Sie sind auf Seite 1von 150

Energy management strategies

THÈSE
pour obtenir le grade de
DOCTEUR DE L’UNIVERSITÉ DE GRENOBLE

for smart grids


ALPES
Spécialité : spécialité
Arrêté ministériel : 7 août 2006

Présentée par
Prénom NOM

Ph.D. thesis presented by:

Peter Paum
Thèse dirigée par Prénom NOM et
codirigée par Prénom NOM
préparée au sein du
laboratoire nom du labo (sigle du labo)
dans l’école doctorale nom de l’école doctorale (sigle de l’ed)

Supervised by:

MonMazen
sujet de Alamir
thèse sur
Mohamed-Yacine Lamoudi & Claude Le Pape
plusieurs lignes

Thèse soutenue publiquement le date de soutenance,


devant le jury composé de :

January 9 2017
Prénom NOM th
Labo de bidule, Rapporteur, Présidente du jury
Prénom NOM
Labo de bidule, Rapporteur
Prénom NOM
Labo de bidule, Examinateur
Prénom NOM
Labo de bidule, Examinateur
Prénom NOM
Labo de bidule, Directeur de thèse
Prénom NOM
Labo de bidule, Encadrant de thèse
Introduction DMPC Randomized Conclusion

Introduction

Challenge by 2050...

×1.5
Energy consumption
will grow by 50%
2
÷
CO2 emissions
need to be halved

We have to become
3 times more efficient

Note: Forecast for 2050 compared to 2009 levels

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 1 / 47
Introduction DMPC Randomized Conclusion

Introduction
The emergence of smart-/micro grids

Production islands
operated by
cities/industries

1900 2000

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 2 / 47
Introduction DMPC Randomized Conclusion

Introduction
The emergence of smart-/micro grids

Production islands
operated by
Centralization for
better availability
cities/industries

1900 2000

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 2 / 47
Introduction DMPC Randomized Conclusion

Introduction
The emergence of smart-/micro grids

Production islands
operated by
Centralization for Decentralization for
better availability "territory" services
cities/industries

1900 2000

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 2 / 47
Introduction DMPC Randomized Conclusion

The AMBASSADOR project

European FP7 project dedicated to the development of autonomous


energy management systems at building- and district level

"Develop ecient energy op-


timization systems for smart
districts"

• 15 partners
• 11 dierent EU countries
• 4 years (2012 - 2016)
• 9.9 Me

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 3 / 47
Introduction DMPC Randomized Conclusion

Thesis objectives

Design control algorithms able to improve energy management in


smart grids
1 Increase energy eciency at district level
2 Improve the robustness of the electricity grid
3 Design generic, scalable and modular solutions

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 4 / 47
Introduction DMPC Randomized Conclusion

Overview of works

Distributed MPC
for smart grids

MPC for district heating


networks

EV charging station
control under
uncertainties

Battery sizing for PV


power plants

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 5 / 47
Introduction DMPC Randomized Conclusion

Overview of works P. Paum, M. Alamir, and M.Y. Lamoudi (2014).  Comparison of


a primal and a dual decomposition for distributed MPC in smart
In: IEEE International Conference on Smart Grid
districts.
Communications
Distributed MPC
P. Paum, M. Alamir, and M.Y. Lamoudi (2015).  Scalability
for smart grids
study for a hierarchical NMPC scheme for resource sharing
problems. In: European Control Conference (ECC),
pp. 14681473

MPC for district heating F. Bourry et al. (2015).  Simulation for the evaluation of energy
management algorithms at the district level - Example of use case
networks
from the AMBASSADOR project. In: PowerTech, 2015 IEEE
Eindhoven, pp. 16

EV charging station P. Paum and M. Alamir (2015b).  Procédé de gestion en énergie


control under d'une station de recharge de véhicules électriques. Patent
uncertainties

Battery sizing for PV P. Paum and M. Alamir (2015a).  Procédé de contrôle d'une
power plants centrale de production électrique. Patent

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 5 / 47
Introduction DMPC Randomized Conclusion

Overview of works

Distributed MPC
for smart grids

MPC for district heating


networks

EV charging station
control under
uncertainties

Battery sizing for PV


power plants

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 5 / 47
Outline

1 Distributed MPC for smart grids

2 Randomized algorithm for EV charging stations

3 General conclusion
Outline

1 Distributed MPC for smart grids


System model
Model Predictive Control
The centralized control problem
Primal and dual decomposition
Bundle method
Validation

2 Randomized algorithm for EV charging stations

3 General conclusion
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

The smart grid

Renewable Building
Production
Residential
Grid
PG ≤ Plim

Electrical
Vehicles Storage

Common grid connection with a limited capacity Plim


No consideration of the distribution grid topology and of stability
issues (voltage-, frequency constraints,...)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 6 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

The smart grid

Simplied view:

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 7 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Building model

Linear rst-order model

ϑt+1 = ϑt + α1 (ϑext − ϑt ) + α2 PBui

ϑ, ϑext indoor- and outdoor temperature


PBui heating/cooling power of the HVAC system
α1 , α2 model coecients

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 8 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Building model

Linear rst-order model

ϑt+1 = ϑt + α1 (ϑext − ϑt ) + α2 PBui

ϑ, ϑext indoor- and outdoor temperature


PBui heating/cooling power of the HVAC system
α1 , α2 model coecients

Wide literature on building modeling & control available


Signicant energy and cost savings are reported
Mohamed Yacine Lamoudi, Mazen Alamir, and Patrick Béguery (2012).  Model predictive
control for energy management in buildings. Part 1: Zone Model Predictive Control. In:
Nonlinear Model Predictive Control. Vol. 4. 1, pp. 2126

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 8 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Battery model

Linear rst-order battery model

if PBat ≥ 0
(
t+1 E t + τ · η + · PBat
E =
E t + τ · 1/η − · PBat otherwise

E stored energy
PBat charging/discharging power
η+, η− charging/discharging eciencies
τ time step

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 9 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Electric vehicle charging station model

EVCS consisting of NCP charging points

NCP
X
PEVCS = PBat,v
v =1
Evt+1 = Evt + η · τ · PBat,v ; PBat,v ≥ 0

PEVCS power consumption of the EVCS


Ev energy delivered to EV at v -th charging point
PBat,v charging power at v -th charging point
η charging eciency
τ time step

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 10 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Renewable energy source model

Non-controllable renewable energy sources


Represented through their predicted production
prole P
b RES only

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 11 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model Predictions

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model Predictions Objective

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model Predictions Objective

Optimization Problem

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model Predictions Objective Solver

Optimization Problem

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Model Predictive Control MPC for energy management in buildings Model Predictive control

MPCMPC for building


- An intuitive conceptEnergy management
The ingredients ...
Anticipate the future to nd the optimal control at each control instant.

Model Predictions Objective Solver

*
Optimal
Optimization Problem solution

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

P. Paum
M.  Schneider
Y. Lamoudi Electric/Gipsa-lab - Energy
- Schneider-Electric/Gipsa-lab management
DMPC for Energy strategies
management in buildingsfor smart grids  09/01/2017
- 11/29/2012 12
12 //5247
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective

Minimize energy costs


Respect the global power
limitation PG ≤ Plim

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective

Minimize energy costs


Respect the global power
limitation PG ≤ Plim

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective

20 Minimize energy costs


powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24

30
20
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24

30
20
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24

30
20
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24 2 Battery supplies to building so

30 that PG ≤ Plim
20
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24 2 Battery supplies to building so

30 thatPG ≤ Plim
20 3 PG = 0
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24 2 Battery supplies to building so

30 thatPG ≤ Plim
20 3 PG = 0
10 ϑindoor comfort bounds
4 battery charges from PV
0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24 2 Battery supplies to building so

30 thatPG ≤ Plim
20 3 PG = 0
10 ϑindoor comfort bounds
4 battery charges from PV
0
0 2 4 6 8 10 12 14 16 18 20 22 24
5 buy expensive energy from grid
0.2
0.15 to keep the battery charged
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Smart grid control objective


1 2 3 4 5 6
20 Minimize energy costs
powers [kW]

0
Respect the global power
Plim PG PB
limitation PG ≤ Plim
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
1
temperature [◦ C] battery SoC

0.8 1 charge battery while energy is


0.6
0.4 cheap
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24 2 Battery supplies to building so

30 thatPG ≤ Plim
20 3 PG = 0
10 ϑindoor comfort bounds
4 battery charges from PV
0
0 2 4 6 8 10 12 14 16 18 20 22 24
5 buy expensive energy from grid
0.2
0.15 to keep the battery charged
tari

0.1
0.05 tari Γ 6 Respond to a planned load
0 virtual tari Γ+λ−µ
shedding event
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategiessdfsdf


for smart grids  09/01/2017 13 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

L` local objective fc.


Minimize L` (x` , r` )
X
rx
{ ` , ` }`∈S
`∈S

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

L` local objective fc.


Minimize L` (x` , r` )
X
rx
{ ` , ` }`∈S
`∈S
x` local variables
(indoor temperatures,...)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

L` local objective fc.


Minimize L` (x` , r` )
X
rx
{ ` , ` }`∈S
`∈S
x` local variables
(indoor temperatures,...)
r` local resource prole
(power consumption,...)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

L` local objective fc.


Minimize L` (x` , r` )
X
rx
{ ` , ` }`∈S
`∈S
x` local variables
(indoor temperatures,...)

Subj. to: r` local resource prole


(power consumption,...)
h` local constraints
h` (x` , r` ) ≤ 0 ∀` ∈ S (sub-system model,...)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

L` local objective fc.


Minimize L` (x` , r` )
X
rx
{ ` , ` }`∈S
`∈S
x` local variables
(indoor temperatures,...)

Subj. to: r` ≤ Rlim r`


X
local resource prole
(power consumption,...)
`∈S
h` local constraints
h` (x` , r` ) ≤ 0 ∀` ∈ S (sub-system model,...)
Rlim global resource limit

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

!
L` local objective fc.
Minimize L` (x` , r` ) + ΓT · r`
X X
rx
{ ` , ` }`∈S
`∈S `∈S
x` local variables
(indoor temperatures,...)

Subj. to: r` ≤ Rlim r`


X
local resource prole
(power consumption,...)
`∈S
h` local constraints
h` (x` , r` ) ≤ 0 ∀` ∈ S (sub-system model,...)
Rlim global resource limit
Γ energy tari

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Resource sharing problem

System composed of many dynamically uncoupled sub-systems


Coupling through a limited shared resource

!
L` local objective fc.
Minimize L` (x` , r` ) + ΓT · r`
X X
rx
{ ` , ` }`∈S
`∈S `∈S
x` local variables
(indoor temperatures,...)

Subj. to: r` ≤ Rlim r`


X
local resource prole
(power consumption,...)
`∈S
h` local constraints
h` (x` , r` ) ≤ 0 ∀` ∈ S (sub-system model,...)
Rlim global resource limit
if `∈S r` ≥ 0
(
Γ+
P
Γ energy tari
Γ=
Γ− otherwise Γ+ = buying, Γ− = selling

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 14 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Centralized MPC controller

Drawbacks of a centralized MPC Centralized MPC


controller:
Scalability
Robustness
Modularity
Privacy

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 15 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Distributed MPC controller

Split centralized controller into


Coordinator several sub-problems and a
coordinator
MPC MPC MPC MPC MPC MPC
Iterative process involving some
communication between the
coordinator and the
sub-problems
Convergence to the global
optimum

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 16 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal- and dual decomposition

Iterative information exchange in the two methods


Primal decomposition Dual decomposition

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 17 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal- and dual decomposition

Iterative information exchange in the two methods


Primal decomposition Dual decomposition

Coordinator
Power
profile Objective value
& (sub-)gradient

MPC MPC

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 17 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal- and dual decomposition

Iterative information exchange in the two methods


Primal decomposition Dual decomposition

Coordinator Coordinator
Power Dual
profile variables Objective value
Objective value
& (sub-)gradient & opt. power profile

MPC MPC MPC MPC

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 17 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Qualitative comparison

Primal decomposition: Dual decomposition:


Feasibility at each iteration Feasibility only after convergence
( `∈S r` ≤ Rlim ) sdfsdfsdfsdfsdfsdfsdf
P

The centralized problem only The centralized problem must be


has to be convex strongly convex
nb. of decision variables nb. of decision variables is
proportional to NS independent of NS

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 18 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal decomposition

Initial centralized problem:


!
Minimize L` (x` ,r` ) + Γ · r`
X X
T
rx
{ ` , ` }`∈S
`∈S `∈S

Subj. to: r` ≤ Rlim


X

`∈S
h` (x` ,r` ) ≤ 0 ∀` ∈ S
if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 19 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal decomposition

Initial centralized problem:


!
Minimize L` (x` ,r` ) + Γ · r`
X X
T
rx
{ ` , ` }`∈S
`∈S `∈S

Subj. to: r` ≤ Rlim


X

`∈S
h` (x` ,r` ) ≤ 0 ∀` ∈ S
if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 19 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal decomposition

Initial centralized problem:


!
Minimize L` (x` ,r` ) + Γ · r`
X X
T
rx
{ ` , ` }`∈S
`∈S `∈S

Subj. to: r` ≤ Rlim


X

`∈S
h` (x` ,r` ) ≤ 0 ∀` ∈ S
if `∈S r` ≥ 0
(
Sub-problem ` Γ+
P
Γ=
(for given r` ): Γ− otherwise

J` (r` ) := Minimize L` (x` ,r` )


x `

Subject to: h` (x` ,r` ) ≤ 0


P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 19 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal decomposition

Initial centralized
Coordinator problem:
problem:
!
Minimize J` (r` ) r`
X X
T
+Γ ·
r
{ ` }`∈S
`∈S `∈S

Subj. to: r` ≤ Rlim


X

`∈S
h` (x` ,r` ) ≤ 0 ∀` ∈ S
if `∈S r` ≥ 0
(
Sub-problem ` Γ+
P
Γ=
(for given r` ): Γ− otherwise

J` (r` ) := Minimize L` (x` ,r` )


x `

Subject to: h` (x` ,r` ) ≤ 0


P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 19 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Primal decomposition
Diculty: J` (r` ) can only
be evaluated point-wise !
Initial centralized
Coordinator problem:
problem:
!
Minimize J` (r` ) r`
X X
T
+Γ ·
r
{ ` }`∈S
`∈S `∈S

Subj. to: r` ≤ Rlim


X

`∈S
h` (x` ,r` ) ≤ 0 ∀` ∈ S
if `∈S r` ≥ 0
(
Sub-problem ` Γ+
P
Γ=
(for given r` ): Γ− otherwise

J` (r` ) := Minimize L` (x` ,r` )


x `

Subject to: h` (x` ,r` ) ≤ 0


P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 19 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition

Initial centralized problem:


!
Min x r r
X X
T
L` ( ` , ` ) + Γ · `
{r` ,x` }`∈S
`∈S `∈S
Subj. to: hX(x , r ) ≤ 0
` ` ` ∀` ∈ S
r ≤ Rlim
`
`∈S

if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition

Initial centralized problem:


!
Min x r r
X X
T
L` ( ` , ` ) + Γ · `
{r` ,x` }`∈S
`∈S `∈S
Subj. to: hX(x , r ) ≤ 0
` ` ` ∀` ∈ S
r ≤ Rlim
`
`∈S

if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


x r
L ( ` , ` , λ)
! !
Max Min x r r r − Rlim
 X X X
T
L` ( ` , ` ) + Γ · ` + λT · `
λ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S

if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise
λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


! !
Max Min x r r r − Rlim
 X X X
T T
L` ( ` , ` ) + Γ · ` +λ · `
λ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S

if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise
λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


! !
Max Min x r r r − Rlim
 X X X
T T
L` ( ` , ` ) + Γ · ` +λ · `
λ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S

NOT decomposable yet


if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise
λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


! !
Max Min x r r r − Rlim
 X X X
T T
L` ( ` , ` ) + Γ · ` +λ · `
λ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S

NOT decomposable yet


if `∈S r` ≥ 0
(
Γ+
P
Γ=
Γ− otherwise Step 3: Introduce
additional variable µ
λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


! !
Max Min x r r r − Rlim
 X X X
+ T T
L` ( ` , ` ) + (Γ − µ) · ` +λ · `
λ,µ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ + −

λ≥0

Step 3: Introduce
additional variable µ

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Step 1: Build Lagrangian L (x` , r` , λ)

Step 2: Dual problem


! !
Max Min x r r r − Rlim
 X X X
+ T T
L` ( ` , ` ) + (Γ − µ) · ` +λ · `
λ,µ {r` ,x` }`∈S
`∈S `∈S `∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ + −

λ≥0 DECOMPOSABLE
Step 3: Introduce
additional variable µ

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition

Step 2: Dual problem

Max Min x r r R
 X  
L` ( ` , ` ) + (Γ+ + λ − µ)T · ` − λT · lim
λ,µ {r` ,x` }`∈S
`∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ + −

λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Sub-problem ` (for given λ, µ)
J` (λ, µ) :=Min L (x , r ) + (Γ
r` ,x`
` ` `
+
+ λ − µ)T · r
`

Step 2: Dual problem Subj. to: h (x , r ) ≤ 0


` ` `

Max Min x r r R
 X  
L` ( ` , ` ) + (Γ+ + λ − µ)T · ` − λT · lim
λ,µ {r` ,x` }`∈S
`∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ + −

λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Sub-problem ` (for given λ, µ)
J` (λ, µ) :=Min L (x , r ) + (Γ
r` ,x`
` ` `
+
+ λ − µ)T · r`

Step 2: Dual problem Subj. to: h (x , r ) ≤ 0


` ` `

Max Min x r r R
 X  
L` ( ` , ` ) + (Γ+ + λ − µ)T · ` − λT · lim
λ,µ {r` ,x` }`∈S
`∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ
Coordinator problem
+ −

λ≥0
Max J (λ, µ) − λ · Rlim
 X 
T
`
λ,µ
`∈S

Subj. to: 0 ≤ µ ≤ Γ+ − Γ−
λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition
Sub-problem ` (for given λ, µ)
J` (λ, µ) :=Min L (x , r ) + (Γ
r` ,x`
` ` `
+
+ λ − µ)T · r`

Step 2: Dual problem Subj. to: h (x , r ) ≤ 0


` ` `

Max Min x r r R
 X  
L` ( ` , ` ) + (Γ+ + λ − µ)T · ` − λT · lim
λ,µ {r` ,x` }`∈S
`∈S
Subj. to: h (x , r ) ≤ 0
` ` ` ∀` ∈ S
0≤µ≤Γ −Γ
Coordinator problem
+ −

λ≥0
Max J (λ, µ) − λ · Rlim
 X 
T
`
λ,µ
`∈S
Diculty: J (λ, µ) can only
` Subj. to: 0 ≤ µ ≤ Γ+ − Γ−
be evaluated point-wise ! λ≥0

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 20 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Recall...

20
powers [kW]

0
λ, µ act like modications
of the energy tari !
PG PG PB
−20 PBat PPV

0 2 4 6 8 10 12 14 16 18 20 22 24
temperature [◦ C] state-of-charge

1
0.8
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20 22 24

30
20
10 ϑindoor comfort bounds

0
0 2 4 6 8 10 12 14 16 18 20 22 24
0.2
0.15
tari

0.1
0.05 tari Γ
0 virtual tari Γ+λ−µ
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 21 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Bundle method to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Primal coordinator problem:
 

Minimize
{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method to solve the coordinator problem

Iteratively
nknown at build cutting
the coordination layer plane approximations of the sub-problems'
cost functions
Jℓ
 

? Minimize
{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
rℓ

er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Minimize
{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S
g`
Function value
Sensitivity
Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(0)
rℓ
rℓ
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Jˇℓ Minimize


{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(0)
rℓ
rℓ
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Jˇℓ Minimize


{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(1) (0)
rℓ
rℓ rℓ
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Jˇℓ Minimize


{r` }`∈S
X
r T
J` ( ` ) + Γ · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(1) (2) (0)
rℓ
rℓ rℓ rℓ
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Jˇℓ Minimize


{r` }`∈S
X
r
Jˇ` ( ` ) + ΓT · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(1) (2) (3) (0)
rℓ
rℓ rℓ rℓ rℓ
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
dle method
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation
approximation

Bundle method
Jℓ gℓ to solve the coordinator problem

Iteratively build cutting plane approximations of the sub-problems'


cost functions
Jℓ
 

epi(Jℓ ) Jˇℓ Minimize


{r` }`∈S
X
r
Jˇ` ( ` ) + ΓT · 
X
r `
`∈S `∈S

Subj. to:
X
r ≤ Rlim
`
`∈S

r
(
Γ+
P
if `∈S ` ≥0
Γ=
Γ− otherwise
(1) (2) (3) (0)
rℓ
rℓ rℓ rℓ rℓ

→ can be solved as a LP
er-Electric/Gipsa-lab - DMPC for Energy management in buildings - 11/29/2012 37 / 52

Source of illustration: DMPC for energy management in buildings,


Small number of iterations due to memory in receding horizon mode
Mohamed-Yacine Lamoudi, 2012

Antonio Frangioni (2002).  Generalized bundle methods. In: SIAM Journal on


Optimization 13.1, pp. 117156
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 22 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Solving the coordinator problem


Nctr number of constraints
NS number of sub-systems
nB number of cuts per
cutting plane model

each cut results in one constraint in the LP: Nctr = nB · NS

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 23 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Solving the coordinator problem


Nctr number of constraints
NS number of sub-systems
nB number of cuts per
cutting plane model

each cut results in one constraint in the LP: Nctr = nB · NS


Problem: Computation time increases with NS

100
tCoordinator [s]

10−1

10−2

10−3
0 100 200 300 400 500 600 700 800 900 1,000
nb. of constraints in dual coordinator LP (Nctr = nB · NS )

Inter(R) Core(TM) i7-3540M CPU @ 3.00 GHz, 16,0 Go RAM, Gurobi 6.0 for LPs and QPs

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 23 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Disaggregated vs. Aggregated Bundle


Precise, but high
Disaggregated bundle method Nctr = NS · nB

Coordinator problem
Coordinator

MPC MPC MPC MPC

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 24 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Disaggregated vs. Aggregated Bundle


Less precise, but
Fully aggregated bundle method Nctr = nB

Coordinator problem
Coordinator

MPC MPC MPC MPC

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 24 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Disaggregated vs. Aggregated Bundle


Compromise between
Semi-aggregated bundle method the other two

Coordinator problem
Coordinator

MPC MPC MPC MPC

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 24 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Optimal trade-o for the bundle aggregation

Trade-o between niter and


tCoordinator !

Computation time: ttotal = niter · (tCoordinator + tcomm + tSubsys )

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 25 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Optimal trade-o for the bundle aggregation

Trade-o between niter and


tCoordinator !

Computation time: ttotal = niter · (tCoordinator + tcomm + tSubsys )


t
ttotal = t1 + t2
t2 = niter × tCoordinator

t1 = niter × (tSubsys + tcomm )

Nctr ,opt Nctr

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 25 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Optimal bundle aggregation level


niter with primal decomposition (NS = 500 sub-systems)

fully aggregated (NC = 1)


2 bundles (NC = 2)
106 5 bundles
objective value

10 bundles
25 bundles
105 50 bundles
100 bundles
250 bundles
104
disaggregated (NC = 500)

103
0 100 200 300 400 500 600 700 800
iterations

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 26 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Optimal bundle aggregation level


niter with primal decomposition (NS = 500 sub-systems)

fully aggregated (NC = 1)


2 bundles (NC = 2)
106 5 bundles
objective value

10 bundles
25 bundles
105 50 bundles
100 bundles
250 bundles
104
disaggregated (NC = 500)

103
0 100 200 300 400 500 600 700 800
iterations

Optimal aggregation level: NC = 25


ttotal = niter · (tCoordinator + tcomm + tSubsys ) = 285 · (1.77 + 0 + 0.1) = 533sec

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 26 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Optimal bundle aggregation level


niter with primal decomposition (NS = 500 sub-systems)

fully aggregated (NC = 1)


2 bundles (NC = 2)
106 5 bundles
objective value

10 bundles
25 bundles
105 50 bundles
100 bundles
250 bundles
104
disaggregated (NC = 500)

103
0 100 200 300 400 500 600 700 800
iterations

Optimal aggregation level: NC = 25


ttotal = niter · (tCoordinator + tcomm + tSubsys ) = 285 · (1.77 + 0 + 0.1) = 533sec

High coordinator computation time with primal decomposition when


the number of sub-systems NS is high
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 26 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Dual decomposition  1000 buildings

District composed of 1000 buildings of dierent size and inertia


Prediction horizon thorizon = 24h, sampling period τ = 15min
Number of decision variables ≈ 800000
Time to solve one building MPC problem: tSubsys = 75ms
tCoordinator = 79ms (aggregated bundle)
Around 30 iterations required in closed loop
Building MPC problems can be solved in parallel

Total computation time to solve the problem:


ttotal = niter · (tSubsys + tCoordinator ) = 30 · (75[ms] + 79[ms]) = 4.6[s]
(Inter(R) Core(TM) i7-3540M CPU @ 3.00 GHz, 16,0 Go RAM, Gurobi 6.0 for LPs and QPs)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 27 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Real-world application

Implementation of the primal- and dual frameworks on two pilot sites:


LTCP (Lavrion, Greece) INES (Chambéry, France)

• 5 buildings • 4 buildings
• 1 H2 storage system • 2 batteries
• 1 battery • 1 EV charging station
• 1 PV installation ' • 1 PV installation
• Wind turbines • 1 PV installation
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 28 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Real-world issues & cost savings

Real-world issues
Data acquisition
Interaction with low-level control systems
Forecast errors/Availability of forecasts

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 29 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Real-world issues & cost savings

Real-world issues
Data acquisition
Interaction with low-level control systems
Forecast errors/Availability of forecasts

Cost savings
Annual cost savings (obtained in simulation): 0% - 30%
Highly context dependent (sub-system types, tari, climate,...)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 29 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Model MPC Problem Decomposition Bundle Validation

Conclusions -DMPC

Summary
Primal- and dual decomposition frameworks for problems with
limited shared resources
buying- and selling taris on the globally consumed resource
Ecient implementation of the bundle method
Comparison of the two approaches
Better scalability with dual decomposition
Primal is less restrictive (convexity required, but not strong convexity)
Feasibility & eciency shown in simulations and real-life experiments
Future works
Asynchronous information exchange
Benet guarantee for each actor

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Outline

1 Distributed MPC for smart grids

2 Randomized algorithm for EV charging stations


Motivation
Problem statement
The approach
Validation through simulation
Conclusion

3 General conclusion

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 30 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation

Model Predictive Control performs poorly when uncertainties are high


This has been put into evidence at the Euref Campus in Berlin

PG ≤ Plim

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 31 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Motivation
Uncertain power consumption of the EVCS consisting of 10 charging points
(10 scenarios)
18
mean power prole PEVCS
16

14

12
power EVCS [kW]

10

0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 32 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

EVCS control under uncertainties

Context
EVCS with M charging points
No forecasts of the individual EV behavior available
But known statistic model of the EV behavior
EVCS occupancy scenario
S = {tarr,v , tdep,v , Ereq,v } v = 1, ..., nEV
The scenario S is the set of EVs which arrive during a given day

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 33 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Objective
Design a controller for the EVCS that
respects an a priori declared power limitation prole p
guarantees the QoS (Quality of Service) at a high probability
(≥ (1 − η) )
randomized direct charging
50 50
power limit p
pEVCS (scenario 1)
40 pEVCS (scenario 2) 40
pEVCS (scenario 3)
pEVCS [kW]

30 30

20 20

10 10

0 0
0 10 20 30 40 0 10 20 30 40
time [h] time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 34 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Objective
Design a controller for the EVCS that
respects an a priori declared power limitation prole p
guarantees the QoS (Quality of Service) at a high probability
(≥ (1 − η) )
randomized direct charging
50 50
power limit p
pEVCS (scenario 1)
40 pEVCS (scenario 2) 40
pEVCS (scenario 3)
pEVCS [kW]

30 30

20 20

10 10

0 0
0 10 20 30 40 0 10 20 30 40
time [h] time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 34 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Quality of Service 8
EVs
QoS limit
6

parking time [hour]


EV satisfaction metric: 4
 
Echarged tpark
≥ csatis × min 1, 2
Ereq Tth

0
0 20 40 60 80 100
Echarged [% of Ereq ]

QoS indicator g
0 if all EVs in a scenario S are satised
(
g (S) :=
1 otherwise

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 35 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Relaxation of the QoS constraint g (S) QoS indicator


(0 := satised)
S uncertainty set

Robust constraint: g (S) = 0 ∀S ∈ S

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 36 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Relaxation of the QoS constraint g (S) QoS indicator


(0 := satised)
S uncertainty set

Robust constraint: g (S) = 0 ∀S ∈ S


overly conservative, hard to solve

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 36 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Relaxation of the QoS constraint g (S) QoS indicator


(0 := satised)
S uncertainty set

Robust constraint: g (S) = 0 ∀S ∈ S


overly conservative, hard to solve A
First relaxation: Pr (g (S) = 1) ≤ η (η := precision)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 36 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Relaxation of the QoS constraint g (S) QoS indicator


(0 := satised)
S uncertainty set

Robust constraint: g (S) = 0 ∀S ∈ S


overly conservative, hard to solve A
First relaxation: Pr (g (S) = 1) ≤ η (η := precision)
Second relaxation: Pr (!A) ≤ δ (δ := condence)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 36 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Relaxation of the QoS constraint g (S) QoS indicator


(0 := satised)
S uncertainty set

Robust constraint: g (S) = 0 ∀S ∈ S


overly conservative, hard to solve A
First relaxation: Pr (g (S) = 1) ≤ η (η := precision)
Second relaxation: Pr (!A) ≤ δ (δ := condence)

Randomized algorithm
If for no more than m out of N scenarios g (S) = 1, then the QoS is
guaranteed with probability > (1 − η) and condence > (1 − δ).

Teodoro Alamo et al. (2015).  Randomized methods for design of uncertain systems:
Sample complexity and sequential algorithms. In: Automatica 52, pp. 160172
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 36 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator

Simulate RT-
controller &
evaluate QoS

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator

Generate
occupancy
scenarios

Simulate RT-
controller &
evaluate QoS

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator
IQoS feasibility indicator

Generate
occupancy
scenarios

Simulate RT-
controller &
evaluate QoS

QoS
guaranteed?

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator
IQoS feasibility indicator
θ design param. vector

Generate
occupancy
scenarios

Parameter- Simulate RT-


ization controller &
evaluate QoS

QoS
guaranteed?

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator
IQoS feasibility indicator
θ design param. vector

Generate
occupancy
scenarios
Define a finite
set of design Parameter- Simulate RT-
parameter ization controller &
vectors evaluate QoS

QoS
guaranteed?

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator
IQoS feasibility indicator
θ design param. vector

Generate N
occupancy
scenarios
Define a finite
set of design Parameter- Simulate RT-
parameter ization controller &
vectors evaluate QoS

Determine
QoS
optimal design
guaranteed?
parameter

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

p
S occupancy scenario
Outline of the approach power limit prole
g QoS indicator
IQoS feasibility indicator
θ design param. vector

Generate N
occupancy
scenarios
Define a finite
set of design Parameter- Simulate RT-
parameter ization controller &
vectors evaluate QoS

Determine
QoS
optimal design
guaranteed?
parameter

30
p(θ(k) )
20
10 Min p(θ
Z
θ(opt) := (k)
)dt
1
k∈{ ...nΘ }
0
0 10 20 30 40
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 37 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Parameterization of the power limit prole p(θ)

Low-dimensional design parameter vector θ := (θ1 , ..., θ4 )T


Allow suciently rich variations in the prole p(θ)
(e.g. two superposed Gaussian distributions)

40 p(θ)
power limit p(θ)

20 θ4
θ3
θ1 θ2

0
0 5 10 15 20 25
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 38 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Randomized algorithm
η precision
δ condence
m positive integer
nΘ card(Θ)

What is the necessary number of scenarios N ?

Teodoro Alamo et al. (2015).  Randomized methods for design of uncertain systems:
Sample complexity and sequential algorithms. In: Automatica 52, pp. 160172
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 39 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Randomized algorithm
η precision
δ condence
m positive integer
nΘ card(Θ)

What is the necessary number of scenarios N ?

1
 
e nΘ 
N≥ ln +m
η e −1 δ

Teodoro Alamo et al. (2015).  Randomized methods for design of uncertain systems:
Sample complexity and sequential algorithms. In: Automatica 52, pp. 160172
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 39 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Statistic model - Euref campus


Euref Campus

E-car sharing station located at the Euref Campus in Berlin

800 charging events


Available information for each event: tarr , tpark , Ereq
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 40 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Statistic model - Euref campus

Fit non-parametric probability distributions to tarr , tpark , Ereq


100
NEV (real data)

80
60
40
20
0
0 2 4 6 8 10 12 14 16 18 20 22 24
arrival time tarr [h]
100
NEV (tted model)

80
60
40
20
0
0 2 4 6 8 10 12 14 16 18 20 22 24
arrival time tarr [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 41 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Statistic model - Euref campus

Fit non-parametric probability distributions to tarr , tpark , Ereq


100
NEV (real data)

80
60
40
20
0
0 2 4 6 8 10 12 14 16 18 20 22 24
arrival time tarr [h]
100
NEV (tted model)

80
60
40
20
0
0 2 4 6 8 10 12 14 16 18 20 22 24
arrival time tarr [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 41 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Statistic model - Euref campus

Fit non-parametric probability distributions to tarr , tpark , Ereq


150
NEV (real data)

100

50

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
150
NEV (tted model)

100

50

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
parking time tpark [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 41 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Statistic model - Euref campus

Fit non-parametric probability distributions to tarr , tpark , Ereq


150
NEV (real data)

100

50

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
150
NEV (tted model)

100

50

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
required energy Ereq [kWh]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 41 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Algorithm conguration

Algorithm parameters
η 0.05 (accuracy)
δ 0.05 (condence)
nΘ 432 (design parameters)
m 20 (unsatisfying scenarios)
NCP 50 (charging points)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 42 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Algorithm conguration

Algorithm parameters Required number of scenarios N


η 0.05 (accuracy)
δ 0.05 (condence) 1

e

nΘ 
nΘ 432 (design parameters) N ≥ ln + m = 920
η e −1 δ
m 20 (unsatisfying scenarios)
NCP 50 (charging points)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 42 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Algorithm conguration

Algorithm parameters Required number of scenarios N


η 0.05 (accuracy)
δ 0.05 (condence) 1

e

nΘ 
nΘ 432 (design parameters) N ≥ ln + m = 920
η e −1 δ
m 20 (unsatisfying scenarios)
NCP 50 (charging points)

Computation time using brute-force:


tcomp = titeration · nΘ = 5.8sec · 432 = 42min

(titeration := time to simulate the real-time controller for N scenarios)


(Inter(R) Core(TM) i7-3540M CPU @ 3.00 GHz, 16,0 Go RAM)

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 42 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Validation through simulation (Euref Campus)

Comparison with a direct charging strategy


randomized direct charging
50 50
power limit p
pEVCS (scenario 1)
40 pEVCS (scenario 2) 40
pEVCS (scenario 3)
pEVCS [kW]

30 30

20 20

10 10

0 0
0 10 20 30 40 0 10 20 30 40
time [h] time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 43 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Validation through simulation (Euref Campus)

Mean over-estimated energy need

40
power limit p
p(i)
EVCS (scenario i)

30
pEVCS [kW]

20
(i)
E
10

0
0 10 20 30 40
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 43 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Validation through simulation (Euref Campus)

Mean over-estimated energy need

40 1X
N
(i)
power limit p E= E
p(i)
EVCS (scenario i)
N
i=1
30
For a single scenario S (i) :
pEVCS [kW]

20
Z tend
(i)
p(t) − p(i)
(i)
E E = EVCS (t)dt
10 t=0

0
0 10 20 30 40
time [h]

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 43 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Validation through simulation (Euref Campus)

Mean over-estimated energy need

Sensitivity of E w.r.t. NCP and η : 1X


N
(i)
E= E
N
η i=1
0.05 0.1 0.15 0.2
25 58.4 56.3 54.2 51.7 For a single scenario S (i) :
NCP 50 37.5 34.1 32.9 31.2
100 30.6 28.8 27.1 25.1 Z tend
p(t) − p(i)
(i)
E = EVCS (t)dt
t=0
η probability of QoS violation
NCP number of charging points

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 43 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Validation through simulation (Company site)


η = 0.05
Comparison with a direct charging strategy NCP = 20
randomized E = 12.5%
200
150
power [kW]

predicted limit
100 scenario 1
scenario 2
50 scenario 3
0
0 2 4 6 8 10 12 14 16 18 20 22 24
direct charging
200
scenario 1
150 scenario 2
power [kW]

scenario 3
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24
time [h]
P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 44 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Conclusion

Summary
Energy management strategy for an EVCS allowing to
respect an upper bound on the power consumption
guarantee the QoS with a congurable probability
Validation on statistic data from the Euref Campus in Berlin
Future extensions
Additional battery to further tighten the upper bound prole
Also compute a guaranteed lower bound prole

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 45 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Conclusion

Summary
Energy management strategy for an EVCS allowing to
respect an upper bound on the power consumption
guarantee the QoS with a congurable probability
Validation on statistic data from the Euref Campus in Berlin
Future extensions
Additional battery to further tighten the upper bound prole
Also compute a guaranteed lower bound prole

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 45 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Conclusion

Summary
Energy management strategy for an EVCS allowing to
respect an upper bound on the power consumption
guarantee the QoS with a congurable probability
Validation on statistic data from the Euref Campus in Berlin
Future extensions
Additional battery to further tighten the upper bound prole
Also compute a guaranteed lower bound prole

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 45 / 47
Introduction DMPC Randomized Conclusion Motivation Problem Approach Validation Conclusion

Conclusion

Summary
Energy management strategy for an EVCS allowing to
respect an upper bound on the power consumption
guarantee the QoS with a congurable probability
Validation on statistic data from the Euref Campus in Berlin
Future extensions
Additional battery to further tighten the upper bound prole
Also compute a guaranteed lower bound prole

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 45 / 47
Outline

1 Distributed MPC for smart grids

2 Randomized algorithm for EV charging stations

3 General conclusion
Introduction DMPC Randomized Conclusion

General conclusion

Summary

1 Distributed MPC design for Large-scale


systems
Renewable Building
Production
Residential

Primal and dual decomposition Grid

Energy eciency in smart grids


Real-time implementation Electrical
Vehicles Storage

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 46 / 47
Introduction DMPC Randomized Conclusion

General conclusion

Summary

1 Distributed MPC design for Large-scale


systems
Renewable Building
Production
Residential

Primal and dual decomposition Grid

Energy eciency in smart grids


Real-time implementation Electrical
Vehicles Storage

2 Randomized algorithms for EV charging


stations
Explicitly deal with uncertainty
Upper bound power prole
Probabilistic QoS guarantee
Validation on real data

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 46 / 47
Introduction DMPC Randomized Conclusion

General conclusion

Future works...
Optimization framework integrating a exibility market model
MPC prototype at the Learning Grid by Grenoble
Integration of MPC into micro grid design tools

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 47 / 47
Introduction DMPC Randomized Conclusion

Thank you for your attention!

Questions ?

P. Paum  Schneider Electric/Gipsa-lab  Energy management strategies for smart grids  09/01/2017 47 / 47

Das könnte Ihnen auch gefallen