Addressing the Challenges of Advanced Compression Ignition Strategies Using Optimization Techniques with Machine Learning

Kavuri, N. K. C. Addressing the Challenges of Advanced Compression Ignition Strategies Using Optimization Techniques With Machine Learning. University of Wisconsin-Madison, 2018.

Advanced compression ignition strategies like reactivity controlled compression ignition (RCCI) and gasoline compression ignition (GCI) have received substantial interest over the past few years. This is due to their potential to achieve reduced emissions, and higher efficiency, relative to conventional diesel combustion. However, most of the benefits seen in past research from these strategies were demonstrated under mid-load conditions. For these strategies to be implemented practically, similar benefits must be demonstrated across the drive cycle. Two particularly challenging areas of operation are high-load-low- speed and low-load-high-speed. Very limited research has been done with advanced compression ignition strategies in these points of the engine operating map. The reason for this is, at these operating conditions, there exists a mismatch between engine and chemistry time scales. The time scale mismatch results in either increased pressure rise rates or high levels of incomplete combustion, both of which make it difficult to operate. The work presented in this dissertation attempts to fill in these research gaps by using a combination of computational fluid dynamics modeling and genetic algorithm optimization.
Initially, targeting high-load-low-speed conditions, a computational optimization study was performed at 20 bar indicated mean effective pressure and 1300 rev/min. with RCCI and GCI combustion strategies. The study was performed on a low compression ratio (12:1) piston with a ?bathtub? geometry, since it was found to be well suited for high-load operation in earlier studies. The optima from the two combustion strategies were compared in terms of combustion characteristics, combustion control, and sensitivity to operating parameter variations. The results showed that both the strategies have similar combustion characteristics, including a two-stage heat release. A near top dead center injection initiated the combustion and its injection timing could be used to control the combustion phasing for both the strategies. Both the strategies required elevated levels of exhaust gas recirculation (EGR) (~55%) at a near stoichiometric global equivalence ratio to control the peak pressure rise rate. This resulted in high sensitivity to variations in EGR. To address this issue, high-load strategies at reduced EGR levels were investigated.
A constraint analysis was performed using the optimization data to identify the constraints preventing operation at lower EGR levels. Results showed that operation at lower EGR rates was
constrained by NOx emissions. Relaxing the NOx constraint enabled lower EGR operation with significant efficiency improvement. Allowing NOx emissions to increase to acceptable levels for selective catalytic reduction after treatment yielded an optimum at a moderate (~45%) level of EGR and a globally lean equivalence ratio of 0.8. This optimum case had near zero soot emissions and a higher net fluid efficiency (which accounted for the pumping loop work and the diesel exhaust fluid mass required to reduce the NOx emissions) compared to the earlier high EGR optima. Furthermore, the optimum case with NOx aftertreatment was compared with the high EGR optima in terms of combustion control and stability to operating condition fluctuations. The optimum with NOx aftertreatment retained the excellent combustion control seen with the high EGR optima, while reducing the sensitivity to operating parameter variations. The improved stability was attributed to operation at a reduced global equivalence ratio (from 0.93 to 0.8), which decreased the sensitivity to fluctuations in EGR rate.
After addressing the issues at the high-load-low-speed operating condition, a low-load-high-speed operating point of 2 bar and 1800 rev/min. was simulated on the same engine used for the high-load studies. The results showed poor thermal efficiency for the low-load point. The poor efficiency was found to be due to an elevated level of incomplete combustion, which was a result of the low compression ratio piston used for the study. This result suggested that an optimum compression ratio should be identified considering the performance at the low-load and high-load conditions simultaneously. In addition, past optimization studies performed at low-load conditions have shown that the optimum bowl and injector design are very different compared to the high-load conditions. Accordingly, an optimization study was performed, considering performance at low- and high-load simultaneously. The optimum from the study was a stepped bowl geometry, with a compression ratio of 13.1:1, which resulted in a gross indicated efficiency of ~46% at both the loads. The study showed that the optimum design obtained from prioritizing one load deteriorates the performance at the other load. The results highlight the importance of considering multiple modes of the drive cycle simultaneously, when optimizing the engine design for advanced combustion strategies.

It was shown that multiple modes of the drive cycle should be considered in optimization studies for advanced combustion strategies; however, the optimization with just two operating points took three months to complete. To consider all the modes of a drive cycle in the optimization, the computational time must be reduced. To address this issue, machine learning through Gaussian process regression was coupled with a genetic algorithm optimization to speed up the optimization process. Including machine learning within the optimization process reduced the computational time of optimization by 62%. The optimization process was further improved by using the Gaussian process regression model to check for the sensitivity of the designs to operating parameter variations during the optimization. The approach was tested with existing optimization data and it was shown that adding the stability check resulted in a reliable and stable optimum solution.