Riassunto analitico
In the present work a complete procedure concerning the optimization of fuel surrogates’ composition for high performance applications is presented.
Starting from the analysis of kinetic schemes governing chemical reactions involving different species it has been possible to investigate their characteristics and behaviour in terms of autoignition and detonation resistance. The whole procedure has been carried out thanks to Cantera, a bunch of object-oriented software tools available in Python and specifically suited for problems regarding chemical kinetics and thermodynamics. The selected schemes have been validated comparing obtained results with data coming from literature and concerning single species or mixtures with the same composition of the ones tested.
Machine Learning (ML) algorithms have been introduced to predict autoignition time of novel blends obtained imposing limits on the percentage of each specie class (Aromatics, Olefines, Ethanol) following the ongoing legislation. Different ML techniques, as for example ensemble methods (AdaBoost) or Neural Networks (NN), have been tested and compared to determine the one more appropriate for the specific purpose of this work. Optimization procedures involving hyperparameters of each algorithm have been performed.
Then, different methodologies to predict Research Octane Number (RON) for different mixtures have been presented, taking advantage from the previously introduced ML algorithms and coupling them to the Livengood and Wu correlation, providing an estimation of knock occurrence. A simplified model of a CFR (Cooperative Fuel Research) engine has been introduced to simulate real experiments for the determination of RON, following a precise standardization.
Subsequently, a zero-dimensional engine model has been introduced, gradually updated to take into account more and more complex phenomena occurring in real engines and allowing to obtain more precise results. In particular, this model provides the most relevant engine-related parameters (pressure and fresh mixture temperature) that are then fed into the neural network and allow to determine knock occurrence for a specific gasoline blend. The results obtained with this method are then compared to the ones coming from a chemical reactor model, to better understand if the simplification hypothesis introduced could be considered fair.
Finally, an optimization procedure has been performed to find the best possible surrogate composition in terms of percentage of each chemical specie available. To do so, it has been introduced a function which emulates an engine cycle, providing as output a target parameter specifically selected for this purpose, and taking as input the various percentages of chemical groups and single species introduced before, also in this case taking care to respect the limits reported in the ongoing legislation. Moreover, the function also considers eventual detonation occurrence, modifying the spark timing in order to provide the highest possible performances without falling into knock conditions.
|
