Upload
trinhtruc
View
213
Download
0
Embed Size (px)
Citation preview
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010 RENAULT PROPERTY
DEVELOPMENT OF A FAST RUNNING LEAN NOx TRAP MODEL FOR
SIMULATION OF REAL CUSTOMER DRIVE CYCLESD.MAROTEAUX, P.DARCY - F.LORMIER - E.PAUTASSO - W.WANG, S.WAHIDUZZAMAN
2RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
01 BAKGROUND
02 PROCEDURE
03 NEURAL NETWORK MODEL HYBRID NEURAL NETWORK & ARCHITECTURE
04 NEURAL NETWORK MODEL RESULTS
05 CONCLUSION
SUMMARY
3RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
BACKGROUND
01
4RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Renault NOx Trap Technology
� First application : Renault Espace dCi 175 Euro 5
� Sold in Europe since mid 2009
� A technology that will be applied on Euro 6 high-end veh icles.
5RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Design of the NOx Trap system (1/2)
DPF
DPF
DPF
NOx Trap
NOx Trap
NOx Trap
1/ Architecture: NOx Trap position in the exhaust line
6RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
2/ Catalyst characteristics:
� NOx Trap catalyst volume,
� NOx Trap coating, …
3/ Constraints :
� Emissions on NEDC / resp. FTP75 – US06 test cycles
� Compatibitily with real customer use (city, road, highway) and serviceability targets
Design of the NOx Trap system (2/2)
7RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Real Customer use
� Real customer use = various drive cycles
� City (including heavy traffic conditions & slow average speed)
� Road
� Highway
� Mixed
On distances up to several x 10.000 km
� A challenge for simulation experts :
� Even at a real-time speed simulation duration is prohibitive !
� Accuracy and numerical efficiency are often conflic ting requirements
=> A methodology is needed that is capable of simul ating at speeds exceeding 10 3 x Realtime sacrificing minimum accuracy.
8RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
PROCEDURE
02
9RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Procedure (1/2)
=> Neural Network reduction of a detailed model, b ased on previous Gamma Technologies’ experience (with DOC and SCR – cf FISITA 2010 paper).
1/ Creation of the detailed model of a Lean NOx Trap .
2/ Model validation and calibration by LMM, based on Renault experimental data.
3/ Definition of operating ranges of the input para meters and creation of a design of experiments. Training of a set of static neural network to reproduce the kinetic behavior of the catalyst usin g the resulting data .
4/ Training of the NN models to predict the catalys t output conditions given the inlet conditions and the catalyst state (temper ature and NOx coverage)
10RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Procedure (2/2)
� Catalyst states could not be trained by the static Neural Network as they depend on past history. Dynamic Neural Network is r uled out as it cannot be easily made autonomous.
LNT Heat Flux
LNT Wall
Temperature
LNT
LNT NOx
CoverageNN Models
Physical Models
� State variables are deduced from two physical models that rely on outputs from the Neural Network models.
inletoutlet
11RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
NEURAL NETWORK MODEL
HYBRID NEURAL NETWORK & ARCHITECTURE
03
12RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
� The considered detailed model consists in a Lean NO x Trap catalyst operating under lean and rich conditions
Detailed Model Overview
13RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Detailed Model Overview
� The LNT detailed model behavior is characterized by the following 12 reactions :
1. NO + BaO + 1/2 O2 →→→→ BaONO2
2. NO2 + BaO →→→→ BaONO2
3. BaONO2 →→→→ NO2 +BaO
4. C10H22 + 15.5 BaONO2 →→→→ 15.5BaO + 7.75N2 + 11H2O + 10CO2
5. C3H6 + 4.5BaONO2 →→→→ 4.5BaO + 2.25N2 + 3H2O + 3CO2
6. C7H8 + 9BaONO2 →→→→ 9BaO + 4.5N2 + 4H2O + 7CO2
7. 10CO + 5BaONO2 →→→→ 5BaO + 2.5N2 + 10CO2
8. C10H22 + 15.5O2 →→→→ 10CO2 + 11H2O
9. C3H6 + 4.5O2 →→→→ 3CO2 + 3H2O
10.C7H8 + 9O2 →→→→ 7CO2 + 4H2O
11.CH4 + 2O2 →→→→ CO2 + 2H2O
12.CO + 0.5O2 →→→→ CO2
14RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Methodology - Neural Network Architecture
� The proposed neural network architecture is a three -layer feed-forward neural network with 2 hidden layers whose activatio n functions are tan-sigmoid and 1 output layer with linear activation
15RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
LNT NOx Coverage
� The used approach consists in using a NN to predict derivative of coverage and then to integrate it along the simulation. Thus the coverage calculation during transient simulation is equal to:
� This procedure can be improved by using conservatio n principles.
16RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
LNT wall temperature
� The LNT mass is treated as a single mass
� Model is considered to be adiabatic with respect to the heat loss to the environment
� Calculation based on the heat flux of energy enteri ng/leaving the catalyst wall predicted by a NN
17RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Methodology - DoE Setup
� Two DoE models were set up (each 5.000 experiments) : one for lean and one for rich conditions
� Results are used to train Neural Network: 8 input v ariables, 11 predicted quantities
18RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Neural Network model setup
� During the driving cycle transient simulation, the appropriate NN model is activated (switching from lean NN model to rich NN model and vice versa) based on the inlet condition.
19RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
NEURAL NETWORK MODEL RESULTS
04
20RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Neural Network Model Results - Hydrocarbon (inst. and cumul.)
Inlet (measured)
Detailed model results
Neural Network results
CH
4M
ole
Fra
ctio
n
CH
4C
umul
ativ
e M
ole
Fra
ctio
n
C10
H22
Mol
e F
ract
ion
C10
H22
Cum
ulat
ive
Mol
e F
ract
ion
21RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Neural Network Model Results – NO2 and CON
O2
Mol
e F
ract
ion
NO
2C
umul
ativ
e M
ole
Fra
ctio
n
CO
Mol
e F
ract
ion
CO
Cum
ulat
ive
Mol
e F
ract
ion
Inlet (measured)
Detailed model results
Neural Network results
22RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Neural Network Model Results – Temperatures and NOx CoverageLN
T W
all T
empe
ratu
re
Gas
Tem
pera
ture
(LN
T
mid
dle
poin
t)
Gas
outle
ttem
pera
ture
NO
xC
over
age
Inlet temp. (measured)
Detailed model results
Neural Network results
23RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
CONCLUSION
05
24RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Conclusion (1/2)
� Detailed After-treatment models can be reduced to v ery fast running Neural Network model with a more than acceptable lo ss of accuracy.
� Neural Network model was improved using a hybrid me thodology that improved the fidelity of the system model.
� For an overall duration of about 33h of real time. The computational time on a laptop with Intel 2.53 GHz CPU and 3 GB ram ( using a single core) were:
� Detailed model: 10’ 15”
� NN model: 01’ 25”
25RENAULT PROPERTY
RENAULTRESEARCH, ADVANCED STUDIES & MATERIALS ENGINEERING OCTOBER 25, 2010
Conclusion (2/2)
� It was noted that the Neural Network model runs 130 0 times faster than real time which is about 7 times faster than detail ed model.
� It should be pointed out that the performance of th e Neural Network model depends on the number of independent variable s (look ups) and the number of dynamic physical models. Thus, the nu merical efficiency of the NN model will be maintained even in cases wh ere the detailed model numerical efficiency degrades due to stiffer and/or more complex kinetics.
� Further improvements in computation speed can be expect ed, e.g. by using parallelized computing.