dynamic emulation modelling for the optimal operation of water systems: an overview
TRANSCRIPT
Andrea Castelletti1, Stefano Galelli2, Matteo Giuliani1
1 Dipartimento di Elettronica, Informazione, e Bioingegneria, Politecnico di Milano, Milano, Italy 2 Pillar of Engineering Systems and Design, Singapore University of Technology and Design, Singapore
Improving Computational Efficiency in Modeling Complex Environmental Systems
Dynamic emulation modelling for the optimal operation of water systems: an overview
Pollock n. 31
H41J-01
Emulation modelling: reconciling science and decision making …
DM High fidelity Accuracy
Credibility Simplicity
DATA DRIVEN PROCESS BASED
DM
Emulation modelling: reconciling science and decision making …
Emulator: A low-order, computationally efficient model identified from an original large high fidelity model and then used to replace it in computationally intensive applications.
DM High fidelity Accuracy
Credibility Simplicity
DATA DRIVEN PROCESS BASED
DM
EMO’s application tree
Model Emulation
MODEL DIAGNOSTIC
Data assimilation
DECISION MAKING
Model identification
Sensitivity analysis
Optimal planning
What-if analysis
Optimal control
[Castelletti et al. 2012a]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
Non-dynamic emulator [Razavi et al. 2013]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
N-DEMo
Non-dynamic emulator [Razavi et al. 2013]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
N-DEMo
Non-dynamic emulator [Razavi et al. 2013]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
Dynamic emulator [Castelletti et al. 2012a]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
DEMo
Dynamic emulator
time
red
uc
ed
sta
te
[Castelletti et al. 2012a]
Non-dynamic vs dynamic emulation
decision 1
de
cisi
on
2
objective 1
ob
jec
tive 2
time
sta
te
ext
driv
ers
time
HIGH FIDELITY MODEL
DEMo
Dynamic emulator
time
red
uc
ed
sta
te
[Castelletti et al. 2012a]
How to build a Dynamic Emulator
Transform the state and input/decision vectors into lower dimension vectors through adoption of a suitable aggregation (PCA, SVD, etc)
2. Variable aggregation
Selection of the most relevant aggregated variables in explaining the model output (IVS, PMI, etc)
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo Design a sequence of simulation runs to construct a high dimension sample data-set.
1. DOE and simulation runs
[Castelletti et al. 2012a]
REAL TIME CONTROL OF MARINA BARRAGE
SINGAPORE
CASE STUDY
Singapore’s 4 national taps strategy
Malaysia
Indonesia
0 100 km
0 10 km
Singapore Strait
Marina Reservoir
(source: URA)
The 4 TAPS [Kristiana et al., 2011] 1. Local catchments water
2. Imported water
3. Reclaimed water (NEWater)
4. Desalinated water
Marina Barrage
Low tide High tide
Seepage
Pumps
Gates
Pipes
Actuators:
§ 7 pumps § 9 weirs § 2 bottom pipes
Water quantity objectives:
§ Water supply § Flood control § Energy usage
Water quality objectives:
§ Maintain low salinity
[Galelli et al. 2014]
SEA SEA
The high fidelity model
2D view 3D view
barrage
cross-section
observation point
m0-1-2-3-4-5-6-7
The DELFT3D-FLOW hydrodynamic model calculates non-steady flow and transport phenomena (i.e., temperature and salinity conditions)
§ 5 states per cell = 5,500 state variables § Real-to-run time ratio 100:1
[Zijl and Twigt 2007]
The real time control framework
DEFLT3D FLOW
DYNAMIC EMULATOR
MODEL PREDICTIVE CONTROL
- STATE - OBJECTIVES
RELEASE DECISIONS
MODEL REDUCTION
- OPTIMAL POLICY HYDROMETEO
DRIVERS
EVALUATION via SIMULATION
-1.71
2.54
12.47
2.06
2.12
6.81
Temp1
log(Lev) Algae
2 54
7
0666
Legend
6.2.45 10 6.3.84 10
7.14.26 107.4.01 10
size: Irr1
orientation: Sed
color: approachNSPCAbased
expertbased
PARETO Front
FAST SIMULATION
- REDUCED STATE - OBJECTIVES
- STREAMFLOW PREDICTIONS
- OBJECTIVES
STREAMFLOW FORECAST
(SOBEK)
RELEASE DECISION
GENERATOR
Building a dynamic emulator of salinity @ the dam
Transform the state and input/decision vectors into lower dimension vectors through adoption of a suitable aggregation.
2. Variable aggregation
Selection of the most relevant aggregated variables in explaining the model output.
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo Design a sequence of simulation runs to construct a high dimension sample data-set.
1. DOE and simulation runs
[Castelletti et al. 2012a]
Building a dynamic emulator of salinity @ the dam
Transform the state and input/decision vectors into lower dimension vectors through adoption of a suitable aggregation.
2. Variable aggregation
Selection of the most relevant aggregated variables in explaining the model output.
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Selection of the most relevant aggregated variables in explaining the model output.
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Variable aggregation
Cluster 1 (Salinity): saline layer.
m0-1-2-3-4-5-6-7
Variable aggregation
m0-1-2-3-4-5-6-7
Cluster 2 (Salinity): saline layer.
Variable aggregation
m0-1-2-3-4-5-6-7
Cluster 3 (Salinity): saline layer.
Variable aggregation
m0-1-2-3-4-5-6-7
Cluster 4 (Salinity): ‘buffer zone’ at a depth of 4 – 5 m, between the saline layer and the freshwater area.
Variable aggregation
m0-1-2-3-4-5-6-7
Cluster 5 (Salinity): upper layers of the reservoir (max depth of about 4 m), with a uniform salinity of 3 ppt.
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Selection of the most relevant aggregated variables in explaining the model output.
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Recursive Variable Selection [Galelli and Castelletti, 2013a] = 32 to 1 state
3. Variable selection
Selection of a family of models for the emulator, calibration, validation, and physical interpretation.
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Recursive Variable Selection [Galelli and Castelletti, 2013a] = 32 to 1 state
3. Variable selection
Extremely Randomized Trees [Galelli and Castelletti, 2013b]
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Variable selection
Cluster 2 salinity
Salinity @ dam
Release from pipes
Groundwater seepage
CONTROL
EXTERNAL DRIVER
STATE OUTPUT
Seepage
Pumps
Gates
Pipes SEA
Emulator calibration and validation
R2 – cross-validation (April – December
2009)
0.989
R2 – validation (January – December
2010)
0.970
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Recursive Variable Selection [Galelli and Castelletti, 2013a] = 32 to 1 state
3. Variable selection
Extremely Randomized Trees [Galelli and Castelletti, 2013b]
4. Emulator calib. & validation
MODEL REDUCTION
The dynamic emulator is used in the task is was designed for.
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Building a dynamic emulator of salinity @ the dam
Hierarchical Agglomerative Time Series Clustering [Magni et al., 2008] = 5,500 to 32 states
2. Variable aggregation
Recursive Variable Selection [Galelli and Castelletti, 2013a] = 32 to 1 state
3. Variable selection
Extremely Randomized Trees [Galelli and Castelletti, 2013b]
4. Emulator calib. & validation
MODEL REDUCTION
Model Predictive Control [Scattolini, 2007]
Model use
DEMo 300,000 samples by pseudo random sampling = 5,500 state variables
1. DOE and simulation runs
[Castelletti et al. 2012a]
Real time control with and without quality target
Period: April 2009 – December 2010
Objective (minimize) Without wq With wq
Water deficit [Mm3/year] 75.04 74.81
Flood control [hours/year] 305.67 302.15
Energy usage [Mm3/year] 17.28 16.19
Min salinity [ppt] 9.95 5.63
Max salinity [ppt] 30.97 29.52
Mean salinity [ppt] 28.41 22.24
Salinity simulated with DELTF3d Flow at the observation point (water column)
Obj.: water quantity Obj.: water quantity
Obj.: water quantity + quality Obj.: water quantity + quality
April – December 2009 January – December 2010
cross-section
observation point
Real time control with and without quality target
Salinity simulated along the cross-section (dry period vs. wet period)
Obj.: water quantity
Distance from the barrage Distance from the barrage
Water intake Water intake
cross-section
observation point
Obj.: water quantity + quality
Real time control with and without quality target
Conclusions
DEMo as a tool to put more science into decision-making, by …
… preserving the accuracy of the original high fidelity model
… providing some explanatory capability and thus credibility
… allowing to solve complex task such as real time control.