DYNAMIC RESOURCE ALLOCATION THROUGH

SEMI-STRUCTURED ADAPTATION

Sander van Splunter, Elth Ogston and Frances Brazier

Dynamic Adaptive Systems Design group, Delft University of Technology, Jaffalaan 5, Delft, The Netherlands

Keywords:

Multi-agent systems, Autonomy, Distributed energy management, Reconﬁguration.

Abstract:

Many of today’s systems are complex, distributed and networked, often situated in very dynamic environments.

Such systems are often designed to adapt to change autonomically, to manage themselves autonomously. The

Smart Energy Grid is an example of a large scale distributed system for which Distributed Energy Resource

Management is crucial. This paper proposes a loosely coordinated management structure for Virtual Power

Stations (VPS): hierarchical conﬁguration. Within VPSs individual consumers and producers each with their

own goals and responsibilities also share responsibility for collective goals such as reliability. Hierarchic self-

management combines the strengths of centralised approaches with clear contracts and dependencies, with the

strength of a fully decentralised approach within which distributed parts of a system adapt autonomously.

Agent-based simulation experiments illustrate the potential of a hierarchical approach for distribution of re-

sources within and between Virtual Power Stations as conditions change. Comparisons to centralised manage-

ment and to fully decentralised management show that performance of the hierarchical approach is close to a

centralised approach, whilst ﬂexibility and scaleability are comparable to a fully decentralised approach.

1 INTRODUCTION

In dynamic distributed systems the ability for a sys-

tem to adapt to a changing environment, whilst re-

specting global requirements for which it has been

designed, is a challenge. The Smart Energy Grid is

an example of a distributed system with Distributed

Energy Resource Management. Distributed Energy

Resource Management (DER) is designed to fulﬁl

global goals such as reliability. Many of today’s DER

systems are based on the creation of virtual groups

(Kok et al., 2005; Braun and Strauss, 2008; James

et al., 2008) of producers and consumers within which

power production and power consumption are rela-

tively balanced. In most cases dynamic virtual or-

ganisations are needed to cope with ﬂuctuation in de-

mand and production. A virtual organisation with a

large number of solar panels, for example, may pro-

vide sufﬁcient power for an ofﬁce block on cloudy

days, but will, most likely overproduce when it is

sunny. Another virtual organisation may rely on other

sources and have too little production power when the

sun shines. These virtual organisations, referred to as

“Virtual Power Stations” (VPS), may agree to jointly

regulate power exchange as a shared responsibility,

e.g. to regulate a certain reliability of service across

their overall network. To this purpose VPSs, for

example, may collectively participate in whole-sale

electricity markets but also in local markets in which

small scale producers can sell energy that would oth-

erwise be wasted. The ability to match demand to

supply within and between VPSs has many potential

advantages: distributing responsibility and overhead

within and between VPSs. There is, however, a need

for coordination between the parties involved: in par-

ticular coordination of both capacity and production

targets across groups.

This paper extends previous research on distribu-

tion of autonomy and control of resources (Ogston

and Brazier, 2009), exploring different approaches

to system architecture. An agent-based experimen-

tal comparison of (1) a centralised, (2) a fully de-

centralised and (3) a hierarchical approach to DER

management is presented. The centralised approach

is shown to be limited in scalability and ﬂexibility.

The fully decentralised approach is limited in its abil-

ity to assess the impact of local changes. The hier-

archical approach provides scalability and ﬂexibility,

supporting automated assessment of the impact of lo-

cal changes to overall (global) system behaviour.

Two sets of experiments are performed: one to as-

sess the speed of adaptation, and the second to assess

the impact of the ability to redistribute responsibil-

ity. Section 2 describes the design of the experiments,

302

van Splunter S., Ogston E. and Brazier F..

DYNAMIC RESOURCE ALLOCATION THROUGH SEMI-STRUCTURED ADAPTATION.

DOI: 10.5220/0003188303020310

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 302-310

ISBN: 978-989-8425-41-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Section 3 presents experimental results for artiﬁcial

data for solar panels. Section 4 presents the results of

the same experiments with solar panel models based

on real world data. Section 5 discusses the results and

directions for future research.

2 EXPERIMENTAL SETUP

The purpose of experimentation is to explore the ef-

fect of system design, i. e. division of control and re-

sponsibility on system performance. Section 2.1 de-

scribes the scenario, the context within which the ef-

fect of division of responsibility is explored. Section

2.2 describes the services and agents involved.

Section 2.3 described the heuristics used within

the VPS to adapt to changes in production. The data

sets used within the simulation are described in detail

in Section 2.4.

2.1 Scenario

A scenario has been designed to explore the effect

of system designs, ie o compare different ways of

dividing resources and responsibilities between one

or more virtual power stations. A group of virtual

power stations is given a (global) target amount of

power to produce. This target is divided between the

VPSs to give each a local production target. Each

VPS then selects energy resources to contribute to its

local production. The goal is to meet the global tar-

get as well as possible. Failing to produce enough

power results in failing to meet an external service

level agreement, or additional costs to acquire power

from external sources. Producing too much results in

the extra power being wasted. The energy produced

by the resources varies over time, so that any single

conﬁguration may not always be optimal.

The amount of power produced is managed in two

ways. First each VPS manages its own composition,

adding and removing resources from an external pool,

to best meet its local target. Second, production tar-

gets can be redistributed. Each VPS can have a dif-

ferent set of resources from which to choose. This

means that when one VPS is unable to meet its target

because it no longer has resources available in its ex-

ternal pool, another VPS may still be able to increase

its production using resources available in its external

pool. Redistributing production targets can result in a

better match of resources within each location.

The experimental model distinguishes VPS man-

agers and Resource managers. A VPS manager rep-

resents a VPS and is responsible for managing the

composition of the VPS, coordination between Re-

source managers within the VPS and interactions be-

tween the VPS and the external environment within

which the VPS exists. Resource managers represents

a small-scale prosumer, and are responsible for local

management of that prosumer’s energy resources. A

simple VPS consists of one VPS manager and a num-

ber of Resource managers. Each Resource manager

can only be a member of one VPS at a time. A com-

pound VPS consist of one VPS manager coordinating

multiple sub-VPS manager. A compound VPS is rep-

resented as a hierarchy of VPS managers, in which all

the leaves are resources, such the compound VPS de-

picted in Figure 1(c). Each VPS manager involved

manages a single level within the hierarchy of the

overall VPS.

For a given experiment a “map” deﬁnes which

VPS has access to which energy resources. This map

includes a geographical location for each resource,

and a geographical area covered by each VPS. A re-

source may only join a VPS if its position falls within

the VPS’s geographical area. VPS areas may be over-

lapping, meaning that some resources can be targeted

by more than one VPS.

More precisely, the scenario consists of a global

energy production target C(t) distributed among a set

of N virtual power stations, V = {v

, . . . , v

} such

that

∑

i=1

(t) = C(t) at all times, where C

(t) is the

target production of v

at time t. R = {r

, . . . , r

}

denotes the set of all resources. Each virtual power

station has a pool of energy resources, R

⊆ R from

which resources are selected. The set of active re-

sources at time t at VPS v

is denoted by A

(t) ⊆ R

where A

(t) ∩ A

(t) is assumed to be empty for all

i 6= j. Each energy resource r

has a variable energy

production e

(t). The goal of each virtual power sta-

tion is to assure that at all times its total production

meets its target production, E

(t) :=

∑

∈A

(t)

(t) ≥

(t). The goal of the system as a whole is to as-

sure that the total production of all the VPSs meets

the global target at all times,

∑

i=1

(t) ≥ C(t). This is

done either by changing the selected energy resources

A(t) := ∪

(t) and by redistributing the target pro-

duction C(t).

2.2 Simulation

The agent implementation used for the experiments

in this paper distinguishes: (1) VPS managers, that

maintain and manipulate a single level in a VPS con-

ﬁguration. (2) energy resource managers, that man-

age the available resources, in this experiment - solar

panels; (3) a resource monitor, that monitors the pro-

duction of the managed energy resources and aggre-

DYNAMIC RESOURCE ALLOCATION THROUGH SEMI-STRUCTURED ADAPTATION

303

gation of information on production within a single

level of a VPS; and (4) a directory service, that main-

tains a list of the resources that are available at any

given time, resources that could join a VPS.

Each VPS manager manages a single level within

a VPS. A VPS manager is either a simple or a com-

pound VPS manager. A simple VPS manager main-

tains a level in a VPS containing only resource man-

agers, such as the VPS manager depicted in Figure

1(a). A compound VPS manager maintains a level

in a VPS containing only other VPS managers, for

example the top two layers of VPS managers de-

picted in Figure 1(c). A complete VPS is represented

as the VPS manager of all of the levels in the VPS

together with the active energy Resource managers.

Each VPS manager performs modiﬁcations on its own

level: adding or dropping resources, or redistribut-

ing target capacities over subVPS manager. The fre-

quency at which these modiﬁcation are performed, m,

determines the stability of the overall VPS. In this

paper m = 30 minutes is used, preventing the over-

all VPS from becoming too instable. Decisions to

select modiﬁcations are made based on information

provided by a local resource monitor and heuristics

further described in Section 2.3.

An energy resource manager represents a solar

panel, to allow simple contracts to be negotiated with

VPS managers, and to exchange information on the

current output of its represented energy resource. The

solar panels used within this experiment are described

in section 2.4.2.

A resource monitor provides a VPS manager in-

formation on the current energy output of its managed

level in the VPS. To monitor the energy output of a

VPS with multiple VPS managers, as depicted in Fig-

ure 1(c), monitoring needs to distributed. Each VPS

manager within a VPS has an associated resource

monitor. The Resource monitor interacts with the en-

ergy Resource managers and with Resource managers

of subVPS managers to construct reports on current

energy output of the managed level of its associated

VPS manager. Similar to the hierarchy of VPS man-

agers, a hierarchy of resource monitors is created in

which each node in the tree periodically reports its ag-

gregate output of its underlying resources. Each VPS

manager has access to an estimate of the total current

output of its members. The frequency with which ag-

gregation updates are made, f , determines the accu-

racy of these values. In this paper f = 5 minutes is

used, which corresponds to the rate at which the out-

put data for the solar panels changes. This gives a

fairly accurate measurement, though updates are not

synchronised and thus the measurement is not precise.

Increasing f allows us to test the effect of less accu-

rate, less communication intensive, monitoring.

The directory service maintains a list all energy

Resource managers not currently assigned to a VPS.

The directory records resource capacities and loca-

tions. When a VPS manager requires a new resource

it queries the directory to get a list of available re-

source management agents. When a VPS drops a re-

source it is put back into the directory.

The simulation is programmed in AgentScope, of

which the initial version is described in (Oey et al.,

2010), with Platform 9 3/4 as backend.

2.3 Reconﬁguration Process

A simple or compound VPS manager locally adapts

the system conﬁguration using heuristics. These

heuristics are for local assessment of over- and un-

derproduction. Choice of reconﬁguration actions is

determined on the basis of these local assessments,

and consist of adding and dropping resources, or re-

distribution of target capacities.

Over- and underproduction are determined as fol-

lows. The VPS manager retrieves the current out-

put of its aggregated resources from its local resource

monitor. The difference between its production tar-

get, C

(t), and the current output of its underlying

resources E

(t) is determined. A set of managed re-

sources is underproducing if it does not meet its pro-

duction target, E

(t) − C

(t) < 0. A set of managed

resources is overproducing when its production target

is overshot with exactly or more than the production

of its smallest contributing resource, E

(t) − C

(t) ≥

min

∈A

(t)

(t).

A simple VPS manager adds a resource to com-

pensate underproduction, or removes a resource to

compensate overproduction. VPS manager can only

modify the status of a single resource during each re-

conﬁguration step. For the sake of simplicity all re-

sources in a managed location are assumed to be sim-

ilar to each other (homogeneous) The responsibility

for the production of the VPS is divided equally over

each of the resources.

A compound VPS managers relies in most cases

on the VPS managers of its managed subVPSs for re-

solve under- and overproduction. Overproduction can

always be handled by subVPSs, as underlying active

resources can be dropped. However, if a conﬁguration

manager of a subVPS is not able to resolve under-

production, i. e. it has no further available resources

to extend the production, the target production for

this subVPS needs to be reduced. The dependencies

within the level of the compound VPS are modiﬁed to

redistribute the target capacities. The heuristics used

for this redistribution are to remove 20% of the tar-

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

304

get production set for the subVPS unable to resolve

underproduction, and redistribute this evenly over the

remaining subVPSs, that are still able to extend their

production. If none of the remaining subVPSs is able

to extend its production, then the reconﬁguration pro-

cess fails, for which the conﬁguration manager of the

parent VPS is informed, that also this compound VPS

manager is unable to resolve its underproduction.

2.4 Data Sets

Two types of data are important for the purpose of

experimentation: the map dataset and the solar panel

dataset. The map datasets specify (1) the number of

resources, and VPS managers, (2) the organisation

of the VPS, in terms of relations between VPS man-

agers, (3) the the geographical locations/areas of both

resources, and VPS managers (4) the initial VPS con-

ﬁguration specifying allocation of resources, and of

responsibilities. The solar panel dataset speciﬁes the

models to determine the output of each solar panel

within the experiment.

2.4.1 Map Data Sets

Each map has in total 31 nodes, of which 7 virtual

nodes (V = {v

, . . . , v

}) representing VPS managers,

and 24 representing a solar panel (R = {r

, . . . , r

}).

The number of virtual nodes that are activated during

a simulation is dependent on the organisation of the

virtual network. In a centralised network only 1 of the

7 virtual nodes is used, in the decentralised version 4

virtual nodes are used, and in the hierarchical network

7 nodes are used. Each virtual node has an overall

target production, dependent on the speciﬁc map.

The solar panels are distributed over four geo-

graphical areas. For these experiments, the panels are

distributed evenly over the locations.

Organisational Variations. Dependent on the dif-

ferent organisations the maps are varied. Three varia-

tions are considered, depicted in Figure 1: centralised,

decentralised, and hierarchical.

In a centralised VPS only 1 of the 7 virtual nodes

is used (V = {v

}). The centralised VPS has access

to the resources in all four geographic areas, allowing

its resource pool to contain all resources: R

= R.

In the decentralised VPS 4 virtual nodes are used

V = {v

, . . . , v

}, where each virtual node has its own

separate geographical area. The solar panels are dis-

tributed evenly over these four areas, so that each area

has six resources:

• R

= {r

, . . . , r

}

• R

= {r

, . . . , r

}

VPS

panel

(a)

VPS

panel

VPS

panel

VPS

panel

VPS

panel

(b)

VPS

panel

(c)

Figure 1: Visualisations of different organisations of VPS

networks. (a) Centralised VPS network; (b) Decentralised

VPS network; (c) Hierarchical VPS network.

• R

= {r

, . . . , r

}

• R

= {r

, . . . , r

}

The hierarchical VPS is equal to the decentralised

VPS, however with the addition of three VPS nodes,

in which v

manages {v

, v

}, v

manages {v

, v

}

and v

manages {v

, v

}. As {v

, v

} only man-

age virtual nodes (compound VPS managers), their

resource pools are empty: R

= R

Allocation of Resources and Responsibilities.

This paragraph describes (1) setting production tar-

gets for the virtual nodes, and (2) determining the de-

ﬁned initial conﬁguration. Note that the initial conﬁg-

urations match the conﬁgurations depicted in Figure

The experiments uses a well-balanced map as a

base case, representing an ideal case, in which the

overall system has been designed for an equal division

of responsibilities. The initial situation/conﬁguration

is also deﬁned as such.

The production target for the central VPS equals

the overall production target (C

(t) = C(t)). The pro-

duction target at initialisation (t = 0) is: C

(0) = 24.

The initial conﬁguration is: A

(0) = {r

. . . r

The production target for the decentralised VPS is

equally distributed so that the total production meets

the overall production. The production targets at ini-

tialisation are: C

(0) = C

(0) =

6. The initial conﬁguration is: A

(0) = {r

. . . r

(0) = {r

. . . r

}. A

(0) = {r

. . . r

}. A

(0) =

. . . r

The scenario with the hierarchical VPS is similar

to the scenario with the decentralised VPS, with an

extension for production targets. The production tar-

gets of the additional VPS nodes are such that they

equal the summation of the production targets of the

VPS that each node manages: C

(0) = C

(0) = 12,

and C

(0) = 24.

2.4.2 Solar Panel Data Sets

The scenario includes 3 data sets to examine the out-

DYNAMIC RESOURCE ALLOCATION THROUGH SEMI-STRUCTURED ADAPTATION

305

put of the solar panels. Three data sets are used: two

artiﬁcial and one based on real data. The artiﬁcially

generated data sets used are stepwise upward output,

stepwise downward output. The model based on real

data is the measured output panel. The artiﬁcial pan-

els have been constructed to simulate substantial sud-

den changes.

In stepwise upward output panel, illustrated in

Figure 2(b), a constant output is produced which in-

creases in a single step to a higher constant output.

This panel switches from an average output, to high

output.

In stepwise downward output panel, illustrated in

Figure 2(b), a constant output is produced which de-

creases in a single step to a higher constant output.

This panel switches from an average output, to low

output.

In measured output panel, as illustrated in Fig-

ure 2(c), a solar panel with dynamically gradually

changing power output is simulated, based on mea-

surements of a solar panel over several hours during

noon. It starts with a low power output at the start,

rising to its maximum output at the middle of the day,

and then decreasing its output, on the remainder of

the session. The measurements have been scaled to

the range of the other panels .

(a) (b) (c)

Figure 2: Outputs of the solar panel models. (a) Stepwise

upward output panel; (b) Stepwise downward output panel;

3 EXPERIMENTAL RESULTS

This section presents the experimental results of two

sets of experiments: one to explore adaptation speed,

and one to explore the exchange of responsibilities.

For each of the system designs (organised centrally,

decentrally and hierarchically) the resulting output is

presented for the two experiments as are the points

in time at which the VPS nodes perform management

tasks, e. g. adding, dropping or redistributing respon-

sibilities.

3.1 Adaptation Speed Experiment

The adaptation speed experiment illustrates differ-

ences in speed of adaptation comparing a single au-

tonomous point of control versus multiple points of

Figure 3: Resulting output when only simple adaptations

are performed. The goal of the overall VPS is to return to

the initial total output, after a sudden increase of production.

control. This experiment is designed to not require re-

distribution of production targets. In this experiment

all solar panels are of type stepwise upward output

panel. In the initial situation each VPS is able to meet

its target output. The stepwise improvement causes

an overproduction which needs to be countered by re-

moving solar panels from the VPSs.

The overall production graphs, in Figure 3, show

a steep increase in overall production when the so-

lar panels increase production: the overall production

increases from 36 to 60. This causes an overproduc-

tion, given that the production target of each approach

is 36. The hierarchical, decentralised and the cen-

tralised approach decrease their production to better

meet their production targets.

The centralised approach has a gradual decrease

in production. It requires 4 sequential steps to reduce

its production from 60 to 40.

The hierarchical and the decentralised approach

are able to quickly reduce their production within a

single time steps. The resulting overall output to

which both approaches converge is 40. The graph

displaying the management interactions of each VPS

node, in Figure 4(b) and 4(c), shows one time steps

in which four VPS nodes concurrently perform ac-

tions to drop solar panels, in order to decrease pro-

duction. The VPS nodes performing these changes

are the lower VPS nodes that are directly managing

the solar panels (v

, v

). When the production

of the solar panels increase, all VPSs v

, v

have

an overproduction which needs to be reduced. These

VPSs have a local target of 9, and locally their pro-

duction is 15. Each of these VPS managers drop 1 of

their 3 panels.

3.2 Responsibility Exchange

Experiment

This experiment is designed such that half of the net-

work increase production, while the other half de-

creases production. The netto overall output is stable.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

306

(a) Central Events (b) Hierarchic Events (c) Decentral Events

Figure 4: Resulting events related to the adaptation speed. The centralised approach is limited to sequential modiﬁcations,

due to a single VPS manager agent, whereas the hierarchical and decentralised approach use multiple VPS manager agents.

(a) Central events (b) Hierarchic events (c) Decentral events

Figure 5: Resulting events related to the exchange of responsibilities. Half of the solar panels decrease output, and the other

half of the solar panels increase output. The VPS manager in the centralised approach does not observe a change in its overall

output, and performs no actions The VPS managers in decentralised approach only react locally. The VPS managers in the

hierarchical approach react locally and redistribute responsibilities.

Figure 6: Resulting output when responsibilities are ex-

changed. The goal of the overall VPS is to return at least

to the initial total output, after half of the solar panels de-

creased output, and the other half increased output.

However, one part of the network will underproduce

while another part overproduces. In this experiment

half of the solar panels (r1...r12) are of type stepwise

upward output and the other half (r13...r24) are of

type stepwise downward output.

In the initial situation each VPS is able to meet

its target output. The change in production causes

a shift in production at the solar panels. In the cen-

tralised VPS, the VPS manager observes no change

in the overall total output, as the decrease in produc-

tion of half of its managed resources is compensated

by an increase of the other half of its managed re-

sources. For the decentralised and hierarchical VPSs

the change in production results in a strong decrease

in the production on locations R6, R7 and an increase

in production in locations R4, R5. This production

shift results in that local production targets are not

met: locally there is either overproduction or under-

production.

The overall production graph, Figure 6 shows that

the three approaches behave differently. The cen-

tralised approach shows, as its netto production is not

affected, a stable constant output at 36 kWh. The cen-

tralised approach requires no management interven-

tions. The decentralised and hierarchical approaches

show more varying behaviour.

The decentralised VPS is unable to redistribute

production responsibilities, and is limited to meet-

ing its initial local production responsibilities. The

management interactions of the VPS nodes are visu-

alised in Figure 5(c). The VPS nodes managing loca-

tions R4, R5, each have the responsibility to produce

9kWh, which they meet at the initial situation. At

the point of change, their local production increases

from 9 kWh to 15 kWh, which leads to dropping one

of the solar panels reducing the production from 15

kWh to 10kWh, and effectively reducing most local

overproduction. The VPS nodes managing locations

R6, R7, each have the responsibility to produce 9kWh,

which they meet at the initial situation. At the point of

change, their local production decreases from 9 kWh

to 3 kWh, which leads to adding three panels each in

three consecutive steps, after which all panels in the

location are in use. These additions increase produc-

tion from 3 kWh to 6 kWh, still leaving an under-

production of 3 kWh for each of both locations. The

DYNAMIC RESOURCE ALLOCATION THROUGH SEMI-STRUCTURED ADAPTATION

307

overall production eventually results in 32 kWh, re-

sulting in an overall underproduction of 4 kWh, and

not adhering to the overall production target.

The hierarchical VPS is able to redistribute pro-

duction responsibilities, and ﬁrst adapts to its ini-

tial local production responsibilities, and, after fail-

ing to do so, redistributes the production responsibil-

ities. The management interactions of the VPS nodes

are depicted in Figure 5(b). The VPS nodes man-

aging locations R4, R5, each have the responsibility

to produce 9kWh, which they meet at the initial sit-

uation. At the point of change, their local produc-

tion increases from 9 kWh to 15 kWh, which leads to

dropping one of the solar panels reducing the produc-

tion from 15 kWh to 10kWh, and effectively reducing

most local overproduction.

The VPS nodes v6, v7 managing locations R6, R7,

each have the responsibility to produce 9kWh, which

they meet at the initial situation. At the point of

change, their local production decreases from 9 kWh

to 3 kWh, which leads to adding three panels each in

three consecutive steps, after which all panels in the

location are in use. These additions increase produc-

tion from 3 kWh to 6 kWh, still leaving an underpro-

duction of 3 kWh for each of both locations.

As it is not possible to meet the local production

targets, one of the two underproducing VPSs, in this

case v6 request a redistribution of production target to

v3. Note that both v6 and v7 are requesting redistribu-

tion of production, however, due to the distribution of

control these request will not arrive simultaneously.

The supervising VPS 3 designates part of the target

production of v6 to v7. The production target of v6

is lowered from 9kWh to 7.2 kWh, while the produc-

tion target of v7 is increased from 9 kWh to 10.8 kWh.

This is immediately followed by the request from v7

to redistribute capacity. As the supervising VPS man-

ager v3 is now aware that local redistribution is not

sufﬁcient, it escalates to the root node v1, to redesig-

nate part of the capacity of v3. This effectively assigns

part of the production target from v6 and v7 through

v3 − v1 − v2 to v4 and v5. This is done twice, af-

ter which enough of the production targets is shifted

from the locations with a lowered production, to the

locations with an increased production. There is a

small rebalancing between v6 and v7, which compen-

sates for the ﬁrst redistribution of production, before

v3 was aware that both subVPSs were unable to meet

their targets.

The overall production eventually results in 42

kWh, resulting in an overproduction of 6 kWh, but

adhering to the overall production target.

3.3 Discussion

The results of the Adaptation speed experiment show

that both the distribution of productional and decen-

tralised approach are able to converge in less time

after substantial changes due to the effect of having

multiple autonomous management nodes, able to re-

act to changes in parallel. The centralised approach is

slower in its convergence, as it has a single point of

control.

The results of the responsibility exchange experi-

ment show that in the non-centralised approach, adap-

tations are required in production targets to be able to

handle non-homogeneous behaviour in different parts

of the network. A purely decentralised approach, with

complete partitioned control, and pre-deﬁned set of

local targets (the initial set), is therefore incapable of

handling this behaviour. The hierarchical approach

that can adapt production responsibilities using only

local reasoning, supports scalable adaptivity. Non-

centralised approaches often lead to an increase in-

efﬁciency in terms of overproduction, due to the dis-

tribution of the central production target into multi-

ple distributed targets. The requirement that all local

production needs to be equal or larger than the local

production targets, causes more overproduction in the

overall system.

4 EXPERIMENTAL RESULTS

BASED ON REAL WORLD

DATA

The two experiments described in the previous sec-

tion are based on artiﬁcial data sets. These data sets

provide a means to analyse the impact of substantial

sudden change within VPSs with different organisa-

tion structures. As the artiﬁcial models for the solar

panels are not realistic, two experiments are included

using solar panel models based on measured output of

solar panels.

The main differentiating factors are that the mea-

sured output changes gradually, instead of switching

between two static output values. This section de-

scribes two experiments relating to: the adaptation

speed, and to the exchange of responsibilities with the

real data.

The ﬁrst experiment, related to the adaptation

speed, creates a situation in which enough spare ca-

pacity is available. All resources are equally dis-

tributed within the overall VPS, and the overall target

output, 24, is lower than the total output of the initial

conﬁguration. The results of this ﬁrst experiment are

depicted in Figures 7 and 8.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

308

(a) Central events (b) Hierarchic events (c) Decentral events

Figure 8: Resulting events with real world data,based on VPSs with an equal distribution of resources.

Figure 7: Resulting output with real world data, based on

VPSs with an equal distribution of resources.

The second experiment, related to the exchange

of responsibilities, creates a situation that strains the

overall system. The resources are not equally divided

over the location, and the overall target output, 43,

is set high compared to the total available capacity

within the system. The results of this ﬁrst experiment

are depicted in Figures 9 and 10.

In general, due to the gradual change, the cen-

tralised approach meets its production target during

most of both experiments. This changes during the

last phase of both experiments. The centralised ap-

proach meets its overall output target slightly more

effectively than the other two approaches. In the last

hour of the two experiments the total reduction in pro-

duction increases despite the addition of a single solar

panel.

Distribution of production targets over multiple

VPS agents in both the decentralised and hierarchical

approaches, effects the time needed to meet the target

total output. The ability to redistribute responsibilities

is clearly illustrated in the second experiment: the de-

centralised approach fails to meets its set target, the

hierarchical approach is able to recover from the un-

equal distribution of resources.

Figure 9: Resulting output with real world data, based on

VPSs with an equal distribution of resources.

5 DISCUSSION AND FUTURE

WORK

This paper focuses on coping with changes in large

distributed systems, and, more speciﬁcally, in Dis-

tributed Energy Management. This work extends re-

search described in (Ogston and Brazier, 2009), in-

troducing an agent-based experimental framework to

compare different approaches to appointing auton-

omy and control of resources for Distributed Energy

Resource Management.

The approaches for appointing autonomy and con-

trol explored in this paper are based on centralised,

decentralised, and hierarchical control. The experi-

ments explore the speed at which the overall system

reacts to changes, the ability to reason and act on the

responsibilities of the overall system.

The results show that the centralised approach is

best in handling responsibilities of the overall sys-

tem, due to its knowledge of the overall system, and

its modiﬁcations directly effect the satisfaction of the

overall responsibilities. On the other hand, a cen-

tralised approach is, by deﬁnition, limited in scala-

bility due to communication overhead.

Distributed approaches adapt the overall system

simultaneously at different nodes, resulting in a

higher ﬂexibility, and a higher speed to complete

complex modiﬁcations. Static distribution of respon-

sibilities over subparts of the system provides less

ﬂexibility.

A hierarchical approach supports redistribution

of the responsibilities of the overall system over

subparts, while also responding quickly to system

changes. It combines the strengths of the centralised

and decentralised approach. The hierarchical ap-

proach is scaleable supporting local redistribution of

responsibilities.

More general, the experiments indicate that the

distribution of autonomy and control supports better

management in large-scale distributed systems, with

hierarchical adaptation as an example.

Future work will focus on increasing heterogene-

ity of the scenario integrating other data sets based on

measurements of power output of solar panels, inte-

DYNAMIC RESOURCE ALLOCATION THROUGH SEMI-STRUCTURED ADAPTATION

309

(a) Central events (b) Hierarchic events (c) Decentral events

Figure 10: Resulting events with real world data,based on VPSs with an unequal distribution of resources.

gration of batteries, handling rapidly changing power

outputs, and integration of local power demand. As-

pects of further exploring autonomy and control will

exploring dynamic clustering of pools and VPSs, and

introducing heterogeneous strategies for VPSs.

ACKNOWLEDGEMENTS

The authors thank NGI for their support of this work,

part of the NGI project Self-Managed Dynamic Insti-

tutions in Power Grids: Sharing the Cost of Reliabil-

ity.

REFERENCES

Braun, M. and Strauss, P. (2008). A review on aggrega-

tion approaches of controllable distributed energy unts

in electrical power systems. International Journal of

Distributed Energy Resources, 4(4):297 – 319.

James, G., Peng, W., and Deng, K. (2008). Managing

household wind-energy generation. IEEE Intelligent

Systems, 23(5):9–12.

Kok, J. K., Warmer, C. J., and Kamphuis, I. G. (2005). Pow-

ermatcher: multiagent control in the electricity infras-

tructure. In AAMAS ’05: Proceedings of the fourth

international joint conference on Autonomous agents

and multiagent systems, pages 75–82, New York, NY,

USA. ACM.

Oey, M. A., van Splunter, S., Ogston, E., Warnier, M., and

Brazier, F. M. T. (2010). A framework for devel-

oping agent-based distributed applications. In Pro-

ceedings of the 2010 IEEE/WIC/ACM International

Conference on Intelligent Agent Technology (IAT-10).

IEEE/WIC/ACM. To appear.

Ogston, E. and Brazier, F. M. T. (2009). Apportionment of

control in virtual power stations. In In the proceed-

ings of the international conference on infrastructure

systems and services 2009: Developing 21st Century

Infrastructure Networks.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

310