A superheated steam temperature wide load model prediction method based on a PGNN method for a coal-fired unit

Abstract

The application provides a kind of PGNN method-based wide load model prediction method for superheated steam temperature of coal-fired unit.Proposed a method for establishing prediction model based on physical guided neural network (PGNN), and based on gap measurement method, the load section distribution of multi-model is determined.The multi-model gap measurement PGNN prediction method adopts multi-model gap measurement method to reasonably divide the load section, combined with the mechanism-guided long short-term memory neural network (LSTM) of superheater, can accurately predict the wide load feedforward signal of superheated steam temperature with strong coupling and large inertia.The results show that when the unit is running in wide load, as the nonlinearity of control object gradually increases with the decrease of load, more models are needed, the average absolute deviation of the prediction results of the multi-model gap measurement PGNN method and the field is within 1℃, and the main steam temperature is optimized and controlled by using the accurate prediction model as the feedforward signal.

Description

Technical Field

[0001] This invention relates to the field of generator set operation control technology, specifically to a method for predicting superheated steam temperature over a wide load range in coal-fired power units based on the PGNN method. Background Technology

[0002] When participating in deep peak shaving of the power grid, generating units typically need to operate at a wide load range of 30-40% of their rated load, with some units capable of operating at 15%-20% of their rated load. During this process, although the superheated steam temperature control system can be put into automatic operation, its control quality is poor, requiring frequent intervention from operators and increasing their workload.

[0003] Under wide-load operating conditions, for complex systems with variable parameter nonlinearity, the reasons for this include not only the numerous, frequent, and large disturbances causing changes in superheated steam temperature, but also the large delay, large inertia, and nonlinearity of the superheated steam temperature dynamic characteristics. More importantly, while protecting the unit equipment, it is required to maximize the unit's thermal efficiency, thus the superheated steam temperature must operate within the range of ±5℃ of the rated temperature, increasing the difficulty of superheated steam temperature characteristic modeling. Currently, there is limited research on the superheated steam temperature characteristics in the 20% to 50% load range of the unit. To address the above issues, it is necessary to improve the identification speed and accuracy of the prediction model. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a wide-load model prediction method for superheated steam temperature in coal-fired power units based on the PGNN method. By employing a fusion of mechanism and long short-term memory neural networks, the nonlinear temperature subsystem of the superheated steam temperature control system in the lag zone under different load ranges is modeled and predicted. Based on this, gap metric theory is used to evaluate the nonlinearity, and the divided subsystems are corrected to finally obtain a characteristic model under wide load conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting superheated steam temperature over a wide load range in coal-fired power units based on the PGNN method includes the following steps:

[0007] 1) Determine the operating condition range to be studied based on the actual operation of the system, and divide the system into N subspaces within this range at reasonable intervals.

[0008] 2) The loss function of the prediction model is obtained by mechanistic analysis. Combined with the LSTM-DNN network architecture, PGNN modeling is performed on each subspace segment to obtain N subsystem models.

[0009] 3) Calculate the gap metric value between each subsystem. For each subsystem, record its gap metric value with other (N-1) subsystems to obtain a gap metric value matrix M with a space of N×N.

[0010] 4) Based on the gap measurement values ​​between adjacent subsystems, calculate the average of these (N-1) gap measurement values ​​to determine a reasonable gap threshold and a reasonable range D for the gap threshold.

[0011] 5) Starting from the first subsystem with the lowest load, determine whether the gap metric between it and the adjacent subsystem is within a reasonable interval D. If it is not within the interval D, continuously adjust the input range of the PGNN network until the distance between two adjacent subsystems falls within the reasonable interval D. Then, it can be considered that the distance between adjacent subsystems is approximately equal.

[0012] 6) Update the data in the matrix recording the gap measurement values, and use the newly divided subsystem space to model the load characteristics using PGNN to obtain the prediction model.

[0013] As a further improvement of the present invention, in step 1), the range of 30%-100% full load is equally divided into N subspaces as N typical working condition intervals, and the load P and coal quantity under the corresponding working conditions are obtained. air volume q w Main steam pressure p main and superheater inlet temperature T f1 , with superheater outlet temperature T f2 data.

[0014] As a further improvement of the present invention, step 2) includes the following steps:

[0015] 2-1) Dimension transformation of the serialization network output is performed using a multi-layer deep neural network (DNN) to represent the load P and coal quantity. air volume q w Main steam pressure p main and superheater inlet temperature T f1 As input, the superheater outlet temperature T f2 To produce the output, an LSTM-DNN network architecture is established.

[0016] 2-2) The expressions for the mass conservation relationship and energy conservation relationship of the water spray desuperheater are as follows:

[0017]

[0018]

[0019] The inlet and outlet working fluid flow rates are both q. m,f , mg and M m These represent the mass of superheated steam stored in the superheated tubes, the mass of flue gas stored in the superheated tubes, and the mass of the superheated tube metal, respectively; h1 and h2 are the enthalpy values ​​of the superheated steam inlet and outlet, respectively; c pm T m These are the specific heat of the superheated tube metal and its wall temperature, respectively; c pg1 c pg2 These are the specific heat capacities at constant pressure for the flue gas inlet and outlet; Q1, Q2, and Q... loss These represent the heat exchange between the superheated tubes and the flue gas, the heat exchange between the superheated tubes and the superheated steam, and the heat dissipation of the equipment, respectively.

[0020] 2-3) During model training, the differential equation needs to be discretized, and the implicit Euler method is used to differ Δh0:

[0021] Vh0=h f1 q m，f +M f h f2 ′+Q a -h f2 ×(M f +q m，f )

[0022] Among them, h f1 h f2 These are the enthalpy values ​​of the working fluid at the inlet and outlet of the superheater, Q. a Q represents the heat absorbed by the superheater pipes from the flue gas. a =Ah(T) m -T f A represents the heat transfer area, h represents the heat transfer coefficient, and T represents the heat transfer coefficient. f This refers to the average temperature of the working fluid at the inlet and outlet. The superheater outlet pressure is p2 = p1 - K. p W f K p This is the pressure loss coefficient per unit pipe. Let τ be the pressure loss coefficient. t This indicates the maximum allowable computational energy error in the lag region.

[0023] 2-4) Calculate the loss function of the PGNN prediction model for the superheater outlet temperature:

[0024]

[0025] Where γ is the weight of the loss constraints of the data model and the corresponding physical model, and its magnitude indicates the degree of influence in the final model bias loss function; The root mean square error loss function of the PGNN deep network framework; The output loss of the PGNN model represents the superheater outlet temperature.

[0026] As a further improvement of the present invention, step 3) includes the following steps:

[0027] 3-1) Calculate the gap metric values ​​of the two subsystems P1 and P2. According to the definition of gap metric, the gap metric value between two operators is equal to the absolute value of all directed gap metrics between them. To avoid the influence of random errors in the neural network, the obtained values ​​are summed. The expression is:

[0028]

[0029] 3-2) According to the relevant definition of Chebyshev distance, the gap metric between two closed systems in Hilbert space is equal to the gap metric between the graphs of these two systems, expressed as:

[0030]

[0031] 3-3) Model the subsystems divided in step 1) according to step 2), and calculate the gap metric between each subsystem. Record the calculated gap metric in an N×N matrix M of gap metric values:

[0032]

[0033] As a further improvement of the present invention, step 4) includes the following steps:

[0034] 4-1) The average value of the clearance metric between subsystems is calculated as follows:

[0035]

[0036] 4-2) with δ ave The clearance metric between standard constraint subsystems is allowed an error range of 5%, meaning the clearance metric between adjacent subsystems should be within the region D ∈ (95% δ). ave 105% δ ave If the gap metric between two adjacent systems exceeds this range, it indicates that the distance between the two systems is too great, and a new load point needs to be redefined between them to build a new model for better prediction results.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] By recording and training historical operating data of the power plant, model values ​​are obtained through recurrent neural network training on the historical data. This method can effectively extract the characteristics of main steam temperature from the historical operating data, greatly reducing the impact of nonlinearity of the controlled object on model establishment. The average absolute error of the prediction model is less than 1℃, which effectively improves the accuracy of main steam temperature characteristic prediction, provides support for precise control of unit operation under wide loads, and improves the automation level of thermal power unit operation. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the LSTM-DNN network architecture for the superheater prediction model of this invention.

[0040] Figure 2 Flowchart of the multi-model PGNN load characteristic prediction method based on gap metric Detailed Implementation

[0041] The invention will now be further described with reference to the accompanying drawings.

[0042] like Figure 1 As shown, this invention provides a method for predicting superheated steam temperature over a wide load range in coal-fired power units based on the PGNN method, specifically including the following steps:

[0043] 1) Divide the 30%-100% full load range into N equal subspaces as N typical operating condition ranges, and obtain the load P and coal quantity under the corresponding operating conditions. air volume q w Main steam pressure p main Superheater inlet temperature T f1 With superheater outlet temperature T f2 data.

[0044] 2) The loss function of the prediction model is obtained through mechanistic analysis. Combined with the LSTM-DNN network architecture, PGNN modeling is performed on each subspace segment to obtain N subsystem models. The specific steps are as follows:

[0045] 2-1) Dimension transformation of the serialization network output is performed using a multi-layer deep neural network (DNN) to represent the load P and coal quantity. air volume q w Main steam pressure p main and superheater inlet temperature T f1 As input, the superheater outlet temperature T f2 To produce the output, an LSTM-DNN network architecture is established.

[0046] 2-2) The expressions for the mass conservation relationship and energy conservation relationship of the water spray desuperheater are as follows:

[0047]

[0048]

[0049] The inlet and outlet working fluid flow rates are both q. m,f , m g and M m These represent the mass of superheated steam stored in the superheated tubes, the mass of flue gas stored in the superheated tubes, and the mass of the superheated tube metal, respectively; h1 and h2 are the enthalpy values ​​of the superheated steam inlet and outlet, respectively; c pm T m These are the specific heat of the superheated tube metal and its wall temperature, respectively; c pg1 c pg2 These are the specific heat capacities at constant pressure for the flue gas inlet and outlet; Q1, Q2, and Q... loss These represent the heat exchange between the superheated tubes and the flue gas, the heat exchange between the superheated tubes and the superheated steam, and the heat dissipation of the equipment, respectively.

[0050] 2-3) During model training, the differential equation needs to be discretized, and the implicit Euler method is used to differ Δh0:

[0051] Vh0=h f1 q m，f +M f h f2 ′+Q a -h f2 ×(M f +q m，f )

[0052] Among them, h f1 h f2 These are the enthalpy values ​​of the working fluid at the inlet and outlet of the superheater, respectively. Q1 represents the heat absorbed by the flue gas in the superheater pipes, Q1 = Ah(T m -T f A represents the heat transfer area, h represents the heat transfer coefficient, and T represents the heat transfer coefficient. f This refers to the average temperature of the working fluid at the inlet and outlet. The superheater outlet pressure is p2 = p1 - K. p W f K p This is the pressure loss coefficient per unit pipe. Let τ be the pressure loss coefficient. t (2-4) The loss function of the PGNN prediction model for calculating the superheater outlet temperature indicates the maximum allowable energy error in the hysteresis region.

[0053]

[0054] Where γ is the weight of the loss constraints of the data model and the corresponding physical model, and its magnitude indicates the degree of influence in the final model bias loss function; The root mean square error loss function of the PGNN deep network framework; The output loss of the PGNN model represents the superheater outlet temperature.

[0055] 3) Calculate the gap metric value between each subsystem. For each subsystem, record its gap metric value with other (N-1) subsystems to obtain a gap metric matrix M with an N×N space. The specific steps are as follows:

[0056] 3-1) Calculate the gap metric values ​​of the two subsystems P1 and P2. According to the definition of gap metric, the gap metric value between two operators is equal to the absolute value of all directed gap metrics between them. To avoid the influence of random errors in the neural network, the obtained values ​​are summed. The expression is:

[0057]

[0058] 3-2) According to the relevant definition of Chebyshev distance, the gap metric between two closed systems in Hilbert space is equal to the gap metric between the graphs of these two systems, expressed as:

[0059]

[0060] 3-3) Model the subsystems divided in step 1) according to step 2), and calculate the gap metric between each subsystem. Record the calculated gap metric in an N×N matrix M of gap metric values:

[0061]

[0062] 4) Based on the gap measurement values ​​between adjacent subsystems, calculate the average of these (N-1) gap measurement values ​​to determine a reasonable gap threshold and a reasonable range D for the gap threshold.

[0063] 4-1) The average value of the clearance metric between subsystems is calculated as follows:

[0064]

[0065] 4-2) with δ ave The clearance metric between standard constraint subsystems is allowed an error range of 5%, meaning the clearance metric between adjacent subsystems should be within the region D ∈ (95% δ). ave 105% δ ave If the gap metric between two adjacent systems exceeds this range, it indicates that the distance between the two systems is too great, and a new load point needs to be redefined between them to build a new model for better prediction results.

[0066] 5) Starting from the first subsystem with the lowest load, determine whether the gap metric between it and the adjacent subsystem is within a reasonable interval D. If it is not within the interval D, continuously adjust the input range of the PGNN network until the distance between two adjacent subsystems falls within the reasonable interval D. Then, it can be considered that the distance between adjacent subsystems is approximately equal.

[0067] 6) Update the data in the matrix recording the gap measurement values, and use the newly divided subsystem space to model the load characteristics using PGNN to obtain the prediction model.

[0068] The above description is merely a preferred embodiment of the present invention; other effective embodiments are also possible. Any effective improvements proposed by those skilled in the art based on the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting superheated steam temperature over a wide load range in coal-fired power units based on the PGNN method, characterized in that, Includes the following steps: 1) Determine the operating condition range to be studied based on the actual operation of the system, and divide the system into N subspaces within this range at reasonable intervals; 2) The loss function of the prediction model is obtained by mechanistic analysis. Combined with the LSTM-DNN network architecture, PGNN modeling is performed on each subspace segment to obtain N subsystem models. 3) Calculate the gap metric between each subsystem. For each subsystem, record its gap metric with other (N-1) subsystems to obtain a matrix M of gap metrics with a space of N×N. 4) Based on the gap measurement values ​​between adjacent subsystems, calculate the average of these (N-1) gap measurement values ​​to determine a reasonable gap threshold and a reasonable range D for the gap threshold; 5) Starting from the first subsystem with the lowest load, determine whether the gap metric between it and the adjacent subsystem is within the reasonable interval D. If it is not within the reasonable interval D, continuously adjust the input range of the PGNN network until the distance between two adjacent subsystems falls within the reasonable interval D, that is, consider the distance between adjacent subsystems to be approximately equal. 6) Update the data in the matrix recording the gap measurement values, and use the newly divided subsystem space to model the load characteristics using PGNN to obtain the prediction model.

2. The method for predicting superheated steam temperature across a wide load range in coal-fired power units based on the PGNN method according to claim 1, characterized in that, In step 1), the range of 30%-100% full load is divided into N equal subspaces as N typical operating condition intervals, and the load P and coal quantity under the corresponding operating conditions are obtained. air volume q w Main steam pressure p main and superheater inlet temperature T f1 , with superheater outlet temperature T f2 data.

3. The method for predicting superheated steam temperature across a wide load range in coal-fired power units based on the PGNN method according to claim 1, characterized in that, Step 2) includes the following steps: 2-1) Dimension transformation of the serialization network output is performed using a multi-layer deep neural network (DNN) to represent the load P and coal quantity. air volume q w Main steam pressure p main and superheater inlet temperature T f1 As input, the superheater outlet temperature T f2 For the output, an LSTM-DNN network architecture is established; 2-2) The expressions for the mass conservation relationship and energy conservation relationship of the water spray desuperheater are as follows: The inlet and outlet working fluid flow rates are both q. m,f ; m g and M m These represent the mass of superheated steam stored in the superheated tubes, the mass of flue gas stored in the superheated tubes, and the mass of the superheated tube metal, respectively; h1 and h2 are the enthalpy values ​​of the superheated steam inlet and outlet, respectively; c pm T m These are the specific heat of the superheated tube metal and its wall temperature, respectively; c pg1 c pg2 These are the specific heat capacities at constant pressure for the flue gas inlet and outlet; Q1, Q2, and Q... loss These are the heat exchange between the superheated tubes and the flue gas, the heat exchange between the superheated tubes and the superheated steam, and the heat dissipation of the equipment, respectively. 2-3) During model training, the differential equation is discretized, and the implicit Euler method is used to differ Δh0: Vh0=h f1 q m，f +M f h f2 +Q a -h f2 ×(M f +q m，f ) Among them, h f1 h f2 These are the enthalpy values ​​of the working fluid at the inlet and outlet of the superheater, Q. a Q represents the heat absorbed by the superheater pipes from the flue gas. a =Ah(T) m -T f A represents the heat transfer area, h represents the heat transfer coefficient, and T represents the heat transfer coefficient. f This refers to the average temperature of the inlet and outlet working fluids; the superheater outlet pressure is p2 = p1 - K. p W f K p It is the pressure loss coefficient per unit pipe; denoted by τ. t This indicates the maximum allowable computational energy error in the lag region. 2-4) Calculate the loss function of the PGNN prediction model for the superheater outlet temperature: Where γ is the weight of the loss constraints of the data model and the corresponding physical model, and its magnitude indicates the degree of influence in the final model bias loss function; The root mean square error loss function of the PGNN deep network framework; The output loss of the PGNN model represents the superheater outlet temperature.

4. The method for predicting superheated steam temperature across a wide load range in coal-fired power units based on the PGNN method according to claim 1, characterized in that, Step 3) includes the following steps: 3-1) Calculate the gap metric values ​​of the two subsystems P1 and P2. According to the definition of gap metric, the gap metric value between two operators is equal to the absolute value of all directed gap metrics between them. To avoid the influence of random errors in the neural network, the obtained values ​​are summed. The expression is: 3-2) According to the relevant definition of Chebyshev distance, the gap metric between two closed systems in Hilbert space is equal to the gap metric between the graphs of these two systems, expressed as: 3-3) Model the subsystems divided in step 1) according to step 2), and calculate the gap metric between each subsystem. Record the calculated gap metric in an N×N matrix M of gap metric values:

5. The method for predicting superheated steam temperature across a wide load range in coal-fired power units based on the PGNN method according to claim 1, characterized in that, Step 4) includes the following steps: 4-1) The average value of the clearance metric between subsystems is calculated as follows: 4-2) with δ ave The clearance metric between standard constraint subsystems is allowed an error range of 5%, meaning the clearance metric between adjacent subsystems should be within the region D ∈ (95% δ). ave 105% δ ave If the gap metric between two adjacent systems exceeds this range, it proves that the distance between the two systems is too far, and a new load point needs to be redefined between them to build a new model to achieve better prediction results.

Images