PHEV Control Strategy Optimization Using MATLAB Distributed Computing: From Pattern to Tuning MathWorks Automotive Conference 3 June, 2008 S. Pagerit, D. Karbowski, S. Bittner, A. Rousseau, P. Sharer Argonne National Laboratory Sponsored by Lee Slezak, U.S. DOE G. Sharma The MathWorks
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 2
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 3
New Constraints = More Complex Vehicle Fuel Economy Environment HEVs and PHEVs Cost and Complexity Higher use of math-based tools before and during design: Model-Based design Physical modeling Monte Carlo analysis Caveat: Increased detail of modeling, complexity Increased number of simulations => Longer calculation, analysis and development time 4
More Complex Vehicle = More Sensitive Control Higher Electric Energy Higher Electric Power Higher Control Freedom Fuel Savings Potential Depending on various driven distance, several modes are possible during charge depleting: Electric-only (EV) and Blended km/h Common Strategy = EV+CS % km/h Blended Strategy % SOC Veh. speed η high ICE η high ICE η high ICE off off η low ICE s off off off s EV MODE CS MODE BLENDED BLENDED Optimization Evaluates Control Strategy s Potential 5
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 6
Tool: PSAT Powertrain Systems Analysis Toolkit Setup Simulation Analysis 7
Process: 3-Way Approach to Control Optimization Global Optimization T eng T mc ω eng Backward model Control Design PSAT Various Control Principles ω mc X( 0) PSAT Derivative Free Optimization { P 0 ( t + 1 ), SOC ( t )} J () t 0' + 1, 0,0' { Pm ( t), SOCn( t) } U 0 ( t).. U M ( t). { PM ( t + 1 ), SOC M '( t + 1) }., J M, M '() t.. Bellman Principle X ( T ) Rule-Based Control Design Optimal Control Pattern, Minimal Fuel Cons. Various PSAT Control Strategies Optimally tuned Parameters Control Logic DIRECT Algorithm 8
Hardware: Computation Time Requires Distributed Computing MATLAB Distributed Computing Server - 1 Header Node - 2 Worker Nodes MATLAB and Parallel Computing Toolbox # 1 # 2 # 8 MATLAB Distributed Computing Server - 2 Worker Nodes Server Rack: - PSAT Software - Simulation Results 9
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 10
Finding Control Patterns Fast(er) Robustness: Different Cycles (e.g.: Urban, Highway, ) Different Distances Different Initial SOC => Set of 45 simulations => Sequential computation time ~ 2 weeks Using Distributed Computing: Simulations run in parallel => Running time ~ 12 hours 11
Global Optimization Showed Minimal Fuel Consumption Achieved in Blended Mode SOC 1 0.9 0.8 0.7 0.6 0.5 Japan 10-15 x25 (PSAT) Japan 10-15 x25 (gl. optim) UDDS x10 (gl. optim) NEDC x10(gl. optim) 3 cycles demonstrated blended strategy is optimal when knowing the distance 0.4 0.3 0.2 Example of EV+CS Mode for comparison 0 10 20 30 40 50 60 70 80 90 100 % of total distance 12
Global Optimization Showed Engine Starting Condition Almost Proportional to Electrical Consumption Pwr wheels (kw) above which P(ICE=ON)=.95 35 30 25 20 15 10 5 0 0 50 100 150 200 250 Electric Energy Consumption (Wh/km) Cycle: UDDS HWFET LA92 Driven Distance: 10 miles 20 miles 40 miles 13
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 14
Engine ON/OFF Logic & Engine Torque Request Engine ON / OFF Logic in Stateflow 1 wh trq dmd 2 wh spd dmd 1 eng _pwr_dmd 3 SOC nit.soc_ ess_soc_init ess pwr charging eng pwr dmd divider 6 acc pwr 4 eng spd f(u) Operating region for the engine??? Eng hot low torq curve??? Eng hot opt torq curve eng_trq_opt_dmd 2 engine torque Engine Torque Demand in Simulink??? Eng hot high torq curve 5 eng trq max 15
Blended Control Strategy Design Showed Significant Improvements Over EV Mode Fuel Energy Increase Over EV/CS 2% 0% -2% -4% -6% -8% -10% 0 16 32 48 64 80 96 112 Up to 9% less fuel consumed Trip distance (km) Full Engine Power Optimal Engine Power 10 miles AER vehicle run on several UDDS cycles 16
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 17
Control Parameter Tuning with DIRECT Robustness: Different Cycles Different Distances Iterative Process: Control space is sampled at each iteration One simulation is run for each sample Control space is re-sampled around the best Simulation => Convergence after 30 iterations ~ 400 simulations => Sequential computation time ~ 2 days Using Distributed Computing: Simulations run in parallel => Running time ~ 5 hours 18
The Longer the Electrical Distance, the More Robust Fuel Consumption Ratio 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Influence of cycles (distance=23.7km) Fuel Consumption Ratio 1 Battery Energy Ratio 0.8 UDDS2 UDDS4 0.6 UDDS6 UDDS8 Tuned for a longer distance (6xUDDS) Fuel Consumption Ratio 1.4 1.2 0.4 0.2 0 1.8 1.6 1.4 1.2 Influence of cycles (distance 1 =71.2Km) Fuel Consumption Ratio Battery Energy Ratio 2 0.8 0.6 0.4 0.2 UDDS2 UDDS4 UDDS6 UDDS8 Cycles 0 Battery Energy Ratio Tuned for a short distance (2xUDDS) 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Battery Energy Ratio 19
Outline Introduction Setup Global Optimization for Patterns Real Time Controller DIRECT Optimization for Tuning Conclusion 20
Conclusion Using a combination of optimization techniques and modeling based on PSAT and MATLAB, we were able to: single out control patterns, implement them in Simulink and Stateflow, tune their parameters. Only the use of distributed computing allows this process to be performed in a timely manner: After setting up the MATALB Parallel Computing Toolbox, and Distributed Computing Servers, it took only 1 hour of development to get the first simulations running. The optimization times were reduced from more than 2 weeks to less than a day. 21
References Plug-in Hybrid Electric Vehicle Control Strategy: Comparison between EV and Charge-Depleting Options Sharer, P. / Rousseau, A. / Karbowski, D. / Pagerit, S. SAE World Congress 2008, April 2008 Impact of Component Size on Plug-In Hybrid Vehicle Energy Consumption Using Global Optimization Karbowski, D. / Haliburton, C. / Rousseau, A. EVS 23, December 2007 Plug-in Hybrid Electric Vehicle Control Strategy Parameter Optimization Rousseau, A. / Pagerit, S. / Gao, D. EVS 23, December 2007 Plug-in Vehicle Control Strategy: From Global Optimization to Real Time Application Karbowski, D. / Rousseau, A. / Pagerit, S. / Sharer, P. EVS22, October 2006. Global Optimization to Real Time Control of HEV Power Flow: Example of a Fuel Cell Hybrid Vehicle Pagerit, S. / Rousseau, A. / Sharer, P. EVS 21, April 2005. 22