K., in 2005. He was also awarded thegold medal for best graduating student. He receivedthe Ph.D. degree from Monash University, Victoria,Australia, in 2011.

He is a member of the RMIT Platform Technolo-gies Research Institute and a member of the Cogni-tive and Connectionist Systems Laboratory, MonashUniversity. From 2008 to 2009, he was a Lecturer atthe Faculty of Information Technology (IT), Monash

University. His research interests include cognitive approaches to incrementallearning, autonomous learning algorithms, incremental knowledge acquisition,stream mining, and other data mining applications of incremental learning.

Xinghuo Yu (M’91–SM’98–F’08) received the B.Sc.and M.Sc. degrees from the University of Scienceand Technology of China, Hefei, in 1982 and 1984,respectively, and the Ph.D. degree from South-EastUniversity, Nanjing, China, in 1988.

He is now with RMIT University (Royal Mel-bourne Institute of Technology), Melbourne,Australia, where he is the Director of RMIT Plat-form Technologies Research Institute. He haspublished over 380 refereed papers in technicaljournals, books, and conference proceedings. His

research interests include variable structure and nonlinear control, complex andintelligent systems and industrial applications.

Prof. Yu received an award under the Thousand Talents Program of theChinese Government in 2010, a Chang Jiang Scholar (Chair Professor) Awardfrom the Ministry of Education of China in 2009, the 1995 Central QueenslandUniversity Vice Chancellor’s Award for Research, and was made Emeritus Pro-fessor of Central Queensland University in 2002 for his long term contributions.He is serving as an Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS

AND SYSTEMS PART I, the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS,the IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, and several otherscholarly journals. He is Vice-President (Planning and Development) of theIEEE Industrial Electronics Society and an IEEE Distinguished Lecturer. Heis also a Fellow of the Institution of Engineers Australia and the AustralianComputer Society. He is currently chairing Technical Committee on SmartGrids for the IEEE Industrial Electronics Society.

Damminda Alahakoon graduated with a Degree(first class honors) in computer science from theUniversity of Colombo, Sri Lanka, and receivedthe Ph.D. degree from Monash University, Victoria,Australia. in 2002.

His work has been published in the areas ofdata mining, clustering, neural networks, machinelearning, information fusion, incremental learning,self-organization, and cognitive systems.

Dr. Alahakoon is a member of the AustralianArtificial Intelligence Steering Committee. He was

awarded the Monash Artificial Intelligence Prize in 2000. He is closely involvedwith the Australian artificial intelligence community and Chaired the 22ndAustralasian Artificial Intelligence Conference, December 2009, in Melbourne.


Grahame Holmes (M’87–SM’03) received theB.S. degree and the M.S. degree in power systemsengineering from the University of Melbourne, Mel-bourne, Australia, in 1974 and 1979, respectively,and the Ph.D. degree in pulse width modulation(PWM) theory for power electronic converters fromMonash University, Clayton, Australia, in 1998.

In 1984, he joined Monash University, where heestablished and directed the Power Electronics Groupfor over 25 years. In 2010, he moved to RMIT Uni-versity (Royal Melbourne Institute of Technology) to

take up a professorial chair in Smart Energy. He has a strong commitment andinterest in the control and operation of electrical power converters His researchinterests include fundamental modulation theory and its application to the op-

eration of energy conversion systems, current regulators for drive systems andPWM rectifiers, active filter systems for quality of supply improvement, reso-nant converters, current-source inverters for drive systems, and multilevel con-verters. He has made a significant contribution to the understanding of PWMtheory through his publications and has developed close ties with the interna-tional research community in the area. He has published well over 150 papersat international conferences and in professional journals, and regularly reviewspapers for all major IEEE TRANSACTIONS in his area. He has also coauthored amajor reference textbook on PWM theory with Prof. T. Lipo of the Universityof Wisconsin-Madison.

Prof. Holmes is an active member of the Industrial Power Converterand Industrial Drive Committees of the Industrial Applications Society ofthe IEEE, and is a member-at-large of the Adcom of the IEEE PowerElectronics Society.

Documents

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 7, NO. …