ScienceDirect ScienceDirect - Colorado State University

6 downloads 1817 Views 1MB Size Report
2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights ... simulation-based comparison is a set of real-world driving data.
Engine and Powertrain Control, Simulation and Modeling 4th IFAC23-26, Workshop August 2015. on Columbus, OH, USA 4th IFACand Workshop on Available at www.sciencedirect.com Engine Powertrain Control, Simulation and online Modeling Engine and Powertrain Control, OH, Simulation August 23-26, 2015. Columbus, USA and Modeling August 23-26, 2015. Columbus, OH, USA

ScienceDirect

IFAC-PapersOnLine 48-15 (2015) 271–276 The Effect of Trip Preview Prediction Signal Quality on Hybrid Vehicle Fuel The Signal Economy The Effect Effect of of Trip Trip Preview Preview Prediction Prediction Signal Quality Quality on on Hybrid Hybrid Vehicle Vehicle Fuel Fuel Economy Economy

Thomas Cummings*. Thomas H. Bradley.* Zachary D. Asher*H. Bradley.* Thomas Cummings*. Thomas  Thomas H. Bradley.* Thomas Cummings*. Zachary D. Asher* Zachary  D. Asher* *Colorado State University, Fort  Collins, CO 80523 USA (e-mail: [email protected]) *Colorado State University, Fort Collins, CO 80523 USA (e-mail: *Colorado State University, Fort Collins, CO 80523 USA (e-mail: [email protected]) [email protected]) Abstract: This study seeks to develop an understanding of the sensitivity of sensing and predictionderived vehicle to prediction signal Theoftrip pattern identification Abstract: This fuel studyeconomy seeks toimprovements develop an understanding of the quality. sensitivity sensing and predictionAbstract: This study seeks develop understanding of the sensitivity of sensing and predictiontype of scenario control was to selected foranin-depth study. For thisquality. scenarioThe control, we developed realderived vehicle fuel economy improvements to prediction signal trip pattern identification derivedderived vehicledrive fuel economy improvements to prediction signal quality. The trip pattern identification world cycles to test and demonstrate the effectiveness of the scenario control. Baseline type of scenario control was selected for in-depth study. For this scenario control, we developed realtype of scenario control selected for in-depth Fora this scenario control, wemodel. developed realmodels of the drive Prius HV was were refined used tostudy. develop baseline economy Optimal world derived cycles to test and and demonstrate the effectiveness of fuel the scenario control. Baseline world derived drive cycleswere to test and demonstrate the effectiveness of the scenario control. Baseline scenario control policies derived assuming perfect signal quality and were implemented in the models of the Prius HV were refined and used to develop a baseline fuel economy model. Optimal models of the Prius HV were model refinedtoand used to develop a baseline fuel economycontrol model.under Optimal baseline vehicle fuel economy demonstrate the effectiveness of the scenario ideal scenario control policies were derived assuming perfect signal quality and were implemented in the scenario control policies were derived assuming perfect signal and were in the conditions. Bothfuel the economy optimized and baseline vehicle models werequality thenofsubjected to implemented imperfections baseline vehicle model to demonstrate the effectiveness the scenario control underinideal baseline vehicle fuel economy modeloftoquantifying demonstratethe theabsolute effectiveness of the performance scenario control under ideal prediction signals with the objective and relative of the scenario conditions. Both the optimized and baseline vehicle models were then subjected to imperfections in the conditions. Both the optimized and baseline vehicle models were then subjected to imperfections in the control policies, thethe baseline vehicle control. the absolute and relative performance of the scenario prediction signalsand with objective of quantifying prediction signals with the objective of quantifying the absolute and relative performance of the scenario control and the baseline vehicle control. © 2015,policies, IFAC (International Federation ofoptimal Automatic Control) Hosting Elsevier Ltd.modeling, All rights reserved. Keywords: Hybrid electric vehicle, powertrain control,bypowertrain powertrain control policies, and the baseline vehicle control. simulation, prediction error impacts, dynamic programming, energy management, fuel economy. Keywords: Hybrid electric vehicle, optimal powertrain control, powertrain modeling, powertrain Keywords: Hybrid electric vehicle, optimal powertrain control, powertrain modeling, powertrain simulation, prediction error impacts, dynamic programming, energy management, fuel economy.  simulation, prediction error impacts, dynamic programming, energy management, fuel economy. to the vehicle controller. The trip pattern scenario control  1. INTRODUCTION  relies a prediction of many of the characteristics of the to the on vehicle controller. The trip pattern scenario control 1. INTRODUCTION to the vehicle predictions controller. of Thefuture trip pattern scenario control trip including vehicle location, future relies on a prediction of many of the characteristics of the 1. INTRODUCTION The fuel economy (FE) of modern vehicles is limited by relies onspeed, a prediction many of the characteristics of We the vehicle and tripofdestination al. 2008). trip including predictions of future (Gong vehicleetlocation, future the degree to which the vehicle can understand and The fuel economy (FE) of modern vehicles is limited by trip including predictions of future vehicle location, future study the trip pattern scenario control to understand what vehicle speed, and trip destination (Gong et al. 2008). We The fueltoeconomy (FE) of modern limited by respond itstoenvironment. A wide vehicles variety ofis researchers the degree which the vehicle can understand and vehicle speed, and trip destination (Gong et al. 2008). We the effects of pattern mis-estimation these predictions will be study the trip scenarioofcontrol to understand what the degree to which the vehicle can understand and have found that vehicle FE can be improved perfect respond to its environment. A wide variety ofwith researchers study the trip pattern scenario control to understand what on vehicle fuel economy. the effects of mis-estimation of these predictions will be respond to itsregarding environment. A wide varietyof of the researchers information trajectory have found that vehicletheFEfuture can be improved withvehicle. perfect the effects of mis-estimation of these predictions will be on vehicle fuel economy. have found that vehicle FE can be improved with perfect For example,regarding (O’Keefethe & Markel 2006) found thatvehicle. on vehicle fuel of economy. information future trajectory of the The objective this study is to provide more detailed regarding the future trajectory of the vehicle. information the length and energy intensity of For example, (O’Keefe & Markel 2006) found that perfect understanding of potential for fuel The objective of thisthestudy is to provide moreeconomy detailed For example, (O’Keefe & Markel 2006) found that trips can be used to derive of aperfect pluginformation regarding the FE-optimal length and operation energy intensity of The objective that of this study is to provide more detailed improvements can be achieved through vehicle-level understanding of the potential for fuel economy information regarding the length and energy have intensity of in hybrid electric vehicle. Other researchers sought trips can be used to derive FE-optimal operation of a plugunderstanding of the potential for fuel economy sensing, prediction, andbecontrol. seek to develop a improvements that can achievedWe through vehicle-level tripsoptimize can be used to derive FE-optimal operation of a plugto vehicle operation a forward-looking in hybrid electric vehicle. Otherusing researchers have sought improvements that quantification can be achievedofthrough vehicle-level simulation-based the absolute anda sensing, prediction, and control. We seek to develop in hybridofelectric vehicle. Other (Zhang researchersVahidi have sought window 2010) to optimize perfect vehicleinformation. operation using a & forward-looking sensing, prediction, andtrip control. We seek tocontrol develop a relative benefits of the pattern scenario as simulation-based quantification of the absolute and to optimize vehicle operation using energy a forward-looking use perfect predictions of segment intensity to window of perfect information. (Zhang & Vahidi 2010) simulation-based quantification of the absolute and function of signal quality. relative benefits of the trip pattern scenario control as a window optimal of perfect information. (Zhang & Vahidi 2010) develop hybrid vehicle control strategies, while use perfect predictions of segment energy intensity to relative benefits of the trip pattern scenario control as a function of signal quality. use perfect predictions of segment energy intensity to (Kohut al. 2009) use avehicle windowcontrol of prediction of vehicle developetoptimal hybrid strategies, while function of signal quality. 2. METHODS develop optimal hybrid vehicle control strategies, while speed to propose FE-optimized speed trajectories. (Kohut et al. 2009) use a window of prediction of vehicle 2. METHODS (Kohut et error al. 2009) a window of prediction Predition was use briefly considered by (He trajectories. etofal.vehicle 2012) The trip pattern prediction control aims to optimize fuel speed to propose FE-optimized speed 2. METHODS speed to was propose FE-optimized speed trajectories. but error implemented as a stochastically imperfect economy by recognizing routes aims and tooptimizing the Predition error was briefly considered by (He et al. 2012) The trip pattern prediction control optimize fuel Predition error was briefly by event (He etprediction al. 2012) prediction signal rather thanconsidered a real world The trip pattern prediction control aimsby to second optimizebasis. fuel function of the vehicle on a second but error was implemented as a stochastically imperfect economy by recognizing routes and optimizing the but error was implemented as a stochastically imperfect signal disturbance. economy by recognizing routesallows and the optimizing the Having of theonroute controller to prediction signal rather than a real world event prediction function knowledge of the vehicle a second by second basis. prediction signal rather than a real world event prediction function of the vehicle on a second by second basis. choose the most efficient powertrain operation over the signal disturbance. Having knowledge of the route allows the controller to signal disturbance.therefore exists to model more realistic The opportunity Having knowledgeTooftest the route allows the controller to predicted sensitivity of this scenario choose thedrive. most efficient the powertrain operation over the connections between the prediction error and vehicle choose the most efficient powertrain operation over control to prediction signal quality, we will measure the The opportunity therefore exists to model more realistic predicted drive. To test the sensitivity of this scenario The opportunity exists to the model more realistic control systems sotherefore as tothe understand real-world predicted drive. test the the trained sensitivity of this scenario fuel economy costTo when vehicle controller is connections between prediction error and vehicle vehicle control to prediction signal quality, we will measure the connectionsand between the preference prediction error and vehicle efficiency that control to prediction signal quality, that we will measure the subjected to normal disturbances are a result of control systems consumer so as to understand theimprovements real-world vehicle fuel economy cost when the trained vehicle controller is control systems so as to understand the real-world vehicle can result from sensing and control. We fuel economy cost when thequality. trained vehicle controller is imperfect Thisarecomparison efficiency and improved consumervehicle preference improvements that subjected prediction to normal signal disturbances that a result of efficiencya particular and consumer preference improvements that propose modevehicle of vehicle sensing and control subjected to flowchart normal disturbances that The are input a result of presented in format in Fig.1. to this can result from improved sensing and control. We imperfect prediction signal quality. This comparison is can result from improved vehicle sensing andidentification control. We that is referred to here as the trip pattern imperfect prediction signal quality. This comparison is simulation-based comparison isina Fig.1. set of real-world driving propose a particular mode of vehicle sensing and control presented in flowchart format The input to this propose acontrol. particular mode of vehicle sensing andscenario control scenario The trip pattern identification presented in flowchart format in Fig.1. The input to this data. The predictive controller works from that data to that is referred to here as the trip pattern identification simulation-based comparison is a set of real-world driving that is referred to here as the trip pattern identification control designed to simulate a predictive controller in simulation-based comparison a set of operation. real-world driving construct prediction of isvehicle This scenarioiscontrol. The trip pattern identification scenario data. The apredictive controller works from that data to scenario control. The oftrip patterninformation identification scenario which a large amount detailed is available data. The predictive controller works from that data to prediction is input to the scenario controller, which control is designed to simulate a predictive controller in construct a prediction of vehicle operation. This control is designed to simulate a predictive controller in construct a prediction of vehicle operation. This which a large amount of detailed information is available is input to the scenario controller, which which ©a 2015 largeIFAC amount of detailed information is available 271 prediction prediction is input to the scenario controller, which Copyright 2405-8963 © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Copyright 2015 responsibility IFAC 271Control. Peer review©under of International Federation of Automatic Copyright © 2015 IFAC 271 10.1016/j.ifacol.2015.10.039

IFAC E-COSM 2015 272 August 23-26, 2015. Columbus, OH, USAThomas Cummings et al. / IFAC-PapersOnLine 48-15 (2015) 271–276

provides strategic guidance to the vehicle running controller. The running controller ensures that component level constraints and driver torque commands are met by interrupting or adding commands to the vehicle plant. The output of the vehicle plant is the vehicle FE. We seek to compare the vehicle FE under a perfect information scenario to the vehicle FE when the scenario predictor mis-predicts the route due to normal disturbances that might be encountered during real-world driving. The modeled disturbances to the prediction controller will include normal route changes, traffic, and unplanned stoplights.

considered validated for the purpose of predicting the fuel economy of the HV Prius on the basis of their similarity to the real-world test results. 2.2 Drive Cycle Development For the Trip Pattern Identification scenario control, we have developed a drive cycle that allows for the concentration of features of interest into a short duration driving trace. The drive cycle seeks to concentrate urban, and traffic-sensitive driving conditions into a short drive cycle that can be repeated during simulation. The Trip Pattern Identification scenario control requires a secondby-second speed trace that includes the capability to insert traffic delays and traffic signal events. To develop this baseline drive cycle, the authors drove from the Colorado State University MERC Campus to a parking lot in South Fort Collins, Colorado. The drive cycle dataset was recorded at 1Hz using a handheld GPS device. The speed trace of this drive cycle (MERC drive cycle) is presented in Fig. 2 (labelled as “Trip Pattern Baseline Cycle”). Additional MERC drive cycle details are presented in Fig. 1.

2.1 Baseline Vehicle Fuel Economy Modeling The baseline simulations model the vehicle fuel economy and energy consumption of the HV (hybrid vehicle) Prius as it drives over industry standard and custom drive cycles. The baseline vehicle simulation is a dynamic vehicle fuel economy and energy consumption simulation. The Modelica simulation language is used to develop the baseline simulation models, and to solve for the performance of the vehicles. OpenModelica is a simulation environment using Modelica as an open source non-causal modeling language that includes dynamic differential equation solvers. Modelica is a physicalsystem modeling tool that allows for system transparency, modification and repeatability. The Modelica modeling environment additionally allows for speedy simulation in comparison with other simulation tools such as Matlab/Simulink. The baseline simulation includes custom component models developed by Colorado State University specific to HV Prius simulation (Geller & Bradley 2011). The first output from the development of the baseline simulation is a simulation of the Prius HV fuel economy over standard regulation drive cycles. These results are used to validate the performance of the baseline simulation against publically available fuel economy datasets. A summary of the results of this baseline simulation comparison is shown in Table 1. Table 1. Comparison of baseline HV Prius simulation results over regulation drive cycles. Simulation Type

Baseline HV Prius Simulation

Drive Cycle

Simulated CS FE

ANL-Tested CS FE (ANL 2010)

UDDS

84.2 mpg

75.6 mpg

US06

41.6 mpg

45.3 mpg

HWFET

70.7 mpg

69.9 mpg

Fig. 1. Cycle length, baseline simulated FE, and GPS route location for the MERC drive cycle. For the Trip Pattern Identification scenario control, we seek to model 3 types of prediction signal errors. Recall that the trip pattern identification scenario controller seeks to predict the drive cycle that the vehicle is driving on. We hypothesized that 3 types of prediction errors would commonly confound the prediction of the second-bysecond drive cycle, and each are presented for comparison to the Trip Pattern Baseline Cycle (MERC Cycle) in Fig. 2.

Over three relevant regulation drive cycles, the simulated charge-sustaining (CS) fuel economy (FE) of the baseline vehicle model can be compared to the CS FE as measured by Argonne National Laboratory (ANL) over the same drive cycles. In each case, the results compare the “hot” unadjusted test results over each cycle (SAE 2010). In general, the results of the baseline simulation are

The first type of prediction error is entitled “Traffic Signal.” Under this model of prediction error, the scenario control fails to predict a traffic signal. The

272

IFAC E-COSM 2015 August 23-26, 2015. Columbus, OH, USAThomas Cummings et al. / IFAC-PapersOnLine 48-15 (2015) 271–276

273

S  f (S , u, w)

vehicle is forced to decelerate to a stop and hold at zero speed for approximately 60 seconds. Upon accelerating away from the traffic signal, the vehicle rejoins its previous driving schedule.

(1) The optimization problem is then to determine the control sequence in discrete time,

u(k ), k  0,1,2...N  1

The second type of prediction error is entitled “Route Change.” Under this model of prediction error, the scenario control fails to predict a route change initiated by the driver. In this case, the driver ends the trip before the anticipated location has been reached. This prediction error represents a near-worst-case scenario as SOC is often at a maximum or minimum mid-cycle.

(2)

that minimizes the objective function, N 1

J cost   gcost S k ,u k , wk  k 0

(3)

subject to state and control constraints,

S k   S (k )

The last type of prediction error for the Trip Pattern identification cycle is entitled “Traffic.” Under this prediction error, the scenario controller does not predict traffic which slows down the vehicle. To model the effect of traffic, this cycle forces the vehicle to slow down to 85% of the baseline cycle speed for the first 335 seconds of the drive cycle. At that point, the traffic is assumed to have dissipated and the vehicle resumes travels at normal speeds.

uk  u (k )

(4) In our Prius energy management optimization problem, we consider S(k) to be the vehicle’s SOC at each stage, k, w(k) to be the power required by the vehicle to meet the drive cycle (Pvehicle), and u(k) to be the engine power control sequence (Pengine). With these defined, we can assemble a set of non-linear equations to define the dynamics of the problem.

Pvehicle  Pengine  Pbattery 2

Pbattery  VOC I battery  I battery Rint SOC 

I

battery

dt

C

(5) Based on these equations, we can see that the battery model assumes that the battery coulombic efficiency is 100%. The energetic efficiency of the battery is less than 100% because of losses from ohmic losses during charging and discharging that are modeled using the battery internal resistance. The battery energetic efficiency is defined by the ratio of the electrical energy that enters the battery to the energy extracted from the battery at constant state of charge. The thermal state of the battery is not modeled. The objective function Jcost is a summation of the fuel consumption at each stage gcost(k), so that minimization of Jcost maximizes vehicle fuel economy. The fuel consumption at each stage gcost(k) is calculated as a nonlinear function of Pengine.

Fig. 2. Speed traces of the drive cycle used for the Trip Pattern Identification scenario control. 2.3 Optimal Scenario Controller Development

The state of charge is constrained to remain within a recommended state of charge range, and the initial and final states of charge (SOCi and SOCf) are constrained to ensure that the change in state of charge over the drive cycle is small. The engine power command u(k) is constrained as nonlinear function of a maximum regenerative braking power, a maximum motoring power, and a maximum and minimum battery current.

The next task of developing the scenario controller is to develop a scenario controller that represents the globally optimal policy for driving over the baseline drive cycle. A deterministic dynamic programming algorithm will be used to derive optimal battery/vehicle/engine power flows so as to optimize the vehicle for the baseline drive cycle. The resulting optimal energy management policies can then be used to develop optimal controllers for the Prius that can implement the proposed scenario control under both the baseline and disturbed cycles. This process is illustrated schematically in Fig. 4.

 S (k ) | 23%   S (k )  80%  S (0)  SOCi    S ( N )  SOCf    (6) u (k ) |  f S (k ), w(k )  u (k )  f S (k ), w(k )

The generic dynamic programming problem is defined by a dynamic equation,

273

IFAC E-COSM 2015 274 August 23-26, 2015. Columbus, OH, USAThomas Cummings et al. / IFAC-PapersOnLine 48-15 (2015) 271–276

Deterministic dynamic programming proceeds by splitting the N-stage optimal control problem into a set of recursive 1-stage problems, with discrete states. Working backwards in time, dynamic programming with backward induction defines a cost to go V(S(k),k) at each state S(k). The cost to go defines the minimum cost to proceed from S(k) to each final state S(N). The optimal control policy u(k) satisfies the Bellman principal of optimality:

Pengine) to reach a global optimum in the discretized timespace domain. To implement the optimized control strategy into the Modelica simulation, we simply replace the Engine Command Algorithm of the baseline simulation with the optimized algorithm derived from the dynamic programming routine. The optimized control policy is imported into Modelica as a 2-D lookup table that can be polled at any state (SOC) or stage (time) at which the simulation finds itself. The output of the lookup table is the optimal engine power, from which the optimal engine torque command can be calculated. As illustrated in Fig. 3, whenever the driver command disagrees with the scenario controller recommendation (as might occur during operation over disturbed drive cycles), the running controller overrides the commands from the scenario controller and enables the vehicle to follow driver commands.

V ( S (k  1), k  1)  minJ cost ( S (k  1), u (k  1))  V ( S (k ), k )

(7)

This equation allows for the recursive calculation of the optimal control sequence u(k) beginning from S(N). The state space is discretized into ~30,000 discrete states (a spacing of