Optimization of combustion engine emissions and performance has become a highly complex problem (as negatively illustrated by the Volkswagen software defeat device scandal) involving—but not limited to—variable valve timing, variable valve lift, variable effective compression/expansion ratio (Miller cycle, Atkinson cycle), fuel pressure, fuel injection timing, fuel injection rate, fuel injection events, fuel reactivity, boost pressure / vane control, exhaust recirculation rate, combustion chamber design and exhaust aftertreatment system control.

The effect on performance of adding each new parameter can become harder to predict and the combustion itself can become more difficult to control.

Modern diesel engines can use from 5 to 8 injections during each combustion cycle alone, and in a rate shape controlled injector this would result in up to 4 parameters for each injection, potentially adding 32 variables to the necessary design space. The absurdity of such full parametric studies is quite clear.

Furthermore, the variable devices identified above would appear in the as-built condition of an engine, not its original design parameters, which could easily number into the hundreds. Given such complex systems which need to be tuned and calibrated for operation in the real world it is clear that optimization techniques need to be applied not only to computational design but also to operational testing of such engines. Validation of models, emissions testing, and rapid prototyping are just a few reasons that real engine testing must still occur despite so much progress in the realm of computational modeling.

—Bertram (2014)

Control parameter optimization is thus used to enable the rapid evaluation of new advancements and improvements in the design process.

Modern heuristic methods have been shown to be successful tools for optimizing engine parameters in both simulation and experimental testing. Particle swarm optimization (PSO) and the genetic algorithm (GA) have both been applied successfully to engine optimizations.

• Particle swarm optimization, developed in 1995, is a population-based optimization algorithm in which potential solutions are evaluated and improved using information exchange to produce improved solution values. Advantages of the PSO method include rapid discovery of the approximate optimum, minimal objective evaluations when compared to other evolutionary algorithms, few tuning parameters, and relative simplistic methodology, noted Iowa State lead author Aaron Bertram in his 2014 thesis, which was the basis for the new work.

  Drawbacks of the general PSO method include its single-objective arrangement, a likelihood of premature convergence by clustering in a local optimum, the inability to make vast leaps to isolated regions of the feasible design space, and stagnancy in late stages.

• The genetic algorithm (GA), originally developed in the 1960s and 70s, came into commercial use in the late 1980s with the advent of more accessible commercial tools. The GA broadly is a set of functions inspired by natural evolution—i.e., incorporating mutation, inheritance, selection and crossover—to deliver optimal solutions. Beneficial traits are identified by their retention in the population as it evolves.

  The GA has been proven useful in determining optimal injection strategies in simulated experiments which revealed a prediction of PCCI combustion, Bertram noted. The ability of the GA to find regions of the design space which are not contiguous with the initial search space is one of the primary reasons that the GA is so useful when discovering new modes of combustion, in particular, HCCI, RCCI, PCCI, as well as alternatively fueled engines.

  However, the GA can suffer from scalability issues as well as demanding a large number of evaluations to converge. Other potential shortcomings of the GA include failure to converge, destructive mutation, and genetic stagnancy.

There have been numerous approaches to combining PSO and GA methods simultaneously in computational studies. Other hybrid approaches have also combined non-evolutionary methods with evolutionary ones.

PSO and GA hybrids are often applied in novel ways using problem specific knowledge. As identified by prior studies the PSO method advantage is primarily rapid convergence, especially in locally smooth regions. The GA advantages are primarily exploration and diversity, especially in convoluted and disjoint regions of feasible design space, and with highly complex response surfaces. Identifying which parameters are well suited to optimization by PSO and which are best suited to the GA requires engineering intuition, general problem-specific background knowledge, apparatus-specific knowledge, preliminary studies, or a substantial quantity of a priori decisions derived from prior explorations or research.

This study constructed a PSO-GA stepwise hybrid using a PSO “step” followed by a GA “generation” in order to exploit a settling time difference in the test apparatus. No function evaluation took place between the two different steps so the current objective value of the “parent” particle is unknown when compared to the current generation. In this evaluation, however, the parent particle location is tested to demonstrate the type of improvements expected.

—Bertram (2014)

For the IJER study, engine testing was performed under steady-state conditions at 1400 rpm at 4.15 bar brake mean effective pressure. The basic PSO and the hybrid PSO-GA method were applied to the test apparatus and used to locate the optimum neighborhood of the engine operation. The team used a single-objective function representing nox, particulate matter, hydrocarbon, CO, and fuel consumption.

Resources

• Aaron M Bertram, Qiang Zhang, Song-Charng Kong (2015) “A novel particle swarm and genetic algorithm hybrid method for diesel engine performance optimization” International Journal of Engine Research doi: 10.1177/1468087415611031

• Sahnehsaraei, M. Andalib et al. “A Hybrid Global Optimization Algorithm: Particle Swarm Optimization in Association with a Genetic Algorithm.” In Complex System Modelling and Control Through Intelligent Soft Computations doi: 10.1007/978-3-319-12883-2_2

• Aaron Bertram (2014) “A novel particle swarm and genetic algorithm hybrid method for improved heuristic optimization of diesel engine performance”

• S. Jeong, S. Obayashi, Y. Minemura (2008) “Application of hybrid evolutionary algorithms to low exhaust emission diesel engine design” Engineering Optimization Vol. 40, Iss. 1 doi: 10.1080/03052150701561155