CA2438046A1 - Method for regulating a thermodynamic process in particular a combustion process - Google Patents

Abstract

The invention relates to a method for regulating a thermodynamic process in particular a combustion process, in which the state (St) of the system is measured and compared with optimisation targets (rj) and adjustment actions (ai) are carried out in the system for the purposes of regulation, whereby a process model (PM), independent of the optimisation targets (ri), is determined, describing the effects of actions (at) on the state (St) of the system. A situation assessment (SB), independent of the process model (PM), is carried out by means of quality functions (ut) and the state (st) of the system with regard to the optimisation targets (rj).

Description

PW 1001 WO 20 February 2002 Powitec Intelligent Technologies GmbH, D-45219 Essen, Germany Method for controlling a thermodynamic process, in particular a combustion process The invention concerns a method for controlling a thermodynamic process, in particular a combustion process, having the features of the precharacterizing clause of Claim 1.
In a known method of this type, a quality is obtained from status variables and the possible setting actions through folding, taking into account the optimization targets. In the context of a Monte Carlo method, a setting action is carried out using the previous status and then determining the new status. The resulting change in quality is a measure for the suitability of the setting action carried out to reach the optimization target. With this method, the system adapts to the next closest extreme, even when there are frequent changes in the optimization targets. However, the method is still open for improvement.
The object of the invention is to improve a method of the above type. This object is achieved through a method having the features of Claim 1. Advantageous refinements are the subject matter of the subclaims.
By the fact that a process model is determined which is independent of the optimization targets and which describes the effects of actions on the system status, and by the fact that a situation evaluation by means of quality functions, which is independent of the process model, evaluates the system status with regard to the optimization targets, a model-based control method is available which can continue to use past information both to evaluate the system and also after major changes in the status. In the known method, on the other hand, almost all information must be freshly obtained, i.e. the previous information is unlearned when adapting to a new status.
Due to economized computer power, the "process navigator" according to the present invention can be used preferably to calculate the status in advance and to execute only the appropriate setting actions for attaining the optimization target. In so doing, the process model is constantly fine-tuned in order to reach the optimization target more economically in the future. In a preferred treatment of the process model in a neural network, it is initially preferable to carry out an initialization with selected statuses and a smoothed and weighted time behaviour of the system in order to avoid random setting actions in the future.
By including the activation of a boiler blower (soot blower) as a setting action and/or the particle size of the fuel as a setting variable, there will be more setting variables available for performing the control. At the same time, the new setting variables are no longer a source of interference that would cause a greater control effort during activation. Instead of a combination of both setting variables, it is also possible to include only one of these variables for controlling. To start with, the influence of the new setting variables is preferably measured in an initialization phase in order to train the controlling process; this is done, for example, by means of a neural network.
The inclusion according to the invention can be applied in combustion processes and other thermodynamic processes.
In the case of a boiler blower, the decision on its activation is made preferably by weighing the effort and the success, for example by defining fictitious costs and integrating them during a given time interval, so that the decision can be fully automated. When factoring the particle size of the fuel, one can additionally take into consideration the wear and tear on the assigned mill in order to fine-tune the controlling process. The effect of different particle sizes is preferably measured through the fluctuations in the ray image of a flame. With the information gained on the effect of the activation of the boiler blower, the particle size of the fuel and the wear and tear on the mill, it is also possible to control either the boiler blower and/or the mill independently without having to control the combustion process for this.
If, in systems with sampling devices, the computer connected with the sensors determines the moment for activating the sampling device, this, on the one hand, economizes resources, since only the necessary number of sample analyses have to be carried out; on the other hand, there is more data available in critical situations for a decision process in the computer; this, for example, optimizes the controlling process.
Sampling is thus completely integrated into the monitoring process and preferably into the controlling process. In addition to the sampling device, one could imagine other "off line" sensors feeding data into the computer together with the "on-line"
sensors.
Activating the sampling device can be done manually or automatically.
The following figure shows the invention in detail with reference to one exemplary embodiment with examples of applications.
Fig. 1 is a schematic structure of an exemplary application of the method according to the invention, Fig. 2 is a time diagram of an exemplary process parameter over several cleaning intervals, Fig. 3 is a time diagram of the costs in the soiling and cleaning model, and Fig. 4 is a schematic structure of another example of application.

The exemplary process occurs as a combustion process in a fire boiler of a coal power plant and is to be controlled in such a way as to display, on the one hand, a certain stability and, on the other hand, a certain flexibility, i.e. it adapts to a given situation.
The status in the boiler is described by (time-dependent) status variables st, for example the temperature in different locations inside the boiler, the flame textures and/or concentration of different pollutants in the exhaust air which are determined by suitable sensors. In the formal description of the combustion process, the parameters and variables, such as st, are to be considered as mufti-dimensional vectors. The status st of the combustion process can be changed by different actions at, in particular by changing the setting variables, for example the supply in coal, core air or expulsion air, but also the quality of the coal. There are optimization targets r' for the combustion process, for example to achieve a concentration of nitrogen oxide NOX and carbon monoxide CO
which is below pre-determined limits or approaches a minimum.
For on-line monitoring and controling and for predicting future boiler statuses by means of a neural network according to the invention, one defines, on the one hand, a process model PM reporting the change in the status st of the combustion process as a reaction to actions at. The process model PM is independent of the optimization targets r' and operates on a time window of previous process statuses in order to integrate a time context. The concatenation of the variables contains information on specific properties of the boiler. On the other hand, one defines a quality ut for a situational evaluation SB, in which a specific current status st is evaluated by taking into consideration the optimization targets r'. This definition is done from the point of view of predetermined characteristic curves and a fuzzy logic concept, which means that in the simplest case, the quality at a certain pollutant concentration, for example, is ut = 1 for a minimum pollutant concentration and u~ = 0 when the upper pollutant concentration threshold is reached, with a linear dependence between these two values. The situational evaluation SB is dependent on the actions at. For the evaluation at the time point t and the step towards the time point t+1, the following equation applies:
uc - E u~~'~ st) TSB) st+1 - f(st, ac) (PM) In contrast to the known quality function which folds the system status, the actions and the optimization targets so that the entire quality function must be redetermined in the event of changes in the status, the process model PM remains unaffected in the method according to the invention. In a numerical realization by means of a neural network, this means that it is not necessary to readapt the entire neural network for each action, losing the information from the past, which implies losing processing capacity and time, but that information on the process model is maintained. In an imagined representation of the status space in which any status st is symbolized by a point on a map, the number of points on the map increases without loosing points.
The economized computer power can be used to calculate future system statuses and thus achieve an overall instead of a local optimization. To do so, a total quality Q is defined in which qualities ut for several pre-calculated situational evaluations are taken into consideration. For a number N of pre-calculated time steps, the following equation applies:
N
~''~st~at) - ~ ut+n n=1 By reason of the pre-calculations it is no longer necessary to make the boiler initiate different statuses by itself by means of a random walk model in order to fmd the conditions for the optimal combustion process.

In order to build the process model PM, one first determines the setting variables for the network architecture and the available status variables together with their technically possible and reasonably useable threshold values; preferable extension options are considered. A noise filter function to be applied prior to input into the neural network will be defined for the status variables that are to be measured. By using the method of optimal test planning, an exploration plan is designed which selects some statuses (approximately 500) being considered as reasonable from within the threshold values.
By means of specific setting actions, these statuses are now successively produced in the boiler, at least approximately, and are then stored in the neural network. In this exploration it is useful, following a setting action with the help of stored past status variables which are smoothed and weighted (relation of the status variables before, during and after the setting action), to create a multi-channel time function which can be used to produce orderly changes in the status.
On the basis of this initialized process model PM, the boiler is optimally controlled during operation by first, in the neural network in dependence on the optimization targets r' on the basis of the present status, starting a numerical search for the overall system optimum and the optimal setting actions for reaching said target. The optimal setting actions can be achieved as follows: on the basis of the present status sc, the effects of different actions a' are determined, in particular changes in the significant setting variables, i.e. the statuses s'c+1 are calculated through the process model PM
and the quality u'c+1 is then calculated through the situational evaluation SM, if applicable several time steps in advance. The quality u't+i and/or the total quality Q' then automatically determines which setting variable is to be changed.
st --~ al -~ slt+i -~ u'c+1 ...--~ Q' a' -~ s't+1 --~ ut+i ...--~Q~

Once the optimal setting actions a' have been determined, the control electronics of the boiler carry out said setting actions a' automatically, for example by changing the supply of different fuels that is treated as a setting variable. In so doing, status variable data is continually collected and validated in the boiler in order to increase the density of known statuses st on the map of the status space by training the neural network.
If, once the optimum has been reached, other actions at occur, i.e. intended setting actions or unintended disturbances, renewed optimization using the described steps will be necessary. When training the process model PM, it may also be useful or necessary to extend the network architecture, for example beyond the threshold values considered as useful towards the technically possible threshold values if, while searching for the optimum, the neural network frequently reaches a threshold value that is considered as useful. Significance testing constantly monitors which setting variables have the greatest effects on the system status, so that their changes will be preferably taken into account when searching for the optimum, in particular if a major status change is necessary.
Generally speaking, it would be sufficient for controlling the boiler to take the (usual) setting variables from the actions and some characteristic process parameters from the status variables into account. However, if in this case, disturbance parameters, such as fluctuations in the quality of the coal or wear and tear in the boiler K, are also to be used as input into the process model, it will be necessary to monitor the inside of the boiler, for example by monitoring the flames. The following is a description of some application examples.
In a first application example, a combustion process to be controlled is occurring in a system in a fire boiler 1 of a coal power plant. Upstream from the boiler 1, there is a coal mill 3 grinding the coal to be burned which is then fed into the boiler 1. The boiler 1 also receives core air and expulsion air. The exhaust air produced is evacuated from the boiler. A number of sensors 5 in the feeder lines to the boiler l, in the evacuation lines and on the inner walls of the boiler measure the relevant process parameters, for example the supplied air and coal, pollutant concentrations in the exhaust air, temperatures in the boiler 1, flame characteristics, etc. A so-called boiler blower 6 is provided for cleaning the inner boiler wall; when activated, it burns the residues on the inner boiler walls.
The sensors 7 are connected to a computer 9 to which different setting devices 13 are also connected, thereby creating a feedback loop or control loop. The combustion process is influenced by changing the setting variables, for example the coal stream and the air stream, i.e. the process parameters change. The activation of a setting device 13 such as a valve in an air feeder line will be referred to as an action in the following text.
A self calibrating and self learning, recurrent or time-delayed neural network of the type described above is implemented in the computer 9. The system configured this way is used for optimizing the combustion process, i.e. the computer 9 evaluates the input it receives and then polls the setting devices 13 accordingly.
The activation of the boiler blower 6 will be treated in an exemplary fashion as an action.
For this purpose, it will be examined in an initialization phase how quickly the boiler 1 becomes soiled with different coal varieties and loads and how thorough the cleaning is after activating the boiler blower 6 in dependance on different time intervals Ot between two activations of the boiler blower 6. The cleaning result can be defined by different action results such as the change in pollutant concentration C in the exhaust air normed for the air flow or the change in efficiency.
The neural network learns this technical aspect of the soiling and cleaning model, i.e the kind of effects from two possible actions: activating the boiler blower 6 and not activating it. In addition, the neural network learns a cost aspect of the soiling and cleaning model, i.e. costs for using the boiler blower 6 are determined, for example by measuring energy loss or down-times, in the same way as they are determined for low efficiency or for exceeding the permissible pollutant thresholds in the exhaust air.
Once the neural network has learned the soiling and cleaning model, the boiler blower 6 S can be used for controlling the combustion process. Using current process parameters, the neural network estimates the effect of activating or not activating the boiler blower 6 on the process parameters, and estimates the costs in these two cases. The costs K for both cases are integrated in dependence on a pre-determined planning interval fit. In Fig. 3 the cost integral with activated boiler blower is obliquely hatched and the cost integral without activated boiler blower is vertically hatched. The computer 9 will then decide in dependence on this estimate whether the boiler blower 6 is to be activated or not.
The particle size of the ground coal is treated as another setting variable, i.e. actions to change said setting variable are different setting actions in the controllable coal mill 6, which have an effect on the fineness of the ground coal, e.g. the engine speed of the drums, the angle of the flaps, the down pressure of the drums or the temperature of the evacuation air. The corresponding setting devices are connected to the same computer 9 in which the neural network is implemented.
Accordingly, in the initialization phase one examines the change in the process parameters in dependence on the coal particle size. For example, one can perform a frequency analysis in the local time/space on a (spatially and) time-resolved radiation measurement of the flames in the boiler 1, for example by means of a time-delay neural network or so-called w-flat. The shape, height, breadth and, if applicable, a peak shift can provide information on fluctuations and thereby on the coal particle size which cause said fluctuations in the flame image.

The neural network learns this mill model obtained from the initialization phase and improves it during the operation of the boiler 1 by means of the data provided by the sensors 7. A measuring device is installed in the coal mill 3 as a further source of information, for example a (special) video camera, such as a CCD camera, a depth tactile measuring device or a structure sample measuring device for calibration which, at every stop of the coal mill 3, for example during maintenance, provides a status report which is then subjected to a characteristics analysis in the computer 9, in particular to determine wear and tear. The information obtained this way is correlated with the other information on the combustion process and is available for the learning process of the neural network. The mill model is therefore also used for controlling the combustion process.
In a modified form, controlling can include - in addition to the usual setting variables -the soiling and cleaning model with the boiler blower 6, but without considering the coal particle size, or the the mill model without the boiler blower 6.
In another exemplary application, the combustion process in a boiler 1 of a garbage incinerator is constantly monitored with several sensors 5, for example with a camera 7 which captures the flame image, and with exhaust sensors 8 which monitor the gaseous combustion products. The different sensors 5 are connected to a computer 9 in which a neural network of the type described above is implemented determining the status in the garbage incinerator 1 by means of the signals coming from the sensors 5. Said status, for example, is displayed on a monitor 11, among other things, for example, as a live image.
To control the combustion process, for example with a view to obtaining the lowest possible pollutant concentration in the combustion products, the computer 9 can poll different setting devices 13 which define, for example, the air supply or the supply of additional fuel. The status in the boiler 1 then changes through a change in said setting variables which is calculated by the computer 9.

In addition to the aforementioned "on-line" sensors 5, there is a sampling device 15 with a sample evaluation device 17 connected to it, together working as an "off line" sensor.
Once the computer 9 detects that the status in the boiler 1 is approaching a critical level and additional data will be useful for the controlling strategy, the computer 9, at a time it itself determines, emits a signal for activating the sampling device 15.
In the simplest case, the sampling device 15 is manually activated once the sampling order has appeared on the monitor 11, and the sample taken is forwarded to the sample evaluation device 17. In an automated version, the sampling device 15 polled by the computer 9 is machine-activated and the sample taken automatically forwarded to the sample evaluation device 17. In all cases, the sample evaluation device 17 sends the result of the analysis, preferably a chemical analysis, to the computer 9 which includes the data in the controlling process.

Claims (12)

Claims

1. Method for controlling a thermodynamic process, in particular a combustion process, in which the system status (s t) is measured and then compared with optimization targets (r j), and in which setting actions (a i) suitable for controlling are performed in the system, characterized in that a process model (PM) is determined which is independent of the optimization targets (r j) and which describes the effects of actions (a t) on the system status (s t), and in that a situational evaluation (SB) which is independent of the process model (PM) evaluates the system status (s t) by means of quality functions (u t) with regard to the optimization targets (r j).

2. Method according to Claim 1, characterized in that before actually carrying out setting actions (a i) in the system, the effects of different setting actions (a i) on the system status (s t) are numerically calculated using the process model (PM) and are evaluated individually using the situational evaluation (SB), whereupon the the optimal setting actions (a i) in the system are performed.

3. Method according to Claim 2, characterized in that the effects of different setting actions (a i) on the system status (s t) are calculated several time steps (t) in advance and that an overall quality (Q) is determined.

4. Method according to one of the preceding claims, characterized in that the process model (PM) is constantly fine-tuned with information from the ongoing measurements of the system status (s t).

5. Method according to Claim 4, characterized in that the results of the ongoing measurements, prior to feeding them into the process model (PM) are subjected to a noise filter.

6. Method according to one of the preceding claims, characterized in that the process model (PM) is stored in a neural network which is constantly trained.

7. Method according to one of the preceding claims, characterized in that prior to continuous use of the system, an exploration of a number of selected system statuses (s t) is performed to initialize the process model (PM).

8. Method according to Claim 7, characterized in that the exploration takes into account the smoothed and weighted time behaviour of the system on a setting action (a i).

9. Method according to one of the preceding claims, characterized in that the quality functions (u t) are set up according to the fuzzy logic rules.

10. Method according to one of the preceding claims, characterized in that for controlling a combustion process in a boiler (1) fired with solid fuel, the activation of a boiler blower (6) and/or the particle size of the fuel are used as setting variables.

11. Method according to one of the preceding claims, characterized in that a combustion process taking place in a boiler (1) is constantly monitored by at least one sensor (5) and evaluated by a computer (9) connected to the sensor (5), in that for this, a sample can be taken from the boiler (1) by means of a sampling device (15), and in that the computer (9) determines the time for activating the sampling device (15).

12. Device for carrying out a method according to one of the preceding claims, having sensors (5) for measuring the system status (s t), a computer (9) for applying the process model (PM) for situational evaluation (SB), and feedback to the system for performing setting actions (a i).