Subcritical unit load real-time control method based on multivariable model predictive control

Abstract

The invention relates to a subcritical unit load real-time control method based on multivariable model predictive control. The subcritical unit load real-time control method comprises the following steps: S1, establishing a time sequence prediction model of subcritical unit power generation load and main steam flow based on an LSTM model; s2, on the basis of the data predicted in the S1 and historical operation data, working condition online identification and load section division are firstly carried out, then state-space equations of different load sections of the subcritical unit are identified in real time by using a least square method, and a multi-section linear state-space equation set is established; and S3, designing a controller based on the multi-stage state space equation set in the S2, performing optimization solution through quadratic programming to obtain the optimal control quantity at the future moment, correcting the current actual control quantity in advance, and improving the load response rate. The method can effectively improve the load response characteristic of the unit and improve the economic performance of the unit, is suitable for large-range working condition changes, and meets the peak regulation and frequency modulation requirements.

Description

Technical field:

[0001] The present invention relates to the field of computer artificial intelligence technology, and in particular to a subcritical unit load real-time control method based on multivariable model predictive control. Background technology:

[0002] Renewable energy sources such as wind and solar energy are characterized by randomness, intermittency and volatility. In recent years, the problem of wind power grid connection and absorption has become increasingly prominent. Therefore, the flexibility and dynamic energy efficiency of generator sets have become the research focus of relevant researchers.

[0003] The load control of thermal power units mainly controls the generator set from one steady state to another. Model predictive control can handle multiple variables and various constraints. It is a robust control method and is widely used in the field of power plants. In order to achieve the desired closed-loop performance, the key is to establish a reliable boiler turbine model that can accurately predict the future state evolution. Most of the existing mature model predictive control technologies use linear models, and it is very difficult to use nonlinear models for actual control. The deep peak-shaving flexible operation and changes in coal quality cause the generator set to exhibit strong time-varying nonlinearity. In addition, the large lag and large inertia it has make it difficult to characterize the relationship between input and output through an explicit linear model, which leads to slow load response of the unit and the unit flexibility cannot meet the requirements of deep peak-shaving.

[0004] It should be noted that the above contents belong to the technical knowledge of the inventor and do not necessarily constitute prior art. Summary of the invention:

[0005] The purpose of the present invention is to solve the problems existing in the prior art and provide a real-time control method for subcritical unit load based on multivariable model predictive control, which can effectively improve the load response characteristics of the unit and enhance its economic performance, adapt to a wide range of operating conditions, and meet the needs of peak load regulation and frequency regulation.

[0006] The present invention achieves the above-mentioned purpose by adopting the following technical solutions:

[0007] The method for real-time control of subcritical unit load based on multivariable model predictive control includes the following steps:

[0008] S1. Establish a time series prediction model for subcritical unit power generation load and main steam flow based on LSTM model;

[0009] S2, based on the data predicted by S1 and historical operation data, first perform online identification of operating conditions and load segment division, then use the least squares method to identify the state space equations of different load segments of the subcritical unit in real time, and establish a multi-segment linear state space equation group;

[0010] S3, design a controller based on the multi-segment state space equations of S2, optimize and solve the optimal control quantity at future moments through quadratic programming, make advance corrections to the current actual control quantity, and improve the load response rate.

[0011] In S1, the power plant operation data is first read, and outlier processing, noise removal and delay time calculation are performed on the data. The data includes state variables and control variables. The state variables include power generation load and main steam flow. The control variables include the opening of the fuel flow control valve, the opening of the feed water flow control valve and the total air volume.

[0012] x=[x1,x2] T

[0013] u=[u1,u2,u3] T

[0014] Where x is the time series of state variables, x1 is the time series of power generation load, x2 is the time series of main steam flow; u is the time series of control variables, u1 is the time series of fuel flow control valve opening, u2 is the time series of feed water flow control valve opening, and u3 is the time series of total air volume;

[0015] Secondly, build an LSTM model and train it with the data collected above. Take RMSE as the evaluation standard, take the input of the control variable at the current moment and the state variable output at the previous moment as the input of the LSTM model, and the output of the LSTM model at the next moment as the predicted data of the state variable.

[0016] In S2, the specific steps are as follows:

[0017] S21. Divide the load segments in advance according to the power generation load of the subcritical unit:

[0018] Setting load bands based on 25% of the rated maximum generation load of subcritical generating units;

[0019] Setting load bands based on 50% of the rated maximum generation load of subcritical generating units;

[0020] Setting load bands based on 75% of the rated maximum generation load of subcritical generating units;

[0021] Setting load bands based on 100% of the rated maximum generating load of subcritical generating units;

[0022] S22, based on the predicted data obtained in S1, perform online identification of the working condition and divide the data into the above-mentioned different load segments;

[0023] S23, based on the predicted data of power generation load and main steam flow obtained in S1, establish the state space equations of different load segments based on least squares identification, and find the optimal parameters by minimizing the residual square sum between the actual value and the predicted value of the prediction model:

[0024]

[0025] In the formula, Y is the observed output matrix, that is, the state variable output value x(k+1), X is the input value, that is, the state variable x(k) and the control variable u(k), θ is the system matrix to be estimated, including A and B. Through model identification, the following state space equations can be established:

[0026]

[0027] In the formula, A i , B i It is the state transfer matrix of the i-th local model obtained after linearization of different operating points.

[0028] In S3, the specific implementation steps are as follows:

[0029] S31. Set the objective function as:

[0030] MinJ=J r +J e (3)

[0031]

[0032]

[0033] In the formula, J r is the system adjustment index, J e is an economic indicator, x(k+j) is the predicted value of the state variable at time k+j, and x d (k+j) is the state variable tracking value at the time k+j, u(k+j) is the control variable at the time k+j, x1(k+j) is the power generation load at the time k+j, u1(k+j) is the fuel flow control valve opening at the time k+j, S is the terminal error weight, Q is the operation error weight, R is the input weight, E is the weight coefficient of the power generation load, and G is the weight coefficient of the fuel amount;

[0034] S32. Optimization solution of objective function:

[0035] In model predictive control, it is necessary to calculate the optimal control input in the future through the prediction model at each moment. The state space equation according to the load segment is:

[0036] x(k+1)=Ax(k)+Bu(k) (6)

[0037] In the formula, x(k) is a 2×1 vector, A is a 2×2 matrix, B is a 2×3 matrix, and u(k) is a 3×1 vector. xd The augmented form of the state space equation is obtained:

[0038]

[0039] A D The state transfer matrix of the tracking target is:

[0040]

[0041] The state space equation can be transformed into:

[0042] x a (k+1)=A a x a (k)+B a u(k) (9)

[0043] The state prediction for the future moment can be expressed as follows:

[0044]

[0045] It can be expressed as:

[0046] X a (k) = Mx a (k)+NU(k) (11)

[0047] And because:

[0048]

[0049]

[0050] The optimization index can be transformed into:

[0051]

[0052]

[0053] The cost of the initial value Extracted from the sum, the optimization objective is expanded to obtain:

[0054]

[0055]

[0056]

[0057] We can get:

[0058]

[0059] Substituting formula (11) into formula (17), we can get the final expansion of the optimization objective:

[0060]

[0061] The first term is determined by the initial state and is not affected by the control variable U(k). It can be ignored during optimization calculations and is defined as:

[0062]

[0063] The final indicator can be transformed into the standard form of quadratic programming:

[0064]

[0065] Finally, the MATLAB software toolkit is used to solve the optimization value.

[0066] The present invention adopts the above method to bring the following beneficial effects:

[0067] (1) Using LSTM neural network as a prediction model can effectively solve the problems of nonlinearity, large capacity and time lag of boiler combustion system; (2) Using multilinear model to approximate the nonlinear model can not only accurately characterize the input and output relationship but also facilitate practical application control; (3) Designing a controller for multi-segment state space equations, optimizing and solving the optimal control quantity at future moments through quadratic programming, and correcting the current actual control quantity in advance. The method of this application effectively improves the load response characteristics of the unit and improves its economic performance, adapts to a wide range of operating conditions, and meets the needs of peak load regulation and frequency regulation. Description of the drawings:

[0068] Figure 1 is an overall flow chart of the control method of the present invention;

[0069] Figure 2 This is a schematic diagram of the multivariable model predictive control principle of the present invention;

[0070] Figure 3 This is a prediction effect diagram based on the LSTM prediction model of the present invention;

[0071] Figure 4 For the present invention Figure 4 This is a load control comparison effect diagram; Specific implementation method:

[0072] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0074] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0075] like Figure 1-2 As shown, the real-time control method of subcritical unit load based on multivariable model predictive control includes the following steps:

[0076] S1. Based on the LSTM model, a time series prediction model for the power generation load and main steam flow of the subcritical unit is established. The specific process is as follows:

[0077] S11, read the power plant operation data, and perform outlier processing, noise removal and delay time calculation on the data, the data includes state variables and control variables, the state variables include power generation load and main steam flow, the control variables include fuel flow control valve opening, feed water flow control valve opening and total air volume, set:

[0078] x=[x1,x2] T

[0079] u=[u1,u2,u3] T

[0080] Where x is the time series of state variables, x1 is the time series of power generation load, x2 is the time series of main steam flow; u is the time series of control variables, u1 is the time series of fuel flow control valve opening, u2 is the time series of feed water flow control valve opening, and u3 is the time series of total air volume;

[0081] S12. Construct an LSTM model. The LSTM model solves the gradient vanishing and explosion problems of traditional RNN by introducing three gates (i.e., forget gate, input gate, and output gate), and can be effectively used to predict long time series information. In addition, LSTM can process data step by step, and can make predictions without processing the entire sequence, which is more in line with the real-time prediction needs of power plants. Therefore, this application uses a time series prediction method based on the LSTM model to establish a reliable boiler turbine model that can accurately predict future state evolution, so as to achieve the desired closed-loop performance.

[0082] Then, the LSTM model is trained with the above collected data, and RMSE is used as the evaluation standard. The input of the control variable at the current moment and the state variable output at the previous moment are used as the input of the LSTM model, and the output of the LSTM model at the next moment is the predicted data of the state variable.

[0083] S2, based on the data predicted by S1 and historical operating data, first perform online identification of operating conditions and load segment division, then use the least squares method to identify the state space equations of different load segments of the subcritical unit in real time, and establish a multi-segment linear state space equation group. The specific process is as follows:

[0084] S21. Divide the load segments in advance according to the power generation load of the subcritical unit:

[0085] Setting load bands based on 25% of the rated maximum generation load of subcritical generating units;

[0086] Setting load bands based on 50% of the rated maximum generation load of subcritical generating units;

[0087] Setting load bands based on 75% of the rated maximum generation load of subcritical generating units;

[0088] Setting load bands based on 100% of the rated maximum generating load of subcritical generating units;

[0089] S22, based on the predicted data obtained in S1, perform online identification of the working condition and divide the data into the above-mentioned different load segments;

[0090] S23, based on the predicted data of power generation load and main steam flow obtained in S1, the state space equation group of the load segment is established based on the least squares identification method, and the optimal parameters are found by minimizing the residual square sum between the actual value and the predicted value of the prediction model:

[0091]

[0092] In the formula, Y is the observed output matrix, that is, the state variable output value x(k+1), X is the input value, that is, the state variable x(k) and the control variable u(k), θ is the system matrix to be estimated, including A and B. Through model identification, the following state space equations can be established:

[0093]

[0094] In the formula, A i , B i It is the state transfer matrix of the i-th local model obtained after linearization of different operating points.

[0095] S3, based on the multi-segment state space equations of S2, the controller is designed, and the optimal control quantity at the future moment is obtained through optimization and solution through quadratic programming, and the current actual control quantity is corrected in advance to improve the load response rate. The specific process is as follows:

[0096] S31. Set the objective function as:

[0097] MinJ=J r +J e (3)

[0098]

[0099]

[0100] In the formula, J r is the system adjustment index, J e is an economic indicator, x(k+j) is the predicted value of the state variable at time k+j, and x d (k+j) is the state variable tracking value at the time k+j, u(k+j) is the control variable at the time k+j, x1(k+j) is the power generation load at the time k+j, u1(k+j) is the fuel flow control valve opening at the time k+j, S is the terminal error weight, Q is the operation error weight, R is the input weight, E is the weight coefficient of the power generation load, and G is the weight coefficient of the fuel amount;

[0101] S32. Optimization solution of objective function:

[0102] In model predictive control, it is necessary to calculate the optimal control input in the future through the prediction model at each moment. The state space equation according to the load segment is:

[0103] x(k+1)=Ax(k)+Bu(k) (6)

[0104] In the formula, x(k) is a 2×1 vector, A is a 2×2 matrix, B is a 2×3 matrix, and u(k) is a 3×1 vector. xd The augmented form of the state space equation is obtained:

[0105]

[0106] A D The state transfer matrix of the tracking target is:

[0107]

[0108] The state space equation can be transformed into:

[0109] x a (k+1)=A a xa (k)+B a u(k) (9)

[0110] The state prediction for the future moment can be expressed as follows:

[0111]

[0112] It can be expressed as:

[0113] X a (k) = Mx a (k)+NU(k) (11)

[0114] And because:

[0115]

[0116]

[0117] The optimization index can be transformed into:

[0118]

[0119]

[0120] The cost of the initial value Extracted from the sum, the optimization objective is expanded to obtain:

[0121]

[0122]

[0123]

[0124] We can get:

[0125]

[0126] Substituting formula (11) into formula (17), we can get the final expansion of the optimization objective:

[0127]

[0128] The first term is determined by the initial state and is not affected by the control variable U(k). It can be ignored during optimization calculations and is defined as:

[0129]

[0130] The final indicator can be transformed into the standard form of quadratic programming:

[0131]

[0132] Finally, the MATLAB software toolkit is used to solve the optimization value.

[0133] Conduct experimental analysis:

[0134] The data collected comes from the Huaguang subcritical power generation unit in Shanxi Province. A total of 44,000 sets of I / O data were selected for training, and the load varied between 0 and 600MW. With a sampling rate of 60s, 43,000 sets of training data represent a period of nearly one month. Another 1,000 sets of data were selected for verification, which was different from the training data selection. The root mean square error (RMSE) was used as the evaluation criterion. The following results were obtained using the method of this application:

[0135] Figure 3 This is a prediction effect diagram based on the LSTM prediction model. Figure 3 (a) is a comparison chart of the predicted value and actual value of power generation load. Figure 3 (b) is a comparison chart of the predicted value and the actual value of the main steam flow. The root mean square errors are 0.06168 and 0.08496 respectively. The prediction accuracy is high and the prediction effect is good, which overcomes the lag and nonlinearity of the boiler combustion system.

[0136] Figure 4 This is a load control comparison diagram. Figure 4 (a) is the state quantity control effect diagram of the unit load increase. Figure 4 (b) is the change diagram of the control quantity when the load increases. Figure 4 (c) is the state quantity control effect diagram of unit load reduction, Figure 4 (d) is the change diagram of the control quantity when the load decreases.

[0137] The model predictive controller (MPC) is compared with the quadratic regulator (LQR). The overall experiment is carried out in the range of 25% to 50% of the maximum rated power generation load of the unit. The terminal error coefficient, operating error coefficient and input coefficient are set as follows:

[0138]

[0139] The constraint range is: 0≤u1≤100, 0≤u2≤2000, 0≤u3≤100, the prediction domain is set to 20, the optimization domain is 250, and the prediction effect is as follows Figure 4 As shown in the figure, both controllers can achieve the control effect. When the control quantity meets the given constraints, the power generation load and main steam flow have reached the control requirements in about 50 steps, and have remained stable since then. The tracking speed and stability of the MPC controller are slightly better than those of the LQR controller, and it has obvious advantages in fuel consumption. It can quickly track the load while reducing coal consumption.

[0140] The above specific implementation manner cannot be used as a limitation on the protection scope of the present invention. For those skilled in the art, any substitution, improvement or change made to the implementation manner of the present invention falls within the protection scope of the present invention.

[0141] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

Claims

1. A real-time control method for subcritical unit load based on multivariable model predictive control, characterized in that: The following steps are involved: S1. Establish a time series prediction model for subcritical unit power generation load and main steam flow based on LSTM model; S2, based on the data predicted by S1 and historical operation data, first perform online identification of operating conditions and load segment division, then use the least squares method to identify the state space equations of different load segments of the subcritical unit in real time, and establish a multi-segment linear state space equation group; S3, design a controller based on the multi-segment state space equations of S2, optimize and solve the optimal control quantity at future moments through quadratic programming, make advance corrections to the current actual control quantity, improve the load response rate and optimize its economic performance.

2. The method for real-time control of subcritical unit load based on multivariable model predictive control according to claim 1, characterized in that: In S1, the power plant operation data is first read, and outlier processing, noise removal and delay time calculation are performed on the data. The data includes state variables and control variables. The state variables include power generation load and main steam flow. The control variables include the opening of the fuel flow control valve, the opening of the feed water flow control valve and the total air volume. Set: x=[x1,x2] T u=[u1,u2,u3] T Where x is the time series of state variables, x1 is the time series of power generation load, x2 is the time series of main steam flow; u is the time series of control variables, u1 is the time series of fuel flow control valve opening, u2 is the time series of feed water flow control valve opening, and u3 is the time series of total air volume; Secondly, build an LSTM model and train it with the data collected above. Take RMSE as the evaluation standard, take the input of the control variable at the current moment and the state variable output at the previous moment as the input of the LSTM model, and the output of the LSTM model at the next moment as the predicted data of the state variable.

3. The method for real-time control of subcritical unit load based on multivariable model predictive control according to claim 1 or 2, characterized in that: In S2, the specific steps are as follows: S21. Divide the load segments in advance according to the power generation load of the subcritical unit: Setting load bands based on 25% of the rated maximum generation load of subcritical generating units; Setting load bands based on 50% of the rated maximum generation load of subcritical generating units; Setting load bands based on 75% of the rated maximum generation load of subcritical generating units; Setting load bands based on 100% of the rated maximum generating load of subcritical generating units; S22, based on the predicted data obtained in S1, perform online identification of the working condition and divide the data into the above-mentioned different load segments; S23, based on the predicted data of power generation load and main steam flow obtained in S1, the state space equation group of the load segment is established based on the least squares identification method, and the optimal parameters are found by minimizing the residual square sum between the actual value and the predicted value of the prediction model: In the formula, Y is the observed output matrix, that is, the state variable output value x(k+1), X is the input value, that is, the state variable x(k) and the control variable u(k), θ is the system matrix to be estimated, including A and B. Through model identification, the following state space equations can be established: x(k+1)=A i x(k)+B i u(k) (2) i＝1,2…4 In the formula, A i , B i It is the state transfer matrix of the i-th local model obtained after linearization of different operating points.

4. The method for real-time control of subcritical unit load based on multivariable model predictive control according to claim 3, characterized in that: In S3, the specific implementation steps are as follows: S31. Set the objective function as: MinJ=J r +J e (3) In the formula, J r is the system adjustment index, J e is an economic indicator, x(k+j) is the predicted value of the state variable at time k+j, and x d (k+j) is the state variable tracking value at the time k+j, u(k+j) is the control variable at the time k+j, x1(k+j) is the power generation load at the time k+j, u1(k+j) is the fuel flow control valve opening at the time k+j, S is the terminal error weight, Q is the operation error weight, R is the input quantity weight, E is the weight coefficient of the power generation load, and G is the weight coefficient of the fuel quantity; S32. Optimization solution of objective function: In model predictive control, it is necessary to calculate the optimal control input in the future through the prediction model at each moment. The state space equation according to the load segment is: x(k+1)=Ax(k)+Bu(k) (6) In the formula, x(k) is a 2×1 vector, A is a 2×2 matrix, B is a 2×3 matrix, and u(k) is a 3×1 vector. xd The augmented form of the state space equation is obtained: A D The state transfer matrix of the tracking target is: The state space equation can be transformed into: x a (k+1)=A a x a (k)+B a u(k) (9) The state prediction for the future moment can be expressed as follows: It can be expressed as: X a (k)=Mx a (k)+NU(k) (11) And because: The optimization index can be transformed into: The cost of the initial value Extracted from the sum, the optimization objective is expanded to obtain: We can get: Substituting formula (11) into formula (17), we can get the final expansion of the optimization objective: The first term is determined by the initial state and is not affected by the control variable U(k). It can be ignored during optimization calculations and is defined as: The final indicator can be transformed into the standard form of quadratic programming: Finally, the MATLAB software toolkit is used to solve the optimization value.

Images