A Time of Transition

Increasing environmental pressure has brought advanced control techniques into the power industry arena. Results from the Ostroleka power plant in Poland have proven the suitability of hybrid neural networks in NOx control applications. With similar projects underway in the USA, this technology looks set to become a standard.

The control of coal-fired boilers poses many challenges for power plant operators in today`s industry. Inconsistencies in coal properties, burner arrangement and varying load demands are parameters which affect and alter unit performance in terms of efficiency and emissions.

Although in recent years there have been significant advances in process control systems, there still remain problems in achieving high control quality, particularly in the fine-tuning of control parameters. Fine tuning requires experts with knowledge of both control theory and process dynamics.

More advanced control techniques have therefore evolved, based on soft computing techniques such as artificial neural network and fuzzy logic controllers. In an industry where environmental compliance is rising in importance, these systems enable an operator to enhance boiler efficiency, control emissions of flue gases, and control steam parameters.

A company which has pioneered this technology in the power industry is Transition Technologies Ltd., of Warsaw, Poland. Established in 1991, the company has quickly become the biggest subcontractor to Westinghouse Process Control Inc. (WPC), a Fisher-Rosemount company, with which it has a license agreement for the sale of technical solutions and products.

Transition Technologies` Advanced Control Department provides research and product applications in the field of advanced control. It focuses on specialized control structures using novel approaches such as predictive and adaptive control combined with soft computing methods.

The company has implemented a number of projects using its advanced control technologies in NOx, SOx, opacity and steam temperature control applications. The Ostroleka power plant in Poland is one of the first successful applications in the world of artificial neural network based control of the boiler. This project was completed in 1998 and based on its success, Transition Technologies has since undertaken a similar project in the USA.

A question of reliability

Neural network technology, such as Model Predictive Control (MPC) systems, is an approach to describing a physical system from process data using mathematical algorithms and statistical techniques.

The potential of neural networks for control application lies in the following properties: they can be used to approximate any continuous mapping; they achieve this approximation through “learning”; parallel processing and fault tolerance are easily accomplished; and it is easy to implement control structures with an internal non-linear model, for instance in the predictive or adaptive structures.

Neural networks are an excellent modelling tool and have been used for 20 years in the petrochemical industry for linear modelling techniques.

While neural network based model predictive control can be applied for nonlinear processes, stability problems can arise because its predictive capability is restricted to a certain range of data. If the process moves outside this range, the model becomes unstable and a reliable control system is prevented.

However, Transition Technologies` Aspen Target system is a non-linear MPC controller which uses a multi-step, constrained, Newton-type optimization algorithm to overcome this problem and provide robust and stable operation.

Transition Technologies describes Aspen Target as a hybrid neural network model, where a linear model is first built and then a neural network correction is applied. Thus when the process moves outside the range of data collected, the hybrid model reverts to a linear model, ensuring stability of control.

To understand this in simple terms, take the example of a steam attemperator control, where water is sprayed to maintain the temperature. A pure neural network model may actually decrease water flow on an increase of steam temperature because the water-to-steam temperature relationship may be incorrect outside the data bounds. A hybrid model in the same situation will increase the water flow, and although the effective change made in water will not be exactly the desired value, it can be trimmed by prediction error correction.

With the Aspen Target software, which has been installed at the Ostroleka power plant, process data is taken in an Ascii format and conditioned using various smoothing and filtering techniques. The process gains are determined in two stages: the linear gains are established first followed by a neural network based non-linear correction. The model is considered adaptive due to “gain adjustments” using extended kalman filter based error tracking, which allows gain changes within limits.

The Path Optimizer plans the control moves based on the targets established by the Economic Optimizer. The input data can be processed through an optional sensor validation module. Process model and controller configuration is created offline and loaded onto the online controller.

The controller is implemented through an existing DCS system, Westinghouse WDPF II in the case of Ostroleka, on a special dedicated engineering station connected to the process highway (WesNET Data Highway). The special interface programme every sampling time reads the process points from the highway, and then after performing calculations, sends controls to the actuators via the highway.

The primary goal of this model-based control is to enhance unit efficiency without violating environmental contraints. The project at Ostroleka is the first reported closed loop operation in Europe using neural network technology and one of the first in the world. Here, it is designed to improve boiler efficiency and control NOx emissions.

The Ostroleka project

Ostroleka began modernization several years ago by installing the Westinghouse WDPF control system in all of its three units. Plant performance was significantly improved in terms of boiler operation and reduced maintenance.

As a result of this, the plant owners decided to implement advanced control techniques in order to gain further economic benefits.

Transition Technologies undertook the Ostroleka project in conjunction with the Warsaw University of Technology and with the cooperation of plant managers and engineering staff. With the Aspen Target system now controlling NOx emissions on all of its units, the plant is one of the best practice coal-fired power plants in Poland, and probably Eastern Europe.

Poland introduced regulations for emissions from power plants covering NOx, SOx and dust in the early 1990s, with the aim of bringing plant emission levels in line with those in western Europe. Under the legislation, all new boilers were obliged to install flue gas desulphurization (FGD) equipment, low NOx burners and electrostatic precipitators by a deadline of 1998.

The law was updated in 1998, however, with less restrictive limits for NOx emissions, and enabling some boiler operation without FGD until 2005. Ostroleka is equipped with electrostatic precipitators and low NOx burners, but no FGD. The change in the law has also meant that SO2 control has not been an objective of the Ostroleka project.

The Ostroleka power plant is located near Warsaw, Poland. It consists of three boilers, each with a capacity of 200 MWe and producing 650 t/h steam, and uses pulverized coal for fuel with four mills per boiler. Coal is crushed in the mills and transported to the combustion chamber using secondary air.

The coal has the following properties:

• Heat rate: 21 000 kJ/kg

• Ash content: 16 per cent

• Moisture content: 12 per cent

• Sulphur content: less than 0.8 per cent.

Low NOx burners are used in four elevations with six burners at each level. The burners are distributed on the front wall of the chamber.

The two main control objectives at Ostroleka are:

• Combustion process: control the distribution of coal and air in the combustion chamber. Due to the proper conditions the controller has to enable more efficient and environmentally friendly combustion and reduce emissions as a result.

• Steam temperature control: temperature stabilization (higher efficiency) together with economic use of cooling water.

Two different model based controllers were created to accomplish each of these objectives. The first installation of the system, on Unit 3, was completed in late 1997, while Units 2 and 1 were completed in September 1998.

The project was implemented in the following steps:

• Plant testing followed by data collection

• Process model identification and controller configuration

• Aspen Target integration with the WDPF control system

• Advisory mode implementation (operators can compare current set points with optimal/neural and use them in manual control)

• Comprehensive tests on model predictive control performance

• Tests on hybrid model based control in closed loop

• Enhancement of control with mill system load optimization

• Superheater steam temperature control.

The neural network model structure for the boiler optimization control objective is as follows:

Input:

• Manipulated variables: total air flow; total air dampers (in the open advisory mode only); secondary air dampers; OFA dampers; fuel conveyors.

• Disturbances: O2 concentration; flue gas temperature first superheater; net power generated; combustion chamber pressure; flue gas fan motor power; total air fan motor power; energy produced in steam; mills data

Output:

• NOx emission level

• CO emission level

• Outlet flue gases temperature

• Primary steam temperature.

Encouraging results

The project has been successfully commissioned with the neural supervision level using the Aspen Target controller applied in both operator advisory mode and in closed loop supervision.

Operation in the operator advisory mode showed the system`s high accuracy, stability and robustness, with good application behaviour. During tests performed, NOx emissions decreased below set limits of 460 mg/Nm3, sometimes even reaching the value of 350 mg/Nm3, after the controller was turned on. At the same time no increase in the LOI (loss of ignition) was observed, and a slight increase in average boiler efficiency was seen.

Closed-loop results have also been encouraging:

• The decrease in the NOx emission level by about 15 to 25 per cent

• Approximately 0.1 to 0.5 per cent increase in the average boiler efficiency

• Lower level of LOI – about three per cent versus five per cent

• Stable boiler action without any non predicted actions

• Robustness to the extraordinary events (rapid load changes in the limits of 135 to 200 MW, mill changes, computer reset, communication failures)

• Stable cooperation with existing WDPF control system and LDC controller

• Better steam temperature with less cooling water consumption (up to five per cent)

• Acceptance by the qualified power plant control staff.

Improvements in NOx emissions and boiler efficiency are shown in Table 1. Only a small amount of data was used to compile results for the low and high loads shown in the Table.

New contracts

In addition to Ostroleka, Transition Technologies has undertaken several similar projects in Poland as a WPC subcontractor. In January this year, a NOx control neural network system was commissioned at Polaniec power plant`s Unit 4, a 200 MW unit running on pulverised coal. NOx control testing is also being carried out on the 150 MW lignit-fired Adamow Unit 5.

Transition Technologies is also involved in two projects in the USA with WPC. In January 1999, an opacity and NOx control project was started at Florida Power and Light`s Riviera Plant. The goal of this project, implemented on the plant`s 300 MW gas and oil-fired Unit 3, is to use advanced software control to improve power generation and environmental conditions while optimising opacity and NOx control. A neural network model based predictive control system will be used in conjunction with sensor validation for NOx measurement and a burner management advisor based on Kohonen neural networks.

Again in January this year, WPC and Transition Technologies, in conjunction with EPRI, began a steam temperature control project on a 500 MW pulverised coal-fired unit at Neal North power station in Iowa. This project involves fuzzy neural networks prediction to control steam temperature and neural network-based supervision.

Fuzzy logic controller systems are capable of successfully coping with non-linear problems. Such systems have been installed at the Zeran heat plant and Kozienice power plant, both in Poland.

Click here to enlarge image

Click here to enlarge image

Figure 1. Hybrid model structure: a linear model is built and then a neural network correction is applied

Click here to enlarge image

Figure 2. The Aspen Target controller within the DCS system

Click here to enlarge image

Click here to enlarge image

Figure 3. Control performance results at the Ostroleka power plant