A Data-Driven Adaptive Speed ​​Control Method for Gas Turbines Based on DQN Algorithm

Abstract

This invention discloses a data-driven adaptive speed control method for gas turbines based on the DQN algorithm. Addressing the issue of traditional gas turbine control methods' strong dependence on models, this method proposes online establishment of a tight-format dynamic linearized model of the gas turbine under large-scale conditions. Furthermore, it designs a cascaded model-free adaptive controller for the gas turbine incorporating the DQN algorithm. The outer loop is a PID control loop for the power turbine speed, and the inner loop is a gas turbine speed MFAC loop with limit protection and multiple-path switching. This invention employs a scheme of cascaded inner loop control parameter tuning and outer loop command tracking based on the DQN algorithm. The state is designed as overshoot and error integral, and the action is set as four controller parameters. This invention not only achieves good tracking of commanded speed under gas turbine disturbance and performance degradation scenarios but also exhibits stronger robustness and stability to fuel flow and load.

Description

Technical Field

[0001] This invention belongs to the field of propulsion system control, specifically relating to a data-driven adaptive speed control method for gas turbines based on the DQN algorithm. Background Technology

[0002] The performance of a gas turbine's control system under varying operating conditions determines the economy and safety of its corresponding power plant. Among these, the gas turbine speed control system is a crucial component. The operating state of the load system affects the gas turbine's operating state. During gas turbine operation, the most basic goal of the control system is to stably operate at a certain speed while maintaining operational reliability to provide the required load. To achieve this control objective, traditional gas turbine control generally employs cascaded dual-loop control, and most still use conventional PI controllers. However, due to the strong nonlinearity and time-varying nature of gas turbines, the PI parameters tuned under a certain operating condition may not be applicable to all operating states of the gas turbine. Furthermore, because traditional PI control law design is highly dependent on the model, the accuracy of the gas turbine model greatly limits the improvement of control quality. System uncertainties caused by environmental changes and performance degradation will also lead to control performance falling short of expectations. Using a data-driven approach to achieve adaptive control of the gas turbine is another viable solution.

[0003] In data-driven control methods, Professor Hou Zhongsheng proposed Model-Free Adaptive Control (MFAC) in 1994. With the rapid development of computer technology, in-depth research on MFAC has been conducted in recent years. However, research on applying model-free adaptive control algorithms to gas turbine control is relatively limited, and many problems remain to be solved in gas turbine speed control. On the one hand, the single-loop control scheme of model-free adaptive control is a basic speed control scheme, widely used due to its simple structure. However, it cannot simultaneously meet control quality requirements under multiple operating conditions and suffers from slow response to internal engine disturbances. On the other hand, model-free adaptive control algorithms require tuning of multiple parameters. For a new research object under different operating conditions and in the face of environmental changes and performance degradation, repeated debugging is required before application, making the operation complex.

[0004] The cascaded dual-loop control scheme reduces the control system's dependence on the model. When faced with system uncertainties caused by environmental changes and performance degradation, it can quickly and stably operate at a certain speed. To address the problem of blind adjustment of control parameters in the inner-loop modelless adaptive controller, a control parameter tuning method based on Deep Q Network (DQN) was developed. This reduces the blindness of control parameter adjustment, accelerates online learning, thereby improving the system's speed and optimizing dynamic performance indicators. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the problem that traditional gas turbine control methods are highly dependent on models. It provides a data-driven adaptive speed control method for gas turbines based on the DQN algorithm, which reduces the dependence of the control system on models, accelerates online learning, and thus improves the system's speed. When faced with system uncertainties caused by environmental changes and performance degradation, it can quickly and stably operate at a certain speed and maintain control quality.

[0006] The present invention adopts the following technical solution, including the following steps:

[0007] Step A) Based on the input and output data of the gas turbine, a tight-format data model of the internal loop of the gas turbine controller is established online in real time using the dynamic linearization method;

[0008] Step B) Design a data-driven series hybrid control scheme based on a compact format data model. The outer loop of the controller uses PID control with the power turbine speed as the controlled signal. The high-pressure rotor speed command of the gas turbine is calculated based on the deviation between the command value and the current actual power turbine speed. The inner loop uses the high-pressure speed of the gas turbine as the controlled variable. Based on the deviation between the current gas turbine speed and the given value, MFAC control is used to adjust the fuel flow. The DQN algorithm is used to self-learn and adjust the control parameters of MFAC.

[0009] Step C) addresses the uncertain disturbances in the input and output data of the gas turbine under various operating conditions and conditions. The series hybrid control scheme and DQN algorithm from step B) are adopted to achieve adaptive control of performance degradation within the life cycle, further verifying the control performance and robustness.

[0010] As a data-driven adaptive speed control method for gas turbines based on the DQN algorithm, the specific steps of step A) are as follows:

[0011] Step A1) The step response of the nonlinear model of the gas turbine is linearized by the least squares fitting method to obtain the high pressure speed and power turbine speed curves, thereby obtaining the transfer function G2(s) of the gas turbine gas generator and the transfer function G3(s) of the load system.

[0012] Step A2): Considering the dynamic characteristics of the actuator, a tight-format data model is established online based on the gas turbine output and input data using a dynamic linearization method.

[0013] A data-driven adaptive speed control method for gas turbines based on the DQN algorithm, wherein the transfer functions of the gas generator and load system identified in step A1) are:

[0014]

[0015] s is the Laplace operator;

[0016] Furthermore, the specific steps of step A2) are as follows:

[0017] The inner loop control system of the cascade control loop is described by the following discrete-time single-input single-output nonlinear system:

[0018] y(k+1)=f(y(k),··,y(kn y ),u(k),··,u(kn u (3)

[0019] Where u(k) and y(k) are the gas turbine input and output at time k, respectively, and for gas turbine speed control, they represent the fuel flow rate W. f and the high-pressure rotor speed Ng; n y ,n u It is the unknown order of the system.

[0020] Based on the gas turbine's output and input data, a compact-format dynamic linearized data model of the gas turbine at the current operating point is established using a dynamic linearization method.

[0021] Δy(k+1)=φ(k)Δu(k) (4)

[0022] Where Δu(k) is the change in fuel flow rate, Δy is the change in the high-pressure rotor speed of the gas turbine, which is the controlled variable in the inner loop, and φ(k)∈R is the pseudo-partial derivative PPD of the system.

[0023] Based on the established dynamic linearization model, after finding the extrema of its criterion function and introducing a step size factor, the control law of the compact-format dynamic linearization model-free adaptive control is obtained as follows:

[0024]

[0025] Where λ > 0 is the penalty factor, which can limit the change of the control quantity u(k) and indirectly limit the change of the pseudo-partial derivative; ρ ∈ (0,1] is the step size factor of the control rate; y rThis represents the desired output. Further, a PPD estimate without matrix inversion is provided:

[0026]

[0027] Where η∈(0,2] is the step size factor for PPD estimation, and μ>0 is the weight factor. These two factors are adjustable during the control process. Furthermore, for application purposes, the pseudo-partial derivative is reset when the pseudo-partial derivative estimation is clearly unreasonable: if sign(φ(k))≠sign(φ(1)), or |φ(k)|≤ε, or |Δu(k-1)|≤ε, then φ(k)=φ(1). Here, ε is an arbitrarily small constant.

[0028] The present invention discloses a data-driven adaptive speed control method for gas turbines based on the DQN algorithm. The specific steps of step B) are as follows:

[0029] Step B1) assumes the transfer function of the actuator is a first-order inertial element:

[0030]

[0031] The s-domain transfer function in equations (1), (2), and (7) is used to represent the relationship between the system's input and output in the z-transform form applicable to discrete-time systems.

[0032] Step B2) Design a limit protection controller module, employing MIN-MAX switching logic. To prevent engine overheating and overpressure, control the high-pressure compressor outlet total pressure P3 and low-pressure turbine outlet total temperature T5 to remain within limit values. The limit protection controller calculates and maintains the limit parameters within the limits. PI control is used for the limit protection controller. The fuel flow rate obtained from the limit protection controller and the fuel flow rate obtained from the main loop model-free adaptive controller are compared (high / low selection), and then passed through the actuator to obtain the final fuel flow rate acting on the gas turbine.

[0033] Step B3) Design a data-driven series hybrid control scheme, with an outer loop of a PID controller and an inner loop of a compact-format model-free adaptive controller based on the design in step A) that includes over-limit protection.

[0034] Step B4) uses the DQN algorithm to self-learn and adjust the control parameters of the MFAC. The input to the outer loop PID controller is the difference between the commanded value and the actual value of the power turbine speed.

[0035] e(t)=Npr(t)-Np(t) (8)

[0036] Add e(t) to the state parameter and set the state S. t=[maxe(t),∫e(t)dt], and limit it to ensure that the gas turbine always operates within a safe range; set the action as the four parameters A of the inner loop MFAC controller. t = [eita,miu,rou,lamda], where eita, miu, rou, and lamda correspond to the step size factor η, weight factor μ, control rate step size factor ρ, and penalty factor λ of PPD estimation, respectively. A greedy strategy selects four controller parameters from the action set as output actions. By assigning a corresponding execution probability to each possible action, all possible actions are attempted. The range of the designed model-free adaptive controller parameters must satisfy the following constraints: eita∈(0,2], miu>0, rou∈(0,1], lamda>0, so that the gas turbine control system remains stable.

[0037] The reward function is set to be the sum of the two state variables:

[0038] R t =maxe(t)+∫e(t)dt (9)

[0039] The smaller the expected reward value, the better, that is, the overshoot and steady-state performance are guaranteed at this time; the gas turbine outputs the power turbine speed, calculates the difference between the speed and the command value, and then calculates the reward function. The state, action and reward are stored in the memory data playback library, the loss function required for updating is calculated, some empirical data is randomly sampled from the memory library for training, the action is selected based on the Epsilon-Greedy policy, and the output is used for MFAC control law design and CFDL dynamic linearization calculation.

[0040] The specific steps of step C1) of the gas turbine data-driven adaptive speed control method based on the DQN algorithm described in this invention are as follows:

[0041] Based on the data-driven series hybrid control scheme designed in step B), simulation was performed under rated operating conditions. The simulation required a certain period of time to stabilize before it began. Therefore, the system simulation process was first maintained at a certain fuel quantity for 10 seconds before the controller was connected to control the fuel quantity. The expected power turbine speed signal in the simulation process was to step from 95% of the rated speed to the rated speed at the 10th second. The actuator transfer function G1(s) and the load system transfer function G3(s) were subjected to two interference conditions: gain skewed by two times and phase lag not exceeding 60 degrees. The stability and tracking effectiveness of the series hybrid control scheme based on the DQN algorithm were verified.

[0042] Furthermore, the specific scheme for step C2) is as follows:

[0043] Under design point conditions, open-loop control was implemented to observe the impact of five rotating components—low-pressure compressor, high-pressure compressor, high-pressure turbine, low-pressure turbine, and power turbine—on gas turbine performance under different operating modes, including no degradation, low degradation level, and medium degradation level. The same simulation process was used for all three cases. Simulations were conducted on the data-driven series hybrid control scheme designed in step B) at three operating points: 100%, 90%, and 80% operating conditions, to verify the adaptive holding effect of the designed control loop under various operating conditions of the gas turbine and different levels of performance degradation.

[0044] The present invention has the following beneficial effects:

[0045] (1) This invention utilizes only the input and output quantities of the gas turbine, enabling the controller parameters to be independently tuned online without relying on the model. Since the response speed and control quality of the control loop depend on the parameter selection of the MFAC controller, a control parameter tuning method based on the deep Q-network algorithm has been developed to reduce the blindness of control parameter adjustment.

[0046] (2) Compared with traditional control, the present invention has stronger anti-interference and adaptive capabilities. It can track the command speed well under different operating conditions and different degrees of performance degradation. It also has stronger robust stability to fuel flow and load. Attached Figure Description

[0047] Figure 1 This is a structural diagram of the MFAC cascade control scheme;

[0048] Figure 2 This is a schematic diagram of the inner loop control system;

[0049] Figure 3 The fuel flow rate W after tuning using MFAC and DQN algorithms. f Simulation results;

[0050] Figure 4 The simulation results of the power turbine speed Np after tuning using MFAC and DQN algorithms;

[0051] Figure 5 The fuel flow W designed in this invention is the fuel flow rate W controlled by the PI control method when the actuator phase is deflected. f Simulation results;

[0052] Figure 6 The simulation results of the power turbine speed Np of the control method and PI control designed in this invention when the actuator phase is pulled;

[0053] Figure 7 The fuel flow W designed in this invention is the fuel flow rate W controlled by the PI control method when the phase of the controlled object is deflected. f Simulation results;

[0054] Figure 8 The simulation results of the power turbine speed Np of the control method and PI control designed in this invention when the phase of the controlled object is deflected;

[0055] Figure 9 The fuel flow rate W of the control method designed in this invention is when different degrees of degradation occur under operating condition 1.0. f Simulation results;

[0056] Figure 10 The simulation results of the power turbine speed Np of the control method designed in this invention under different degrees of degradation under the 1.0 operating condition are:

[0057] Figure 11 The fuel flow rate W of the control method designed in this invention is when different degrees of degradation occur under operating conditions of 0.9. f Simulation results;

[0058] Figure 12 The simulation results of the power turbine speed Np of the control method designed in this invention under different degrees of degradation at the 0.9 operating condition are:

[0059] Figure 13 The fuel flow rate W of the control method designed in this invention is measured when different degrees of degradation occur under operating conditions of 0.8. f Simulation results;

[0060] Figure 14 The simulation results of the power turbine speed Np of the control method designed in this invention under different degrees of degradation at the 0.8 operating condition are shown. Detailed Implementation

[0061] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0062] The gas turbine used in this invention has main components including a low-pressure compressor, a high-pressure compressor, a combustion chamber, a high-pressure turbine, a low-pressure turbine, a power turbine, and an exhaust system. The nonlinear mathematical model of the engine is obtained by using the component method based on the C language, and then encapsulated into a dynamic link library for digital simulation verification in the MALTAB environment.

[0063] This invention proposes a data-driven adaptive speed control method for gas turbines based on the DQN algorithm, the structure of which is shown in the figure below. Figure 1 As shown in the figure, Npr is the commanded speed of the power turbine; Np is the actual speed of the power turbine; Ng is the actual speed of the high-pressure rotor of the gas turbine; W fLet d1 be the fuel flow rate, d2 be the fuel flow rate disturbance, and d2 be the load disturbance. Different subscripts of G(s) represent the transfer functions of different components or system modules. G1(s) is the transfer function of the actuator, i.e., the transfer function from the controller-calculated command to the fuel flow rate. G2(s) is the transfer function from the fuel flow rate to the high-pressure rotor speed of the gas generator. G3(s) is the transfer function from the high-pressure rotor speed of the gas generator to the rotor speed of the power turbine. C1(s) is the inner loop MFAC controller, and C2(s) is the outer loop PID controller. The specific steps include:

[0064] Step A) Based on the input and output data of the gas turbine, a tight-format data model of the internal loop of the gas turbine controller is established online in real time using the dynamic linearization method;

[0065] Step A1) The step response of the nonlinear model of the gas turbine is linearized using the least squares fitting method to obtain the high-pressure speed and power turbine speed curves, thereby obtaining the transfer function G2(s) of the gas turbine gas generator and the transfer function G3(s) of the load system:

[0066]

[0067] s is the Laplace operator;

[0068] Step A2): Considering the dynamic characteristics of the actuator, a tight-format data model is established online based on the gas turbine output and input data using a dynamic linearization method. The inner loop control system of the cascade control loop is described by the following discrete-time single-input single-output nonlinear system:

[0069] y(k+1)=f(y(k),··,y(kn y ),u(k),··,u(kn u (3)

[0070] Where u(k) and y(k) are the gas turbine input and output at time k, respectively, and for gas turbine speed control, they represent the fuel flow rate W. f and the high-pressure rotor speed Ng; n y ,n u It is the unknown order of the system.

[0071] Based on the gas turbine's output and input data, a compact-format dynamic linearized data model of the gas turbine at the current operating point is established using a dynamic linearization method.

[0072] Δy(k+1)=φ(k)Δu(k) (4)

[0073] Where Δu(k) is the change in fuel flow rate, Δy is the change in the high-pressure rotor speed of the gas turbine, which is the controlled variable in the inner loop, and φ(k)∈R is the pseudo-partial derivative PPD of the system.

[0074] For model-free adaptive control of compact schemes in general nonlinear systems, the controlled object must satisfy the following assumptions:

[0075] Assumption 1: Except for finite time points, f(·..·) for the (n)th time point y The partial derivatives of the +2) control input variables are continuous.

[0076] Assumption 2: Except at finite time points, the generalized Lipschitz condition is satisfied, that is, for any k1≠k2, k1,k2≥0 and u(k1)≠u(k2), ||y(k1+1)-y(k2+1)||≤b||u(k1)-u(k2)||, where y(k1)≠u(k2)||. i +1)=f(y(k i ),...,y(k i -n y ),u(k i ),...,u(k i -n u )), i=1,2; b>0 is a constant, that is, a bounded change in input energy should produce a bounded change in output energy within the system.

[0077] Based on the established dynamic linearization model, after finding the extrema of its criterion function and introducing a step size factor, the control law of the compact-format dynamic linearization model-free adaptive control is obtained as follows:

[0078]

[0079] Where λ > 0 is the penalty factor, which can limit the change of the control quantity u(k) and indirectly limit the change of the pseudo-partial derivative; ρ ∈ (0,1] is the step size factor; y r This represents the desired output. Further, a PPD estimate without matrix inversion is provided:

[0080]

[0081] Where η∈(0,2] is the step size factor, and μ>0 is the weight factor, both of which are adjustable during the control process. Furthermore, for application purposes, the pseudo-partial derivative is reset when the estimation of the pseudo-partial derivative is clearly unreasonable: if sign(φ(k))≠sign(φ(1)), or |φ(k)|≤ε, or |Δu(k-1)|≤ε, then φ(k)=φ(1). Here, ε is an arbitrarily small constant.

[0082] Step B) Design a data-driven series hybrid control scheme based on a compact format data model. The outer loop of the controller uses PID control with the power turbine speed as the controlled signal. The high-pressure rotor speed command of the gas turbine is calculated based on the deviation between the command value and the current actual power turbine speed. The inner loop uses the high-pressure speed of the gas turbine as the controlled variable. Based on the deviation between the current gas turbine speed and the given value, MFAC control is used to adjust the fuel flow. The DQN algorithm is used to self-learn and adjust the control parameters of MFAC.

[0083] Step B1) assumes the transfer function of the actuator is a first-order inertial element:

[0084]

[0085] The s-domain transfer function in equations (1), (2), and (7) is used to represent the relationship between the system's input and output in the z-transform form applicable to discrete-time systems.

[0086] Step B2) aims to ensure the turbine speed continuously tracks the command value by changing the fuel quantity. A limit protection controller module is designed, employing MIN-MAX switching logic. To prevent engine overheating and overpressure, the total pressure P3 at the high-pressure compressor outlet and the total temperature T5 at the low-pressure turbine outlet are controlled to remain within limits. The limit protection controller uses PI control to maintain these parameters within the limits. The fuel flow rate obtained from the limit protection controller and the fuel flow rate obtained from the model-free adaptive controller in the main loop are compared (high / low selection), and then passed through the actuator to obtain the final fuel flow rate acting on the gas turbine.

[0087] Step B3) Design a data-driven series hybrid control scheme, with an outer loop of a PID controller and an inner loop of a compact-format model-free adaptive controller with over-limit protection designed in step A).

[0088] Step B4) uses the DQN algorithm to self-learn and adjust the control parameters of the MFAC. In the DQN algorithm, two...

[0089] Value function approximation networks with identical structures, where the target network... The parameter w in - It will be fixed in one batch training and used to generate the target Q value to calculate the TD target y. j This updates the required loss function (y). j -Q(S j A j ;w)) 2The online network Q(s,a;w) is used to evaluate the policy, and its parameters are updated in each iteration. After the online network has received a certain number of updates, the latest network weight parameters are directly used to update the target network, serving as the fixed parameters for the target network in the next round.

[0090] The specific solution steps are as follows: The agent perceives the state S of the environment. t It randomly takes actions, interacts with the environment for a period of time, and executes action A according to a greedy strategy. t And the reward R was observed t and the new state S t+1 Then, take the empirical data [S] over a period of time. t A t ,R t ,S t+1 Store the data in the experience replay library, and then randomly sample a small sample set from it. j A j ,R j ,S j+1 This is used to update the online network. The state of the next time step is input to the target network. The maximum value is selected from the Q values ​​of each action to calculate the TD target, and it is determined whether it is a terminated state. After a certain number of updates to the online network, for (y... j -Q(S j A j ;w)) 2 Update the online network parameters using gradient descent, and continue repeating the above steps, changing the current environment state to a new state, until the maximum number of iterations is reached and the result is output.

[0091] In the DQN algorithm, state S t It directly reflects the cascade control objective of the gas turbine, namely, to ensure that the power turbine speed of the gas turbine can still operate stably at a certain speed while meeting the premise of operational reliability. The input of the outer loop PID controller is the difference between the commanded value and the actual value of the power turbine speed:

[0092] e(t)=Npr(t)-Np(t) (8)

[0093] Add e(t) to the state parameter and set the state S. t =[maxe(t),∫e(t)dt], and limit it to ensure that the gas turbine always operates within a safe range; set the action as the four parameters A of the inner loop MFAC controller. t= [eita, miu, rou, lamda], which selects four controller parameters as output actions from the action set through a greedy strategy. By assigning a corresponding execution probability to each possible action, it attempts to complete all possible actions. The range of the model-free adaptive controller parameters is restricted to meet the following conditions: eita∈(0,2], miu>0, rou∈(0,1], lamda>0, so as to keep the gas turbine control system stable.

[0094] Additionally, reward R t The rationality of the design directly affects the convergence effect and control accuracy of the neural network. The reward condition and control objective must also have a corresponding relationship to effectively guide network training. This design improves the reward function, setting it as the sum of two state variables:

[0095] R t =maxe(t)+∫e(t)dt (9)

[0096] The smaller the expected reward value, the better, as this ensures both overshoot and steady-state performance. A schematic diagram of the overall structure of the inner-loop control system is shown below. Figure 2 As shown. The calculation process is as follows: based on the gas turbine input data W... f The output data Ng is obtained; the pseudo-partial derivative PPD is estimated according to the formula, and online compact-form dynamic linearization is performed to construct a compact-form dynamic linearization data model at the current operating point of the gas turbine; the fuel flow rate of the gas turbine is calculated according to the CFDL control law formula, and the control input is applied to the gas turbine. Since the four parameters involved need to be adjusted through multiple trials during the control law design and dynamic linearization process, control parameter tuning is included. The gas turbine outputs the power turbine speed, and the difference between the output and command values ​​is calculated to calculate the reward function. The state, actions, and rewards are stored in the memory data playback library. The experience library is set to store 5000 sets of data. The loss function required for updating is calculated, and a portion of the experience data is randomly sampled from the memory library for training. The batch size is set to 1000. Actions are selected based on the Epsilon-Greedy policy, and the output is used for MFAC control law design and CFDL dynamic linearization calculation.

[0097] Step C) addresses the uncertain disturbances in the input and output data of the gas turbine under various operating conditions and conditions. The series hybrid control scheme and DQN algorithm in step B) are adopted to achieve adaptive control of performance degradation within the life cycle, and to further verify the control performance and robustness.

[0098] Step C1) Based on the gas turbine's strong dependence on the model and slow response to internal disturbances, a data-driven series hybrid control scheme is designed. The robustness of the scheme under multiple operating conditions is analyzed for disturbances in gain and phase bias at different positions of the actuator and the controlled object.

[0099] Step C2) Under the design point conditions, open-loop control is used to observe the impact of rotating components on gas turbine performance under different operating modes, including no degradation, low degradation level, and medium degradation level. For the gas turbine under multiple operating conditions, the adaptive holding effect of the designed control loop is verified when the gas turbine experiences different levels of performance degradation.

[0100] The four parameters involved in the model-free adaptive controller are obtained through parameter tuning using the DQN algorithm. Besides the multiple hyperparameters required for the DQN algorithm, Table 1 shows the other training parameter settings. The initial controller parameters are set as follows: eita = 1, miu = 2, rou = 0.5, lambda = 2; after tuning: eita = 0.81, miu = 1.92, rou = 0.41, lambda = 2.11. Therefore, the parameter settings for the MFAC controller are shown in Table 2.

[0101] Table 1 Training parameter settings

[0102]

[0103] Table 2 MFAC Controller Parameter Settings

[0104]

[0105] Simulations were conducted under 100% rated operating conditions for 25 seconds with a sampling step size of 0.025s to verify the MFAC controller's ability to track the turbine rotor speed command in the cascade control. The simulation required a certain amount of time to stabilize before starting. Therefore, the system simulation was first run for 10 seconds at a specific fuel level before the controller was connected to control the fuel quantity. The desired turbine speed signal in the simulation was a step jump from 95% of the rated speed to the rated speed at the 10th second. The changes in fuel quantity and turbine speed throughout the simulation were observed. To analyze the anti-interference effect of the MFAC controller, simulations were performed comparing the cascade control processes with an inner loop PI controller and an inner loop CFDL-MFAC controller. For both fuel flow and load disturbances, the phases of the transfer functions G1(s) and G3(s) were biased. The parameter settings for the inner loop PI controller are shown in Table 3 (subscript 1 represents the inner loop PI controller parameters, subscript 2 represents the outer loop PI controller parameters). The fuel flow simulation results when the actuator phase was biased are as follows: Figure 5 Simulation results of power turbine speed are as follows: Figure 6As shown; when the phase of the controlled object is deflected, the simulation results of fuel flow are as follows. Figure 7 Simulation results of power turbine speed are as follows: Figure 8 As shown; it should be noted that the simulation environment runs on an Intel Core i7-12700 CPU, and the system results are given after normalization.

[0106] Table 3 PI Controller Parameter Settings

[0107]

[0108] Simulation results show that when the actuator phase is pulled, the response settling time of the power turbine speed is 7s, the overshoot is 0.6%, and there is no steady-state error. When the controlled object phase is pulled, the response settling time of the power turbine speed is 8.8s, the overshoot is 0.5%, and there is no steady-state error. Furthermore, compared to the traditional PI control law method, the PI controller exhibits chattering when fuel flow and load disturbances occur. The balance between system response speed and chattering depends on parameter selection, requiring detailed design of controller parameters to achieve optimal control quality. Therefore, MFAC has better anti-interference performance than the traditional PI control law method, proving that the model-free adaptive control algorithm can improve control quality and automatically compensate for system performance changes caused by the uncertainty of process object parameters and environment. This verifies the stability and tracking effectiveness of the MFAC scheme based on CFDL.

[0109] Gas turbine compressors and turbines operate continuously under harsh conditions such as high temperature, high pressure, and polluted environments, which can lead to temporary or permanent performance degradation. The causes and forms of failure vary, but they can all be characterized by changes in the flow coefficient and efficiency coefficient of these components. The thermal efficiency and output power of the gas turbine decrease, thus affecting the economic feasibility and safety of gas turbine power plants. To quantitatively represent the degree to which component characteristics deviate from their design state, a set of health parameters is defined to measure the degree of performance degradation of each engine component.

[0110] To ensure that the control system designed for the gas turbine under rated conditions can maintain good dynamic and steady-state performance after degradation, it is necessary to study an adaptive control performance maintenance method. By adding health parameters to the gas turbine component-level model, the performance changes after degradation are simulated. i = 1, 2, 3, 4, 5 represent the efficiency coefficients of the five rotating components: low-pressure compressor, high-pressure compressor, high-pressure turbine, low-pressure turbine, and power turbine: [SE1, SE2, SE3, SE4, SE5]. The specific settings of the health parameters for the three degradation scenarios are shown in Table 4.

[0111] Table 4 Gas Turbine Health Parameter Settings (%)

[0112]

[0113] Different degrees of performance degradation will affect the control of the gas turbine, preventing it from performing at its full potential. Based on the designed cascade dual-loop control scheme, the impact of components on gas turbine performance under no degradation, low degradation level, and medium degradation level is observed, thereby verifying the robustness of model-free adaptive control for gas turbine speed control under different degradation scenarios. The performance degradation of the gas turbine is simulated by skewing the efficiency parameters of five rotating components. The cascade control process of the MFAC controller is simulated at three operating points: 1.0 (output power at rated power), 0.9 (output power at 90% of rated power), and 0.8 (output power at 80% of rated power). Figures 9-14 As shown, the adaptive holding effect of the control loop designed in the face of gas turbine performance degradation is verified.

[0114] Simulation results show that, when facing performance degradation, the present invention, under the action of a model-free adaptive controller, can basically maintain the power turbine speed at the commanded value at three operating points, and can control the power turbine speed to the commanded value within 7 seconds, with no steady-state error and an overshoot within 1%. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: The overall simulation process of the present invention shows that, considering the performance degradation of the gas turbine, the MFAC control has smaller overshoot and no chattering at each operating point under disturbance conditions, exhibiting better dynamic performance and robustness than the PID controller.

Claims

1. A data-driven adaptive speed control method for gas turbines based on the DQN algorithm, characterized in that: Includes the following steps: Step A) Based on the input and output data of the gas turbine, a tight-format data model of the internal loop of the gas turbine controller is established online in real time using the dynamic linearization method; Step B) Design a data-driven series hybrid control scheme based on a compact format data model. The outer loop of the controller uses PID control with the power turbine speed as the controlled signal. The high-pressure rotor speed command of the gas turbine is calculated based on the deviation between the command value and the current actual power turbine speed. The inner loop uses the high-pressure speed of the gas turbine as the controlled variable. Based on the deviation between the current gas turbine speed and the given value, MFAC control is used to adjust the fuel flow. The DQN algorithm is used to self-learn and adjust the control parameters of MFAC. Step C) addresses the uncertain disturbances in the input and output data of the gas turbine under various operating conditions and conditions. The series hybrid control scheme and DQN algorithm from step B) are adopted to achieve adaptive control of performance degradation within the life cycle, further verifying the control performance and robustness.

2. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 1, characterized in that, The specific steps of step A) are as follows: Step A1) The step response of the nonlinear model of the gas turbine is linearized by the least squares fitting method to obtain the high pressure speed and power turbine speed curves, thereby obtaining the transfer function G2(s) of the gas turbine gas generator and the transfer function G3(s) of the load system. Step A2): Considering the dynamic characteristics of the actuator, a tight-format data model is established online based on the gas turbine output and input data using a dynamic linearization method.

3. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 2, characterized in that, The transfer functions of the gas generator and load system identified in step A1) are as follows: s is the Laplace operator.

4. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 2, characterized in that, The specific steps of step A2) include: The inner loop control system of the cascade control loop is described by the following discrete-time single-input single-output nonlinear system: y(k+1)=f(y(k),··,y(k-n y ),u(k),··,u(k-n u )) (3) Where u(k) and y(k) are the gas turbine input and output at time k, respectively, and for gas turbine speed control, they represent the fuel flow rate W. f and the high-pressure rotor speed Ng; n y ,n u It is the unknown order of the system; Based on the gas turbine's output and input data, a compact-format dynamic linearized data model of the gas turbine at the current operating point is established using a dynamic linearization method. Δy(k+1)=φ(k)Δu(k) (4) Where Δu(k) is the change in fuel flow rate, Δy is the change in the high-pressure rotor speed of the gas turbine, which is the controlled variable in the inner loop, and φ(k)∈R is the pseudo-partial derivative PPD of the system; Based on the established dynamic linearized data model, after finding the extremum of its criterion function and introducing a step size factor, the control law of the compact-format dynamic linearized model-free adaptive control is obtained as follows: Where λ > 0 is the penalty factor, which can limit the change of the control quantity u(k) and indirectly limit the change of the pseudo-partial derivative; ρ ∈ (0,1] is the step size factor of the control rate; y r The expected output is given; further, a PPD estimate without matrix inversion is provided: Where η∈(0,2] is the step size factor of PPD estimation, and μ>0 is the weight factor. These two factors are adjustable during the control process. Moreover, for application considerations, the pseudo-partial derivative is reset when the pseudo-partial derivative estimation is obviously unreasonable: if sign(φ(k))≠sign(φ(1)), or |φ(k)|≤ε, or |Δu(k-1)|≤ε, then φ(k)=φ(1), where ε is a constant.

5. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 3, characterized in that, The specific steps for designing a data-driven series hybrid control scheme based on a tight-format data model in step B) are as follows: Step 1), assume the transfer function of the actuator is a first-order inertial element: The s-domain transfer function in equations (1), (2), and (7) is used to represent the relationship between the system's input and output in the z-transform form suitable for discrete-time systems; Step 2) Design a limit protection controller module, using MIN-MAX switching logic. To prevent engine overheating and overpressure, control the total pressure P3 at the high-pressure compressor outlet and the total temperature T5 at the low-pressure turbine outlet to not exceed the limit values. The limit protection controller calculates and keeps the limit parameters within the limit values. The limit protection controller uses PI control. The fuel flow rate obtained by the limit protection controller and the fuel flow rate obtained by the model-free adaptive controller in the main loop are selected as high / low, and then passed through the actuator to obtain the final fuel flow rate acting on the gas turbine. Step 3) Design a data-driven series hybrid control scheme, with an outer loop of a PID controller and an inner loop of a compact-format model-free adaptive controller based on the design in step A) that includes over-limit protection.

6. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 4, characterized in that, The specific steps in step B) of adjusting the control parameters of the MFAC using the DQN algorithm through self-learning include: The input to the outer loop PID controller is the difference between the commanded and actual turbine speed values. e(t)=Npr(t)-Np(t) (8) Add e(t) to the state parameter and set the state S. t =[maxe(t),∫e(t)dt], and limit it to ensure that the gas turbine always operates within a safe range; set the action as the four parameters A of the inner loop MFAC controller. t = [eita,miu,rou,lamda], where eita, miu, rou, and lamda correspond to the step size factor η, weight factor μ, control rate step size factor ρ, and penalty factor λ of PPD estimation, respectively. A greedy strategy selects four controller parameters from the action set as output actions. By assigning a corresponding execution probability to each possible action, all possible actions are attempted. The range of the designed model-free adaptive controller parameters must satisfy the following constraints: eita∈(0,2], miu>0, rou∈(0,1], lamda>0, so that the gas turbine control system remains stable. The reward function is set to be the sum of the two state variables: R t =maxe(t)+∫e(t)dt (9) The smaller the expected reward value, the better, that is, the overshoot and steady-state performance are guaranteed at this time; the gas turbine outputs the power turbine speed, calculates the difference between the speed and the command value, and then calculates the reward function. The state, action and reward are stored in the memory data playback library, the loss function required for updating is calculated, some empirical data is randomly sampled from the memory library for training, the action is selected based on the Epsilon-Greedy policy, and the output is used for MFAC control law design and CFDL dynamic linearization calculation.

7. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 1, characterized in that, The specific steps of step C) are as follows: Step C1) Perform robustness analysis under multiple operating conditions on the disturbances of gain and phase bias at two different positions of the actuator and the controlled object. Step C2) Under the design point conditions, open-loop control is used to observe the impact of rotating components on gas turbine performance under different operating modes, including no degradation, low degradation level, and medium degradation level. For the gas turbine under multiple operating conditions, the adaptive holding effect of the designed control loop is verified when the gas turbine experiences different levels of performance degradation.

8. The gas turbine data-driven adaptive speed control method based on the DQN algorithm as described in claim 7, characterized in that, The specific steps of step C1) are as follows: Based on the data-driven series hybrid control scheme designed in step B), simulation was performed under rated operating conditions. The simulation required a certain period of time to stabilize before it began. Therefore, the system simulation process was first maintained at a certain fuel quantity for 10 seconds before the controller was connected to control the fuel quantity. The expected power turbine speed signal in the simulation process was to step from 95% of the rated speed to the rated speed at the 10th second. The actuator transfer function G1(s) and the load system transfer function G3(s) were subjected to two interference conditions: gain skewed by two times and phase lag not exceeding 60 degrees. The stability and tracking effectiveness of the series hybrid control scheme based on the DQN algorithm were verified.

9. The data-driven adaptive speed control method for gas turbines based on the DQN algorithm as described in claim 7, characterized in that, The specific steps of step C2) are as follows: Under design point conditions, open-loop control was implemented to observe the impact of five rotating components—low-pressure compressor, high-pressure compressor, high-pressure turbine, low-pressure turbine, and power turbine—on gas turbine performance under different operating modes, including no degradation, low degradation level, and medium degradation level. The same simulation process was used for all three cases. Simulations were conducted on the data-driven series hybrid control scheme designed in step B) at three operating points: 100%, 90%, and 80% operating conditions, to verify the adaptive holding effect of the designed control loop when facing different levels of performance degradation in a large-scale, multi-operating-condition environment.

Images