Comment

1 July 2006

Netting a model predictive combo

This clever optimization scheme used model predictive control, neural nets, and expert systems

By Steven Coker, Don Labbe, and Andy Speziale

The pressures of Entergy’s coal-fired plant are many-fold.

The complex nature of the twin furnace boiler with eight coal mills, 152 air dampers, burner tilts, and multiple spray controls challenged the operator and DCS to optimize performance. 

The further need to meet dispatch requirements added to the dynamic needs.

Entergy operates two 800 MW units at their Independence Station in Independence County, Ark. The units have tangentially fired drum type twin furnace boilers with eight coal mills supplying PRB coal. Design turbine throttle conditions are 1000°F/1000°F and 2400 psig. 

The units recently had retrofits including distributed control systems for the boiler and auxiliary controls. These modifications enhanced unit reliability, improved thermal performance, and provided continuous dispatch capability.

With the DCS as a solid high performance control platform, Entergy embarked on a project to further reduce their NOx emissions, improve heat rate, and lower peak steam temperatures. 

Since the project requirements combined optimization and dynamic control of a wide range of variables, a solution combining model predictive control, neural nets, and expert systems implemented and securely interfaced to the DCS platform. 

NOx, heat rate, and steam 

With the successful installation of the DCS to meet the regulatory requirements, Entergy envisioned a wide scope optimization system that would further enhance boiler performance in each area: emissions, heat rate, and dynamics. The complex nature of the twin furnace design made it very challenging for the operator to achieve minimum NOx even with the power of the DCS.

Therefore, the first objective of the optimization system was to capitalize on the “flexibility” of the twin furnace controls and drive the unit to minimize NOx formation on a continuous basis including dispatch operation. A neural net approach was the choice to effectively link the complex interrelationships.   

The second objective was to improve unit heat rate, thereby reducing fuel costs with secondary benefits of reduced emissions of SO2 and CO2. The twin furnace design provided optimization opportunities for excess air, exit gas temperature, and superheat/reheat steam temperature and spray. Since heat-rate optimization needed to be continuous during both fixed and dispatch operation, engineering selected a model predictive control approach with its inherent advantages in dynamic control.

The third objective was to trim superheat and reheat steam temperature peaks during dispatch operation. These peaks probably contributed to tube failures, and any reduction in forced outages greatly improves the economics of the units. At high loads, the units had insufficient superheat tube capacity to meet full steam temperature requirements. 

In addition to the heat-rate performance impact, this also challenged the regulatory steam-temperature control system. During dispatch, the steam temperatures would quickly accelerate from a low temperature condition through set point to a high steam temperature as the spray system responded. 

A model predictive approach offered a clear method to anticipate increasing steam temperature and take appropriate control steps in advance, and thereby prevent high peaks without excessive spray at lower steam temperatures.

Each furnace had a full set of soot blowers to clear tubes of excessive soot deposits. The standard application of soot blower sequences was quite effective in removing these deposits, but it often sacrificed steam temperatures and increased tube erosion. 

To meet the objectives of removing deposits and favorably controlling energy distribution within the boiler, an expert system based soot blow system went into the overall system. 

This system also helped in NOx mitigation as well as heat rate improvement. The system was labeled smart soot blow automation.

Control combo formulate

Model predictive control uses time dependent relationships between controlled variables and manipulated variables to predict response to control key process parameters near set point or within constraint. 

The manipulated variables are independent of each other and do not affect the process response of other manipulated variables. The control response from model predictive control comes out of a combination of past-manipulated variable moves, current values, and future predictions over a design horizon. 

The models come together from response data, so they are representative of the process during actual operation. The transparency of the models and flexibility in tuning parameters lend well to controlling challenging processes such as steam temperature.

Model predictive control models are well suited for processes in which the controlled variables and manipulated variables connect through thermodynamic, chemical, or control relationships. 

For example, final steam temperature relates to superheat spray with a particular gain based on steam properties and a time response based on metal mass and steam flow. As another example, O2 and CO regulation are related to air flow bias (typically referred to as O2 control) through a combination of chemical, control, and analyzer response. Model predictive control provides rapid response and stable control for such processes, allowing operation close to constraint limits for maximum heat rate and NOx benefits.

Neural networks also apply time dependent relationships between controlled and manipulated variables. However, the manipulated variables set up as interactive such that the response to each manipulated variable relates to the position of all other manipulated variables. For example, air dampers form a parallel network and are interactive, suggesting a neural network approach. 

Due to the more complex model structure, neural networks require much more computer resources than model predictive control and may introduce undesirable interrelationships between manipulated variables. For example, increased superheat spray is typically in response to an increase in firing rate, so relationships between spray flow control and combustion control variables could result from a neural net model. In addition, the fundamental relationships may skew over parts of the operating range due to unmeasured process disturbances, such as soot blowing or slag falls. Additionally, forcing a neural net solution on all variables for large complex systems increases computational requirements and may force larger time steps that would compromise dynamic control. For these reasons, we implemented a neural net architecture, which combines with model predictive control for selected controlled variables.

In the first pass of the controller design, plant data developed models for later use. Then we developed dynamic neural network models for controlled variables specifically selected by the user. The model predictions were compared to determine the trade-offs between the modeling approaches. We retained the neural net models for control variables, which benefited from the neural nets and model predictive control applied for variables. 

Steam temperature control

These units have a history of operating at low superheat steam temperatures and with large temperature imbalance. This is apparently due to the undersized superheater sections and twin furnace design. 

As a result, the superheat sprays are usually closed during steady load operation and load reductions. However, during load increases, firing drives steam temperatures upwards at a fast rate. Sprays must come in quickly to slow the rate of increase and prevent excessive temperature increases. 

The challenge for steam temperature control was clipping of the high peak steam temperatures and then cutting back sprays as temperatures turned. 

As such, the distinct objectives of the steam temperature control were to:

• Clip the high peaks of steam temperature.
• Bring the A and B side temperatures into better balance.
• Increase the average superheat and reheat steam temperatures towards set point for better unit heat rate.

Load was under dispatch control for most of the 16-day period with brief periods of steady high load. Dispatch operation causes large variations in unit firing and swings in steam temperatures. 

Over the entire period, there were only a few occurrences of superheat steam temperatures greater than 1010°F with a maximum superheat steam temperature of 1015°F. Prior to the installation of the optimization system, steam temperatures in excess of 1020°F were quite common. 

In addition, the optimization system balanced the A and B-side steam temperatures and brought them closer to set point. This is a major component in the heat rate benefits. 

Neural net NOx control

The tangentially fired twin furnace design provides many opportunities for NOx reduction. There are 152 air dampers and eight coal mills providing fuel through 64 primary air ports in the furnaces. 

This complex design necessitated the collection of operational data from a large tag database to identify the dominant manipulated variables for NOx mitigation. Data from operations went into a neural network, and engineers developed control models. 

Many tags showed no NOx correlation, and DCS control handled them. Those manipulated variables, which provided some NOx merit, were included within the control design and became a part of the neural network models. 

The neural network technology utilizes time dependent Radial Basis Function networks, which offer faster training times than other technologies such as Multi Layer Perceptron.

This is an advantage in controller design, when screening of manipulated variables is required to identify the significant players. This is also an advantage for online updating of neural net models.

There are several drawbacks associated with the use of traditional neural networks in control, notably their inability to extrapolate to new operating regimes, and the fact that information on the process hides in the “black box” of the neural network. 

Since the objective of the neural net system is robust control and optimization, prior to control action the gains resulting from the neural network are evaluated, logged, and range checked. This prevents the application of destabilizing gains within the controller and effectively opens the “black box” for closer evaluation.

The neural net system lowered NOx levels approximately 25% from original levels. 

Smart soot blow automation

Soot blowing was a challenging problem for these units. Over the years, instrumentation had added on to assist in the decision process for soot blowing. The additional instrumentation included extra temperature measurements along the superheat, reheat paths, gas side pressure drop measurements, and strain gage measurements of tube bundle weights. 

Soot blowing served several purposes, including maintaining sufficient gas path openings to meet the air requirements for full load, the prevention of large accumulations of slag that could have destructive effects during slag falls, and the favorable distribution of energy for highest cycle efficiency.

The large number of soot blowers coupled with the complex data base of effects made it very challenging for the operators to select individual blowers for activation, so soot blower group sequences typically undertaken on a time basis or in response to process alarm conditions. This approach was effective in responding to the issues of gas path pressure drop and large accumulations of slag. However, it was not always favorable to the distribution of energy. For example, unnecessary soot blowing of the water wall area would typically lower superheat steam temperature and adversely affect unit heat rate.

To provide an automatic system for soot blower activation, which met all three objectives, engineering employed the expert system feature of the optimization system. The expert system processes control variable data, such as steam temperatures and burner tilts, steam temperature distribution, gas path pressure drops, gas temperature distribution, and strain gage weights for the various tube bundles. 

Applying model data for soot blower effectiveness, the expert system then selects the most effective soot blower to meet the existing soot blowing requirements. Following an evaluation of satisfactory permissive logic at the DCS level, the soot blower activates.

The smart soot blow automation system reduces the burden on the operator for the time consuming task of soot blower monitoring and sequence control and prevents most process alarms associated with the build up of soot deposits. The operators have the support data for the selection and can activate soot blowers of their own choice at any time. The system is popular with operations.

Further, the system has reduced soot blower steam consumption and soot blower tube erosion while increasing boiler efficiency and superheat steam temperatures. This has contributed significantly to the heat rate improvements realized by the optimization system.

Delta heat rate methodology

Following the completion of an optimization project, the final challenge is to quantify the realized benefits. NOx reductions are typically available through comparisons of before and after operations. However, heat rate improvements are much more difficult to directly measure, particularly for coal fired units. For this reason, this project used the Delta Heat Rate Methodology online within the optimization system.

The delta heat rate methodology applies a heat rate value to the critical performance measurements influenced by the optimization system. The heat rate value for each performance measurement is consistent with industry standard published values. 

The online calculation compares the current value to baseline values for these critical performance measurements using the actual data during the period. Since most performance measures have a strong dependence to load, the baseline values are load dependent values.

The delta heat rate value for each parameter is determined from the product of the delta heat rate factor and the difference between the actual value and the baseline value. One derives the dynamic delta heat rate impact by summing the delta heat rate values for all of the parameters. Throttle pressure was not included since a fixed set point was applied for the operating range.

This approach does not give full consideration to turbine issues, such as valve point operation. It does provide a consistent measure of performance. By tracking this index, one can identify improvements or degradation in performance.  

Since the calculations are online, they can apply dynamically. During dispatch operation, the values will vary due to changes in the measured parameters, such as steam temperatures, spray flows, and gas temperatures. 

Although the base line values were determined from steady load operation, prior to the installation of the optimization system, the comparison of dispatch operation is still valuable. The desire would be that the optimization system provides merit during both steady load conditions and dispatch operation.

Therefore, the optimization needs for coal-fired units are varied and may reach beyond steady load NOx mitigation. The combination of neural net/model predictive control provides a means to economic benefits of an optimization system. 

This case illustrates the large heat rate benefits and NOx reductions realized from an optimization of a complex twin furnace design. The pay back in fuel saving alone was several months. The system provides dynamic steam temperature control, which reduces high temperature stress levels, potentially reducing tube outages. The inclusion of Smart Soot Blow Automation within the optimization system completes the integrated overall optimization.

ABOUT THE AUTHORS

Steven Coker (scoker1@entergy.com) is senior engineer at Entergy Arkansas. Donald Labbe (Donald.Labbe@ips.invensys.com) and Andy Speziale (Andy.Speziale@ips.invensys.com) are consulting engineers at Invensys Systems Inc.