Composite modeling method for steam temperature of thermal power secondary reheating unit
By combining the composite modeling method of mechanism model and neural network, the problem of nonlinear and strong coupling characteristics in steam temperature control of thermal power secondary reheating unit is solved, and high-precision steam temperature modeling and control are realized, satisfying thermal power generation high performance and safety requirements.
Patent Information
- Application Number
- CN202410936198.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-05
AI Technical Summary
The steam temperature control method of traditional thermal power secondary reheating units is difficult to deal with sudden changes in working conditions, and it fails to effectively solve the nonlinear and strong coupling characteristics of steam temperature, resulting in low model accuracy and prone to functional degradation.
Using a composite modeling method combining mechanism model with neural network, the nonlinear composite model is established through the neural network architecture of ordinary differential equations, combining heat transfer and fluid mechanics prior knowledge, embedding the prior mechanism and simplifying the model through neural network compensation Residue, achieving high-precision modeling of steam temperature.
It improves the modeling accuracy of steam temperature and the interpretability of the model, can accurately predict steam temperature under different operating conditions, meets the high performance and safety needs of thermal power generation, and improves energy utilization efficiency.
Smart Images

Figure CN118897461A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of automatic control technology, and in particular to a composite modeling method for steam temperature of a thermal power secondary reheat unit. Background Art
[0002] The demand for electricity is increasing year by year. Large-scale supercritical and ultra-supercritical thermal power units have gradually taken a dominant position in the field of power production. Their control systems play a vital role in improving energy efficiency and ensuring safe production. With the upgrading of material technology, thermal power secondary reheat units with higher parameters and larger capacity have also put forward higher requirements for corresponding modeling technology and control technology.
[0003] Traditional thermal power unit control usually includes linear model predictive controller or model-free cascade PI control of water-coal ratio, semi-manual control of flue gas valve opening and other solutions. However, in industrial processes, model-free control methods are usually unable to effectively handle sudden changes in operating conditions; simplified models that do not consider system nonlinearity will experience model drift during long-term production, and then cumulative errors will cause functional degradation. Therefore, traditional control methods are difficult to meet the performance and economic requirements of production. At the same time, steam temperature has strong nonlinear and strong coupling characteristics. In order to achieve precise control of steam temperature, a high-precision model of the steam temperature object is particularly important.
[0004] For thermal power units, traditional modeling methods can be roughly divided into two categories: mechanism models and data-driven models. The mechanism modeling method uses the laws of conservation of mass and energy to analyze the power production process, and then establishes a simplified physical model based on necessary fluid mechanics and heat transfer assumptions. The model established by the mechanism modeling method has appropriate model accuracy and interpretability, and its parameters and functions have relatively clear physical meanings. However, due to the complexity of industrial processes, it is difficult for the mechanism modeling method to accurately construct the production process based on limited knowledge, and its reasoning time and computational cost are often high. On the other hand, the model established by the data-driven modeling method has good accuracy, but due to the lack of prior mechanism analysis, the model performance may degrade under abnormal conditions.
[0005] Therefore, if a full-operating nonlinear model of the steam temperature system is constructed by combining mechanism and data-driven methods to achieve high-precision modeling of the reheated steam of the thermal power secondary reheat unit, it is expected to meet the high performance, safety and reliability requirements of thermal power generation in the subsequent controller design. Summary of the invention
[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a composite modeling method for steam temperature of a thermal power secondary reheat unit in view of the deficiencies of the prior art, which can accurately infer the state parameters of the unit under different working conditions. The method of the present invention establishes a mechanism model of the unit steam temperature by analysis, and designs an excitation signal to collect data. On this basis, a nonlinear composite model combining a mechanism model and a neural network model is established through an ordinary differential equation neural network architecture, which specifically includes the following steps:
[0007] Step 1: Establish a component-level model based on prior knowledge of heat transfer and fluid mechanics, and allocate the work proportion of the ultra-high pressure cylinder, high pressure cylinder, and medium pressure cylinder based on the ideal cycle process of the thermal power secondary reheat unit;
[0008] Step 2, according to the designed pressure drop ratio and the given heat exchange efficiency, calculate the inlet and outlet total pressure, total temperature and required flue gas flow rate of the superheater and reheater respectively;
[0009] Step 3, through the mass conservation and energy conservation equations, a nonlinear model of the mechanism relationship between the three input variables of water-coal ratio, given load, and valve opening and the three output variables of main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature is established;
[0010] Step 4, simplifying the nonlinear model and expressing it with a state space equation to construct the mechanism part of the composite model;
[0011] Step 5, design the experimental signal to stimulate the input and output data set D1 of water-coal ratio, given load, valve opening, main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature, change the amplitude of the experimental signal to stimulate the data set D2, pre-process the data sets D1 and D2 and divide them into training sets and test sets;
[0012] Step 6, establish the nonlinear mapping relationship between the state steam temperature and the input through the ordinary differential equation neural network, compensate the residual part of the mechanism model and the real model, and learn the continuous system dynamics through back propagation of the adjoint sensitivity method;
[0013] Step 7: Select the number of training rounds, the number of batches for one training, and the sample duration of each batch;
[0014] Step 8, pre-training is performed on the training set based on the data set D1. In one round of training, the prediction value of the mechanism model and the prediction value of the network model are calculated respectively, and the prediction value of the mechanism model and the prediction value of the network model are linearly combined to obtain the prediction value of the composite model;
[0015] Step 9: Calculate the training error between the predicted value and the actual value, and update the network model parameters through back propagation; at the same time, calculate the test error between the predicted value and the actual value on the test set, and record the minimum test error;
[0016] Step 10, repeat the training process on the data set D2 for fine-tuning to obtain the optimal network parameters, and obtain the steam temperature composite model of the square thermal power secondary reheat unit composed of a mechanism part and a network part.
[0017] Step 1 includes: given load and operating conditions, based on the ideal cycle process of the thermal power secondary reheat unit and its temperature and entropy diagram, allocating the work of the ultra-high pressure cylinder, the work of the high pressure cylinder, and the work of the medium pressure cylinder.
[0018] Step 2 includes: calculating the inlet and outlet pressures and the main steam temperature T1, the first reheat steam temperature T2, and the second reheat steam temperature T3 according to the design pressure drop ratio of the ultra-high pressure cylinder, the design pressure drop ratio of the high pressure cylinder, and the design pressure drop ratio of the medium pressure cylinder; given the superheater heat exchange efficiency, the first reheat heat exchange efficiency, the second reheat heat exchange efficiency and the specific heat capacity of steam, the required flue gas flow rates are calculated in combination with the work done.
[0019] Step 3 includes: establishing a nonlinear model of the mechanism relationship between the water-coal ratio r, the given load Ne, the valve opening O and the main steam temperature T1, the primary reheat steam temperature T2, and the secondary reheat steam temperature T3 through the mass conservation and energy conservation equations, which is expressed as:
[0020]
[0021] Where, matrix x = [T1 T2 T3] T , matrix u r =[r Ne O] T , f m (·) is a nonlinear mapping, T represents matrix transposition, represents the derivative with respect to x.
[0022] Step 4 includes: Simplifying the nonlinear model And expressed in state space equation, the state space equation is:
[0023]
[0024] Among them, the state matrix A and the control matrix B reflect the nonlinear relationship between the input and output of the thermal power unit.
[0025] Step 5 includes:
[0026] Step 5-1, select the design point condition, give the excitation time T and the sampling time interval Δt, and select the pseudo-random signal as the input signal u r (t), input signal u r The mean value of (t) is Eu r (t) = 0, the amplitude switches between V and -V, V>0 satisfies the safety constraint, and the switching probability P at each sampling time t is:
[0027] P[u r(t) = -V] = P t ; P[u r (t) = V] = 1-P t
[0028] Among them, P t is the switching probability, 0<P t <1; record the corresponding input signal u r (t) and the output x constitute the data set D1;
[0029] Step 5-2, repeat step 5-1, change the signal amplitude and record the data set D2, which is used for pre-training and model fine-tuning respectively;
[0030] Step 5-3: Divide the data sets D1 and D2 into a training set and a test set for network training according to the ratio.
[0031] In step 5-2, the excitation signal amplitude is changed to be within ±(5-10)% of the original amplitude;
[0032] In step 5-3, the data sets D1 and D2 are divided into training sets and test sets in a ratio of 4:1, respectively.
[0033] Step 6 includes: establishing a nonlinear mapping relationship between the state reheat steam temperature and the input f nn (·) is an ordinary differential equation neural network, which is used to compensate for the residual part between the mechanism model and the real model, and learn the dynamics of the continuous system through back propagation of the adjoint sensitivity method, f nn The structural design of (·) includes the number of neural network layers l, the number of neurons in a single layer n, the activation function and the connection method.
[0034] Step 8 includes: pre-training on a training set based on data set D1. In one round of training, the prediction value sequence of the physical model is calculated for different batches. and the network model predicts the value sequence And the mechanism model prediction value sequence and the network model predicts the value sequence Perform linear combination to obtain the composite model prediction value sequence The calculation formula is:
[0035]
[0036] Among them, i represents the sequence label, x i Represents the data value of the output at time i; predicted value and It is usually solved by an ordinary differential equation solver using the Euler method or the Runge-Kutta method; w and b are the weight and bias of the linear combination, respectively, and are converted into and Linear combination; dt is the differential symbol.
[0037] Step 9 includes: according to the predicted value sequence Calculate the training error L between the predicted value and the actual value train , update the network model parameters through back propagation; the optimization goal is:
[0038]
[0039] The constraints are:
[0040]
[0041] Among them, θ is a part of the network model parameters;
[0042] On the test set, calculate the test error L between the predicted value and the actual value test , without back propagation, record the minimum test error;
[0043] Step 10 includes: repeating steps 8 to 9 on the data set D2 to fine-tune the network model parameters to obtain L test The network weight corresponding to the minimum; from the mechanism part f m (·) and the network part f nn (·) Composite model of steam temperature for thermal power secondary reheat units.
[0044] Beneficial effects:
[0045] The present invention discloses a composite modeling method for the steam temperature of a thermal power secondary reheat unit. This method aims at solving the problems that the control method in the conventional steam temperature control system of the thermal power secondary reheat unit is easily disturbed, the modeling accuracy is not high, and the degradation is easy to occur. Through the composite modeling method of mechanism analysis and neural network learning, the nonlinear characteristics are effectively compensated, and an accurate model of the steam temperature of the thermal power secondary reheat unit is established. This method improves the interpretability of the model by embedding a priori mechanism in the neural network, and at the same time, the nonlinear part of the simplified mechanism model is compensated by the neural network to improve the accuracy of the model. This is of great value for accurately predicting the thermal power steam temperature, designing the subsequent controller, and improving the energy utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.
[0047] Figure 1 The present invention is a flow chart of a composite modeling method for steam temperature of a thermal power secondary reheat unit.
[0048] Figure 2This is a schematic diagram of the process flow of a thermal power secondary reheat unit.
[0049] Figure 3 This is a graph of the model fitting experimental results.
[0050] Figure 4 This is a graph of the model testing experimental results.
[0051] Figure 5 Training loss plot for the model fitting experiments.
[0052] Figure 6 This is a graph of the model prediction experimental results. DETAILED DESCRIPTION
[0053] The present invention provides a composite modeling method for steam temperature of a thermal power secondary reheat unit. The method of the present invention establishes a mechanism model of the unit steam temperature by analysis, designs an excitation signal to collect data, and on this basis, establishes a nonlinear composite model combining the mechanism model and the neural network model through an ordinary differential equation neural network architecture. The method specifically includes the following steps:
[0054] Step 1: Establish a component-level model based on prior knowledge of heat transfer and fluid mechanics, and allocate the work proportion of the ultra-high pressure cylinder, high pressure cylinder, and medium pressure cylinder based on the ideal cycle process of the thermal power secondary reheat unit;
[0055] Step 2, according to the designed pressure drop ratio and the given heat exchange efficiency, calculate the inlet and outlet total pressure, total temperature and required flue gas flow rate of the superheater and reheater respectively;
[0056] Step 3, through the mass conservation and energy conservation equations, a nonlinear model of the mechanism relationship between the three input variables of water-coal ratio, given load, and valve opening and the three output variables of main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature is established;
[0057] Step 4, simplifying the nonlinear model and expressing it with a state space equation to construct the mechanism part of the composite model;
[0058] Step 5, design the experimental signal to stimulate the input and output data set D1 of water-coal ratio, given load, valve opening, main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature, change the amplitude of the experimental signal to stimulate the data set D2, pre-process the data sets D1 and D2 and divide them into training sets and test sets;
[0059] Step 6, establish the nonlinear mapping relationship between the state steam temperature and the input through the ordinary differential equation neural network, compensate the residual part of the mechanism model and the real model, and learn the continuous system dynamics through back propagation of the adjoint sensitivity method;
[0060] Step 7: Select the number of training rounds, the number of batches for one training, and the sample duration of each batch;
[0061] Step 8, pre-training is performed on the training set based on the data set D1. In one round of training, the prediction value of the mechanism model and the prediction value of the network model are calculated respectively, and the prediction value of the mechanism model and the prediction value of the network model are linearly combined to obtain the prediction value of the composite model;
[0062] Step 9: Calculate the training error between the predicted value and the actual value, and update the network model parameters through back propagation; at the same time, calculate the test error between the predicted value and the actual value on the test set, and record the minimum test error;
[0063] Step 10, repeat the training process on the data set D2 for fine-tuning to obtain the optimal network parameters, and obtain the steam temperature composite model of the square thermal power secondary reheat unit composed of a mechanism part and a network part.
[0064] Step 1 includes: given load and operating conditions, based on the ideal cycle process of the thermal power secondary reheat unit and its temperature and entropy diagram, allocating the work of the ultra-high pressure cylinder, the work of the high pressure cylinder, and the work of the medium pressure cylinder.
[0065] Step 2 includes: calculating the inlet and outlet pressures and the main steam temperature T1, the first reheat steam temperature T2, and the second reheat steam temperature T3 according to the design pressure drop ratio of the ultra-high pressure cylinder, the design pressure drop ratio of the high pressure cylinder, and the design pressure drop ratio of the medium pressure cylinder; given the superheater heat exchange efficiency, the first reheat heat exchange efficiency, the second reheat heat exchange efficiency and the specific heat capacity of steam, the required flue gas flow rates are calculated in combination with the work done.
[0066] Step 3 includes: establishing a nonlinear model of the mechanism relationship between the water-coal ratio r, the given load Ne, the valve opening O and the main steam temperature T1, the primary reheat steam temperature T2, and the secondary reheat steam temperature T3 through the mass conservation and energy conservation equations, which is expressed as:
[0067]
[0068] Where, matrix x = [T1 T2 T3] T , matrix u r =[r Ne O] T , f m (·) is a nonlinear mapping, T represents matrix transposition, represents the derivative with respect to x.
[0069] Step 4 includes: Simplifying the nonlinear model And expressed in state space equation, the state space equation is:
[0070]
[0071] Among them, the state matrix A and the control matrix B reflect the nonlinear relationship between the input and output of the thermal power unit.
[0072] Step 5 includes:
[0073] Step 5-1, select the design point condition, give the excitation time T and sampling time interval Δt, and select the pseudo-random signal as the input signal u r (t), input signal u r The mean value of (t) is Eu r (t) = 0, the amplitude switches between V and -V, V>0 satisfies the safety constraint, and the switching probability P at each sampling time t is:
[0074] P[u r (t) = -V] = P t ; P[u r (t) = V] = 1-P t
[0075] Among them, P t is the switching probability, 0<P t <1; record the corresponding input signal u r (t) and the output x constitute the data set D1;
[0076] Step 5-2, repeat step 5-1, change the signal amplitude and record the data set D2, which is used for pre-training and model fine-tuning respectively;
[0077] Step 5-3: Divide the data sets D1 and D2 into a training set and a test set for network training according to the ratio.
[0078] In step 5-2, the excitation signal amplitude is changed to be within ±(5-10)% of the original amplitude;
[0079] In step 5-3, the data sets D1 and D2 are divided into training sets and test sets in a ratio of 4:1, respectively.
[0080] Step 6 includes: establishing a nonlinear mapping relationship between the state reheat steam temperature and the input f nn (·) is an ordinary differential equation neural network, which is used to compensate for the residual part between the mechanism model and the real model, and learn the dynamics of the continuous system through back propagation of the adjoint sensitivity method, f nn The structural design of (·) includes the number of neural network layers l, the number of neurons in a single layer n, the activation function and the connection method.
[0081] Step 8 includes: pre-training on a training set based on data set D1. In one round of training, the prediction value sequence of the physical model is calculated for different batches. and the network model predicts the value sequence And the mechanism model prediction value sequence and the network model predicts the value sequence Perform linear combination to obtain the composite model prediction value sequence The calculation formula is:
[0082]
[0083] Among them, i represents the sequence label, x i Represents the data value of the output at time i; predicted value and It is usually solved by an ordinary differential equation solver using the Euler method or the Runge-Kutta method; w and b are the weight and bias of the linear combination, respectively, and are converted into and Linear combination; dt is the differential symbol.
[0084] Step 9 includes: according to the predicted value sequence Calculate the training error L between the predicted value and the actual value train , update the network model parameters through back propagation; the optimization goal is:
[0085]
[0086] The constraints are:
[0087]
[0088] Among them, θ is a part of the network model parameters;
[0089] On the test set, calculate the test error L between the predicted value and the actual value test , without back propagation, record the minimum test error;
[0090] Step 10 includes: repeating steps 8 to 9 on the data set D2 to fine-tune the network model parameters to obtain L test The network weight corresponding to the minimum; from the mechanism part f m (·) and the network part f nn (·) Composite model of steam temperature for thermal power secondary reheat units.
[0091] In a specific embodiment of the present invention, taking a 1000MW ultra-supercritical secondary reheat unit of a thermal power plant as an example, the process flow is as follows: Figure 2 As shown in the figure, the boiler on the flue gas side burns, and the steam on the water feed side drives the turbine to drive the generator to generate electricity. The two are coupled with each other in heat exchange, thus forming an ideal thermodynamic cycle consisting of a constant entropy process and a constant pressure process.
[0092] Since the subsystem with the secondary reheat steam temperature as output has strong nonlinearity, this embodiment takes the mechanism relationship submodel of water-coal ratio, given load, valve opening to secondary reheat steam temperature as the object.
[0093] The implementation process of this method is as follows Figure 1 As shown, the specific process is as follows:
[0094] Step S1, building a component-level model based on prior knowledge of heat transfer and fluid mechanics, designing the relationship between parameters and variables, and simplifying them on this basis to increase the calculation speed, thereby constructing the mechanism part of the composite model;
[0095] Step S1.1, given a load of 1000MW and a normal temperature of 27°C, based on the ideal cycle process of the thermal power secondary reheat unit and its temperature-entropy diagram, the work proportions of the ultra-high pressure cylinder, high pressure cylinder, and medium pressure cylinder are allocated to be 65%, 20%, and 15% respectively;
[0096] Step S1.2, calculate the inlet and outlet pressures and temperatures according to the design pressure drop ratio of the ultra-high pressure cylinder, the design pressure drop ratio of the high pressure cylinder, and the design pressure drop ratio of the medium pressure cylinder; given the superheater heat exchange efficiency, the primary reheat heat exchange efficiency, the secondary reheat heat exchange efficiency and the steam specific heat capacity c=1.84kJ / (kg·K), calculate the required flue gas flow rates respectively in combination with the work done;
[0097] Step S1.3, through the mass conservation and energy conservation equations, establish the mechanism relationship model of the water-coal ratio r, the given load Ne, the valve opening O and the main steam temperature T1, the primary reheat steam temperature T2, and the secondary reheat steam temperature T3; its expression form is:
[0098]
[0099] Where x = [T1 T2 T3] T ,u r =[r Ne O] T , f m (·) is a nonlinear mapping;
[0100] Step S1.4: Simplify the nonlinear model And expressed by the state space equation, this embodiment is specifically:
[0101]
[0102] Where A = [-0.116156]; B = [-0.04011 0.49211-0.000037];
[0103] Step S2, designing experimental signals to obtain fully stimulated input and output data, and dividing them into a training set and a test set;
[0104] Step S2.1, select the design point condition, give a sufficiently long excitation time T = 60000s and a sampling time interval Δt = 1s, and select a pseudo-random signal as the input signal u r(t), whose amplitudes are 7, 1000MW, and 0.67, respectively, and the corresponding input and output data set D1 is recorded, with a total of 60,000 data points;
[0105] Step S2.2, repeat the steps described in step S2.1, change the signal amplitudes to 7.5, 980MW, and 0.7 respectively, and record the data set D2;
[0106] Step S2.3: Divide D1 and D2 into a training set and a test set for network training in a ratio of 4:1.
[0107] Step S3: construct the network model part of the composite model through the ordinary differential neural network framework, and then establish a composite model of high-precision mechanism and neural network.
[0108] Step S3.1: Establish a nonlinear mapping relationship between the state reheat steam temperature and the input f nn (·) is an ordinary differential equation neural network, which is used to compensate the residual part between the mechanism model and the real model, and learn the dynamics of the continuous system through back propagation of the adjoint sensitivity method. Its structural design includes the number of neural network layers l=6, the number of single-layer neurons n=16, the activation function is Leaky Relu, and the connection method is residual block connection;
[0109] Step S3.2, select the number of training rounds as 500, the number of batches for one training as 16, and the sample duration of each batch as 32 seconds;
[0110] Step S3.3: Pre-train on the training set of D1. In one round of training, the prediction value sequence of the physical model is calculated for different batches. and the network model predicts the value sequence The results of the two are linearly combined to obtain the composite model prediction value sequence The calculation method is:
[0111]
[0112] Among them, it is usually solved by the ordinary differential equation solver of Euler method or Runge-Kutta method; w and b are the weight and bias of the linear combination, which are connected through the fully connected layer. and Linear combinations;
[0113] Step S3.4: Calculate the training error L between the predicted value and the actual value based on the predicted value sequence train , and then update the network model parameters through back propagation; its optimization goal is:
[0114]
[0115] The constraints are:
[0116]
[0117] Among them, θ is a part of the network model parameters;
[0118] Step S3.5: Calculate the test error L between the predicted value and the actual value on the test set test , without back propagation, record the minimum test error;
[0119] Step S3.6: Repeat steps S3.3 to S3.5 on D2 to fine-tune the network model parameters, and then obtain L test The minimum is 12.128, and the average error is 0.025. The corresponding network weights are shown in Table 1. m (·) and the network part f nn (·) constitutes the steam temperature composite model of the thermal power secondary reheat unit, and its linear combination weight is w = [0.0014, -0.2846], and the bias is b = 0.0732; on the data set D2, Figure 3 The figure shows the model fitting test results from 20,000 to 30,000 seconds. It can be seen that the composite model can achieve better fitting performance based on the original data. The model test results are as follows: Figure 4 As shown in the figure, the test results from 40,000 to 50,000 seconds are intercepted. It can be seen that the composite model can also achieve good test performance based on the original data. At the same time, the training loss graph is as follows Figure 5 As shown, the neural network can reach convergence after about 100 rounds of training;
[0120] Table 1
[0121]
[0122] Step S3.7: Further, change the excitation signal u r The amplitude of (t) and the excitation time T are 8000 seconds. The prediction performance of the established composite model is evaluated. The model prediction experimental results are shown in Figure 6 As shown in the figure, the prediction error is 43.015, the average error is 0.081, and the determination coefficient of the prediction index is 0.9769. It can be seen that the composite model can also make good predictions under unknown inputs, indicating that the composite modeling method is effective and has good generalization ability.
[0123] The present invention provides a composite modeling method for steam temperature of a thermal power secondary reheat unit. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.
Claims
1. A composite modeling method for steam temperature of a thermal power secondary reheat unit, characterized in that: The mechanism model of the unit steam temperature is analyzed and established, and the excitation signal acquisition data is designed. On this basis, a nonlinear composite model combining the mechanism model and the neural network model is established through the ordinary differential equation neural network architecture. The specific steps include the following: Step 1: Establish a component-level model based on prior knowledge of heat transfer and fluid mechanics, and allocate the work proportion of the ultra-high pressure cylinder, high pressure cylinder, and medium pressure cylinder based on the ideal cycle process of the thermal power secondary reheat unit; Step 2, according to the designed pressure drop ratio and the given heat exchange efficiency, calculate the inlet and outlet total pressure, total temperature and required flue gas flow rate of the superheater and reheater respectively; Step 3, through the mass conservation and energy conservation equations, a nonlinear model of the mechanism relationship between the three input variables of water-coal ratio, given load, and valve opening and the three output variables of main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature is established; Step 4, simplifying the nonlinear model and expressing it with a state space equation to construct the mechanism part of the composite model; Step 5, design the experimental signal to stimulate the input and output data set D1 of water-coal ratio, given load, valve opening, main steam temperature, primary reheat steam temperature, and secondary reheat steam temperature, change the amplitude of the experimental signal to stimulate the data set D2, pre-process the data sets D1 and D2 and divide them into training sets and test sets; Step 6, establish the nonlinear mapping relationship between the state steam temperature and the input through the ordinary differential equation neural network, compensate the residual part of the mechanism model and the real model, and learn the continuous system dynamics through back propagation of the adjoint sensitivity method; Step 7: Select the number of training rounds, the number of batches for one training, and the sample duration of each batch; Step 8, pre-training is performed on the training set based on the data set D1. In one round of training, the prediction value of the mechanism model and the prediction value of the network model are calculated respectively, and the prediction value of the mechanism model and the prediction value of the network model are linearly combined to obtain the prediction value of the composite model; Step 9: Calculate the training error between the predicted value and the actual value, and update the network model parameters through back propagation; at the same time, calculate the test error between the predicted value and the actual value on the test set, and record the minimum test error; Step 10, repeat the training process on the data set D2 for fine-tuning to obtain the optimal network parameters, and obtain the steam temperature composite model of the square thermal power secondary reheat unit composed of a mechanism part and a network part.
2. The method according to claim 1, characterized in that Step 1 includes: given load and operating conditions, based on the ideal cycle process of the thermal power secondary reheat unit and its temperature and entropy diagram, allocating the work of the ultra-high pressure cylinder, the work of the high pressure cylinder, and the work of the medium pressure cylinder.
3. The method according to claim 2, characterized in that Step 2 includes: calculating the inlet and outlet pressures and the main steam temperature T1, the first reheat steam temperature T2, and the second reheat steam temperature T3 according to the design pressure drop ratio of the ultra-high pressure cylinder, the design pressure drop ratio of the high pressure cylinder, and the design pressure drop ratio of the medium pressure cylinder; given the superheater heat exchange efficiency, the first reheat heat exchange efficiency, the second reheat heat exchange efficiency and the specific heat capacity of steam, the required flue gas flow rates are calculated in combination with the work done.
4. The method according to claim 3, characterized in that Step 3 includes: establishing a nonlinear model of the mechanism relationship between the water-coal ratio r, the given load Ne, the valve opening O and the main steam temperature T1, the primary reheat steam temperature T2, and the secondary reheat steam temperature T3 through the mass conservation and energy conservation equations, which is expressed as: Where, matrix x = [T1 T2 T3] T , matrix u r =[r Ne O] T , f m (·) is a nonlinear mapping, T represents matrix transposition, represents the derivative with respect to X.
5. The method according to claim 4, characterized in that Step 4 includes: Simplifying the nonlinear model And expressed in state space equation, the state space equation is: Among them, the state matrix A and the control matrix B reflect the nonlinear relationship between the input and output of the thermal power unit.
6. The method according to claim 5, characterized in that Step 5 includes: Step 5-1, select the design point condition, give the excitation time T and sampling time interval Δt, and select the pseudo-random signal as the input signal u r (t), input signal u r The mean value of (t) is Eu r (t) = 0, the amplitude switches between V and -V, V>0 satisfies the safety constraint, and the switching probability P at each sampling time t is: P[U r (T)=-v]=P t ;P[u r (t)=V]=1-P t Among them, P t is the switching probability, 0<P t <1; record the corresponding input signal u r (t) and the output x constitute the data set D1; Step 5-2, repeat step 5-1, change the signal amplitude and record the data set D2, which is used for pre-training and model fine-tuning respectively; Step 5-3: Divide the data sets D1 and D2 into a training set and a test set for network training according to the ratio.
7. The method according to claim 6, characterized in that In step 5-2, the excitation signal amplitude is changed to be within ±(5-10)% of the original amplitude.
8. The method according to claim 7, characterized in that Step 6 includes: establishing a nonlinear mapping relationship between the state reheat steam temperature and the input f nn (·) is an ordinary differential equation neural network, which is used to compensate for the residual part between the mechanism model and the real model, and learn the dynamics of the continuous system through back propagation of the adjoint sensitivity method, f nn The structural design of (·) includes the number of neural network layers l, the number of neurons in a single layer n, the activation function and the connection method.
9. The method according to claim 8, characterized in that Step 8 includes: pre-training on a training set based on data set D1. In one round of training, the prediction value sequence of the physical model is calculated for different batches. and the network model predicts the value sequence And the mechanism model prediction value sequence and the network model predicts the value sequence Perform linear combination to obtain the composite model prediction value sequence The calculation formula is: Among them, i represents the sequence label, x i Represents the data value of the output at time i; predicted value and It is usually solved by an ordinary differential equation solver using the Euler method or the Runge-Kutta method; w and b are the weight and bias of the linear combination, respectively, and are converted into and Linear combination; dt is the differential symbol.
10. The method according to claim 9, characterized in thatStep 9 includes: according to the predicted value sequence Calculate the training error L between the predicted value and the actual value train , update the network model parameters through back propagation; the optimization goal is: The constraints are: Among them, θ is a part of the network model parameters; On the test set, calculate the test error L between the predicted value and the actual value test , without back propagation, record the minimum test error; Step 10 includes: repeating steps 8 to 9 on the data set D2 to fine-tune the network model parameters to obtain L test The network weight corresponding to the minimum; from the mechanism part f m (·) and the network part f nn (·) Composite model of steam temperature for thermal power secondary reheat units.
