01 October 2003

Chinese steel plant clears air with software

Model-free adaptive control puts Ling-Yuan Iron and Steel in a league of its own.

By George Cheng and Li-Qun Huo

In an iron and steel complex, operating units including blast furnaces, basic oxygen furnaces, and coking ovens all produce gases as by-products.

Some plants still discharge these gases into the atmosphere, wasting valuable energy and causing severe air pollution. A gas plant mixes these gases to produce fuel for the furnaces in metal casting and rolling mills.

One measures the quality of the mixed gas by its heating value. Gases with inconsistent heating values can cause major control, quality, and production problems due to over- or underheating.

Even during normal production, gas supply and demand can change randomly. Major operating units such as blast furnaces and reheating furnaces may go online and off-line periodically causing huge disturbances in gas flow, pressure, and heating value.

Online heating value analyzers are available but are seldom used during normal operations because they are difficult to maintain and very expensive. It is desirable to monitor and control the gas-heating value automatically in all operating conditions.

Here is what it takes to automatically measure and control the mixed gas-heating value. Here is how model-free adaptive (MFA) control technology enables users to improve production efficiency and yield.



A simplified gas mixing system has coking furnace gas (PT1 for pressure, FT1 for flow) mixing with blast furnace gas (PT2 for pressure, FT2  for flow) to produce the mixed gas.

An online analyzer (AT) measures the heating value. However, because it is difficult to maintain, the analyzer is operational only periodically.

For plant safety, maximizing energy usage, and minimizing pollution, the gas mixing process has these production objectives.

  • Use as much coking furnace gas as possible, because its heating value is five times higher than the blast furnace gas with which it mixes. In normal conditions, there is no need to control its flow.
  • Mix blast furnace gas with coking furnace gas to form a consistent heating value.
  • Because the coking furnace and blast furnace are working in a changing production environment, the amount of gas they produce can change significantly. When PT1 is too high, the blast furnace gas line (FT2) must be shut down to prevent the high-pressure coking furnace gas from flowing into the low-pressure gas grid. This could cause a plant accident. On the other hand, when PT1 is too low, FT2 must increase to keep a sufficient gas supply going to the downstream processes.




When the gas mixing system runs under manual control, operators leave FT1 wide open to use as much coking furnace gas a possible; then they manually adjust the pressure control valves to control PT1 and PT2 .

Afterwards, they manually adjust FT2  based on experience and measurement from the heating value analyzer to control the mixed gas.

Unfortunately, the large disturbances in the gas grid, frequent changes in gas supply and demand, and the nonlinearity of the gas and flow loops cause the old proportional, integral, derivative (PID) and manual control system to have various problems, resulting in poor heating value consistency.

The product quality and production efficiency suffers in the downstream processes.

The project objective is to automatically measure and control the gas-heating value of the mixed gas during all operating conditions. CyboSoft offers a solution for heating value measurement and control. Using its soft-sensor technology, the system calculates the heating value online.

Using MFA controllers, one can effectively direct gas flow and pressure loops. An MFA controller controls the gas-heating value by cascading with the gas flow controllers.

Model-free adaptive control is a patented technology that readily embeds into a variety of control equipment. Derivations of the core technology address specific control problems.

Most industrial processes still use manual control or the sixty-year-old PID controllers.

Starting from the same oscillating control condition, the system will continue to oscillate under PID control, while the MFA system will quickly adapt to an excellent control condition.

If both controllers start from a sluggish situation, MFA will control the process faster and better. Better control means improved process stability, higher production efficiency and yield, consistent product quality, and reduced material and energy waste.

MFA controls complex systems without requiring process models and is effective for the tough processes such as nonlinear, pH, multivariable, and processes with large time delays. Because there is no model training required, it can launch at any time to immediately control the process. Once installed, no controller tuning is required.


MFA consists of a nonlinear dynamic block that performs the tasks of a feedback controller. A dynamic block is just a dynamic system with inputs and outputs. The control objective is to produce an output to minimize the error between the set point (SP) and the controlled process variable.

A multilayer perceptron artificial neural network (ANN) is part of the design of the controller. The ANN has one input layer, one hidden layer with N neurons, and one output layer with one neuron.

Within the dynamic block there is a group of weighting factors (wij and hi ) that can be changed as needed to vary the nonlinear functions of the block.

The learning algorithm for updating the weighting factors is based on the goal of minimizing the error between the set point and process variable. This effort is the same as the control objective. Therefore, the adaptation of the weighting factors can help the controller minimize the error.

Based on its design, the MFA controller remembers a portion of the process data, providing valuable information for the process dynamics. In comparison, a digital version of the PID controller remembers only the current and previous two samples.


The solution for the Ling-Yuan Iron and Steel situation used this design.

Because the control objective is to provide a consistent mixed gas-heating value (AT), a cascade control system is used to control AT as its outer loop and the ratio of the FT1 and FT2  as its inner loop.

An antidelay controller (AC) was used to control the quality variable of the system-mixed gas-heating value. This controller has a built-in predictor that can effectively handle large time delays.

A soft-sensor calculates the heating value online as the process variable (PV) for the AT controller. Operators turn on the analyzer periodically to check and calibrate the calculated value. The soft-sensor keeps the system running in the automatic mode twenty-four hours a day, seven days a week without depending on the analyzer.

A robust controller (FC1) is the coking furnace gas flow controller for protecting the mixed gas load. During normal operation, the control valve is wide open. When the gas demand load goes below a certain range, the FC1 will start to close the valve to reduce FT1. Based on the flow ratio control, FT2  will reduce accordingly. The combined control effort will reduce the total supply of the mixed gas and also suppress its pressure to keep the system working within its safety range.

PC2 controls the blast furnace gas pressure (PT2) and works as a supporting controller for the heating value loop. If PT1 is too low during gas mixing, it can cause the heating value-AT-loop to be sluggish, and if it is too high it can cause the AT loop to oscillate. Controller PC2 is able to keep PT2  within a defined working range. This improves the stability and controllability of the heating value loop.

Another controller protects and supports the flow control (FC1 ) loop. By manipulating the pressure differential of the control valve, it is able to stabilize the coking furnace gas line. It allows the gas pressure to fluctuate within a defined range.

A feedforward controller configures inside controller FC1  with the load pressure as the feedforward signal. It is able to compensate for the load changes quickly.

This control software, monitoring, and optimization package was in-stalled in a PC running Windows 2000 and connected to the distributed control system through an OPC interface.

A soft-sensor program embeds into the human interface to calculate the mixed gas-heating value and also serves as the process variable for the heating value controller.

The heating value analyzer runs periodically to check and calibrate the calculated value. After several months of operation, the soft-sensor proved to be accurate, which enables automatic control of heating value and saves analyzer life.

This patent-pending robust MFA controller is not a control system design method. We use the term robust here, because the MFA control system dramatically improves the control system robustness.

Without redesigning a controller, using feedforward compensation, or retuning the controller parameters, this controller maintains the process through normal conditions and those where there are significant disturbances or system dynamic changes. IT

Behind the byline

George Cheng has B.S., M.S., and Ph.D. degrees in electrical engineering and is the chairman and chief technical officer of CyboSoft. He holds six U.S. patents. Dr. Cheng is a senior member of ISA and IEEE. Li-Qun Huo has a B.S. degree in automation and is the deputy general manager of CyboSoft Automation Technology (Beijing) Co. Ltd. He is a member of the Chinese Instrument Society.



Looking beyond PID control

For most loops, a simple proportional, integral, derivative (PID) algorithm controller works.

Besides the standard PID-the ideal or ISA controller-Process Software and Digital Networks (ISA and CRC Press, 2002) lists at least eight variations that work more aggressively on processes that have frequent set point changes and disturbances, that need averaging and surge help, and others.

  • Smith predictor: This controller receives feedback from a model of the process, without the dead time, rather than from the real process. The real process variable compares with models to correct errors and handle disturbances. It is good for processes with large dead time that have set point changes and no overshoot on set point changes. It is as fast or only slightly faster than normal PID control load changes.
  • Dahlin controller/Model-predictive controller: The controller behaves as one over the process (process · (1/process) = 1), giving a closed-loop response that has the same form as the open-loop response. This is similar to lambda tuning, beneficial for loops that can tolerate no overshoot on a set point change. There is no overshoot on set point changes, and it is very sluggish on load changes.
  • Adaptive model controller: Same as model-predictive controllers, but the model receives continuous updates. It is very good in mechanical systems, but does not handle process problems very well. Hysteresis, stiction, and nonlinearity cause the system to remodel continuously.
  • Model-free controller: Leading edge technology . . . none that truly work . . . until now.