The present invention relates generally to methods and apparatus for operation of a combined cycle power plant (also referred to herein as “CCPP”) which combines the use of both gas and steam turbines in a single power generating plant, and more specifically to methods and apparatus for optimizing operation of a CCPP.
In recent years, CCPPs have become more common for generation of electric power due to the high efficiencies achieved with CCPPs, as compared with conventional power generating plants. A typical CCPP includes, but is not limited to: at least one gas turbine (also referred to as a “gas combustion turbine”), at least one steam turbine, and at least one heat recovery steam generator (HRSG), which is also referred to as a “heat recovery system” or a “boiler”. The gas turbine (GT) produces electric power from a fuel source, the HRSG generates steam by capturing heat from the exhaust of a GT, and the steam turbine (ST) produces electric power using steam produced by the HRSG. To boost performance of the HRSG, the HRSG may also include gas burners (also referred to as “duct burners”) to increase the amount of heat entering the HRSG.
In general, a gas turbine operates by pulling in air from the outside and pressurizing the air to provide compressed air to a combustion chamber. In the combustion chamber the compressed air is ignited by burning fuel (e.g., low sulphur fuel oil or natural gas). As the compressed air is heated it expands, thereby causing a turbine to rotate. The rotating turbine turns a generator that generates electric power. Heat exiting the gas turbine as exhaust gas (i.e., “waste heat”) is used as an energy input to the HRSG. As indicated above, the HRSG may also include duct burners to provide additional energy input. The output of the HRSG is high temperature steam that is supplied to a steam turbine that also generates electric power in connection with a generator.
The above-mentioned components of the CCPP may be arranged in a variety of configurations. For example, the components may be arranged in a single train where a gas turbine feeds a HRSG with a steam turbine, or two GT/HRSG pairs may share a single steam turbine. Furthermore, the components of a CCPP may be configured as elements of one or more “blocks.” Each block is comprised of at least one ST and at least one GT. The number of “blocks” in a CCPP is determined by the number of STs in the plant, while the number of “units” in the plant is determined by the number of GTs in the plant. Each GT and each ST produces electric power. A CCPP may also include one or more air pollution control (APC) devices for removal of pollutants from flue gas; at least one stack for release of flue gas; and at least one water cooling system for condensing high temperature steam. An example of one typical CCPP will be described in detail below.
CCPPs are currently the second leading source of electric power in the United States. A CCPP can provide higher efficiency and can often ramp up and down on load more rapidly than coal-fired power generating units. For this reason, CCCPs are often used to provide load balancing to the electric grid as load across the grid increases and decreases.
Allocation of load to individual power plants in the electric grid is determined automatically by regional power authorities (e.g., an independent system operator (ISO) or a regional transmission organization (RTO)). Typically, a power generating plant is connected to a system referred to as Automatic Generation Controller (AGC). The AGC determines the load required for a given plant. In cases where an AGC is not available, another system is used to determine the overall generation requirements for the plant.
Given the load requirements, determination of the load within a power generating plant may need to be done at the site of the power generating plant. For a CCPP, it is common for the AGC system to establish a plant-wide load for the entire CCPP, and the CCPP then makes an on-site allocation of the established plant-wide load among the various turbines or blocks of the CCPP. The most common method of allocation of the plant-wide load is based upon simple “rules of thumb” or lookup tables. For example, a first GT may take the lowest low range and ramp up production of electric power to a certain point after which a second GT may be turned on and ramp up to the next level of electric power production.
Although such “rules of thumb” or look up tables are easy to implement, they do not provide optimal performance of the CCPP. For example, if the first GT is more efficient than the second GT, then it would be more economical to start the first GT prior to starting the second GT. Similarly, if starting the burners on a first GT is more efficient that starting the burners on a second GT, then it would be more economical to start the burners on the first GT prior to starting the burners on the second GT.
The present invention overcomes drawbacks of the prior art, and provides methods and apparatus for optimizing operation of a combined cycle power plant.
In accordance with the present invention, there is provided an optimization system for a combined cycle power plant having one or more gas turbines, one or more steam turbines, and one or more boilers associated with the one or more steam turbines, wherein a duct burner is associated with at least one of said boilers, said optimization system comprising: a load prediction model for determining a predicted maximum load for the plant; a plant optimization model including: (i) a plant power model for determining a predicted plant power produced by the plant, wherein said predicted plant power is determined by summing a total predicted gas turbine power produced by the one or more gas turbines and a total predicted steam turbine power produced by the one or more steam turbines, and (ii) a duct burner power model for determining a predicted duct burner power indicative of plant power due solely to one or more duct burners that are associated with the one or more boilers for producing steam for the one or more steam turbines; and an optimizer for determining optimal setpoint values for manipulated variables associated with operation of the plant, given (a) a goal associated with operation of the plant and (b) constraints associated with operation of the plant, wherein the optimizer uses said predicted maximum load for the plant, said predicted plant power produced by the plant and said predicted duct burner power to determine the setpoint values.
In accordance with another aspect of the present invention, there is provided a method for optimizing operation of a combined cycle power plant having one or more gas turbines, one or more steam turbines, and one or more boilers associated with the one or more steam turbines, wherein a duct burner is associated with at least one of said boilers, said method comprising the steps of: determining a predicted maximum load for the plant; using a plant optimization model to (i) determine a predicted plant power produced by the plant and (ii) determine a predicted duct burner power indicative of plant power due solely to one or more duct burners that are associated with the one or more boilers for producing steam for the one or more steam turbines; and using an optimizer to determine optimal setpoint values for manipulated variables associated with operation of the plant, given (a) a goal associated with operation of the plant and (b) constraints associated with operation of the plant, wherein the setpoint values are determined by the optimizer using said predicted maximum load for the plant, said predicted plant power produced by the plant and said predicted duct burner power.
An advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant.
Another advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve optimal allocation of load within the plant.
Another advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve greater efficiency in the allocation of fuel within the plant.
Still another advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve optimal performance of the plant.
Still another advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve improved reliability of the plant.
A still further advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve improved capacity of the plant.
Yet another advantage of the present invention is the provision of an optimization system for optimizing operation of a power generating plant to achieve reduced plant emissions.
These and other advantages will become apparent from the following description taken together with the accompanying drawings and the appended claims.
The invention may take physical form in certain parts and arrangement of parts, embodiments of which will be described in detail in the specification and illustrated in the accompanying drawings which form a part hereof, and wherein:
Referring now to the drawings wherein the showings are for the purposes of illustrating embodiments of the invention only and not for the purposes of limiting same,
In general, for a CCPP, fuel (typically natural gas) is input into a set of gas turbines and boilers. As indicated above, a boiler can also be referred to as Heat Recovery Steam Generator (HRSG). The gas turbines produce both electric power and heated flue gas that flows into a boiler. Duct burners in the boilers (or near the front of the boilers) can be used to add additional heat to the flue gas entering the boilers. The boilers produce steam which is used to power steam turbines that produce electric power.
In the illustrated plant 170, there are four gas turbines (identified as GT1, GT2, GT3 and GT4), four boilers (identified as B1, B2, B3 and B4) having associated duct burners (identified as DB1, DB2, DB3 and DB4), and two steam turbines (identified as ST1 and ST2). As can be seen in
The power respectively produced by the gas turbines GT1-GT4 is identified as GT1 Power, GT2 Power, GT3 Power and GT4 Power. The power respectively produced by the steam turbines ST1, ST2 is identified as ST1 Power and ST2 Power. The total overall power produced by plant 170, known as plant load, is the summation of the power produced by the four gas turbines GT1-GT4 and the two steam generators ST1, ST2.
In a typical DCS, control elements are not only located in a central location, but are also distributed throughout a system with each component sub-system controlled by one or more controllers. The entire system of controllers is connected by networks for communication and monitoring. A DCS also includes input and output modules. The controllers receive data from input modules and sends data to output modules. The input modules receive data from input components (e.g., sensors 215) at plant 170, and the output modules transmit instructions to output components at the plant (e.g., actuators 205). The inputs and outputs can be either analog signals which are continuously changing or discrete signals which are two-state (either on or off). Buses connect the controllers and modules through multiplexer or demultiplexers. The buses also connect the controllers with a central controller and finally to an operator interface or control console (not shown). The operator interface provides means for an operator to communicate with DCS 150. DCS 150 may also communicate with a historian (not shown).
As described with respect to
Model(s) 120 will now be broadly described. In this respect, model(s) 120 are used to represent the relationship between (a) manipulated variables (MV) and disturbance variables (DV) and (b) controlled variables (CV). Manipulated variables (MVs) may be changed by the operator or optimization system 100 to affect the controlled variables (CVs). As used herein, disturbance variables refer to variables (associated with components of the power generating plant) that affect the controlled variables, but cannot be manipulated by the operator (e.g., ambient conditions at the power generating plant). Optimizer 110 determines an optimal set of setpoint values for the manipulated variables given (1) a desired goal associated with operation of the power generating plant (e.g., minimizing fuel consumption) and (2) constraints associated with operation of power generating plant (e.g., meeting required power demand).
At a predetermined frequency (e.g., every 10-60 seconds), optimization system 100 obtains the current values of manipulated variables, controlled variables and disturbance variables from DCS 150. An “optimization cycle” commences each time the current values for the manipulated variables, controlled variables and disturbance variables are read out from DCS 150.
As will be described in further detail below, optimization system 100 uses model(s) 120 to determine an optimal set of setpoint values for the manipulated variables based upon current conditions of plant 170. The optimal set of setpoint values are sent to DCS 150. An operator of plant 170 has the option of using the optimal set of setpoint values for the manipulated variables. In most cases, the operator allows the computed optimal set of setpoint values for the manipulated variables to be used as setpoint values for control loops. Optimization system 100 runs in a closed loop adjusting the setpoint values of the manipulated variables at a predetermined frequency (e.g., every 10-60 seconds) depending upon current operating conditions of power generation block 200. Optimization systems are described in U.S. Pat. No. 8,295,953 to Piche (“System for Optimizing Power Generating Unit”), issued Oct. 23, 2012, which is fully incorporated herein by reference.
It should be understood that the optimization system (including optimizer and model(s) described herein) may be implemented in various different ways well known to those skilled in the art. These implementations include the use of one or more programmed computer systems. Each computer system may include one or more processors, one or more controllers, data storage devices (e.g., memory, hard drive, etc.), input devices (e.g., keyboard, mouse, touch screen and the like), and output devices (e.g., display devices such as monitors and printers). The computer system may communicate with components of the power plant via any suitable data communications medium including, but not limited to, a wired network, a wireless RF network, a fiber optic network, telephone lines, the Internet, or combinations of these mediums.
To properly capture the relationship between the manipulated/disturbance variables and the controlled variables, model(s) 120 may have the following characteristics:
Given the foregoing requirements, a neural network based approach is presently the preferred means for implementing models in accordance with the present invention. Neural networks are developed based upon empirical data using advanced regression algorithms. See, for example, C. Bishop, Pattern Recognition and Machine Learning, Springer, New York, N.Y., 2006, fully incorporated herein by reference. Neural networks are capable of capturing the nonlinearity commonly exhibited by boilers. Neural networks can also be used to represent systems with multiple inputs and outputs. In addition, neural networks can be updated using either feedback biasing or on-line adaptive learning. Finally, neural networks can be developed to take disturbance in account, as described in U.S. Pat. No. 7,123,971 to Piche (“Non-Linear Model With Disturbance Rejection”), issued Oct. 17, 2006.
Dynamic models can also be implemented in a neural network based structure. A variety of different types of model architectures have been used for implementation of dynamic neural networks, as described in S. Piche, “Steepest Descent Algorithms for Neural Network Controllers and Filters,” IEEE Trans. Neural Networks, vol. 5, no. 2, pp. 198-212, 1994 and A. Barto, “Connectionist Learning for Control,” Neural Networks for Control, edited by W. Miller, R. Sutton and P. Werbos, MIT Press, Cambridge, Mass., pp. 5-58, Jan. 3, 1990, both of which are fully incorporated herein by reference. Many of the neural network model architectures require a large amount of data to successfully train the dynamic neural network. A novel neural network structure, which may be trained using a relatively small amount of data, was developed in the late 1990's. Complete details on this dynamic neural network based structure are provided in S. Piche, B. Sayyar-Rodsari, D. Johnson and M. Gerules, “Nonlinear Model Predictive Control Using Neural Networks,” IEEE Control Systems Magazine, vol. 20, no. 2, pp. 53-62, June 2000, which is fully incorporated herein by reference.
Given a model of a power generating plant, it is possible to determine the effects of changes in the manipulated variables on the controlled variables. Furthermore, since the model is dynamic, it is possible to determine the effects of changes in the manipulated variables over a future time horizon (i.e., multiple changes rather than a single change).
Empirical modeling techniques are described above for developing relationships between input and outputs of systems to be modeled. Empirical models work well in data rich and knowledge poor situations, i.e., the relationship between input and output is not well understood, but there are large amounts of data available to “learn” the relationship.
In situations where the relationship between the input and output is well known, that relationship can be used directly in the model. For example, equations representing performance of a feedwater heater, a type of heat exchanging commonly used in power generating plants, are well known and published in textbooks on the subject. These equations can be written directly into a model of a feedwater heater. The parameters of such models would be based upon the design parameters of the equipment which is commonly available. Thus, given the design equations, a very precise model of a piece of equipment can be developed based upon “first principle” knowledge of the system. Such models are referred to as “first principles models.”
In power generating plants, well known thermodynamic equations can be used for developing models of many of the components of both coal-fired and combined cycle power plants. See K. C. Cotton, Evaluating and Improving Steam Turbine Performance, Second Edition, Cotton Fact, Rexford, N.Y., 1998. This may include models of the gas turbine, steam turbine, HRSG, condenser, feedwater heaters and other major components. Since these models are based upon thermodynamic equations that are based upon well known equations, they are also referred to as rigorous thermodynamic models.
These models may be interconnected to form a larger model of an entire unit, entire power generating block or an entire power generating plant. Because there is feedback among the components, it may be necessary to use an optimizer to solve for the overall model.
An optimizer is used to minimize a “cost function” subject to a set of constraints. The cost function is a mathematical representation of a desired goal or goals. For instance, to minimize fuel flow, the cost function includes a term that decreases as the level of fuel flow decreases. One common method for minimizing a cost function is known as “gradient descent optimization.” Gradient descent is an optimization algorithm that approaches a local minimum of a function by taking steps proportional to the negative of the gradient (or the approximate gradient) of the function at the current point.
Since the model is dynamic, the effects of changes must be taken into account over a future time horizon. Therefore, the cost function includes terms over a future horizon, typically 30 minutes for CCPP optimization. Since the model is used to predict over a time horizon, this approach is commonly referred to as model predictive control (MPC). Model Predictive Control is described in detail in S. Piche, B. Sayyar-Rodsari, D. Johnson and M. Gerules, “Nonlinear Model Predictive Control Using Neural Networks,” IEEE Control Systems Magazine, vol. 20, no. 2, pp. 53-62, 2000, which is fully incorporated herein by reference. Also see E. Camacho and C. Alba, Model Predictive Control, Springer, New York, N.Y., 2007.
Constraints may be placed upon both the inputs (MVs) and outputs (CVs) over the future time horizon. Typically, constraints that are consistent with limits associated with the DCS are placed upon the manipulated variables. Constraints on the outputs (CVs) are determined by the problem that is being solved.
A nonlinear model can be used to determine the relationship between the inputs and outputs of a plant. Accordingly, a nonlinear programming optimizer is used to solve the optimization problem in accordance with an embodiment of the present invention. However, it should be understood that a number of different optimization techniques may be used depending on the form of the model and the costs and constraints. For example, it is contemplated that the present invention may be implemented by using, individually or in combination, a variety of different types of optimization approaches. These optimization approaches include, but not limited to, linear programming, quadratic programming, mixed integer non-linear programming (NLP), stochastic programming, global non-linear programming, genetic algorithms, and particle/swarm techniques. See R. Baldick, Applied Optimization: Formulation and Algorithms for Engineering Systems, Cambridge University Press, Cambridge, UK, 2009.
Given the cost function and constraints, a non-linear program (NLP) optimizer typically solves problems with 20 manipulated variables and 10 controlled variables in less than one second. This is sufficiently fast for most applications since the optimization cycle is typically in the range of 10-60 seconds. Additional details on the formulation of the cost function and constraints are provided in the above-mentioned reference S. Piche, B. Sayyar-Rodsari, D. Johnson and M. Gerules, “Nonlinear model predictive control using neural networks,” IEEE Control Systems Magazine, vol. 20, no. 2, pp. 53-62, 2000, which is fully incorporated herein by reference.
The optimizer computes the full trajectory of manipulated variable moves over the future time horizon, typically 30 minutes. For an optimization system that executes every 60 seconds, 30 values are computed over a 30 minute future time horizon for each manipulated variable. Since the model or goals/constraints may change before the next optimization cycle, only the first value in the time horizon for each manipulated variable is output by the optimization system to the DCS as a setpoint value for each respective manipulated variable.
At the next optimization cycle, typically 60 seconds later, the model is updated based upon the current conditions of the plant. The cost function and constraints are also updated if they have changed. Typically, the cost function and constraints are not changed. The optimizer is used to recompute the set of values for the manipulated variables over the time horizon and the first value in the time horizon, for each manipulated variable, is output to the DCS as the setpoint value for each respective manipulated variable. The optimization system repeats this process for each optimization cycle (e.g., every 60 seconds), thus, constantly maintaining optimal performance as the boiler is affected by changes in such items as load, ambient conditions, boiler conditions, and fuel characteristics.
An alternative approach to solving the optimization problem described above is to use a rules-based approach which is reliant upon using an “expert system” to determine the optimal setpoints for the MVs to achieve the desired goals of the system. It is contemplated that the optimization system of the present invention may include the use of a rules-based system.
The load demand to a power generating plant provides a real-time desired mega-watt (MW) generation demand for the plant. However, it does not contain any information about the potential future demand for the plant.
It is advantageous to know the future load profile in order to better allocate resources in the power generating plant. Given that a future load profile is not available, it is necessary to estimate such a profile. With respect to one embodiment of the present invention, one future load profile that is of interest is the maximum possible load for the plant at some time in the future. In accordance with this embodiment of the present invention, it is desired to know the maximum value of the load for the plant at a predetermined time in the future, e.g., 30 minutes in the future. Furthermore, it is an objective in this embodiment to guarantee that the prediction of maximum load in 30 minutes is greater than the actual load 30 minutes later at least 97.5% of the time. The prediction of the maximum load at 30 minutes in the future is used as part of the optimization process to determine the fuel flows to the gas turbines and duct burners, as will be described in detail below.
Empirical data is available for load at power generating plants for the past several of years. This data can be used to train a neural network to predict the future load profile. Typically, the following inputs can be used in the prediction: calendar information, grid frequency, time of day, ambient conditions, as well as current and past loads.
Data for the inputs to model 330 is collected from current time (t) and used to predict the load at time (t+1) and the standard deviation of the predicted load at time (t+1). In the illustrated embodiment of load prediction model 330, the model inputs include: Day Of Week, which is represented by the integers 1-7; Month Of The Year, represented by an integer from 1-12, Hour Of Day, represented by an integer from 1-24; current frequency of the grid (Grid Frequency) in units of Hertz; current load in mega-watts (MW) of the plant (Load Current); and delayed versions of the load from the previous 5 minutes (Load 5 Minutes Ago), 15 minutes (Load 15 Minutes Ago), 30 minutes (Load 30 Minutes Ago), 1 hour (Load 1 Hour Ago), 2 hours (Load 2 Hours Ago), 6 hours (Load 6 Hours Ago), 12 hours (Load 12 Hours Ago), 24 hours (Load 24 Hours Ago), and 48 hours (Load 48 Hours Ago). The above-mentioned Day of Week, Month of Year and Hour of Day information is collectively referred to as “calendar data.”
It should be appreciated that the above-identified inputs for model 330 are for the purpose of illustrating an embodiment of the present invention, and not for limiting same. It is contemplated that alternative inputs may be used in load prediction model 330 in connection with the present invention.
In accordance with one embodiment of the present invention, load prediction model 330 is a neural network model that is trained using historical data from the plant. Model 330 is trained using data collected over the past year at 30 minute sampling periods of the inputs and output of interest (plant load). Data associated with plant downtime when the plant was not producing a load is not included in the training data. Neural network training algorithms, such as backpropagation (C. Bishop, Pattern Recognition and Machine Learning, Springer, New York, N.Y., 2006), are used to train model 330 to predict the load at time (t+1) (i.e., 30 minutes in advance) using data at time (t).
Once model 330 has been trained, the Standard Deviation of Predicted Load can be determined using the standard deviation of the error between the predicted and actual data over the training data. Alternatively, Bayesian techniques as describe in C. Bishop, Pattern Recognition and Machine Learning, Springer, New York, N.Y., 2006, can be used to determine both Predicted Load and Standard Deviation Of Predicted Load based upon the values of the inputs to model 330. In this approach, the standard deviation varies as a function of the inputs to model 330.
Predicted Load (t+1) and Standard Deviation Of Predicted Load (t+1) are used to predict the maximum load in the future at time (t+1) (i.e., 30 minutes in the future). Referring now to
Load allocation within a power generating plant can be achieved using optimization system 100 shown in
In one embodiment of the present invention, the plant model used for optimization is derived by developing component models for the gas turbines GT1-GT4, as well component models for the combination of the steam turbines ST1, ST2 and associated boilers B1-B4. These component models are then connected together to form models of each power generation block (i.e., BLOCK1 and BLOCK2), and then the block models are connected together to provide a model of the power generating plant. Building of the plant model will now be described in detail with reference to
It will be appreciated that various approaches described above can be used to create gas turbine power model 350. In one embodiment of the present invention, gas turbine power model 350 is created by training a neural network with historical data. In this case, 3 months of data of the inputs and outputs is collected at 15 minute samples. Samples that are not at steady state (i.e., inputs held approximately constant over the past 15 minutes) and samples where the gas turbine is off are removed from the data set. The remaining data is used to train the neural network model and provides an accurate prediction of the gas turbine power at time (t), given the GT fuel flow and ambient conditions.
Referring now to
Similar to gas turbine power model 350 described above, historical data over the past 3 months is used to train a neural network model of the steam turbine power. Once again, steady state, non-zero load, 15 minute samples of the inputs and outputs are used to train steam turbine power model 360. Once gas turbine power model 350 and steam turbine power model 360 have been trained, they may be combined to provide a block power model and a plant power model, as will be described in detail below.
Referring now to
The inputs to each gas turbine power model 350A, 350B are GT Fuel Flow and ambient conditions (i.e., Temperature, Pressure and Relative Humidity). The inputs to ST1 power model 360A are the predicted gas turbine power from both gas turbines (as determined by models 350A and 350B), the duct burner fuel flows associated with boilers B1 and B2, and the ambient conditions. It is important to note that the prediction of the gas turbine power is used as an input to the steam turbine model instead of the actual gas turbine power. The effects of changes in the GT fuel flow on steam turbine power can be predicted by chaining together GT power models 350A, 350B and steam power model 360A, as shown in
It should be appreciated that for plant 170 shown in
Referring now to
Again for simplification, plant power model 400 of
Optimization system 100 described in detail below not only uses the power plant model 400 of
Referring now to
Again for simplicity, plant power models 400 shown in
Optimizer 110 of optimization system 100 can be used to manipulate the fuel flows in the future (i.e., at time t+1) in order to produce changes to the predicted plant power and predicted duct burner power in the future (i.e., at time t+1). Once optimization system 100 determines a solution, setpoints can be sent to plant 170 for the fuel flows, and the plant power and duct burner power should move approximately to the values predicted by model 430.
Given steady state plant optimization model 430 shown in
The “goal” of the optimization system according to one embodiment of the present invention is to minimize the fuel usage while maintaining the required demand for power from the plant. It should be noted that the power plant receives a demand signal from a regional system provider and is required to meet this demand or otherwise pay a penalty. In addition, the gas turbines are used for fast ramping of the plant, while the duct burners are used to provide a minimum amount of power that is greater than or equal to the predicted maximum load at a predetermined time in the future (e.g., 30 minutes) minus predicted maximum plant power with no duct burners in use (if this value is greater than 0). By using the duct burners to provide the power associated with the difference between the predicted maximum load at the predetermined time in the future (e.g., 30 minutes) and the maximum plant power with no duct burners in use (again, assuming this is greater than zero), the power plant has the ability to deliver power to the electric grid at a fast ramp rate allowed by gas turbines (but not by duct burners). Since electric power producers get paid a premium for being able to quickly ramp their power generating plant, it is more profitable to be able to deliver power at a fast ramp rate even if it means using the more inefficient duct burners to provide that power. The goals and constraints of the optimizer of the optimization system will now be described in detail.
According to one embodiment of the present invention, the goal of the optimizer is to minimize the following cost function, J, given as:
Min(J(GT Fuel Flow,DB Fuel Flow)) (1)
where
J=Σ
i=1
4GT Fuel Flowi(t+1)+Σi=14DB Fuel Flowi(t+1) (2)
subject to the constraints:
Corrected Predicted Plant Power(t+1)=Demand(t) (3)
Predicted Duct Burner Power(t+1)≧Max(0,Predicted Max Load(t+1)−Predicted Max Power with no Duct Burners(t)) (4)
In addition, each of the fuel flows must be maintained within the allowable range of operation between typically 0% and 100% flow. In equations 1 and 2, the cost of fuel flow to the duct burners and gas turbines is minimized subject to the constraints in equations 3 and 4. In equation 3, the corrected predicted plant power must equal demand (i.e., the demand for power) where the Corrected Predicted Plant Power (t+1) is formally defined in
The optimizer is used to minimize the fuel flows, maintain the load of the plant, and keep the duct burner power above the difference between the predicted maximum load at time t+1 and the predicted maximum plant power with no duct burners at time t, if that value is greater than 0. After the optimization run, the setpoints of the fuel flows to the gas turbines and duct burners are output to the plant. To keep the plant up-to-date with changing conditions such as demand, the optimization system may need to run an optimization cycle at a faster frequency than once every 30 minutes. Accordingly, the optimizer described above may be run more frequently or a model predictive control technique which uses dynamic models may be used to solve the optimization problem, as will now be described.
In an alternative embodiment of the present invention, a dynamic component is added to the plant model shown in
Referring now to
With reference to
As shown in
The goal of the dynamic optimization, also referred to as model predictive control, is to minimize fuel flow over the future trajectory from time t+1 to t+M, while meeting the demand and maintaining the duct burner power to be greater than the predicted maximum load at time t+M (e.g., in 30 minutes) minus the predicted maximum plant power with no duct burners over the future trajectory from time t+1 to t+M. Since the system is dynamic, it is not possible to guarantee that the duct burner power will be greater than the predicted maximum load at time t+M minus the predicted max plant power with no duct burners over the entire trajectory. In this respect, it may take some time for this constraint to be achieved. To accommodate the dynamic response, in the dynamic optimization, the hard constraint used in the steady state version (equation 4) is replaced by a soft constraint which is included in the cost function.
The goal of the dynamic optimizer is to minimize the following cost function, J, given as:
Min(J(GT Fuel Flow,DB Fuel Flow)) (5)
where
J=Σ
m=1
MΣi=14GT Fuel Flowi(t+m)+Σm=1MΣi=14DB Fuel Flowi(t+m)+Σm=1MG(Predicted Duct Burner Power(t+m)−(Max(0,Predicted Max Load(t+M)−Predicted Max Power with no Duct Burners(t))) (6)
where G(x)=0 if x>=0 and G(x)=x2 if x<0, subject to the constraints:
for all m in {1, . . . ,M},Corrected Predicted Plant Power(t+m)=Demand(t)) (7)
In addition, each of the GT and DB fuel flows over the trajectory from t+1 to t+M must be maintained within the allowable range of operation between typically 0% and 100% flow. In equation 5 and 6, the cost of fuel flow to the duct burners and gas turbines is minimized along with a penalty term associated with the Predicted Duct Burner Power over the trajectory from t+1 to t+M being less than the Predicted Max Load at time t+M (in 30 minutes) minus the Predicted Max Plant Power with no Duct Burners at current time. In equation 7, the corrected predicted plant power over the future trajectory from t+1 to t+M must equal demand. The Corrected Predicted Plant Power (t+1, . . . , t+M) is formally defined in
One of the advantages of a dynamic optimizer (model predictive controller) is that it can be executed at a high frequency. In one embodiment of the present invention, the optimizer is executed every 1 minute and the results of the optimization at time (t+1) are output to the DCS. It is important to note that only the first value in the trajectory of fuel flows is actually used for control of the plant. After 1 minute, the dynamic optimization is rerun using the most current values from the plant, and the optimal trajectories are recomputed. Since the Demand, as well as the actual load, is changing minute-by-minute, this will result in a slightly different computation of the trajectories for the fuel flows at each optimization cycle. Again, only the first value of the fuel flow trajectories is output to the DCS. This cycle continues every minute with the optimizer constantly determining new values for the fuel flows.
The above description shows a variety of different techniques that may be used to solve the load allocation problem within a power generating plant (i.e., a combined cycle power plant).
The present invention described above uses adaptive on-line learning neural networks in combination with rigorous thermodynamic modeling to employ the most efficient possible firing regime (i.e., loading across combustion turbines, HRSGs, duct burners, and steam turbines) while meeting the maximum capacity and load-following. Neural network modeling and optimization can be used to both capture knowledge in plant data to find and apply in real-time optimal firing regimes across the entire range of total plant output and relevant ambient conditions, such as temperature, humidity, and barometric pressure. Neural network modeling can also be employed to provide a short-term forecast of mega-watt (MW) demand for the plant's output, to ensure the any given firing regime is capable of meeting ancillary services commitments for ramp rate and Automatic Generation Control (AGC). Real-time rigorous thermodynamic modeling is used to inform the empirical modeling and optimization of key subsystem interactions (such as that between combustion turbine, HRSG, and steam turbine efficiencies) and dynamics, using condition-based rules that exploit this thermodynamic knowledge. Such knowledge also helps inform overall operations strategies and tuning of the loops in the underlying distributed control system (DCS).
Other modifications and alterations will occur to others upon their reading and understanding of the specification. For instance, at coal-fired power generating plants, it is not unusual for the AGC system to determine the load for each coal-fired unit of the power generating plant. Although it is common for coal-fired units to be dispatched externally, in some cases, the coal-fired units may individual be allocated within the site of the power generating plant. In such cases, it is contemplated that the present invention may also find utility in connection with a coal-fired power generating plant. It is intended that all such modifications and alterations be included insofar as they come within the scope of the invention or the equivalents thereof.
This application claims the benefit of U.S. Provisional Application No. 61/874,652, filed Sep. 6, 2013, which is hereby fully incorporated herein by reference.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/US14/54242 | 9/5/2014 | WO | 00 |
Number | Date | Country | |
---|---|---|---|
61874652 | Sep 2013 | US |