subway stations energy and air quality management
TRANSCRIPT
![Page 1: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/1.jpg)
Subway stations energy and air quality management
using discrete time stochastic optimal control
Tristan Rigaut1,2,4,Advisors: F. Bourquin1,4, P. Carpentier3, J.-Ph. Chancelier2,
M. De Lara1,2, J. Waeytens1,4
EFFICACITY1
CERMICS, ENPC2
UMA, ENSTA3
LISIS, IFSTTAR4
June 21, 2016
Stations Optimal Management June 21, 2016 1 / 26
![Page 2: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/2.jpg)
Optimization for subway stations
Paris urban railway transport system energy consumption =1
3subway stations + 2
3traction system
Subway stations present a signi�cantly high particulate
matters concentration
We use optimization to harvest unexploited energy
ressources and improve air quality.
Stations Optimal Management June 21, 2016 2 / 26
![Page 3: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/3.jpg)
Outline
1 Subway stations optimal management problemEnergyAir qualityEnergy/Air management system
2 Two methods to solve the problemWe are looking for a policyDynamic programming in the non Markovian caseModel Predictive Control
3 Numerical resultsRandom variables modelingMethodsResults
Stations Optimal Management June 21, 2016 3 / 26
![Page 4: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/4.jpg)
Outline
1 Subway stations optimal management problemEnergyAir qualityEnergy/Air management system
2 Two methods to solve the problemWe are looking for a policyDynamic programming in the non Markovian caseModel Predictive Control
3 Numerical resultsRandom variables modelingMethodsResults
Stations Optimal Management June 21, 2016 3 / 26
![Page 5: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/5.jpg)
Energy
Stations Optimal Management June 21, 2016 3 / 26
![Page 6: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/6.jpg)
Subway stations typical energy consumption
Stations Optimal Management June 21, 2016 4 / 26
![Page 7: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/7.jpg)
Subway stations have unexploited energy ressources
Stations Optimal Management June 21, 2016 5 / 26
![Page 8: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/8.jpg)
Energy recovery requires a bu�er
Stations Optimal Management June 21, 2016 6 / 26
![Page 9: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/9.jpg)
Air quality
Stations Optimal Management June 21, 2016 6 / 26
![Page 10: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/10.jpg)
Subways arrivals generate particulate mattersRails/brakes wear and resuspension increase PM10 concentration
Recovering energy improves air qualityStations Optimal Management June 21, 2016 7 / 26
![Page 11: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/11.jpg)
Energy/Air managementsystem
Stations Optimal Management June 21, 2016 7 / 26
![Page 12: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/12.jpg)
Subway station microgrid concept
Stations Optimal Management June 21, 2016 8 / 26
![Page 13: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/13.jpg)
Objective: We want to minimize energy consumption and
particles concentration
A parameter λ measures the relative weights of the 2 objectives:
T∑t=0
Costt (ESupplyStationt + E
SupplyBattery t
)︸ ︷︷ ︸Grid supply
+λ C InP t︸︷︷︸
PM10
Stations Optimal Management June 21, 2016 9 / 26
![Page 14: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/14.jpg)
We control the battery
Stations Optimal Management June 21, 2016 10 / 26
![Page 15: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/15.jpg)
We have many uncertainties
Let Wt the random variables vector of uncertainties at time t:
Outdoor particles concentration : COutP t
Regenerative braking : EAvailableTrain t
Station consumption : EDemandStation t
Cost of electricity : Costt
Particles generation : QPt
Resuspension rate : ρRt
Deposition rate : ρDt
Stations Optimal Management June 21, 2016 11 / 26
![Page 16: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/16.jpg)
We set a stochastic optimal control problem
minU∈U
E( T∑
t=0
Costt(ESupplyStationt + E
SupplyBattery t
) + λC InP t
) }Objective
s.t
Soct+1 = Soct −1
ρdcUB−
t + ρc(UB+t + E
SurplusTrain t)
}Battery dynamics
C InP t+1
= C InP t +
dVentilt
V(COut
P t − CInP t)
+ρRt
SCFloorP t −
ρDt
VC InP t +
QPt
V
CFloorP t+1
= CFloorP t +
ρDt
SC InP t −
ρRt
VCFloorP t
Particles dynamics
UB+t + E
SurplusTrain t = E
SupplyBattery t
+ ERecoveredTrain t
EDemandStation t = E
SupplyStationt +UB−
t
}Supply/demand balance
SocMin ≤ Soct ≤ SocMax
C InP t ≥ 0
}Constraints
Stations Optimal Management June 21, 2016 12 / 26
![Page 17: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/17.jpg)
Summary of the equations
State of the system: Xt =
Soct
C InP t
CFloorP t
Controls: Ut =
UB−t
UB+t
dVentilt
,
And the dynamics:
Xt+1 = ft(Xt ,Ut ,Wt+1)
We add the non-anticipativity constraints:
Ut � σ(W1, ...,Wt)
Stations Optimal Management June 21, 2016 13 / 26
![Page 18: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/18.jpg)
Compact stochastic optimal control problem
We obtained a stochastic optimization problem consistent with the generalform of a time additive cost stochastic optimal control problem:
minX ,U
E
(T−1∑t=0
Lt(Xt ,Ut ,Wt+1) + K (XT )
)
s.t. Xt+1 = ft(Xt ,Ut ,Wt+1)Ut � σ(X0,W1, ...,Wt)
Ut ∈ Ut
Stations Optimal Management June 21, 2016 14 / 26
![Page 19: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/19.jpg)
Outline
1 Subway stations optimal management problemEnergyAir qualityEnergy/Air management system
2 Two methods to solve the problemWe are looking for a policyDynamic programming in the non Markovian caseModel Predictive Control
3 Numerical resultsRandom variables modelingMethodsResults
Stations Optimal Management June 21, 2016 14 / 26
![Page 20: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/20.jpg)
We are looking for a policy
Stations Optimal Management June 21, 2016 14 / 26
![Page 21: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/21.jpg)
What is a solution?In the general case an optimal solution is a function of past uncertainties:
Ut � σ(X0,W1, ...,Wt)⇒ Ut = πt(X0,W1, ...,Wt)
This is an history-dependent policy
In the Markovian case (noises time independence) it is enough to restrictthe search to state feedbacks:
Ut = πt(Xt)
In the Markovian case we can introduce value functions:
∀x ∈ Xt , Vt(x) =minπ
E( T−1∑
t′=t
Lt′(Xt′ , πt′(Xt′),Wt′+1) + K (XT ))
s.t Xt = x and dynamics
and use Bellman equationStations Optimal Management June 21, 2016 15 / 26
![Page 22: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/22.jpg)
Dynamic programming inthe non Markovian case
Stations Optimal Management June 21, 2016 15 / 26
![Page 23: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/23.jpg)
Dynamic programming in the general case
Bellman equation does not hold in the non Markovian case.Let P be the probability s.t (Wt)t∈[|1,T |] are time independent but keep thesame marginal laws.
Algorithm
O�ine: We produce value functions with Bellman equation using thisprobability measure:
Vt(x) = minu∈Ut
EPt
(Lt(x , u,Wt+1) + Vt+1(ft(x , u,Wt+1))
)Online: We plug the computed value functions as future costs at time t:
ut ∈ argminu∈Ut
E˜Pt
(Lt(xt , u,Wt+1) + Vt+1(ft(xt , u,Wt+1))
)
Stations Optimal Management June 21, 2016 16 / 26
![Page 24: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/24.jpg)
We produce history-dependent controls
With ˜Pt the probability updating Wt+1 marginal law taking into accountall the past informations: ∀i ≤ t, Wi = wi .
If the (Wt)t∈1..T+1 are independent the controls are optimal and ˜Pt = Pt
Stochastic Dynamic Programming su�ers the well known "curse ofdimensionality".
Stations Optimal Management June 21, 2016 17 / 26
![Page 25: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/25.jpg)
Model Predictive Control
Stations Optimal Management June 21, 2016 17 / 26
![Page 26: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/26.jpg)
Rollout algorithms
To avoid value functions computation we can plug a lookahead future costfor a given policy:
ut ∈ argminu∈Ut
Et
(Lt(xt , u,Wt+1) + Jπ
t
t+1(ft(xt , u,Wt+1)))
It gives the cost of controlling the system in the future according to thegiven policy:
∀x ∈ Xt+1, Jπt
t+1(x) =Et
( T−1∑t′=t+1
Lt′(Xt′ , πt′(Xt′),Wt′+1) + K (XT ))
s.t Xt+1 = x , and the dynamics
Stations Optimal Management June 21, 2016 18 / 26
![Page 27: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/27.jpg)
Model Predictive ControlChoosing πt in the class of open loop policies minimizing the expectedfuture cost:
∀i ≥ t + 1, ∃ui ∈ Rn, ∀x , πti (x) = ui
ut ∈ argminu∈Ut
min(ut+1,...,uT−1)
Et
(Lt(xt , u,Wt+1) +
T−1∑t′=t+1
Lt′(Xt′ , ut′ ,Wt′+1))
With Et replacing noises by forecasts, we obtain a deterministic problem.
Algorithm
Online: At every MPC step t, compute a forecast (wt+1, ..., wT+1) usingthe observations ∀i ≤ t, Wi = wi . Then compute control ut :
ut ∈ argminu∈Ut
min(ut+1,...,uT−1)
Lt(xt , u, wt+1) +T−1∑
t′=t+1
Lt′(xt′ , ut′ , wt′+1)
MPC is often de�ned with a rolling horizon.Stations Optimal Management June 21, 2016 19 / 26
![Page 28: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/28.jpg)
Outline
1 Subway stations optimal management problemEnergyAir qualityEnergy/Air management system
2 Two methods to solve the problemWe are looking for a policyDynamic programming in the non Markovian caseModel Predictive Control
3 Numerical resultsRandom variables modelingMethodsResults
Stations Optimal Management June 21, 2016 19 / 26
![Page 29: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/29.jpg)
Random variables modeling
Stations Optimal Management June 21, 2016 19 / 26
![Page 30: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/30.jpg)
Some random variables are taken deterministic
Outdoor particles concentration : COutP t
Regenerative braking : EAvailableTrain t
Station consumption : EDemandStation t
Cost of electricity : Costt
Particles generation : QPt
Resuspension rate : ρRt
Deposition rate : ρDt
Stations Optimal Management June 21, 2016 20 / 26
![Page 31: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/31.jpg)
Stochastic models
We have multiple equiprobable scenarios:
Braking energy and outside PM10 concentration every 5s
We deduce the discrete marginal laws from these scenarios.
Stations Optimal Management June 21, 2016 21 / 26
![Page 32: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/32.jpg)
Details on the methods
Stochastic Dynamic Programming:
We compute value functions every 5s. We can compute a control every 5s.The algorithm is coded in Julia.
Model Predictive Control:
The deterministic problem is linearized leading to a MILP. It is solved every15 min with a 2 hours horizon. We use two forecasts strategies:
MPC1: Expectation of each noise ignoring the noises dependence
MPC2: Scenarios where the next outside PM10 concentration is nottoo far from the previous one
Stations Optimal Management June 21, 2016 22 / 26
![Page 33: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/33.jpg)
Results
Stations Optimal Management June 21, 2016 22 / 26
![Page 34: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/34.jpg)
Air quality comparaisonReference case:
Optimized with SDP:
Stations Optimal Management June 21, 2016 23 / 26
![Page 35: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/35.jpg)
Battery control over a scenario
Result produced using SDP with a regular day
Stations Optimal Management June 21, 2016 24 / 26
![Page 36: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/36.jpg)
Energetic results
Assessor: 50 scenarios of 24h with time step = 5 secReference: Energy consumption cost over a day without battery andventilation control
MPC1 MPC2 SDP
O�ine comp. time 0 0 12hOnline comp. time [10s,200s] [10s,200s] [0s,1s]Av. economic savings -26.2% -27.4% -30.7%
Stations Optimal Management June 21, 2016 25 / 26
![Page 37: Subway stations energy and air quality management](https://reader034.vdocuments.site/reader034/viewer/2022042614/6263c90dd86b574dc35af89e/html5/thumbnails/37.jpg)
Conclusion & Ongoing work
Our study leads to the following conclusions:
A battery and a proper ventilation control provide signi�cant economicsavings
SDP provides slightly better results than MPC but requires moreo�ine computation time
We are now focusing on:
Using other methods that handle more state/control variables (SDDP,RL...)
Taking into account more uncertainty sources
Calibrating air quality models for a more realistic concentrationdynamics behavior
Ultimate goal: apply our methods to laboratory and real size
demonstrators
Stations Optimal Management June 21, 2016 26 / 26