Con en s lis s a ailable a ScienceDi ec
In o ma ion Fusion
jou nal homepage: www.else ie .com/loca e/in us
Spa io- empo al da a usion amewo k based on la ge language model o
enhanced p edic ion o elec ic ehicle cha ging demand in sma g id
managemen
Yi ong Shang a, Wen-Long Shang b,c,∗, Dingsong Cui c, Peng Liu d,e, , Haibo Chen c,
Dongdong Zhang g, Runsen Zhang h, Chengcheng Xu i, Ye Liu c, Chenxi Wang c,
Mohannad Alhazmi j
aDepa men o Ci il and En i onmen al Enginee ing, The Hong Kong Uni e si y o Science and Technology, Hong Kong, China
bDepa men o Ci il and En i onmen al Enginee ing, Impe ial College London, UK
cIns i u e o T anspo S udies, Uni e si y o Leeds, 34-40 Uni e si y Road, Leeds LS2 9JT, UK
dCollabo a i e Inno a ion Cen e o Elec ic Vehicles, Beijing, China
eNa ional Enginee ing Resea ch Cen e o Elec ic Vehicles, Beijing Ins i u e o Technology, Beijing 100081, China
Beijing Ins i u e o Technology Chongqing Inno a ion Cen e, Chongqing 401120, China
gRenewable Ene gy School, Inne Mongolia Uni e si y o Technology, O dos Ci y, China
hG adua e School o F on ie Sciences, The Uni e si y o Tokyo, Kashiwa 2778563, Japan
iSchool o T anspo a ion, Sou heas Uni e si y, Nanjing 210096, China
jElec ical Enginee ing Depa men , College o Applied Enginee ing, King Saud Uni e si y, Riyadh, Saudi A abia
A R T I C L E I N F O
Keywo ds:
Elec ic ehicle
Cha ging demand p edic ion
Spa io empo al da a usion
La ge language models
Model usion
Low- ank adap a ion
A B S T R A C T
Accu a e p edic ion o elec ic ehicle (EV) cha ging demand is pi o al o e ec i e sma g id managemen
and enewable ene gy in eg a ion. Howe e , p edic ing spa io- empo al EV cha ging pa e ns emains chal-
lenging due o complex da a usion equi emen s a ising om he e ogeneous empo al, spa ial, and con ex ual
ea u es, as well as di icul ies in e ec i ely in eg a ing mul iple modeling app oaches. This pape in oduces
EV-STLLM, a no el spa io- empo al da a usion amewo k based on La ge Language Model explici ly designed
o accu a e sho - e m EV cha ging demand o ecas ing h ough inno a i e in eg a ion o da a-le el and
model-le el usion echniques. A he da a le el, a mul i-sou ce embedding module is de eloped o seamlessly
use empo al ea u es (e.g., ime slo s, weekdays), spa ial he e ogenei y (e.g., geog aphical loca ion), and
con ex ual cha ging beha io s in o a uni ied ep esen a ion ia embedding con olu ional ne wo k. A he model
le el, a la ge language model (LLM) is employed o cap u e global spa io empo al dependencies, enhanced
wi h Low-Rank Adap a ion (LoRA) o pa ame e -e icien ine- uning, subs an ially educing compu a ional
cos s while main aining p edic ion obus ness. Using a comp ehensi e eal-wo ld da ase comp ising o e
830,000 EV cha ging eco ds ac oss 16 dis ic s and 331 subdis ic s in Beijing, we alida e EV-STLLM
ac oss mul iple o ecas ing scena ios (dis ic and subdis ic le els, one-s ep and wo-s ep ahead p edic ions).
Ex ensi e compa a i e e alua ions demons a e ha EV-STLLM consis en ly ou pe o ms classical, g aph-based,
and deep lea ning baselines. Speci ically, in one-s ep ahead dis ic -le el o ecas ing, EV-STLLM achie es up o
a 15.41% educ ion in MAE and a 53.51% educ ion in MAPE compa ed o he leading baseline, unde sco ing
i s po en ial o signi ican ly enhance da a-d i en sma g id ope a ions.
1. In oduc ion
Wi h he accele a ed global ansi ion owa ds ca bon neu ali y,
Elec ic Vehicles (EVs) a e shi ing om being an ‘‘op ion’’ in he
anspo a ion e olu ion o a ‘‘necessi y’’ in u ban low-ca bon ans o -
ma ion [1]. As o 2023, he numbe o EVs in China has su passed 20
∗Co esponding au ho .
E-mail add ess: [email p o ec ed] (W.-L. Shang).
million, and i is expec ed o each 160 million by 2030 [2]. Howe e ,
he apid g ow h o EV owne ship has led o unp eceden ed scheduling
p essu es on u ban powe sys ems [3]. EV cha ging beha io exhibi s
signi ican spa io empo al cha ac e is ics, wi h high ola ili y and con-
cen a ion in di e en ime pe iods and spa ial egions, c ea ing new
challenges o g id s abili y [4]. Pa icula ly du ing peak elec ici y
h ps://doi.o g/10.1016/j.in us.2025.103692
Recei ed 26 Ap il 2025; Recei ed in e ised o m 14 Augus 2025; Accep ed 1 Sep embe 2025
In o ma ion Fusion 126 (2026) 103692
A ailable online 9 Sep embe 2025
1566-2535/© 2025 The Au ho s. Published by Else ie B.V. This is an open access a icle unde he CC BY-NC-ND license ( h p://c ea i ecommons.o g/licenses/by-
nc-nd/4.0/ ).
Y. Shang e al.
usage pe iods o in speci ic a eas, concen a ed EV cha ging may cause
localized load su ges, e en leading o powe supply bo lenecks [5].
In his con ex , conduc ing spa io empo al p edic ions o sho - e m
EV cha ging demand is o g ea impo ance o sma g id managemen
and p icing s a egies [6]–[7]. By accu a ely p edic ing he cha ging
demand dis ibu ion in di e en egions o e he nex hou o nex
ew hou s, powe companies can de ise ime-o -use p icing s a egies,
guiding use s o a oid peak load pe iods and imp o ing he ope a-
ional e iciency o he powe sys em [8]. Addi ionally, he p edic ion
esul s can suppo g id scheduling, enabling dynamic load adjus men
and op imized ene gy alloca ion, educing he peak– alley di e ence,
and enhancing g id secu i y and lexibili y [9]. Mo eo e , sho - e m
p edic ions can imp o e use expe ience, such as by in elligen ly ec-
ommending cha ging imes o loca ions, educing queuing ime, and
imp o ing cha ging con enience and cos -e ec i eness [10]. The e-
o e, sho - e m spa io empo al p edic ion o EV cha ging demand is
no only an e ec i e means o add essing g id ope a ional p essu e,
bu also an indispensable echnological suppo in he de elopmen
o g een anspo a ion and sma ene gy [11]. Howe e , despi e ex-
ensi e esea ch on modeling and p edic ing EV cha ging demand,
achie ing high-accu acy sho - e m spa io empo al p edic ions a he
u ban scale s ill aces se e al challenges, especially in da a usion and
model usion [12]. Speci ically, da a-le el usion ocuses on p ep ocess-
ing and in eg a ing aw o in e media e da a om di e en sou ces
o c ea e a uni ied ep esen a ion o p edic ion models, while model-
le el usion emphasizes he combina ion o ou pu s o in e media e
ep esen a ions om mul iple models o imp o e o e all accu acy and
obus ness. The de ails o da a-le el usion and model-le el usion a e
desc ibed as ollows.
On he one hand, al hough exis ing esea ch has ex ac ed a la ge
numbe o ea u es om empo al, spa ial, and use beha io dimen-
sions, e icien ly in eg a ing hese mul i-dimensional and mul i-modal
ea u es emains one o he co e challenges in sho - e m EV cha ging
demand o ecas ing. Fi s ly, cha ging demand exhibi s e iden empo-
al pe iodici y (e.g., daily and weekly cycles) and empo al bu s iness
(e.g., du ing commu ing hou s). T adi ional ime se ies modeling ap-
p oaches o en s uggle o cap u e such nonlinea ends and sudden
luc ua ions. Fo example, one s udy epo ed ha adi ional me h-
ods (e.g., ARIMA) unde pe o med hei p oposed me hod (TSAGE)
by a ac o o 3.25 in e ms o RMSE [8]. Addi ionally, his o ical
da a o en include anomalous ime poin s such as holidays, ex eme
wea he , o unexpec ed e en s, u he complica ing he modeling and
usion o empo al ea u es [13]. Secondly, he spa ial dis ibu ion o
cha ging s a ions in u ban en i onmen s is highly une en. Cha ging
beha io s ac oss di e en egions a e in luenced by a ious ac o s such
as geog aphic loca ion, a ic condi ions, and su ounding acili ies.
E ec i ely modeling spa ial co ela ions (e.g., in luence p opaga ion
be ween neighbo ing egions) and spa ial he e ogenei y (e.g., beha -
io al di e ences be ween u ban cen e s and subu ban a eas) is a key
challenge o spa ial ea u e usion [14]. Finally, undamen al ea u es
such as use ype, cha ging s a ion ca ego y, and p icing mechanisms
ep esen s a ic in o ma ion whose in luence a ies unde di e en
spa io empo al con ex s. These s a ic ea u es in e ac in complex ways
wi h dynamic spa io empo al ea u es. The e o e, dynamically adjus -
ing he weigh s o s a ic in o ma ion in he modeling p ocess is an
u gen issue in cu en da a usion esea ch. Consequen ly, cons uc ing
a da a usion amewo k ha can lexibly adap o mul i-sou ce he -
e ogeneous da a and dynamically adjus ea u e weigh s is c ucial o
imp o ing p edic ion accu acy.
On he o he hand, in spa io empo al o ecas ing asks o sho -
e m EV cha ging demand, model usion has eme ged as a key s a egy
o enhancing p edic ion accu acy and sys em obus ness. The cu en
mains eam p edic ion models include small models, la ge models, and,
mo e ecen ly, la ge language models (LLMs). Each model ca ego y
o e s ad an ages in ea u e modeling, ep esen a ion capaci y, and
compu a ional e iciency, bu also aces speci ic challenges in p ac ical
applica ion. Small models, such as linea eg ession (LR), decision
ees, and suppo ec o machines, o e high aining e iciency, low
compu a ional cos , simple s uc u e, and ease o deploymen . How-
e e , hey a e limi ed in cap u ing complex nonlinea ela ionships
o long- e m dependencies, making hem less sui able o la ge-scale,
mul i-dimensional u ban-le el cha ging demand o ecas ing [15]. La ge
models, such as con olu ional neu al ne wo ks (CNNs), ecu en neu-
al ne wo ks (RNNs), and T ans o me s, possess signi ican ad an ages
in spa io empo al modeling, capable o unco e ing deepe sequen ial
pa e ns and spa ial dependencies in cha ging beha io s. Ne e heless,
hese models equi e subs an ial da a and compu a ional esou ces o
aining, a e cos ly o implemen , and a e p one o o e i ing [16].
Recen ly, LLMs, such as GPT and BERT, ha e achie ed ema kable
p og ess in na u al language p ocessing, especially in unde s anding
complex seman ics and gene a ing con ex -awa e con en . Howe e ,
hese models we e no o iginally designed o spa io empo al sequence
p edic ion asks. Di ec ly applying LLMs o s uc u ed ime se ies mod-
eling may encoun e s uc u al misma ches and ask adap a ion chal-
lenges [17]. In summa y, al hough in eg a ing small models, la ge
models, and LLMs could le e age hei complemen a y s eng hs, in
p ac ice, model usion aces se e al di icul ies, including inconsis en-
cies in model a chi ec u es, he e ogenei y in ea u e inpu o ma s, and
complexi y in designing e ec i e usion s a egies.
To add ess he a o emen ioned challenges, his pape p oposes a
spa io empo al usion amewo k ha in eg a es bo h da a-le el and
model-le el s a egies, aiming o enhance he accu acy and obus ness
o sho - e m elec ic ehicle cha ging demand o ecas ing a he u ban
scale. Le e aging eal-wo ld EV cha ging da a om Beijing—co e ing
331 sub-dis ic s ac oss 16 adminis a i e dis ic s— his s udy demon-
s a es how mul i-sou ce, ci y-scale da a usion can e ec i ely suppo
in elligen demand o ecas ing and u ban ene gy scheduling, p o iding
c i ical insigh s o he de elopmen o sma g ids and low-ca bon
ci ies. The main con ibu ions o his pape include:
(1) Uni ied spa io empo al da a usion h ough collabo a i e em-
beddings. To add ess he he e ogenei y o mul i-sou ce da a
in EV cha ging demand o ecas ing, we design a collabo a i e
embedding mechanism ha enables e ec i e da a-le el usion
ac oss empo al, spa ial, and beha io al dimensions. Speci i-
cally, empo al pa e ns a e encoded ia ime-slo and weekday
embeddings; spa ial he e ogenei y is cap u ed h ough nonlinea
ans o ma ions o u ban egion ea u es; and localized cha ging
beha io s a e ep esen ed using poin wise con olu ion. These
componen s a e in eg a ed ia an Embedding Con olu ional Ne -
wo k (ECN), cons uc ing a uni ied spa io empo al ep esen-
a ion capable o cap u ing bo h s a ic and dynamic ea u e
in e ac ions ac oss ime and space.
(2) Model-le el usion ia LoRA-enhanced la ge language model. To
ully le e age he complemen a y s eng hs o adi ional ea u e
embeddings and la ge-scale p e ained models, we p opose a
model usion s a egy ha in eg a es ligh weigh spa io empo al
encodings wi h a pa ame e -e icien ine- uned LLM. By adop -
ing Low-Rank Adap a ion (LoRA), we inse ainable low- ank
ma ices in o he T ans o me ’s Que y and Value p ojec ions
while keeping he p e ained backbone ozen. This app oach
signi ican ly educes aining o e head, enhances scalabili y,
and enables he LLM o adap o s uc u ed o ecas ing asks
wi hou comp omising i s gene aliza ion abili y.
(3) Ex ensi e eal-wo ld alida ion wi h supe io quan i a i e pe -
o mance ac oss scales. We cons uc a la ge-scale EV cha ging
da ase spanning 16 dis ic s and 331 subdis ic s in Beijing, and
e alua e he p oposed model ac oss mul iple o ecas ing sce-
na ios. Compa ed o s ong baselines (e.g., GRU, G aphSAGE),
EV-STLLM achie es up o 15.45% educ ion in MAE and 53.51%
educ ion in MAPE o dis ic -le el 1-s ep p edic ion, and up o
19.61% and 26.39% educ ions in MAE and RMSE espec i ely
In o ma ion Fusion 126 (2026) 103692
2
Y. Shang e al.
a he subdis ic le el. These esul s demons a e he model’s su-
pe io accu acy, obus ness, and ine-g ained adap abili y ac oss
bo h coa se and ine spa ial esolu ions.
The es o his pape is o ganized as ollows. Sec ion 2 e iews
he ela ed li e a u e on EV cha ging demand p edic ion, spa io em-
po al modeling, and he applica ion o LLMs in o ecas ing. Sec ion 3
o mally de ines he p oblem o sho - e m EV cha ging demand p e-
dic ion. Sec ion 4 de ails he p oposed me hodology, including he
o e all amewo k, he spa io empo al ea u e embedding p ocess, he
T ans o me -based a chi ec u e, and he pa ame e -e icien ine- uning
s a egy. Sec ion 5 desc ibes he expe imen al se up, p esen s he e-
sul s o compa a i e and abla ion s udies, and analyzes he model’s
sensi i i y. Finally, Sec ion 6 concludes he pape , summa izing he key
indings and discussing he p ac ical implica ions and limi a ions o he
wo k.
2. Rela ed wo k
2.1. EV cha ging demand p edic ion
EV cha ging demand p edic ion has become an impo an esea ch
di ec ion in he ields o in elligen anspo a ion and sma ene gy [18].
Exis ing EV cha ging demand p edic ion me hods can be b oadly di-
ided in o h ee ca ego ies: s a is ical models, adi ional machine
lea ning me hods, and deep lea ning me hods. S a is ical models, such
as au o eg essi e in eg a ed mo ing a e age (ARIMA) [19] and sea-
sonal ARIMA (SARIMA) [20] a e cha ac e ized by simplici y in mod-
eling and s ong in e p e abili y, making hem sui able o ela i ely
s able ime se ies p edic ion asks. Howe e , hei abili y o model
nonlinea ela ionships is limi ed, and hey pe o m poo ly when deal-
ing wi h complex and a iable cha ging beha io s. Machine lea ning
me hods, such as suppo ec o eg ession (SVR), andom o es ,
and g adien boos ing decision ees (GBDT) a e capable o cap u ing
nonlinea ea u es o some ex en . These me hods a e sui able o
medium- o small-scale da ase s, bu hey hea ily depend on ea u e
enginee ing [21]. In ecen yea s, deep lea ning models based on
RNN [22], CNN [23], and hei a ian s ha e been widely applied
in cha ging demand p edic ion. These models possess s ong ea u e
ex ac ion and pa e n ecogni ion capabili ies, especially in handling
high-dimensional complex da a and cap u ing empo al and spa ial
dependencies.
The a ia ion in EV cha ging demand is d i en by mul iple ac o s,
and cons uc ing a high-quali y p edic ion model equi es comp e-
hensi e conside a ion and modeling o key in luencing ac o s [24].
Tempo al ea u es, such as hou s, days o he week, holidays, e c.,
e lec ing he pe iodici y and egula i y o cha ging beha io . Spa ial
ea u es, including dis ic s and subdis ic s, which e lec egional
di e ences in popula ion, a ic, and in as uc u e, he eby in lu-
encing cha ging beha io . Cen al u ban a eas end o ha e highe
cha ging demands due o dense commu ing, while subu ban a eas
a e signi ican ly in luenced by esiden ial dis ibu ion and cha ging
s a ion co e age. P ope modeling o spa ial hie a chy helps imp o e
spa ial accu acy and gene aliza ion abili y in p edic ions. En i on-
men al ac o s, such as wea he ( empe a u e, p ecipi a ion, e c.) and
ai quali y, which may indi ec ly a ec ehicle a el equency and
cha ging demand. The e o e, he usion o mul i-sou ce he e ogeneous
da a has become an impo an way o enhance model pe o mance, and
es ablishing e ec i e co ela ions be ween di e en da a ypes emains
a c i ical esea ch opic.
2.2. Spa io empo al p edic ion
Cha ging demand p edic ion alls unde he b oade ca ego y o spa-
io empo al p edic ion. The e o e, when cons uc ing p edic ion mod-
els, ime and space ea u es should no be ea ed sepa a ely bu should
be unde s ood in e ms o hei coupling ela ionship. Du ing he de-
elopmen o spa io empo al modeling me hods, esea che s ha e p o-
posed a ious a chi ec u es o enhance he abili y o cap u e complex
spa io empo al ea u es. Ea ly me hods o en employed combina ions
o CNNs and RNNs, whe e he o me was used o cap u e spa ial
dependencies and he la e was used o model empo al dynamics.
Models such as CNN-long sho - e m memo y ne wo ks (LSTM) [25]
and Con LSTM [26] ha e imp o ed he accu acy o spa io empo al
sequence modeling o some ex en bu ha e limi a ions when dealing
wi h non-Euclidean spa ial s uc u es, such as u ban oad ne wo ks.
To add ess hese issues, g aph neu al ne wo ks (GNNs) ha e been
in oduced in o spa io empo al modeling in ecen yea s. Rep esen-
a i e models such as spa io empo al g aph con olu ional ne wo ks
(STGCN) [27] and di usion con olu ional ecu en ne wo ks
(DCRNN) [28] build spa ial g aph s uc u es o cap u e non-Euclidean
ela ionships be ween nodes and in eg a e ime se ies modeling ech-
niques o e icien ly p edic spa io empo al da a like a ic low and
ene gy consump ion. Xing e al. p oposed a spa io empo al usion ne -
wo k ha in eg a es GCN wi h LSTM-A en ion, speci ically designed
o u ban ail ansi OD low p edic ion [29]. Xu e al. in oduced a
no el anspo a ion p edic ion model, he HSTGODE, which employs
a dual-laye s uc u e combined wi h spa io empo al o dina y di e en-
ial equa ion modules o add ess he o e -smoo hing p oblem in GNNs
and e ec i ely cap u e hie a chical spa io empo al ea u es a bo h
egional and node le els [30]. These me hods ha e achie ed signi ican
esul s in u ban a ic, powe load, and o he ields, u he expanding
he echnical bounda ies o spa io empo al p edic ion.
Wi h b eak h oughs in he T ans o me model in na u al language
p ocessing (NLP) o sequence modeling, esea che s ha e begun o
in oduce hese models in o spa io empo al p edic ion, de eloping spa-
io empo al T ans o me a chi ec u es [31]. These models u ilize sel -
a en ion mechanisms, dynamically cap u ing global spa io empo al
dependencies, and a e be e a modeling long- ange ea u es. Fo
example, models like spa io- empo al a en ion ne wo k (STAN) ha e
shown supe io pe o mance in mul iple anspo a ion and ene gy
p edic ion asks compa ed o adi ional CNN–RNN s uc u es [32]. Yu
e al. p oposed he MGSF o me , which u ilizes a esidual edundancy
elimina ion module o emo e in o ma ion edundancy ac oss di e en
g anula i ies o da a. Fu he mo e, i inco po a es spa io empo al a en-
ion and dynamic usion modules o achie e e icien ai quali y p edic-
ion [33]. Zhang e al. in oduced he EF- o me , a deep lea ning-based
mul is ep passenge low p edic ion model. By in eg a ing he pa allel
in e ac i e a en ion module and mul i-scale causal mul i-Head sel -
a en ion module, EF- o me ex ac s bo h global and local empo al
dependencies, enabling p ecise modeling o spa io empo al dynamics
du ing la ge-scale e en s and ealizing accu a e mul is ep o ecas ing
o passenge low [34].
2.3. P edic ion-o ien ed applica ions o la ge language models
In ecen yea s, LLMs such as he GPT se ies and Llama se ies ha e
achie ed ema kable success in he ield o NLP. As he capabili ies
o LLMs con inue o expand, esea che s ha e begun explo ing hei
po en ial applica ions in s uc u ed da a modeling, especially in spa-
io empo al p edic ion asks. These applica ions can be summa ized
in o he ollowing h ee pa hways [35].
The i s is knowledge ex ac ion and ea u e enhancemen [36].
This app oach uses LLMs o pe o m deep seman ic unde s anding
and implici knowledge ex ac ion om mul imodal da a (such as
ex , images, e c.). The ex ac ed seman ic ea u es a e hen inpu
In o ma ion Fusion 126 (2026) 103692
3
Y. Shang e al.
in o adi ional p edic ion models (such as LSTM, GNN, e c.) o en-
hance he model’s abili y o pe cei e backg ound in o ma ion, beha -
io al pa e ns, and en i onmen al ac o s. In his p ocess, LLMs ac as
knowledge encode s, e ec i ely supplemen ing and enhancing ea u e
seman ics wi hou changing he o iginal model s uc u e.
The second is ex - o-LLM app oach [37]. In his me hod, aw
spa io empo al s uc u ed da a (such as ime se ies, spa ial loca ions,
wea he , e c.) is ans o med in o na u al language desc ip ions, which
a e hen okenized and inpu in o ozen o ine- uned LLMs o p e-
dic ion. This app oach has good gene ali y and in e p e abili y, as i
can lexibly adap o any p e- ained language model and imp o e he
human eadabili y o model ou pu s. Howe e , due o limi a ions in
exp essing complex spa io empo al dependencies, his me hod aces is-
sues such as high oken usage cos s, limi ed con ex windows, and insu -
icien exp ession accu acy, making i unsui able o high-dimensional
and mul i-scale complex p edic ion asks. Kang e al. p oposed an en-
hanced mul i-le el heal h e en p edic ion amewo k, LLM-DG, which
imp o es p edic ion accu acy and obus ness by seman ically enhanc-
ing he ep esen a ion o pa ien s and discha ge summa ies, injec ing
domain knowledge, cap u ing highe -o de co ela ions, and in eg a -
ing dynamic and s a ic ea u es o simul aneously model in e -pa ien
clus e ing and in a-pa ien disease e olu ion cha ac e is ics [38]. Shen
e al. in oduced a amewo k ha employs ein o cemen lea ning o
p o ide decisi e subg aph in o ma ion o G aph LLMs. By u ilizing
a ein o cemen subg aph de ec ion module and a node-guidance ne -
wo k, he amewo k sea ches o and deli e s c i ical neighbo hood
and node in o ma ion in ex ual o m o assis LLM p edic ions, wi hou
equi ing model e aining [39].
The hi d is ine- uning LLMs o spa io empo al asks [40]. This
me hod encodes spa io empo al da a as s uc u ed oken sequences
(such as imes amps, egion codes, nume ical a ibu es, e c.) and ine-
unes LLMs unde supe ision, allowing hem o unde s and he spa-
io empo al meaning behind hese okens and pe o m p edic ion asks.
Compa ed o na u al language desc ip ions, his app oach has highe
oken e iciency and exp ession accu acy, be e modeling highe -o de
empo al dependencies and spa ial ela ionships, and s onge adap -
abili y. Howe e , he ine- uning p ocess ypically equi es high com-
pu a ional esou ces and da a quali y.
2.4. Resea ch gap
Al hough signi ican p og ess has been made in EV cha ging de-
mand p edic ion, spa io empo al modeling, and he applica ion o
LLMs, se e al in e connec ed challenges emain. In EV cha ging de-
mand p edic ion, exis ing me hods s uggle o e ec i ely model he
nonlinea and high-dimensional in e ac ions be ween empo al, spa ial,
and beha io al ea u es, pa icula ly in sho - e m u ban-scale asks.
While deep lea ning models ha e imp o ed p edic i e pe o mance,
hey o en ail o adequa ely use mul i-sou ce he e ogeneous da a,
such as spa ial hie a chies, empo al pe iodici y, and en i onmen al
ac o s, leading o subop imal gene aliza ion and scalabili y. In he
b oade domain o spa io empo al p edic ion, adi ional a chi ec u es
like CNN–RNN hyb ids and Con LSTM ace di icul ies in explici ly
cap u ing complex ea u e in e ac ions and balancing he ade-o be-
ween global gene aliza ion and local de ail i ing. Ad anced models,
such as GNNs and spa io empo al T ans o me s, ha e add essed some
o hese issues bu s ill s uggle wi h mul i-scale and highly dynamic
scena ios, pa icula ly in non-Euclidean spa ial s uc u es like u ban
oad ne wo ks. Meanwhile, he eme ging applica ion o LLMs in p edic-
ion asks highligh s hei po en ial o seman ic knowledge ex ac ion
and ea u e enhancemen , ye challenges such as e icien okeniza ion,
ep esen a ion o spa io empo al dependencies, and he compu a ional
cos o ine- uning emain signi ican ba ie s. These gaps collec i ely
indica e he need o a uni ied amewo k ha in eg a es he s eng hs
o spa io empo al a chi ec u es wi h he seman ic unde s anding ca-
pabili ies o LLMs. The ocus should be on achie ing e icien da a
usion ac oss mul i-sou ce he e ogeneous da ase s and model usion,
p o iding high-pe o mance p edic ion sys ems o in elligen ene gy
and u ban-scale applica ions.
3. P oblem de ini ion
This wo k ocuses on he p oblem o sho - e m EV cha ging de-
mand p edic ion, which aims o es ima e he u u e cha ging demand
ac oss mul iple spa ial loca ions in a ci y o e disc e e ime in e als.
This is a ypical spa io empo al o ecas ing ask, whe e bo h empo al
dynamics and spa ial he e ogenei y mus be join ly modeled.
Fo mally, le he ci y be pa i ioned in o 𝑁 spa ial uni s (dis ic
o subdis ic ), and ime be di ided in o disc e e in e als (30 min).
Fo each loca ion and ime slo , we obse e a se o ea u es desc ibing
cha ging beha io and con ex ual ac o s. The objec i e is o p edic
he EV cha ging demand o each loca ion in he nex ew ime s eps
based on his o ical da a. Le ∈R𝑇𝑖𝑛×𝑁𝑛𝑜𝑑𝑒𝑠×𝐶𝑖𝑛 deno e he his o ical
mul i a ia e inpu enso , whe e: 𝑇𝑖𝑛 is he numbe o his o ical ime
s eps, 𝑁𝑛𝑜𝑑𝑒𝑠 is he numbe o spa ial loca ions, 𝐶𝑖𝑛 is he numbe o
ea u es pe loca ion pe ime s ep (e.g., pas cha ging demand, ime
slo index, week o day, e c.).
Le
𝑌∈R𝑇𝑜𝑢𝑡×𝑁𝑛𝑜𝑑𝑒𝑠 be he p edic ed cha ging demand o e a
o ecas ing ho izon o 𝑇𝑜𝑢𝑡 ime s eps o 𝑁𝑛𝑜𝑑𝑒𝑠 loca ions. The goal is
o lea n a mapping unc ion 𝑓(⋅) such ha :
𝑌𝑇+1∶𝑇+𝐻=𝑓(1,2,…,𝑇)(1)
He e, 𝑓 is a p edic i e model capable o lea ning complex spa-
io empo al dependencies om he inpu da a. This p edic ion unc ion
should cap u e: Tempo al dependencies, such as daily/weekly cha ging
pa e ns, peak s. o -peak hou s, and holiday e ec s. Spa ial dependen-
cies, such as mobili y pa e ns ac oss dis ic s, cha ging in as uc u e
densi y, and popula ion dis ibu ion. Ex e nal ac o s, such as wea he
condi ions, public e en s, and oad a ic s a us, which may impac ip
equency and cha ging beha io .
4. Me hodology
4.1. F amewo k
The p oposed amewo k o EV cha ging demand p edic ion u ilizes
a Spa io empo al La ge Language Model (EV-STLLM) as shown in
Fig. 1. This amewo k combines spa io empo al embeddings wi h a
ans o me -based a chi ec u e o model complex dependencies ac oss
ime and space, aiming o enhance he accu acy and obus ness o
sho - e m cha ging demand o ecas s in u ban en i onmen s. T adi-
na ional spa io- empo al GCN-based p edic ion me hods ypically ely
on ixed g aph s uc u es o p ede ined a en ion pa e ns, which limi
hei abili y o adap o dynamic spa ial ela ionships. Ou p oposed
amewo k is g aph- ee and le e ages posi ion-awa e embeddings and
T ans o me a en ion o model global spa io empo al dependencies
wi hou assuming any igid s uc u e. Fu he mo e, by in eg a ing
LoRA o e icien ine- uning, EV-STLLM achie es a be e balance
be ween adap abili y and compu a ional e iciency, which is especially
impo an o eal-wo ld u ban compu ing scena ios
The his o ical inpu da a, ∈R𝑇𝑖𝑛×𝑁𝑛𝑜𝑑𝑒𝑠×𝐶𝑖𝑛 , is con e ed in o o-
kens ep esen ing spa ial and empo al ea u es. Th ee ypes o
embeddings—Auxilia y, Tempo al, and Spa ial—a e applied and used
in o a uni ied spa io empo al ep esen a ion. The spa io empo al em-
beddings a e eshaped and passed in o a ans o me a chi ec u e
wi h a masked mul i-head sel -a en ion mechanism. This sel -a en ion
enables he model o cap u e bo h global and local dependencies
in ime and space, imp o ing he model’s abili y o lea n complex
pa e ns. To adap he p e ained language model o he EV cha ging
demand ask while minimizing o e i ing, he LoRA echnique is used.
LoRA allows he model o adjus only a small se o pa ame e s,
e ec i ely ans e ing knowledge wi h ewe compu a ional esou ces.
A e p ocessing h ough he ans o me laye s, he model p oduces
he inal p edic ions using a eg ession con olu ion laye . The loss
unc ion combines he p edic ion e o wi h a egula iza ion e m o
p e en o e i ing.
In o ma ion Fusion 126 (2026) 103692
4
Y. Shang e al.
Fig. 1. F amewo k o his wo k, (a) EV-STLLM, (b) T ans o me ne wo k,(c) decode -based T ans o me , and (d) LoRA mechanism.
4.2. Spa io empo al ea u e embedding
As illus a ed in Fig. 1(a), we conside he inpu da a
∈R𝑇𝑖𝑛×𝑁𝑛𝑜𝑑𝑒𝑠×𝐶𝑖𝑛 as a sequence composed o spa ial and empo al
in o ma ion. To comp ehensi ely cap u e i s unde lying pa e ns, we
de ise h ee dis inc embedding mechanisms.
Auxilia y ea u e embedding. We i s p ocess he aw inpu h ough a
poin wise con olu ional laye o gene a e an auxilia y ea u e embed-
ding:
𝐇𝑎𝑢𝑥 =Poin wiseCon (;𝛩𝑎𝑢𝑥),𝐇𝑎𝑢𝑥 ∈R𝑁𝑛𝑜𝑑𝑒𝑠×𝑑𝑚𝑜𝑑𝑒𝑙 (2)
whe e Poin wiseCon (⋅) deno es a 1 × 1 con olu ion ope a ion wi h
ainable pa ame e s 𝛩𝑎𝑢𝑥. 𝐇𝑎𝑢𝑥 is he esul ing node embedding ma-
ix, and 𝑑𝑚𝑜𝑑𝑒𝑙 de ines he dimension o he embedding ec o s. We
employ poin wise con olu ion due o i s excep ional compu a ional
e iciency. I e ec i ely cap u es local in e - ea u e ela ionships and
comp esses he high-dimensional inpu in o a mo e compac embedding
space, p o iding a e ined ea u e ep esen a ion o he subsequen
spa io empo al usion.
Tempo al in o ma ion embedding (hou and day-o -week). To encode
empo al pe iodici ies, we c ea e lea nable embedding ec o s o di -
e en ime slo s o he day and days o he week:
𝐇𝑡𝑖𝑚𝑒 =𝐖ℎ𝑜𝑢𝑟(ℎ𝑜𝑢𝑟) + 𝐖𝑑𝑎𝑦(𝑑𝑎𝑦),𝐇𝑡𝑖𝑚𝑒 ∈R𝑁𝑛𝑜𝑑𝑒𝑠×𝑑𝑚𝑜𝑑𝑒𝑙 (3)
He e, ℎ𝑜𝑢𝑟 and 𝑑𝑎𝑦 a e he hou -o -day and day-o -week indices o
each node, espec i ely. 𝐖ℎ𝑜𝑢𝑟 ∈R𝑇ℎ𝑜𝑢𝑟×𝑑𝑚𝑜𝑑𝑒𝑙 and 𝐖𝑑𝑎𝑦 ∈R𝑇𝑑𝑎𝑦×𝑑𝑚𝑜𝑑𝑒𝑙
a e wo lea nable embedding lookup ables. 𝐇𝑡𝑖𝑚𝑒 is he inal empo al
embedding ma ix.
Spa ial in o ma ion embedding. We ex ac spa ial in o ma ion di ec ly
om he inpu ea u es using a ully connec ed laye :
𝐇𝑠𝑝𝑎𝑐𝑒 =𝜙(⋅𝐖𝑠𝑝𝑎𝑐𝑒 +𝐛𝑠𝑝𝑎𝑐𝑒),𝐇𝑠𝑝𝑎𝑐𝑒 ∈R𝑁𝑛𝑜𝑑𝑒𝑠×𝑑𝑚𝑜𝑑𝑒𝑙 (4)
whe e 𝐖𝑠𝑝𝑎𝑐𝑒 ∈R𝐶𝑖𝑛×𝑑𝑚𝑜𝑑𝑒𝑙 and 𝐛𝑠𝑝𝑎𝑐𝑒 ∈R𝑑𝑚𝑜𝑑𝑒𝑙 a e he weigh ma ix
and bias ec o o he laye , espec i ely. 𝜙(⋅) ep esen s a non-linea
ac i a ion unc ion (e.g., ReLU), and 𝐇𝑠𝑝𝑎𝑐𝑒 is he spa ial embedding
ma ix de i ed om he aw ea u es.
Finally, we in eg a e he h ee a o emen ioned embeddings o o m
a uni ied, mul i- ace ed ea u e ep esen a ion:
𝐇𝑓𝑢𝑠𝑒𝑑 =FusionLaye (𝐇𝑎𝑢𝑥||𝐇𝑡𝑖𝑚𝑒||𝐇𝑠𝑝𝑎𝑐𝑒;𝛩𝑓𝑢𝑠𝑒),𝐇𝑓 𝑢𝑠𝑒𝑑 ∈R𝑁𝑛𝑜𝑑𝑒𝑠 ×3𝑑𝑚𝑜𝑑𝑒𝑙 (5)
In his equa ion, || deno es he conca ena ion ope a ion along he
ea u e dimension. FusionLaye (⋅) is an ECN unc ion used o ea-
u e usion, wi h 𝛩𝑓 𝑢𝑠𝑒 as i s lea nable pa ame e s. 𝐇𝑓 𝑢𝑠𝑒𝑑 is he inal
used spa io empo al embedding ep esen a ion. We op o ECN as
he spa io empo al ea u e usion me hod because i e icien ly models
local dependencies, has low compu a ional complexi y, and ea u es
lexible lea nable pa ame e s. In ou expe imen s, i demons a ed su-
pe io pe o mance compa ed o bo h mo e complex and simple usion
al e na i es.
In o ma ion Fusion 126 (2026) 103692
5
Y. Shang e al.
4.3. T ans o me -based decode -only a chi ec u e
As depic ed in Fig. 1(b) and (c), we adop a T ans o me -based
decode -only a chi ec u e o model he complex spa io empo al depen-
dencies in EV cha ging demand. The model akes he used spa io em-
po al embeddings 𝐇𝑓𝑢𝑠𝑒𝑑 ∈R𝑁𝑛𝑜𝑑𝑒𝑠×3𝑑𝑚𝑜𝑑𝑒𝑙 and eshapes hem in o a
sequence o p ocessing:
𝐙(0) =Reshape(𝐇𝑓 𝑢𝑠𝑒𝑑 ) ∈ R𝑆×𝑑,whe e 𝑆=𝑁𝑛𝑜𝑑𝑒𝑠, 𝑑 = 3𝑑𝑚𝑜𝑑𝑒𝑙 (6)
Each decode block consis s o a masked mul i-head sel -a en ion
laye and a eed- o wa d ne wo k (FFN). Du ing spa io empo al mod-
eling, we le e age he sel -a en ion mechanism o he T ans o me
o cap u e global spa io empo al dependencies by compu ing a en-
ion weigh s ac oss di e en ime s eps and spa ial posi ions in he
inpu sequence. Speci ically, o he 𝑙 h laye , he a en ion mechanism
compu es he ollowing esul :
A en ion(𝐐,𝐊,𝐕) = so max (𝐐𝐊⊤
√𝑑𝑘
+𝐌)𝐕(7)
𝐙(𝑙)
𝑎𝑡𝑡𝑛 =Conca (head1,…,head𝐻)𝐖𝑂+𝐛𝑂(8)
𝐙(𝑙)
𝑟𝑒𝑠1=Laye No m(𝐙(𝑙)
𝑎𝑡𝑡𝑛 +𝐙(𝑙−1))(9)
𝐙(𝑙)
𝑓𝑓𝑛 =𝜙(𝐙(𝑙)
𝑟𝑒𝑠1𝐖(𝑙)
1+𝐛(𝑙)
1)𝐖(𝑙)
2+𝐛(𝑙)
2(10)
𝐙(𝑙)=Laye No m(𝐙(𝑙)
𝑓𝑓𝑛 +𝐙(𝑙)
𝑟𝑒𝑠1)(11)
Th ough his mechanism, he model is able o simul aneously cap-
u e long- ange dependencies ac oss mul iple ime s eps and spa ial
egions. Mo eo e , o enable he LLM o handle he e ogeneous in o ma-
ion, we se ialize empo al con ex , spa ial iden i ie s, and auxilia y be-
ha io al ea u es in o oken sequences as inpu s o he model, he eby
implici ly in eg a ing in o ma ion ac oss scales. A e 𝐿 laye s, he inal
ou pu is:
𝐇(𝐿)=𝐙(𝐿)∈R𝑆×𝑑𝑠𝑒𝑞 (12)
We hen apply a eg ession laye o gene a e he inal p edic ions:
𝐘𝑇+1∶𝑇+𝐻=RCon (𝐇(𝐿);𝛩𝑟)(13)
4.4. Pa ame e -e icien LoRA ine- uning
To adap he p e ained model o he cha ging demand p edic ion
ask while e aining i s p io knowledge, we employ Low-Rank Adap-
a ion (LoRA) o e icien ine- uning, as shown in Fig. 1(d). LoRA
injec s low- ank ma ices in o speci ic weigh ma ices (e.g., he 𝐐
and 𝐕 p ojec ions in mul i-head a en ion) and exclusi ely ains hese
newly in oduced ma ices while keeping he p e ained weigh s ozen.
Speci ically, o a gi en weigh ma ix 𝐖0∈R𝑑×𝑑, LoRA modi ies i
as:
𝐖LoRA =𝐖0+𝛼⋅𝐀𝐁 (14)
whe e 𝐀∈R𝑑×𝑟 and 𝐁∈R𝑟×𝑑 a e low- ank ma ices wi h 𝑟 ≪ 𝑑, and 𝛼
is a scaling ac o . The o iginal weigh ma ix 𝐖0 emains ozen, and
only 𝐀 and 𝐁 a e ainable. The key ad an ages o his design a e as
ollows:
•Compu a ional E iciency: By signi ican ly educing he numbe
o ainable pa ame e s (less han 1% o he ull model), LoRA
d as ically dec eases GPU memo y usage and aining ime.
•Robus ness: LoRA p ese es he o iginal knowledge and capaci y
o he p e ained model, mi iga ing he isk o o e i ing in
low-da a scena ios.
•S abili y: By only ine- uning he Que y and Value ma ices while
keeping o he componen s (e.g., 𝐾 and FFN) ozen, LoRA u he
educes he isk o ca as ophic o ge ing.
In ou model, we apply LoRA ine- uning o he Que y and Value
p ojec ion ma ices o each T ans o me block:
𝐐LoRA =𝐐0+𝛼𝑞⋅𝐀𝑞𝐁𝑞,𝐕LoRA =𝐕0+𝛼𝑣⋅𝐀𝑣𝐁𝑣(15)
The emaining ma ices, such as o he Key (𝐊), Ou pu (𝐎), and he
FFN, a e kep ozen o minimize he numbe o ainable pa ame e s
and p e en ca as ophic o ge ing. Th ough his app oach, we achie e
e icien cus omiza ion o domain-speci ic asks wi hou sac i icing
model pe o mance.
4.5. Loss unc ion
A e ob aining he ou pu ep esen a ions 𝐇(𝐿)∈R𝑁𝑛𝑜𝑑𝑒𝑠 ×𝑑𝑠𝑒𝑞 om
he T ans o me backbone, a eg ession con olu ion laye is applied o
p oduce he inal p edic ion:
𝐘𝑇+1 =RCon (𝐇(𝐿);𝛩𝑟)(16)
The loss unc ion is de ined as he sum o he p edic ion e o and a
egula iza ion e m on he LoRA pa ame e s:
=‖‖‖
𝐘𝑇+1 −𝐘𝑇+1‖‖‖
2
2+𝜆⋅‖𝛩LoRA‖2
2(17)
whe e 𝛩LoRA deno es he se o all ainable LoRA pa ame e s and 𝜆 is
he egula iza ion coe icien .
5. Expe imen s
5.1. Expe imen se up
5.1.1. Da ase desc ip ion
In his sec ion, we demons a e he e ec i eness o he p oposed
EV-STLLM amewo k h ough a case s udy based on Beijing, China. As
one o he leading ci ies in China o EV adop ion, Beijing has de eloped
a obus elec ic anspo a ion ne wo k. We u ilized EV cha ging log
da a om Decembe 2020, collec ed a a 30 min in e al, co e ing
16 dis ic s and 331 subdis ic egions ac oss he ci y. The da ase
includes 833,439 cha ging e en s, each wi h an a e age o 22.06 kWh
o ene gy consumed.
The spa ial dis ibu ion o cha ging demand ac oss a ious dis ic s
and subdis ic egions is depic ed in Fig. 2. Speci ically, Fig. 2(a) shows
he o al EV cha ging demand agg ega ed by dis ic . Da ke colo s
indica e dis ic s wi h highe o al ene gy consump ion, highligh ing
ha cen al dis ic s such as Chaoyang and Haidian exhibi signi ican ly
highe demand. Fig. 2(b) p o ides a ine g anula i y by illus a ing he
spa ial dis ibu ion o cha ging demand a he subdis ic le el. B igh e
colo s ep esen highe demand densi ies, e ealing ho spo s o EV
ac i i y wi hin he ci y.
Table 1 summa izes he maximum ins an aneous cha ging demand
and he cumula i e o al cha ging demand o each dis ic du ing he
obse a ion pe iod. In addi ion, Fig. 3 displays he empo al dynam-
ics o EV cha ging demand h oughou Decembe 2020. Speci ically,
Fig. 3(a) illus a es he cha ging demand o e ime a each dis ic ,
segmen ed in o aining, alida ion, and es ing pe iods, wi h a a io
o 8:1:1. The spli is empo ally consis en , meaning ea lie days a e
alloca ed o he aining se , and mo e ecen days a e assigned o he
es ing se . This s a egy p e en s da a leakage and ensu es he model
is ained on his o ical da a and e alua ed on u u e da a, mi o ing
eal-wo ld o ecas ing scena ios. The aining se is shown in blue,
he alida ion se in o ange, and he es ing se in g een. Dis inc
daily pa e ns and weekly seasonali y can be obse ed. Fo ea u e
s anda diza ion, we employed he z-sco e no maliza ion me hod. This
In o ma ion Fusion 126 (2026) 103692
6
Y. Shang e al.
Fig. 2. Dis ibu ion o he EV cha ging demand in (a) di e en dis ic s and (b) di e en subdis ic .
Table 1
Maximum and o al EV cha ging demand by Beijing dis ic s.
Index 1 2 3 4 5 6 7 8
Name Dongcheng Xicheng Chaoyang Haidian Shijingshan Feng ai Men ougou Fangshan
Mean (kWh) 245.16 434.66 2123.46 2021.02 490.84 1329.92 260.63 693.42
S d 219.65 379.34 1766.87 1695.40 418.29 1129.79 230.64 597.07
Max (kWh) 1776.53 2875.16 13993.56 14218.98 3367.72 10184.80 1991.60 4878.33
Sum (kWh) 364546.60 646 338.50 3157578.00 3005260.00 729 876.70 1977598.00 387558.10 1031116.00
Index 9 10 11 12 13 14 15 16
Name Changping Daxing Shunyi Tongzhou Miyun Pinggu Huai ou Yanqing
Mean (kWh) 1094.24 1451.24 748.59 1232.15 143.99 123.60 212.20 103.14
S d 918.04 1211.85 641.65 1030.00 137.96 119.03 191.22 105.14
Max (kWh) 8057.20 9853.18 6074.08 8354.02 1539.35 966.94 1786.56 878.67
Sum (kWh) 1627137.00 2158001.00 1113159.00 1832207.00 214116.00 183791.30 315541.20 153372.30
app oach ans o ms each ea u e o ha e ze o mean and uni a iance,
calcula ed as:
𝑧=𝑥−𝜇
𝜎(18)
whe e 𝑥 is he o iginal ea u e alue, 𝜇 is he mean, and 𝜎 is he
s anda d de ia ion o he ea u e ac oss he aining da ase . Fig. 3(b)
p esen s iolin plo s o no malized cha ging demand o each o he 16
dis ic s. Each iolin plo shows he dis ibu ion, median, and in e qua -
ile ange o he no malized demand, highligh ing he a iabili y and
cen al endency ac oss di e en zones.
5.1.2. E alua ion me ics
To e alua e he accu acy o pa king demand p edic ion, we em-
ploy ou widely used me ics: mean absolu e e o (MAE), oo mean
squa ed e o (RMSE), and mean absolu e pe cen age e o (MAPE).
These me ics a e de ined as ollows:
MAE =1
𝑁×𝑇′
𝑇′
∑
𝑡=1
𝑁
∑
𝑖=1 |||
𝑌𝑡,𝑖 −𝑌𝑡,𝑖|||(19)
RMSE =√
√
√
√1
𝑁×𝑇′
𝑇′
∑
𝑡=1
𝑁
∑
𝑖=1 (
𝑌𝑡,𝑖 −𝑌𝑡,𝑖)2(20)
MAPE =100%
𝑁×𝑇′
𝑇′
∑
𝑡=1
𝑁
∑
𝑖=1 |||||
𝑌𝑡,𝑖 −𝑌𝑡,𝑖
𝑌𝑡,𝑖 |||||
(21)
whe e 𝑇′ deno es he numbe o ime s eps in he e alua ion pe iod
( alida ion o es ing se ), 𝑁 is he numbe o pa king loca ions,
𝑌𝑡,𝑖 is
he p edic ed demand, and 𝑌𝑡,𝑖 is he g ound u h demand a ime 𝑡 and
loca ion 𝑖.
5.1.3. Benchma k models
To assess he obus ness and p edic i e powe o he p oposed
model, we compa e i wi h a wide ange o classical and deep lea ning-
based o ecas ing me hods:
•RNN [41]: A neu al ne wo k a chi ec u e designed o sequen ial
da a modeling, capable o cap u ing empo al dynamics h ough
ecu en connec ions, sui able o asks such as pa king demand
p edic ion.
•LSTM [42]: An enhanced RNN a chi ec u e ha in oduces mem-
o y cells and ga ing mechanisms o e ec i ely p ese e long- e m
dependencies, widely used in complex ime se ies modeling asks.
•Ga ed Recu en Uni (GRU) [43]: A ligh weigh a ian o LSTM
ha e ains ga ing mechanisms wi h ewe pa ame e s, o e ing
e icien aining and e ec i e modeling o sho - o medium- e m
dependencies.
•G aph Con olu ional Ne wo k (GCN) [44]: Cap u es spa ial de-
pendencies among pa king loca ions using a g aph s uc u e and
ex ac s spa ial ea u es ia g aph con olu ion ope a ions.
•G aph A en ion Ne wo k (GAT) [45]: Ex ends GCN by inco -
po a ing a en ion mechanisms ha assign di e en weigh s o
neighbo ing nodes, allowing mo e lexible modeling o spa ial
dependencies.
•G aphSAGE [46]: A scalable g aph neu al ne wo k app oach ha
employs neighbo hood sampling and agg ega ion s a egies, en-
abling e icien ep esen a ion lea ning on la ge-scale g aphs.
•Tempo al G aph Con olu ional Ne wo k (TGCN) [47]: Combines
GCN and GRU o join ly model spa ial and empo al dependen-
cies, sui able o spa io- empo al sequence p edic ion asks.
In o ma ion Fusion 126 (2026) 103692
7
Y. Shang e al.
Fig. 3. Dis ibu ion o he EV cha ging demand in one mon h by Beijing dis ic s, (a) empo al dis ibu ion o cha ging demand, (b) iolin plo o no malized
cha ging demand by zone.
•Tempo al G aph A en ion Ne wo k (TGAT) [48]: Enhances TGCN
by in oducing ime encoding and a en ion mechanisms o cap-
u e he empo al dynamics a ec ing node ep esen a ions mo e
p ecisely.
•Tempo al G aphSAGE (TSAGE) [8]: An ex ension o G aphSAGE
o empo al g aphs, using ime-awa e sampling and agg ega ion
o cap u e dynamic ea u es in e ol ing g aph s uc u es.
5.1.4. Model se ings
The main hype pa ame e s o he p oposed model a e shown in Ta-
ble 2. Besides, o ensu e a comp ehensi e and unbiased compa ison, we
s uc u ed he expe imen as ollows: (1) The model’s ou pu comp ises
he cha ging demand o all dis ic s o subdis ic s o he subsequen
1 imeslo and 2 imeslo s. (2) Fo ai compa ison, all baseline models
use he same hype pa ame e s as he p oposed model. Addi ionally,
ou p edic ion scena ios (S) a e p oposed, as ollows:
•S1: P edic ing he cha ging demand o he nex ime in e al
ac oss 16 dis ic s;
•S2: P edic ing he cha ging demand o he nex wo ime in e -
als ac oss 16 dis ic s;
•S3: P edic ing he cha ging demand o he nex ime in e al
ac oss 331 subdis ic s;
•S4: P edic ing he cha ging demand o he nex wo ime in e -
als ac oss 331 subdis ic s.
(3) The p oposed model is implemen ed using PyTo ch on a wo ks a-
ion equipped wi h a GeFo ce RTX 3090 Ti GPU. The model u ilizes
a mean squa ed e o loss unc ion. The Range op imize , which
in eg a es he RAdam and LookAhead s a egies, is known o i s abili y
o e ain he e icien con e gence p ope ies o Adam while enhancing
model gene aliza ion h ough he LookAhead mechanism. In his s udy,
he lea ning a e is se o 0.001, wi h a weigh decay coe icien o
0.0001. The aining p ocess is conduc ed o e 100 epochs. The li-
b a ies u ilized in code and hei exac e sions used in he expe imen s
a e speci ied in Table 2.
5.2. Compa ison esul s
The compa ison o pe o mance o di e en models on a ious me -
ics a dis ic and subdis ic scales a e shown in Table 3. In he ask
o mul i-scale EV cha ging demand o ecas ing, he EV-STLLM model
demons a es ou s anding pe o mance, pa icula ly in dis ic -le el
In o ma ion Fusion 126 (2026) 103692
8
Y. Shang e al.
Table 2
De ailed Py hon lib a ies and hei exac e sions.
Lib a y Ve sion
numpy 1.26.4
pandas 2.2.3
Py hon 3.12.3
o ch 2.3.0+cu121
o ch ision 0.18.0+cu121
ans o me s 4.49.0
ma plo lib 3.9.0
and subdis ic -le el p edic ions ac oss di e en o ecas ing ho izons.
Fo he dis ic -le el one-s ep p edic ion (S1), EV-STLLM achie es a
mean absolu e e o (MAE) o 218.13, ep esen ing a educ ion o
app oxima ely 15.4% compa ed o he nex -bes model, GRU, which
eco ds an MAE o 257.86. In e ms o oo mean squa e e o (RMSE),
EV-STLLM sco es 680.54, signi ican ly ou pe o ming all o he models.
The mean absolu e pe cen age e o (MAPE) is only 1.39%, a educ ion
o 53.5% compa ed o G aphSAGE’s 2.99%. These esul s indica e ha
EV-STLLM is mo e capable o cap u ing complex spa ial dependencies
and sho - e m empo al dynamics ac oss egions. Fu he mo e, in he
mo e challenging dis ic -le el wo-s ep p edic ion (S2), EV-STLLM con-
inues o lead wi h an MAE o 242.35, which is app oxima ely 34.0%
lowe han GRU’s 366.93. The RMSE eaches 693.39, again ou pe o m-
ing all g aph- and sequence-based models. The MAPE u he d ops o
1.29%, mo e han hal ing ha o G aphSAGE (2.94%). This con olled
inc ease in e o ac oss ime s eps highligh s EV-STLLM’s supe io
abili y in long- e m empo al dependency modeling and spa io empo-
al end lea ning, su passing adi ional RNN/LSTM and g aph-based
me hods in mul i-s ep o ecas ing asks.
A a ine g anula i y, EV-STLLM also exhibi s s ong gene aliza ion
capabili y in subdis ic -le el p edic ions. Fo one-s ep o ecas ing a
he subdis ic scale (S3), EV-STLLM achie es an MAE o 33.80, which
is 19.6% lowe han ha o RNN (42.04), and an RMSE o 53.00,
26.4% lowe han RNN’s 71.99. Al hough MAPE sligh ly exceeds ha
o GRU (1.82% s. 1.21%), his can be a ibu ed o he smalle
magni ude o subdis ic -le el demand, making MAPE mo e sensi i e
o small alues. Consequen ly, MAE and RMSE emain he mo e eli-
able indica o s o ope a ional decision-making in his con ex . In he
wo-s ep subdis ic -le el p edic ion ask (S4), EV-STLLM main ains i s
ad an age, wi h MAE and RMSE o 36.91 and 66.13, espec i ely—
ep esen ing educ ions o app oxima ely 16.7% and 14.0% when com-
pa ed o GRU. While he MAPE is sligh ly highe (1.78% s. 1.23%),
i s p ac ical impac can be mi iga ed h ough weigh ed business me ics
ha emphasize absolu e e o con ol and s abili y in scheduling.
F om a compa a i e pe spec i e, adi ional sequence models such
as RNN, LSTM, and GRU exhibi ce ain s eng hs in empo al de-
pendency modeling. Howe e , hey o en su e om in o ma ion loss
and s uggle o simul aneously ex ac bo h local and global ea u es
in complex u ban spa io empo al in e ac ions. G aph-based models
like GCN, GAT, and G aphSAGE, while e ec i e in cap u ing spa ial
adjacency, end o lack he capaci y o model empo al dynamics,
ocusing p ima ily on s a ic opological ela ionships. Al hough spa-
io empo al g aph models such as TGCN, TGAT, and TSAGE a emp
o in eg a e bo h spa ial and empo al in o ma ion, hei eliance on
localized con olu ional o ga ed mechanisms cons ains hei abili y o
comp ehensi ely use global dependencies. This can be a ibu ed o
se e al inhe en limi a ions:
•S a ic G aph S uc u es: These models ely on ixed, p ede ined
spa ial g aphs, which ail o cap u e dynamic in e - egion ela-
ionships a ising om empo al shi s in EV cha ging demand.
Fo ins ance, egions ha a e spa ially disconnec ed migh exhibi
s ong empo al co ela ions ha s a ic adjacency ma ices canno
ep esen .
•Local Dependency Bias: T adi ional GNNs agg ega e in o ma ion
om local neighbo hoods, lacking he capaci y o model global
pa e ns. While ex ensions like TGAT a emp o inco po a e em-
po al dynamics, hei pe o mance is cons ained by he localized
na u e o hei g aph con olu ion mechanisms.
•Dependency on G aph Cons uc ion: The e ec i eness o hese
models hea ily depends on he quali y o he cons uc ed g aphs.
In eal-wo ld u ban scena ios, accu a ely modeling human mo-
bili y o ehicle low h ough adjacency ma ices is challenging
and o en leads o subop imal g aph s uc u es. Such inaccu acies
p opaga e h ough he model, deg ading i s pe o mance.
In con as , ou p oposed EV-STLLM amewo k elimina es he depen-
dency on explici g aph s uc u es by adop ing an a en ion mech-
anism capable o cap u ing long- ange spa io empo al dependencies
di ec ly om aw da a. This no only add esses he limi a ions o
g aph-based models bu also allows EV-STLLM o gene alize e ec i ely
ac oss egions wi h ambiguous o dynamically shi ing ela ionships.
The inco po a ion o embedding echniques u he enhances i s abil-
i y o in eg a e he e ogeneous ea u es, o e ing a mo e comp ehen-
si e unde s anding o u ban dynamics. Mo eo e , he inco po a ion o
LoRA enables e icien pa ame e uning, acili a ing lexible adap a ion
ac oss di e se applica ion scena ios. O e all, EV-STLLM achie es lowe
e o s and highe obus ness ac oss mul iple o ecas ing le els and ime
ho izons, showcasing no only echnical sophis ica ion in model a chi-
ec u e and lea ning mechanisms bu also p ac ical alue in suppo ing
la ge-scale u ban demand o ecas ing.
Fig. 4 p esen s a compa ison be ween he EV-STLLM p edic ed
cha ging demand cu es (o ange) and he ac ual obse ed cu es (blue)
ac oss 16 di e en dis ic s. I can be obse ed ha he o e all end
i ing is sa is ac o y, wi h he model accu a ely cap u ing he pe iodic
a ia ions in daily cha ging demand, such as he ise in he mo ning
and he decline a nigh . The model also exhibi s s ong adap abili y
o pa e n shi s be ween weekdays and weekends. Du ing majo peak
pe iods (e.g., mo ning and e ening peaks) and ough pe iods, he
p edic ed cu es closely align wi h he g ound u h, demons a ing he
model’s excellen capabili y in ex ac ing empo al ea u es. Al hough
ce ain de ia ions occu du ing ex eme su ges (e.g., a ound holidays),
he magni ude emains well-con olled, and he luc ua ion ends a e
consis en , indica ing s ong obus ness in handling abno mal demand
a ia ions. In addi ion, he model is capable o p omp ly esponding
o sudden load changes, accu a ely e lec ing u ning poin s in he
demand cu es, which highligh s i s sensi i i y and apid adap abil-
i y o load luc ua ions. In summa y, a he dis ic scale, EV-STLLM
e ec i ely cap u es he mac oscopic empo al ends o elec ic e-
hicle cha ging demand ac oss wide u ban spaces, p o iding eliable
suppo o ci y-le el ene gy scheduling and cha ging in as uc u e
op imiza ion.
Fig. 5 illus a es he p edic ion esul s o 16 consecu i e subdis-
ic s om zone 279 o zone 294. Compa ed o he dis ic scale, he
subdis ic da a exhibi g ea e andomness and spa si y, wi h mo e
equen small-scale luc ua ions. Due o he smalle base alues wi hin
subdis ic s, mino a ia ions a e ampli ied, esul ing in sha pe and
mo e i egula cu es. Ne e heless, EV-STLLM main ains good end
i ing ac oss mos subdis ic s, demons a ing s ong s abili y. Fo sub-
dis ic s wi h ex ensi e pe iods o ze o o e y low demand, he model
e ec i ely a oids o e i ing o ze o alues and p ese es easonable
nonlinea i ing, e lec ing sound egula iza ion capabili ies. The p e-
dic ed cu es gene ally synch onize wi h he ac ual changes a u ning
poin s o sudden demand su ges o d ops, al hough sligh smoo hing is
obse ed in ex emely spa se egions. When acing occasional demand
spikes (e.g., caused by local e en s), he model pa ially cap u es he
su ges bu ends o sligh ly unde i , sugges ing ha u u e imp o e-
men s could inco po a e anomaly de ec ion mechanisms. O e all, he
subdis ic -scale e alua ion e i ies EV-STLLM’s gene aliza ion abili y
In o ma ion Fusion 126 (2026) 103692
9
Y. Shang e al.
Fig. 9. Sensi i i y analysis wi h MAE (1) and MAPE (2) ac oss di e en ime sequences leng hs, (a) S1, (b) S2, (c) S3, (d) S4.
Table 7
Compa ison o p edic ion pe o mance o di e en pa ame e e icien ine- uning me hods.
Zone Dis ic Subdis ic
Ou pu 1 2 1 2
Me ic MAE RMSE MAPE MAE RMSE MAPE MAE RMSE MAPE MAE RMSE MAPE
IA3 234.17 690.95 1.53 256.22 695.52 1.73 37.56 68.50 1.97 38.44 70.65 1.84
P-Tuning- 2 585.26 998.78 8.08 580.53 993.93 7.55 39.67 72.70 1.97 38.68 70.04 1.89
P-Tuning 579.21 997.06 7.82 572.94 993.51 7.00 38.78 70.57 2.12 39.70 72.79 1.80
P e ix 574.49 996.52 7.53 569.65 993.61 6.85 40.16 73.13 2.16 40.07 72.01 2.06
Bi Fi 242.02 688.16 1.74 255.79 705.31 1.41 38.10 69.72 1.97 38.13 70.28 1.79
LoRA 218.13 680.54 1.39 242.35 693.39 1.29 33.80 53.00 1.82 36.91 66.13 1.78
IA3, which in oduces a di e en pa ame e injec ion mechanism,
pe o ms compa ably o LoRA a he Dis ic le el, bu i s RMSE and
MAPE a he Subdis ic le el a e sligh ly wo se, indica ing limi ed
gene aliza ion when modeling a ine spa ial scales. Bi Fi , despi e
modi ying only a small subse o bias pa ame e s, pe o ms mod-
e a ely o e en second-bes in many se ings. I shows be e MAE
pe o mance han he P-Tuning amily, highligh ing i s ela i ely high
In o ma ion Fusion 126 (2026) 103692
16
Y. Shang e al.
Table 8
Compa ison o aining and in e ence ime (in seconds) o di e en pa ame e -e icien ine- uning me hods.
Zone Dis ic Subdis ic
Ou pu 1 2 1 2
Type T aining In e ence T aining In e ence T aining In e ence T aining In e ence
IA3 0.4605 0.0547 0.4532 0.0528 8.5828 0.8325 8.5588 0.8330
P-Tuning- 2 0.7531 0.0865 0.7498 0.0860 8.3898 0.8287 8.3490 0.8274
P-Tuning 0.6846 0.0822 0.6853 0.0824 8.3174 0.8257 8.2929 0.8241
P e ix 0.6769 0.0815 0.6744 0.0816 8.3057 0.8233 8.2796 0.8228
Bi Fi 0.4332 0.0491 0.4299 0.0489 8.1539 0.7955 8.1389 0.7962
LoRA 0.5042 0.0581 0.5062 0.0582 8.9191 0.8816 8.2929 0.8241
cos -e ec i eness in cons ained pa ame e -upda e scena ios. The pe -
o mance o P-Tuning and i s a ian P-Tuning- 2 is ela i ely poo
in his expe imen al se up, especially a he Dis ic le el, whe e hei
MAE and RMSE sco es a e signi ican ly highe han hose o o he
me hods—some imes app oaching he pe o mance o an unadap ed
base model. This may be due o he hea y eliance o hese me hods on
ex ensi e p omp oken uning, which is sensi i e o ask s uc u e and
less e ec i e. P e ix Tuning, while enhancing p omp exp essi eness
ela i e o P-Tuning, s ill alls sho o ma ching he pe o mance o
LoRA o IA3. No ably, a he Subdis ic le el in 1 ou pu leng h, i
yields he highes MAPE o 2.16 among all me hods, indica ing po en-
ial limi a ions in modeling complex hie a chical spa ial dependencies.
Fu he analysis o p edic ion e o s ac oss di e en le els e eals ha
all me hods exhibi signi ican ly lowe MAE and RMSE a he Subdis-
ic le el compa ed o he Dis ic le el. This may pa ially e lec
he ac ha ine-g ained spa ial p edic ion asks a e associa ed wi h
smoo he a ge unc ions o a e easie o i . Howe e , he MAPE a he
Subdis ic le el shows g ea e a iabili y, sugges ing ha no malized
e o me ics a e mo e sensi i e o p edic ion a ge s wi h low magni-
ude. This highligh s he need o ca e ully choose e alua ion me ics
based on speci ic business equi emen s in eal-wo ld applica ions.
In addi ion o p edic ion accu acy, compu a ional e iciency is a
c ucial ac o when selec ing PEFT s a egies, pa icula ly o deploy-
men in esou ce-cons ained en i onmen s. Table 8 p esen s a de-
ailed compa ison o aining and in e ence ime. We obse e ha
Bi Fi consis en ly exhibi s he as es aining and in e ence imes
ac oss all scena ios, owing o i s minimal pa ame e upda e design—
only uning bias pa ame e s. IA3 also demons a es low compu a ional
o e head, pa icula ly in he Dis ic -le el asks, wi h aining imes
unde 0.5 s pe epoch and in e ence imes a ound 0.05 s. LoRA, while
no he as es , s ikes a compelling balance be ween e iciency and
pe o mance. A he Dis ic le el, i s aining and in e ence imes
(app oxima ely 0.5 and 0.058 s espec i ely) a e only ma ginally highe
han hose o IA3 and Bi Fi , bu i a su passes all o he me hods in
p edic ion accu acy (see Table 7). Fo ins ance, LoRA achie es he low-
es MAE and RMSE ac oss all egions and ou pu leng hs, and deli e s
pa icula ly s ong esul s a he Subdis ic le el—demons a ing i s
abili y o model ine-g ained spa ial he e ogenei y.
A he Subdis ic le el, LoRA’s aining ime ( 8.9 s pe epoch o 1
ou pu leng h) is sligh ly highe han o he me hods, bu his o e head
is jus i iable gi en i s subs an ial gains in p edic i e pe o mance.
In e ence imes emain compa able wi h o he me hods (e.g., 0.8816 s
s. 0.8325 o IA3), ensu ing ha LoRA emains p ac ical o eal-
ime o nea - eal- ime applica ions. On he o he hand, he P-Tuning
amily and P e ix Tuning me hods, despi e hei exp essi eness h ough
p omp -based pa ame e iza ion, incu longe aining imes (a ound
0.68–0.75 s a he Dis ic le el and o e 8.3 s a he Subdis ic le el)
and ail o o e compe i i e p edic ion accu acy. This sugges s ha
hei compu a ional cos is no well-compensa ed by co esponding
gains in model pe o mance, making hem less a o able in his con ex .
In summa y, combining bo h p edic i e accu acy and compu a ional
e iciency, LoRA eme ges as he mos balanced and e ec i e PEFT
me hod o he s udied spa ial– empo al p edic ion asks.
6. Conclusions
In his pape , we p opose a no el spa io empo al lea ning ame-
wo k, EV-STLLM, o sho - e m EV cha ging demand p edic ion in
u ban en i onmen s o ackle he co e challenges o da a usion and
model usion. A he da a le el, ou model cons uc s a collabo a i e
embedding mechanism ha uses oken-le el, empo al, and spa ial
ea u es, enabling ine-g ained modeling o he nonlinea and dynamic
pa e ns inhe en in EV cha ging beha io . A he model le el, we
in eg a e a p e ained LLM as he backbone o deep spa io empo al de-
pendency modeling while signi ican ly educing aining cos h ough
LoRA, which eezes he bulk o model pa ame e s and only unes a
small se o low- ank ma ices.
Ex ensi e expe imen s conduc ed on a la ge-scale eal-wo ld da ase
om Beijing—co e ing 16 dis ic s and 331 subdis ic s, wi h o e
830,000 cha ging eco ds—demons a e he supe io pe o mance o
EV-STLLM ac oss mul iple e alua ion me ics and p edic ion scena ios.
Compa ed o classical sequence models (RNN, LSTM, GRU), g aph-
based models (GCN, GAT, G aphSAGE), and spa io empo al g aph mod-
els (TGCN, TGAT, TSAGE), EV-STLLM achie es consis en imp o e-
men s ac oss all asks and scales: In dis ic -le el one-s ep p edic ion,
EV-STLLM educes MAE by 15.41% and MAPE by 53.51% compa ed
o he bes -pe o ming baseline. In subdis ic -le el p edic ion, despi e
he g ea e spa ial g anula i y, EV-STLLM main ains a signi ican lead
in bo h MAE and RMSE, showcasing i s s ong gene aliza ion capac-
i y and obus ness a ine spa ial esolu ions. To be e unde s and
he impac o inpu sequence leng h on p edic ion pe o mance, we
conduc a empo al sensi i i y analysis by a ying he leng h o his-
o ical ime windows used in he model inpu . The esul s e eal he
impo ance o selec ing an app op ia e empo al window ha balances
con ex dep h and ele ance. They also sugges po en ial o adap i e
sequence lea ning, whe e he model dynamically adjus s i s ecep i e
ield based on o ecas ho izon o local empo al pa e ns. Addi ionally,
ou abla ion s udy con i ms he e ec i eness o each key componen .
Remo ing he LLM componen (NonLLM) leads o subs an ial pe o -
mance deg ada ion, especially in MAPE, indica ing he c i ical ole o
LLMs in cap u ing global dependencies. Compa ed wi h ull ine- uning
(FFT) and ully ozen (FF) se ups, ou LoRA-based app oach achie es
compa able o e en be e p edic ion accu acy while signi ican ly e-
ducing compu a ional cos , alida ing i s p ac ical alue o scalable
deploymen .
6.1. P ac ical implica ions and limi a ions
The p oposed EV-STLLM amewo k holds signi ican p omise o
eal-wo ld deploymen in sma g id managemen and u ban ene gy
sys ems. Howe e , i s p ac ical applica ion also en ails se e al consid-
e a ions and limi a ions ha mus be ca e ully add essed.
On he one hand, o p ac ical implica ions, Fi s ly, EV-STLLM en-
ables accu a e sho - e m o ecas s o cha ging demand a bo h dis ic
and subdis ic le els. This allows g id ope a o s o p oac i ely alloca e
elec ici y esou ces, mi iga e peak loads, and implemen dynamic load
balancing s a egies. The ine-g ained spa ial esolu ion also suppo s
zonal demand- esponse mechanisms. Secondly, by p edic ing empo al
In o ma ion Fusion 126 (2026) 103692
17
Y. Shang e al.
a ia ions in cha ging ac i i y, EV-STLLM can in o m adap i e ime-
o -use p icing s a egies. This enables u ili y p o ide s o incen i ize
o -peak cha ging, educe g id s ess, and align use beha io wi h
sys em-le el op imiza ion objec i es. And hen, EV-STLLM acili a es
ho spo de ec ion and demand clus e ing a subdis ic scale, suppo -
ing op imal si ing o new cha ging s a ions o mobile cha ging uni s.
Long- e m deploymen planning can bene i om sho - e m demand
dynamics, especially in apidly e ol ing u ban en i onmen s. Finally,
gi en i s e icien a chi ec u e wi h LoRA-based ine- uning, EV-STLLM
can be in eg a ed in o eal- ime decision suppo sys ems, such as
cha ging s a ion managemen pla o ms o u ban ene gy digi al wins,
o e ing imely and localized p edic ions.
On he o he hand, o limi a ions and u u e wo ks, Fi s ly, while
EV-STLLM pe o ms well on Beijing da a, i s gene aliza ion o o he
ci ies wi h di e en u ban opologies, cha ging beha io s, o in as uc-
u e densi ies may be limi ed. Domain adap a ion o ede a ed lea ning
app oaches may be equi ed o c oss-ci y deploymen . Secondly, he
e ec i eness o EV-STLLM depends hea ily on he a ailabili y and
accu acy o ine-g ained cha ging logs, spa ial me ada a, and con ex ual
ea u es. In da a-sca ce egions, pe o mance may deg ade. Syn he ic
da a gene a ion o ans e lea ning could be explo ed o mi iga e his
issue. Thi dly, al hough LLM-based models o e s ong ep esen a ion
capabili ies, hei decision-making p ocess emains ela i ely opaque.
Fo high-s akes applica ions (e.g., g id eliabili y o public sa e y), aug-
men ing he model wi h explainable AI modules is c ucial o enhance
anspa ency and s akeholde us . Las bu no leas , u ban cha ging
pa e ns a e dynamic, a ec ed by policy changes, in as uc u e up-
g ades, and beha io al shi s. The model equi es pe iodic e aining
o online lea ning capabili ies o emain accu a e o e ime. Au o-
ma ed model upda ing pipelines should be conside ed in la ge-scale
deploymen s.
CRediT au ho ship con ibu ion s a emen
Yi ong Shang: W i ing – o iginal d a , Visualiza ion, Me hodol-
ogy, Concep ualiza ion. Wen-Long Shang: W i ing – e iew & edi ing,
W i ing – o iginal d a , P ojec adminis a ion, Me hodology, Funding
acquisi ion, Concep ualiza ion. Dingsong Cui: In es iga ion, Concep-
ualiza ion. Peng Liu: Da a cu a ion, Concep ualiza ion. Haibo Chen:
Valida ion, P ojec adminis a ion, In es iga ion. Dongdong Zhang:
Visualiza ion, Valida ion, So wa e. Runsen Zhang: W i ing – e iew &
edi ing. Chengcheng Xu: W i ing – e iew & edi ing, In es iga ion. Ye
Liu: W i ing – e iew & edi ing, Me hodology. Chenxi Wang: W i ing
– e iew & edi ing. Mohannad Alhazmi: Valida ion.
Decla a ion o compe ing in e es
The au ho s decla e ha hey ha e no known compe ing inan-
cial in e es s o pe sonal ela ionships ha could ha e appea ed o
in luence he wo k epo ed in his pape .
Acknowledgmen s
This wo k was pa ially suppo ed by he ZEV-UP and ePowe Mo e
p ojec s co- unded by he Eu opean Union unde G an ag eemen ID:
101138721 and 101192753. Besides, his esea ch is suppo ed in pa
by he Beijing Na u al Science Founda ion, China (No. 9232003). Also,
his wo k would like o acknowledge he suppo p o ided by Re-
sea che s Suppo ing P ojec (P ojec numbe : RSPD2025R635),
KingSaud Uni e si y, Riyadh, Saudi A abia.
Da a a ailabili y
The au ho s do no ha e pe mission o sha e da a.
Re e ences
[1] Ma ino T an, Da id Banis e , Jus in D.K. Bishop, Malcolm D. McCulloch,
Realizing he elec ic- ehicle e olu ion, Na . Clim. Chang. 2 (5) (2012)
328–333.
[2] Aqib Zahoo , Yajuan Yu, Saima Ba ool, Muhammad Id ees, Guozhu Mao, The
ca bon-neu al goal in China o he elec ic ehicle indus y wi h solid-s a e
ba e y’s con ibu ion in 2035 o 2045, J. En i on. Eng. 149 (12) (2023)
04023082.
[3] Zhaoyao Bao, Jiapei Li, Xuehan Bai, Chi Xie, Zhibin Chen, Min Xu, Wen-Long
Shang, Hailong Li, An op imal cha ging scheduling model and algo i hm o
elec ic buses, Appl. Ene gy 332 (2023) 120512.
[4] Hang Yu, Songyan Niu, Yi ong Shang, Ziyun Shao, Youwei Jia, Linni Jian,
Elec ic ehicles in eg a ion and ehicle- o-g id ope a ion in ac i e dis ibu ion
g ids: A comp ehensi e e iew on powe a chi ec u es, g id connec ion s anda ds
and ypical applica ions, Renew. Sus ain. Ene gy Re . 168 (2022) 112812.
[5] Yanchong Zheng, Songyan Niu, Yi ong Shang, Ziyun Shao, Linni Jian, In eg a ing
plug-in elec ic ehicles in o powe g ids: A comp ehensi e e iew on powe
in e ac ion mode, scheduling me hodology and ma hema ical ounda ion, Renew.
Sus ain. Ene gy Re . 112 (2019) 424–439.
[6] Zaoli Yang, Qin Li, Yamin Yan, Wen-Long Shang, Washing on Ochieng, Examin-
ing in luence ac o s o Chinese elec ic ehicle ma ke demand based on online
e iews unde mode a ing e ec o subsidy policy, Appl. Ene gy 326 (2022)
120019.
[7] Wen-Long Shang, Junjie Zhang, Kun Wang, Hangjun Yang, Washing on Ochieng,
Can inancial subsidy inc ease elec ic ehicle (EV) pene a ion—e idence om
a quasi-na u al expe imen , Renew. Sus ain. Ene gy Re . 190 (2024) 114021.
[8] Yi ong Shang, Duo Li, Yang Li, Sen Li, Explainable spa io empo al mul i- ask
lea ning o elec ic ehicle cha ging demand p edic ion, Appl. Ene gy 384
(2025) 125460.
[9] Alexis Peng ei Zhao, Shuangqi Li, Zhengmao Li, Zhaoyu Wang, Xue Fei, Zechun
Hu, Mohannad Alhazmi, Xiaohe Yan, Chenye Wu, Shuai Lu, e al., Elec ic
ehicle cha ging planning: A complex sys ems pe spec i e, IEEE T ans. Sma
G id (2024).
[10] Zhijie Lai, Sen Li, Towa ds a mul imodal cha ging ne wo k: Join planning o
cha ging s a ions and ba e y swapping s a ions o elec i ied ide-hailing lee s,
T ansp. Res. Pa B: Me hodol. 183 (2024) 102928.
[11] Wenhao Wang, Aihong Tang, Feng Wei, Huiyuan Yang, Li Xin an, Jiao Peng,
Elec ic ehicle cha ging load o ecas ing conside ing wea he impac , Appl.
Ene gy 383 (2025) 125337.
[12] Chahinez Ounoughi, Sadok Ben Yahia, Da a usion o ITS: A sys ema ic li e a u e
e iew, In . Fusion 89 (2023) 267–291.
[13] Lau a Melga -Ga cía, Da id Gu ié ez-A ilés, C is ina Rubio-Escude o, Alicia
T oncoso, A no el dis ibu ed o ecas ing me hod based on in o ma ion usion
and inc emen al lea ning o s eaming ime se ies, In . Fusion 95 (2023)
163–173.
[14] Shams Fo uque Ahmed, Swee y Angela Kuldeep, Sabiha Janna Ra a, Ja e ia
Fazal, Mah a a Hoque, Gang Liu, Ami H. Gandomi, Enhancemen o T a ic Fo e-
cas ing Th ough G aph Neu al Ne wo k-Based In o ma ion Fusion Techniques,
Else ie , 2024.
[15] Niyaz Ahmad Wani, Ra inde Kuma , Ja in Bedi, Imad Rida, e al., Explainable
AI-d i en IoMT usion: Un a elling echniques, oppo uni ies, and challenges
wi h explainable AI in heal hca e, In . Fusion (2024) 102472.
[16] Fa ima Hassan, Syed Fawad Hussain, Saeed Mian Qaisa , Fusion o mul i a ia e
EEG signals o schizoph enia de ec ion using CNN and machine lea ning
echniques, In . Fusion 92 (2023) 466–478.
[17] Hui Huang, Bing Xu, Xinnian Liang, Kehai Chen, Muyun Yang, Tiejun Zhao,
Conghui Zhu, Mul i- iew usion o ins uc ion mining o la ge language model,
In . Fusion 110 (2024) 102480.
[18] Lingshu Zhong, Ziling Zeng, Zikang Huang, Xiaowei Shi, Yiming Bie, Join
op imiza ion o elec ic bus cha ging and ene gy s o age sys em scheduling,
F on ie s o Enginee ing Managemen 11 (4) (2024) 676–696.
[19] M. Hadi Amini, Amin Ka ga ian, O kun Ka abasoglu, ARIMA-based decoupled
ime se ies o ecas ing o elec ic ehicle cha ging demand o s ochas ic powe
sys em ope a ion, Elec . Powe Sys . Res. 140 (2016) 378–390.
[20] Fei Ren, Chenlu Tian, Guiqing Zhang, Chengdong Li, Yuan Zhai, A hyb id me hod
o powe demand p edic ion o elec ic ehicles based on SARIMA and deep
lea ning wi h in eg a ion o pe iodic ea u es, Ene gy 250 (2022) 123738.
[21] Eiman ElGhanam, Mohamed Hassan, Ahmed Osman, Machine lea ning-based
elec ic ehicle cha ging demand p edic ion using o igin-des ina ion da a: A UAE
case s udy, in: 2022 5 h In e na ional Con e ence on Communica ions, Signal
P ocessing, and hei Applica ions, ICCSPA, IEEE, 2022, pp. 1–6.
[22] Shengyou Wang, Chengxiang Zhuge, Chun u Shao, Pinxi Wang, Xiong Yang, Shiqi
Wang, Sho - e m elec ic ehicle cha ging demand p edic ion: A deep lea ning
app oach, Appl. Ene gy 340 (2023) 121032.
[23] Zhiyan Yi, Xiaoyue Ca hy Liu, Ran Wei, Xi Chen, Jiangpeng Dai, Elec ic ehicle
cha ging demand o ecas ing using deep lea ning model, J. In ell. T ansp. Sys .
26 (6) (2022) 690–703.
In o ma ion Fusion 126 (2026) 103692
18
Y. Shang e al.
[24] Lijing Zhu, Wen-Long Shang, Jingzhou Wang, Yixin Li, Chulung Lee, Washing on
Ochieng, Xunzhang Pan, Di usion o elec ic ehicles in Beijing conside ing
indi ec ne wo k e ec s, T ansp. Res. Pa D: T ansp. En i on. 127 (2024)
104069.
[25] Nikolaos Tsalikidis, Pa aske as Kouka as, Dimos henis Ioannidis, Dimi ios Tzo-
a as, Hyb id CNN-LSTM o ecas ing model o elec ic ehicle cha ging demand
in sma buildings, in: 2024 6 h Global Powe , Ene gy and Communica ion
Con e ence, GPECOM, IEEE, 2024, pp. 590–595.
[26] Zhiju Chen, Kai Liu, Jiangbo Wang, Toshiyuki Yamamo o, H-Con LSTM-based
bagging lea ning app oach o ide-hailing demand p edic ion conside ing im-
balance p oblems and spa se unce ain y, T ansp. Res. Pa C: Eme g. Technol.
140 (2022) 103709.
[27] Shengyou Wang, An hony Chen, Pinxi Wang, Chengxiang Zhuge, P edic ing
elec ic ehicle cha ging demand using a he e ogeneous spa io- empo al g aph
con olu ional ne wo k, T ansp. Res. Pa C: Eme g. Technol. 153 (2023) 104205.
[28] Shu Wang, Yang Yang, Yisong Chen, Xuan Zhao, T ip p icing scheme o elec ic
ehicle sha ing ne wo k wi h demand p edic ion, IEEE T ans. In ell. T ansp. Sys .
23 (11) (2022) 20243–20254.
[29] Xue Xing, Bing Wang, Xin Ning, Gang Wang, P ayag Tiwa i, Sho - e m OD low
p edic ion o u ban ail ansi con ol: A mul i-g aph spa io empo al usion
app oach, In . Fusion (2025) 102950.
[30] Tao Xu, Jiaming Deng, Ruolin Ma, Zixiang Zhang, Yingying Zhao, Zhilong
Zhao, Jun ao Zhang, Hie a chical spa io- empo al g aph ODE ne wo ks o a ic
o ecas ing, In . Fusion 113 (2025) 102614.
[31] Zhenghong Wang, Yi Wang, Fu ong Jia, Fan Zhang, Niki a Klimenko, Leye
Wang, Zhengbing He, Zhou Huang, Yu Liu, Spa io empo al usion ans o me
o la ge-scale a ic o ecas ing, In . Fusion 107 (2024) 102293.
[32] Junyi Gao, Rakshi h Sha ma, Cheng Qian, Lucas M. Glass, Je ey Spaede , Jus in
Rombe g, Jimeng Sun, Cao Xiao, STAN: spa io- empo al a en ion ne wo k o
pandemic p edic ion using eal-wo ld e idence, J. Am. Med. In o m. Assoc. 28
(4) (2021) 733–743.
[33] Chengqing Yu, Fei Wang, Yilun Wang, Zezhi Shao, Tao Sun, Di Yao, Yongjun
Xu, MGSF o me : A mul i-g anula i y spa io empo al usion ans o me o ai
quali y p edic ion, In . Fusion 113 (2025) 102607.
[34] Jinlei Zhang, Shuai Mao, Shuxin Zhang, Jia eng Yin, Lixing Yang, Ziyou Gao,
EF- o me o sho - e m passenge low p edic ion du ing la ge-scale e en s in
u ban ail ansi sys ems, In . Fusion 117 (2025) 102916.
[35] Bo Zhang, Hui Ma, Jian Ding, Jian Wang, Bo Xu, Hong ei Lin, Dis illing implici
mul imodal knowledge in o la ge language models o ze o- esou ce dialogue
gene a ion, In . Fusion (2025) 102985.
[36] Hao Xue, Flo a D. Salim, P omp cas : A new p omp -based lea ning pa adigm o
ime se ies o ecas ing, IEEE T ans. Knowl. Da a Eng. 36 (11) (2023) 6851–6864.
[37] Hai eng Zhao, Shengchao Liu, Ma Chang, Hannan Xu, Jie Fu, Zhihong Deng,
Lingpeng Kong, Qi Liu, Gimle : A uni ied g aph- ex model o ins uc ion-
based molecule ze o-sho lea ning, Ad . Neu al In . P ocess. Sys . 36 (2023)
5850–5887.
[38] Yan Kang, Mingjian Yang, Yue Peng, Jingwen Cai, Lei Zhao, Zhan Gao, Ningshu
Li, Bin Pu, LLM-DG: Le e aging la ge language model o enhanced disease
p edic ion ia in e -pa ien and in a-pa ien modeling, In . Fusion 121 (2025)
103145.
[39] Tiesunlong Shen, E ik Camb ia, Jin Wang, Yi Cai, Xuejie Zhang, Insigh a
he igh spo : P o ide decisi e subg aph in o ma ion o g aph LLM wi h
ein o cemen lea ning, In . Fusion 117 (2025) 102860.
[40] Chenxi Liu, Sun Yang, Qianxiong Xu, Zhishuai Li, Cheng Long, Ziyue Li, Rui
Zhao, Spa ial- empo al la ge language model o a ic p edic ion, in: 2024 25 h
IEEE In e na ional Con e ence on Mobile Da a Managemen , MDM, IEEE, 2024,
pp. 31–40.
[41] Wei Zhao, Hao an Zhang, Jianqin Zheng, Yuanhao Dai, Liqiao Huang, Wenlong
Shang, Yong u Liang, A poin p edic ion me hod based au oma ic machine
lea ning o day-ahead powe ou pu o mul i- egion pho o ol aic plan s, Ene gy
223 (2021) 120026.
[42] Zhile Yang, Tianyu Hu, Juncheng Zhu, Wenlong Shang, Yuanjun Guo, Aoi e
Foley, Hie a chical high- esolu ion load o ecas ing o elec ic ehicle cha ging:
A deep lea ning app oach, IEEE J. Eme g. Sel. Top. Indus ial Elec on. 4 (1)
(2022) 118–127.
[43] Haosen Yang, Xin Shi, Linyun Xiong, Ziqiang Wang, Zipeng Liang, Robus
hie a chical g ouping lea ning immune o missing da a o ol age s abili y
assessmen , IEEE T ans. Powe Sys . (2024).
[44] Wen-Long Shang, Xiaoming Tao, Huibo Bi, Yanyan Chen, Hui Zhang, Wash-
ing on Y. Ochieng, Audio ela ed quali y o expe ience e alua ion in u ban
anspo a ion en i onmen s wi h b ain inspi ed g aph lea ning, IEEE T ans.
In ell. T ansp. Sys . 24 (12) (2023) 13841–13851.
[45] Wenlong Liao, Dechang Yang, Qi Liu, Yixiong Jia, Chenxi Wang, Zhe Yang, Da a-
d i en eac i e powe op imiza ion o dis ibu ion ne wo ks ia g aph a en ion
ne wo ks, J. Mod. Powe Sys . Clean Ene gy 12 (3) (2024) 874–885.
[46] Jielun Liu, Ghim Ping Ong, Xiqun Chen, G aphSAGE-based a ic speed o e-
cas ing o segmen ne wo k wi h spa se da a, IEEE T ans. In ell. T ansp. Sys .
23 (3) (2020) 1755–1766.
[47] Lijun Sun, Mingzhi Liu, Guan eng Liu, Xiao Chen, Xu Yu, FD-TGCN: Fas and
dynamic empo al g aph con olu ion ne wo k o a ic low p edic ion, In .
Fusion 106 (2024) 102291.
[48] Chao an Dai, Qideng Tang, Huahua Ding, TGAT: Tempo al g aph a en ion ne -
wo k o blockchain phishing scams de ec ion, in: 2024 In e na ional Con e ence
on Compu e , In o ma ion and Telecommunica ion Sys ems, CITS, IEEE, 2024,
pp. 1–7.
In o ma ion Fusion 126 (2026) 103692
19
