Adaptive Neuro-Fuzzy Inference System (ANFIS) for Controlling Level and Pressure on Deaerator

Deaerator is one of the most widely used plants in the chemical industry and marine steam power plant. Deaerator is used to eliminate oxygen in water that enters the boiler to avoid corrosion of the boiler pipes. Control of pressure and level in deaerator needs to be done to keep the process well. The purpose of the research is to design a control system that can keep pressure and level of deaerator on the set point un the presence of changes in the load and input systems. Deaerator should be controlled to keep its safety and efficiency. Adaptive Neuro-Fuzzy Inference System (ANFIS) is a combination of fuzzy logic control and neural network. Design of ANFIS requires an input-output data set obtained from a PI controller that is considered as a "teacher" for ANFIS for the learning process. The results of the simulation show that the system using ANFIS controller for controlling pressure and level deaerator in normal set point can produce very small maximum overshoot that is equal to 0% and small IAE value that is 7.898 for pressure, and 157.7 for level compared to PI.


Introduction
Deaerator is usually used to remove the oxygen and carbon monoxide in condensation water and heat the condensation water to saturation temperature for boiler (Opriş, 2013;Wang, Meng & Ji, 2014;Wang, Meng & Dong, 2015;Zhao, Yao, & Sun, 2014).Deaerator serves to eliminate the content of O 2 from make-up water which will be pumped into heat exchangers and boilers.The process of removing these gases is by entering steam as a heater that will evaporate the gases in the water.After that, the water that has been through the heating process will be accommodated on the deaerator tank which is part of the deaerator before it flows on the boiler.With the heating process in the deaerator, it changes the vapor pressure inside.In addition to controlling the pressure, the water level in the deaerator also needs to be controlled below the permitted charge.Moreover, the water supply to the boiler can continue to be fulfilled.With the multivariable interrelated process, it is difficult to get satisfactory control effect because of the strong coupling between deaerator pressure and deaerator water level.It is necessary to take a corresponding decoupling measure (Wang, Meng & Ji, 2014).Using decoupling in the plant can reduce the interaction between both controlled variables and can produce better control results than without using decoupling.
Many control methods have been applied to plant deaerators such as research by Peng Wang that used PID neural network decoupling.The results prove that the PID neural network decoupling strategy is more effective in deaerator pressure and water level decoupling control than PID (Wang, Meng & Ji, 2014).Another research using PID neural network decoupling strategy can significantly reduce the overshoot and settling time of condenser water level and deaerator water level and can make the control process much more stable.It is more effective in decoupling control than PID (Wang, Meng & Dong, 2015).In addition, Fuzzy tuning control methods PID and IMC were also used for deaerator control.It could reach settling time faster, i.e. 10s for level and 8 s for pressure, than using IMC control method (Gomathy & Anitha, 2015).Another research comparing Neural Network (NN) with the Extreme Learning Machine (ELM) method found that ELM is better and superior when compared with NN.Based on the time of training, ELM is very fast compared with NN.ELM accuracy in training and validation results has an average error smaller than NN predicted outputs (Mahardhika et al., 2017).
Neural Networks have a well-defined training methodology based on their ability to learn but are limited by the inability to make decisions.In order to best exploit the capabilities of both types of controller, it is possible to utilize a neural network which is trained to perform as a fuzzy logic controller.One system which has been shown to combine the positive aspects of both types of AI controller is the Adaptive Neuro-Fuzzy Inference System (ANFIS) (Darvill, Tisan, & Cirstea, 2015;Jang, Sun, & Mizutani, 1997;Narayan, 2017).This research designs the ANFIS controller to control the level and pressure in the deaerator.The ANFIS learning method uses a PI controller that acts as a "teacher" for the ANFIS controller.The implementation of the controller in this research, as well as the simulation result, is presented in the following sections.

Research Method 2.1 Deaerator
Deaerator works by the character of oxygen whose solubility in water will be reduced by the rise in temperature, The O 2 content of the deaerator needs to be removed in order to avoid corrosion but also to avoid fire.From P & ID (Piping and Instrumentation Diagram) deaerator 101U PT.Petrokimia Gresik, degassing section deaerator get input from the makeup water, as shown in Figure 1.There is a process of binding oxygen by giving steam input with a small O 2 content which will bind the large oxygen content contained in makeup water.Steam will be thrown entirely into the air along with the oxygen content that has been tied up.The deaerator also has a condensate water input.The water content of the condensate is H 2 O without any mixture of O2.Water output from degassing section deaerator enters storage water tank deaerator.In this paper, the tray type deaerator is used that consists of a domed deaeration section mounted on a horizontal vessel cylinder that serves as a water storage tank from the boiler (Figure 1).

Mathematical Modeling of Deaerator
The mathematical model equations for pressure and level can be seen in Equations ( 1) and (2).3), ( 4) and ( 5).

Decoupling
Decoupling is a dynamic element that is added to the MIMO (Multiple Input Multiple Output) control systems which aims to minimize the interaction effects between two loops resulting in two loops that do not interact with each other.Decoupling transforms the MIMO model into SISO (Single Input Single Output) form to facilitate the analysis and design of the controller (Stephanopoulos, 2001;Zhou, Deng, & Duan, 2017;Zhang et al., 2016).The decoupling control for the process of controlling the system with two inputs and two outputs is shown in Figure 2. The decoupling equation used is then: (8)

Adaptive Neuro-Fuzzy Inference System (ANFIS)
Following is the membership functions and rules that are used to control the level and pressure on the deaerator.The membership function used is the generalized bell-type for both pressure and level as shown in Figure 3

Results and Discussions
The test was performed by determining the normal set point of 1.825 kg / cm 2 for pressure control and 1.8 m for level control.Variations of testing conducted were testing of increasing set point, decreasing set point, and interference test.The tests aim to determine the ANFIS controller performance against the changes that often occur in the deaerator system.

Normal Set Point
This test was performed by giving the system input with a normal set point value of 1.825 kg / cm 2 for pressure control and 1.8 m for level control.The result of the system response is shown in Figures 7 and 8.The response of each control to the deaerator plant by testing of the normal set point is summarized in Table 1.From Table 1 it can be seen that the pressure control using ANFIS controller has a rise time of 5.0311s, 13.9643 s for settling time, maximum overshoot approaches 0 %, and IAE 7.898.On the other hand, PI controller with and produces rise time of 0.0834s, settling time of 10.9975 s, maximum overshoot 7.87 %, and IAE 7.93.
Moreover, the level control using ANFIS yields rise time of 196.6964 s, settling time of 371.2379 s, maximum overshoot approaches 0 %, and IAE 157.

Increasing Setpoint
In this test, the set point value is increased by 0.2 kg/cm 2 from the normal set point to 2.025 kg / cm 2 for pressure control.While to control the level, set point value is increased by 0.2 m to 2.0 m.The result of system response can be seen in Figure 11 and Figure 12 while the response of control for increasing set point can be seen in Table 2. Table 2 shows that using ANFIS controller for pressure control has rise time of 3.

Decreasing Setpoint
In this test, set point value is decreased by 0.2 kg / cm 2 from the normal set point to 1.625 kg / cm 2 for pressure control.While to control the level, set point value is decreased by 0.2 m to 1.6 m.The result of system response is shown in Figure 11 and 12 while each response of control for decreasing set point can be seen in Table 3. Table 3 shows that using ANFIS controller for pressure control produces rise time of 3.0573 s, 1004 s for settling time, maximum overshoot 11.849 %, and IAE 18.55.While the PI controller has a rise time of 3.0116 s, settling time 1004 s, maximum overshoot 19.896 %, and IAE 27.61.
Then for controlling the level using ANFIS, it yields rise time of 128.4964 s, settling time of 1831.1 s, maximum overshoot 12.955 %, and IAE 180.Moreover, using a PI controller produces rise time of 19.7702 s, settling time of 1831.1 s, maximum overshoot 96.386%, and IAE 260.7.

Test of Step Signal Disturbance
Testing of the disturbance is done on the system with the normal set point with additional disturbance in the form of step signal at 1500 seconds.The purpose of this test is to know the performance of the controller in response to disturbances that occur in the system when a disturbance exists in the form of step signals of 1.825 kg / cm 2 for pressure and 1.8 m for the level.The system responses are shown in Figures 13 and  14, while the response of each control by interference test is presented in Table 4. Table 4 shows that ANFIS controller for pressure control has a rise time of 0.0561 s, 1502.9 s for settling time, maximum overshoot 42.4944 %, and IAE 149.

Conclusions
Pressure and level control on the deaerator yield very small maximum overshoot that approaches 0% and small IAE value of 7.898 for pressure, 157.7 for level compared to PI in normal set point (with , for pressure and , for level).Pressure and level control response with increasing set point using ANFIS can achieve faster settling time, i.e. 1003.9s for pressure and 1364.7s for level, smaller maximum overshoot of 0.16% for pressure and 0% for the level and smaller IAE that is 16.15 for pressure and 178.4 for level than PI.The test of decreasing set point by using ANFIS controller in pressure and level control can achieve faster settling time of 1004s for pressure and 1697.9s for level, smaller maximum overshoot of 11.849% for pressure and 12.955% for level and smaller IAE that is 18.55 for pressure and 180 for level compared with PI.In the test with step signal disturbance on pressure control, the ANFIS controller cannot reach set point after responding disturbance even though it is moving toward set point, while at level control by using ANFIS controller can reach faster settling time with smaller maximum overshoot and IAE value than PI.For normal set point testing, using ANFIS controller for pressure can reach settling time faster that is 13.9643s in comparison with PID Neural Network that is 160s.Controlling pressure and level using ANFIS need a longer time to reach a time, i.e. 13.9643s for pressure and 371.2379s for level, compared by using Fuzzy-PID which requires 10s for level and 8s for pressure to reach settling time.As a result, the usage of ANFIS has been successfully improving the deaerator performance.

Figure 1 .
Figure 1.Deaerator process vapor) flow rate = water out (flow out) flow rate = condensate flow rate = make up water flow rate By inserting the physical parameter values of the deaerator, the transfer function is shown in Equations (

Figure 2 .
Figure 2. Decoupling System on MIMI 2x2 and 4 while the rules are shown in Figure5and 6.

Figure 3 .Figure 5 .Figure 6 .
Figure 3. Membership function for pressure 7. The usage of PI controller with and produces rise time of 21.5045 s, settling time of 836.341 s, maximum overshoot 75.3193%, and IAE 233.1.

Figure 8 .
Figure 8. Level control response with SP=1.8 m 0573 s, 1004 s for settling time, maximum overshoot 11.849%, and IAE 18.55.While the PI controller produces rise time of 3.0116 s, settling time 1004 s, maximum overshoot 19.896%, and IAE 27.61.In case of the level controlling, ANFIS control yields rise time of 128.4964 s, settling time of 1831.1s,maximum overshoot 12.955%, and IAE 180.PI controller produces rise time of 19.7702 s, settling time of 1831.1s,maximum overshoot 96.386%, and IAE 260.7.

Figure 11 .Figure 12 .
Figure 11.Pressure control response with decreasing set point 2. While the PI controller produces rise time of 0.0565 s, settling time 1503,9 s, maximum overshoot 41.5512 %, and IAE 10.77.Then for controlling the level using ANFIS, it yields rise time of 199.0588 s, settling time of 1799 s, maximum overshoot 28.4234 %, and IAE 218.7.PI controller produces rise time of 21.0301 s, settling time of 2082 s, maximum overshoot 77.5745 %, and IAE 342.5.

Table 1 .
Comparison of each control response with normal setpoint

Table 2 .
Comparison of each control response for increasing setpoint Figure 9. Pressure control response with increasing set point Figure 10.Level control response with increasing set point

Table 3 .
The comparison of each response control for decreasing set point

Table 4 .
Comparison of each response control with step signal disturbance Figure 13. Pressure control response with step signal disturbance Figure 14.Level control response with step signal disturbance