The performance of sub-loop controls and especially the unit control in steam power plants is of critical importance to their cost effectiveness. Improved control quality results in:
• Higher unit efficiency – by enabling the power plant owner to operate the unit closer to the allowable limits, reducing LOI (loss on ignition) in coal-fired power plants and reducing local firing inefficiencies and peaks in flue gas temperature.
• Higher unit flexibility – by enabling the unit to perform faster load ramps and reducing the minimum sustainable load.
• Higher unit availability – by achieving smoother unit operation and therefore reduced wear and tear.
• Reduced emissions – through reduced fuel consumption due to increased unit efficiency and, specifically, lower NOx emissions due to the reduction in local firing inefficiencies and therefore temperature reductions in the furnace.
Effective control concepts are the basis for achieving these goals, one such being the SPPA-P3000 process optimisation package, which was developed by Siemens Energy. This incorporates:
• model-based, predictive feed-forward controls;
• use of inherently stable processes; and
• decoupling of highly intermeshed sub-processes.
See Figure 1.
Based on these modern control principles a high unit operating performance and therefore a high unit profitability can be achieved.
In Figure 2 the basic structure of the SPPA-P3000 unit control is shown. The optimal trends for the turbine and boiler load setpoints are calculated in the model-based feed-forward structure, depending on the dynamic setpoints for the electrical load and the main steam pressure.
Since the unit control performance increases with the accuracy of the feed-forward calculation, it is important to make sure that the model parameters are well determined. One of the most important model parameters is the heating value of the fuel currently being used.
Therefore, the more accurately the heating value of the fuel is known, the more accurate the feed-forward calculation.
This is particularly significant for load changes like load ramps or frequency disturbances. If the heating value is inaccurate, the correction controller has to step in. And in this case the control performance will not be optimal.
Therefore, the heating value correction concept is of high importance. The SPPA-P3000 unit control uses the boiler as a calorimeter and calculates the heating value of the current fuel by solving the mass- and energy-balance across the boiler. This method is very accurate and fast.
Another important input value for the feed-forward calculation is the unit efficiency. There are a lot of factors that have an influence on efficiency, such as:
• unit load;
• fuel type and fuel parameters;
• main steam parameters (temperature, pressure);
• cooling water temperature; and
• outside air temperature.
By using a water-/steam-cycle calculation program, the influence of these different factors on the unit efficiency can be calculated. However, a major effort is required to set up and especially to parameterise such a unit model. Furthermore, if the unit changes over time due to wear and tear and fouling the efficiency calculation is rendered inaccurate once more.
Therefore in most cases in a unit control system the efficiency value depends on load only, or a constant value might even be used. It is obvious that, due to this inaccuracy, reduced unit control performance must be accepted.
However, with a neural unit control structure the model-based feed-forward calculation of the boiler load setpoint can be done with very high accuracy. In this structure a neural network is used to calculate the unit efficiency, taking into account several factors based on on-line data measurements.
A two-phase approach is required to introduce the neural network concept into the unit control. The two phases are:
• training of the neural network by using measurements from the data archive; and
• implementation and on-line operation of the neural network.
Every now and then the training of the neural network can be repeated with actual data, thus the accuracy of the unit control does not degrade over time. This can be done by a member of the power plant staff.
The neural unit control can run on-line as a fully integrated part of the SPPA-T3000 control system, or it can be implemented as external feature of other distributed control systems.
The Velsen 25 experience
The SPPA-P3000 neural network was tested in Nuon’s Velsen 25 power plant unit in the Netherlands, see Figure 3.
Velsen 25 is equipped with a triple fuel once through boiler and the unit’s peak power is 403 MW net.
Unlike natural-gas-fired-only power stations, Nuon’s Velsen 25 uses gas produced during steel production at the Corus steel factory. Residual gases such as blast furnace gas or coke gas are produced as waste in the furnaces of the steel manufacturer and then sent to the power station. Usage of residual gases in a power plant leads to substantial environmental benefits, as it prevents the need to burn these in flares unused.
The heating value of blast furnace gas is low, so it is hard to burn just these gases in the boiler. Therefore high-calorific value natural gas is burned as well. Natural gas serves as a support fuel for stable combustion of the blast furnace gas. The natural gas is also used as control fuel for the unit load control and as regulator for fluctuations in the blast furnace gas and coke gas supply.
Velsen 25 provides a particularly good case study of the implementation of the neural network approach to efficiency calculation. The efficiency of the unit depends highly on the actual fuel mixture, which is the mass flow of blast furnace gas and coke gas in relation to the flow of natural gas. But also the actual heating values of the gases, which are variable, at least for the blast furnace gas and the coke gas, have an influence on the efficiency. A calculation of these interrelations based on a physical model can be considered impossible. Therefore, it was necessary to find an alternative solution.
Training of the neural network
During an initial investigation the dependency of the Velsen 25 unit efficiency on the following variables was checked:
• load;
• flow of blast furnace gas;
• heating value of blast furnace gas;
• flow of coke gas;
• heating value of coke gas;
• heating value of natural gas;
• cooling water temperature;
• main steam temperature; and
• main steam pressure.
It became apparent that it is not necessary to use all these variables for the efficiency calculation. Instead, after analysis, the variables shown in Figure 4 were used.
For the training of the neural network a whole year’s worth of data is used. Thus the effects of the different seasons, weather and outside temperature changes, as well as different shifts and settings in the power station, are eliminated. The data is scanned for periods with steady state unit operation. For each of these periods values for the input variables as well as for the calculated unit efficiency are saved. This data selection is done automatically by using a MATLAB script. The script also makes sure that data that is not meaningful, such as data stored in the DCS archive during a unit shutdown, is not used.
By this means, a number of data sets are created which include information on how the efficiency depends on the input variables. These data sets are used for the training of the neural network.
Figure 5 shows how the unit efficiency of Velsen 25 depends on the unit load and the flow of the blast furnace gas, assuming constant values for the blast furnace gas heating value as well as for all the other input variables. The high degree of non-linearity can be seen and one can imagine how complex a fully multi-dimensional function for the efficiency would be in reality.
Implementation and operation of the neural network
Velsen 25 is equipped with the SPPA-T2000 control system. The neural network runs on a separate industrial PC which exchanges signals with the DCS via the plant bus. With a modern control system such as SPPA-T3000 the neural network can run in the control system itself and is then a fully integrated part of the unit control.
Via the DCS the neural network is supplied with on-line measurements and delivers a new value for the unit efficiency in each calculation cycle of the system. Once new archive data is available in the DCS, further training of the neural network can be carried out. Adjustment of the neural network parameters ensures that calculation of the unit efficiency is adapted to the actual unit performance, see Figure 6. It is recommended that neural network training is done once every one to three months.
Results for Velsen 25
Due to the fact that the unit efficiency of Velsen 25 depends heavily on a variety of factors, such as fuel mixture and heating values, it is very difficult to determine an accurate value. However, it can be estimated that using the neural efficiency calculation results in a maximum inaccuracy of the calculated boiler load setpoint of approximately 0.5%. This high accuracy leads to a higher unit control quality. The following resulting economic benefits can be noted for Velsen 25:
• Smoother unit operation with a corresponding reduction in wear and tear.
• More stable combustion with a corresponding positive effect on unit efficiency.
• More precise boiler operation resulting in a more precise control of the electrical power output. And
• Less limitations on electrical power output. The unit efficiency is used to calculate the maximum allowable load for the boiler. With inaccurate efficiency values, false maximum boiler loads are applied, which leads to a invalid limitations on the maximum electrical power output.
Neural networks: making the seemingly impossible possible
In summary, the control performance of steam power plant units is of critical importance for their cost effectiveness and high quality feed-forward controllers have the advantage over feedback loops that the tracking behaviour for relevant signals improves dramatically leading to corresponding economic benefits.
An accurate feed-forward calculation of the boiler load demand in the unit control requires knowledge of the unit efficiency. Since the efficiency is highly non-linear and dependent on several internal and external process values, it is very difficult to accurately determine what the actual current efficiency is. Therefore a concept based on actual historical data, fed into a neural network model, has been applied. This has proved to be a cost effective way of modeling with a short lead time.
By using a neural network, trained on actual measured data, it has proved possible to arrive at an accurate estimate of unit efficiency. Repeated training with data based on new measurements ensures that the neural network accounts fully for changes in plant behaviour, for example arising from wear and tear.
The calculation can be done on-line as a fully integrated function of the Siemens SPPA-T3000 control system. It can also be used as a third party supplied addition to other control systems.
Siemens SPPA-P3000 neural networks can be used in various ways and for different purposes. Whenever the mathematical description of a process becomes too complex, or even impossible, a neural network can be employed. Other possible applications are the calculation of:
• main steam flow without efficiency reducing differential pressure measurements;
• energy content of combined cycle plant exhaust gases, for a better control of heat recovery steam generators; and
• boiler emissions, leading to improved control of combustion.
As the neural network is a fully integrated module within the distributed control system and can be freely parametrised and monitored, no adjustment to the specific application is required.
