Excessive consumption of fossil fuels, as well as the greenhouse gas issue, is one of the major challenges facing the world.
This issue has led to some efforts to achieve cleaner and more efficient internal combustion engines beside the new combustion strategies such as HCCI combustion strategy.
Since in the HCCI engines, the air-fuel mixture is homogeneous, particular matters are much less than those of conventional engines.
Also, because of the lower in-cylinder temperature in these engines, due to the lean mixture, the produced nitrogen oxides are not considerable.
The use of various fuels in HCCI combustion creates opportunities for HCCI engine applications under low to high load operations.
Although numerous merits have been observed by HCCI, stabilizing HCCI combustion between misfiring and knock criteria in a wide range of speed/load points is known as a significant challenge due to High Heat Release Rate (HRR) and lack of direct capability to control the start of combustion (SOC).
Combustion phasing or ignition timing is a critical factor in determining engine operating characteristics such as power density, HC emission, pressure rise rate and consequently HRR.
Since there is no direct operator for the start of the combustion, homogeneous air-fuel mixture properties entering the combustion chamber, determines the ignition timing in HCCI engines.
To control SOC, first, there should be an accurate model to predict the SOC based on the entering air-fuel mixture properties.
Later than, a controller must be designed to ensure the ignition timing is within a suitable range.
It makes the controller more effective having an accurate model for HCCI engine performance prediction.
In fact, instead of testing the engine and obtaining experimental data in the laboratory to survey the behavior of the engine and to achieve its optimum operating conditions, the obtained validated model can be used as a virtual engine and a platform for collecting the required data.
Numerical investigations include detailed thermo-kinetic models (TKM) were used in subjects like analyzing the influence of Exhaust Gas Recirculation (EGR) on the combustion timing, and combustion duration of a biogas-fueled HCCI engine, as well as parametric study of ignition delay prediction, and HC and NOx emissions characteristics of ethanol-fueled HCCI engine,.
Also, TKM models were used to investigate the effect of various engine parameters on the combustion phasing and duration likely the phasing of the Low-Temperature Reaction (LTR) and Main Combustion Stage (MCS) in an HCCI engine fueled with n-heptane.
Study the knock phenomenon during HCCI mode, based on Large Eddy Simulation (LES) coupled with methanol chemical kinetics was the other application of TKM to control the SOC of HCCI engines.
TKM also was used to simulate the in-cylinder pressure, peak pressure (Pmax), gross Indicated Mean Effective Pressure (gross IMEP), cumulative heat release (CHR), CO, NO and UHC emissions, and the crank angle when 50% of heat is released (CA50) by Poorghasemi et al.
These types of models (i.e., TKM) are relatively accurate, but have high computational cost, and require computational resources which are not suitable for real-time engine control.
Another group of models is physics-based models.
These models are based on empirical equations that are obtained from experimental data of the engine.
Control-oriented model and Mean Value Method (MVM) can be assumed as a subset of these types of models.
These experimental-based approaches were used for prediction of the combustion phasing investigation of engine performance, like in-cylinder pressure, pressure rise rate (HRR) and CO, HC and NOx emissions as well as operating range determination, IMEP variation analysis, ringing intensity analysis and combustion efficiency.
These types of investigation require substantial experimental data and the conventional control-oriented models which utilize the empirical relations and equations, lack the accuracy for predicting the parameters that are out of experimental data.
On the other hand, physics-based models are not able to predict some complicated phenomena such as emissions.
ANN as a powerful tool of artificial intelligence (AI) is used widely to predict different engine operating conditions of spark ignition diesel and HCCI engines.
Janakiraman et al. developed a multi-input single-output model preceded by a principal component analysis (PCA) to predict IMEP, combustion phasing, maximum in-cylinder pressure-rise rate, and equivalent air-fuel ratio.
A combination of experimental data analysis and Generalized Regression Neural Network (GRNN) modeling were used to evaluate and predict the Brake Thermal Efficiency (BTE), Exhaust Gas Temperature (EGT), and the emission such as unburned hydrocarbons (UHC), carbon monoxide (CO), nitric oxides (NOx), and smoke opacity by Bendu et al..
A hybrid GRNN–particle swarm optimization (PSO) model was designed to optimize three input parameters, including the charge temperature, engine load, and EGR rate by Bendu et al..
Bahri et al. used experimental results to design an ANN model to predict Combustion Noise Level (CNL) for identifying normal and ringing regions.
Syed et al. developed a robust ANN model with smaller data set generated from the experiment, which can efficiently predict the performance and emission characteristics of H2 dual fueled diesel engine.ANN models can also be used to optimize engine operating parameters for the best performance.
GA is applied broadly for optimization purposes.
An optimization method based on ANN and GA was investigated to optimize all geometric parameters and operating variables of the range-extended engine under a single operating point considering the constraints of operating points and efficiencies of target generators by Zhao et al..
Previous researches showed that the ANN which is optimized by GA could not only get better network configurations but also improve the efficiency and stability of prediction and also prediction error of the ANN model with using the GA is lower than that without using GA.
Radial Basis Function Neural Network (RBFNN) were developed, and GA was used to optimize the ANN to predict the performance and emissions of the HCCI engine by Anarghya et al.
and found that ANN-GA optimized model was more efficient and estimated the values close to the experimental results.
As thoroughly discussed above, even though ANN, GA, and a combination of them have been used in the study of HCCI engines, an ANN-GA approach which predicts the SOC of HCCI, based on the mixture properties, which consider delays for inputs and target of the network, has not been investigated.
Therefore, the main objective of the present investigation is to develop an optimized multi-input model which predicts the SOC of the HCCI engine based on entering mixture properties.
The developed optimized model can be used as a control-oriented model, which is used to design a controller, particularly to design a Model-Based Neural Network Controller, in this case.
For achieving this aim, three popular architectures namely the Nonlinear Autoregressive Network with Exogenous Inputs (NARXNET), Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) were designed, trained and compared with respect to architecture, prediction accuracy and simulation time and costs for the transient HCCI modeling.
Appropriate delays for inputs and target are considered in the networks architectures to simulate the dynamic behavior of the engine.
The engine model developed using experimental data taken from a one-cylinder Ricardo engine.
The parameters of the networks were optimized using a GA method.
As an alternative to numerical optimization methods, GA is less likely to get trapped in local minima and is not restricted by continuity or differentiability requirements on the cost function.
In the optimization approach, the ANNs were assumed as the objective function that is used to optimize their parameters of the network within a GA optimization.
Using GA ensures that the resulted network has the best architecture which is the number of layers and neurons of each layer and the likely delays of the networks.
In this paper, ANN was used for estimation the SOC of a Ricardo HCCI engine. At first, MLP, RBF, and NARX were trained and validated over a wide range of HCCI engines operations. Results show that the MLP and RBF networks can estimate the SOC fairly accurate and the result of NARX shows estimation with<1% error. GA was then used in order to optimize the parameters of the presented networks.
By comparing the results, the summary of the results of the current study is presented in the following:
• GA method is a powerful approach which seems to be very useful for enhancing neural networks performance by optimizing their architecture parameters, which is used to optimize RBF, MLP and NARX neural networks in this investigation.
• The architecture of the MLP with the best performance is a three-layer MLP with 25, 20 and 20 neurons in the first, second and third hidden layer respectively and one delay for the inputs and one and two delays for the feed backed outputs. With this architecture, a precise prediction of the SOC of HCCI can be possible.
• It was observed that the accuracy of the MLP network usually increases by adding more neurons and layers but necessarily there is no direct relationship between them and there are some networks with more neurons in their layers which had smaller values of R for the test data.
• The optimal values for the number of neurons and bandwidth of Gaussian kernel for the RBF network are 0.2 and 800, respectively.
• The NARX with the best performance was a two-layer network with 10 and 5 neurons in the first and second layers respectively, 0, 1 and 2 delays for the inputs and 1 and 2 delays for the feed backed targets.
• It can be said that the NARX network optimized with the GA has the best performance and the least computational cost among all of the presented networks and is the best choice for estimating the SOC.
This work can be extended by optimization of the weights in different layers of the network using GA and use the GA to optimize the learning process of the network or use GA as an alternative learning technique in place of gradient descent methods. This will enable the learning and training process of the network to escape from entrapment in local minima.
Reference: M. Taghavi, A. Gharehghani, F.B. Nejad, M. Mirsalim, Developing a model to predict the start of combustion in HCCI engine using ANN-GA approach, Energy Convers. Manag. (2019) 57–69. doi:10.1016/j.enconman.2019.05.015.