Player is loading...
Roberto Boghetti - Simulating District Heating Networks’ hydraulics with residual physics neural networks
Roberto Boghetti, Idiap Research Institute
Monday, Nov. 15, 2021 · 10 a.m. · 13m 22s
District heating networks are energy systems allowing the heat produced to be delivered to consumers through a network of insulated pipes. Simulating their hydraulic behaviour is a complex task, which involves finding a solution to a system of discontinuous non-linear equations. As a result, existing numerical simulation methods take a significant amount of time to converge, limiting the number of design and control choices that engineers can test. We present an alternative approach based on neural networks and residual physics that has the potential to dramatically reduce the time needed to simulate the hydraulic behaviour of district heating networks while retaining a good accuracy.
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:01
h. t.
00:00:04
okay so uh hello everyone there were can presenting today is part of a big
00:00:10
a big project it's a small part of a larger project which is calling would skip
00:00:15
meet supported then financed by several suisse companies
00:00:18
and institutions which i would like to tank
00:00:22
um so i will just start by giving you some key facts about the switching networks
00:00:29
uh the speedy networks are sees themself uh our networks off
00:00:33
insulated underground pipes that distributes the heat from uh centralise the or
00:00:39
more than one actually centralise locations to sever customer buildings
00:00:44
as you can see in this illustration the you have two lines of feed line that goes from
00:00:48
the heating station in this case to the buildings
00:00:52
uh bringing hot water and therefore energy to those buildings
00:00:56
and then the return line we've cold weather that goes back to the uh central station movies get it up again
00:01:03
the system can become easily very complex and the this is actually
00:01:07
not one of the biggest one that you can already see some complexity
00:01:11
uh it serves a lot of customers it has a lot of loops and it has several eating stations also on
00:01:18
uh she meeting do systems he's kind of difficult why they
00:01:22
are relevant to the city networks are relevant because they are um
00:01:26
larry fiction systems and they you have the potential to become an eyes space and what
00:01:31
do you think so they are central to many uh a climate change strategies and reports
00:01:39
uh this is an example of the use this forecasted for industry
00:01:44
in a twenty thirty value opinion and in some countries
00:01:49
uh there will be a ah the reason i am
00:01:52
expected growth at least of new system to match the
00:01:57
targets for a twenty fifty and whatever one big problem of these uh
00:02:02
systems is that they are uh i'm kind of a hard to simulate
00:02:08
because the will require a two loops to do a physical
00:02:13
simulation saying not going to annoy you too much with the
00:02:16
equations but basically they will that uh the usually
00:02:19
the and physical based simulation methods follows the same frameworks
00:02:24
with different the questions but it's exactly the same framework so
00:02:28
we make some guesses on the opening of the bard's which is the t. v. and the
00:02:33
rotational speed of the pomp which is the and you can see there and then
00:02:37
we start i'm computing the ideal weeks of the system so again and make some guesses
00:02:44
usually on the most flow or on the pressure differences in the network
00:02:49
we built a system of the creation using those uh i guess it's it's
00:02:53
i'm not a complex system of lonely in our and this continues sick specially equations
00:03:00
uh the to the their c. friction factor uh one day
00:03:04
then we uh try to so the system using um
00:03:09
then you to rub some method which is a a
00:03:12
and i parity if a process to basically we
00:03:15
computed jacoby and metrics use these jacoby metrics do
00:03:20
um build a a linear system and find a correction factor let's say for our guesses
00:03:26
uh and the to apply to the must flow or not popular to guest then
00:03:31
we compute the error and we start again this that uh the from the first
00:03:37
uh we we start again the hydraulics a simulation entity converts
00:03:41
then we can see if our guesses on the t. v.s soda bottles and the rotational
00:03:46
speed of the palms were right otherwise again we change those guesses and this other came
00:03:52
from the start so that you take away here is that basically the simulation of the
00:03:56
system is very complex and computationally expensive so the problem is that these little about this
00:04:02
a difficult it's out so that's why we
00:04:06
might want to use machine learning to uh try
00:04:10
yeah to have a faster way of doing uh not
00:04:13
really simulations but at least influence on what could be d.
00:04:18
a a a state of our network even some constraints
00:04:23
so the naive let's say approach of machine learning in this
00:04:26
case would be to pick a case study fixed some uh
00:04:30
architecture of our this we did a network so we will take a fixed
00:04:36
number of pipes and fixed position we fixed length and picks diameter we um
00:04:42
decide some days that we want to simulate so some energy demands
00:04:46
basically from the the things that are connected weren't those simulations we've um
00:04:53
a physical based software sexy dissimilar on the bytes in this case and then we use the same inputs
00:05:00
that we have use for those models and the outputs of
00:05:03
those models to train the neural network or maturity model in general
00:05:07
so that hopefully it will be able to make prediction on new
00:05:11
data when new uh energy demand in this case for the singapore
00:05:17
so this approach as i've said the as the
00:05:20
advantage of being very fast what's that the model restraint
00:05:24
however it does some disadvantages so it requires a lot of data
00:05:28
it's very hard to junior allies such model mm because uh what we have done is to take
00:05:35
a fixed network architecture and make a model for that architecture for that
00:05:40
particular this to the network so it doesn't really generalised one new network
00:05:46
and uh it's second computer yeah black box so we don't know what is happening inside our
00:05:51
neural network our maturity model which is know what is going in and what's going out to
00:05:56
there are some techniques that we have an idea of what is happening but
00:06:00
still it's kind of a black box for us and um the main is about
00:06:06
but these advantage is also that we are not exploiting the domain knowledge it we have so we know
00:06:11
what are the physical constraints of the problem but we are just
00:06:14
is regarding these knowledge and not using it so that's a waste
00:06:19
it's another maybe smarter approach would be to
00:06:24
use what's called receiver physics neural networks or
00:06:28
or um physics informed kind of neural network
00:06:33
uh it has several names but basically what we do is again we pick a case study fixed an uh
00:06:39
we generate some random inputs for the energy demands so we have
00:06:43
some fixed inputs which are the network architecture it characteristics and some variable
00:06:49
uh um inputs that is the energy that is requested by all those
00:06:55
uh i'm a substation of those buildings actually we can
00:06:59
do something smarter and not really generated randomly completely randomly
00:07:03
but let's take it like it is now so we put those inputs into a neural network and then what
00:07:12
and neural network does usually doing this painting go to minimise let's say a function
00:07:19
in uh the previous case the function was the error between our pretty uh our simulations and the production of the network
00:07:25
in this case i wear a loss function our cars would be some kind of
00:07:30
a physical constraints that we need to
00:07:33
uh minimise so the advantages of this approach
00:07:37
is that again the insurance once the the training of the neural network restraint would
00:07:42
be very fast compared to physical models it
00:07:46
doesn't requires a to do simulations before hand
00:07:51
and it's a little more explainable than the previous approach
00:07:54
because we can have for example intermediate quantities that we
00:07:58
uh compute in our network we don't really need to have only are
00:08:03
a neural network full of new rooms uh that we don't know what they are doing
00:08:09
uh but still it's not a hundred percent explainable so it's halfway there
00:08:17
and again the disadvantages of this approach is mainly that it's very hard to general lice
00:08:22
of course we can change the inputs and devise
00:08:24
a network that can take a any kind of
00:08:30
uh this critic network as an input but then it would be a battery very complex modern it would be super hard
00:08:36
to train so that's hopefully our goal but it will take
00:08:40
time so for now let's say that it's hard to denver lies
00:08:45
so rude just give you a simple example in these keys actually you
00:08:48
may not need to do the double loop simulation uh to get the simulation
00:08:54
because it's just a three bayes net too but it's just to give you an idea of sand and we would build up from here
00:09:00
so again we will have some inputs defeats network topology characteristic of the pipes and
00:09:05
the bible energy demand of the consumer that's the only thing that changes basically from one
00:09:10
a a sample it say to the other the outputs the two you
00:09:14
want are the most full inch by and the pressure losses in each bite
00:09:19
the constraints that we we use are the conservation of mass so basically
00:09:24
whatever enters one no one has to accept that note not more not less
00:09:30
and the uh and we won the the energy
00:09:32
delivered at the consumers it's what they asked for basically
00:09:37
so for the conservation of mass we can do something pretty interesting uh so instead of
00:09:43
putting this constraining the loss function we can put this constraint into the abstract network so um
00:09:50
i don't know if any of you knows about cycles in a graph
00:09:55
but basically the network is exactly a graph made out of nodes and edges
00:10:00
if the edges being the pipes and eating stations and some and the buildings
00:10:04
and the nodes being the junction snazzy um
00:10:09
so we can find some cycles in a graph which are passed the
00:10:13
goes from one point with cell without meeting the same edge yeah twice
00:10:20
and we can also find a ah subgroup
00:10:23
let's say a subset of these uh um
00:10:28
uh cycles from week we can be uh the
00:10:31
by super imposition all the other uh cycles autograph
00:10:35
and if we can and estimate the mass plough into uh these uh uh
00:10:43
elementary cycles as they are called by spreading position we can find the
00:10:48
mass flow in each element of our network and it we'll uh um
00:10:54
um automatically a preserve the conservation of mass basically
00:10:59
the second constrain that we would even this is actually in the last function that we want putting ice for
00:11:05
is uh uh to satisfy the energy demand of the consumer
00:11:09
in this case we can and x. p. seat this uh a constraint in different ways
00:11:16
could be um mean squared error on the values uh
00:11:19
it could be whatever it's a mm but what is the
00:11:24
key take away here that we are not using any simulation
00:11:27
that that just an uh those constraints to train our network
00:11:32
so as you can see their training of these network is pretty fast the next again the network
00:11:37
is very small it's getting network and it's very easy because at three days it has no loops
00:11:44
but uh uh we've only four to five minutes of training
00:11:49
uh we have a model that can output a full year of
00:11:53
simulations how the simulation for one full year in less than one second
00:11:58
zero point twenty five seconds while laughing text based
00:12:02
uh approach will take a three minutes around thing it's
00:12:07
we can extend this to look to network as i was saying by introducing new
00:12:12
uh and the more um physical constraints in this case we have the
00:12:18
constant um another who's of the cutest chick of second low
00:12:23
in the electrics you could so basically we know that in our pipes and just
00:12:28
in general we have pressure losses or pressure least especially if you have a pop
00:12:33
and if we take our newly meant to recycle
00:12:36
recycle of our network the directed some of all those
00:12:40
pressures losses hand leaves should be equal to zero so
00:12:44
that's something that we can uh with the mice for
00:12:47
because of course when you have a loop um you don't know a priori what the flow is going so when you have
00:12:53
but the loop starts you can grant lefty can go right or it can split and it depends on
00:12:58
the pressure differences inside your network so that's what we have done as you can see the reason um
00:13:05
but uh learning crew put a network which takes a bit more because it's
00:13:10
more complex than the previous case but still it's a reasonable amount of time
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
Flowsheets and process unit models
Prof. François Maréchal, EPFL
March 8, 2012 · 10:16 a.m.
106 views