Transcriptions
Note: this content has been automatically generated.
00:00:01
so good morning meaning is her back to him from my my student at the university of
00:00:05
these am i would search about to be morphology energy needs an artificial intelligence and in other words
00:00:13
i want to show you how we can estimate the energy consumption the other group of buildings
00:00:20
uh using artificial intelligence and how your bumper meters uh can help
00:00:26
us in the process in the process like for having more precise results
00:00:32
so imagine for example he wants to estimate the
00:00:35
energy consumption all that colour drop group of buildings uh
00:00:41
we we need to use uh some uh no one data some
00:00:46
no uh energy demanded from uh the other buildings in the city
00:00:52
so just well why do we want to do this well um the first reason is optimisation so
00:00:59
we want to have a a models that are more a more precise in order to
00:01:05
i know the for planners and designers to make a a more conscious decision in
00:01:12
the project and of course uh in this
00:01:15
way reducing d. energy waste uh and uh
00:01:21
together with the energy ways that would like to also we'd use the winners guess emission cars
00:01:27
as you might know a building sector is the
00:01:31
may or one of the main your um codes obvious
00:01:35
yeah i'll just giving the greenhouse gas emissions so the second reason is complexity uh the problem
00:01:42
is really complex 'cause there are a lot of a complex phenomena in boulder like for example
00:01:48
he thailand's or a wooden canyons are are to model a and the result so
00:01:54
i'm a high level of buttons are thinking that is given by the user behaviour
00:01:59
so basically we don't know how the user will behave inside our building we can build some model to estimate it
00:02:05
but uh it's an estimation and it uh and those
00:02:09
two factors user really to a big performance gap between the
00:02:13
measured and the predicted energy consumption of buildings the third
00:02:19
reason is the lack of data uh it's more like a lack of details 'cause if we look at the
00:02:25
problem uh at you been scale uh we usually have
00:02:28
some g. i. yes they today lack some information so like
00:02:32
uh the getting ratio of the buildings or a um delay years that the heavenly walls
00:02:38
and so we need to make some assumption we've physics these model to uh fill those gaps
00:02:45
and on his obsession that maybe just wrong and increases the the performance gap
00:02:53
so on this is the typical work flow
00:02:55
wharf uh on machine learning model so we need
00:03:01
to have the some data for as an input um the model with
00:03:05
them prepare these data uh it means that you will clean it ten
00:03:10
try to find meaningful information trying to extract meaning full information from it
00:03:15
then the model is selected ah 'cause you we have um the choice between
00:03:21
a more multiple models and uh depending on the case study
00:03:25
some models may be better or some models maybe worse but
00:03:29
we don't know it a priori so we need to figure it out before using the
00:03:33
actual model that the model the string and wages are an output that will be our
00:03:40
uh energy demand so we basically uh as you said we have
00:03:44
a training set the production said they had this same database structure
00:03:49
uh and of course the only difference is that for
00:03:51
our training set we know would be upset energy consumption
00:03:56
uh 'cause the model need to learn from it so on and we can call these
00:04:01
like the basic requirements so the basic information we need to have a precise a model
00:04:08
if you miss one of those information uh you re lab are really bad estimation down that's
00:04:15
um the geometry uh it's the it's a the
00:04:21
g. i. yes um all the of the uh
00:04:25
uh all the c. d. where you have a all the points of the footprint of the buildings
00:04:30
uh the type of buildings that is uh and cody
00:04:34
doesn't number buddies like for example you can assign a
00:04:38
um hospital two one um church to to or whatever the
00:04:44
year of construction the hate and the number of floors of
00:04:47
course you can give more information if you have it and
00:04:50
uh hopefully the model will be more precise with more informations
00:04:55
uh so is as we said we can take this part of the seating
00:04:59
the previous example as the training set and that part of the provision set
00:05:07
um so the second step is to prepare the that as we said uh the uh um
00:05:15
i mean what do we take we extract some information from d. uh
00:05:19
from our inputs uh we we had some the scale parameters that i'm
00:05:24
not gonna cover because they are pretty self explanatory so we get the
00:05:28
perimeter from the put green to the every our biggest volume et cetera
00:05:33
and then we have some um human scale parameters
00:05:37
them some morphological parameters at neighbourhood scale needs are
00:05:42
small area around the building and that you've been scale that is um wider area around the building
00:05:49
so then it would scale uh it's mostly used to make a is to
00:05:54
as the meat the uh how much sharper will building get
00:05:59
in every direction so i will take um i take um
00:06:04
and now around that this was arrayed use is the right
00:06:08
from the latitude of this you get from the highest building and
00:06:11
the city so yeah that's basically used to tackling today and
00:06:17
maximum length of the shadow they you can have on the building
00:06:21
yeah in that the billing is you can cast and um
00:06:27
uh i divide is uh i ran into slices according to some parent number of slices is
00:06:33
it's not a fix it number but its main any good according to some parameter but i'm not gonna cover it right now
00:06:40
and um basically with the former we can uh calculated the a shot
00:06:46
with portion of every uh uh on every face on building in every direction
00:06:53
uh uh and of course for every slice we take the maximum bad is good we want tunnel
00:06:58
uh uh the maximum a shot with portion of the face
00:07:03
but you're been scale and uh uh we take an already these
00:07:06
three times the previous area uh the radius it's three times a um
00:07:13
we um we're trying to you you must get parameters there are a bit coverage ratio
00:07:19
is the proportion uh of these they're ready does all complied by the footprint of buildings
00:07:25
the aspect ratio is the ratio between the hate all the uh building that we're considering
00:07:31
and the mean we've all the road so it's the
00:07:35
stance from other buildings and entered building a it's a it's
00:07:40
the average building the average rate of the buildings in that area um we
00:07:45
study those parameters we've the pretty technical editor you know it's a investing tuning
00:07:51
and uh we're uh we're currently uh making
00:07:56
a research to estimate near impact on the end
00:07:59
and on a different models and we're trying to find other meaning full
00:08:05
with ascii parameters but uh for now and i just implementing this three
00:08:11
so the next step is to select the model uh i'm going to using a matter the
00:08:16
school the shop full speed so we basically take the training set we divide this training set tina
00:08:22
uh are are into some set one for training and one for testing
00:08:27
and we pass the and and um randomly we
00:08:31
could these two random this too so subsets randomly
00:08:36
and we estimate the uh precision of every algorithm on those two subsets
00:08:42
uh and we do this multiple time in a case like one hundred times
00:08:46
and we said the and which parameters a score the better uh
00:08:52
results so i the better performance something which she choose the that parameter
00:08:57
to uh for our uh estimation
00:09:02
then these uh let me use the full training set to train the model and the
00:09:09
we and finally have some outputs may case i i i've bow to examples to
00:09:16
case studies one is bro a a small town in the canton everybody in switzerland
00:09:22
um and the other is doing a big city in italy
00:09:27
so uh it was interesting to see the difference between using a big c. d. and a small town
00:09:34
uh also because in this small town you have
00:09:36
of course a lot less training data while uh
00:09:42
in this case for example we had a um twenty buildings
00:09:45
for the training and the ten buildings for our estimations d. um
00:09:53
use also very good but uh are uh
00:09:56
really dependent in this case on which building we
00:10:00
choose for training and which meeting views for
00:10:03
our provision 'cause we have really needed data um
00:10:10
in the second case because of touring it's a little bit
00:10:13
different we have um some very good results uh and uh
00:10:18
the other room lot less dependent from what we choose what
00:10:22
me would we choose to uh um as a case study
00:10:27
uh because of course having more data uh help us a lot
00:10:31
in uh and maintaining these good performances uh as you can see the
00:10:40
uh and uh you can see the uh arrow or that
00:10:45
is a performance arrows of ah physics base
00:10:50
model uh essentially um basics phase uh based software
00:10:56
uh i'm in those two cases the act option intelligence as being
00:11:00
a more precise in the single cases away more precise we can say
00:11:07
and the possible reasons for these uh our uh as you said a
00:11:11
beginning a lack of the kind of you been a a scale uh
00:11:16
and they can uh of course we can of meat some
00:11:20
wrong assumptions one uh considering the the uh basically physically is model
00:11:28
i'm also the fuzzy matches 'cause made
00:11:32
a bayes software uh could be uh in
00:11:36
some cases it depend on the software but usually can to need a using some
00:11:42
the only time in this case we didn't tune it we didn't uh
00:11:48
feed the uh we didn't have it we anything else than the input data
00:11:54
and i'm also the a and the third reason could be um
00:11:59
uh the fact that the model might not grasp some original difference you so for example in the
00:12:04
use of behaviour and there are some places where
00:12:08
user are more conscious on the importance off uh
00:12:12
keep in for example the windows closed when you are using
00:12:15
your room and not wasting energy them someplace in might be uh
00:12:21
in some places people may be less aware of this problem so inmate waste more energy
00:12:28
um uh finally uh the influence of the context
00:12:33
uh so the influence of d. u. been uh parameters
00:12:38
uh when the h. two probably studies we need more data cause
00:12:46
we have a a lot of four meters sec specially for the child with portion
00:12:51
so the more like the more um details we add the more
00:12:56
data we have the better but in my case we had um
00:13:01
a pretty the persian was improved by a one or a between one and seven percent uh
00:13:07
especially in the case of doing we bro he was kind off i told me is depending on
00:13:14
which uh building we were choosing chorus uh of course we had battery uh we had
00:13:20
jess twenty buildings a wedding today we had
00:13:24
five hundred buildings so that's a big difference but
00:13:29
uh creates some problem when you will uh extending eating a lot of features
00:13:35
it does get more features than a case study them that the point
00:13:40
um uh improvement uh uh there was a big
00:13:44
improvement we've uh buildings that and other big error so
00:13:49
in some cases we had the an improvement of the twenty five to around twenty five
00:13:55
percent and that's really a lot uh and this could be an interesting uh thing to study
00:14:03
um and yeah that's it so i'll thank you very much