Joint use of artificial neural networks and particle swarm optimization to determine optimal performance of an ethanol SI engine operating with negative valve overlap strategy

Abstract

A SI engine fueled with anhydrous ethanol and using NVO was modeled by ANN and hybridization of ANNs with PSO with two standard training methods: Levenberg-Marquardt and Bayesian Regularization. Engine testing results were obtained from a single cylinder Ricardo Hydra camless research engine equipped with a side mounted direct injector. NVO was achieved through early closing of exhaust valves and late opening of intake valves, while engine load was controlled through early intake valve closure. Results showed that both the standard ANN and the ANNs with PSO could accurately predict the outputs, with R2 above 0.93 and mean relative error below 12% for almost all networks. PSO-based ANNs showed an advantage in the prediction of COVIMEP and NOX emissions in case of few neurons in the hidden layer. Input and output mapping of the best ANN showed that SI engine higher indicated efficiencies could be achieved with more advanced spark timings and longer NVO periods in each load block tested. Additionally, it was possible to find an engine part load operating condition around 8.5 bar IMEP where NOX emissions would decrease whereas CO and unburned hydrocarbons emissions would not significantly change, with indicated engine efficiency higher than those found experimentally.