Prediction of Emission Characteristics of Generator Engine with Selective Catalytic Reduction Using Artificial Intelligence

Abstract

Eco-friendliness is an important global issue, and the maritime field is no exception. Predicting the composition of exhaust gases emitted by ship engines will be of consequence in this respect. Therefore, in this study, exhaust gas data were collected from the generator engine of a real ship along with engine-related data to predict emission characteristics. This is because installing an emission gas analyzer on a ship has substantial economic burden, and, even if it is installed, the accuracy can be increased by a virtual sensor. Furthermore, data were obtained with and without operating the SCR (often mounted on ships to reduce NOx), which is a crucial facility to satisfy environment regulation. In this study, four types of datasets were created by adding cooling and electrical-related variables to the basic engine dataset to check whether it improves model performance or not; each of these datasets consisted of 15 to 26 variables as inputs. CO₂ (%), NOx (ppm), and tEx (°C) were predicted from each dataset using an artificial neural network (ANN) model and a support vector machine (SVM) model with optimal hyperparameters selected by trial and error. The results confirmed that the SVM model performed better on smaller datasets, such as the one used in this study compared to the ANN model. Moreover, the dataset type, DaCE, which had both cooling and electrical-related variables added to the basic engine dataset, yielded the best overall prediction performance. When the performance of the SVM model was measured using the test data of a DaCE on both no-SCR mode and SCR mode, the RMSE (R²) of CO₂ was between 0.1137% (0.8119) and 0.0912% (0.8975), the RMSE (R²) of NOx was between 17.1088 ppm (0.9643) and 13.6775 ppm (0.9776), and the RMSE (R²) of tEx was between 4.5839 °C (0.8754) and 1.5688 °C (0.9392).

1. Introduction

At the beginning of 2021, 99,800 ships were sailing worldwide on seas to transport logistics and energy; this was an increase of 3% compared to January 2020. From 1970 to 2020, international maritime trade steadily increased. Most of the exhaust from a ship is generated from the marine main engine, marine generator engine (G/E), and boiler. The volume of greenhouse gases (GHG), including carbon dioxide (CO₂), methane, and nitrogen oxides (NOx), increased from 977 million tons in 2012 to 1076 million tons in 2018. Moreover, the share of ships in the global anthropogenic emissions increased from 2.76% in 2012 to 2.89% in 2018 [1,2]. According to the study, the allometric relationship between global fleet size and NOx from shipping was observed to be positive [3]. Therefore, the amount of NOx will be increased gradually, and adverse effects cannot be ignored. NOx in GHG affects human health and the environment, in addition to depleting the ozone layer and causing acid rain [4]. Furthermore, shipping emissions contribute to 1–14% of PM and 7–24% of NOx in European coastal areas [5]. Accordingly, the International Maritime Organization (IMO) imposes sanctions mandating the reduction in NOx, and emission control areas have been defined in several places. Therefore, ships have started implementing fuel–water emulsion, exhaust gas recirculation (EGR), selective catalytic reduction (SCR), and common rail fuel injection to reduce PM and NOx. Alternative fuels such as biodiesel, methanol, and LNG are another option to reduce emissions and ammonia, as well as hydrogen, which are considered as potential future fuels to meet the regulations [6,7]. Even if such eco-friendly technology is applied, if emission is not monitored, it is impossible to know whether environmental regulations are satisfied or whether the equipment is operating normally. Therefore, emission measurement is important in the mid- to long-term management aspect.

Compared to a ship, the power plant has a small number of funnels per unit of energy generated; hence, expensive equipment, such as a continuous emission monitoring system (CEMS) can be installed to check the emissions. However, even when such equipment is installed, an adequate measurement accuracy cannot be guaranteed owing to problems [8] such as low flow rates, leaks, sample condensation, different lengths of the flue gas sampling pipeline, and possibility of damage to or rapid aging of the equipment [9,10]. Therefore, machine learning (ML) studies using an artificial neural network (ANN) model [9], attention mechanism long short-term memory model [10], and least squares support vector machine (LS-SVM) model [9,11] were conducted to compensate for the drawbacks of power plant CEMSs in performing emission predictions.

In recent years, artificial intelligence (AI) has been applied to diverse fields, including maritime applications. Studies related to marine diesel engines have mainly focused on fault detection and diagnosis [12,13,14], performance analysis and prediction [15,16], and optimization [17,18]. Studies on other maritime machinery include prediction and monitoring of machinery system conditions [19,20], energy management systems [21], and fuel consumption [22].

In this study, we created an emission prediction model of a G/E because it is very important, as it is always running whether the ship is sailing or moored in the port. The pertinent literature on using ML to predict the emissions generated by a diesel engine was explored and is summarized in Table 1. However, ML studies on G/E emissions are very limited; furthermore, most of the studies are related to diesel engine emissions from vehicles or emissions of single-cylinder diesel engines at laboratory scales. Therefore, this study can be considered necessary and valuable for predicting maritime emissions of a generator engine. Moreover, considering that industrial generator engines used in various places such as power plants and buildings as emergency generators are the same type as the G/E, this study can be applied to several fields of research.

According to previous studies, it can be seen that ANNs, which are currently in vogue, were used for emission predictions, followed by support vector machines (SVMs). It can be estimated that the SVM, which has a great advantage in solving small sample regression problems, was used in many studies because it was not possible to obtain long-term big data by installing emission measurement devices in laboratory experiments or on cars [23].

Among those mentioned in Table 1, four studies [24,25,26,27] predicted emissions using nonconventional fuel oil for diesel engines. However, in this study, marine gas oil (MGO), which is widely used according to the 2020 IMO sulfur content regulation, was used. Similar to this study, the authors of [28] performed predictions on the performance and emissions of a diesel electric generator (DG). However, only an ANN model was used, with the number of input features limited to three, which might not be sufficient to reflect various characteristics of the DG. In addition, only carbon-based emission was predicted, and hyperparameter tuning, a process for finding an optimal model, was not implemented. Another study [29], similar to the above, predicted the performance and emissions using data generated from a verified thermodynamic model of a G/E. However, in this study, NOx and Soot were predicted for emissions, but not for carbon-based emissions. In addition, four input features and simulation data were used instead of the actual data. This method has drawbacks in terms of time and energy required to generate simulation data for general application to numerous ships. Therefore, an approach that can be applied to all kinds of ships in a simple way is required.

Table 1. Reviewed articles related to engine exhaust emissions using AI.

Authors	Research Content	Field
[30]	-Predicted the exhaust gas temperature and compared it with that from four different algorithms, namely, the ANN, random forest, SVM, and gradient boosting regression trees.	Not clearly stated
[23]	-Studied the effects of the model parameters and the training sample size on the prediction accuracy of the SVM regression model for an HCNG engine.	Vehicle
[24]	-Reviewed engine modeling based on statistical and ML methodologies through response surface and ANN techniques for various alternative fuels in both SI and CI engines.	Not clearly stated
[25]	-Built an ANN to predict the engine performance and emission characteristics for different injection timings, using waste cooking oil as a biodiesel blended with diesel.	Not clearly stated
[26]	-Analyzed the performance and the emissions of a four-stroke SI engine operating on ethanol–gasoline blends with the aid of ANN.	Vehicle
[27]	-Investigated ANN to predict the SI engine performance and exhaust emissions for methanol and gasoline.	Vehicle
[28]	-Modeled an ANN to predict CO2, CO/CO2 ratio, flue gas temperature, and gross efficiency in three-phase, 415 V, DG sets of different capacities operated at different loads, speeds, and torques.	Industry
[29]	-Investigated ANN and SVM based on Taguchi orthogonal array owing to the availability of limited experimental data of CRDI-assisted G/E for emissions prediction.	Maritime
[31]	-Established a NOx emissions prediction model of a diesel engine for both steady and transient operating states with an ensemble method based on principal component analysis, genetic algorithm, and SVM.	Vehicle
[32]	-Utilized an ANN algorithm with engine speed and load as the model inputs, and fuel consumption and emission as the model outputs.	Vehicle
	-Compared experimentally measured data and model predictions for forecasting engine efficiency and emissions.
[33]	-Modeled performance and emission parameters of single-cylinder four-stroke CRDI engine coupled with EGR by GEP and compared the results with those from an ANN model.	Not clearly stated
[34]	-Explored the potential of ANN to predict the performance and emissions with load, fuel injection pressure, EGR, and fuel injected per cycle as input data for a single-cylinder four-stroke CRDI engine under varying EGR strategies.	Not clearly stated

Table 1. Reviewed articles related to engine exhaust emissions using AI.

Authors	Research Content	Field
[30]	-Predicted the exhaust gas temperature and compared it with that from four different algorithms, namely, the ANN, random forest, SVM, and gradient boosting regression trees.	Not clearly stated
[23]	-Studied the effects of the model parameters and the training sample size on the prediction accuracy of the SVM regression model for an HCNG engine.	Vehicle
[24]	-Reviewed engine modeling based on statistical and ML methodologies through response surface and ANN techniques for various alternative fuels in both SI and CI engines.	Not clearly stated
[25]	-Built an ANN to predict the engine performance and emission characteristics for different injection timings, using waste cooking oil as a biodiesel blended with diesel.	Not clearly stated
[26]	-Analyzed the performance and the emissions of a four-stroke SI engine operating on ethanol–gasoline blends with the aid of ANN.	Vehicle
[27]	-Investigated ANN to predict the SI engine performance and exhaust emissions for methanol and gasoline.	Vehicle
[28]	-Modeled an ANN to predict CO2, CO/CO2 ratio, flue gas temperature, and gross efficiency in three-phase, 415 V, DG sets of different capacities operated at different loads, speeds, and torques.	Industry
[29]	-Investigated ANN and SVM based on Taguchi orthogonal array owing to the availability of limited experimental data of CRDI-assisted G/E for emissions prediction.	Maritime
[31]	-Established a NOx emissions prediction model of a diesel engine for both steady and transient operating states with an ensemble method based on principal component analysis, genetic algorithm, and SVM.	Vehicle
[32]	-Utilized an ANN algorithm with engine speed and load as the model inputs, and fuel consumption and emission as the model outputs.	Vehicle
	-Compared experimentally measured data and model predictions for forecasting engine efficiency and emissions.
[33]	-Modeled performance and emission parameters of single-cylinder four-stroke CRDI engine coupled with EGR by GEP and compared the results with those from an ANN model.	Not clearly stated
[34]	-Explored the potential of ANN to predict the performance and emissions with load, fuel injection pressure, EGR, and fuel injected per cycle as input data for a single-cylinder four-stroke CRDI engine under varying EGR strategies.	Not clearly stated

Since the installation of a marine emission analyzer on ships is not obligatory, shipping companies do not use it while paying a high cost; however, even if it were installed on the ship ignoring the cost aspects, ML methods would have to be applied to increase the prediction accuracy of the analyzer. Furthermore, emission measurements entail significant time and cost, and they suffer from disadvantages, such as a harmful effect on the environment during the experiments. Therefore, it would be ideal to measure the emissions during the period of preparing the performance report for each load of the G/E, which is routinely performed to check the condition of the G/E by engineer. Assuming that the ship is not equipped with an expensive emission analyzer for engine, an emission prediction model can be created and distributed when the ship is built as a virtual sensor, and it would be desirable to update this model with regular measurements during the dry-dock period or when the ship is berthed at a port. This method can be easily implemented in next-generation smart generator engines equipped with data collection devices and is expected to be widely used as a measure to respond to environmental regulations. Therefore, in this study, we aimed to verify whether the aforementioned aspects can be leveraged by measuring the emissions while the G/E performance report is being compiled; moreover, we tested the model on data measured on a different day (after and before turbocharger overhauling).

In this study, we attempted to create models that could predict emissions with the SCR (an eco-friendly technology for reducing the NOx emissions). To the best of our understanding, studies related to G/Es with the SCR to predict emissions by ML are scarce or nonexistent; thus, this research is meaningful and valuable for future studies on emission predictions of generator engines. A novel aspect of this study involves improving upon the currently existing results by using and comparing ANN and SVM models; furthermore, 15 to 26 input features acquired from the ship’s database were directly used for datasets. Additionally, the prediction results of the models were compared using datasets considering not only the variables of the G/E, but also the cooling and electrical-related variables. Lastly, we tried to prove the superiority of the DaCE dataset, which added cooling and electrical-related variables to the existing G/E dataset, through performance evaluation in ANN and SVM models.

2. Materials and Methods

2.1. Description of Machinery and System

2.1.1. G/E and Cooling System

The specifications of the G/E used in this study are listed in Table 2. The G/E is a critical component that generates all the electricity required for various machines and the occupants of a ship. The G/E receives oil from a fuel tank via a pump and sprays it through a set of nozzles into the engine cylinders. The fuel, thus atomized, mixes with the air supplied by a T/C and generates power from combustion; this power is then transmitted through the reciprocating motion of a piston. With this power, the crankshaft rotates, and the rotor of the generator connected to the crankshaft also rotates to produce electricity. The turbine wheel of the T/C receives the energy of the exhaust gas generated from G/E and rotates the central shaft; thus, the compressor wheel of the T/C supplies air into the cylinder. An air cooler cooled by cooling water is installed between the T/C and engine cylinder, thus increasing the power output, reducing smoke, and improving the fuel consumption [35]. However, the viscosity of the LO, whose purpose is to lubricate the engine is lowered by the heat generated from the engine; therefore, it passes through a lube oil cooler. In the past, sea water was supplied as a coolant to the air cooler and lube oil cooler. However, currently, most of the engines use a central cooling system, as it increases the operational window between the downtimes for cleaning the cooler; thus, essentially, only the central cooling heat exchanger needs to be cleaned. As shown on the bottom left side of Figure 1, the central cooling heat exchanger exchanges heat between the seawater and freshwater, and this CFW is supplied to the cooler of each machine; in the case of G/E, it is supplied to the air cooler and lube oil cooler.

Table 2. Engine specifications.

Type of Engine	4-Stroke, Vertical, Direct Injection Single-Acting, and Trunk Piston Type with T/C and Intercooler
Cylinder configuration	Inline
Number of cylinders	6
Rated speed	900 rpm
Power per cylinder	200 kW
Cylinder bore	210 mm
Piston stroke	320 mm
Swept volume per cylinder	11.1 dm3
Mean piston speed	9.6 m/s
Mean effective pressure	24.1 bar
Compression ratio	17:1
Cylinder firing order	1–4–2–6–3–5

Figure 1. Schematic diagram of the experimental setup.

2.1.2. SCR

The specifications of the SCR used in this study are listed in Table 3. SCR is an eco-friendly technology that can reduce NOx, as can be confirmed from Figure 2. The figure shows that NOx was reduced by up to 749 ppm; moreover, it was further reduced after turbocharger overhaul compared to values before turbocharger overhaul. As shown in Figure 1, the SCR device consists of a urea supply unit, urea dosing unit, control unit, injection pipe, SCR chamber, and soot blowing unit. Urea solution is supplied from the urea tank to the urea dosing unit through the pump of the urea supply unit. The urea dosing unit supplies the urea solution to the injection pipe and is automatically activated by the control unit. Additionally, an automatic purging system is provided to clean the urea line between the urea dosing unit and the injection pipe when the SCR is stopped. The injection pipe atomizes the urea aqueous solution through the urea injection nozzle and static mixer and supplies it to the exhaust gas line. The SCR chamber consists of a catalyst layer. The urea solution is used as a reducing agent and is decomposed as given by Equation (1) into ammonia and CO₂ in the hot exhaust gas stream. Ammonia decomposed in the aqueous urea solution is converted into nitrogen molecules and water by a chemical reaction with NOx on the surface of the catalyst layer, as given by Equation (2). The soot blowing unit sprays compressed air on the surface of the catalyst layer to remove dust and soot particles from the engine combustion process to maintain the NOx reduction performance.

$${\left({\mathrm{NH}}_2\right)}_2\mathrm{CO}+{ \mathrm{H}}_2\mathrm{O} \to 2{\mathrm{NH}}_3+{\mathrm{CO}}_2$$

(1)

$$\left\{\begin{array}{c}4\mathrm{N}\mathrm{O}+4\mathrm{N}{\mathrm{H}}_3+{\mathrm{O}}_2 \to 4{\mathrm{N}}_2+6{\mathrm{H}}_2\mathrm{O}\\ 6\mathrm{N}{\mathrm{O}}_2+8\mathrm{N}{\mathrm{H}}_3 \to 7{\mathrm{N}}_2+12{\mathrm{H}}_2\mathrm{O}\end{array}\right.$$

(2)

Table 3. SCR specifications.

NOx emission value after the SCR	2.31 g/kWh
Pressure drop across the SCR	≤150 mmAq
Ammonia slip	≤10 ppm
Sulfur content of the fuel oil for SCR operation	≤0.1%
Maximum allowable exhaust gas temperature	≤400 °C

Figure 2. Comparison of NOx ppm levels between no-SCR and SCR modes.

2.2. Description of Workflow

2.2.1. Overview of the Training Sequence

In this study, as shown in Figure 3, ship data and exhaust gas data were acquired on two different days, with an interval of 7 days between their acquisitions. One was termed BTC, to be used as the test set, and the other was termed ATC to be used as the training/validation set. In most cases, the G/E performance is measured when the G/E load is maintained at 0%, 25%, 50% and 75%; in this study, the exhaust gas data were acquired as the load was increased from 25% to 50% and 75%; then, again after operating the SCR, the emissions were measured while decreasing the load from 75% to 50% and 25%. These data were preprocessed by merging and filtering. Thus, datasets for four cases as input data were created on the basis of different variables to predict CO₂, NOx, and tEx. The ATC was used for training until the loss functions of the ANN and SVM models were minimized through a hyperparameter tuning. Subsequently, the performances of the models were measured using the BTC. Because the machine’s performance worsens as the engine is driven, the models were trained with the good case and tested on the relatively bad case. If the training and test sets were made from a combination of BTC and ATC, there would be a high possibility that the information of the test set would be diluted in the training set. Therefore, we tried to evaluate the performance of the models with completely new and bad condition data collected before the T/C overhaul to obtain conditions similar to the case of developing and distributing a model at the time of shipbuilding, followed by predicting the emissions after some time passes by; this is the hypothesis that was established a priori. Python, a popular computer programming language in AI research, was used, and data analysis was performed on a Jupyter notebook. The open-source libraries used included Numpy, Pandas, Matplotlib, Seaborn, Scipy, Scikit-Learn, Tensorflow, and Keras. Numpy and Pandas were used for data analysis; Matplotlib and Seaborn were used for data visualization; Scikit-Learn was used for data analysis, preprocessing, metrics, and AI model learning; Scipy was also used for metrics; lastly, Tensorflow and Keras were used to create the ANN.

Figure 3. Flowchart for emission prediction modeling.

2.2.2. Data Acquisition

Both BTC and ATC were gathered using the method described in Section 2.2.1. As shown in Figure 1, the exhaust gas data were collected for each load, while creating the G/E performance report by attaching the gas sampling probe of the exhaust gas analyzer to the IMO flange at the top of the exhaust gas pipe of the G/E. The type of fuel used was MGO, and the fuel type for the initial setting of the exhaust gas analyzer was set by referring to the specifications listed in Table 4. After raising or lowering the load of the G/E for each stage, it was stabilized for approximately 10 min, and then the exhaust gas data were collected after every 1 s for 10 min. The collected data were then saved to the local computer as an Excel spreadsheet through the acquisition software after all the measurements were completed. Table 5 can be referred to for the accuracy of the exhaust gas analyzer. The manufacturer of the gas analyzer stated that the accuracy of flue gases satisfies the MARPOL Annex VI and NOx Technical Code. Appendix 4 (calibration of the analytical and measurement instruments) of the NOx Technical Code 2008 states that the true concentration of a calibration and span gas shall be within ±2% of the nominal value.

Table 4. Fuel specifications.

API gravity, 60 °F	35.6
Specific gravity, 15/4 °C	0.8464
Flash point	66.0 °C
Sulfur	0.0340 wt.%
Kinematic viscosity	2.8940 mm2/s
Net heat of combustion	10,220 kcal/kg
Gross heat of combustion	10,891 kcal/kg

Table 5. Accuracy of exhaust gas analyzer.

Flue gas CO2	±2% ppm
Flue gas NOx	±2% ppm
Exhaust gas temperature	±0.4 °C (−100 to +200.0 °C)
	±1.0 °C (200 to +1370.0 °C)

The ship data were automatically saved; therefore, after the measurements, the MS Access file was extracted from the database and converted into an Excel file. The data comprised 8640 samples and 1000 features, and it contained all the featured data of the ship over a 24 h period at 10 s intervals. Therefore, only the G/E-related features during the measurement were selected, and the 1000 features were compressed to 15–26 features.

2.2.3. Data Preprocessing and Dataset Generation

The variables used in this study and their ranges are summarized in Table 6. In the case of exhaust gas data, CO₂ (%), NOx (ppm), and funnel exhaust gas temperature (tEx, °C) of the G/E were selected as the values to be predicted. However, the values were unstable for approximately 20 s initially; therefore, only the data gathered subsequently were used. The ship data had a 10 s interval for each sample; therefore, the exhaust gas data were also arranged in 10 s intervals to coincide with the ship data, and these two were merged with each other to create the complete dataset. Among the G/E-related features, constant or unnecessary variables were deleted. In addition, each of the six cylinder exhaust gas temperature variables and each of the two T/C exhaust gas inlet temperature variables had a high correlation; therefore, it was difficult to determine the effect of those variables in isolation, and the significance of specific variables may be lost. Therefore, average values of the variables were created and substituted for the individual variables of the cylinder exhaust gas temperature and T/C exhaust gas inlet temperature. Lastly, a pure G/E dataset for the no-SCR mode, named Da-sx, was created. Here, sx means no-SCR mode.

Table 6. Parameters and their ranges for the emission prediction model.

Parameter	Range
Common data for all datasets
LO inlet temperature (°C)	63.98–66.96
Cylinder exhaust gas outlet temperature—average (°C)	386.07–440.86
FO inlet temperature (°C)	14.99–19.98
T/C exhaust gas inlet temperature—average (°C)	422.95–530.89
LO inlet pressure (kg/cm2)	4.65–4.89
T/C exhaust gas outlet temperature (°C)	357.9–408.89
FO inlet pressure (kg/cm2)	6–6.2
Exhaust gas compensation temperature (°C)	24.96–36
LO filter inlet pressure (kg/cm2)	5.1–5.3
Charge air outlet temperature (°C)	36–38
FO filter inlet pressure (kg/cm2)	6–6.6
Charge air inlet pressure (kg/cm2)	0.5–2.4
T/C LO inlet pressure (kg/cm2)	3–3.5
T/C rpm pick-up (rpm)	22,603.18–40,771.68
Engine rpm pick-up (rpm)	897–901.98
Additional data for DaC
LT CFW outlet temperature (°C)	34.97–40.98
LT CFW temperature difference (°C)	0–5.03
Central CFW cooler temperature difference (°C)	0.47–1.11
Additional data for DaE
Alternator winding phase temperature-average (°C)	37.75–65.64
G/E generator current (A)	336–1186
G/E generator power (kW)	242–791.7
G/E current phase-average (A)	331–1176.33
G/E bus net used power (kW)	341–864
G/E non-drive-end bearing temperature (°C)	39.18–54.09
Additional data for SCR mode
G/E load to G/E SCR (%)	23.05–75.32
Urea injection (l/h)	4.7–14.68
Prediction data
CO2 (%)	5.48–7.44
NOx (ppm)	51–1016
tEx (°C)	158.1–382.2

The CFW, after exchanging heat in the central cooler with the seawater, was supplied to the air cooler, and dense air was supplied into the cylinder, thus affecting the exhaust gas; therefore, three variables related to the CFW were selected to generate DaC-sx. The variable “LT CFW temperature difference” represents the difference between the inlet and outlet temperatures of the CFW flowing through the air cooler and lube oil cooler of the G/E, and “central CFW cooler temperature difference” represents the difference in the temperatures of the CFW between inlet and outlet of the central cooler.

Moreover, in addition to the mechanical process variables, the values of the electrical-related variables also change according to the G/E load, even though they are not directly affected by the exhaust gas. Therefore, six variables were selected to generate DaE-sx. In this dataset, two average variables were created to reduce redundancy: one was the average temperature of the R, S, and T phases of the alternator windings, and the other was the average G/E current in the R, S, and T phases. Lastly, DaCE-sx was created by adding both cooling and electrical-related variables to Da-sx.

To create the SCR mode dataset, only two variables “G/E load to G/E SCR” and “urea injection” were added to the set of variables of the no-SCR mode dataset; the designator “so” was added at the end of the name for the dataset instead of “sx”. As an example, the correlation heatmap of the dataset Da-sx, wherein the correlation value varied from –1 to 1, is shown in Figure 4. A closer correlation value to −1 indicates a more negative correlation, while a closer value to 1 indicates a more positive correlation. In this figure, the names of the variables are abbreviated; however, they are exactly the same as those described in the “common data for all datasets” in Table 6. From the heatmap, it can be seen that the target values of CO₂, NOx, and tEx were strongly positively correlated with the LO inlet temperature, cylinder exhaust gas out temperature—average, T/C exhaust gas in temperature—average, and charge air outlet temperature. Therefore, it can be assumed that the dataset DaC would help increase the model prediction performance, because the CFW of DaC affected the lube oil cooler and air cooler, consequently affecting the temperature of the LO, charge air, and exhaust gas. However, other pressure variables were inversely correlated with the target values, except for the charge air inlet pressure. This can be determined by considering that the target values increased as the engine load increased. As the load increased, the heat generated by the engine increased and the viscosity of the pressure variables decreased, thereby decreasing the pressure. Conversely, the T/C driven by the heat of the exhaust gas increased the charge air pressure, as the load increased.

Figure 4. Correlation heatmap of Da-sx.

After extracting 75% of the data for the training set and 25% of the data for the validation set from each load, i.e., 25%, 50%, and 75% for both the no-SCR mode and SCR mode for the ATC, a total of eight different types of datasets were created, as shown in Figure 3. They were combined to create 16 datasets including eight training sets and eight validation sets having an even distribution of the samples of each load. The training set was used to train the model, while the validation set was used to prevent overfitting of the trained model and tune the hyperparameters. Lastly, the performance of the model was evaluated with the test set. The reason for assigning a ratio of 75% to 25% to the training and validation sets is that this study used a small dataset, and, if the training set is used in a large proportion, it could lead to inconsistent results, as indicated in [36]. In addition, when dividing the training and validation sets to compare the performance of different datasets in the same model, the random seed was similarly set to have the same row of data. In the case of eight BTCs, this division into training and validation sets was not performed because the entire dataset was a test set. Before training the ANN and SVM, standard scaling, which is one of the normalization methods, was used to improve the learning performance of the model, and its formula is shown in Equation (3); here, the mean of all data was made zero and the variance was made 1 by subtracting the mean and scaling to unit variance.

$$z=\frac{x-\mu }{\sigma },$$

(3)

where x is the training sample, μ is the mean of the training samples, σ is the standard deviation of the training samples, and z is the standardized training sample.

3. Theory/Modeling

3.1. Artificial Neural Network

ANN is an ML model inspired by a BNN. The ANN structure was first introduced by neurophysiologist Warren McCulloch and mathematician Walter Pitts in 1943 [37]. Subsequently, the simplest ANN structure, called a perceptron, was proposed by Frank Rosenblatt in 1957 [38]. The description of the operation of an artificial neuron in a BN is shown in the upper right side of Figure 5.

Figure 5. ANN depicted as a BNN.

The BNN is composed of a cell body containing a nucleus, dendrite, axon, and axon terminals. The input data, multiplied by weights, are delivered to the dendrite, and a weighted sum of the input data with a bias added is transferred to an activation function in the cell body. The data then flow along the axon and are emitted from the axon terminals as output data. These output data are transferred as the input data to the dendrites of other neurons, and a set of these neurons is called an ANN.

However, the perceptron is composed of one layer and cannot solve problems, such as XOR; therefore, the MLP, in which several perceptrons are stacked, was proposed by Marvin Minsky and Seymour Papert [39]. The MLP is composed of an input layer, several hidden layers, and an output layer. It is also called a DNN, and its structure is shown in Figure 5. However, the initially proposed MLP suffered from a disadvantage in that it could not learn weights, while the backpropagation training algorithm introduced by David Rumelhart, Geoffrey Hinton, and Ronald Williams in 1986 could solve this problem [40]. The backpropagation is calculated in the forward direction from the input of the input layer to the output of the output layer, and it stores the intermediate calculated values for the backward calculation. Then, the error between the output and target values is calculated, and the error gradient is measured for the weight between the layers by applying the chain rule, which is called backpropagation, as it propagates from the output to the input layer. Finally, by performing a gradient descent, the weights are adjusted to reduce the error; this process is repeated several times until the cost function is minimized, and each of these iterations is named an epoch.

Several hyperparameters exist in the design of an ANN, e.g., the number of hidden layers, number of nodes in each layer, activation function applied to each node, kernel initializer for weight initialization, optimizer as the training method of the model, learning rate, batch size, and epochs. However, there is no single correct method for selecting the hyperparameters; therefore, a model was designed by trying various combinations of hyperparameters according to the characteristics of the data to find the optimal result. After several trial-and-error attempts, a model with a structure of 64–32–16–8–4–1, where each number represents the number of nodes in the hidden and output layers, was designed, as shown in Figure 6. There were three models for each dataset to predict CO₂, NOx, and tEx separately. As the activation function, Swish proposed by the Google Brain team, which is a variant of the ReLU, was used, and its formula is given by Equation (4) [41]. This function is calculated by multiplying the sigmoid by variable x, and it has a shape similar to that of ReLU.

Figure 6. Example of the designed ANN with a structure of 64–32–16–8–4–1 layers.

However, as it allows negative values it has the advantage of solving the problem of dying ReLU, which replaces negative values with zeros. To solve the problem of vanishing or exploding gradients, a kernel initializer was proposed [42]; among various available kernel initializers, He-uniform, in which the weights follow a uniform distribution within the range of Equation (5), was used to randomly initialize the weights of each layer. Here, fan_in is the number of input units in the weight tensor [43]. Nadam, which applies the Nesterov method to Adam, was used as an optimizer; it should be noted that Nadam updates the weights, as shown in Equation (6). A detailed explanation of the process involved is beyond the scope of this study, and readers are referred to [44] for additional details. In addition, because it is necessary to make predictions with new data, a dropout was applied to every hidden layer for generalization and preventing overfitting by setting the ratio to 10% [45]. The hyperparameters used in designing the ANN model are summarized in Table 7.

$$f\left(x\right)=\frac{x}{1+e^{-x}}.$$

(4)

$$\left[-limit, limit\right], \mathrm{where} limit=\sqrt{\frac{6}{fa{n}_{in}}}.$$

(5)

$${\theta }_t \leftarrow {\theta }_{t-1}-\frac{{\alpha }_t}{\sqrt{{\widehat{n}}_t+ϵ}}{\widehat{m}}_t.$$

(6)

Table 7. Hyperparameters for the ANN model.

Hyperparameter	Value
ANN structure	64–32–16–8–4–1
Activation function	Swish
Kernel initializer	He-uniform
Optimizer	Nadam
Learning rate for the optimizer	0.0001
Dropout rate	10%
Patience for early stopping	300
Epochs	3000
Batch size	16

3.2. SVM

The SVM was developed by Vapnik and colleagues at AT&T Bell Laboratories [46,47,48,49,50,51] and has been one of the most popular models in ML to date. The SVM can be used for both classification with SVC and regression with SVR, as illustrated in Figure 7. The SVM works well on small to medium-sized datasets and has been utilized in many studies owing to its good performance [52,53,54]. As this study also constructed a model with a small dataset, the SVM was used to build a model in this study. The basic concept involves the SVC learning to maximize the width of the road between the classes within the error margin, as well as the SVR learning to fit as many samples as possible in the road. The previously mentioned road is the sum of the margins, which represent the distances from the decision boundary to the support vectors. The SVM solves nonlinear problems through linear regression by mapping low-dimensional data to high-dimensional data using a kernel function. There are four representative kernel functions, as given by Equations (7)–(10), and, in this study, the commonly used RBF was utilized. Here, $x_i, x_j$ are the input vectors, and γ (gamma) is a hyperparameter that plays the role of regulation. It should decrease when the model is overfitted and increase when the model is underfitted. Assuming that $x_i$ is the input data, w is the weight vector, b is a scalar, and $f\left(x_i\right)$ is the output of the SVM in Equation (11), the SVM finds the optimal linear regression function by minimizing the loss function of Equation (12). ${\xi }_i^{+},{\xi }_i^{-}$ are slack variables for each sample, which determine how much margin is violated by the i-th sample; ε (epsilon) is the margin. However, a conflicting problem is faced with the loss function; one solution is to make $\frac{1}{2}{w}^{T}w$ so small as to increase the margin, and the other is to make ${\xi }_i^{+}+{\xi }_i^{-}$ so small as to minimize the margin error. Therefore, the sum of the slack variables is multiplied by hyperparameter C, which acts as a penalty factor [55]. There are several hyperparameters, such as C, ε, and γ, in SVM, because optimal hyperparameter values applicable to all the problems do not exist, and it is common to find hyperparameters with the best performance through a grid search process. In this study, C = 100, ε = 0.0005, and γ = 0.01 for the no-SCR mode, and C = 10, ε = 0.01, and γ = 0.002 for the SCR mode were selected to train the model. A detailed description of the SVM theory can be found in [56,57,58].

Figure 7. Support vector regression.

-Linear:

$$K\left(x_i, x_j\right)=x_i^T{x}_j.$$

(7)

-Polynomial:

$$K\left(x_i, x_j\right)={(\gamma x_i^T{x}_j+r)}^d, \gamma >0.$$

(8)

-RBF:

$$K\left(x_i, x_j\right)=\exp \left(-\gamma \Vert x_i-x_j\Vert {}^2\right), \gamma >0.$$

(9)

-Sigmoid:

$$K\left(x_i, x_j\right)=\tanh \left(\gamma x_i^T{x}_j+r\right), \gamma >0.$$

(10)

$$f\left(x_i\right)=w^T{x}_i+b.$$

(11)

$$\begin{gathered} \mathrm{Minimize} \frac{1}{2}{w}^{T}w+C{\displaystyle \sum }_{i=1}^n\left({\xi }_i^{+}+{\xi }_i^{-}\right), \\ \mathrm{subject} \mathrm{to} \left\{\begin{array}{c}-\epsilon -{\xi }_i^{-} \le y_i-f\left(x_i\right) \le \epsilon +{\xi }_i^{+}\\ {\xi }_i^{+},{\xi }_i^{-} \ge 0\end{array}\right.. \end{gathered}$$

(12)

3.3. Performance Measurement Metrics

In this study, to evaluate the performance of the models, the metrics RMSE, MAE, MAPE, R², and $r_{y\widehat{y}}$ were adopted, and their formulae are given by Equations (13), (14), (15), (16), and (17), respectively. When training the model, because the data were standardized in both ANN and SVM, the predicted values were inverse-transformed, as shown in Equation (18). Then, the model performance was assessed by comparing the predicted values (${\widehat{y}}_i$) and actual values ($y_i$) using the abovementioned metrics.

$$\mathrm{RMSE}=\sqrt{\frac{1}{n} {\displaystyle \sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}.$$

(13)

$$\mathrm{MAE}=\frac{1}{n} {\displaystyle \sum }_{i=1}^n\left|y_i-{\widehat{y}}_i\right|.$$

(14)

$$\mathrm{MAPE}=\frac{100\%}{n} {\displaystyle \sum }_{i=1}^n\left|\frac{y_i-{\widehat{y}}_i}{y_i}\right|.$$

(15)

$$R^2=1-\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-\overline{y}\right)}^2}.$$

(16)

$$r_{y\widehat{y}}=\frac{{{\displaystyle \sum }}_{i=1}^n\left(y_i-\overline{y}\right)\left({\widehat{y}}_i-\overline{\widehat{y}} \right)}{\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-\overline{y}\right)}^2}\sqrt{{{\displaystyle \sum }}_{i=1}^n{\left({\widehat{y}}_i-\overline{\widehat{y}} \right)}^2}}.$$

(17)

$$x=\left(z\times \sigma \right)+\mu .$$

(18)

4. Results and Discussion

4.1. Comparison of Emission Predictions of ANN and SVM Models for the No-SCR Mode

4.1.1. Comparison of Dataset Types for the No-SCR Mode

Da-sx was generated by merging the G/E data and exhaust gas data; DaC-sx and DaE-sx were generated by adding cooling and electrical-related variables to Da-sx, respectively; DaCE-sx was generated by adding both cooling and electric variables to Da-sx. In the case of ANN, the predictions were slightly different for each trained model owing to random weight initialization. Therefore, 15 instances of training were performed to create 15 different models for each emission characteristic, and evaluations with the validation set were performed for each model to check the influence of different datasets. Table 8 summarizes the results of each dataset for the 64–32–16–8–4–1 ANN model with optimal hyperparameters. The values shown in Table 8a are the averages of the calculated metrics tested from the validation set with 15 models trained on the training set. In the case of CO₂, it can be seen that the error was almost negligible, as reflected by the RMSE (MAE) value, which is less than 0.0461% (0.0338%); this is almost the same as that of the other tables. Therefore, in this study, NOx and tEx were the main concerns because the CO₂ prediction performance was confirmed with excellent results. For NOx, the RMSE (MAE) values decreased by 0.6159 ppm (0.6812 ppm), 1.7881 ppm (1.3073 ppm), and 1.2385 ppm (1.0419 ppm) for DaC-sx, DaE-sx, and DaCE-sx, respectively. In the case of tEx, unlike NOx, the RMSE (MAE) value in DaCE-sx compared to Da-sx yielded the largest error reduction of 1.5326 °C (1.084 °C), and DaC-sx and DaE-sx showed better performance than Da-sx. Therefore, it can be said that DaCE-sx was the best choice because the difference between DaCE-sx and DaE-sx was not large for CO₂ and NOx, as listed in Table 8a. In Table 8b, it can be seen that, when the cooling parameters were added to the test dataset BTC, the prediction performance of NOx and tEx was slightly decreased. However, when DaCE-sx was compared with Da-sx, the RMSE (MAE) decreased by 5.5982 ppm (4.1466 ppm), indicating that the NOx prediction performance was the best. In the case of CO₂ and tEx, the lowest RMSE (MAE) was obtained in with DaE-sx; however, it was slightly different from that of DaCE-sx. Therefore, DaCE-sx offered the best prediction with ANN in the no-SCR mode; the same can be confirmed with other metrics, such as MAPE, R², and Pearson r. Furthermore, the superiority of DaCE-sx can be checked by referring to Table 8c and Figure 8. In Figure 8, data points of DaCE-sx got closer to the best prediction line compared to others. Especially for tEx, it could be precisely confirmed that the prediction performance improved when cooling and electrical variables were added. The cooling and electrical variables added to the Da dataset changed in proportion to the engine load. In addition, the emissions and tEx also changed with engine load. Therefore, the DaCE dataset, which has a high correlation with the target variables, generally showed good performance.

Table 8. Metrics from different datasets for the ANN model with the no-SCR mode (a) trained 15 times from the ATC training set with average taken of the 15 metrics from the validation set; (b) 15 trained models tested from the BTC with average taken of the 15 metrics; (c) the best model for the validation set tested from the BTC and its metrics.

		Da-sx			DaC-sx			DaE-sx			DaCE-sx
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(a)	RMSE	0.0461	13.5937	5.5857	0.0460	12.9778	5.1007	0.0425	11.8056	5.2933	0.0449	12.3552	4.0531
	MAE	0.0338	10.9727	4.0019	0.0335	10.2915	3.7966	0.0313	9.6654	3.6559	0.0325	9.9308	2.9179
	MAPE	0.5474	1.3560	1.8322	0.5434	1.2845	1.7303	0.5071	1.1913	1.7301	0.5259	1.2404	1.3278
	R2	0.9424	0.9853	0.9918	0.9428	0.9860	0.9923	0.9510	0.9891	0.9915	0.9453	0.9878	0.9955
	Pearson r	0.9735	0.9950	0.9977	0.9740	0.9960	0.9980	0.9769	0.9961	0.9986	0.9755	0.9966	0.9989
		Da-sx-test			DaC-sx-test			DaE-sx-test			DaCE-sx-test
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(b)	RMSE	0.1208	25.8067	6.7677	0.1148	26.8612	7.0220	0.1010	21.3484	5.1222	0.1093	20.2085	5.5835
	MAE	0.0915	20.1967	5.2798	0.0883	20.5117	5.2452	0.0756	16.1402	4.1195	0.0841	16.0501	4.4798
	MAPE	1.4360	2.3328	1.5056	1.3834	2.3573	1.4881	1.1810	1.8805	1.1692	1.3139	1.8724	1.2706
	R2	0.7846	0.9477	0.8583	0.8075	0.9413	0.8491	0.8498	0.9637	0.9155	0.8245	0.9653	0.9042
	Pearson r	0.9090	0.9795	0.9327	0.9213	0.9773	0.9340	0.9335	0.9858	0.9621	0.9266	0.9866	0.9550
(c)	RMSE	0.1100	20.6742	11.8977	0.1083	20.7300	7.1590	0.0991	23.8558	6.7880	0.1162	19.1140	4.2768
	MAE	0.0835	16.5848	8.9862	0.0872	16.2398	5.2717	0.0742	18.9137	5.1237	0.0929	14.9196	3.5232
	MAPE	1.3052	1.9477	2.5613	1.3618	1.8357	1.4758	1.1608	2.1030	1.4325	1.4376	1.6747	0.9975
	R2	0.8242	0.9673	0.5905	0.8294	0.9671	0.8518	0.8573	0.9564	0.8667	0.8036	0.9720	0.9471
	Pearson r	0.9229	0.9858	0.7856	0.9162	0.9858	0.9342	0.9326	0.9865	0.9417	0.9335	0.9910	0.9752

Figure 8. Comparison of dataset performances for ANN model with the no-SCR mode.

As shown in Table 9a, in the case of NOx, the RMSE (MAE, MAPE) of DaE-sx was the lowest at 5.9282 ppm (4.7832 ppm, 0.5672%); however, it was only slightly different from that of DaCE-sx. In the case of tEx, DaCE-sx yielded the best performance, with its RMSE (MAE, MAPE) being the lowest at 0.4150 °C (0.3143 °C, 0.1324%). When the electrical-related parameters were added, the R² and Pearson’s r were 1.0. As shown in Table 9b, the RMSE (MAE, MAPE) of NOx was larger at 45.9671 ppm (39.2833 ppm, 4.7777%) with Da-sx. However, the predictions with DaC-sx and DaE-sx revealed that the RMSE (MAE, MAPE) was dramatically reduced by 17.171 ppm (14.3075 ppm, 1.8415%) and 30.2676 ppm (25.7416 ppm, 3.2877%), respectively. As shown in Figure 9, it could be confirmed that data points of DaE-sx and DaCE-sx centered to the best prediction line, while those of Da-sx and DaC-sx did not. In conclusion, with the SVM model, the prediction results of DaE-sx for CO₂, NOx, and tEx were superior to those of the other datasets.

Table 9. Metrics from different datasets for the SVM model with the no-SCR mode: (a) training from the ATC training set and metrics from the validation set; (b) trained model tested from the BTC and its metrics.

		Da-sx			DaC-sx			DaE-sx			DaCE-sx
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(a)	RMSE	0.0356	8.9050	2.2663	0.0336	6.8926	2.1422	0.0343	5.9282	0.6333	0.0324	6.0193	0.4150
	MAE	0.0274	6.3848	1.6711	0.0272	5.0855	1.4682	0.0266	4.7832	0.3586	0.0264	5.0455	0.3143
	MAPE	0.4421	0.7546	0.6191	0.4366	0.5960	0.5583	0.4283	0.5672	0.1548	0.4234	0.6001	0.1324
	R2	0.9657	0.9939	0.9987	0.9695	0.9963	0.9988	0.9682	0.9973	0.9999	0.9716	0.9972	1.0000
	Pearson r	0.9829	0.9970	0.9994	0.9851	0.9982	0.9994	0.9854	0.9987	1.0000	0.9877	0.9986	1.0000
		Da-sx-test			DaC-sx-test			DaE-sx-test			DaCE-sx-test
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(b)	RMSE	0.1312	45.9671	8.9578	0.1136	28.7961	9.5169	0.0901	15.6995	4.1195	0.1137	17.1088	4.5839
	MAE	0.1073	39.2833	7.6497	0.0936	24.9758	8.4966	0.0659	13.5417	3.5823	0.0836	14.5928	4.1775
	MAPE	1.6889	4.7777	2.2293	1.4584	2.9362	2.4750	1.0243	1.4900	0.9989	1.2924	1.7120	1.1871
	R2	0.7499	0.8382	0.7679	0.8124	0.9365	0.7380	0.8819	0.9811	0.9509	0.8119	0.9776	0.9392
	Pearson r	0.8995	0.9243	0.9199	0.9058	0.9711	0.9276	0.9399	0.9920	0.9754	0.9303	0.9962	0.9712

Figure 9. Comparison of dataset performances for SVM model with the no-SCR mode.

4.1.2. Comparison of Model Performances with the No-SCR Mode

For the ANN model, owing to the weight initialization, the predictions were slightly different for each trained model. Therefore, a total of 15 models were generated through 15 training sessions. By testing the validation set with this model, the model with the lowest RMSE (MAE) was selected, the performance was predicted with the test dataset BTC, and the results were compared with those of the SVM model. As shown in Table 8c, DaCE-sx had the best performance with the ANN model, apart from a negligible CO₂. Referring to Figure 10, it can be seen that the error in CO₂ prediction with the ANN model was almost zero. In the case of tEx, the red line indicating the real values did not overlap with the predicted values, except for DaCE-sx, which had an almost zero error. In the case of NOx, the error was the smallest in DaCE-sx; therefore, the orange line overlapped the royal blue line well. In addition, the values for each load can be seen by referring to the NOx graph, which had a step-like change. In Figure 11, the error of DaE-sx for all the target data with the SVM model was close to zero compared to other datasets. From the prediction results of DaCE-sx with the ANN model and DaE-sx with the SVM model, the RMSE (R²) values were 0.1162% (0.8036) and 0.0901% (0.8819), respectively, for CO₂, 19.1140 ppm (0.9720) and 15.6995 ppm (0.9811), respectively, for NOx, and 4.2768 °C (0.9471) and 4.1195 °C (0.9509), respectively, for tEx, thus indicating that the SVM model was superior to the ANN model.

Figure 10. Comparison of real and predicted values for different datasets with ANN models in no-SCR mode.

Figure 11. Comparison of real and predicted values for different datasets with SVM models in no-SCR mode.

4.2. Comparison of Emission Predictions of ANN and SVM Models with the SCR Mode

4.2.1. Comparison of Dataset Type with the SCR Mode

For the dataset DaE-so, as presented in Table 10a, the results of the ANN model when the SCR was operational, the RMSE (MAE) values were 13.0547 ppm (9.4762 ppm) and 0.3538 °C (0.2760 °C) for NOx and tEx, respectively, thus indicating the best prediction performance. For CO₂, DaCE-so exhibited the best performance. As can be seen from Table 10b, DaC-so performed the best in predicting CO₂ and tEx emissions in terms of the RMSE, R², and Pearson r. In the case of NOx, Da-so, which was the most basic type among all the datasets, yielded the best performance. However, the difference in the metrics for CO₂ and tEx was insignificant among the datasets. In the case of NOx, the difference in the RMSE (MAE) between DaCE-so and Da-so was 0.4879 ppm (0.0301 ppm), which is not considered significant. Referring to Table 10c, the differences in CO₂ and tEx between datasets were not large. However, as shown in Figure 12, DaC-so-test and Da-so-test performed the best in predicting CO₂ and tEx, respectively. In the case of NOx, DaCE-so-test had the smallest RMSE (MAE) at 15.1043 ppm (12.5964 ppm); referring to Figure 12, it can be seen that data points were more located at the center than others.

Table 10. Metrics from different datasets for the ANN model with the SCR mode: (a) trained 15 times from the ATC training set with average taken of the 15 metrics from the validation set; (b) 15 trained models tested from BTC with average taken of the 15 metrics; (c) the best model for the validation set tested from BTC and its metrics.

		Da-so			DaC-so			DaE-so			DaCE-so
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(a)	RMSE	0.0498	13.3101	0.7196	0.0481	13.8279	0.6893	0.0478	13.0547	0.3538	0.0471	13.7466	0.3843
	MAE	0.0351	9.6977	0.5599	0.0345	10.0285	0.5240	0.0346	9.4762	0.2760	0.0342	9.8260	0.3080
	MAPE	0.5633	8.6440	0.1502	0.5539	8.6388	0.1407	0.5545	8.4854	0.0738	0.5486	8.7009	0.0825
	R2	0.9479	0.9699	0.9701	0.9514	0.9675	0.9725	0.9519	0.9710	0.9928	0.9534	0.9679	0.9910
	Pearson r	0.9757	0.9867	0.9880	0.9772	0.9848	0.9909	0.9773	0.9877	0.9972	0.9780	0.9857	0.9972
		Da-so-test			DaC-so-test			DaE-so-test			DaCE-so-test
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(b)	RMSE	0.1036	18.5344	1.7806	0.1015	20.2399	1.7621	0.1017	19.6393	1.8163	0.1047	19.0223	1.8988
	MAE	0.0698	15.1720	1.2828	0.0682	16.1593	1.4040	0.0678	16.4335	1.4341	0.0707	15.2021	1.5524
	MAPE	1.0647	7.6687	0.3393	1.0362	8.1386	0.3716	1.0283	8.1405	0.3793	1.0766	7.6966	0.4103
	R2	0.8669	0.9315	0.8339	0.8725	0.9202	0.8407	0.8713	0.9245	0.8322	0.8641	0.9286	0.8166
	Pearson r	0.9350	0.9768	0.9197	0.9361	0.9674	0.9205	0.9348	0.9752	0.9179	0.9322	0.9699	0.9086
(c)	RMSE	0.1093	16.7953	1.5430	0.0954	16.3705	1.8551	0.0998	18.9179	1.7513	0.1138	15.1043	1.9636
	MAE	0.0713	14.0085	1.0606	0.0590	13.8268	1.4716	0.0671	16.3756	1.2498	0.0848	12.5964	1.6660
	MAPE	1.0901	6.9846	0.2800	0.8975	6.9991	0.3892	1.0114	7.6940	0.3303	1.2954	6.3993	0.4401
	R2	0.8526	0.9462	0.8795	0.8877	0.9489	0.8258	0.8772	0.9318	0.8448	0.8404	0.9565	0.8049
	Pearson r	0.9262	0.9867	0.9411	0.9447	0.9796	0.9155	0.9378	0.9764	0.9291	0.9207	0.9828	0.9031

Figure 12. Comparison of dataset performances for ANN model with the SCR mode.

For the SVM model, from Table 11 and Figure 13, it can be observed that DaCE-so yielded the best performance for all metrics in terms of CO₂, NOx, and tEx. In Figure 13, when the value of CO₂ was 6.0%, it can be seen that the data points of Da-so and DaC-so were far from the best prediction line compared to DaE-so and DaCE-so. In the prediction of NOx, when it was 300 ppm, the data points of DaCE-so were most concentrated along the best prediction line. When predicting tEx, data points near 380 °C were best distributed along the best prediction line for DaCE-so.

Table 11. Metrics from different datasets for the SVM model with the SCR mode: (a) training from ATC training set and the value of metrics from the validation set; (b) trained model tested from BTC and its value of metrics.

		Da-so			DaC-so			DaE-so			DaCE-so
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(a)	RMSE	0.0401	11.3561	1.0949	0.0402	11.2468	1.0645	0.0387	10.9078	0.9494	0.0381	10.8215	0.9516
	MAE	0.0308	8.4896	0.7926	0.0307	8.4305	0.6722	0.0286	8.2835	0.7048	0.0284	8.1028	0.6140
	MAPE	0.4979	7.6329	0.2114	0.4959	7.5758	0.1791	0.4628	7.4243	0.1881	0.4593	7.3156	0.1636
	R2	0.9648	0.9776	0.9267	0.9646	0.9780	0.9307	0.9672	0.9793	0.9449	0.9681	0.9797	0.9446
	Pearson r	0.9829	0.9889	0.9647	0.9829	0.9891	0.9670	0.9842	0.9898	0.9733	0.9845	0.9899	0.9736
		Da-so-test			DaC-so-test			DaE-so-test			DaCE-so-test
		CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)	CO2 (%)	NOx (ppm)	tEx (°C)
(b)	RMSE	0.0987	13.9492	1.7759	0.0987	13.7728	1.7947	0.0911	13.9251	1.5729	0.0912	13.6775	1.5688
	MAE	0.0597	11.7822	1.4015	0.0581	11.5351	1.3697	0.0516	11.8302	1.2459	0.0508	11.4612	1.1570
	MAPE	0.9014	5.8510	0.3717	0.8757	5.7186	0.3627	0.7772	5.9312	0.3301	0.7636	5.7718	0.3061
	R2	0.8797	0.9629	0.8404	0.8799	0.9638	0.8370	0.8975	0.9630	0.8748	0.8975	0.9643	0.8754
	Pearson r	0.9471	0.9909	0.9309	0.9458	0.9904	0.9271	0.9481	0.9916	0.9429	0.9481	0.9912	0.9431

Figure 13. Comparison of dataset performances for SVM model with the SCR mode.

4.2.2. Comparison of Model Performance in the SCR Mode

As shown in Table 10c, in the ANN model, CO₂, NOx, and tEx performed best in DaC-so-test, DaCE-so-test, and Da-so-test, respectively. From Figure 14, It can be observed that the error of DaCE-so with the ANN model was small compared to that of other datasets. Even though CO₂ and tEx performed well in other datasets, it can be seen from the graph that this was negligible. In Figure 15, it is evident that, in the SCR mode, the SVM model predictions were superior for all the datasets. Even though it is difficult to visually assess them, a comparison of the RMSE (R²) values shows that DaCE-so was the best. Furthermore, in Figure 15, it can be intuitively seen that the area between the DaCE-so error line and the zero-value baseline was the smallest. Lastly, considering that the prediction errors of CO₂ and tEx among the datasets were insignificant, the performances of the ANN and SVM models were compared for DaCE-so. The RMSE (R²) values for ANN and SVM were 0.1138% (0.8404) and 0.0912% (0.8975), respectively, for CO₂, 15.1043 ppm (0.9565) and 13.6775 ppm (0.9643), respectively, for NOx, and 1.9636 °C (0.8049) and 1.5688 °C (0.8754), respectively, for tEx, indicating that the SVM model was superior to the ANN model.

Figure 14. Comparison of real and predicted values for different datasets with ANN models in SCR mode.

Figure 15. Comparison of real and predicted values for different datasets with SVM models in SCR mode.

5. Conclusions

To predict the emissions and exhaust gas temperatures with and without an SCR for a G/E, the exhaust gas variables were measured and ship data were collected during the period of documenting the G/E performance. Most of the previous studies predicted the emissions or performance using only the engine variables as the input data. However, the CFW supplied from the outside flows through the engine and is directed to an air cooler. Therefore, the CFW affects the charging air and eventually affects the exhaust gas; thus, a dataset called DaC was created in this study. In addition, another dataset DaE was created by adding the electrical variables proportional to the engine load; lastly, the dataset DaCE was created by adding both the cooling and the electrical variables. For the predictions for the SCR mode, the amount of urea injected and percentage of the G/E load were added to the four datasets created above. To compare the model performances, after several trial-and-error attempts, an ANN model of the 64–32–16–8–4–1 structure with optimal hyperparameters and an SVM model using the RBF kernel trick were created. They were trained with the data after the T/C overhaul, verified with the validation set, and tested with the data before the T/C overhaul.

From the prediction results, it can be observed that NOx had a larger variation between the datasets compared to CO₂ and tEx; therefore, the NOx prediction performance was mainly focused on. When comparing the performance of the dataset in the no-SCR mode, with the ANN model, the results of the validation set presented in Table 8a show that DaE had an excellent predictive performance for CO₂ and NOx. However, the difference in the prediction performance of NOx compared to that of DaCE was not significant, and the prediction of tEx was better with DaCE. From DaC in Table 8b and DaE in Table 8c, it can be seen that the performance was inferior to that of Da with regard to NOx prediction; however, the best performance was achieved in DaCE. In the SVM model as shown in Table 9, DaE performed the best, and there was no significant difference from that of DaCE.

For the ANN model with the SCR mode, DaE performed the best with the validation set. Table 10b shows that Da had the best performance; however, there was no significant difference from that of DaCE. From Table 10c, it can be observed that the performance of DaE was inferior to that of Da; however, DaCE yielded the best performance. In the case of the SVM model, as shown in Table 11, DaCE exhibited the best performance. The performances of DaC and DaE were worse than that of Da in some cases; however, overall, DaCE generally yielded good performances. Even if the performance was worse than that of other datasets in some cases, the differences were insignificant; thus, it is preferable to use DaCE.

In conclusion, according to the evaluation of the performance of the models, it can be seen that SVM outperformed ANN regardless of whether SCR was operated, which explains that SVM is powerful in small datasets. When the performance of the SVM model was measured using the test data of DaCE in both no-SCR mode and SCR mode, the RMSE (MAPE, R²) of CO₂ was between 0.1137% (1.2924%, 0.8119) and 0.0912% (0.7636%, 0.8975), the RMSE (MAPE, R²) of NOx was between 17.1088 ppm (5.7718%, 0.9643) and 13.6775 ppm (1.7120%, 0.9776), and the RMSE (MAPE, R²) of tEx was between 4.5839 °C (1.1871%, 0.8754) and 1.5688 °C (0.3061%, 0.9392). Accordingly, it can be seen that satisfactory results could be obtained when the DaCE dataset and SVM were used. Therefore, when applying ML as a concept of a virtual sensor for engine emission measurement equipped with a data collection device, it would be desirable to apply SVM. This is because, initially, the amount of data is insufficient. However, when the number of collected datasets increases with time, the prediction performance is expected to improve further, and ANN or other ML models can also be considered.

As a follow-up study, AI modeling based on data acquired under long-term operating conditions is required. In addition, in order to satisfy TIER 3, which is a strong NOx regulation for marine engines, it is also necessary to study an engine equipped with EGR, which is the only alternative to SCR that can be used.