Energy EfficiencyDOI 10.1007/s12053-013-9223-9

ORIGINAL ARTICLE

Integrating renewables economic dispatch with demandside management in micro-grids: a geneticalgorithm-based approach

Ahmer Arif · Fahad Javed · Naveed Arshad

Received: 14 February 2012 / Accepted: 6 August 2013© Springer Science+Business Media Dordrecht 2013

Abstract Economic dispatch and demand sidemanagement are two of the most important toolsfor efficient energy management in the grid. Itis a casual observation that both these processesare intertwined and thus complement each other.Strategies aiming to optimize economic dispatchhave implications for demand side managementtechniques and vice versa. In this paper, wepresent a genetic algorithm-based solution whichcombines economic dispatch and demand sidemanagement for residential loads in a micro-grid.Our system collects preferences of demand datafrom consumers and costs of energy of varioussources. It then finds the optimal demand schedul-ing and energy generation mix for the given timewindow. Our evaluations show that the given ap-proach can effectively reduce operating costs in asingle- and multiple-facility micro-grids for bothsuppliers and consumers alike.

A. Arif · F. Javed · N. Arshad (B)Department of Computer Science,School of Science and Engineering,Lahore University of Management Sciences,DHA, Sector U, Lahore, Pakistane-mail: naveedarshad@lums.edu.pk

A. Arife-mail: ahmer.arif@lums.edu.pk

F. Javede-mail: fahadjaved@lums.edu.pk

Keywords Demand side management ·Economic dispatch · Smart grids · Micro-grids ·Intelligent energy management

Nomenclature

VC Cut-in speed of wind generatorVR Rated speed of wind generatorVF Cutoff speed of wind generatorC The set of consumers under the micro-gridS The set of currently available energy

suppliers to the micro-gridμx Total energy demand by consumer xμx,h Total energy demanded by consumer x at

hour hαy,h Total energy available from supplier y at

hour hλy,h Price of energy from supplier y at hour hβx,y,h Total energy allocated to consumer x from

supplier y at hour h

Introduction

Economic dispatch and demand side management(DSM) are two of the most important tools forefficient energy management in the grid. Whereaseconomic dispatch is the task of finding the rightand optimal mix of energy for the given demand,demand side management aims at managing end

Energy Efficiency

user demand to reduce cost. There are variousways to achieve this goal, the most common wayis by shifting movable loads to off-peak time ortimes of lower cost. The goal is to reduce en-ergy cost by balancing energy resources againstdemand.

It is a casual observation that both theseprocesses are interlinked. Economic dispatch triesto reduce cost according to available demand andDSM tries to manage demand to reduce cost. Itstands to reason that if both tasks are consid-ered together, where DSM and economic dispatchplanning include their interdependent constraints,a better solution may emerge.

Specifically, in renewable solution imple-mented on a small scale by the consumers, thisaspect becomes more evident. In these systems,typically, the excess low-cost energy from renew-able sources is either stored in battery systems orsold back to energy retailers up the grid. These ap-proaches miss the opportunity to take advantageof the localized aspects of micro-grids and incursignificant costs associated with often expensive,inefficient energy storage systems and transmis-sion losses upstream to the grid. By more immedi-ately utilizing this excess energy, these losses areminimized and equal or lower energy costs canbe achieved with potentially fewer distributed re-newable resources, creating savings in the installa-tion, maintenance, and operating costs associatedwith them.

In this paper, we present a strategy to inte-grate DSM and economic dispatch for renewablesin a localized setting. In our strategy, user pref-erences and generation capacity are encoded asconstraints of the system. With these constraints,an optimization objective function is constructedfrom the cost of generation for each unit for eachhour. Afterwards, a genetic algorithm is used tofind the optimal solution for the problem. Ourresults show that this integrated methodologyis up to 15 % more cost-effective than state-of-the-art independent DSM and economic dis-patch models for a neighborhood as compared toour combined DSM and economic dispatch im-plementation for an integrated neighborhoodapproach (Gudi et al. 2011).

Our study is focused on the abstraction ofhouses, renewable energy resources (RES), and

devices and their average generation and con-sumption similar to the works of Shivakumar et al.(2013), Livengood and Larson (2009), Ranade andBeal (2010), and Conejo et al. (2010). Study ofintegration of power flows in the micro-grid andintegration of different energy sources for a robustpower delivery is the next step of deploying thissystem but is beyond the scope of the currentwork.

The remaining sections of the paper are orga-nized as follows: In “Economic dispatch and DSMopportunities in smart grids,” we discuss the op-portunities for economic dispatch and DSM thathave emerged through research in smart gridsand micro-grids. In “Operational system,” theoperation system is discussed and in “Problemformulation,” the problem present in this opera-tional system is modeled and design objectivesare discussed. In “Genetic algorithm,” the ge-netic algorithm formulation and its algorithm isdiscussed. “Model implementation” specifies im-plementation details for the various steps for thegenetic algorithm, and finally, in “Evaluations andresults,” evaluation of the numerical results and asummary of the findings are presented.

Economic dispatch and DSM opportunitiesin smart grids

There are two developments in recent times whichhave provided new opportunities for economicdispatch and DSM. On one hand, smart grid ini-tiatives have provided greater surgical control ofuser-side power consumption (Coll-Mayor et al.2007; Farhangi 2010; Ipakchi and Albuyeh 2009).This is being used as an opportunity by variousresearchers to define fine-grained DSM (Conejoet al. 2010; Gudi et al. 2011; Livengood andLarson 2009; Ranade and Beal 2010) as opposedto the blanket DSM employed in traditional grids(Amjady 2007; Gellings and Chamberlin 1993;Shahidehpour et al. 2002; Weron 2006). On theother hand, distributed generation and renewablesources have enabled the emergence of micro-grids within which more efficient and exact eco-nomic dispatches are possible (Jiayi et al. 2008;Lasseter et al. 2002; Tsikalakis and Hatziargyriou2008; Yalcinoz et al. 2001).

Energy Efficiency

Demand-side management programs in tradi-tional grids usually involved manual control ofend user devices (Gellings and Chamberlin 1993).Automation is proposed by some authors, but thecontrol was blanket in the sense that the par-ticipating devices were either switched on or off(Albadi and El-Saadany 2007). There is no con-cept of intelligent planning for the operation ofthese devices.

Fine-grained DSM strategies, on the otherhand, can perform complicated automated devicescheduling based on consumer preferences andelectricity prices. Furthermore, with the adventof more dynamic forms of pricing, they have be-come increasingly flexible in order to adjust forfluctuations in the cost of electricity. These DSMtechniques usually manage energy within a sin-gle household and typically involve capitalizingon variable pricing by energy providers and in-tegration of house-based renewables for cheaperenergy utilization. Several papers have alreadybeen published which relate to such techniques.An example is the work presented by Livengoodand Larson (2009) which uses a stochastic pro-gramming solution to optimize user-side demand.However, the approach focuses on optimizingfor a single household at the granularity of in-dividual consumer devices. On the other end ofthe spectrum is the probabilistic demand shapingand peak-shaving approach outlined by Ranadeand Beal (2010) which concerns itself with reduc-ing demand by multiple consumers during peakhours but does not focus on actual scheduling ofactivities.

Economic dispatch in the scope of micro-gridshas also generated a fair amount of interest. Themain contention here has been to incorporaterenewables (Jiayi et al. 2008). While Celli andcolleagues have proposed methods to use micro-grid economic dispatches to maximize benefit inintegration with the rest of the grid, Tsikalakissuggests a system where intelligent controllersat consumer level collaborate with a micro-gridcentral controller (Tsikalakis and Hatziargyriou2008). In this case, the central controller is respon-sible for dispatch whereas local controllers are re-sponsible for DSM (Tsikalakis and Hatziargyriou2008). In addition to these solutions, Lasseter andcolleagues have also designed a framework for

economic dispatch in micro-grids (Lasseter et al.2002). A genetic algorithm-based approach byYalcinoz and colleagues can also be found in thecurrent literature (Yalcinoz et al. 2001). However,all of these systems cater to economic dispatchindependent of DSM.

Recently, a series of researchers have de-veloped economic dispatch solution incorporat-ing the demand response saving. For instance,Behrangrad et al. presented an economic dispatchmodel which uses the demand response as spin-ning reserve (Behrangrad et al. 2011). Parvaniaand Firuzabad presented an economic dispatchmodel which used the demand response planners(DRPs) as an integral part of grid and mod-eled the saving realized by DRPs in planningoptimal economic dispatch (Parvania and Fotuhi-Firuzabad 2010). However, in both the instances,economic dispatch does not actually plan house-hold consumption in the same way as Conejo et al.(2010), Livengood and Larson (2009), or Ranadeand Beal (2010) do.

As demonstrated later in our experiments,treating DSM and economic dispatch as indepen-dent modules can result in suboptimal power pric-ing across the grid. Instead, in this paper, we com-bine fine-grained DSM with economic dispatch ofmicro-grids and attempt to find an optimal sched-ule for both under a single optimization model.This way we are able to leverage the power ofDSM to arrive at better power dispatches, result-ing in a lower cost than independent DSM andeconomic dispatch systems.

We specifically consider the micro-grid devel-oped within the EU R&D MG project and pre-sented by Jiayi et al. in their survey (Jiayi et al.2008). In this micro-grid is a hierarchical structurewhere a micro-grid controller (MGCC) acts as abroker between load controllers (LC) and micro-sources (MS). In the survey, it has been shownthat LC and MS are two independent entitiesand optimize their operations independently. Weshow that a combined model at MGCC level, atleast for the renewable MS, provides a better op-timization than independent LC and MS planning.

This is under the assumption that micro-gridimplements an economic operational cycle similarto that in the work of Dimeas and Hatziargyriou(2005). In this setup, MGCC announces the

Energy Efficiency

beginning of market period and MS and LC pre-dict their production and demand, respectively,and bid for the resources under MGCC. MGCCthen resolves the market using some algorithm.When we consider DSM within this setup, wesee a point of combined optimization for MSand LC.

To our knowledge, such solutions for micro-grids are rare. Approaches that attempt to bringtogether both sides of this relationship are limitedin their scope. For example, the linear program-ming optimization solution utilized by Conejoand colleagues is only applicable to a single-load,single-supplier scenario (Conejo et al. 2010). Sim-ilarly, the particle swarm optimization of Wangand colleagues handles multiple sources but isdesigned to function on the scale of single-facilitymicro-grids only (Gudi et al. 2011). A need existsfor a more holistic approach that can encom-pass both these aspects of energy managementwhile accommodating multiplicity of suppliers andconsumers.

In this paper, we satisfy this need by present-ing a genetic algorithm-based solution for com-bined power dispatch and DSM for residentialusers in a micro-grid. Our proposed strategy has abottom–up approach where individual consumersin the micro-grid communicate their power re-quirements as a set of constraints over a neigh-borhood area network to the local micro-grid.The energy management system (EMS) of themicro-grid then collates this information and usesit in conjunction with information from availablepower supplies to generate an optimum powerdistribution schedule. This schedule satisfies themaximum number of consumer constraints at thelowest possible total cost to all the consumers.It is important to note that while this algorithmcan be applied on an individual consumer level,the primary focus is to optimize across multi-ple small-scale consumers located within a singlemicro-grid.

An important question within the scope of re-newable integration is their intermittency. It hasalready been established that integration of MSin micro-grid and for RES requires an interme-diate storage. However, the sizing of battery isproblematic and adds cost to the system. In thispaper, we assume that a storage of sufficient size is

available to mitigate the response time of MS andintermittency of RES. However, unlike other DRsystems (Gudi et al. 2011; Livengood and Larson2009), we do not consider this storage as an activesource in our planning. The storage is placed asa backup measure to safeguard against failures.Using this storage for any purpose can place oursystem in a compromised state which we wouldlike to avoid under any circumstance. Further-more, for our optimization, we only consider thehourly average wind speed and solar irradiance.This, in conjunction with the storage, provides uswith sufficient buffer to plan DR. This treatmentof renewable is consistent with similar works inthe literature (Gudi et al. 2011; Livengood andLarson 2009).

To test the feasibility of our approach versusexisting solutions, we have benchmarked it on twolevels. At the first level, we construct a powerdispatch and DSM system for a single house andcompare it with a state-of-the-art energy man-agement system (Gudi et al. 2011) at the samescale. Our results show that our genetic algorithm-based solution is up to 10 % more effective. Wethen extrapolate to multiple houses that wouldbe present in a micro-grid and show that ourtechnique is able to utilize the combination ofDSM and economic dispatch to deliver highersavings than DSM and economic dispatch systemsworking independently. In this way, our contri-bution is unique that it brings the best of thesetwo strategies together and forms a succinct so-lution for energy management at the level of amicro-grid.

Operational system

In this section, we discuss the operational sys-tem for which our technique is designed. Ourproposed strategy is for a micro-grid setup withvariety of micro-generation and renewable energyresources. Houses have some home area networkor smart devices setup which can monitor andcontrol energy consumption.

The planning module operates on a 24-h plan-ning horizon for its scheduling. It is assumed thatpricing and availability details of energy suppliesare known 24 h in advance due to day-ahead pric-

Energy Efficiency

ing. It is also assumed that consumption schedulesfor the consumers are available. These schedulesare in the form of demand of energy in each hourand the elasticity bounds for movable loads.

The proposed algorithm can readily be imple-mented in the EMS of any micro-grid network.Moreover, it possesses the required speed andscalability to enable its practical use in such a case,as demonstrated later in the paper with the help ofexperiments.

Problem formulation

The objective of this approach is to determine aneffective distribution of sources of energy betweenthe multiple consumers in a micro-grid such thatthe total cost across all consumers is minimizedand all consumer-specified and design constraintsare satisfied. Specifically, we consider a daily 24-htime window and must generate a mapping ofconsumers to suppliers at the resolution of hourlyintervals for certain amounts of power.

The objective function selected for be mini-mized is defined below:

∑

x∈C

∑

y∈S

24∑

h=1

λy,h × βx,y,h (1)

Subject to:

∑

x∈C

∑

y∈S

24∑

h=1

βx,y,h =∑

x∈C

μx (2)

∑

y∈S

24∑

h=1

boole

(αy,h <

∑

x∈C

βx,y,h

)= 0 (3)

∑

x∈C

24∑

h=1

boole

⎛

⎝μx,h >∑

y∈S

βx,y,h

⎞

⎠ = 0 (4)

Constraints in Eqs. 2 and 3 define the op-erational limits of the intended system, whichshould be satisfied throughout system operationfor any feasible solution. Constraint in Eq. 2 stip-ulates that the load assigned to different housesshould be equal to the total system demand. Thisdoes not limit generation but limits the assign-ment of that generation to demands. Constraint inEq. 3 stipulate that for each hour and for each

generation unit, the energy dispatch from thatunit should be less than or equal to its projectedgeneration. Each of these <supplier,hour> tuplesis evaluated as a Boolean. The sum of the contra-positive of the previous statement is used as theconstraint. That is, we evaluate that generation islesser than dispatch. The constraint is that for allthe <supplier,hour> tuple, this statement shouldbe false. This contrapositive is helpful in making ageneric statement since otherwise the right-handside of the equation will be equal to the number ofsuppliers × hours in planning window. These twoequations in combination ensure that the energysupplied does not surpass energy demand anddoes not exceed what suppliers can provide ingiven intervals, respectively. Solutions that violatethese constraints cannot be considered valid.

Constraint in Eq. 4 defines the consumer-specified constraints which individual consumersover the micro-grid communicate to the micro-grid EMS. These constraints help ensure that anysolution meets consumer requirements by allow-ing consumers to specify how much power theyneed within a certain time interval. The timeinterval can be rigid (consumer 5 must receive5,000 kWh between 3:00 p.m. and 4:00 p.m.) orflexible to allow for cost savings via scheduling(consumer 6 must receive 10,000 kWh through any2 h between 12:00 a.m. and 5:00 a.m. in chunksof 5,000 kWh). Solutions that violate only Eq. 2can theoretically be considered valid—subject toa penalty function as described by Michalewicz(1995) that penalizes based on how many con-straints were violated and by what margin—butfor the purposes of this work, we attempt to en-sure that every consumer constraint is satisfied ifsupply allows for it—which, in many ways, is aharder problem. Consumers must be motivated toopt into DSM schemes and fulfilling their require-ments as far as possible plays an essential role inthis function.

Genetic algorithm

Genetic algorithms are a particular class of evo-lutionary algorithms that are based on evolution-ary processes, such as reproduction, crossover,and mutation. An initial population of strings

Energy Efficiency

(chromosomes), which encode possible solutions(phenotypes) to an optimization problem, evolvestoward better solutions. The evolution starts froma population of randomly generated chromo-somes and happens across a prespecified numberof generations. For every generation, the fitnessof all chromosomes in the population is calculatedusing the objective function. Multiple chromo-somes are stochastically selected from the currentpopulation (based on fitness to some degree) toform the gene-pool for the next population. Thenext population is populated using a combinationof the very best solutions from the previous gen-eration (elitism) and modified chromosomes fromthe gene pool (cross-over and possibly randomlymutated). The new population is then used inthe next iteration of the algorithm. In this way,fitter solutions have a higher probability of beingreproduced in the next generation and mutationand crossover help avoid falling into the trap oflocal minima of results by ensuring that the chro-mosomes do not become too similar.

The computational procedure used in the ap-proach is laid out as follows:

Algorithm 1 DSM–economic dispatch plannerStep 1: Input supplier, consumer, and consumer

constraint tuples and specify genera-tion shift factors: population size, num-ber of generations, mutation probability,elitism percentage, and tournament size.

Step 2: Pseudo-randomly generate the initialpopulation.

Step 3: Rank chromosomes using the fitnessfunction described in Weron (2006).

Step 4: Apply elitism by reserving slots in thenext generation for the highest scoringchromosomes of the current generation.

Step 5: Build gene pool for the next generationby the Tournament Selection scheme.

Step 6: Apply Crossover and Mutation opera-tors on gene pool members to repopulatethe next generation.

Step 7: Repeat steps 3–6 until some terminatinggeneration number is reached or timelimit is exceeded.

Step 8: Output the fittest chromosome as theproposed optimum solution.

Model implementation

In this section, we will discuss the details of eachstep of our genetic algorithm implementation. Wewill discuss steps 1 to 6 since steps 7 and 8 aretrivial.

Input representation

Generation shift factors are specified as simplereal number values. Supplier information is mod-eled as a 3-tuple which holds the supplier iden-tifier, energy retail price, and a list of how muchkilowatts of energy is available for each hour ofthe 24-h window in question. Consumer informa-tion is modeled as a simple 2-tuple that containsthe consumer identifier and total energy required.Similarly, consumer constraints are modeled as a5-tuple holding a consumer identifier representingwhom the constraint applies to, the amount ofenergy to be delivered, and the range of hoursit must be delivered in. The fifth field denotesthe basic quantity of energy the algorithm shouldallocate when handling the constraint. It makesno sense to allocate 9 kWh in 1 h and 1 kWh inthe next if the constraint was built to model thefunctioning of a pair of appliances that required5 kWh each and could be scheduled in separatehours or together. Specifying 5 in this field wouldensure that only useful allocations are made.

Phenotype representation

A fundamental aspect of genetic algorithms(GAs) is the encoding of solutions appearing inthe population. This encoding, along with the as-sociated decoding to return to the natural problemspace, is essential to the GA operations. Eachphenotype consists of a set of elements calledgenes that are 4-tuples holding the consumer iden-tifier, the supplier identifier, the energy allocated,and the hour in question (1–24). When encodedas a chromosome, each of these fields is repre-sented as a real number. While binary encoding istraditionally the most common method, we haveutilized the real-valued representation schemefor the benefits it offers in numerical functionoptimization (Gomez-Villalva and Ramos 2003;

Energy Efficiency

Philpott and Pettersen 2006). Using real-numberencoding increases the efficiency of the GA sinceconversion of the chromosomes to binary is nolonger needed, avoids loss in precision due todiscretization to binary, requires less memory asit relies on the computer’s internal floating-pointarchitecture, and offers the choice to use a greatervariety of genetic operators (Borenstein 2005).

Generating the initial population

One open problem faced by evolutionary algo-rithms that handle constraints for numerical opti-mization problems such as this one is that of build-ing an initial population of chromosomes that isactually feasible at all. When the fitness function iscomplex and noncontinuous in combination witha search space that is relatively large, GAs canstruggle. This is the case here where a solutionmust always fulfill Eqs. 2 and 3 to be valid at alland must fulfill Eq. 4 in order to maintain thequality of service threshold at acceptable levelsfor consumers. The cardinality of these consumer-specified constraints can reach over 200 for a rela-tively large micro-grid. Every consumer has fixedbut different minimum consumption for everyhour of the day along with variable energy re-quirements that lie within specific time intervals.Generating even a single solution randomly thatis capable of satisfying each of these constraintscan be time-prohibitive.

To overcome this challenge, our approach usesa pseudo-random technique that ensures user-specified constraints are satisfied while generatingan initial population of size N:

Algorithm 2 Population generationStep 1: Select the first constraint from the list of

user-specified constraints.Step 2: Randomly select an hour from within the

time range given in the constraint.Step 3: Pick a supplier by filtering for the set

of suppliers that are selling energy atthat hour and randomly selecting one ofthem.

Step 4: Determine how much power to allocateby randomly picking a multiple of thebasic quantity of energy defined in theconstraint (See section A) up to the total

energy required. However, if the basicquantity exceeds the total energy re-quired, randomly pick a value from 1 upto the total energy required.

Step 5: Consider the selected hour, supplier, andpower to allocate—along with the con-sumer identifier from the constraint—asa gene and append it to the chromosome.

Step 6: If the total energy required by the con-straint has not been satisfied, update thisquantity by subtracting the power allo-cated in step 4 and repeat from step 2.

Step 7: Pick the next constraint and repeat fromstep 2 until all constraints have beensatisfied.

This ensures that any generated solution sat-isfies Eqs. 2 and 4. Additional energy requiredby the consumer that was not covered by anyconstraint is then handled by generating genes re-peatedly through a completely random procedureuntil all requirements have been met. Solutionsgenerated in this fashion can still violate Eq. 3since overallocation on a particular supplier canoccur so checks should be evaluating the fitness ofa solution.

Elitism

A fraction of the fittest chromosomes Ne are guar-anteed a place in the next generation directly with-out any mutation or crossover operators applied.This helps ensure that if an optimum chromosomeis found, it remains a competitive candidate. Notethat these elitist chromosomes in the original pop-ulation are also eligible for selection and subse-quent recombination in the gene pool.

Selection scheme

After elitism has been applied, a subset of theN chromosomes is selected to be used as parentsfor the succeeding generation. It is essential thatpriority is not monopolized by chromosomes withthe highest fitness values since this tends to re-duce the diversity, often causing premature con-vergence in a local optimum. To avoid this, our

Energy Efficiency

approach uses tournament selection. Chromo-somes are inducted into (N − Ne) “tournaments”by random sampling (with replacement) from theoriginal population. The winner of each of thesetournaments (the one with the best fitness) isplaced in the gene pool as a parent for the nextgeneration of chromosomes. Stronger chromo-somes can enter the pool multiple times, but bykeeping the tournament size small, weaker chro-mosomes have a reasonable chance to be selectedas well (Holland and Mansur 2006).

Reproduction scheme

The (N − Ne) parent chromosome entries fromthe gene pool are paired together to help breedthe (N − Ne) new chromosomes needed to repop-ulate the next generation. Pairs are subjected to acrossover operator with a prespecified probabilitywhich produces two offspring. If the crossoveroperator is not applied, the offspring will be di-rect clones of the parents. Since a given popula-tion might not contain enough diversity to findthe solution via crossover operations alone, thealgorithm also uses a mutation operator on anygiven offspring with a prespecified probability inan attempt to generate novel solutions. Details onthese two operators are given below:

1. Crossover operator: If crossover does takeplace, then the two offspring are produced

according to an interchange of parts of thechromosome structure of the two parents. Weaccomplish this by one-point crossover, wherethe numbers appearing after a randomly cho-sen dividing (splice) point in the genes of thetwo chromosomes are interchanged.

2. Mutation operator: For a chromosome sub-ject to mutation, we take a randomly chosensubset of its genes. For each of these genes,mutation is accomplished by adding a small in-dependent normal to the hour at which powerwas to be supplied and switching the supplieridentifier to another randomly selected validvalue.

Evaluations and results

Experimental setup

In this section, we describe the set of experimentsdesigned to evaluate the feasibility of our ap-proach. Figure 1 shows the flow of data in the sys-tem. There are two input providers to the system.The suppliers provide their day-ahead approxi-mate availability and costs. The consumer’s smarthome system provides the planner with consump-tion constraints according to user’s preferences.The GA optimizer keeping the constraints andcost in view creates a consumption schedule for

Consumers Suppliers

GA Optimizer

Consumption requirements /

constraints

Day-ahead availability

& costs

Consumption schedule

Supply schedule

Fig. 1 Flow of data in economic dispatch–DSM system.The suppliers will provide their day-ahead approximateavailability and costs. The consumer smart home systemwill provide the system with consumption constraints. The

genetic algorithm will in return provide a plan to house-holds to consume energy and will provide plans to genera-tors for optimal dispatch

Energy Efficiency

Fig. 2 Unit commit forrenewables: Solid linewith triangle markersrepresents the daily costof power when X sets orrenewables are installedin combinedconfiguration. Dashedline represents the cost of10 renewables indistributed controllerconfiguration

the consumers and a dispatch schedule for thegenerating units (Fig. 2).

The GA optimization engine is set up withthe following parameters: We set up our geneticalgorithm with a population size of 1,000 across100 generations with a tournament size of 3,and elitism, mutation, and crossover probabilitieswere set at 0.02, 0.05, and 0.5, respectively. Thechoice of these parameters was guided by Gomez-Villalva and Ramos (2003).

The daily cost rates for consumption of electric-ity used in both experiments are part of a cost ratestructure plan obtained from Gudi et al. (2011)and shown in Table 1.

For renewable sources of energy, we utilizethe wind turbines and photovoltaic cell. Thespecifications for these sources were also acquiredfrom Gudi et al. (2011) and set as follows:

Photovoltaic cell: Area 20 m2

Efficiency 12.5 %Wind turbine: Rotor diameter 4.75 m

VC 2 m/sVR 8 m/sVF 25 m/sTurbine efficiency 35 %

Table 1 Energy rates for experiment (Gudi et al. 2011)

Time Cost (cents/kWh)

7 a.m. to 9 a.m. 0.459 a.m. to 4 p.m. 0.304 p.m. to 6 p.m. 0.376 p.m. to 10 p.m. 0.4510 p.m. to 7 a.m. 0.30

The wind speed and solar radiation datautilized during evaluations are delineated inTables 2 and 3. This information was obtainedfrom Lasseter and Paigi (2004) and Zareipouret al. (2004), respectively, and mirrors the valuesused in the case study outlined by Gudi et al.(2011). All the experiments were run on a com-modity computer with core i5 processor and 8 GBof memory.

Experiment 1: single-house distributedconfiguration benchmarking

To evaluate the feasibility of our approach ver-sus existing solutions, we benchmarked our GAagainst the state-of-the-art energy managementsolution presented in the study by Gudi et al.(2011).

The case study therein incorporates distributedrenewable resources and utilizes detailed con-sumption information of sample household appli-ances to output a consumption schedule that min-imizes total cost to the micro-grid. The properties(number of appliances, time of operation, powerconsumption, and priority) of the appliances wereset as shown in Table 4. It should be noted that ap-pliances with multiple-duty-cycle operation havevariation in their peak power consumption. Thegoal of the DSM is to reduce the cost of energyand is constrained by the practical considerationthat all the operations for the devices must becompleted within the planning window. The se-lection of device is according to the prioritiesset by the user. This is the experimental setup

Energy Efficiency

Table 2 Solar insolationvalues

Hour of day 1 2 3 4 5 6 7 8Insolation (W/m2) 0 0 0 0 0 0 56 214Production (W) 0 0 0 0 0 0 1.4 5Hour of day 9 10 11 12 13 14 15 16Insolation (W/m2) 346 703 955 1,044 989 949 899 789Production (W) 9 18 23 26 25 27 22 20Hour of day 17 18 19 20 21 22 23 24Insolation (W/m2) 581 352 73 0 0 0 0 0Production (W) 15 8 2 0 0 0 0 0

of Gudi and colleagues and is consistent withthe experimental setup designed by Livengoodand Larson (2009) and Ranade and Beal (2010).We see possible improvements in this scheme;however, we would first like to present compar-ison with the existing work to benchmark ourresults.

Based on the cost rate plan defined in Table 1,the results for optimal DSM planning using Gudiet al. (2011) and our genetic algorithm imple-mentation are shown in Table 5. There are threeresults for comparison. First is the base systemwithout DSM, and the second is when DSM isapplied but without any renewable resources. Thethird result is the savings of application of DSMusing renewable sources. The percent cost savingsfor direct operation with renewable resource inte-gration is 10.39 %.

Results

The results for operation with renewable re-sources are competitive but not immediately com-parable directly since our results do not accountfor the use of a 100 % efficient energy storagesystem that begins with 40 % state of charge at theonset of simulations in the study by Vittal (2010).We opt out of using such energy storage systemsin active DSM planning in favor of sharing excess

energy production from on-site renewable sourcesacross multiple consumer loads in the micro-gridas will be shown in experiment 2. In this way,we maximize the use of these resources duringoff-peak hours and create savings by overcomingenergy losses to storage inefficiencies and mini-mizing the installation and operational costs asso-ciated with batteries and energy transformations.

Experiment 2: distributed controllerconfiguration versus combined controllerconfiguration

Having established the feasibility of our algorithmwith respect to alternate solutions at a single-consumer level, we extended the scale to simulatea multiple-facility micro-grid. For this, we usedenergy consumption data collected from 10 houses(Table 6). Due to privacy requirements, we addeda small amount of random noise in the data whilemaking sure that the characteristics of the data arenot affected. Our comparison is between a DSMin a distributed controller configuration whereeach house has its own DSM planner against aDSM where loads of each house are combinedinto a single system. This single system acts asthe planner combining the energy produced andthe devices available for scheduling. The goal is tomaximize the renewable energy usage across the

Table 3 Hourly averagewind speeds

Hour of day 1 2 3 4 5 6 7 8Wind speed (kmph) 45.2 16.8 5.9 34.4 20.7 27.2 36.8 12.2Production (W) 0 831.25 325.53 0 831.25 0 0 831.25Hour of day 9 10 11 12 13 14 15 16Wind speed (kmph) 6.5 37.6 3 31.6 19.7 44.2 11.3 11.4Production (W) 439.74 0 31.33 0 831.25 0 831.25 831.25Hour of day 17 18 19 20 21 22 23 24Wind speed (kmph) 13.7 27.6 34.9 20.9 31.1 3.24 35.1 2Production (W) 831.25 0 0 831.25 0 42.90 0 0

Energy Efficiency

Table 4 Usageconsumption pattern ofthe user

Appliance name Time of Time of operation No. of Consumption Priorityoperation without DSM appliances (W)

3,500Air conditioner Full day Full day 3 2,500 2

1,500Clothes dryer 120 min 9 p.m.–11 p.m. 1 650 6Dishwasher 180 min 9 a.m.–10 a.m. 1,200 4

9 p.m.–11 p.m. 1 1,200800

Refrigerator Full day Full day 2 600 1450

Pool pump 240 min 4 p.m.–8 p.m. 1 2,000 7Washing machine 60 min 8 p.m.–9 p.m. 1 800 5

6 a.m.–11 a.m. 700Water heater Full day* 11 a.m.–6 p.m. 1 500 3

6 p.m.–1 a.m. 700

Table 5 Experimental results: without DSM is the basesystem running without any renewables and without apply-ing any DSM

Scenario Cost (cents/day)

Benchmark Experiment

Without DSM 131 –With DSM 88.7 80.35With DSM & 22.5–26.9 23.47

renewable sources

With DSM is a system with a DSM but no renewable.With DSM and renewable is when houses are fitted withrenewable sources of energy, a DSM is applied to managethe energy loads

Table 6 Independent run DSM results

House no. Cost (cents/day)

1 20.499662 34.800373 10.45514 20.642985 21.054566 13.894497 39.527338 25.224959 29.8226710 12.23465

Total 228.15676

micro-grid and avoid storage or inefficient backselling.

To conduct this experiment, we executed ouralgorithm for each of the houses individually andadded the values to obtain the total cost for theneighborhood. Table 7 lists the cost of power incents per day for each house while each house usesthe DSM strategy of experiment 1. Each housepossesses its own wind turbine and photovoltaiccell. The specifications for these were the same asthose used in experiment 1. We use this result asthe benchmark and compare it against our com-bined configuration model.

In the combined model, as compared to the pre-vious one, the houses could share excess energyfrom their renewable resources with each other.In this way, we leverage the combined energy pro-duction and larger device management domain to

Table 7 Combined DSM results

Renewable sets Cost (cents/day)

1 519.91682 438.07973 358.31614 278.98735 254.51086 245.0057 237.19878 227.03739 224.875410 217.5591

Energy Efficiency

Fig. 3 Total demand for10 houses and availabilityof renewable over the24-h period

increase system efficiency. As an illustration of theoutcome, Fig. 3 shows the total demand for 10houses and the generation through wind and solarpanels.

Results

Our first result compares the total energy cost forboth configurations. Whereas our total cost withdistributed controllers was 228.16 cents per day,the cost using combined configuration came outto 217.6. This is a 10.56 cents or 5 % of savingsby using exactly the same devices but a differentcontroller methodology.

Discussion on unit commit problem

As a side note, we further investigated if we cansolve the unit commit problem with this setup. Westarted with no renewables and calculated a planby adding one set of renewables in each iteration.Figure 2 shows a graph of this gradual increase ofrenewables in combined configuration against thetotal cost of distributed controller configuration.We consider 10 sets of renewables in distributedcontroller configuration since combining powerof more than one renewable set is not possiblein distributed configuration. We see that for thesame demand pattern, seven renewable pairs willbe sufficient to provide energy at the same energycost as compared to each house having its ownrenewables and controllers. This means that wecan save as much at 30 % of setup cost or save

5 % of operational expenses through a combinedconfiguration controller.

Conclusion and future work

This paper provides a genetic algorithm to beintegrated into the EMS of a multi-facility micro-grid. The algorithm offers a scalable means ofminimizing energy costs to the micro-grid’s con-sumers while meeting their time-based require-ments. Simultaneously, it assists energy suppliersby determining the optimum configuration ofgenerational facilities that must be deployed tomeet the stipulated demand. The interaction takesplace on a daily basis where energy consumption,user-specified constraints, and supplier day-aheadpricing information is. A pair of case studiesdemonstrates the feasibility and effectiveness ofthe proposed algorithm to minimize total cost toboth the suppliers and consumers within a micro-grid that integrates the proposed procedure inits EMS.

There are many areas in which the algorithmmight be further improved. Constraint-capturingmechanisms are one obvious place: where inte-gration with negotiation algorithms that bridgethe gap between algorithms that operate on theindividual consumer level such as those of thestudy by Livengood and Larson (2009) andthe larger multi-facility micro-grid level wouldmake the overall system more dynamic in theface of changing consumer behavior. Similarly,

Energy Efficiency

ensuring fairness to individual consumers couldbe handled in a more sophisticated process usinga tiered-pricing schema from individual suppliersafter distributing less-expensive energy resourcesevenly. More lenient penalty functions for notmeeting user-specified constraints are possible aswell—with users being better able to prioritizecost-cutting measures over device usage and viceversa. Finally, an obvious next step is to deploythe algorithm on a small multi-facility micro-grid,validating that it behaves as expected in a realnetwork environment before moving toward testdeployments with actual consumers.

Acknowledgements We wish to thank our donors theNational ICT R&D Fund of Pakistan, Higher EducationCommission of Pakistan, and Department of ComputerScience at LUMS for partially funding this work.

References

Albadi, M.H., & El-Saadany, E.F. (2007). Demand re-sponse in electricity markets: An overview. In Powerengineering society general meeting, 2007 (pp. 1–5).IEEE

Amjady, N. (2007) Short-term bus load forecasting ofpower systems by a new hybrid method. IEEE Trans-actions on Power Systems, 22(1), 333–341.

Behrangrad, M., Sugihara, H., Funaki, T. (2011). Effect ofoptimal spinning reserve requirement on system pol-lution emission considering reserve supplying demandresponse in the electricity market. Applied Energy,88(7), 2548–2558.

Borenstein, S. (2005). The long-run efficiency of real-timeelectricity pricing. The Energy Journal, 26(3), 93–116.

Coll-Mayor, D., Paget, M., Lightner, E. (2007). Futureintelligent power grids: analysis of the vision in theEuropean Union and the United States. Energy Pol-icy, 35(4), 2453–2465.

Conejo, A.J., Morales, J.M., Baringo, L. (2010). Real-timedemand response model. IEEE Transactions on SmartGrid, 1(3), 236–242.

Dimeas, A.L., & Hatziargyriou, N.D. (2005). Operationof a multiagent system for microgrid control. IEEETransactions on Power Systems, 20(3), 1447–1455.

Farhangi, H. (2010). The path of the smart grid. IEEEPower and Energy Magazine, 8(1), 18–28.

Gellings, C.W., & Chamberlin, J.H. (1993). Demand-side management: Concepts and methods (1st edn.).Lilburn: Fairmont.

Gomez-Villalva, E., & Ramos, A. (2003). Optimal energymanagement of an industrial consumer in liberalizedmarkets. IEEE Transactions on Power Systems, 18(2),716–723.

Gudi, N., Wang, L., Devabhaktuni, V., Depuru, S.S.S.R.(2011). A demand-side management simulation plat-

form incorporating optimal management of distrib-uted renewable resources. In Power systems con-ference and exposition (PSCE), 2011 (pp. 1–7).IEEE/PES

Holland, S.P., & Mansur, E.T. (2006). The short-run effectsof time-varying prices in competitive electricity mar-kets. The Energy Journal, 27(4), 127–156.

Ipakchi, A., & Albuyeh, F. (2009). Grid of the future. IEEEPower and Energy Magazine, 7(2), 52–62.

Jiayi, H., Chuanwen, J., Rong, X. (2008). A reviewon distributed energy resources and microgrid. Re-newable and Sustainable Energy Reviews, 12(9),2472–2483.

Lasseter, R., Akhil, A., Marnay, C., Stephens, J., Dagle, J.,Guttromson, R., Sakis Meliopoulous, A., Yinger, R.,Eto, J. (2002). Integration of distributed energyresources: the certs microgrid concept. Berkeley,California: Lawrance Buckley National Lab.

Lasseter, R.H., & Paigi, P. (2004). Microgrid: A concep-tual solution. In 2004 IEEE 35th annual power elec-tronics specialists conference, 2004. PESC 04. (Vol. 6,pp. 4285–4290).

Livengood, D., & Larson, R.C. (2009). Energy box: locallyautomated optimal control of residential electricityusage. Service Science, 1(1), 1–16.

Michalewicz, Z. (1995). Genetic algorithmsnumerical op-timizationand constraints. In Proceedings of the6th international conference on genetic algorithms(pp. 151–158). San Francisco, CA: Morgan KaufmannPublishers Inc.

Parvania, M., & Fotuhi-Firuzabad, M. (2010). Demand re-sponse scheduling by stochastic scuc. IEEE Transac-tions on Smart Grid, 1(1), 89–98.

Philpott, A.B., & Pettersen, E. (2006). Optimizing demand-side bids in day-ahead electricity markets. IEEETransactions on Power Systems, 21(2), 488–498.

Ranade, V.V., & Beal, J. (2010). Distributed control forsmall customer energy demand management. In 4thIEEE international conference on self-adaptive andself-organizing systems (SASO), 2010, 27 Sept–1 Oct2010 (pp. 11–20).

Shahidehpour, M., Yamin, H., Li, Z. (2002). Market opera-tions in electric power systems: Forecasting, scheduling,and risk management (1st edn.). New York: Wiley–IEEE Press.

Shivakumar, A., Abeysekera, M., Silva, C.A., Pina, A.(2013). Unit commitment model with demand re-sponse and wind energy integration. In ICAE ’2013:Proceedings of the fifth international conference on ap-plied energy.

Tsikalakis, A.G., & Hatziargyriou, N.D. (2008). Cen-tralized control for optimizing microgrids operation.IEEE Transactions on Energy Conversion, 23(1), 241–248.

Vittal, V. (2010). The impact of renewable resources on theperformance and reliability of the electricity grid. InBridge (Vol. 40). Washington DC: National Academyof Engineering of the National Academies.

Weron, R. (2006). Modeling and forecasting electricityloads and prices: a statistical approach (the Wileyf inance series). England: Wiley.

Energy Efficiency

Yalcinoz, T., Altun, H., & Uzam, M. (2001). Economicdispatch solution using a genetic algorithm based onarithmetic crossover. In Power tech proceedings, 2001IEEE Porto (Vol. 2, p. 4).

Zareipour, H., Bhattacharya, K., Canizares, C. (2004).Distributed generation: current status and challenges.In 36th annual north americal power symp. (NAPS),Moscow, ID.