On the Robustness of Adaptive Nonlinear Model

7 downloads 0 Views 2MB Size Report
Nonlinear MPC in PCM. • 23 kB ROM, 59 kB RAM. • Model Adaptation (RLS). • 0.6-2.4% FE Benefit. *. *. **. 2018-01-1360. 3. Project History. SAE 2016-01-0155.
On the Robustness of Adaptive Nonlinear Model Predictive Cruise Control O. Santin1, J. Beran1, O. Mikuláš1, J. Pekar1, J. Michelini2, S. Szwabowski2, S. Mohan2, D. Filev2, J. Jing3, U. Ozguner3 1

2

3

Content Introduction Algorithm High-level Overview Robustness Simulation Study Road Grade Accuracy Impact Analysis Road Grade Phase Error Correction Algorithm Conclusion

SAE INTERNATIONAL

2018-01-1360

2

Project History Adaptive Nonlinear Model Predictive Controller (ANLMPC) Project History SAE 2017-01-0090

SAE 2016-01-0155 * • • • • •

Lincoln MKT Nonlinear MPC in PCM 23 kB ROM, 59 kB RAM Model Adaptation (RLS) 0.6-2.4% FE Benefit

SAE 2018-01-1360

* • • • • •

Lincoln MKT + 1600 kg trailer Downshifting Avoidance 24 kB ROM, 6 kB RAM Constrained + Dead-zone RLS 6.3-8.8% FE Benefit

** • • • •

Ford F-150 + trailer Robustness Study Grade Phase Error Correction Up to 11% FE Benefit (4500 kg trailer)

*https://cars.usnews.com/ ** https://ford.com/ SAE INTERNATIONAL

2018-01-1360

3

MPC: Model Predictive Control PCC: Predictive Cruise Controller

Introduction # passengers, fuel type, Trailer mass & drag

Weather Conditions

Ford F-150

GPS Location Inaccuracy

Grade Map Inaccuracies

Goal: Evaluation of robustness of MPC based PCC SAE INTERNATIONAL

2018-01-1360

Real grade Maps grade 4

Algorithm High-level Overview Slow changes (drag, rolling resistance, mass, bias/scale in grade, fuel type)

Fast changes (e.g. head wind)

Grade phase offset

SAE INTERNATIONAL

2018-01-1360

5

5

Robustness Simulation Study - Multipliers Fuel Type (octane number)

Weight

Aerodynamic Drag

Grade

Tire Pressure • •

Ford F-150 Model Baseline: ACC controller

SAE INTERNATIONAL

Test Road I94 M39 US12

Length [miles] 11 12 8.5

2018-01-1360

Road grade [%] (-2.0, 2.5) (-3.3, 3.8) (-5.5, 5.5)

Real grade Maps grade

Speed [mph] (61, 69) (51, 59) (46, 54)

limits

6

Weight

Aero

Weight

Aero

MPG benefit [%] MPG benefit [%]

US12

I94

Robustness Simulation Study - Results

Grade Tire Press Fuel (octane) Fuel effect • Low octane + ACC higher torque spark timing further from MBT higher fuel Aero effect • Depends on grade Grade Tire Press Fuel (octane) symmetricity (uphill / downhill sections) • US12 closes to grade symmetricity (positive avg.) Grade effect • +/- Avg. grade  -/+ FE benefit

Baseline: ACC controller (no preview utilization) with the same average speed SAE INTERNATIONAL

2018-01-1360

7

Motivation for Road Grade Error Study

SAE INTERNATIONAL



In-vehicle data: • grade phase error Δd • control performance.



Robustness improvement  a grade correction method is developed.

2018-01-1360

8

Road Grade Accuracy Impact Analysis • Road Grade Preview Errors 𝜙𝐵 𝑡 = 𝜸𝜙 𝑑 + 𝜟𝒅 + 𝜙𝐵𝑖𝑎𝑠 ,

Location Inaccuracy Grade Map

• Handled by Adaptation of Model Parameters (RLS) 𝑑𝑣(𝑇, 𝑣, 𝜙ሻ Scale / Bias = 𝛽1 𝑇 + 𝛽2 𝑣 2 + 𝛽3 𝜙𝐵 + 𝛽4 𝑑𝑡 error 𝑔 • Scale Error: 𝛽3 = − 100𝜸

• Bias Error: 𝛽4 = 𝛽4 𝑛𝑜𝑚 + 𝛽3 𝜙𝐵𝑖𝑎𝑠 • Phase Error: Not possible to separate the effect SAE INTERNATIONAL

2018-01-1360

9

MPH benefit [%]

𝑣ሶ = 𝛽1 𝑇 + 𝛽2 𝑣 2 + 𝛽3 𝜙𝐵 + 𝛽4

Model Parameters for 𝜟𝒅

𝜟𝒅 Grade Phase Error [m]

𝛽3

MPH benefit [%]

𝜙𝐵𝑖𝑎𝑠 Grade Bias Error [%]

MPH benefit [%]

𝛽2

𝜸 Grade Scale Error [-]

MPG benefit [%]

MPG benefit [%]

MPG benefit [%]

Road Grade Accuracy Impact Analysis – Results*

𝜟𝒅 = 0 m 𝜟𝒅 = -15 m 𝜟𝒅 = -10 m 𝜟𝒅 = 10 m 𝜟𝒅 = 15 m

 Large Sensitivity to Phase Error

*Ford F-150 + 10,000lbs trailer, baseline is ACC controller, US12 road (+/-5.5 %) SAE INTERNATIONAL

2018-01-1360

𝜙𝐵 𝑡 = 𝜸𝜙 𝑑 + 𝜟𝒅 + 𝝓𝑩𝒊𝒂𝒔 , 10

Road Grade Phase Error Correction Algorithm • • • • • •

Method basis: Minimizing residuals in model representation Grade Error: 𝜙𝑒 𝑘, Δ𝑑 = 𝜙 𝑘 − 𝜙෨ 𝑘, Δ𝑑 Vehicle acceleration 𝑎𝑘 = 𝛽1 𝑇k + 𝛽2 𝑣𝑘2 + 𝛽3 𝜙(𝑘ሻ + 𝛽4 𝑎ොk (Δ𝑑ሻ = 𝛽1 𝑇k + 𝛽2 𝑣𝑘2 + 𝛽3 𝜙෨ k, Δ𝑑 + 𝛽4 Predicted acceleration Error in predicted acceleration ek (Δ𝑑ሻ = 𝑎k − 𝑎ො𝑘 (Δ𝑑ሻ Implementation: Collect backward looking vector of acceleration errors 𝐞𝐤 (Δdሻ = [𝑒𝑘 , 𝑒𝑘−1 , 𝑒𝑘−2 , ⋯ , 𝑒𝑘−𝑁𝑝 ]′

• Solve the optimization problem min 𝐞𝐤 (Δ𝑑ሻ

2

Δ𝑑

SAE INTERNATIONAL

2018-01-1360

11

Road Grade Phase Error Correction Algorithm • Task Flow Diagram

𝛽 1,⋯,4 ,𝑘−1

SAE INTERNATIONAL

2018-01-1360

12

Road Grade Phase Error Correction Algorithm - Results

SAE INTERNATIONAL

2018-01-1360

20m grade phase delay

MPH

MPG

Gear Shifts

Fuel Saving

With

49.21

16.20

38

10.46%

Without

48.66

15.57

120

6.13% 12

Conclusion • Simulation study for Ford F-150 vehicle • Achievable FE benefit up to 11% (10000lbs trailer) • Positive impact from vehicle mass • Grade and Aerodynamic drag depends on symmetricity of the road grade profile • Road Grade Inaccuracies • Offset and scale error handled by RLS • Phase error handled by the iterative estimation of the phase lag

SAE INTERNATIONAL

2018-01-1360

14

Thank you Ondrej Santin – [email protected] Shankar Mohan – [email protected]

M39 (~+/-3.5%)

I94 (~+/-2.5%)

Bias Error

Scale Error

US12 (~+/-5.5%)

SAE INTERNATIONAL

2018-01-1360

16

M39 (~+/-3.5%)

I94 (~+/-2.5%)

Phase Error Comp.

Phase Error

US12 (~+/-5.5%)

SAE INTERNATIONAL

2018-01-1360

17

Suggest Documents