Intelligent Dynamic Load Management Based on Solar Panel Monitoring

Gwendolin Wilke

, Marc Schaaf

, Erik Bunn

, Topi Mikkola

, Remo Ryter

, Holger Wache

and Stella Gatziu Grivas

Institute for Information Systems, University of Applied Sciences and Arts Northwestern Switzerland,

Riggenbachstr. 16, 4600 Olten, Switzerland

BaseN Corporation, Helsinki, Salmisaarenaukio 1, FI-00180 Helsinki, Finland

Keywords:

Smart Grid, Complex Event Processing, Dynamic Data Handling.

Abstract:

The Smart Grid will largely increase the amount of measurement data that needs to be processed on distribution

grid level in order to fulﬁll the promised smart behavior. Many modern information systems are capable of

handling the produced data amounts quite well. However they are usually highly specialized systems that are

costly to change or limited to very basic analytical tasks. We aim to overcome these limitations by utilizing an

optimized event processing based framework that can easily be customized to a certain application scenario.

In the paper we outline our approach by applying it to one of our motivational scenarios from the area of

intelligent dynamic load management.

1 INTRODUCTION AND

PROBLEM STATEMENT

Smart Grids promise to improve the efﬁciency, reli-

ability and sustainability of today’s power grids by

utilizing communication and information technology.

To this end, detailed information on supply, transmis-

sion and consumption is made accessible, which leads

to a massive increase in measurement data to process

in particular at the distribution grid level. This holds

most notably for dynamic load management (DLM)

approaches, cf., e.g., (Koch et al., 2009; Dam et al.,

2008).

Dynamic load management is a special form of

demand-side management that refers to the ability to

inﬂuence load curves dynamically as a reaction to

an operational state, such as transformer overload-

ing or the measured level of energy production of a

photovoltaic unit (PV). DLM will become particu-

larly important in the future due to the proliferation of

stochastic and distributed renewable energy sources

such as solar panels and wind turbines (Molderink

et al., 2010): The high ﬂuctuation in energy pro-

duction leads to higher grid volatility, which jeopar-

dizes the stability and security of power supply. DLM

can be used to balance ﬂuctuations by complement-

ing long-term storage units, such as pumped-storage

hydropower plants, cf., e.g., (Koch et al., 2009). To

implement DLM in a smart grid, it is necessary to

introduce highly ﬂexible monitoring and information

systems that are able to monitor and rapidly process

the huge amounts of data from advanced metering in-

frastructure (AMI) and external sources. To provide

the necessary decision making capability to automat-

ically trigger load shifting, complex relations in the

produced data must be analysed. Many modern in-

formation systems are capable of handling these data

volumes quite well. Yet, there is usually a trade off

between two aspects:

• the capability to monitor and process big data

amounts in near real time, and

• the amount of ﬂexibility in performing complex

analytical tasks.

One approaches that allows for a partial reconcilia-

tion of these two aspects is complex event processing

(CEP). It allows for monitoring and rapidly analysing

Big Data streams. Yet, to ensure scalability, the con-

ventional CEP processing model requires the incom-

ing event streams to be pre-partitioned. The static

pre-partitioning limits the ﬂexibility of the approach

in analysing complex relationships in the data, since

data streams can only be related to each other if they

share the same partition. This can pose a problem for

DLM solutions that rely on highly dynamic and com-

plex patterns for grid state estimation. To overcome

this problem, we propose an extension of the conven-

tional CEP framework which we call dynamic com-

plex event processing (DCEP) framework. It adds a

Wilke G., Schaaf M., Bunn E., Mikkola T., Ryter R., Wache H. and Grivas S..

Intelligent Dynamic Load Management Based on Solar Panel Monitoring.

DOI: 10.5220/0004847300760081

In Proceedings of the 3rd International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2014), pages 76-81

ISBN: 978-989-758-025-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

situation speciﬁc dynamic partition to the static CEP

partitions. We outline the proposed DCEP framework

by applying it to one of the use cases identiﬁed in the

EUROSTARS project DYNE

. Here, we consider the

issue of increased voltage variations along branches

of the distribution grid, which stem from the stochas-

tic infeed of decentralized PVs. The voltage varia-

tions can be balanced with DLM if intraday PV pro-

duction forecasts can be provided. Since they depend

on fast changing weather conditions such as moving

clouds, these patterns must be analysed to estimate

the grid state, and we show that they usually break

with any static partitioning of sensor measurement

streams.

The paper is structured as follows: Section 2

brieﬂy outlines similar approaches to dealing with big

data in DLM applications. Section 3 describes the ref-

erence use case that is used as a running example in

the remainder of the paper. The processing model of

the DCEP framework is described in detail in section

4. The paper ends with conclusions and an outline of

future work.

2 APPROACHES TO HANDLING

BIG SMART GRID DATA

The highly dynamical nature of the power grid, the

proliferation of distributed generation (DG), and the

massive amounts of measurement data available in

a smart grid pose considerable challenges to DLM

processing. This holds in particular for DLM appli-

cations that aim at avoiding switching off DG units

when the frequency or voltage deviation in the grid

exceeds the allowable threshold, cf., e.g., (Lu and

Hammerstrom, 2006). To meet these challenges,

DLM algorithms often use demand prediction to esti-

mate future grid states. Yet, accuracy of estimations

is limited, since short-term forecasting of PV produc-

tion curves is usually omitted due to the stochastic

nature of the infeed. Many DLM models employ dif-

ferential equations and optimization functions that re-

quire signiﬁcant time to be solved numerically. As

a consequence, grid state estimations can be updated

only in intervals of several minutes (insetad of several

seconds), and the according load management opera-

tions operate at a similar level of detail. E.g., (Koch

et al., 2010) employ a Model Predictive Control strat-

egy to operate controllable thermal household appli-

ances for minimizing the “balance error” of a group

of end customers at distribution grid level. In this

case, a 15-minute update cycle leaves enough time to

www.project-dyne.eu

pre-calculate the estimated grid state. As another ex-

ample, (Ringwood et al., 2001) use neural networks

for peak demand prediction, but the level of detail ac-

quired omits other cardinal points of the load curves

such as valleys, which are necessary for short time-

scale dynamic load management.

In order to allow near-real-time DLM operations

that operate at a higher level of detail, the accuracy

of grid state estimation must be increased. This can

be done by complementing the above mentioned de-

mand forecast approaches with short time scale pro-

duction forecasts, and we address this challenge in

our reference use case, cf. section 3. To achieve this

goal, the computing performance of DLM solutions

must be increased. As of today, only few approaches

exist that address the problem, and they mostly do

not directly address DLM. E.g., (Yin et al., 2011)

propose a scalable and reliable middleware layer for

real time data management in smart grids. The pro-

posed software is tailored to smart grid requirements

and thereby eliminates overheads of other data mid-

dleware such as latencies and unpredictable perfor-

mance stemming from e.g. the use of generalized

APIs. Yet, the software has not been tested with DLM

applications. Another approach is given by (Huang

and Nieplocha, 2008) who propose a parallelization

approach with high performance computing (HPC).

Unlike these solutions, the proposed DCEP approach

does not rely on costly HPC hardware. We increase

the computing efﬁciency by separating simple pattern

matching tasks (that can operate rapidly on Big Data

using conventional CEP) from dynamically changing

complex analysis tasks that only required a small por-

tion of the overall data stream for their deductions.

3 USE CASE SCENARIO

The introduction of decentralized energy production

causes increased variability of the voltage level along

distribution lines at the lowest grid layer, and, as a

result, the voltage level sometimes exceeds the maxi-

mum allowable threshold (Vovos et al., 2007), cf. Fig-

ure 1.

This problem can be solved by switching off gen-

erators when the threshold is met, by increasing the

capacity of power lines, or by installing distribution

grid transformers that are switchable on load. Our

reference use case addresses the latter scenario: here,

DLM peak shaving can be used to avoid unnecessary

switching operations of the adjustable transformer,

switching being costly and reducing equipment life

span.

Figure 1A) illustrates the problem of increased

IntelligentDynamicLoadManagementBasedonSolarPanelMonitoring

Figure 1: Increased variability of voltage levels caused by

PV infeed, A) with conventional transformers, B) with ad-

justable transformers.

variability of the voltage level when conventional

transformers are used. In rural areas, distribution

power lines are often connected to the grid by only

one feeder. If additional decentralized PV units are

connected to such power lines, the voltage level pro-

gression along a line depends not only on the power

consumption of the households, but also on the PVs’

power production. The ﬁgure shows the voltage level

progression along a power line in times of high power

consumption. The continuous line depicts the volt-

age curve when the PVs production is average. In or-

der to account for the additional PV infeed, the trans-

former’s level of infeed is lower than normal, and cho-

sen so that the resulting (continuous) voltage curve is

well within the tolerance band of 230 ± 23V. In con-

trast, the dotted line depicts the voltage curve when

PVs production is low, e.g., at nighttime or when PVs

are shaded by fog or passing clouds. In this case, PVs

do not add to the voltage level. Since the infeed level

is low and can not be changed, the dotted curve falls

below the lower threshold of 207V. Choosing a higher

infeed level does not solve the problem, since the con-

tinuous voltage curve would exceed the upper thresh-

old of 253V.

Figure1B) illustrates the situation when adjustable

transformers are used. Here, the infeed level can be

changed on load. Whenever PV production is low,

smart metering devices at the households measure a

drop in the voltage level, and report it to the au-

tonomous management unit of the adjustable trans-

former, which steps up the infeed level. The dot-

ted line shows that the resulting voltage curve stays

Figure 2: Use case scenario.

within the tolerance band.

DLM can be used to in order to avoid unneces-

sary switching operations that are, e.g., caused by fast

travelling clouds, cf. Figure 2. As a cloud passes over

a PV P installed along the power line L, it temporar-

iliy casts a shadow on P and causes the transformer T

to switch the infeed level up and then down againg in

fast succession. Whenever the time span in question

is less than than 30 minutes, DLM is well suited to

temporarily shift or reduce consumer loads on house-

hold level, e.g., by delaying the charging of electric

vehicles or by temporarily interrupting the operation

of heat pumps, cf., e.g., (Koch et al., 2009).

Yet, the intervals between switching operations

are often much smaller, and DLM algorithms must

react quickly. To provide the required rapid reaction

time, DLM algorithms must include short-term de-

mand and production forecasts. Our use case focuses

on predicting sudden drops in PV production that are

caused by traveling clouds in areas of high PV den-

sity. Here, smart meter readings of neighbouring PVs

can be used to estimate a cloud’s trajectory, which in

turn can be used to predict the time and duration of

shading of a given PV, cf. Figure 2.

Complex Event Processing (CEP) is a scal-

able method for analysing and combining Big Data

streams from multiple sources to infer events or pat-

terns that suggest more complicated circumstances.

Its goal is to identify and track meaningful events

and respond to them as quickly as possible. It hence

provides the capability to process and analyze big

data amounts in near real time, and this is a nec-

essary requirement to implement the cloud tracking

task described above. Unfortunately, CEP does not

provide sufﬁcient ﬂexibility to indeed implement the

task, since CEP requires the overall event stream to be

partitioned in advance to ensure the scalability of the

approach. The partitions can not be changed during

processing, and complex patterns can only be recog-

nized if they are composed of event streams of the

same partition. As a consequence, cloud tracking can

not be implemented with conventional CEP.

To see this, consider Figure 3, which sketches a

sensor ﬁeld consisting of PV units, partitioneded into

SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems

Figure 3: Fiexed partitions prevent cloud tracking.

four squares. A cloud that is traveling above the

PV ﬁeld may have an arbitrary trajectory that poten-

tially crosses the boundaries of any predeﬁned parti-

tion along its way. E.g., as shown in the ﬁgure, the

cloud may occlude parts of all four partitions of the

sensor ﬁeld at a given point in time. In this case, the

CEP platform can not recognize the cloud as one in-

tegral object, but instead recognizes four different ob-

jects, one in each of the four partitions. This would

not a problem per se, if the trajectories of the four

“cloud parts” could be calculated to estimate their fu-

ture movements. Unfortunately, this is not the case:

when trying to derive an object’s trajectory, a tracking

algorithm must consider changes of sensor readings in

the object’s neighbourhood. This is not always possi-

ble, since the neighbourhood of a cloud part overlaps

with another partition, preventing the algorithm from

analyzing its movement.

4 PROCESSING MODEL AND

PROTOTYPE ARCHITECTURE

In order to address the problem of prepartitioning in

conventional CEP architectures, we extend it with

the possibility of instantiating dynamically chang-

ing partitions. We refer to such a dynamic parti-

tion as a focus area, and the processing tasks per-

formed within are referred to as focused processing.

Our focused processing framework consists of three

phases: The situation indication phase is based on

normal CEP stream processing; the focused process-

ing initialization phase determines the selection of

data streams necessary for a given complex analysis

task (e.g., cloud recognition and tracking); ﬁnally, the

focused processing phase performs the the dynami-

cally changing and computationally intensive analy-

sis tasks in a separate processing environment that

is decoupled from the normal CEP processing with

its ﬁxed pre-partitioning. The prototype architecture

has been developed in a cooperation of the University

of Applied Sciences and Arts Northwestern Switzer-

land (FHNW) and the Finnish technology company

Figure 4: DCEP Architecture Overview.

BaseN

as part of the DYNE project. BaseN’s real-

time monitoring and stream processing platform pro-

vides built-in components for preprocessing the raw

event stream data from PV measurements. The pre-

processing includes semantic enrichment with con-

text information such as measurement units and de-

vice parameters, as well as ﬁlters for statistical evalua-

tion and data concentration. In particular, data quality

issues such as noisy PV measurement data, varying

latencies and packet loss observed in smart metering

systems can already be partially dealt with in the pre-

processing step.

The processing system architecture deﬁnes several

loosely coupled components that can be deployed on

multiple computers to allow for basic scalability (Fig-

ure 4). The main components are deﬁned as follows:

The DCEP Manager acts as an overseer of compo-

nents that manages the life cycle of the speciﬁed use

cases including the initial situation monitoring setup

and the coordination of the focused processing. The

Continuous Processor is conﬁgured by the DCEP

Manager to continuously monitor for use case spe-

ciﬁc situation indicators in the incoming event data.

The event data is provided by the Data Stream Man-

ager which allows each component to subscribe itself

to event streams provided by the BaseN Monitoring

platform. A specialized adapter component mediates

between the BaseN communication system and our

DCEP system. If an indicator for a complex situation

of interest is detected during normal CEP processing,

the DCEP Manager is notiﬁed, which conﬁgures the

Focused Processor component to initialize a new fo-

cused processing and to subscribe to the required data

streams. Both processing components, the continu-

ous and the focused processing, are intended to be in-

stantiated multiple times on several computers to pro-

vide scalability. The resource provisioning of the in-

frastructure is handled by the BaseN Platform and is

similar to a typical cloud infrastructure as a service

offering. To support the processing tasks with addi-

tional domain speciﬁc knowledge, the Relationship

Manager provides access to a specialized knowledge

https://www.basen.net/

IntelligentDynamicLoadManagementBasedonSolarPanelMonitoring

base speciﬁed by a RDF schema.

To illustrate the focused processing mechanism,

we outline the processing ﬂow for a cloud tracking

use case in simpliﬁed form.

Identifying relevant event streams. For every

use case, a processing model speciﬁc representation

of the use case is provided: the so-called use case tem-

plate. It refers to domain speciﬁc background knowl-

edge that is available through the relationship man-

ager. Since not all event streams are relevant for ev-

ery use case, the knowledge contained in the use case

template is used to query the relationship manager for

relevant data streams. For each of the single event

streams contained in the result set, the situation in-

dication is initialized as a normal processing task by

the DCEP Manager. In our example use case, relevant

event stream sources are the PV sensors. The query is

speciﬁed in SPARQL:

PREFIX :< h t t p : / / www. w3 . o rg / 2 0 0 0 / 0 1

r df −schema #>

PREFIX r d f :< h t t p : / / www. w3 . org / 1 9 9 9 /

02/22 − r d f −s yn t ax −n s #>

PREFIX s g :< h t t p : / / p r o j e c t −dyne . eu /

s m a r t g r i d #>

s e l e c t d i s t i n c t ? s t r e a m i d where

{? s t r e a m i d r d f : t y p e sg : s o l a r P a n e l }

Situation Indication Phase. The use case tem-

plate provides a set of rules that deﬁnes an indication

of the possible occurrence of a situation of interest.

It is evaluated against each of the relevant individual

event streams. In our example use case, a sudden pro-

duction drop in a single PV sensor can serve as an

indicator that it may be shaded by a cloud. The indi-

cation rule is speciﬁed in the event stream processing

language used by the JBoss Drools Fusion engine

pa c ka ge eu . p r o j e c t −dyne ;

i m p o r t eu . p r o j e c t −dyne . Measure ment ;

d e c l a r e Measur em ent

@role ( e v e n t )

end

r u l e ” Pr o d u c t io n D r o p ”

d i a l e c t ” mvel ”

when

Number ( $avg : d o u bl e V al u e )

from a c c u m ul a t e (

Measur ement ( $ v a l : v al u e , v a lu e >0)

o ve r window : t i me ( 1m)

from e n t r y −p o i n t

” d a t a s t r e a m 0 ” ,

a v e r a g e ( $ v a l ) )

Number ( $avg <30)

Ref to JBoss hp

t h e n

/ / t r i g g e r s i t u a t i o n

end

Focused Processing Initialization Phase. Once a

situation indication rule ﬁres , the DCEP Manager is

informed so it can determine an initial focus area for

a focused processing instance. To do this, a SPARQL

query to the Relationship Manager retrieves a set of

single event stream identiﬁers to be monitored in the

beginning of the focused processing stage. In the ex-

ample, the focus area consists of all PV sensor event

streams that are in the immediate neighbourhood of

the event stream that triggered the indication rule, cf.

Figure 3. The exact query would exceed the size of

this paper. It is of a similar structure as the previous

SPARQL query and uses the additional WGS84 Geo

Positioning vocabulary

. To avoid reduntant instanti-

ations of processes in the focused processing phase,

the DCEP Manager checks if the retrieved focus area

coincides with other focus areas of already instanti-

ated focused processing tasks. If so, the process is

aborted. Otherwise the DCEP Manager instantiates a

new focused processing instance.

Focused Processing. An algorithm is speciﬁed

that performs the actual focused processing. Since we

consider highly dynamical problems that can not be

solved with the conventional CEP pre-partitioning of

the event stream, the focus area is expected to change

dynamically over time during focused processing. To

account for this fact, the algorithm is designed as

an iterative process that allows for the adaptation of

the focus area after each iteration step. Furthermore,

it generates external notiﬁcations after each iteration

steps and checks for termination criteria. In the ex-

ample, the focused processing algorithm tries to de-

termine the border of a cloud by looking at the mea-

surement streams within the focus area. In case a co-

herent shape can not be found, the focused processing

is terminated. Otherwise the generated shape is val-

idated against background knowledge to distinguish

it from other patterns such as fog or malfunction of

PVs. If the shape’s distance from the focus area bor-

der falls below a threshold, the focus are must be up-

dated, i.e., moved and/or extended such that the iden-

tiﬁed shape has sufﬁcient padding. The DCEP Man-

ager is informed about the update and again checks

for redundancies. Each processing iteration produces

as its result the current shape of the cloud which is

published within the DYNE system to be usable by

other DLM components, such as a forecasting service

that predicts the PV production curves.

http://www.w3.org/2003/01/geo/wgs84 pos

SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems

5 CONCLUSIONS AND NEXT

STEPS

The extension of the conventinal CEP architecture by

decoupled dynamic partitions overcomes the problem

of limited ﬂexibility due to pre-partitioning. It thereby

allows for applying the proposed DCEP approach to

highly dynamic and high volume DLM scenarios. By

seperating simple CEP pattern matching tasks from

dynamically changing complex analysis tasks, we in-

crease the computing efﬁciency of DLM processing.

In a next step, the prototypical realization of the

DCEP architecture will be evaluated using several

data sets of AMI data from a smart grid test ﬁeld

in northwestern Switzerland. The provided PV mea-

surement data will be combined with simulated cloud

movement data, based on empirically tested assump-

tions about the signal behaviour of the PV devices

under the effect of cloud shading. Existing studies

show that PV arrays under partially shaded conditions

exhibit characteristic signal patterns, cf., e.g., (Ve-

muru et al., 2012). These signal characteristics can be

used for cloud recognition when combined with back-

ground domain knowledge on typical cloud shapes,

shape change, movement patterns and the geographic

charactersitics of PV locations that inﬂuence the shad-

ing patterns. Other challenges that must be addressed

in the generalized DCEP framework design such as

the fuzzy matching of time lags between situations

of interest will be addressed in future work, cf. also

(Schaaf, 2013).

ACKNOWLEDGEMENTS

Parts of the work presented here were done under Eu-

rostars Project E!7377.

REFERENCES

Dam, Q., Mohagheghi, S., and Stoupis, J. (2008). Intelli-

gent demand response scheme for customer side load

management. In Energy 2030 Conference, 2008. EN-

ERGY 2008. IEEE, pages 1–7.

Huang, Z. and Nieplocha, J. (2008). Transforming power

grid operations via high performance computing. In

Power and Energy Society General Meeting - Con-

version and Delivery of Electrical Energy in the 21st

Century, 2008 IEEE, pages 1–8.

Koch, S., Barcenas, F. S., and Andersson, G. (2010). Using

controllable thermal household appliances for wind

forecast error reduction. In Ferrera, P. M., editor, IFAC

Conference on Control Methodologies and Technol-

ogy for Energy Efﬁciency, volume 1, pages 261–266.

IFAC.

Koch, S., Meier, D., Zima, M., Wiederkehr, M., and Ander-

sson, G. (2009). An active coordination approach for

thermal household appliances - local communication

and calculation tasks in the household. In PowerTech,

2009 IEEE Bucharest, pages 1–8.

Lu, N. and Hammerstrom, D. (2006). Design considerations

for frequency responsive grid friendlytm appliances.

In Transmission and Distribution Conference and Ex-

hibition, 2005/2006 IEEE PES, pages 647–652.

Molderink, A., Bakker, V., Bosman, M. G. C., Hurink, J.,

and Smit, G. J. M. (2010). Management and control

of domestic smart grid technology. Smart Grid, IEEE

Transactions on, 1(2):109–119.

Ringwood, J. V., Bofelli, D., and Murray, F. T. (2001). Fore-

casting electricity demand on short, medium and long

time scales using neural networks. J. Intell. Robotics

Syst., 31(1-3):129–147.

Schaaf, M. (2013). Event processing with dynamically

changing focus. To be published in the Proceedings

of the Seventh IEEE International Conference on Re-

searcg Challenges in Information Science 2013 (RCIS

2013).

Vemuru, S., Singh, P., and Niamat, M. (2012). Anal-

ysis of photovoltaic array with reconﬁgurable mod-

ules under partial shading. In Photovoltaic Specialists

Conference (PVSC), 2012 38th IEEE, pages 001437–

001441.

Vovos, P., Kiprakis, A., Wallace, A., and Harrison, G.

(2007). Centralized and distributed voltage control:

Impact on distributed generation penetration. Power

Systems, IEEE Transactions on, 22(1):476–483.

Yin, J., Kulkarni, A., Purohit, S., Gorton, I., and Akyol, B.

(2011). Scalable real time data management for smart

grid. In Proceedings of the Middleware 2011 Industry

Track Workshop, Middleware ’11, pages 1:1–1:6, New

York, NY, USA. ACM.

IntelligentDynamicLoadManagementBasedonSolarPanelMonitoring