Link to the thesis :
http://urn.kb.se/resolve?urn=urn%3Anbn%3Ase%3Amdh%3Adiva-49378
Main contributions:
- A dynamic physics-based BEM that is detailed enough to provide actionable insights, yet directly calibratable
- Methods and tools to incorporate information from meteorological reanalysis, remote sensing and on-site metering data sources
- A fully probabilistic BEM calibration framework that make also inputs probabilistic, enables data and knowledge fusion and propagates uncertainty
A Framework for Probabilistic Building Energy Modeling
1. A Framework for Probabilistic
Building Energy Modeling
Lukas Lundström
Doctoral thesis defense 2020-09-10
Opponent: Angela Sasic Kalagasidis (Chalmers)
Evaluation board: Joakim Widén (UU), Natasa Nord (NTNU), Sture Holmberg (KTH) and
reserve Ning Xiong (MDH)
Supervisors: Erik Dahlquist, Fredrik Wallin and Jan Akander
2. Content
2
Background & data Gap & research questions Methods
Case studies Contribution, conclusions & future work
3. • More efficient energy use and a larger share of renewables are needed to
mitigate global warming
• Existing buildings play a key role in this transition
• They stand for about 40% of the global energy use
• More flexible energy use could enable a larger share of intermittent
energy supply from renewables
3
Background
4. Heat balance of a Swedish multifamily building
4
Solar heat gains
Internal gains
Heating system
Space heating
Ventilation
Transmission
Internal mass
Infiltration
Manual venting
AHU
Thermal bridges
Roof
Windows
External walls
Ground floor
5. Total heat supply
5
District heating
Solar heat gains
Internal gains
Heating system
Internal DHWC
Space heating
Ventilation
Transmission
Pipe network
Internal mass
Infiltration
Manual venting
AHU
Thermal bridges
Roof
Windows
External walls
Ground floor
DHW
16. Data tuned decision making
16Understanding
Data
Information
Knowledge
Wisdom
Relations
Patterns
Insights
Context
17. Focus on the machinery of creating data calibrated
knowledge
17
Wisdom
Context
Understanding
Information
Knowledge
Relations
Patterns
Data
Insights
18. Two approaches to BEM (Building Energy Modeling)
18Complexity
Simulation software,
e.g. IDA ICE
Grey-box
models
data-driven / inverse
law-driven / forward
19. Detailed models potentially provide more insights
19Complexity
<-LessMore->
Insights
Simulation software,
e.g. IDA ICE
Grey-box
models
21. A calibrated model provides more trustworthy insights
21Complexity
<-LessMore->
Identifiable /
Directly calibratable
Insights
Simulation software,
e.g. IDA ICE
Grey-box
models
23. Push the limit of directly calibratable BEM
23Complexity
<-LessMore->
Identifiable /
Directly calibratable
Insights
Simulation software,
e.g. IDA ICE
Grey-box
models
Proposed BEM
24. Research questions
24
Q2. What level of detail is required to obtain a BEM that can provide
actionable insights on the building’s actual energy performance, yet
allow calibration in a scalable way?
VI
Q1. What information from meteorological reanalysis, remote sensing
and on-site metering data sources can be utilized for BEM calibration in
a scalable way?
IV,
VI
Q3. What are the advantages and disadvantages of using a probabilistic
BEM calibration approach?
II-
VI
25. 25
Included publications
Heat Demand Profiles of Energy Conservation Measures in Buildings and Their
Impact on a District Heating System (Lundström L., Wallin F.)
Mesoscale Climate Datasets for Building Modelling and Simulation (Lundström L.)
Adaptive Weather Correction of Energy Consumption Data
(Lundström L.)
Development of a Space Heating Model Suitable for Automated Model Generation
of Existing Multifamily Buildings (Lundström L., Akander J., Zambrano J.)
Bayesian Calibration with Augmented Stochastic State-Space Models of District-
Heated Multifamily Buildings (Lundström L., Akander J.,)
Uncertainty in Hourly Readings from District Heat Billing Meters (Lundström L.,
Dahlquist E.)
Applied
Energy 2016
Proceedings
CLIMA 2016
I
II
III
IV
V
VI
Energies
2019
Energies
2019
Proceedings
SIMS 2019
Proceedings
ICAE 2016
32. 32
Deterministic BEM
• Dynamic and physics-based
• Modelling district-heated multifamily buildings
𝜽
𝜋 𝜽 𝒚 𝒖 𝜋 𝒚 𝜽 𝒖 𝜋 𝜽
Prediction
eat
meter
lo
meter
return
supply
ipe net orold ater
supply
adiators
aps
33. 33
Thermal space
• A resistance-capacitance (RC) network
• Lumped version of ISO 52016:2017
eat
meter
lo
meter
return
supply
ipe net orold ater
supply
adiators
aps
35. 35
Thermal network
• A resistance-capacitance (RC) network
• Lumped version of ISO 52016:2017
• Building geometry represented with 5 elements
• 14 temperature nodes (states)
External
walls
Windows
Ground
floor
Roof
Internal
mass
36. System of differential equations
36
A = -
x =
𝑑𝒙 𝑡 = 𝑨𝒕 𝒙 𝒕 + 𝑩 𝒕 𝒖 𝒕 𝑑𝑡
External
walls
Windows
Ground
floor
Roof
Internal
mass
42. Case study Paper IV
Comparison with IDA ICE
• Replicates heat & temperatures patterns well
42
𝜽
𝜋 𝜽 𝒚 𝒖 𝜋 𝒚 𝜽 𝒖 𝜋 𝜽
Prediction
43. Case study
Comparison with IDA ICE
• Replicates heat & temperatures patterns well
• Fast: ~50 whole-year simulations per second
43
𝜽
𝜋 𝜽 𝒚 𝒖 𝜋 𝒚 𝜽 𝒖 𝜋 𝜽
Prediction
44. Case study
Comparison with IDA ICE
• Replicates heat & temperatures patterns well
• Fast: ~50 whole-year simulations per second
• pre-processing
• lumping
• discretization scheme for time-varying matrices
• floor area normalization, less precision required
• linear algebra when possible (Eigen C++)
• KF only on a subset of the system
• automatic differentiation
44
𝜽
𝜋 𝜽 𝒚 𝒖 𝜋 𝒚 𝜽 𝒖 𝜋 𝜽
Prediction
46. Case study
Location of unknown parameters θ
District heating
Solar heat gains
Internal gains
Heating system
Internal DHWC
Space heating
Ventilation
Pipe network
Internal mass
Infiltration
Manual venting
AHU
DHW
Fint
ggl
Hhyd, θset Urf
κrf
Uew
κew
Ugf
κgf
qV;ahu
Cinf
Uwi
Htb
Fdhw
Φdhwc
Ψp
Htot
Ctot, Htr
Hve
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floorκim
46
47. Case study: Heat transfer
coefficient for transmission (Htr)
Htr
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floor
47
48. Case study
Probability density distribution
0.4 0.5
MLE
HMC
Prior
W/K,mfl
2
Htr
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floor
48
49. Case study
Prior set at lowest level, pushed
forward
0.4 0.5
MLE
HMC
Prior
Uew
0.2 0.4
Ugl
1.0 1.5 2.0
Ugf
0.2 0.4 0.6
Htb
0.04 0.06 0.08
Urf
0.10 0.1
W/K,mfl
2
Htr
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floor
49
50. Case study
Evidence from data pushed back
Uew
0.2 0.4
Ugl
1.0 1.5 2.0
Ugf
0.2 0.4 0.6
Htb
0.04 0.06 0.08
Urf
0.10 0.15
0.4 0.5
MLE
HMC
Prior
W/K,mfl
2
Htr
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floor
50
51. Case study
Full picture
Urf
κrf
Uew
κew
Ugf
κgf
Uwi
Htb
Thermal bridges
Roof
Transmission
Windows
External walls
Ground floor
Ventilation
AHU
District heating
Solar heat gains
Internal gains
Heating system
Internal DHWC
Space heating
Pipe network
Internal mass
Infiltration
Manual venting
DHW
Fint
ggl
Hhyd, θset
qV;ahu
Cinf
Fdhw
Φdhwc
Ψp
Htot
Ctot, Htr
Hve
κim
51
52. 52
Contribution to Knowledge
A dynamic physics-based BEM that is detailed enough to
provide actionable insights, yet directly calibratable
Methods and tools to incorporate information from
meteorological reanalysis, remote sensing and on-site
metering data sources
A fully probabilistic BEM calibration framework that make also
inputs probabilistic, enables data and knowledge fusion and
propagates uncertainty
Framework
BEM
Data
53. Conclusions I –
data
• Reanalysis and satellite-based solar irradiance data are homogenous,
consistent and provide variables that are seldom metered locally
• Aerial surveying scales well, provides information on geometry and
shading/sheltering
• District-heating billing metering scales well, but still quality issues
• Domestic-hot-water metering is not available in a systematic way,
domestic-cold-water metering has potential as as a substitute
• Indoor temperature readings carry much information, but scalable
methods cannot relay on such data being available
Q1. What information from meteorological reanalysis,
remote sensing and on-site metering data sources can be
utilized for BEM calibration in a scalable way?
53
54. Conclusions II –
Building Energy Model
• Proposed model fills a gap
• More detailed and realistic than the grey-box models commonly used in literature
• Yet, directly calibratable
• Parameters are interpretable (which also allow incorporation of prior knowledge)
• Automatic differentiation is essential for computing the likelihood in a
scalable way
Q2. What level of detail is required to obtain a BEM that
can emulate reality sufficiently well to gain actionable
insights, yet allow calibration in a scalable way?
54
55. Conclusions –
framework
• The uncertainty of the resulting estimates is acknowledged. But requires
quantifying the uncertainty -> can be laborious and unfamiliar
• Full Bayesian inference requires compute-intensive sampling. However,
used BEM is well approximated with the much faster penalized
optimization
• The prior model provides the necessary regulation to identify the relatively
complex model, while also enabling information and knowledge fusion
• Kalman filtering enables data fusion on time-series level
• State augmentation enable treating also inputs as probabilistic
• more implicit knowledge ca be codified into explicit information
• smarter learning as more weight is put on periods of lower uncertainty
Q3. What are the advantages and disadvantages of using
a fully probabilistic BEM calibration approach?
55
56. Future work
• More information from aerial surveys (shading, sheltering, heat island) and
building management systems
• A seamless utilization of both metering- and model-based time-series
• Improve the modelling of the indoor temperature (the between apartment
variations)
• Building archetype modeling approach to incorporate more information
from databases (e.g. building codes) and allow information pooling
between models
56