In the case of the conventional rule-based control strategy used in battery hybrid systems for ships, the system specifications are selected, and the load-sharing section criteria for the generator are set according to the number of generator units and settings of the battery, SOC_MIN (minimum state of charge) and SOC_MAX (maximum state of charge). Subsequently, the status of the generator is categorized into single/parallel and stop/variable/maximum loads. The generator output varies according to the load of the ship and the SOC of the battery, where the prescribed rule is composed of if–then statements and accounts for load sharing. However, if the number of rules increase, the system design becomes more complex [9], and when the optimum specific fuel oil consumption (SFOC) is not taken into account according to the load of the power source, the load sharing is performed arithmetically, leading to inefficient operation.

To address the aforementioned limitations, a control strategy is designed in which the optimum operating point of a power source is set as the reference point for the operating mode change as well as the output reference value of the power source, referred to as the ECMS [5]. Table 1 shows an example design using this strategy [8]. Depending on the system designer, the settings of the load sharing reference point α, SOC_MIN and SOC_MAX of the battery, the minimum/optimum/maximum operating points of the power source differ, and the variable operating point changes in size according to the load condition.

where, P_load is the propulsion motor load, G_max is the maximum output of the genset, and P_batt is the battery power.

The conventional rule-based control strategy is based on the rule of if–then statements, with the battery SOC and load of the ship as inputs and the generator power as the output according to the given conditions. However, with an increasing number of generator units and requirements, the number of rules also increase thereby expanding the system complexity.

A method for training a neural network by calculating the generator output reference value of the expected condition according to the rule-based control strategy was proposed [8]. However, the neural network design can be simplified using the operating mode as a reference value instead of the generator output in the neural network controller design.

As presented in Table 2, the generator operating mode can be classified into three states (1: stop; 2: variable (0%–100%); and 3: optimum (75%)). The different cases in which the two generator units are operated in each mode are defined as modes L1 to L6. As shown in Table 2, the 75% optimum operating point of the generator is set as a reference value for the operating mode section change, and the optimum operating mode is set as the reference value of the generator’s operation. The state of the battery was controlled for these conditions according to the following equation:

where, P_Gn is the output of genset number n.

Based on the proposed rule-based control strategy, the training data for the neural network are represented in terms of the ship load and battery SOC, as shown in Table 3. In the conventional rule-based control strategy reflecting the ECMS, the generator output reference value is used for neural network training. However, in the proposed strategy, the generator operating modes presented in Table 2 are set as the reference values in the neural network.

The training data shown in Table 3 were used to train the neural network, the results of which are shown in Figure 1. The results show good agreement between the predicted values and the design.

Figure 1: 
Training result by neural network controller

To examine the functions of the neural network controller described previously and identify the problems before use in the energy management system of a battery hybrid system under load conditions with real-time changes, MATLAB/Simulink simulations of the system were conducted, and the load sharing performance equivalent to the training data was analyzed under various load conditions. Figure 2 shows a load divider designed to apply the simplified rule-based strategy. First, the real-time incoming load of the ship and SOC of the battery are the two input values, and the neural network provides outputs as the pre-learned modes 1 to 6 as output values.

Figure 2: 
Switch selector for load sharing

Subsequently, to provide a split command to the two generators, it is transmitted to each primary lookup table. As shown in Table 2, when the NN output is changed between modes 1 and 6, the primary lookup table transmits the generator status commands of 1,2,3 to the switch selector. If it is 1, it is stopped; if it is 2, it is a variable load, and if it is 3, it is operated in the optimal operation mode. In the case of stop and optimal operation commands, a fixed value of 0% and 75% as the designer’s decision can be input to the generators through the switch selector regardless of the load; however, it must be transmitted in real time in the case of a variable load. For this purpose, a secondary lookup table was used. Specifically, in the case of a single operation among L3, L5, and L6 requiring variable operation sections, L3 only orders a single operation, and in the case of L5 and L6, it has the function of operating G1 and G2, which are parallel sections. The final output from this is transferred to a lookup table that is converted into a value corresponding to revolutions per minute (RPM) as the output order of the variable-speed generator. Here, the switch selector transmits the operation reference value to the generator according to the scheme in Table 2, which are predetermined signals, and the switch itself is designed to operate only in the stop, variable, and optimum modes.

Figure 3 shows the load profile to verify that the proposed simplified rule-based strategy is applied accurately to the battery hybrid system model using a neural network controller.

Figure 3: 
Operating profile

In the case of a neural network controller using the load of the ship and the SOC of the battery as input values, the neural network output values shown in Figure 4 were obtained according to the training data in Table 3. In this simulation, the SOC of the battery at the beginning was 50% and changed according to the load balancing.

Figure 4: 
Neural network output

Figure 5 shows the verification of the implementation of the rules proposed in Table 2 using the switch selector in the neural network output.

Figure 5: 
Output of switch selectors for load sharing

Figure 6 shows the RPMs of the variable speed generators no.1 (red line) and 2 (blue line), according to the output reference values.

Figure 6: 
RPM of generators

Figure 7 and 8 show the outputs of the generators reflecting the output of the switch selector.

Figure 7: 
Output of generator no.1

Figure 8: 
Output of generator no.2

The changes in the RPM, voltage, and current of the propulsion motor according to the load are shown in Figures 9, 10, and 11, respectively. A torque load proportional to the squared value of the RPM is applied to the propulsion motor, and excellent dynamic characteristics are confirmed by the motor controller using indirect vector control.

Figure 9: 
RPM of propulsion motor

Figure 10: 
Voltage of propulsion motor

Figure 11: 
Current of propulsion motor