Incorporating Drivability Metrics into Optimal Energy Management Strategies for Hybrid Vehicles. Daniel Opila

Transcription

1 Incorporating Drivability Metrics into Optimal Energy Management Strategies for Hybrid Vehicles Daniel Opila

2 Collaborators Jeff Cook Jessy Grizzle Xiaoyong Wang Ryan McGee Brent Gillespie Deepak Aswani, Ameet Deshpande The Team

3 Unifying Themes Control of Energy Systems: Multiple Conflicting Objectives Many Actuators/Sensors Complex Dynamics Model Reduction Time-Dependent Decisions Future Information 200 Engine speed Engine torque CISG torque Engine torque feedback 150 Commanded CISG torque Parallel mode request Parallel mode actual

4 What is Power Management? Key problem: determine the total amount of power to be generated and its split between the multiple power sources/modes; transmission gear number (1 st, 2 nd, etc). Legal Mandate: Charge sustaining Engine Transmission Driver Control Module Eng. Command Gear Command Motor Command Battery Electric Motor

5 Goals: Overall Project Objectives Make power management controller design systematic & optimal Address drivability and real-world driving conditions Compare to industrial controller Hardware implementation 5

6 Two-level control architecture Energy Management Problem Driver Supervisory Powertrain Control System Fuel economy/ emissions/ drivability Long horizon slow dynamics (simple) Engine controller Mot/Battery controller Trans controller Brakes controller Engine Engine Clutch Clutch Motor Motor Battery Battery Transmission Transmission Vehicle Vehicle Transient Response/ drivability/ Fuel economy/ emissions/ Short horizon fast dynamics(detailed) Lin-2002

7 Vehicle Configuration Modified Volvo S-80 Ford developed a baseline controller Front Differential Electric Machine 2 (EM2) Diesel Engine text text Battery Clutch EM 1 te text Transmission Electric Machine 1 7

8 Shortest Path Stochastic Dynamic Programming Dynamics: x k+1 = f(x k,u k,w k ) Driver Behavior: w k Total Cost: k =1 c( x, u) System Dynamics Probabilities Cost Function c(x,u) SDP Feedback Controller Minimizes: Cycle Expectation E k =1 c( x k, u k ) 8

9 Solution Method Driving Cycles Vehicle Data Program Objectives Markov Chain Modeling Vehicle Dynamic Model Cost Function Stochastic Dynamic Programming

10 Solution Method Driving Cycles Vehicle Data Program Objectives Markov Chain Modeling Vehicle Dynamic Model Cost Function Stochastic Dynamic Programming Simplified HEV model Optimal Control Policy, u * (x) Model Simplification Fuel Economy, Drivability Vehicle Response Complete Ford Model Real-Time Implementation

11 Stochastic Driver Model Discrete Time Markov Chain One Step Transition: P(v k+1 v k,a k ) v k,a k p 1 p 2 p 3 Compute by Cardinality k + 1 k 1 v k+1,a 1 v k+1,a 2 v k+1,a Number of occurences of next acceleration for v=5 m/s a=-0.8m/s Speed (Mph) time (s) Next Acceleration (m/s 2 ) 11

12 Statistics-Based (Stochastic) Drive Cycle Animation 12 Animation: Ed Tate

13 Basic Development Concept Design (off-line) Typical Drive Cycles Plant Data Automated Stochastic DP Real-Time Controller High Level Constraints (SOC range, Drivability) Implementation (Real-time) Inputs Velocity Battery SOC Current Gear Engine State Power Demand Table Interpolation & Minimization argmin E{c(x,u) + V(x k+1 )} Outputs Gear Engine Torque Engine State Clutch EM1 Torque EM2 Torque 13

14 Ford Drivability Requirements (from the experts) Average time in gear Gear hunting behavior Engine Speed after a shift Downshifts as a function of pedal angle Short duration engine events Pedal Correlation with Engine Noise Simplified Drivability Metrics Primary Metrics: Total Number of Shifts Total Number of Engine on/off events 14

15 Cost Function: What we care about Fuel Only: c ( x, u) = Fuel Fuel + Final SOC: c ( x, u) = Fuel + κ * SOCfinal Fuel + Final SOC + Drivability: c( x, u) = Fuel + α * Shifts + β * EngOns + κ * SOCfinal 15

16 Tradeoff Analysis Fuel Only: 60 c ( x, u) = Fuel 50 Fuel + Final SOC: c ( x, u) = Fuel + κ * SOCfinal Speed (mph) Fuel + Final SOC + Drivability: c ( x, u) Fuel + α * Shifts + β * EngOns + κ * SOCfinal = time (s) Fixed Mpg Mpg Gear Events SPSDP 93 Gear Events -Ford 25 Gear Events 75 Gear Events 100 Engine Events SPSDP 93 Gear Events -Ford 25 Gear Event Fit 25 Gear Event Data Gear Event Fit 75 Gear Event Data Engine Events

17 Drive Cycle Robustness Drive cycles from University of Michigan Transportation Research Institute (UMTRI) 2700 Cycles, 87 Drivers 2 sets of 100 cycles: Ensembles 1 & 2 Each set of 100 cycles is roughly 1000 mi CDF CDF Full Data Set 0.1 Ensemble 1 Ensemble Trip Length (miles) Full Data Set Ensemble 1 Ensemble FTP Cycle NEDC Cycle Velocity (mph) 17

18 Model Simulation Results: Real-World Cycles 10% Relative Improvement in Fuel Economy on Real-World Drive Cycles New Controllers Better Fuel Econom my Baseline Controller Worse Better Drivability Worse 18

19 19 Statistical Real-World Fuel Economy

20 Limitations of Baseline Controller Goal: Trade off fuel economy vs. engine activity Compare to a baseline industrial controller Can it be tuned differently? Better? Fuel Economy Engine On-Off Activity No possible tuning exists to match performance! 20

21 Theoretical WHY? SPSDP Baseline u u u u u 4 = G( x1, x2, x3, x4, x5) u1 g11( x1, x2) u2 = g21( x2, x3, x4) u (, ) + ( ) 3 g31 x1 x5 g32 x3 u 4 g 41 ( x 2, g 42 ( x 3, x 4 )) x x x x Table Interpolation & Minimization argmin E{c(x,u) + V(x k+1 )} u u u u SOC Wheel Power Optimal Battery Power Battery Power Engine Eng State State Machine Actuator Commands 21 Assuming an a priori structure can limit performance!

22 To hardware! Table Interpolation & Minimization argmin E{c(x,u) + V(x k+1 )} u u u u x x x x

23 Progressive Complexity Engine Torque (Nm) Engine Speed (rad/s) Control-Oriented Model Ford s Fuel Economy Model Model-In-Loop Simulation (MIL) Hardware-in-Loop Real-Time (HIL) Vehicle

24 Multi-Rate Actuation Engine On/Off Gear Faster Engine Torque EM 1 Command EM 2 Command Brake Command 1 hz 3 hz 20 hz

25 80 Driving NEDC Actual Target Engine On 50 Speed (mph) time (s)

26 Driving NEDC Actual Speed Target Speed Engine Speed (RPM/100) Engine Torque (Nm/10) time (s)

27 Normalized Fuel Economy (with hardware failure) Cycle Controller Normalized Corrected Fuel Economy (MPG) FTP72 Baseline 1 Delta FTP72 SDP % NEDC Baseline NEDC SDP % 27

28 Vehicle Testing on FTP with Hardware Fault Simulation Vehicle Tests Mpg Normalized Engine Events Better Drivability Worse 28

29 Vehicle Testing on FTP with Hardware Fault Simulation Vehicle Tests Mpg Normalized Engine Events Better Drivability Worse 29

30 Summary Stochastic optimal control is a Method rather than a controller: Optimizes fuel economy and constraints Highly automated Robust to real-world driving Generates controllers that work in hardware These techniques are applicable in general: Other hybrid configurations Other Systems 30

31 Questions? Engine speed Engine torque CISG torque Engine torque feedback Commanded CISG torque Parallel mode request Parallel mode actual Vehicle speed Engine speed SOC Parallel mode actual