Part Load Engine Performance prediction for a gasoline engine using Neural Networks. Sreekanth R, Sundar S, Rangarajan S, Anand G -System Simulation

Part Load Engine Performance prediction for a gasoline engine using Neural Networks Sreekanth R, Sundar S, Rangarajan S, Anand G -System Simulation CAE-2 System Simulation GT-SUITE User Conference Feb 06, 2017

CONTENTS Outcome of the Seminar 1. Dependency of Transient Engine Thermal Management simulations on engine heat release rates. 2. Prediction of engine part load heat release rates based on full load P-Ө. Conclusions Calibration and validation of Full load performance Training, Implementing and validation of Neural networks Part load P-Ө prediction using trained Neural networks Validation of Engine Thermal Management with predicted P-Ө and Future goals Figure 1: Part load P-θ prediction steps Images courtesy: www.bing.com 2

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 3

1- INTRODUCTION Challenges and Motivation Vehicle Platform concept freeze Performance evaluation of Engine thermal management system components Thermal loads consideration for optimum component sizing. Limited test data motivates to predict the part load heat release rates. Engine Heat release Methods to predict heat release rate Challenges: Method Vs Parameter Inputs complexity Cost Time Accuracy Heat release rates not available in development phases Complexity in measurement Map data from test High High High -- 1D Medium Less Less Medium. 3D High Medium Medium High Figure 2: Engine heat release to Head, Block and Piston 4

1- INTRODUCTION 1D GT-ISE model to predict Heat release rates Engine Thermal management layout 1D Engine model in GT-ISE Engine Metal mass Figure 3: Engine thermal management component layout Deliverables Figure 4: 1D Engine model in GT-ISE Engine Full load and Part load IMEP,FMEP and BMEP Engine Full load and Part load P-θ curves Deriving Heat release rates from Part load P-θ and execute Engine thermal management 5 Images courtesy: www.bing.com

1- INTRODUCTION Methods to predict engine performance Non- Predictive models Burn rate is directly imposed as simulation input. Used for studies which does not impact burn rate. Less computation time. Predictive models Takes into account the cylinder geometry, spark location, timing, air motion and fuel properties. Higher computation time and inputs complexity. Used for studies involving parameters which impact burn rate. Semi predictive models Good substitute for predictive model. Utilizes Wiebe methodology. Utilizes neural networks to calculate Wiebe parameters. Used for studies involving parameters which impact burn rate. Image courtesy: www.bing.com 6

1- INTRODUCTION Incylinder pressure(bar) Incylinder pressure(bar) Simulation Steps for WOT calibration, Neural network training and PLP prediction Assumed data Test data INPUTS Burn rate prediction Calibrated data Model Predicted data 2000 RPM 3250 RPM Burn rate prediction: Air fuel ratio(-) Engine speed(rpm) Injected fuel mass(mg) Measured P-θ curves Spark timing(deg.ca) Volumetric efficiency 60 OUTPUTS Part load P-θ at 2500RPM Simulation data Test data Training Neural networks: Air fuel ratio at IVC(-) Burn fraction(100%) Engine speed(rpm) Gas temperature at IVC(K) Spark timing(deg.ca) Trapped mass at IVC(mg) Training output data: Anchor angle Burn duration Wiebe exponent FTP calibration: Engine speed(rpm) Inlet and outlet boundary conditions(flow, pressure and temp.) Intake and exhaust cam center angles(vvt) Burn fraction(100%) Intake and exhaust valve timings Valve discharge coefficients Calibrated intake and exhaust ports(htms and temperatures) Air fuel ratio(fuel and Air flows) Spark timing Inlet valve closing angle Discretization lengths Engine friction data(rfmep wrt RPM) Wiebe combustion parameters Trained Neural networks for PLP prediction FTP Calibration 0-360 -180 0 180 360 Crank angle(deg.ca) 70 0 Part load P-θ at 3000RPM Simulation data Test data -360-180 0 180 360 Crank angle(deg.ca) 7

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 8

RFMEP(bar) Torque(N-M) Air flow(kg/hr) 2 MODEL CALIBRATION FOR FULL LOAD PERFORMANCE Model calibration for Full load performance(flp/wot/ftp) Full load Calibration steps Volumetric efficiency calibration (± 2%) In cylinder Heat transfer multiplier calibration Parameter Air flow as inlet boundary condition 1.5 250 Full load Air flow and Fuel flow comparison FTP_Air flow_test data FTP_Air flow_simulation FTP_Fuel flow_test data FTP_Fuel flow_simulation 40 Back pressure calibration Torque at highest speed within ± 1% FMEP(Chen-Flynn model) Constant, piston speed factor, Piston speed squared factor from test data. Peak cylinder pressure factor= 0.005 0 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(n-m) Table 1: WOT Calibration steps Graph 1: Air flow and fuel flow comparison 3 Rubbing friction mean effective pressure_breakdown Cylinder head valvetrain_90deg.c Water pump Oil pump_90deg.c Piston group_90deg.c 120 WOT Torque comparison Crank shaft_90deg.c Total Rubbing Friction MEP 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 2: Engine rubbing friction breakdown FTP_Test data FTP_Simulation 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(n-m) Graph 3: Simulation WOT torque comparison with test data 9

2 MODEL CALIBRATION FOR FULL LOAD PERFORMANCE Incylinder pressure(bar) Incylinder pressure(bar) Incylinder pressure(bar) Incylinder pressure(bar) Full load P-θ Validation with test data Full load_p-θ_1500rpm Full load_p-θ_2000rpm Test data 100 Simulation data Test data 100 Simulation data -360-180 0 180 360 Crank angle(deg.ca) -360-180 0 180 360 Crank angle(deg.ca) Graph 4: Full load P-θ comparison with test data Full load_p-θ_3000rpm Graph 5: Full load P-θ comparison with test data Full load_p-θ_4000rpm Test data 100 Simulation data Test data 100 Simulation data -360-180 0 180 360 Crank angle(deg.ca) Graph 6: Full load P-θ comparison with test data -360-180 0 180 360 Crank angle(deg.ca) Graph 7: Full load P-θ comparison with test data 10

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 11

3 Training, implementation and validation of neural networks Neural networks for predicting Anchor angle, Burn duration and Wiebe exponent WOT data at Inlet Valve closing angle Parameter variation in WOT 1 Engine speed(rpm) 1000 to 6000 Sensitive parameters for Input training data Output training data 2 Burn fraction(-) 1 3 A/F ratio(-) 13.5 to 16.5 4 Trapped mass(mg) 310 to 345 5 Temp at IVC(K) 380 to 430 6 Spark timing(deg.ca) 0 to 30 1 2 50% Burned Crank Angle(deg.CA) Burn Duration 10-90%(deg.CA) Table 2: Neural networks training data 5 to 25 18 to 27 3 SI Wiebe Burn Exponent(-) 9 to 1.nno file Figure 5: Model incorporated with trained Neural networks Note: If any parameter is constant, then it has to be exclude from being given as input or output training data to the neural networks. 12

3 Training, implementation and validation of neural networks Temperature(K) Air Fuel ratio(-) Trapped mass(mg) Spark timing(btdc,deg.ca) Validating neural networks control logic at IVC(inlet valve closing) 20 Air fuel ratio at IVC Control strategy input Training input 500 Trapped mass at IVC Control strategy input Training input ± 0.7(-) 0 1000 1500 2000 2500 3000 Engine speed(rpm) 3500 4000 4500 Graph 8: Air-Fuel ratio comparison at IVC ± 3 mg 300 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 9: Trapped mass comparison at IVC 500 Temperature at IVC Control strategy input Training input 30 Spark timing Case setup input Training input 300 ± 2 C 1000 2000 3000 4000 Engine speed(rpm) Graph 10: Trapped gas Temperature comparison at IVC 0 1000 1500 2000 2500 3000 Engine speed(rpm) 3500 4000 4500 Graph 11: Spark timing comparison There is a deviation of ± 4% in Air-Fuel ratio, ± 1% in trapped mass and ±0.5% in temperature compared with the actual input at inlet valve closing angle. 13

3 Training, implementation and validation of neural networks Wiebe exponent(-) Anchor angle(deg.ca) Burn duration(deg.ca) Engine Torque(N-m) Validating neural networks control logic at IVC(inlet valve closing) 20 Anchor angle(50% burn angle from TDC) Neural network output Training output 40 Burn duration(10% to 90%) Neural network output Training output 0 ± 2 deg.ca 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 12: Anchor angle comparison with training output ± 3 deg.ca 20 1000 1500 2000 2500 3000 Engine speed(rpm) 3500 4000 4500 Graph 13: Burn duration comparison with training output 5 Wiebe exponent Neural network output Training output 120 FTP comparison with FTP from neural networks FTP_after calibration FTP_with Neural networks ± 1.5(-) 0 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 14: Wiebe exponent comparison with training output 60 1000 2000 3000 Engine speed(rpm) 4000 ± 6 % Graph 15: FTP comparison with Calibrated model There is a deviation in Anchor angle(± 2 deg.ca), Burn duration(± 3 deg.ca) and Wiebe exponent(± 1.5) from the neural networks with the optimum training model. With the above deviations, Neural network predicts Full load performance within ± 6 %. 14

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 15

4 Part load P-θ prediction using neural networks Part load P-θ prediction input data Part load P-θ is predicted by using the trained neural networks with part load input data. Inputs from part load test data Parameter Unit Description rpm RPM Engine Speed AMBIENT-PRES bar Ambient Pressure AMBIENT-TEMP K Ambient Temperature AFR (-) Fuel Ratio Air_Flow_target kg/s Target Signal Spark_timing deg.ca Spark timing(btdc) IVC_angle deg.ca Inlet valve closing angle Ambient-TManifoldExt K Ambient Temp Surrounding Manifold EvaporationConstant (-) Injected Fuel Vaporization Constant ncyc (-) Simulation Duration TW-Head K Head Zone 1 Temperature TW-Piston K Piston Zone 1 Temperature TW-Liner K Cylinder Zone 1 Temperature IFAM (-) Flow Area Multiplier EFAM (-) Flow Area Multiplier IDR g/s Injector Delivery Rate ITA deg Injection timing angle WallTemp K Wall temperature Table 3: Inputs considered from Part load test data for prediction Part load performance 16

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 17

5 Part load results validation Torque(N-m) Torque(N-m) Comparison between test data and simulation data Part load performance comparison 100 N-m Part load performance comparison 70 N-m 120 Test data Simulation data 80 Test data Simulation data ± 5 % 80 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) ± 9 % 60 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 16: PLP comparison for all speeds at 100N-m Graph 17: PLP comparison for all speeds at 70N-m Engine Part load performance is significantly predicted within ± 5% compared with test data. However for lowest part loads the variation is ± 10-15% due to limited input training data. 18

Torque(N-m) Torque(N-m) 5 Part load results validation Comparison between test data and simulation data Part load performance comparison 50 N-m Part load performance comparison 40 N-m 60 Test data Simulation data 50 Test data Simulation data ± 2 % 40 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 18: PLP comparison for all speeds at 50N-m ± 11 % 30 1000 1500 2000 2500 3000 3500 4000 4500 Engine speed(rpm) Graph 19: PLP comparison for all speeds at 40N-m Engine Part load performance is significantly predicted within ± 5% compared with test data. However for lowest part loads the variation is ± 10-15% due to limited input training data. 19

5 Part load results validation Incylinder pressure(bar) Incylinder pressure(bar) Comparison between test data and simulation data PLP_P-θ 2500RPM PLP_P-θ 3000RPM 50 Simulation data Test data 60 Simulation data Test data 0-360 -180 0 180 360 Crank angle(deg.ca) Graph 20: Part load P-θ Comparison 0-360 -180 0 180 360 Crank angle(deg.ca) Graph 21: Part load P-θ Comparison Parameters Test P-θ Simulation P-θ BMEP(Bar) 4.780 4.646 P max(bar) 36.8 33.38 θ at P max 12 12 θ at dp/dθ max 3-2 Table 4: Part load P-θ characteristics comparison between simulation and test data Parameters Test P-θ Simulation P-θ BMEP(Bar) 7.830 7.840 P max(bar) 46.49 47.03 θ at P max 15 15 θ at dp/dθ max 5 1 Table 5: Part load P-θ characteristics comparison between simulation and test data 20

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 21

6 Validation of Engine thermal management with predicted part load performance Aim of Calibration & Calibration Parameters To calibrate the model using simulation P-θ by finding the optimal combination of calibration parameters. The calibrated loads should be physically realistic. Initial Loads Combustion Heat Load Piston-Oil Heat Load Underhood Air Heat Load Calibration parameters kcombustion Kpiston-oil kair Calibrated Loads Combustion Heat Load Piston-Oil Heat Load Underhood Air Heat Load Effect of Calibration parameters Figure 6: Engine thermal management simulation process layout Calibration Parameter Coolant Temp Oil Temp k combustion Driving Cycle-NEDC (Vehicle Speed, Engine Speed, Engine Torque) k piston-oil - k air Table 6: Effects of calibration parameters 22

6 Validation of Engine thermal management with predicted part load performance Coolant temperature(deg.c) Oil temperature (deg.c) Comparison between test data and simulation data Coolant Temperature (degc) Oil Temperature (degc) Test data Simulation data Test data 120 120 Simulation data 0 200 400 600 800 1000 1200 Time (sec) 0 200 400 600 800 1000 1200 Time (sec) Graph 22: Coolant temperature raise comparison between Simulation data and Test data Graph 23: Oil temperature raise comparison between Simulation data and Test data Coolant temperature and Oil temperature is predicted within ± 5 0 C deviation compared with the test data. 23

CONTENTS Contents 1. Introduction 2. Model calibration and validation for Full load performance 3. Training, implementation and validation of neural networks 4. Part load P-θ prediction using neural networks 5. Part load results validation 6. Validation of Engine thermal management with predicted part load performance 7. Conclusions and future work 24

7 Conclusions and Future work Conclusions and Future work Conclusions: There is a deviation of ± 4% in Air-Fuel ratio, ± 1% in trapped mass and ±0.5% in temperature compared with the actual input at inlet valve closing angle. There is a deviation in Anchor angle(± 2 deg.ca), Burn duration(± 3 deg.ca) and Wiebe exponent(± 1.5) from the neural networks with the optimum training model. With the above deviations, Neural network predicts Full load performance within ± 6 % Neural network approach can be used as a semi predictive model to predict majority of part load performance within ± 5%. However for lowest part loads the variation is ± 10% due to limited input training data. Coolant and oil temperature is predicted within ± 5 0 C. Future goals: Fine tuning of the GT model to predict all the part loads with ± 5% and to predict engine thermal management coolant and oil temperatures between ± 2 0 C compared with test data. 25

Thank you for your attention! Queries? RNTBCI Contact person: R Sreekanth Mail id: Rayavalasa.Sreekanth@rntbci.com Images courtesy: www.google.com 26