Player is loading...
Dr. Giuseppe Peronato - HeatLearn: a convolutional neural network to assess the building energy consumption at regional scale from open land-use maps
Dr. Giuseppe Peronato, Idiap Research Institute
Monday, Nov. 15, 2021 · 10:30 a.m. · 13m 07s
Energy density is an important indicator for urban energy planning. It can be used by planners and decision-makers to identify clusters of energy demand. We have developed a deep learning model to predict the urban energy density using only land-use maps available for the whole European continent as open data. The model was trained with the building heating demand dataset available in the Canton of Geneva. Our approach allows for a quick energy analysis at the regional scale, and is ready to be deployed as a web tool.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
most set the i'm gonna talk about this project uh in in a heap learn a um
00:00:07
the the idea is to go back to the uh one a fundamental which a
00:00:12
research question in a in urban planning which is uh uh how does the orb on
00:00:18
um uh the the the or been the fabric performance and so find a
00:00:25
uh uh uh the the determinants of uh out of the performance of
00:00:29
the city based just on the or bum a morphology so in this case
00:00:33
we're gonna start from a um a body like a energy perspective
00:00:40
uh uh which uh with a with a question that is uh what is
00:00:43
segregating keep building demand of a given urban area that's a a very generic uh
00:00:49
question but it's uh it's very useful a tool
00:00:53
because it's dead the printer preston the should be answer
00:00:56
before planning any large scale energy system in a in
00:01:00
a city for example i distinctively networks we have to
00:01:04
uh we need i'm anymore um he didn't demand uh from
00:01:09
the buildings for the system to be act economically bible um
00:01:15
and then uh there is a a problem and this is a problem that has
00:01:19
been a already addressed thing in the literature usually with our engineering approach and uh
00:01:26
these engineering approach uh have 'em yet the disadvantage that they need us about that assets
00:01:33
and um maybe didn't mention that we are gonna uh the the globe was also
00:01:37
to apply to the large scale uh so the continental scale uh for example europe
00:01:42
and so we need a data available for all spatial location
00:01:46
and seen also the the the the can have a consistent approach
00:01:49
so uh yeah this is just a than example from a from an existing budget from the from
00:01:55
the d. touch over you see uh all they needed a data to perform dingy remodel and uh um
00:02:02
then uh the research question uh is uh another such question with
00:02:07
which is more uh on then much learning perspective so can much earning
00:02:12
on the land use maps which is this kind of marks
00:02:15
the that are um available for the entire a european union
00:02:19
and the including a and some other countries like switzerland uh can nah a machine
00:02:25
on these data sets upper for outperform engineering models as the one be seen that before
00:02:30
so that was the fundamentals is just question and to go for the on the
00:02:35
uh uh sorry on the on the context of this project uh i'm presenting so
00:02:40
um this is a part of a research project called the
00:02:43
enter marks you've seen maybe be a outside the the boot
00:02:47
uh by our colleagues from crime and uh you're gonna have some presentation this afternoon but just to
00:02:53
one introduce why uh we did that so and
00:02:56
what isn't there max's out what platform for mapping
00:03:00
selected energy data sets which are listed in the open air uh research uh
00:03:04
are positive which is a scientific a repository for uh yeah the research community
00:03:09
and um as part of the spectrum we have some recognition models
00:03:12
that make use of these that that's that's available in the platform
00:03:16
to run on the fly analysis and the idea was to
00:03:19
develop like aggression model based on a machine learning using then uh
00:03:23
the europeans of them at as the input which is uh one of the data sets are listed eh in this but from
00:03:29
so yeah uh and europeans that don't map so here we have an an example he's kind of
00:03:36
a noisy and um yeah it's not very high resolution one h. picks a two point five meters
00:03:42
uh basically what we can extract is the morphology of the um
00:03:46
the the buildings all of the all the or been a area
00:03:50
and then i um we train other model using some uh uh that
00:03:55
that's that's from a cheating demand which are available and a condo journey about
00:03:59
where uh buildings are required by law to uh give their
00:04:02
uh on well uh it hitting consumption to the to the state
00:04:06
and then this uh this data is very interesting it's a very unique that the set i
00:04:10
think in in in europe uh and then we can train the model based on on this data
00:04:15
and then of course we have um more uh cholesterol data which divides the the shape of the
00:04:21
of the buildings and the and the footprint of the of the buildings and that other for area
00:04:26
and so we used we train our model on
00:04:29
these the uh um i d. c. uh from french
00:04:32
uh on this the the points to shut up which is the yeah this building keeping demanded a set
00:04:38
and yet the predictors as a side is the european certain matt but then we added
00:04:43
a mask uh oh we're uh to to tell the model where are we don't have data from uh
00:04:49
from the geneva that's set um because yeah uh the men should be of about one third of the of the
00:04:55
buildings that one for which we have to use that you stated which is it still a kind of a lot
00:05:00
but uh yeah uh from uh for each eh area we don't have all of them
00:05:06
and then uh yeah a classic approach uh uh uh
00:05:09
in this kind of work rework the tiles a regular tiles
00:05:14
then the ties are overlap to augment the the data we can train up
00:05:18
and uh then we also feet vista the style uh to uh to have uh yeah more data to train our model
00:05:25
and then uh we trained a channel uh using the so the
00:05:30
two input images um the the mask and you're gonna certain map
00:05:34
and then the the resulting uh uh uh information uh
00:05:38
uh is the one by one information i used it
00:05:41
uh the is the heating a undercutting amount for each time so aggravated at the at the trial level
00:05:47
and then uh with this first approach we compared with uh um the the baseline
00:05:53
model which is appalling on your regression and we see we be uh improve uh the
00:05:59
um the results um this is another three hundred meters style and
00:06:04
the um and then the the question is also a whether a um
00:06:10
these uh what is the question a different title a tie levels
00:06:14
and these are these are results for different resolution
00:06:17
so as you see uh uh the larger uh
00:06:21
the better which is a yeah a condo maybe
00:06:24
into it it um you know uh uh um
00:06:29
the the but the this probably means that we have a kind of a kind of low
00:06:34
uh i work for a five hundred meters silence this probably means that the
00:06:38
geneva organ topic is pretty consistent within this uh this range of uh and uh
00:06:43
uh you average out the the possible differences of the usages that we can find in the in the serious
00:06:50
and uh and then the compared we disagree a reference injury model presented
00:06:54
before so as a set based on an extensive number that assets and accuracy
00:06:58
a it was not available supposed to also the idea to
00:07:01
a to a test accuracy in the image in your area
00:07:05
um the ad should other sets from this project where use one for the
00:07:10
they actually keep density and one uh the uh cross floor area
00:07:14
uh which is uh yeah boss our roster models at the uh one hundred meter a verse or solution
00:07:21
and um yeah b. u. b. first uh uh adapted um uh the
00:07:26
um the data from the syringe your mother to be compared with uh with
00:07:29
our model of from uh from geneva so a question for weather and
00:07:33
the correction for the area which uh uh was not necessarily same between the
00:07:37
um uh the the geneva that the sets and uh and uh and the heart marks that that's it
00:07:44
and then uh we is so these a reference injury mother was on not one other meat of tiles
00:07:49
um we seen before that the performance was not a very good that one of the meter buys for our model
00:07:55
so the idea was to improve the the performance use our context aware mother
00:08:00
where we she that were mother would uh some information about the surrounding area
00:08:04
so we want to predict the the centre of the tie the one on the
00:08:07
meter power by we keep the model the information about the are allowed to tile
00:08:12
and the the results so compared with the um
00:08:16
uh the the our model with a reference engineering model
00:08:21
uh the performance could say it's kind of a similar but uh yeah we have to consider that i
00:08:27
mean it's this not that completely fair comparison because
00:08:30
the engineer mother was uh as a much more up
00:08:33
eh input uh that that's that's well uh we rely in our
00:08:37
for our model just on a on their europeans at the room map
00:08:41
um and so yeah these are the hour they should be spent the for uh uh the the one hundred meter
00:08:47
but we sing before that if we go to a
00:08:49
lot um a larger tiles we have a much better results
00:08:54
and then uh it's um maybe something uh yeah and
00:08:58
uh this work was present in a poster and uh
00:09:01
uh at says but conference in september and uh since than the b. b. b. apply this
00:09:06
model online and so this them by happy i mean to be able to present to you the
00:09:10
this model it says that was uh just for a poster but also to to see uh
00:09:15
well how does he look a like a one not deployed the in the in the maps platform
00:09:21
i'm just a few words about uh the methodology or
00:09:26
to bring this this work online and the idea for uh
00:09:31
a disk aggression models um in a in a marxist to uh to
00:09:36
be able yet to uh to show how we can uh for example use
00:09:40
the um uh the plot is the deployment of maturity model how can we
00:09:44
work online with uh with data sets the um from uh from my remote
00:09:48
so we used to um the doctor um architecture um
00:09:53
and then um the um yeah we have a an
00:09:57
architecture that managed a a and we did maybe i'm managing the inputs and outputs from the and the calculation models
00:10:03
um we um and use a synchronous that's run by the
00:10:08
by the user so we can support with people uh uses
00:10:12
uh and uh run in the running the task and basically yeah the user selects the uh
00:10:18
the the european so the map the as the input the boundary for uh uh their analysis
00:10:24
and then um the the the the heap learn a
00:10:28
language model uh deploys the pencil from model though also
00:10:32
uh uses the heating degree days from the uh the other maps database which come from your start
00:10:38
and then uh yeah um as the output has this
00:10:41
um the map and some summary statistics about the the real
00:10:46
um so there's some indication about the the to this
00:10:51
work this model has been trained on a on in geneva
00:10:54
and so we have to be careful or the difference that might have in other areas
00:10:59
and then we neglect also the usage of the buildings um
00:11:03
i accept what can be a study only from the building morphology
00:11:08
and um and then uh it yet application you is we need to your country so important
00:11:14
to me not to switch around the because they're more heating degrees days in a presbyterian euro start
00:11:20
uh but we are did some uh and data for uh for the geneva area
00:11:26
and then some uh um yellow density tiles are
00:11:29
excluded uh but they would be not much interest anymore
00:11:33
um so maybe yeah uh you can uh show uh
00:11:37
in the minutes of uh i have left i think
00:11:41
uh how it works the online and uh if you won yeah
00:11:44
we select a european set them up as the input with whom
00:11:49
then we see when we to mean uh we can select some
00:11:52
uh boundaries uh for our analysis in this case we see the municipalities
00:11:57
um then uh we selected the every solution of our knowledge is in this case it's a hundred meters
00:12:03
uh the ear to doesn't twenty because b. can use the historical data
00:12:07
from a a heating grates days from us that and then be run
00:12:11
how about work task takes a few seconds depends on the on the size um
00:12:16
of the area and then we'll uh have the results loaded back into our interface
00:12:21
and then uh yeah uh will be we can we can see the
00:12:26
the legend for our data the parameters and back you can download the
00:12:30
the data with the parameters and so they can be reasonable for example
00:12:34
using that uh um we uh j. s. or some other g. s. after
00:12:39
i i think i run out of time so i thank you for your attention and
00:12:43
uh the other references for their uh resources uh you can find online about the this project
00:12:49
i will not acknowledge the the help from the development team of uh
00:12:53
or this project uh and uh also i mean the inner maps
00:12:58
a project that the which is funded by the european union and then the the
Conference Program
Workshop introduction by Jérôme Kämpf & Henning Müller + Keynote speech Pr. Dr. Guglielmina Mutani - Place-based energy models for sustainable cities
Pr. Dr. Guglielmina Mutani, Politecnico di Torino
Nov. 15, 2021 · 9 a.m.
333 views
Q&A (Keynote speech: Pr. Dr. Guglielmina Mutani)
Pr. Dr. Guglielmina Mutani, Politecnico di Torino
Nov. 15, 2021 · 9:49 a.m.
Roberto Boghetti - Simulating District Heating Networks’ hydraulics with residual physics neural networks
Roberto Boghetti, Idiap Research Institute
Nov. 15, 2021 · 10 a.m.
Q&A (Roberto Boghetti)
Roberto Boghetti, Idiap Research Institute
Nov. 15, 2021 · 10:13 a.m.
Amine Weibel - Machine learning forecast for photovoltaic power stations (PV) with spatial interpolation on the COSMO-7 model. A real ‘use-case’ on the Czech Republic
Amine Weibel, HES-SO
Nov. 15, 2021 · 10:16 a.m.
Q&A (Amine Weibel, HES-SO Valais-Wallis)
Amine Weibel, HES-SO
Nov. 15, 2021 · 10:27 a.m.
101 views
Dr. Giuseppe Peronato - HeatLearn: a convolutional neural network to assess the building energy consumption at regional scale from open land-use maps
Dr. Giuseppe Peronato, Idiap Research Institute
Nov. 15, 2021 · 10:30 a.m.
Q&A (Dr. Giuseppe Peronato)
Dr. Giuseppe Peronato, Idiap Research Institute
Nov. 15, 2021 · 10:43 a.m.
Pr. Dominique Genoud & Jérôme Treboux - Integrating machine learning and AI in practical energy applications : a review of these last 10 years activities at the Institute of Information Systems
Pr. Dominique Genoud & Jérôme Treboux, HES-SO Valais-Wallis
Nov. 15, 2021 · 10:46 a.m.
Q&A (Pr. Dominique Genoud & Jérôme Treboux)
Pr. Dominique Genoud & Jérôme Treboux, HES-SO Valais-Wallis
Nov. 15, 2021 · 10:59 a.m.
Dr. Marco Tognoli - Fuel-saving potential simulation and estimation when optimally predicting the supply set-point conditions of a District Heating Network with detailed system data-driven modelling
Dr. Marco Tognoli, Politecnico di Milano
Nov. 15, 2021 · 11:31 a.m.
Q&A (Dr. Marco Tognoli)
Dr. Marco Tognoli, Politecnico di Milano
Nov. 15, 2021 · 11:44 a.m.
Pr. David Wannier & Jean-Marie Allder - Research advance in Energy prediction for electrical vehicles and green mobility hotels, the results of the eVIP project
Pr. David Wannier & Jean-Marie Allder, HES-SO Valais-Wallis
Nov. 15, 2021 · 11:46 a.m.
Q&A (Pr. David Wannier & Jean-Marie Allder)
Pr. David Wannier & Jean-Marie Allder, HES-SO Valais-Wallis
Nov. 15, 2021 · 11:58 a.m.
Dr. Kavan Javanroodi - Extending the concept of energy hub to facilitate sector and spatial coupling
Dr. Kavan Javanroodi, Solar Energy and Building Physics Laboratory (LESO-PB) at EPFL
Nov. 15, 2021 · 11:59 a.m.
143 views
Q&A (Dr. Kavan Javanroodi)
Dr. Kavan Javanroodi, Solar Energy and Building Physics Laboratory (LESO-PB) at EPFL
Nov. 15, 2021 · 12:16 p.m.
Pr. Pierre Roduit - Prediction of the electricity flexibility of households using Machine Learning
Pr. Pierre Roduit, HES-SO Valais-Wallis
Nov. 15, 2021 · 12:18 p.m.
Q&A (Pr. Pierre Roduit)
Pr. Pierre Roduit, HES-SO Valais-Wallis
Nov. 15, 2021 · 12:32 p.m.
Loïc Puthod - An open-data acquisition toolchain for AI applications
Loïc Puthod, Centre de recherche Crem
Nov. 15, 2021 · 1:59 p.m.
Q&A (Loïc Puthod)
Loïc Puthod, Centre de recherche Crem
Nov. 15, 2021 · 2:10 p.m.
Cédric Mugabo Serugendo - EnerMaps: The open data tool empowering your energy transition
Cédric Mugabo Serugendo, Centre de recherche Crem
Nov. 15, 2021 · 2:16 p.m.
&Q (Cédric Mugabo Serugendo)
Cédric Mugabo Serugendo, Centre de recherche Crem
Nov. 15, 2021 · 2:27 p.m.
Recommended talks
Urban morphology, energy needs and artificial intelligence
Roberto Boghetti, University of Pisa, IT
May 7, 2019 · 10:51 a.m.
121 views