Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
High-Fidelity Modeling of Light-Duty Vehicle Emissions and
Fuel Economy Using Deep Neural Networks
Farhang Motallebiaraghi, Aaron Rabinowitz, Jacob Holden, Alvis Fong,
Shantanu Jathar, Thomas Bradley, and Zachary D. Asher
Western Michigan University, Colorado State University, NREL
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 2
Acknowledgements
This material is based upon work supported by the
U.S. Department of Energy’s Office of Energy
Efficiency and Renewable Energy (EERE). The
specific organization overseeing this report is the
Vehicle Technologies Office under award number
DE-EE0008468.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 3
Agenda
Introduction
Methodology
Results
Conclusion
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 4
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Transportation Sector: Major greenhouse gas emission source
Transportation Sector…
• accounts for ~ ⅓ of all energy used in the U.S.
• is the largest source of CO2 production.
From Transportation Energy Data Book Edition 38: https://tedb.ornl.gov and https://www.epa.gov/mobile-source-pollution/research-health-effects-exposure-risk-mobile-source-pollution
● Severe climate change.
● ~15,000 premature deaths every year.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 5
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
ICE Light Duty Vehicles: Fleet dominators
https://www.epa.gov/greenvehicles/fast-facts-transportation-greenhouse-gas-emissions https://www.eia.gov/outlooks/aeo/aeo2019
• Future reduction in vehicular emissions depends on improvements in energy efficiency
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 6
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Gap: High-fidelity emission and fuel economy modeling
Existing emission models:
• MOtor Vehicle Emission Simulator (MOVES)
and EMission FACtor (EMFAC) may
underestimate on-road emissions.
• Previously it is shown that shallow Artificial
Neural Networks (ANNs) have the potential to
predict on-road emissions and better model the
fuel economy.
www.acscm.com/projects/4wd-chassis-dyno-emissions-test-cell-upgrade
a- Annenberg et al., Impacts and mitigation of excess diesel-related NOx emissions in 11 major vehicle markets,2017
Chenna, Shiva Tarun. n.d. “ARTIFICIAL NEURAL NETWORKS FOR FUEL CONSUMPTION AND EMISSIONS MODELING IN LIGHT DUTY VEHICLES.” Colorado State University. Libraries. https://mountainscholar.org/handle/10217/197403.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 7
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Focus: Deep Neural Networks (DNNs) modeling
• Modeling and evaluation of Deep Neural Networks (DNNs) and history-sensitive (recurrent) DNNs for both fuel consumption and tailpipe emissions from a targeted light-duty ICE vehicle.
• In this study we answered how different combination sets of input data can affects on prediction accuracy using different models.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 8
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
PEMS (Portable Emission
Measurement System)
Drive Cycle Development and on-road
Data Collection
● NOx (ppmv, g/s)
● CO (%, g/s)
● CO2(%, g/s)
● PM10 (mg/m3, g/s)
● HC (ppmv, g/s)
PEMS AXIONR/S+ Volkswagen Jetta, 2003
CAN
● Velocity (mph)
● Engine speed (RPM)
● Manifold air pressure (kPa)
● Intake air temperature (F)
● Exhaust fuel rate (g/s)
● Intake fuel rate (g/s)
● Fuel consumption (g/s)
GPS
● Latitude (deg)
● Longitude (deg)
● Altitude (deg)
Drive cycle used for training
● Data was collected from 5 drive cycles in Fort Collins in 2018.
● We selected 3 drive cycles for training, 1 for validation and 1 for
testing the models
Drive Cycle Development and on-road Data Collection
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 9
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Data Classification
Data Classification
Which combination set of these predictors is more accurate?
● Predictors were categorized into two main categories:
○ Externally Observable Variables (EOV)
○ Internally Observable Variables (IOV)
Predictor Name Predictor Symbol Variable Type
Vehicle Velocity V EOV
Vehicle Acceleration ACC EOV
Time since Start time EOV
Vehicle Altitude Alt EOV
Vehicle Specific Power VSP EOV
Engine Velocity RPM IOV
Intake Air Temperature IAT IOV
Manifold Air Pressure MAP IOV
Fuel Consumption FC IOV
Frey, H.C., Zhang, K., and Rouphail, N.M., “Vehicle-specific emissions modeling based
upon on-road measurements,” Environ. Sci. Technol. 44(9):3594–3600, 2010.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 10
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Data Classification
Data Classification
To select the best combination set sufficient number of iterations were
tested using predictors.
● 10 classes were defined for emission modeling.
● 9 classes were defined for fuel consumption modeling.
Class
Input combination sets
IOV needed
Emission prediction Fuel consumption prediction
C1 V, ACC, time V, ACC, time no
C2 V, ACC, time, VSP V, ACC, time, VSP no
C3 V, ACC, time, Alt V, ACC, time, Alt no
C4 V, ACC, time, VSP, Alt V, ACC, time, VSP, Alt no
C5 V, ACC, time, VSP, Alt, RPM V, ACC, time, VSP, Alt, RPM yes
C6 V, ACC, time, VSP, Alt, RPM, IAT V, ACC, time, VSP, Alt, RPM, IAT yes
C7 V, ACC, time, VSP, Alt, RPM, MAP V, ACC, time, VSP, Alt, RPM, MAP yes
C8 V, ACC, time, VSP, Alt, MAP, IAT V, ACC, time, VSP, Alt, MAP, IAT yes
C9 V, ACC, time, VSP, Alt, RPM, IAT,
MAP
V, ACC, time, VSP, Alt, RPM, IAT,
MAP
yes
C10 V, ACC, time, VSP, Alt, RPM, IAT,
MAP, FC--- yes
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 11
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Emissions and Fuel Consumption Modeling
CAN
● Velocity (mph)
● Engine speed (RPM)
● Manifold air pressure (kPa)
● Intake air temperature (F)
● Exhaust fuel rate (g/s)
● Intake fuel rate (g/s)
● Fuel consumption (g/s)
GPS
● Latitude (deg)
● Longitude (deg)
● Altitude (deg)
● Data collected from 5 drive cycles in Fort Collins in 2018.
● 3 for training, 1 for validation and 1 for testing the
models
Emission and fuel consumption
models
Non-Machine
learning
Machine Learning
(ML)
Artificial Neural
Networks (ANN)
Multiple Linear
Regression (MLR)
Deep FFNN Deep CNN
Motor Vehicle Emission
Simulator
Deep LSTM
Artificial Intelligence
Machine Learning
CNNLSTM
FFNN
Artificial Neural Networks
RNN
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 12
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Emissions and Fuel Consumption Modeling
Deep FFNN (Feed Forward Neural
Network)
We selected a three-hidden layer FFNN and “sigmoid” activation
function.
General structure of a Feed Forward Neural Network (FFNN)
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 13
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Deep LSTM (Long Short-term Memory)
General structure of an LSTM
We selected a three-hidden layer LSTM and “ReLu” activation function.
Emissions and Fuel Consumption Modeling
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 14
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Deep CNN (Convolutional Neural
Network)
General structure of an CNN
We selected a two-hidden layer CNN and “ReLu” activation function.
Emissions and Fuel Consumption Modeling
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 15
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
MLR (multivariable linear regression)
General formulation of an MLRThis model changes the coefficients in the fitting process to
reduce the error indicated in the assessment metrics section.
Emissions and Fuel Consumption Modeling
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 16
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
• For evaluation of the model’s accuracy, we used Mean Absolute
Error (MAE) criteria parameter:
Assessment Metric
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 17
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
● Adding IOV predictors increases the accuracy
of the prediction.
● Adding fuel consumption, engine speed and
manifold air Pressure affects the most on the
emission prediction accuracy (C10)
EOV
EOV &IOV
Behavior of adding predictors on emission prediction accuracy
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 18
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Behavior of adding predictors on fuel consumption prediction accuracy
• The effect of adding IOV parameters can be
seen the most in class 7 and class 9.
• class 9 (C9) was selected as the primary input
class for fuel consumption prediction.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 19
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
● CO2 is predicted the best based on visual
analysis of the traces.
● For better CO prediction more drive cycle
data is required, errors on predictions
happened mostly during cold start phase
(first 500 seconds)
Emissions comparison results using LSTM
Emission comparison
MAE (%)
CO2 NOx HC CO PM10
LSTM 0.04 0.03 0.09 0.06 0.02
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 20
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
● LSTM was able to predict HC with very
low error after 500 seconds of drive cycle
and NOx for almost complete drive cycle.
● LSTM was not able to have a good fit on
PM10 on the spikes, though the MAE is
the minimum.
Emissions comparison results using LSTM
Emission comparison
MAE (%)
CO2 NOx HC CO PM10
LSTM 0.06 0.03 0.09 0.04 0.02
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 21
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
MAE (%)
Input classes LSTM CNN FFNN MLR
C9 1.21 11.2 6.02 199.5
Comparison of measured fuel consumption rates to all
models fuel consumption estimations (test) using C9.
Fuel Consumption comparison results using LSTM
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 22
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
Comparison of measured emission rates to MOVES emission rate estimations
Emission
comparison
Average emission rate (g/mi)
CO2 NOx HC CO PM10
Test dataset 626 1.82 0.34 0.49 0.041
MOVES 532 3.27 0.79 5.23 0.017
LSTM 625.72 1.82 0.34 0.49 0.04
CNN 619.99 1.79 0.33 0.48 0.04
FFNN 555.58 0.91 0.26 0.33 0.03
MLR 1297.39 2.80 0.55 0.10 0.05
Relative Error (RE) rate (%)
MOVES 15 26.7 132.3 967 58.5
LSTM 0.05 0.02 0.09 0.05 0.03
CNN 0.96 1.83 1.57 1.06 0.99
FFNN 11.25 50.24 23.15 32.11 30.20
MLR 307.25 254.1 262.2 120.14 217.15
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 23
ConclusionsResultsMethodologyIntroduction ConclusionsResultsMethodologyIntroduction
• The results show that the deep neural network’s performance consistently
improves when given datasets with more variables (EOV and IOV).
• manifold absolute pressure (MAP), engine speed (RPM), and fuel
consumption are the most beneficial parameters categorized as IOV for
emission prediction.
• LSTM had the best performances for both emission and fuel
consumption prediction. This is because LSTM account for both delayed
effects and recurrent effects for more accurate predictions.
• This model, if developed for a vehicle and integrated within the vehicle
controller, may have value for real time vehicle/engine controls
optimization thereby producing real-time reductions in fuel consumption
and emissions.
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 24
Speaker Information
Thank you• Farhang Motallebiaraghi
• Western Michigan University
SAE International®
SAE WCX Digital Summit Paper # 2021-01-0181 25
Image Resources
• https://www.hrw.org/news/2020/04/02/us-car-emissions-rollback-endangers-peoples-health• https://www.vitalstrategies.org/breathing-smoke-city-air-pollution-lung-cancer
• https://www.epa.gov/greenvehicles/fast-facts-transportation-greenhouse-gas-emissions
• https://www.sciencedirect.com/science/article/pii/S0301421520302627?via%3Dihub
• https://www.eia.gov/outlooks/aeo/