Suitability Mapping of Solar Power Plants Using an Explainable AI-Based Approach
Abstract
The conventional approaches to identifying optimal locations for solar power plants are traditionally handled through Multi-Criteria Decision-Making (MCDM) techniques, which often suffer from subjectivity and lack of transparency. Although recent advancements in Machine Learning (ML) have offered potential solutions to MCDM-based methods, traditional ML models struggle to explain their predictions and operate without assuming that current solar power plants are in optimal locations. Thus, this research represents an integrated explainable AI approach with ML methods, aiming to enhance the transparency and comprehensibility of results and introduce a novel paradigm by employing the classified efficiency of existing plants, categorized into five classes, as the dependent variable for ML models, thereby challenging the prevalent assumption inherent in conventional ML models. Twelve independent variables selected through a literature review were used, along with the classified efficiency values, to train five ML models, namely, Random Forest (RF), Support Vector Machines (SVM), Multi-layer Perceptron (MLP), Decision tree (DT), and K-nearest neighbors (k-NN). Following the assessment of the models' accuracies, the RF model, which achieved 88% overall accuracy, was subsequently explained using SHapley Additive exPlanations (SHAP), revealing Solar Radiation (13%) and Cloud Index (12%) as the most influential variables for the resulting predictions. In comparison, Aspect (5%) was identified as the least significant parameter to the model predictions. The final solar suitability map produced with the superior RF model identified approximately 5% of the total land area in the USA as highly suitable for constructing solar power plants, ensuring optimal operational efficiency. Additionally, 55% of the land is moderately suitable for such establishments. Conversely, approximately 9.5% of the total land area, equivalent to 766,654 km2, is deemed permanently unsuitable for solar power plant construction.
Full text
i
SUITABILITY MAPPING OF SOLAR POWER
PLANTS USING AN EXPLAINABLE AI-BASED
APPROACH
Mawanane Hewa Rasanka Mangala De Sil a
ii
SUITABILITY MAPPING OF SOLAR POWER PLANTS USING
AN EXPLAINABLE AI-BASED APPROACH
Disse a ion supe ised by:
Leona do Vanneschi, PhD
NOVA In o ma ion Managemen School
Uni e sidade No a de Lisboa
Lisbon, Po ugal
Co-supe ised by:
Ma co Painho, PhD
NOVA In o ma ion Managemen School
Uni e sidade No a de Lisboa
Lisbon, Po ugal
Co-supe ised by:
Michael Gould, PhD
GEOTEC
Uni e si a Jaume I
Cas ellón, Spain
Feb ua y 20, 2024
iii
ACKNOWLEDGMENTS
I ex end my hea el g a i ude o my supe iso , P o . D . Leona do Vanneschi, o
his s ead as suppo , guidance, and in aluable insigh s du ing he s udy. Addi ionally,
I am deeply hank ul o my co-supe iso s, P o . D . Ma co Painho and P o . D .
Michael Gould, o hei aluable guidance, unwa e ing encou agemen , and
consis en suppo h oughou he s udy. Thei expe ise and mo i a ion ha e played a
pi o al ole in shaping he ajec o y o my wo k.
I exp ess my genuine g a i ude o he acul y and s a o Uni e sidade No a de Lisboa,
Uni e si y o Müns e , and Uni e si a Jaume I o hei p o ision o esou ces,
assis ance, and an en i onmen conduci e o academic explo a ion. Addi ionally, I
would like o hank he E asmus Mundus p og am o i s unding suppo owa ds my
Mas e o Science in Geospa ial Technologies.
I am hank ul o my amily in S i Lanka o hei unwa e ing suppo in e e y aspec ,
p opelling me o wa d. Las , I hank my wi e, P asadi, o he cons an encou agemen
and s ead as suppo h oughou he s udy. You mo i a ion has been indispensable in
enabling me o comple e his s udy success ully.
This hesis ep esen s he culmina ion o he combined e o s, encou agemen , and
suppo om nume ous indi iduals, and o his, I am since ely g a e ul.
i
SUITABILITY MAPPING OF SOLAR POWER PLANTS USING
AN EXPLAINABLE AI-BASED APPROACH
ABSTRACT
The con en ional app oaches o iden i ying op imal loca ions o sola powe plan s
a e adi ionally handled h ough Mul i-C i e ia Decision-Making (MCDM)
echniques, which o en su e om subjec i i y and lack o anspa ency. Al hough
ecen ad ancemen s in Machine Lea ning (ML) ha e o e ed po en ial solu ions o
MCDM-based me hods, adi ional ML models s uggle o explain hei p edic ions
and ope a e wi hou assuming ha cu en sola powe plan s a e in op imal loca ions.
Thus, his esea ch ep esen s an in eg a ed explainable AI app oach wi h ML
me hods, aiming o enhance he anspa ency and comp ehensibili y o esul s and
in oduce a no el pa adigm by employing he classi ied e iciency o exis ing plan s,
ca ego ized in o i e classes, as he dependen a iable o ML models, he eby
challenging he p e alen assump ion inhe en in con en ional ML models. Twel e
independen a iables selec ed h ough a li e a u e e iew we e used, along wi h he
classi ied e iciency alues, o ain i e ML models, namely, Random Fo es (RF),
Suppo Vec o Machines (SVM), Mul i-laye Pe cep on (MLP), Decision ee (DT),
and K-nea es neighbo s (k-NN).
Following he assessmen o he models' accu acies, he RF model, which achie ed
88% o e all accu acy, was subsequen ly explained using SHapley Addi i e
exPlana ions (SHAP), e ealing Sola Radia ion (13%) and Cloud Index (12%) as he
mos in luen ial a iables o he esul ing p edic ions. In compa ison, Aspec (5%)
was iden i ied as he leas signi ican pa ame e o he model p edic ions.
The inal sola sui abili y map p oduced wi h he supe io RF model iden i ied
app oxima ely 5% o he o al land a ea in he USA as highly sui able o cons uc ing
sola powe plan s, ensu ing op imal ope a ional e iciency. Addi ionally, 55% o he
land is mode a ely sui able o such es ablishmen s. Con e sely, app oxima ely 9.5%
o he o al land a ea, equi alen o 766,654 km2, is deemed pe manen ly unsui able
o sola powe plan cons uc ion.
Sus ainable De elopmen Goals (SGD):
i
KEYWORDS
Renewable Ene gy
Geog aphical In o ma ion Sys ems
Machine Lea ning
Explainable A i icial In elligence
Sui abili y Mapping
Uni ed S a es o Ame ica
ii
ACRONYMS
A – Aspec
AHP – Analy ic Hie a chy P ocess
AI – A i icial In elligence
ANN – A i icial Neu al Ne wo k
AT – Ai empe a u e
AUC – A ea unde he Cu e
AUROC – A ea Unde he ROC cu e
CI – Cloud index
CSV – Comma-sepa a ed alues
DT – Decision T ee
E – Ele a ion
EIA – Ene gy In o ma ion Agency
GIS – Geog aphic In o ma ion Sys ems
GW – Gigawa
k-NN – k-Nea es Neighbo
LULC – Land use/ land co e
MCDM – Mul i-C i e ia Decision Making
MDI – Mean Dec ease in Impu i y
ML – Machine lea ning
MLP – Mul i-laye Pe cep on
MW – Megawa
PC – P oximi y o he ci y cen e
PD – Popula ion densi y
iii
PG – P oximi y o g id
PR – P oximi y o oad ne wo k
RF – Random Fo es
ROC – Recei e Ope a ing Cha ac e is ic
S – Slope
SHAP – Shapley addi i e explana ions
SMOTE – Syn he ic Mino i y O e sampling Technique
SMV – Suppo Vec o Machines
SR – Sola adia ion
Ti – Tole ance
TOPSIS – Technique o O de Pe o mance by Simila i y o Ideal Solu ion
USA – Uni ed S a es o Ame ica
VIF – Va iance In la ion Fac o
WS – Wind speed
XAI – Explainable A i icial In elligence
ix
INDEX OF THE TEXT
Page
ACKNOWLEDGMENTS ................................................................................................... iii
ABSTRACT ........................................................................................................................ i
KEYWORDS ....................................................................................................................... i
ACRONYMS ...................................................................................................................... ii
INDEX OF THE TEXT ....................................................................................................... ix
INDEX OF TABLES ........................................................................................................... xi
INDEX OF FIGURES ........................................................................................................ xii
1. INTRODUCTION ............................................................................................. 13
1.1 O e iew o he wo k ................................................................................................... 13
1.2 Resea ch gap ................................................................................................................ 15
1.3 Objec i es ..................................................................................................................... 15
1.4 Thesis o ganiza ion ...................................................................................................... 16
2. LITERATURE REVIEW .................................................................................. 17
2.1 Sola sui abili y mapping ............................................................................................. 17
2.2 A i icial neu al ne wo ks-based app oaches o sola sui abili y mapping ................. 20
2.3 Machine lea ning o e MCDM-based me hods ........................................................... 21
2.4 Explainable AI in sola sui abili y mapping ................................................................. 22
2.5 C i e ia and es ic ion ac o s ...................................................................................... 24
2.6 Renewable ene gy in he Uni ed S a es o Ame ica ..................................................... 27
3. DATA AND METHODOLOGY....................................................................... 28
3.1 S udy a ea and Da a...................................................................................................... 28
3.2 Me hodology ................................................................................................................ 29
3.2.1 Da a p e-p ocessing .................................................................................................. 31
3.2.2 Machine lea ning models .......................................................................................... 33
3.2.2.1 Random Fo es classi ie (RF) ............................................................................ 34
3.2.2.2 Suppo Vec o Machines (SVM) ....................................................................... 34
3.2.2.3 Decision T ee (DT)............................................................................................. 34
3.2.2.4 k-Nea es Neighbo s (k-NN) .............................................................................. 34
3.2.2.5 Mul i-Laye Pe cep on (MLP) .......................................................................... 35
3.2.3 Model e alua ion, alida ion, and selec ion .............................................................. 35
3.2.3.1 Recei e Ope a ing Cha ac e is ic (ROC) cu es .............................................. 36
3.2.3.2 Accu acy ............................................................................................................. 36
3.2.3.3 P ecision ............................................................................................................. 36
3.2.3.4 F1 sco e .............................................................................................................. 36
3.2.4 Explainable AI .......................................................................................................... 37
3.2.5 Sola powe plan sui abili y map ............................................................................. 38
4. RESULTS AND DISCUSSION .................................................................................. 39
4.1 Resul s .......................................................................................................................... 39
Chap e 1. In oduc ion
16
1. Wha a e he key ac o s ha signi ican ly in luence he sui abili y o sola
powe plan si e selec ion in di e en egions in he USA?
2. Which ML-based classi ica ion me hod is he mos accu a e o c ea ing a
sui abili y map o iden i y po en ial sola powe si es using exis ing sola
p oduc i i y da a?
3. Wha is he in luence o each ac o on he inal sui abili y map?
The esul s o his esea ch ha e he po en ial o p o ide aluable insigh s o decision-
make s in ol ed in u ban planning, ene gy policy de elopmen , and en i onmen al
conse a ion ini ia i es. Doing so can p omo e a cleane , mo e sus ainable
en i onmen and a g eene u u e o u u e gene a ions.
1.4 Thesis o ganiza ion
This hesis comp ises a o al o six subsequen chap e s. Chap e 1 is an in oduc o y
chap e , p o iding an o e iew o he esea ch, emphasizing he esea ch gap, and
ou lining he p ojec 's objec i es. Mo ing on o Chap e 2, a comp ehensi e e iew o
he exis ing li e a u e is conduc ed, ocusing on echniques ela ed o sui abili y
mapping o sola powe plan s and ML and explainable AI-based me hods. Chap e 3
del es in o he s udy a ea, explaining he da a se s employed and de ailing he
me hodology and ools u ilized in he esea ch. Chap e 4 is dedica ed o p esen ing
he esea ch indings, o ganized wi h a comp ehensi e analysis and discussion o he
ob ained esul s. Finally, Chap e 5 se es as a concluding chap e , consolida ing he
esea ch by summa izing he key indings, add essing esea ch ques ions, and o e ing
ecommenda ions o u u e endea o s.
17
2. LITERATURE REVIEW
In he mission o educing ca bon emissions and conse ing ini e ene gy esou ces,
such as ossil uels, which a e consumed a a a e highe han hei na u al
eplenishmen , nume ous coun ies a ound he wo ld ha e s a egized he adop ion o
al e na i e enewable ene gy sou ces as a means o mee ing he ene gy demands o he
21s cen u y (He nandez, 2018). Among hese na ions, hose loca ed in a id and semi-
a id clima ic egions ha e exhibi ed a pa icula in e es in consuming sunligh as a
esou ce due o i s abundan and eely a ailable na u e (Halde e al., 2022). Gi en
he complica ed ela ionship be ween he loca ion o sola powe plan s, hei
e iciency, and he associa ed cons uc ion cos s, he p ac ice o sui abili y mapping
o such ins alla ions has ga he ed subs an ial a en ion (Spy idonidou & Vagiona,
2023). This li e a u e e iew syn hesizes exis ing esea ch on he subjec , speci ically
ocusing on using explainable AI echniques o de e mine he mos sui able loca ions
o sola powe plan s.
2.1 Sola sui abili y mapping
Sola sui abili y mapping in ol es he iden i ica ion o geog aphically op imal egions
o es ablishing sola ene gy gene a ion in as uc u e (Sup o a e al., 2020). I
comp ehensi ely analyzes a ious in luen ial ac o s, such as sola i adiance le els,
land a ailabili y, clima ic condi ions, and economic iabili y (Mahe e al., 2015a;
Rida e al., 2017; Te uel-Solano e al., 2013). Changes in hese ac o s can esul in
some egions on Ea h's su ace being mo e o less sui able o es ablishing sola powe
plan s (Alami Me ouni e al., 2018). Typically, he exis ing li e a u e e eals ha a eas
wi h high sola i adiance le els, minimal en i onmen al limi a ions, and p oximi y o
es ablished powe in as uc u e a e a o ed o sola powe plan ins alla ions (Asadi
e al., 2023; Asadi & Pou hossein, 2021; Bayounis & Eldama y, 2022; Halde e al.,
2022). None heless, he p ecise c i e ia o sui abili y can di e ac oss s udies,
Chap e 1. In oduc ion
18
mi o ing he a ia ions in local condi ions and p io i ies (Sup o a e al., 2020).
Iden i ying hese sui able egions is a complex unde aking ha equi es he
simul aneous e alua ion and compa ison o mul iple in luencing ac o s h ough a
sys ema ic p ocedu e (Halde e al., 2022; He nandez, 2018; Husein e al., 2023).
Following he in oduc ion o Geog aphical In o ma ion Sys ems (GIS), adi ional
esea ch me hodologies ha e u ilized GIS-based Mul i C i e ia Decision Making
(MCDM) echnologies o add ess hese inqui ies (Sup o a e al., 2020). This p ocess
en ails he c ea ion o dis inc spa ial laye s ha ep esen a ious in luencing ac o s,
subsequen ly con e ed in o spa ial decisions. A weigh ed o e lay is hen applied o
hese laye s by assigning speci ic weigh s o each inpu . The de e mina ion o hese
weigh s in ol es e alua ing he signi icance o he ac o s using a a ie y o
echniques, such as Analy ical Hie a chical P ocess (AHP) (Halde e al., 2021),
(Alami Me ouni e al., 2018), (Mahe e al., 2015b), Technique o O de o
P e e ence by Simila i y o ideal Solu ion (TOPSIS) (Sa kodie e al., 2022), and
Boolean uzzy logic models (Youse i e al., 2018), as well as he expe opinions om
he ield. Table 2.1 summa izes he me hodologies used in esea ch ha apply MCDM
echnologies o map he sui abili y o sola powe plan s.
Me hods applied
Re e ences
1
Da a En elopmen Analysis (DEA) and G ey
Based Mul iple C i e ia Decision Making
(G-MCDM)
(Wang e al., 2022)
2
Linea Reg ession Modelling (LRM) and
GIS + AHP
(Asadi e al., 2023)
3
AHP + Sensi i i y analysis (SA)
(Rida Azmi, Hicham Ama , 2017)
4
GIS + MCDM
(Islam e al., 2022), (Ghe boudj & Ghedi a,
2016), (Villac eses e al., 2022)
5
Technique o o de o p e e ence by simila i y
o ideal solu ion (TOPSIS), Complex
p opo ional assessmen (COPRAS), Mul i-
objec i e op imiza ion on a io analysis
(MOORA) + Spea man co ela ion coe icien
(SCC)
(Sa kodie e al., 2022)
6
MCDM + TOPSIS
(Sánchez-Lozano e al., 2015), (Te uel-
Solano e al., 2013)
7
MCDM
(Nzelibe e al., 2022), (Akkas e al., 2017)
8
MCDM +AHP
(Munkhba & Choi, 2021), (Tisza, 2014),
(Wa son & Hudson, 2015), (Bayounis &
Chap e 1. In oduc ion
19
Eldama y, 2022), (Wa son & Hudson,
2015)
9
GIS + AHP
(Halde e al., 2021), (Alami Me ouni e
al., 2018), (Mahe e al., 2015b)
10
GIS + MCDM +AHP
(Halde e al., 2022)
11
GIS + Boolean and uzzy model
(Youse i e al., 2018)
Table 2.1: MCDM based Sola powe plan sui abili y me hods used in p e ious s udies
None heless, he esul s o hese esea ches depend upon a ious ac o s, such as he
selec ion o c i e ia, he assignmen o ela i e weigh s o he selec ed c i e ia, and he
pa icula pai s o c i e ia used o assessing hei ela i e impo ance (Asadi &
Pou hossein, 2021; Husein e al., 2023; Sánchez-Lozano e al., 2015). The e o e, he
conclusions d awn om hese weigh -based echniques a e subjec i e and suscep ible
o he esea che 's pe sonal bias, expe ise, and p ac ical expe ience in he ield (Sun
e al., 2023a).
To illus a e his a iabili y, conside wo s udies conduc ed o assess he sui abili y o
sola powe plan loca ions in Mo occo, namely, (Taou ik e al., 2021) and (Alami
Me ouni e al., 2018). These s udies iden i ied 11 and 8 in luencing ac o s,
espec i ely. Due o di e ing judgmen s ega ding he selec ion o in luencing ac o s,
he wo s udies assigned a ying deg ees o in luence o hei p ima y ac o , global
ho izon al sola adia ion. In Taou ik's me hodology, his ac o held a 26% in luence
on sola si e selec ion when applying he AHP me hod. In con as , Alami concluded
ha di ec no mal i adia ion was mo e han 50% signi ican in he AHP model. In
addi ion, hese wo s udies used di e en le els o impo ance o he in luencing
ac o s. Fo example, Taou ik conside ed sola adia ion abo e 5 kW/m² highly
sui able, while Alami conside ed 2.1 kW/m² he highes sui able ange. The e o e, he
indings o hese wo s udies de ia e signi ican ly and canno be eadily compa ed o
de e mine he mos accu a e esul s o decision-making p ocesses. As such, he choice
o pa ame e s and sco ing me hods g ea ly in luences p ojec ou comes and is subjec
o he esea che 's biases, le el o knowledge, and eal-wo ld expe ience in he ield.
Due o hese limi a ions and ad ancemen s in A i icial Neu al Ne wo ks (ANN),
esea che s ha e adop ed ML models o ackle loca ion-based challenges (Ghimi e e
al., 2022; Rangel-Ma inez e al., 2021).
Chap e 1. In oduc ion
20
2.2 A i icial neu al ne wo ks-based app oaches o sola
sui abili y mapping
The popula i y o ANN, inspi ed by human biological neu ons, has inc eased as a ool
o da a analysis in a ious applica ions (Singh, 2019), including sui abili y mapping
in GIS. Machine lea ning is a sub ield o a i icial in elligence ha in ol es he
de elopmen o ma hema ical o s a is ical algo i hms ha can lea n om and make
p edic ions o decisions based on exis ing da a abou a speci ic eal-wo ld inqui y
(Sa anya & Subhashini, 2023). The u iliza ion o machine lea ning models o assess
loca ional sui abili y has demons a ed a success ul his o y o me hodologies ac oss
a ious ields in ecen yea s, including p edic ing sui able si es o dams, hospi als,
solid was e land ills, and enewable ene gy si es. These s udies ha e used se e al
machine lea ning echniques, summa ized in Table 2.2, including decision ees,
Random Fo es s, suppo ec o machines, and neu al ne wo ks.
Me hods applied
Re e ences
1
GIS + Machine lea ning (Fuzzy membe ship + Fuzzy
logic model)
(Hai aa Nasse Hussein, Noo
Hashim Hamed, 2023)
2
ML (Hyb id deep CNN-SVR algo i hm)
(Ghimi e e al., 2022)
3
GIS + ML(ANN)
(Oyewola e al., 2022)
4
MCDM + ANN
(Asadi & Pou hossein, 2021)
5
ML (ANN, Suppo ec o eg ession (SVR), and
Gaussian P ocess Reg ession (GPR))
(Sha i zadeh e al., 2019)
6
ML (G adien Boos Decision T ee (GBDT) and eX eme
G adien Boos ing (XGBoos )
(Hou e al., 2023)
7
ML (RF, DT, SVM, k-NN, ANN)
(Shahab & Singh, 2019)
8
GIS + ML + XAI
(Sachi e al., 2022a)
9
ML (RF, MLP, X eme G adien Boos ing (XGBoos ) +
XAI
(Sun e al., 2023b)
Table 2.2: ML based sola powe plan sui abili y me hods used in p e ious s udies
Fo ins ance, p io esea ch conduc ed by Shahab examines he e ec i eness o i e
machine lea ning algo i hms (Random Fo es (RF), Decision T ee (DT), Suppo
Vec o Machines (SVM), K-Nea es neighbo s (k-NN), and A i icial Neu al Ne wo k
(ANN)) o classi ying he sui abili y o enewable ene gy si es based on geophysical
da a se s (Shahab & Singh, 2019). The s udy's indings e ealed ha he RF (F1 sco e
o 92%) algo i hm demons a ed he mos ema kable pe o mance, whe eas he SVM
(F1 sco e o 85%) showed he leas accu a e esul s. Ano he s udy conduc ed by Asadi
Chap e 1. In oduc ion
21
employed a combina ion o MCDM and Mul i-Laye Pe cep on (MLP) o de e mine
he sui abili y o sola powe plan loca ions in I aq (Asadi & Pou hossein, 2021). The
s udy's indings u he alida e he sui abili y o machine lea ning-based me hods,
o e ing accu a e and s able esul s o implemen ing global sco ing capabili ies in
sui abili y analysis asks. Howe e , he assessmen o he supe io i y o machine
lea ning-based me hods o e adi ional MCDM-based app oaches needs o be
add essed.
2.3 Machine lea ning o e MCDM-based me hods
A s udy conduc ed by Saha examined he e ec i eness o MCDM-based AHP and
Fuzzy Complex P opo ional Assessmen (FCOP-RAS) me hods o e wo machine
lea ning app oaches: The Random Fo es (RF) algo i hm and he Mul ilaye
Pe cep on (MLP) o he sui abili y analysis ask using i een in luencing c i e ia
collec ed h ough g ound su eys. The esul s o he s udy showed ha he accu acy
o he models de eloped using machine lea ning based model achie ed highe AUC
sco es compa ed o he o he me hods, wi h AUC sco es o 0.947 o RF, 0.923 o
AHP, 0.928 o FCOP-RAS, and 0.932 o MLP. Addi ionally, he s udy highligh ed
he ime-sa ing a ibu e o he machine lea ning-based me hod o e MCDM-based
app oaches due o he lowe human in e ac ion o ask execu ion (Saha & Mondal,
2022).
Addi ionally, a ecen s udy conduc ed a s a e-o - he-a analysis o ML-based and
MCDM-based me hods o decision suppo sys ems e ealed ha he ML-based
decision-making sys ems p o ide obus and e icien solu ions o complex p oblems
by e ec i ely handling la ge and complex da ase s compa ed o he adi ional
MCDM-based me hods. Mo eo e , he s udy also highligh ed he abili y o ML
algo i hms o lea n om da a pa e ns and make p edic ions wi hou explici
p og amming, hus educing he subjec i e judgmen s on hose decision-making
s udies (Ali e al., 2023).
Fu he mo e, a s udy conduc ed by Asadi has highligh ed one o he signi ican
disad an ages o MCDM-based app oaches known as local sco ing p ope y, which
e e s o he limi a ion, whe ein sco es in MCDM me hods a e de i ed om a es ic ed
Chap e 1. In oduc ion
22
se o local obse a ions, esul ing in a lack o obus ness and suscep ibili y o changes
in sco es when new candida es a e in oduced. As a solu ion o his issue, Asadi and
colleagues in oduced a no el ML-based app oach using Mul ilaye Pe cep on
(MLP), which o e s obus and global sco ing solu ions o modeling wind/sola a m
si ing in Eas Aze baijan. This s udy u he con i ms he ad an ages o ML-based
me hods o e he adi ional MCDM-based me hods in he con ex o sui abili y
analysis p ojec s. Howe e , ML-based me hods a e p o en o be mo e accu a e and
e icien , and he a ailabili y o accu a ely labeled g ound u h samples is an essen ial
ac o o he success ul aining o machine lea ning models (Ghimi e e al., 2022),
(Singh, 2019), (Sachi e al., 2022a).
Addi ionally, he s udies ha used he ML-based me hod o loca ional sui abili y
mapping ha e achie ed accu a e esul s and a e capable o sol ing he issues inhe i ed
wi h he adi ional MCDM-based me hods; hey o en ail o explain he eason behind
he ou pu esul s o loca ion selec ions (Sachi e al., 2022a). These s udies usually
ely on accu acy measu es such as a ea unde he cu e (AUC) o accu acy unde he
ecei e ope a ing cha ac e is ics (AUROC) calcula ed o he alida ion da a as he
p ima y basis o he selec ions, wi hou p o iding insigh s, such as whe he he chosen
egions ha e highe insola ion o lowe slopes, o ice e sa (P ana R, Shashank T
K, n.d.; Sa anya & Subhashini, 2023). As a po en ial solu ion, Explainable AI-based
app oaches ha e been inco po a ed in o loca ion sui abili y s udies o p o ide a
anspa en unde s anding o how each inpu ac o in luences and con ibu es o he
AI model's ou pu s.
2.4 Explainable AI in sola sui abili y mapping
Explainable A i icial In elligence (XAI) is a concep pionee ed by he De ense
Ad anced Resea ch P ojec s Agency (DARPA) o explain he in e nal p ocess o an
AI model (Sa anya & Subhashini, 2023). I in ol es p o iding use -unde s andable
explana ions o he me hod, he p ocess, and he ou pu o an ML model. I simply
con e s he Blackbox na u e o an ML model o a Whi ebox (Gohel e al., 2021). The
explana ions p o ided by XAI a e ca ego ized in o wo componen s: Knowledge-
d i en XAI and Da a-d i en XAI. Knowledge-d i en XAI is employed o acqui e
Chap e 1. In oduc ion
23
insigh s in o he me hodology and echniques used. A he same ime, da a-d i en XAI
helps unde s and he impac and con ibu ion o each inpu ea u e on he AI models'
ou comes. A s udy conduc ed by Sa anya has iden i ied a wide ange o applica ion
domains whe e XAI has been employed in pas esea ch, including heal hca e,
ag icul u al planning, inance, o ecas ing, social media, and mo e (Sa anya &
Subhashini, 2023). Sachi acknowledged ha hei s udy ma ked he ini ial a emp o
employ an Explainable AI-based app oach in mapping spa ial sui abili y o global
wind and sola sys ems o e alua e he speci ic con ibu ion o each inpu c i e ion
wi hin he ML model (Sachi e al., 2022a). The s udy highligh s compelling scien i ic
e idence suppo ing machine lea ning models' obus ness, accu acy, and applicabili y
in sui abili y analysis p ojec s. The pe o mance me ics o he chosen models,
including Random Fo es (RF), Mul i-Laye Pe cep on (MLP), and Suppo Vec o
Machine (SVM), demons a ed an accu acy exceeding 70% ac oss all models.
No ably, RF eme ged as he mos sensi i e and speci ic algo i hm, achie ing a
sensi i i y o 0.88 and a speci ici y o 0.91. The algo i hm's o e all accu acy and kappa
coe icien we e 0.90 and 0.78, espec i ely, while he A ea Unde he Cu e eached
0.96. These esul s unde sco e he e ec i eness o ML models, mainly RF, in
enhancing he p ecision and eliabili y o sui abili y analyses. Fu he , he s udy also
u ilized he global and local explana ions o he SHapley Addi i e exPlana ions
(SHAP) me hod o explain he ML model. The s udy's indings e ealed ha ac o s
such as dis ance om ci y cen e s and empe a u e signi ican ly in luenced sui abili y
decisions wi hin he ML model.
A sepa a e s udy conduc ed by Sun modeled he loca ion choice o la ge-scale sola
pho o ol aic powe plan s in China using in e p e able machine-lea ning echniques.
This esea ch e ealed ha he wo mos c i ical p edic o s o sui abili y o sola
pho o ol aic ins alla ion loca ions we e consis en ly he ege a ion index and he
dis ance o he powe g id among he selec ed 21 geospa ial condi ioning ac o s (Sun
e al., 2023a). Ne e heless, i is essen ial o no e ha bo h o hese s udies ope a ed
assuming ha he cu en sola powe plan s a e si ua ed in he mos sui able a eas
wi hin hei espec i e egions (Sachi e al., 2022a; Sun e al., 2023a). Howe e , his
assump ion may no hold in many ins ances. The li e a u e men ioned abo e has also
shown ha using XAI echniques in sola powe plan selec ion p ojec s is s ill
Chap e 1. In oduc ion
24
ela i ely unde - esea ched. Only a ew s udies ha e been ca ied ou in he pas ha
in eg a ed he spa ial loca ions o he exis ing plan s o he ML model (Sun e al.,
2023a). To he bes o ou knowledge, no s udies conside ed he e iciency o p e ious
acili ies o he ML model.
2.5 C i e ia and es ic ion ac o s
The ca e ul selec ion o c i e ia and es ic ion ac o s is a c ucial aspec o a sola
powe plan sui abili y mapping p ojec , as i di ec ly in luences he accu acy and
ele ance o he esul s (Sup o a e al., 2020). The e iciency o sola powe plan s is
in luenced by he en i onmen al and module ea u es o he loca ion and he equipmen
deployed in he sola powe plan s (Sahin e al., 2023). Table 2.3 and Table 2.4
summa izes he c i e ia and es ic ion ac o s in he p e ious s udies examined in his
esea ch.
C i e ia/
In luen ial ac o
Re e ences
1
Sola adia ion
(Wang e al., 2022), (Asadi e al., 2023), (Bayounis & Eldama y, 2022),
(Ghe boudj & Ghedi a, 2016), (Shahab & Singh, 2019), (Sánchez-
Lozano e al., 2015), (Te uel-Solano e al., 2013), (Villac eses e al.,
2022), (Nzelibe e al., 2022), (Munkhba & Choi, 2021), (Tisza, 2014),
(Halde e al., 2022), (Alami Me ouni e al., 2018), (Khandaka e al.,
2019), (Akkas e al., 2017), (Sachi e al., 2022a), (Rida Azmi, Hicham
Ama , 2017), (Sahin e al., 2023), (Wa son & Hudson, 2015), (Sun e al.,
2023b), (Youse i e al., 2018), (He nandez, 2018)
2
Ai empe a u e
(Ghe boudj & Ghedi a, 2016), (Shahab & Singh, 2019), (Sánchez-
Lozano e al., 2015), (Te uel-Solano e al., 2013), (Villac eses e al.,
2022), (Munkhba & Choi, 2021), (Halde e al., 2021), (Khandaka e al.,
2019), (Oyewola e al., 2022), (Sachi e al., 2022a), (Rida Azmi, Hicham
Ama , 2017), (Sahin e al., 2023), (He nandez, 2018)
3
Wind speed
(Wang e al., 2022), (Asadi e al., 2023), (Ghe boudj & Ghedi a, 2016),
(Villac eses e al., 2022), (Khandaka e al., 2019), (Asadi & Pou hossein,
2021), (Sachi e al., 2022a), (Sahin e al., 2023), (Wa son & Hudson,
2015)
4
P ecipi a ion
(Wang e al., 2022), (Oyewola e al., 2022)
5
Rela i e humidi y
(Wang e al., 2022), (Ghe boudj & Ghedi a, 2016), (Khandaka e al.,
2019), (Oyewola e al., 2022), (Sahin e al., 2023), (Youse i e al., 2018)
6
Sunshine hou
(Wang e al., 2022), (Akkas e al., 2017), (Sun e al., 2023b)
7
Slope
(Asadi e al., 2023), (Bayounis & Eldama y, 2022), (Islam e al., 2022),
(Shahab & Singh, 2019), (Sánchez-Lozano e al., 2015), (Te uel-Solano
e al., 2013), (Nzelibe e al., 2022), (Munkhba & Choi, 2021), (Halde e
al., 2021), (Alami Me ouni e al., 2018), (Asadi & Pou hossein, 2021),
Chap e 1. In oduc ion
25
(Sachi e al., 2022a), (Rida e al., 2017), (Sun e al., 2023b), (Youse i e
al., 2018), (He nandez, 2018)
8
Cloud Index
(Wang e al., 2022), (Sachi e al., 2022a)
9
Technical
in o ma ion and
assis ance
(Wang e al., 2022)
10
Geology
(Wang e al., 2022)
11
Skilled manpowe
a ailabili y
(Wang e al., 2022)
12
Elec ici y demand
(Wang e al., 2022), (Sun e al., 2023b)
13
Land Cos s
(Wang e al., 2022), (Sun e al., 2023b), (Sun e al., 2023b)
14
Local esiden
a i ude
(Wang e al., 2022)
15
Go e nmen
policies and laws
(Wang e al., 2022), (Sun e al., 2023b)
16
Land use/ land
co e
(Wang e al., 2022), (Islam e al., 2022), (Sánchez-Lozano e al., 2015),
(Te uel-Solano e al., 2013), (Villac eses e al., 2022), (Nzelibe e al.,
2022), (Tisza, 2014), (Halde e al., 2021), (Oyewola e al., 2022), (Asadi
& Pou hossein, 2021), (Sachi e al., 2022a), (Rida Azmi, Hicham Ama ,
2017), (Sun e al., 2023b), (Youse i e al., 2018)
17
Suppo
mechanisms
(Wang e al., 2022)
18
Wildli e and
endange ed species
impac
(Wang e al., 2022)
19
Ha m ul oxin
emission
(Wang e al., 2022)
20
Ene gy sa ing
bene i s
(Wang e al., 2022)
21
T ansmission g id
accessibili y
(Wang e al., 2022), (Asadi e al., 2023), (Bayounis & Eldama y, 2022),
(Islam e al., 2022), (Sánchez-Lozano e al., 2015), (Te uel-Solano e al.,
2013), (Villac eses e al., 2022), (Nzelibe e al., 2022), (Munkhba &
Choi, 2021), (Tisza, 2014), (Alami Me ouni e al., 2018), (Sachi e al.,
2022a), (Rida Azmi, Hicham Ama , 2017), (Wa son & Hudson, 2015),
(Sun e al., 2023b), (He nandez, 2018)
22
P oximi y o oad
ne wo k
(Wang e al., 2022), (Asadi e al., 2023), (Bayounis & Eldama y, 2022),
(Islam e al., 2022), (Sánchez-Lozano e al., 2015), (Te uel-Solano e al.,
2013), (Villac eses e al., 2022), (Munkhba & Choi, 2021), (Tisza, 2014),
(Halde e al., 2021), (Alami Me ouni e al., 2018), (Asadi &
Pou hossein, 2021), (Sachi e al., 2022a), (Rida Azmi, Hicham Ama ,
2017), (Wa son & Hudson, 2015), (Sun e al., 2023b), (Youse i e al.,
2018), (He nandez, 2018)
23
Residen ial a eas/
Popula ion densi y
(Wang e al., 2022), (Nzelibe e al., 2022), (Alami Me ouni e al., 2018),
(Sachi e al., 2022a), (Wa son & Hudson, 2015), (Sun e al., 2023b)
24
P oximi y o ci y/
u ban
(Asadi e al., 2023), (Bayounis & Eldama y, 2022), (Islam e al., 2022),
(Sánchez-Lozano e al., 2015), (Villac eses e al., 2022), (Asadi &
Chap e 3. Da a and Me hodology
32
𝐸𝑓𝑓𝑖𝑒𝑐𝑖𝑒𝑛𝑐𝑦 = 𝐴𝑛𝑛𝑢𝑎𝑙 𝑛𝑒𝑡 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛
365 𝑑𝑎𝑦𝑠∗24 ℎ𝑜𝑢𝑟𝑠∗𝑐𝑎𝑝𝑎𝑐𝑖𝑡𝑦 ∗ 100% ......................... (3.1)
Due o he lack o exis ing li e a u e p o iding cu -o alues o e iciency ca ego ies
in sola powe plan s, hypo he ical assump ions we e made o classi ica ion as
ou lined below.
I he e iciency exceeded 50%, i indica ed ha he powe plan ope a ed op imally
o a leas 12 hou s a day. Based on his, powe plan s wi h an e iciency alue o
mo e han 40% we e ca ego ized as hose loca ed in he mos sui able geog aphical
a eas, and he powe plan s wi h e iciencies below 1% we e conside ed indica i es o
he pe manen ly unsui able egions. Subsequen ly, he emaining sola powe plan s
we e ca ego ized in o h ee classes based on hei e iciency le els. Powe plan s wi h
e iciencies anging om 1 o 14 we e classi ied as ma ginally no sui able, hose wi h
e iciencies om 14 o 27 we e classi ied as ma ginally sui able, and hose wi h
e iciencies om 27 o 40 we e classi ied as mode a ely sui able. The uppe alue o
each class was no included in he ange. Once he geoda abase was es ablished, he
subsequen p ocessing s eps we e implemen ed o ende i sui able o he ML model.
• Handling missing alues:
Missing alues in he geoda abase we e add essed by eplacing hem wi h he mean
and modes o he a ailable da a, u ilizing he mode o land use, land co e , aspec ,
popula ion densi y, and he mean o he emaining a iables.
• Collinea i y and Mul icollinea i y assessmen :
In o de o mi iga e he in luence o edundan a iables on he ML model, an
e alua ion o collinea i y and mul icollinea i y was unde aken. This in ol ed
gene a ing a co ela ion ma ix and compu ing he a iables' Tole ance (Ti) and
Va iance In la ion Fac o (VIF). The analysis signals mul icollinea i y conce ns when
T is below 0.1, and VIF exceeds 10. Addi ionally, collinea i y issues a e iden i ied
when he co ela ion app oaches ±1.
• Da a spli ing:
The da ase was hen spli in o wo classes o he aining and es ing da ase s,
ollowing he a io o 8:2, consis en wi h o he s udies (Sachi e al., 2022a), (Sun e
Chap e 3. Da a and Me hodology
33
al., 2023a). The ML models we e ained using he aining da ase , and he accu acy
o he models was assessed using he es ing da ase .
• Da a balancing:
Due o he unequal numbe o obse a ions ac oss e iciency classes, he Syn he ic
Mino i y O e sampling Technique (SMOTE) was employed o duplica e obse a ions
in he mino i y class, ensu ing he impa iali y o he ML models.
• Da a no maliza ion (T aining da a se ) and Rescaling (Tes ing da a se ):
Gi en he wide a ia ion in uni s among he independen a iables, a minimum-
maximum no maliza ion me hod was u ilized o s anda dize he aining da a a ia ion
om 0 o 1 o each a iable using he calcula ed minimum and maximum alues
ob ained exclusi ely o he aining da a se . Then, he calcula ed minimum and
maximum alues we e used o escale he es ing da a se o educe he a o emen ioned
wide a ia ions o he es ing da a se .
3.2.2 Machine lea ning models
ML models a e designed o lea n pa e ns and ela ionships om labeled da a, enabling
hem o make p edic ions on new, unseen da a. The wo main ypes o ML models a e
supe ised lea ning and unsupe ised lea ning. Supe ised lea ning models, including
Random Fo es , Suppo Vec o Machines, and Mul i-laye Pe cep on, a e
pa icula ly e ec i e in add essing classi ica ion and eg ession challenges. On he
o he hand, unsupe ised lea ning me hods a e employed o clus e ing asks. The
p esen s udy u ilized i e supe ised ML models: Random Fo es Classi ie , Decision
T ee, Mul i-laye Pe cep on, Suppo Vec o Machines, and K-Nea es Neighbo s o
ca ego ize he sui abili y o sola powe plan s in o i e classes as pe manen ly
unsui able, ma ginally no sui able, ma ginally sui able, mode a ely sui able, and
highly sui able. The classi ica ion me hods ha e demons a ed highe accu acies o
classi ica ion asks in p e ious s udies. The Models we e de eloped in a Py hon
en i onmen using he Sciki -lea n package.
Chap e 3. Da a and Me hodology
34
3.2.2.1 Random Fo es classi ie (RF)
A Random Fo es classi ie is an enhanced e sion o a decision ee whe e nume ous
unco ela ed decision ees a e combined in o a single s uc u e called a o es . I is
widely used o add ess he classi ica ion and eg ession applica ion due o i s abili y o
classi y he inpu samples o high dimensional a iables wi hou dimension educ ion
o o e i ing (Sun e al., 2023b). The selec ion o he numbe o ees o his classi ie
is a complex ask ha needs o be ca e ully conduc ed o p e en unnecessa y noises
and o e i ing o he model. The Random Fo es model used in his s udy was
de eloped wi h 1000 unco ela ed ees o ob ain he mos accu a e esul s o he
classi ica ion p oblem.
3.2.2.2 Suppo Vec o Machines (SVM)
SVM is a obus classi ie buil on he heo e ical concep s o s a is ical lea ning o
ind he op imal hype plane in a high-dimensional ea u e space o sepa a e he da a
poin s in o classes wi h he help o a ke nel unc ion (Singh, 2019). This me hod can
uniquely o e come complex non-linea ela ionships wi h he help o he ke nel
unc ion, which maps he inpu da a in o high-dimensional ea u e space (Hou e al.,
2023). In his s udy, he SVM op imiza ion was achie ed by selec ing he Gaussian
adial basis unc ion as he designa ed ke nel o ope a ions, wi h a chosen C alue o
0.2 assigned o con ol he smoo hness o he decision bounda ies.
3.2.2.3 Decision T ee (DT)
Decision ee classi ie s a e algo i hms ha use he ules o make decisions by
ecu si ely pa i ioning he inpu space in o subse s based on he mos signi ican
ea u es. These classi ie s a e easy o in e p e and hus aluable in decision-making
and classi ica ion p oblems. The s abilized and mos accu a e decision ee chosen o
he s udy has con igu ed he Spli e pa ame e 'bes ' and se he Min Impu i y Dec ease
o 0.1.
3.2.2.4 k-Nea es Neighbo s (k-NN)
k-Nea es Neighbo s is an analy ical ML me hod used o classi ica ion and
eg ession-based p oblems. This me hod calcula es he dis ance om each class o he
objec s, and he class wi h he minimum dis ance is assigned as he objec 's class
Chap e 3. Da a and Me hodology
35
(Singh, 2019). The classi ie 's accu acy was imp o ed by ine- uning pa ame e s,
including he numbe o neighbo s, sea ch me hod (g id and andom), and dis ance
me ic. In he p esen s udy, he model's pe o mance was op imized by employing he
Euclidean dis ance, 5 neighbo s, and a andom sea ch me hod.
3.2.2.5 Mul i-Laye Pe cep on (MLP)
Mul i-laye Pe cep on is a widely used a i icial neu al ne wo k o classi ica ion, da a
mining, and modeling p ocesses (Asadi & Pou hossein, 2021). The s uc u e o an
MLP model comp ises an inpu laye , an ou pu laye , and one o mo e in e media e
laye s dedica ed o ecei ing decision ac o s, ou pu ing classi ied esul s, and
pe o ming he compu a ional p ocess o he model, espec i ely (Sachi e al., 2022a).
In supe ised classi ica ion, he MLP model is ained h ough backp opaga ion,
whe ein he model's accu acy is imp o ed in each i e a i e p ocess by adjus ing he
model wi h a speci ied lea ning a e.
The MLP model u ilized in his s udy employed a Mul i-laye Pe cep on (MLP) model
wi h 4 hidden laye s, comp ising 100, 80, 60, and 40 Pe cep on, espec i ely. The
chosen ac i a ion unc ion was he ec i ie linea uni , and he model unde wen
aining wi h a maximum o 5000 i e a ions. An ea ly s opping me hod was
implemen ed o p e en model o e i ing, which moni o ed he alida ion ac ion se
a 0.2.
3.2.3 Model e alua ion, alida ion, and selec ion
Pe o mance e alua ion is an essen ial ask in ML p ojec s, which allows he
assessmen o he accu acy o p edic ions using a new da a se (Hou e al., 2023). A e
implemen ing he me hods using Py hon’s execu able code, he pe o mance o each
me hod was assessed by u ilizing con usion ma ices o de i e pe o mance me ics,
including Accu acy, P ecision, F1 Sco e, and Recei e Ope a ing Cha ac e is ic
(ROC) cu es, which calcula e he a ea unde he cu es. Finally, he model ha
achie ed he highes F1 sco e and he mos signi ican a ea unde he cu e was chosen
as he op imal one o ad ance he esea ch.
Chap e 3. Da a and Me hodology
36
3.2.3.1 Recei e Ope a ing Cha ac e is ic (ROC) cu es
The ROC cu es a e o en used o compa e he pe o mance o classi ica ion asks by
illus a ing he ac ual posi i e a e o p edic ions agains he alse posi i e a e (Sachi
e al., 2022a). The a ea unde he cu e is he quan i a i e measu e ha e alua es he
quali y o he ROC cu es. The AUC alues can ange om 0.5 o 1, 1 indica ing a
pe ec model. Since he s udy da ase con ained i e sui abili y classes as labels,
sepa a e ecei e ope a ing cha ac e is ics we e eco ded o each class o calcula e he
a eas unde he cu es.
3.2.3.2 Accu acy
The measu ing accu acy quan i ies he pe o mance o he ML models by ob aining
he a io o co ec ly classi ied pixels ( ue posi i es) o he o al numbe o pixels,
which is he sum o ue posi i es, ue nega i es, alse nega i es, and alse posi i es
(I za Alejand a e al., 2020). The accu acy o i e ML models was independen ly
compu ed o i e sui abili y classes, u ilizing he de eloped con usion ma ix. Then,
he inal accu acy alue was de i ed by calcula ing he unweigh ed a e age ac oss he
di e en land use classes.
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 ............................................ (3.2)
3.2.3.3 P ecision
P ecision is he a io be ween ue posi i es and all posi i es cap u ed by he model,
namely ue and alse posi i es (Si e el Ricka d, 2004). I was employed in he s udy
o measu e he ML model's capaci y o p edic posi i e ins ances accu a ely. Simila
o he accu acy calcula ions, he p ecision o independen classes was calcula ed o
ob ain he a e age p ecision o each ML model.
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃
𝑇𝑃+𝐹𝑃 ................................................... (3.3)
3.2.3.4 F1 sco e
The F1 Sco e is a ma ix used o measu e a model's o e all pe o mance by combining
p ecision and ecall in o a single ma ix (Sh es ha & Vanneschi, 2018). I measu es he
model's abili y o cap u e all posi i e ins ances by ge ing he ha monic mean o
Chap e 3. Da a and Me hodology
37
p ecision and ecall, whe e p ecision measu es he a io be ween ue posi i es and he
sum o ue and alse posi i es. In con as , ecall measu es he a io be ween ue
posi i es and he sum o ue posi i es and alse nega i es (Abdollahi e al., 2020).
Gi en he mul iclass classi ica ion na u e, his s udy employs he mac o-a e age F1
sco e o he assessmen .
𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2∗𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑟𝑒𝑐𝑎𝑙𝑙
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙 ........................................... (3.4)
3.2.4 Explainable AI
Explainable AI (XAI) is a concep aiming o explain he in e nal p ocess o an ML
model by p o iding use -unde s andable explana ions o he me hod, he p ocess, and
he ou pu o an ML model (Sachi e al., 2022a). In his s udy, he popula XAI
algo i hm, known as SHAP, is used o explain he RF model, which was iden i ied as
he mos accu a e among he i e selec ed models. Shapely in oduced he SHAP
algo i hm (Vajda e al., 1951) in 1951 o e alua e he con ibu ion o indi idual playe s
o a n-pe sons game whe e mo e han one playe is con ibu ing o he game. I was
subsequen ly expanded in a i icial in elligence o in e p e model p edic ions in 2017
by Sco Lundbe g (Lundbe g & Lee, 2017). The SHAP algo i hm is applied in a ious
o ms, such as T eeSHAP, DeepSHAP, and Ke nel SHAP. This s udy uses T eeSHAP,
designed explici ly o ML models based on ee s uc u es (Sachi e al., 2022b).
The SHAP algo i hm has wo explainabili y echniques: global and local explana ions,
whe e he global explana ions desc ibe he model in gene al. In con as , local
explana ions desc ibe e e y obse a ion in he model (Dallanoce, 2022). In his s udy,
he global in e p e a ions o he model we e ob ained by illus a ing he summa y plo s
o he model. These explana ions aim o map he con ibu ion o each independen
a iable in assessing he loca ional sui abili y o sola powe plan s. Then, he local
explana ions we e ob ained h ough o ce plo s, highligh ing he ac o s ha
signi ican ly con ibu ed o he p edic ion by iden i ying he e ec o each independen
a iable on classi ying an indi idual pixel in he sui abili y map.
Chap e 3. Da a and Me hodology
38
3.2.5 Sola powe plan sui abili y map
As he nex s ep o he s udy, a sui abili y map o es ablishing sola powe plan s in
he Uni ed S a es o Ame ica was gene a ed using he ained model. A e emo ing
he p o ec ed a eas, he model's p edic ions we e displayed ca og aphically as a map,
se ing as a aluable esou ce o he decision-making p ocess in he u u e. The
quan i a i e analysis o he spa ial dis ibu ion was subsequen ly ca ied ou o de i e
insigh s ega ding he pe cen ages o land pa cels and a eas sui able o mapping sola
powe plan s. In he inal s age o he me hodology, a compa a i e analysis was
conduc ed be ween he exis ing sola powe plan s classi ied acco ding o hei
e iciency alues, and hose a e ob ained h ough model p edic ions. This assessmen
aimed o iden i y disc epancies be ween he g ound u h da a and he model
p edic ions.
39
4. RESULTS AND DISCUSSION
This chap e ocuses on he isualiza ion and discussion o he ou comes de i ed om
he conduc ed s udy. The ini ial sec ions showcase he ou comes o he p e-p ocessing
s eps. The ollowing sec ion illus a es he esul s ob ained du ing he model selec ion
and accu acy assessmen . The inal wo sec ions p o ide a de ailed accoun o he
esul s ob ained o explainable AI and he inal sola sui abili y maps.
4.1 Resul s
4.1.1 Da a p e-p ocessing
A e applying a ious da a p e-p ocessing s eps, including ep ojec ion, clip, as e -
o- ec o con e sion, and esampling, Figu e 4.1 – 4.12 p esen s he inal as e laye s
o he wel e independen a iables.
Figu e 4.1: Ai empe a u e
Chap e 4. Resul s and Discussion
40
Figu e 4.4: Aspec
Figu e 4.3: Popula ion densi y
Figu e 4.2: Land use / land co e
Chap e 4. Resul s and Discussion
41
Figu e 4.7: Cloud index
Figu e 4.6: Ele a ion
Figu e 4.5: Global ho izon al i adia ion
Chap e 4. Resul s and Discussion
48
The ou comes e eal ha sola adia ion is he p ima y de e minan in sola powe
plan sui abili y modeling, con ibu ing 13% o he model. Following closely, he cloud
index and p oximi y o he ci y cen e eme ge as he second and hi d mos signi ican
ac o s, wi h in luences o 12% and 11%, espec i ely. Subsequen ly, ai empe a u e
and ele a ion sha e equal impo ance a 9% each in he model. The leas impac ul
ac o in he model is he aspec , con ibu ing only 5% o he o e all impo ance o he
RF model.
4.1.6 Sola powe plan sui abili y map
Figu e 4.16 depic s he esul ing map indica ing he sui abili y o he sola powe plan
gene a ed h ough he Random Fo es classi ica ion model. Be o e c ea ing he map,
he p edic ions o he ML model unde wen a il e ing p ocess, which was used o
assign he "Pe manen ly no sui able" class o p o ec ed a eas.
The quan i a i e examina ion o he ul ima e sui abili y map is p esen ed in Table 4.4,
indica ing ha a ound 5% o he o al land a ea in he Uni ed S a es is highly sui able
o cons uc ing sola powe plan s, ensu ing op imal ope a ional e iciency.
Addi ionally, 55% o he land is mode a ely sui able o such es ablishmen s.
Con e sely, app oxima ely 9.5% o he o al land a ea, equi alen o app oxima ely
Figu e 4.15: Weigh s o he ac o s
Chap e 4. Resul s and Discussion
49
766654 km2, is deemed no pe manen ly sui able o sola powe plan cons uc ion.
This limi a ion is p ima ily a ibu ed o en i onmen ally p o ec ed a eas ese ed
h oughou he coun y.
Class
A ea (sq. km)
Pe cen age (%)
Highly sui able
389666
4.9
Mode a ely sui able
4377279
55.07
Ma ginally sui able
2329314
29.3
Ma ginally no sui able
86039
1.08
Pe manen ly no sui able
766654
9.65
Table 4.4: Quan i a i e examina ion o he inal sui abili y map
4.1.7 Compa a i e analysis o g ound u h da a and model p edic ions
Figu e 4.17 illus a es he o e lay o exis ing powe plan s o e he sui abili y map
p oduced using he Random Fo es model. The quan i a i e analysis conduc ed o e
he o e laying map is p esen ed in , displaying he numbe o exis ing sola powe
plan s wi hin each sui abili y class o he map along wi h he ac ual numbe o
he sola powe plan s loca ed in he espec i e class ha was calcula ed using he
me hod desc ibed in he sec ion 3.2.1.
Figu e 4.16: Final sui abili y map
Chap e 4. Resul s and Discussion
50
No:
Class
G ound
u h da a
P edic ion
Map
Di e ence
Model Accu acy
1
Highly sui able
173
187
14
O e es ima e
2
Mode a ely sui able
609
383
26
Unde es ima e
3
Ma ginally sui able
2996
3006
10
O e es ima e
4
Ma ginally no sui able
939
1114
25
Unde es ima e
5
Pe manen ly no sui able
223
250
27
O e es ima e
Table 4.5: The dis ibu ion o p edic ed and exis ing sola powe plan s o e he s udy egion
The obse ed disc epancies sugges ha he p edic ions o e es ima e h ee classes and
unde es ima e wo classes. Howe e , gi en he sligh di e ences in alues, i can be
assumed ha he model pe o med well in spa ial dis ibu ion.
4.2 Discussion
The con en ional app oaches o si e sui abili y assessmen in scien i ic esea ch o en
in ol e using mul ic i e ia-based analysis me hods o iden i y he op imal loca ions
o ins alling sola powe plan s. Howe e , hese con en ional me hods ha e inhe en
limi a ions ela ed o he subjec i e selec ion o c i e ia and he ca ego iza ion o
in luencing ac o s. Such subjec i i y can in oduce biases and pose limi a ions in eal-
wo ld enewable ene gy scena ios, especially when he e is insu icien knowledge o
ma u i y in he me hodology. Ne e heless, he p oposed s udy uses ad ancemen s in
he a ailabili y o eal-wo ld digi ally o med spa ial da a, geog aphical in o ma ion
sys ems, and emo ely sensed condi ional ac o s o de elop mo e ma u e and less
Figu e 4.17: Exis ing sola powe plan s o e he sui abili y map p oduced wi h he RF model
Chap e 4. Resul s and Discussion
51
biased ML-based solu ions ha e ec i ely add ess loca ional sui abili y issues. The
pe o mance measu es o he cu en s udy, whe e h ee ou o he i e selec ed models
achie ed an accu acy exceeding 74%, a i m he sui abili y and e ec i eness o such
me hods in decision-making p ojec s.
The highes accu acy, p ecision, F1 sco e, and AUC we e achie ed using he Random
Fo es classi ie , ollowed by he decision ees, K-nea es neighbo s, and Mul i-laye
Pe cep on in he gi en s udy. These esul s align wi h p e ious indings, emphasizing
he e iciency and accu acy o he Random Fo es model o loca ional sui abili y
esea ch (Sachi e al., 2022), (Aami Shahab & M.P. Singh, 2022), (Sun e al., 2023).
Mo eo e , he esul s o he impo ance o condi ional ac o s calcula ed wi h he
Random Fo es model e eal ha he sola adia ion, cloud index, and p oximi y o he
ci y cen e o he esea ch a ea employ he mos signi ican in luences on he
selec ions. These weigh s o e aluable insigh s o choosing a loca ion o hos a sola
powe plan in an economically e icien en i onmen . The ac o s pinpoin ed as he
mos in luen ial by he model a e consis en wi h he ecommenda ions de i ed om
nume ous s udies in he ield o loca ional choice esea ch (Akkas e al., 2017),
(Youse i e al., 2018), (Ghe boudj & Ghedi a, 2016), (Husein e al., 2023).
Meanwhile, he ou comes de i ed h ough he XAI app oach ha e con ibu ed o a
deepe comp ehension o he selec ed model and he condi ioning ac o s ha
signi ican ly a ec a speci ic loca ion's classi ica ion wi hin a pa icula sui abili y
class. The global explana ions o he Random Fo es (RF) model o e ed insigh s in o
he mean in luence o each p edic ion ac oss i e sui abili y classes. Sola adia ion
exhibi ed he highes mean SHAP alue, ollowed by cloud index and ai empe a u e.
In addi ion, he local explana ions elucida ed he impac o each ac o on he model
conce ning hei assignmen o speci ic ca ego ies. Sola adia ion eme ged as he mos
in luen ial ac o o he highly sui able class, wi h lowe alues nega i ely impac ing
he class and highe alues posi i ely con ibu ing o he class. The analysis o e ed a
ho ough o e iew o each ac o , illus a ing how lowe and highe alues impac
each classi ica ion class. The p oposed s udy concludes ha he ML models a e
sui able and ha Explainable AI app oaches e ec i ely explain he selec ed ML-based
Chap e 4. Resul s and Discussion
52
model o sola powe plan sui abili y mapping a a esolu ion o 1 km2 o he
con inen al USA. Fu he , i is essen ial o acknowledge ha he cu en s udy has
p ima ily ocused on e alua ing he physical, clima ic, and cons uc ion cos - ela ed
pa ame e s a o able o he selec ion p ocess o sola powe plan si es. Howe e , o
a comp ehensi e assessmen , i is necessa y o also accoun o addi ional economic
and poli ical ac o s.
53
5. CONCLUSIONS AND RECOMMENDATION
This chap e summa izes he s udy's conclusion in alignmen wi h he p ede ined
esea ch objec i es. Addi ionally, i add esses he iden i ied limi a ions and
ecommenda ions o u u e esea ch endea o s, aiming o enhance and e ine he
s udy's ou comes.
5.1 Conclusions
The s udy's p ima y objec i e aimed o de elop an explainable AI-based ML model
o mapping he sui abili y o sola powe plan loca ions by conside ing he e iciency
o he plan as he dependen a iable o he ML models ins ead o conside ing all
exis ing plan s in he ideal loca ions. Inco po a ing e iciency in o he model in ol ed
success ul implemen a ion, whe e calcula ions we e execu ed using ene gy p oduc ion
and capaci y da a om he U.S. Ene gy In o ma ion Adminis a ion (EIA). Howe e ,
he absence o cu o alues o classi ying plan s in o sui abili y classes led o
de eloping a hypo he ical assump ion, elabo a ed in Chap e 3, sec ion 3.2.1.
Fi e ML algo i hms (Random Fo es , Decision ee, K-nea es neighbo s, mul i-laye
Pe cep on, and Suppo ec o machines) we e employed in he s udy using eal-wo ld
da a. Resul s con i med he supe io i y o he Random Fo es model o he s udy wi h
high-pe o mance alues o accu acy me ics: F1 Sco e: 0.881, P ecision: 0.881,
Accu acy: 0.881. Consequen ly, he Random Fo es algo i hm can be sugges ed as a
sui able model o he sui abili y assessmen amewo ks o sola powe plan s. The
conclusi e sui abili y map de eloped u ilizing a chosen Random Fo es indica es ha
app oxima ely 4.90% o he o al land a ea in he USA is ecommended o
cons uc ing sola powe plan s o achie e highe e iciency in ene gy p oduc ion.
Fu he mo e, 55.07% o he land is classi ied as mode a ely sui able, wi h 29.30% and
1.08% iden i ied as ma ginally sui able and ma ginally no sui able, espec i ely.
Chap e 5. Conclusions and Recommenda ions
54
These indings p o ide a p ac ical and ac ionable o e iew o decision-make s and
s akeholde s in he sola ene gy sec o , o e ing insigh s in o he mos sui able a eas
o op imal and e icien sola powe plan cons uc ion ac oss he coun y. The
conclusions d awn om add essing he es ablished esea ch ques ions in pu sui o he
p ima y objec i e a e as ollows:
1. Wha a e he key ac o s ha signi ican ly in luence he sui abili y o sola
powe plan si e selec ion p ocess?
In conclusion, examining key ac o s in luencing he sui abili y o sola powe
plan si e selec ion in ol ed an ex ensi e li e a u e e iew, encompassing
s udies beyond and wi hin he s udy a ea. This e iew iden i ied 14 ac o s ha
signi ican ly in luence he si e selec ion p ocess as Sola adia ion (SR), Cloud
index (CI), P oximi y o ci y cen e s (PC), Popula ion densi y (PD), Ele a ion
(E), P oximi y o oad ne wo k (PR), Land use and land co e (LULC),
P oximi y o g id line (PG), Ai empe a u e (AT), Wind speed (WS), Aspec
(A), Slope (S), Humidi y (HD), and Na u al disas e s (ND). Howe e , al hough
humidi y and occu ences o na u al disas e s in he a ea we e iden i ied as
signi ican , hose we e excluded om he s udy due o he da a una ailabili y.
This highligh s he impo ance o he iden i ied ac o s and unde sco es he need
o comp ehensi e da a a ailabili y o a mo e exhaus i e analysis in u u e
s udies.
2. Which ML-based classi ica ion me hod is he mos accu a e o c ea ing a
sui abili y map o iden i y po en ial sola powe si es using exis ing sola
p oduc i i y da a?
In conclusion, he in es iga ion in o ML-based classi ica ion me hods aimed a
c ea ing a sui abili y map o iden i ying po en ial sola powe si es yielded
aluable insigh s. The s udy employed i e dis inc me hods: Random Fo es ,
Decision ees, K-nea es neighbo s, Mul i-Laye Pe cep on, and Suppo
Vec o machines. Th ough igo ous e alua ion, Random Fo es eme ged as he
mos accu a e classi ica ion me hod, demons a ing supe io pe o mance in
e ms o accu acy. The selec ion o Random Fo es as he p e e ed model o
gene a ing a sui abili y map unde sco es i s e icacy in le e aging exis ing sola
Chap e 5. Conclusions and Recommenda ions
55
p oduc i i y da a o p ecise iden i ica ion o po en ial sola powe si es. This
conclusion p o ides p ac ical guidance o u u e endea o s in sola si e
selec ion, emphasizing he impo ance o employing Random o es s o
op imal accu acy and eliabili y.
3. Wha is he in luence o each ac o on he inal sui abili y map?
In conclusion, he SHAP analysis, conduc ed h ough explainable AI
echniques, highligh ed he c i ical ole o indi idual en i onmen al,
geomo phological, spa ial, and clima ic c i e ia in he spa ial decision-making
p ocess o p edic ing he sui abili y class o sola powe plan loca ions. These
ac o s we e anked in o de o impo ance, om highes o lowes , as ollows:
SR - 13%, CI - 12%, PC - 11%, AT - 9%, E - 9%, PD - 8%, WS - 7%, PR - 7%,
LULC - 7%, PG - 6%, S - 6%, and A - 5%
5.2 Limi a ions and Recommenda ions
While his s udy con ibu es aluable insigh s in o Explainable ML-based sola powe
plan selec ion me hodologies, i is essen ial o acknowledge ce ain limi a ions
inhe en in he me hodology and da a sou ces. These limi a ions may impac he
gene alizabili y and applicabili y o he indings.
1. Una ailabili y o da a:
Limi ed da a a ailabili y o humidi y and he occu ence o na u al disas e s
es ic ed hei inclusion in he analysis, emphasizing he need o
comp ehensi e da ase s in u u e s udies.
2. Limi ed li e a u e on e iciency-based classi ica ion:
Sca ce li e a u e on classi ying sola powe plan s based on e iciency
cons ained he s udy's dep h, highligh ing he impo ance o expanding
esea ch in his a ea.
3. Insu icien eco ded sola powe plan s:
The s udy aced limi a ions due o insu icien sola powe plan s wi h eco ded
p oduc ion and consump ion da a, pa icula ly o la ge-scale si e sui abili y
mapping. Collabo a i e e o s a e needed o compile a mo e ex ensi e da ase
o u u e esea ch.
Chap e 5. Conclusions and Recommenda ions
56
Fo u u e ecommenda ions, he s udy would unde sco e he impo ance o
inco po a ing sola panels' il and sun o ien a ion as in luen ial ac o s, hus
enhancing he comp ehensi e unde s anding o si e sui abili y o sola powe
plan s.
57
6. BIBLIOGRAPHIC REFERENCES
Abdollahi, A., P adhan, B., Shukla, N., Chak abo y, S., & Alam i, A. (2020). Deep
lea ning app oaches applied o emo e sensing da ase s o oad ex ac ion: A
s a e-o - he-a e iew. Remo e Sensing, 12(9).
h ps://doi.o g/10.3390/RS12091444
Akinsola, J. E. T., Awodele, O., Kuyo o, S. O., & Kasali, F. A. (2019). Pe o mance
E alua ion o Supe ised Machine Lea ning Algo i hms Using Mul i-C i e ia
Decision Making Techniques. In e na ional Con e ence on In o ma ion
Technology in Educa ion and De elopmen (ITED), Mcdm, 17–34.
h ps://www.academiain o ma ion echnology.o g/i ed2019/uploads/8135_File_0
3ITED19041 IEEE Pape Fo ma Pe o mance E alua ion o Supe ised
Machine Lea ning Algo i hms Using MCDM Techniques NEW (1).pd
Akkas, O. P., E en, M. Y., Cam, E., & Inanc, N. (2017). Op imal Si e Selec ion o a
Sola Powe Plan in he Cen al Ana olian Region o Tu key. In e na ional
Jou nal o Pho oene gy, 2017(7 June 2017).
h ps://doi.o g/10.1155/2017/7452715
Al-Ruzouq, R., Shanableh, A., Yilmaz, A. G., Id is, A. E., Mukhe jee, S., Khalil, M.
A., & Gib il, M. B. A. (2019). Dam si e sui abili y mapping and analysis using
an in eg a ed GIS and machine lea ning app oach. Wa e (Swi ze land), 11(9).
h ps://doi.o g/10.3390/w11091880
Alami Me ouni, A., Elwali Elalaoui, F., Mez hab, A., Mez hab, A., & Ghennioui, A.
(2018). La ge scale PV si es selec ion by combining GIS and Analy ical
Hie a chy P ocess. Case s udy: Eas e n Mo occo. Renewable Ene gy, 119, 863–
873. h ps://doi.o g/10.1016/j. enene.2017.10.044
Ali, R., Hussain, A., Nazi , S., Khan, S., & Khan, H. U. (2023). In elligen Decision
Suppo Sys ems—An Analysis o Machine Lea ning and Mul ic i e ia Decision-
Making Me hods. Applied Sciences, 13(22), 12426.
h ps://doi.o g/10.3390/app132212426
Almansi, K. Y., Sha i , A. R. M., Abdullah, A. F., & Ismail, S. N. S. (2021). Hospi al
si e sui abili y assessmen using h ee machine lea ning app oaches: E idence
om he gaza s ip in Pales ine. Applied Sciences (Swi ze land), 11(22), 1–22.
h ps://doi.o g/10.3390/app112211054
Asadi, M., & Pou hossein, K. (2021). Neu al ne wo k-based modelling o wind/sola
a m si ing: a case s udy o Eas -Aze baijan. In e na ional Jou nal o Sus ainable
Ene gy, 40(7), 616–637. h ps://doi.o g/10.1080/14786451.2020.1833881
Asadi, M., Pou hossein, K., Noo ollahi, Y., Ma zband, M., & Iglesias, G. (2023). A
New Decision F amewo k o Hyb id Sola and Wind Powe Plan Si e Selec ion
Using Linea Reg ession Modeling Based on GIS-AHP. Sus ainabili y
(Swi ze land), 15(10). h ps://doi.o g/10.3390/su15108359
Annexes A
64
Figu e A.3: ROC cu e o mode a ely sui able class
Figu e A.4: ROC cu e o ma ginally sui able class
Annexes A
65
Figu e A.5: ROC cu e o highly sui able class
66
Annexes B
Figu e B.2: Signi icance o condi ioning ac o s o he p edic ions o ma ginally no sui able class
Figu e B.1: Signi icance o condi ioning ac o s o he p edic ions o pe manen ly no sui able class
Annexes B
67
Figu e B.3: Signi icance o condi ioning ac o s o he p edic ions o ma ginally sui able class
Figu e B. 4: Signi icance o condi ioning ac o s o he p edic ions o highly sui able class
68
Sui abili y mapping o sola powe plan s using an
explainable AI-based app oach
2024
Mawanane Hewa Rasanka Mangala De Sil a
Guia pa a a o ma ação de eses Ve são 4.0 Janei o 2006
69