Anda di halaman 1dari 6

Comparison of Early-stage Design Methods for a Two-mode Hybrid Electric Vehicle

Kukhyun Ahn+, J Whitefoot+, V. P Atluri*, E Tate*, and P Papalambros+


+

University of Michigan * General Motors Company information to converge the solution. Another difficulty is that decomposition typically sets different objectives for each subsystem (sub-problem), which may not be well aligned with the overall system objective. For example, optimizing an electric motor to minimize power use may not guarantee minimizing fuel consumption of the vehicle. In this paper, we examine how solutions from a simple decomposition-based strategy compare with solutions obtained through AIO in the specific context of two-mode hybrid powertrain design. Large design organizations tend to be naturally decomposed and it is preferable to maintain this decomposition for system optimization, if the solutions thus obtained do not significantly sacrifice optimality. On the other hand, if the decomposed design solution is far from the AIO solution, additional integration effort and cost might be warranted. In the present formulation we are focusing on a few main subsystems and use relatively simplified system and subsystem models. However, we employ physics-based models rather than scalable maps for the engine and motor so that adequate fidelity is preserved for the results. Reviewing some of published work points to a few recognizable trends. Component models with higher fidelity are replacing off-the-shelf selection or simple scaling of baseline designs [6]. Further, increasingly advanced optimization techniques are employed for design [6-8]. A common practice is to combine global and local search methods in a sequential way. One expects that decomposition-based design optimization strategies will play an increasingly important role as higher-fidelity models are integrated and powertrain configurations become more complicated. Lastly, hybrid architectures in the literature have been shifting from series and parallel types to power-split and other advanced ones. Few optimization studies exist for power-split and multi-mode power-split architectures. Future research efforts are likely to consider these systems and improvements to emerging architectures, such as plug-in hybrid and range-extended electric vehicles. The present work aims to advance the discussion on the aforementioned research trends. High-fidelity component models are integrated in the simulation with geometric design variables. A design optimization framework based-on

Abstract: A challenge in the early-stage design of hybrid propulsion system architectures is dealing with a vast design space consisting of distinct subsystems, such as engine, motor, transmission, and battery. In this paper, two design optimization methods for a two-mode hybrid vehicle are examined: The first integrates the subsystems into a single virtual system and solves an All-In-One (AIO) problem; the second maintains system decomposition and updates individual subsystem designs through iterative exchange of information. The analysis models are detailed high-fidelity simulations of engine and electric motor components combined with two-mode transmission and overall vehicle models. Results indicate that the AIO systemoptimal solution provides little additional benefit, suggesting that the industry practice to use ad-hoc decomposition-based optimization is sufficient for early-stage design in this class of problems.

I.

INTRODUCTION

The vehicle design process involves coordination of separate groups designing their own specific subsystems. For example, in designing the propulsion system of a hybrid electric vehicle (HEV), engine, transmission, motor, battery, power electronics, and other component groups work both individually and together. A common way to organize such a design process is to have the groups work separately but exchange their design information, such as operation load distribution, efficiency index, and infeasible operation requirements, as needed. This is a simple form of decomposition-based optimization that is mathematically ad hoc, as it usually admits no convergence proofs. However, it is very practical in that it involves much lower cost for managing the information flow or integrating individual processes than theoretically advanced coordination methods or so-call All-In-One (AIO) optimization formulations where all systems are considered together as one single problem [1]. Ad-hoc decomposition-based optimization strategies are known to result in suboptimal solutions compared to, say, those obtained through AIO optimization [1-5]. One difficulty with ad-hoc strategies is that the interdependence among the subsystems cannot be accounted for by exchanging static information. As soon as a design change takes place in any of the subsystems, the shared static information becomes outdated, and designs based on this information are likely to be suboptimal. Thus, such strategies generally include an iterative process of design updates and the exchange of new

metamodels is presented to solve the component sub-system and propulsion system problems. The two-mode hybrid system presented has an advanced architecture with fairly high complexity and nonlinearity. This study contributes to the literature for advanced architectures and provides a design framework that can be applied to other systems of comparable complexity. The remainder of the paper is as follows. Section 2 presents the models of components and system used in the optimization study; Section 3 shows the optimization formulations; Section 4 presents and discussed the results obtained, and Section 5 offers conclusions. II.
COMPONENT AND PROPULSION SYSTEM MODELING

shape, air-gap, pole/slot numbers, slot shape, and shaft radius. These fixed design parameters and sine-wave drive information are summarized in Table 1.

A spark-ignition (SI) engine was simulated for this study. The engine is a single-cylinder simulation, though the baseline engine is a 2.0L, 4 cylinder engine. Thus, all torque and power results are scaled by a factor of four to represent the four-cylinder engine. All engine simulations were performed in GT Power v6.2 [9]. The baseline model developed at the University of Michigan for the SI engine is a single-cylinder, four valve, port-injected SI engine from the GT Power library. The model was parameterized to allow for modification of the bore and stroke as well as the valve diameters and intake runner diameters. Twenty-five engine maps were created from a combination of five bores and five strokes from 77.4mm to 86.0mm, representing engines that ranged from 1.46L to 2.0L in displacement. Using these data, maximum and minimum torque curves were recorded along with the BSFC as a function of torque and speed. In addition, the engine motoring torque was predicted for each engine speed to allow for engine-off operation of the vehicle (zero fuel mass injected) but account for the motor energy necessary to perform engine-off operation (All-EV mode). The technology considered for the two-mode hybrid system is the interior permanent magnet (IPM) AC synchronous motor. For higher-fidelity modeling, commercial electric motor design software, SPEED, was used [10]. The software provides a range of analysis methods, and we chose to use the embedded finite element analysis (FEA) solver. This strategy runs FEA to calculate adjustment factors to account for magnetic saturation and then performs static analysis using the phasor diagram and the calculated factors. Fig. 1 shows the flux density distribution calculated in the FEA along with the outline design of the model developed at the University of Michigan. The design variables were stator outer radius, stack length, and coil turn number. The stator inner radius and the rotor outer radius were set to vary according to the stator outer radius, but we fixed the rest of the design, such as magnet

Figure 1. Flux density distribution in the motor Table 1. Motor simulation parameters
Pole / slot numbers Shaft radius Magnet length/width Air-gap Maximum current DC drive voltage Winding configuration 12 / 18 50 mm 35 X 5 mm 1 mm 88 Arms 350 V Y-connected, concentrated

To generate each designs static map with maximum torque and efficiency information, a tabulation process collected shaft torque, core/copper loss and drive voltage values at drive current and shaft speed grid points. Inverter losses were not modeled in detail, but accounted for by applying a constant efficiency of 97 %. Only the first quadrant of motor operation is analyzed, and the others are obtained by mirroring the map. For building a metamodel, we generated maps for 100 Latin hypercube design samples that covered the three-dimensional design space. Metamodels were built for the motor and engine using feedforward artificial neural networks and radial basis function networks. As described earlier, they are built based on the physics-based simulations in GT Power and SPEED, but there is no need to run these programs any longer to generate a map for an arbitrary design of interest once the metamodels are built. Thus, prediction by the metamodel drastically reduces computation time and enables the use of continuous design variables beyond an off-the-shelf selection. The challenge is achieving adequately high prediction accuracy for the studys purposes. Fig. 2 shows a sample case of the metamodel prediction for the motor. Fig. 2(a) is the map from SPEED analysis, and Fig. 2(b) the predicted map. Note that the design used here was not used in building the metamodel, and this is pure prediction of completely new designs.

The battery model for this vehicle is based upon a known NiMH battery cell and the size of the battery pack is linearly scaled. We decided not to use a physics-based model for the battery because the design of the battery pack is highly dependent on packaging, cooling, and other practical constraints that are not modeled here and are beyond the scope of this study. There are many options for estimating an HEVs fuel economy. Here, the most important consideration is that simulations need to take a reasonably short time so that a large number of simulations can be conducted to explore the design space. This makes it hard to use forward-looking calculation or optimal control-based strategies. Since most of the commercial software uses forward-looking calculation, a simulator was developed that is based on backward-looking calculation and equivalent energy consumption minimization. At each instant, a selection is made by finding the optimal operation that minimizes equivalent energy consumption, which is a linearly-weighted sum of fuel and electric energy consumptions. The weight works as a conversion factor from one energy property to another and is held constant throughout a simulation run. A charge-sustaining logic is developed, finding a conversion factor that leads to no or tolerably little change between the initial and final SOC (state of charge) values. This also ensures charge-sustaining (SOCcorrected) fuel consumption for fair comparison of different vehicle designs. Figure 2. Metamodel prediction of motor maximum torque and efficiency As shown, the prediction matches analysis results fairly closely. The transmission of choice for this study is a two-mode transmission with four fixed gears. See the schematic diagram in Fig. 3. It has three planetary gear (PG) sets and four wet clutches to realize two electrically variable transmission (EVT) modes and four fixed gear (FG) modes [11-13]. To assess a vehicles performance, the full-load capacity of its powertrain needs to be calculated. This full-load information can be used to determine acceleration, passing, climbing, and towing performances. It is obtained when an output torque maximization problem is solved for every output speed in the vehicles speed range [13]. Additional challenges exist in the case of multi-mode hybrids because the process also involves mode selection as does the energy management during normal driving situations. For each speed at which the optimization problem is solved, all six modes are compared to find the mode that leads to the maximum output torque. In addition to this instantaneous calculation, it is also desirable to consider seamless shifting or shifting with minimal shocks. III.
OPTIMIZATION METHODS

The overall problem formulation is presented below. The objective is to minimize fuel consumption on the US-06 drive cycle. The upper bound on the 0-60mph acceleration time constraint was varied from 8.5sec to 10.5sec to create a Pareto set for the results: Figure 3. Cross-sectional schematic view of two-mode transmission Objective: Minimize Fuel Consumption on US-06 drive cycle

Constraints:

0-60 mph time 8.5, 9.0, 9.5, 10.0, 10.5 sec

In addition, upper and lower bounds were placed on the nine design variables selected for this problems are shown in Table 2. Table 2. Design variables and bounds Lower Bound Upper Bound Plan. Gear Ratio 1 Plan. Gear Ratio 2 Plan. Gear Ratio 3 Final Drive Ratio Motor stack length (mm) Motor outer radius (mm) Number of coil turns Engine Stroke (mm) 0.33 0.33 0.33 3.0 30 120 60 77.4 0.67 0.67 0.67 4.5 50 140 80 86

transmission and battery subsystems are integrated in the top level powertrain system design, see Fig. 4. Solving this system-level optimization problem uses the most recent engine and motor designs to compute the transmission and battery designs. The engine and motor use histories are passed down to the respective subsystems to construct new optimization sub-problems. Feeding the respective optimal solutions to the system-level problem completes one iteration loop. This iteration is repeated until the system-level objective converges. It is also possible to iterate until the design changes are within tolerance to avoid local optima. The sub-problem formulation is detailed below. The engine subsystem receives engine operating points (torque/speed points) from the system-level simulation. The subsystem optimization then uses these points to design the engine to minimize fuel consumption at these points while maintaining the acceleration performance of the vehicle. During this optimization, all other vehicle design variables are kept fixed. Specifically, the engine subsystem optimization is as follows: Minimize With respect to Subject to
140 120 100 250 Torque [N-m] 80 60 40 20 400 500 0 1000 280 280 300 350 400 500 1500 2000 300 350 400 500 500 260 250 260 280 300 350 400 BSFC Max torque curve Operation points 240 250 250 260 280 300 350 400 500 4000 250 260 280 300 350 400 500 4500 5000 260 280 300 350 260

Engine Bore (mm) 77.4 86 Note that the number of coil turns was treated as a continuous variable although it is discrete. Since this is a relatively large number and the only discrete design variable in the problem, the discrete optimal solution can be found by evaluating and comparing the designs with the optimal continuous value rounded up and down. In the AIO method, as the name implies, all subsystem models are integrated into a single optimization model. The optimization process does not involve iterative exchange of information, and the subsystem designs are determined simultaneously by solving the single optimization problem. At every step during the numerical optimization, the changes in subsystems are adjusted with no delay. From a design space perspective, the barriers between the subsystem design spaces are removed, and the merged design space contains all information of the subsystems.

(Fuel consumption + Out-of-bounds operation penalty) Engine bore and stroke 0-60mph acceleration time t
target

240 250

2500 3000 3500 Speed [RPM]

Figure 5. Optimized engine subsystem with some out-ofbounds operation points The out-of-bounds operation penalty allows for some of the requested operating points to occur at torque levels above the maximum torque of the engine. This is done to allow for engine downsizing within the decomposition-based strategy. Otherwise, the engine would not have the freedom to downsize past its initial point, thus potentially limiting the ability to reach the globally optimal design. An example

Figure 4. Decomposed optimization design flow In the specific variation of decomposition-based optimization that we examined, engine and motor design takes place separately from the remaining subsystems, and the

result from the engine subsystem optimization is shown in Fig. 5, with engine operating points plotted to show out-ofbounds operation above the maximum torque curve. The motor subsystem is managed in almost the same way as the engine subsystem. The subsystem optimization can be formulated similarly as follows. Minimize With respect to Subject to (Power consumption + Out-ofbounds operation penalty) Stator radius, length, and coil turns 0-60mph acceleration time t
target

differences between the two methods ranged from 0.6% to 6.7%.

The objective to minimize is power consumption instead of fuel consumption. In the literature, metamodel-based design optimization has been found to be a very attractive method for HEV powertrain problems, especially when each simulation run takes considerable time and a large number of runs is needed to search the large design space [6,8,14]. In our specific situation, use of metamodels not only reduces computational cost but also greatly helps to find better solutions by reducing numerical noise in the simulations. Calculating chargesustaining fuel consumption using interpolation creates irregular responses that do not necessarily reflect the physics of the design, and metamodels smooth out such irregularities. Quadratic polynomial regression was the chosen metamodeling form, despite its limited flexibility compared to other metamodel techniques, because it showed adequate accuracy over the entire design space. For calculating the 55 quadratic coefficients including cross terms, 3,000 optimal Latin hypercube samples were used, and all nine design variables were normalized between zero and one. IV. RESULTS

Figure 6. Performance of AIO and decomposed optimization on US06 cycle Similar results were seen when optimizing on the FTP75 drive cycle. The differences in this case were smaller, ranging from 0% to 1.7%.

The optimization results on the US06 drive cycle are shown in Fig. 6 for all five levels of the 0-60mph acceleration constraints. The figure shows each iteration of the decomposition-based method, which converges within 4 to 5 iterations. The cloud of dots represents the attainable set from the 3,000 design samples. The line consisting of the AIO solutions can be regarded as the theoretical Pareto frontier. The decomposition-based (DB in the legend) methods did not start from very good designs because the subsystem designs were fixed at the lowest bounds to avoid out-of-bounds operation problems. However, the method was able to quickly drive the optimization to move the solutions to the boundary of the cloud or even outside of the cloud. Then the solutions approached the AIO solutions, and the fuel consumption

Figure 7. Performance of AIO and decomposition-based optimization on FTP75 cycle V. CONCLUSIONS

Overall, the decomposition-based strategy performed well. In most of the cases, the solutions were fairly close to the AIO solutions, and the differences did not justify the additional effort and cost to integrate naturally-decomposed design groups into a single, unified design process. The optimality was found to be marginally sacrificed in merely two cases (4.7% and 6.7%). Still, robustness issues seem to be the key obstacle for this ad hoc strategy to remain as the industry practice. In the present study, initial design and the objectiveconstraint correlation were major factors that affected robustness. To make this ad hoc strategy perform almost as well as the AIO strategy without failing cases, these issues

need to be properly addressed; coordination strategies and mathematical convergence proofs are topics of future research. Here are a few additional difficulties that we encountered during the optimization, and they are mostly from using static information for each subsystem optimization to represent the rest of the system. First, out-of-bounds operations kept the sequential process from converging faster. For example, downsized designs usually caused these operations, and the penalty imposed on them was not able to provide a clear direction of optimization. Second, the system-level constraint, i.e. 0-60mph acceleration, could not be well translated down to the subsystem level. In our ad hoc strategy, we used the same acceleration as in the system level, assuming the rest of the system is fixed, and this strongly limited the feasible design space for each sub-problem. Addressing these systemsubsystem consistency issues is also part of our future research. In conclusion, separation of the subsystems generally had minimal effect on the eventual design with just one exception, at least within the accuracy of the simulations. Thus, for the early-design stage of vehicle development, the ad hoc decomposition-based strategy is sufficient. In actual practice, the use of subsystem experts enhances the value of a decomposition-based approach, as such expertise tends to get lost in an AIO type of approach.

approach to hybrid electric propulsion system design. Mechanics of Structures and Machines, 27(4), 393-421, 1999. [7] Filipi, Z. S., Louca, L. S., Daran, B., Lin, C.-C., Yildir, U., Wu, B., Kokkolaras, M., Assanis, D. N., Peng, H., Papalambros, P. Y., and Stein, J. L., Combined optimization of design and power management of the hydraulic hybrid propulsion system for a 6x6 medium truck. International Journal of Heavy Vehicle Systems, 11(3-4), 371-401, 2004. [8] Kokkolaras, M., Mourelatos, Z., Louca, L., Filipi, Z., Delagrammatikas, G., Stefanopoulou, A., Papalambros, P., and Assanis, D., Design under uncertainty and assessment of performance reliability of a dual-use medium truck with hydraulic-hybrid powertrain and fuel cell auxiliary power unit. SAE paper 2005-01-1396, 2005. [9] GammaTechnologies, Inc., http://www.gtisoft.com. [10] The SPEED Laboratory, http://www.speedlab.co.uk. [11] Grewe, T., Conlon, B., and Holmes, A. Defining the General Motors 2-mode hybrid transmission. SAE paper 2007-01-0273, 2007. [12] Ahn, K. and Cha, S. Developing mode shift strategies for a two-mode hybrid powertrain with fixed gears. SAE Int. J. Passeng. Cars Mech. Syst., 2009, 1(1), 285-292. [13] Ahn, K. and Papalambros, P. Design optimization of motor/generator full-load characteristics in two-mode hybrid vehicles. SAE Int. J. Passeng. Cars Electron. Electr. Syst., 2009, 2(1), 389-396. [14] Tate, E. and Boyd, S. Finding ultimate limits of performance for hybrid electric vehicles. SAE paper 2000-013099, 2001.

References
[1] Cramer, E. J., Dennis, J. E., Frank, P. D., Lewis, R. M., and Shubin, G. R., Problem Formulation for Multidisciplinary Optimization. SIAM J. of Optimization, 4, pp. 754776, 1994. [2] Sobieszczanski-Sobieski, J., and Haftka, R. T., Multidisciplinary Aerospace Design Optimization: Survey of Recent Developments. Structural Optimization, 14(1), pp. 1 23, 1997. [3] Kim, H. M., Michelena, N. F., Papalambros, P. Y., and Jiang, T., Target Cascading in Optimal System Design. J. of Mechanical Design, 125(3), pp. 474480, 2003. [4] Tosserams, S., Etman, L. F. P., and Rooda, J. E., Augmented Lagrangian Coordination for Distributed Optimal Design in MDO. Int. J. Numerical Methods Engineering, 73 (13), pp. 18851910, 2008 [5] Allison, J. T., Optimal partitioning and coordination decisions in decomposition-based design optimization. Ph.D. Dissertation, Department of Mechanical Engineering, University of Michigan, USA, 2008. [6] Assanis, D., Delagrammatikas, G., Fellini, R., Filipi, Z., Liedtke, J., Michelena, N., Papalambros, P., Reyes, D., Rosenbaum, D., Sales, A., and Sasena, M., An optimization

Anda mungkin juga menyukai