Neural Network Applications in Environmental and Engineering Research : Prediction of Power Output from a Solar Power Generating System

(1)

Kiryu Junior College

NII-Electronic Library Service KiryuJunior College

Neural

Network

Applications

in

Environmental

and

Engineering

Research

-Prediction

of

Power

Output

from

a

Solar

Power

Generating

System-vengr*liffli

e:

ts

Vt6=i-

7V3

v

F

V-

tr

7 7'

]J

()itwa)ki{l&tzi

X7AOM"EOi}nv)

fr-

ei

g

)i

Herchel

Thaddeus

C.

*Gunma

_University

Ma

¢

hacon,

Seiichi

Shiga*

,

Faculty

of

Engineering

Abstract

A

neural network model consisting of an

input

layer,

two

hidden

Iayers,

and an output

layer

was

built

to_predictthe_power

output of a

_photovoltaic

_power

_generation

system,

Satisfactory

_predictions

of the

_power

output were obtained after the

netwerk was trained.

Keywerds:

Neural

network,

Photovoltaic

_power _generationsystem,

Neuron,

Network

architecture,

Back

_propagation

lntroduction

The biologicalnetwork of thehuman

brain

consists of

billionsof cells called neurons. These neurons carry

information through interconnectionswhich enable us to

learn,

recognize, and

_predict.

An

artificialneural network

can

be

built

from

hundreds

of simulated neurons that

mimic the

human

brain's

abiiity to

learn

from

a

_given

input

or

inputs.

Unlike

expert systems, neural networks are

incapable

of

fo11owing

rules.

However,

theycan

be

trained

by

_providing

them with

input

and output

information.

Thus,neural networks are capable ofrecognizing _patterns.

Neural

network applications range

from

_pattern and

character recognitioni'Z)

forecasting3･`),

and

decision

makingS'

.

The

authors

have

_publisheda _paper on theuse of neural networks

for

diesel

engine combustion and

performance analysisfi).

Although theultimate _goalof thisresearch

is

to_predict

the

_photovoltaic_power output

from

a ]OkW solar _power

generating system

_given

a certain date and time, the

immediate _goalat thisstage of theresearch isto develop

and traina neural network to_predictthesystem _generated

power output from temperature, humidity,and irradiance

l63

data.

Data

Acquisition

Temperature

(

℃

)

and

humidity

(%)

readings were

acquired

from

sensors attached toa wireless

data

logging

apparatus,

Irradiance

(Wlm2)

was measured

by

a

pyranometer

with a spectral range of

300

to

2800

nrn and

a

directional

response of

less

than +1-

25 W

f

m7.

The

pyranometer was also connected toawireless

data

logger.

Data

collected

by

the wireless

loggers

were sent to a communication base which was attached to a computer.

The electrical _power output _generated from a 10 kW

photovoltaicpower generationsystem was monitorecl and

recorded through a real-time dataacquisition system. Fig.

1

shows the _pyranometer

located

at the

base

of thesolar module. Specificationsof the solar module are shown in

Table

1,

DC _power _generated

by

the solar module

is

converted intoAC _power which issupplied to the utility

grid.

The

TraininglLearning

Data

Inorder to train

the

neural network,

it

has

to

be

fed

with both input and output data. The

inputs

are

iNtskmaJS(\SEee.

NII-Electronicng17I;-,

2006

(2)

NII-Electronic Library Service KiryuJumor College

Fig.1The_pyranometer[ocatedatthebaseofthesolar module.

Tabie.1 So[armodulespecification$,

Model

HIP-190B2Sanyo

RatedPowerCPmax) 190W MaximumPowerVoltage(Vpm)54.8V

MaximumPowerCurrent(Ipm)3.47A

STC:

Cell

temp. 250C,

AM

1.5,1000 Wlm2

temperature(℃

),

humidity(%),

and

irradiance(Wlm2)

.

The

output isthe

_generated

electrical

_power

(kW)

,

The

Neural

Network

A

neural netwerk consists of

input,

output, and

hidden

layersof neurons connected to each other. Here,we have 3

input

neurons, and 1output neuron representing our

input

and output

data.

The

hidden

layer

serves as a

bridge

connecting

the

flow

of

information

from

the

input

layer

to

the

output

layer.

The

neural network model

designed

here

is

the

back

_propagation

model which

is

a multi-layer

feed

forward

network thatutilizesthe

_generalized

delta

rule,

Since

we

know

the

number of

input

and output neurons,

thenext step

is

to

determine

how

many

hidden

neurons are

necessary

for

arobust network.

t

/fr

th

w

lhbo1r

Basically,

we started totraintheneural network, which

from

now on will

be

referred toas

NN,

with only one

hidden

layer.

Here,

we

have

3 input

neurons, a single

hidden

layer

with

10

neurons, and a single output,

The

neural network architecture

is

shown

in

Fig.

2. The

trainingstatus

is

_presented

below:

MtstemaJit71Ept,

eg17T-,

2oo6 164

WaitingTime

(min:sec)

6:39Learning

Tolerance

1.000 Facts

106 TrainingTolerance

O.100 Bad

18 Good

88

Bad-to-Good

Ratio

O.2 Runs(stoppedat) 15,O15

See

Fig.3.

The

training

_progress

was monitored through

histograms

oftheRMS error values as theNN trains.The

top

_graph

is

a

histogram

which shows

the

distribution

of

errorsover theentirerun, while thelower

_graph

shows the

progress

of theerrors as

the

network

is

being

trained,At

first,

largeerrors can be recognized which istypical of a

NN

as

its

starts

its

learning

_process.

Typically,

theNN

stops training until the specified acceptable tolerance is

reached.

In

the

first

test,thetrainingcontinued

fOr

_quite

some time

but

convergence was not attained.

Asa

consequence, the

NN

training

fbr

this

_particular

data

set

was stopped at

15,O15

runs or

iterations.

Even

at this

stage,

18 bad

facts

remain

fOr

a

bad-to-good

ratioof

O,20.

It

is

obvious then

that

thenetwork specifications need to

be

modified. neuron --.v-v).F L L･

>"

output

hidcteoJa'yer

Fig2 The neuralnetwork architecture

(single

hidden layer),

,t2.II@aLag-y!dttil!pmden-tqzeti

7'h

ththddlers

s

Due tothe difficultyof training the_previous network

model, one more hidden layerwas added to the network architecture. The revised NN architecture

is

shown

in

Fig.

(3)

NII-Electronic Library Service KiryuJunior College = ek/l.vf[,xS 1.UULiU B.URO ffr"¢IE,SSEOo

'

o.gog o.Ego 1.MUU

ltwts

Rumslas2ztoIEo2asltown

Fig.3Ne±work training_progressatover 15,OOOiterations

(sing]e

hiddenlayer}.

Fig,4

4.Thus, there are now two hidden layersof

10

neurons ::'et"""""'{"

each. The number of

input

and output neurons remains

the

same. The training status is

_presented

below:

WaitingTime

(min:sec)

7:06Learning

Tolerance

1.000

Facts 106 TrainingTolerance

O.100 Bad

5 Good

101 Bad-to-Good

Ratio

O.05Runs(stoppedat)

15,O15

Even

though

the

bad-to-good

ratio exhibited asignificant

improvement,

thetrainingof thisnetwork architecture still

failed

toattain convergence with thetrainingtolerance of

O,100.

A

toleranceof

O.1

means thattheoutput value must

be within 10% of the output range in order to be

considered correct. Fig. 5 shows thenetwork training

progressat 15,O15iterations.Itisobvious thatthenetwork cannot attain convergence at a training tolerance of

O,1

even

beyond

15,OOOiterations.Thus,thetraining tolerance

hasto

be

increased

toahighervalue,

3.77Tainin

wr'thtwohicldenla

ers

7)Tainin

I.Bmgo O.EOD

--:/

'x

,f'....

output r2

The Revisedneura] network architecture

(two

hiddenlayers).

Ff,:fi.re#:

foterance

increased

to

O.2

Here, the network architecture remains unchanged, while

the training tolerance was initiallyincreased to O.4.

Although convergence was rapidly attained, analysis of

the histograms of the input-hidden-output connections

showed that the network still

has

the capacity to learn.

Thus, the tolerance was decreased to 0.2,The training

sm

a.Qell o.sua 1.gao

klxlesl

Fig.5

Rufisl4S9Hto15M9Bshown

Networktraining_progressatover 15,QOOiterations

_(two

hiddenlayers,toleranceatO,1).

/"etveus'ft

30

.roDl

u.iOD Stl[t,;'eva.'･./t', a.gaD D.SOi 1.01Z t,argM Fig.6

Runs

t

toZOOshown

Networktrainingprogress atover 160 iterations

(two

hidden layers,toleranceatO.2).

165

mtsntvajlvreet.

ng17e.

2oo6

(4)

NII-Electronic Library Service

Kiryu 　Junior 　College

●

status 　

is

　_presented　

below

：

Wai

廿ng 　

Time

（min ：sec_）

4

：

02Leaming

Tolerance

1．

000

Facts

105

TrainingTolerance0

．

200 Bad

0

Good 105

Bad−

to

−Good

Ratio

0．

0 Runs

（COnVergenCe 　at_）

160

This　time

，　the　network 　trained　fairly　well 　with 　a　bad

−

to

−

good　ratio 　of　

O，

0．

　Convergence 　was 　attained 　at　l60

iterations

．

　Fig

．

6shows 　the　network 　training　

progress

　at

l60　iterations

．

　At　this　

point，

　the　training　was 　deemed

sufficient　for　testing

．

一

networl （mo （

fel

Twelve　fact　files　with 　input　and 　output 　data　were 　used 　to

test　the　trained　NN 　model

．

The　actual 　and 　the　

predicted

　output 　from　the　NN 　model were 　compared

．

　Figure　7　shows 　a　satisfactory 　agreement

between

between

the　actual　and　

predicted

　neural　network 　output

，

it

is

　clear

that　the　

perfomiance

　of　a　

NN

for

forecasting

　or　prediction

is

proportionate

　to　the　nu皿

ber

　of　training　

facts

　or　

data．

This

is

important

　to　enhance 　the　neural　networks 　

learning

process

．

Also，

　we　need 　more 　testing　

data

　to　evaluate　the

trained　

NN ．

桐生短期大学紀要

．

第17号

．

2006 166

Since

　at　this　stage 　of　the　research 　where 　the　training

data

　used 　

in

　the　

NN

learning

process

　was 　rather　

limited

，

there　

）

1

） 2 ） 3 ） 4 ） 5

）

6 References

H ．

Wechsler ：

Optimal

Sampling

Neural

Network

Performance ，　and 　Pattern　Recognition

．

World

Congress　on 　

　Said

Exchange 　Rates

．

　World　

Congress

　on　Neural　

Networks，

19951nternational

　Neural　

Network

　Society　Annual

Meeting

_，

（1）：212

−

215

，

1995

．

A ．

N ．

Networks ，

19951nternatienal 　Neural

Network

Society

Annual

Meeting，

（

1

）：

189−192，1995，

B ．Zhou，

H ．

Machacon，

　et

．

al：ニュ

ー

_ラルネットワ

ー

クの手法を用いて吸気ガス組成によるエンジン性能制御とその予測

．

日本機械学会集（B 編），

66

（

644

）：

297−302，

2000．

Acknowiedgment

The

　authors 　wish 　to　acknowledge 　the　efforts 　of　the

following

　stlldents　of　

Kiryu

Junior

College

：

C ，

Uchiyama，

H ．Eguchi，

M ，

Suzuki，

E ．

Takahashi

，

K

Matsumura，

　and

，

M ．Muraoka．

(5)

NII-Electronic Library Service Kiryu 　Junior 　College

環境

工

_{学研究}

に

お

け

る

ニュ

ー

ラ

ル

ネ

ッ

ト

ワ

ー

クア

プ

リ

ケ

ー

シ

ョン

　　　　　　　　

（

太

陽

光

発

電

シス

テ

ムの

出

力

量

の

_{予測}

_）

　　　　　　　　　

マチャコン

　

ヘ _ルチェル，

志

賀

　

聖

一

＊

　　　　　　　

＊

_群

_馬

_大

_学

_工

_{学部}

　　　　　　

要

　約

　

ニュ

ー

_{ラル}ネ_ッ _トワ

ー

クとは

_，

人間の脳の_神経細胞をモデルとしての［青報処理システムである

第17号

．

2006 　 N工工

一

Eleotronio _Library