A dual-drive calculation method for dynamic efficiency of a steam turbine in a CSP power station under variable operating conditions

Abstract

The present invention relates to the field of tower solar thermal power generation and discloses a dual-drive calculation method for the variable-operating-condition dynamic efficiency of a solar thermal power station steam turbine. The method comprises the following steps: collecting real-time meteorological data from the solar thermal power station and real-time data from steam turbine equipment; utilizing the collected real-time data from the steam turbine equipment and steam turbine design parameters to reconstruct the Flügel formula, introduce a temperature correction term, and establish a new physical model for dynamic calculation of the variable-operating-condition efficiency of the steam turbine to obtain the actual operating-condition efficiency of the steam turbine; and introducing the collected real-time meteorological data and the actual operating-condition efficiency of the steam turbine calculated using the physical model into a trained neural network prediction model for the variable-operating-condition efficiency constraints of the steam turbine to obtain the constrained and corrected dynamic efficiency of the steam turbine. The method disclosed by the present invention reduces the error in the calculation of steam turbine efficiency, ensures that the turbine efficiency prediction result complies with the laws of thermodynamics, and improves the accuracy of the real-time efficiency calculation of the solar thermal power station steam turbine.

Description

Technical Field

[0001] The present invention relates to the field of tower solar thermal power generation, and in particular to a dual-drive calculation method for the variable operating condition dynamic efficiency of a steam turbine in a solar thermal power station. Background Art

[0002] Solar thermal power generation, with its unique advantages of large-scale energy storage, stable output, and peak-shaving capabilities, has become one of the key technologies supporting the safe and stable operation of new power systems. Compared to the intermittent nature of photovoltaic power generation, CSP stations can achieve continuous power generation through heat storage systems, effectively compensating for the volatility of renewable energy generation and playing an irreplaceable role in grid frequency regulation, peak shaving, and ensuring power supply reliability. However, the power generation output of CSP stations is dynamically affected by multiple factors, including meteorological conditions (such as solar irradiance and temperature) and equipment operating conditions (such as mirror field concentration efficiency, absorber heat loss, and turbine operating conditions). Therefore, accurate prediction of power generation within a predetermined timeframe is of great significance.

[0003] As the core energy conversion equipment in a CSP plant, the steam turbine's operating efficiency directly determines the plant's power generation and energy utilization, making it an essential parameter in power prediction models. In actual operation, steam turbine efficiency is not a fixed value but varies dynamically with operating parameters such as load factor, inlet steam pressure, inlet steam temperature, and exhaust pressure (back pressure). Therefore, accurately capturing the dynamic efficiency characteristics of the steam turbine under varying operating conditions is a key prerequisite for improving the accuracy of CSP power prediction.

[0004] Traditional CSP power prediction models typically set turbine efficiency to a fixed value or rely on static lookup tables for estimation. These methods completely ignore the dynamic fluctuations of operating parameters during actual operation, leading to increased errors in turbine efficiency calculations. Some studies have used purely data-driven models, such as neural networks, to predict turbine efficiency. While these models can learn nonlinear correlations in the data, they lack thermodynamic constraints, often resulting in predictions that exceed physical limits. These unrealistic results severely reduce the reliability and engineering practicality of the models.

[0005] Therefore, in view of the variable operating characteristics of the steam turbine in the CSP power station, it is necessary to develop a dynamic efficiency calculation method that combines physical mechanism constraints and data fitting capabilities to solve the problems of large static errors in traditional methods and insufficient physical rationality of pure data models. This is of great significance to improving the accuracy of power generation prediction in CSP power stations and promoting high-quality operation and maintenance of CSP stations. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a dual-drive calculation method for the dynamic efficiency of a solar thermal power station turbine under variable operating conditions, which solves the accuracy problem caused by the fixed efficiency of the steam turbine in the traditional model and significantly improves the accuracy of solar thermal power generation power prediction.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A dual-drive calculation method for dynamic efficiency of a steam turbine in a CSP power station under variable operating conditions includes the following steps:

[0009] Step 1: Collect real-time meteorological data of the CSP plant and real-time data of the steam turbine equipment;

[0010] Step 2: Using the collected real-time data of the steam turbine equipment and the turbine design parameters, the Flügel formula is reconstructed, and a temperature correction term is introduced to establish a new physical model for dynamic calculation of the turbine's variable operating efficiency, thereby obtaining the actual operating efficiency of the turbine.

[0011] In step three, the collected real-time meteorological data and the actual operating efficiency of the turbine calculated by the physical model are introduced into the trained neural network prediction model of the turbine variable operating efficiency constraint to obtain the dynamic efficiency of the turbine after constraint correction.

[0012] In the above scheme, in step 1, real-time meteorological data is obtained from the meteorological station of the CSP plant, including real-time normal solar direct irradiance, temperature, air pressure, humidity, wind speed, and wind direction, with a collection frequency of 1 minute / time; real-time data of the steam turbine equipment is obtained from the CSP plant information management system, including steam mass flow, steam inlet pressure, exhaust pressure, and main steam temperature.

[0013] In the above scheme, in step 2, a new physical model for dynamic calculation of turbine efficiency under variable operating conditions is established as follows:

[0014] ;

[0015] in, is the actual operating efficiency of the steam turbine, η0 is the design operating efficiency of the steam turbine; m is the actual operating steam mass flow rate of the steam turbine, m0 is the design operating steam mass flow rate of the steam turbine; is the flow area variation index; is the actual operating steam inlet temperature of the turbine, is the steam inlet temperature of the steam turbine under design conditions; β is the entropy increase coefficient;

[0016] The reconstructed Flügel formula is as follows:

[0017] ;

[0018] in, is the actual steam inlet pressure of the turbine under working conditions, is the actual exhaust pressure of the steam turbine. is the steam inlet pressure of the turbine under design conditions, It is the exhaust pressure of the steam turbine under design operating conditions.

[0019] In the above solution, in step 3, the neural network prediction model includes an input layer, a hidden layer and an output layer, and the hidden layer includes three layers;

[0020] The eigenvector of the input layer is:

[0021] ;

[0022] Among them, DNI is the direct solar irradiance, and Metdata is meteorological data including temperature, air pressure, humidity, wind speed and wind direction. is the actual operating steam inlet pressure, is the actual operating steam inlet temperature of the steam turbine, m is the actual operating steam mass flow rate of the steam turbine, is the actual operating efficiency of the steam turbine;

[0023] First hidden layer:

[0024] ;

[0025] ;

[0026] in, is the linear transformation output of the first hidden layer; W [1] is the weight matrix of the first hidden layer, ; b [1] is the bias vector of the first hidden layer, with dimension R 128 ;ReLU is the activation function; a [1] It is z [1] The output after ReLU activation has a dimension of R 128 ;

[0027] Second hidden layer:

[0028] ;

[0029] ;

[0030] in, is the linear transformation output of the second hidden layer; W [2] is the weight matrix of the second hidden layer, ; b [2] is the bias vector of the second hidden layer, with dimension R 64 ;a [2] It is z [2] The output after ReLU activation has a dimension of R 64 ;

[0031] Third hidden layer:

[0032] ;

[0033] ;

[0034] in, is the linear transformation output of the third hidden layer; W [3] is the weight matrix of the third hidden layer, ; b [3] is the bias vector of the third hidden layer, with dimension R 32 ;a [3] It is z [3] The output after ReLU activation has a dimension of R 32 ;

[0035] The output layer:

[0036] ;

[0037] Among them, W [4] is the weight matrix of the output layer, with dimension R 1×32 ;a [3] is the activation output vector of the third hidden layer, with a dimension of 32; b [4] is the bias vector of the output layer; is the turbine efficiency predicted by the neural network prediction model.

[0038] In a further technical solution, the turbine efficiency predicted by the neural network prediction model is subjected to constraint correction, and the turbine dynamic efficiency after constraint correction is obtained as follows:

[0039] ;

[0040] in, is Carnot efficiency;

[0041] ;

[0042] Among them, T cond is the saturation temperature corresponding to the condenser pressure, obtained by the following formula:

[0043] ;

[0044] Among them, P cond is the condenser pressure.

[0045] In the above scheme, in step 3, during the training of the neural network prediction model, the dynamic efficiency of the steam turbine after constraint correction of the neural network output is controlled by the loss function. The actual operating efficiency η of the steam turbine calculated by the physical model in step 2 t The deviation is within the set range;

[0046] Loss Function as follows:

[0047] ;

[0048] in, is the final turbine efficiency of the i-th sample after being output by the neural network and passing the Carnot efficiency constraint; is the actual operating efficiency of the steam turbine calculated by the physical model in step 2 for the i-th sample; n is the number of training samples; is the dynamic weight coefficient:

[0049] ;

[0050] Among them, βs is the adjustment intensity coefficient, and the initial value is set to 5;

[0051] ε is the threshold:

[0052] ;

[0053] in, is the actual operating steam inlet temperature of the turbine, It is the steam inlet temperature of the turbine under design operating conditions.

[0054] In a further technical solution, the loss function is minimized by gradient calculation, and the data fitting term gradient is used to drive the neural network prediction result to approach the true efficiency value; the physical constraint term gradient is used to limit the deviation between the neural network prediction result and the value calculated by the physical model in step 2;

[0055] Gradient of the data fitting term:

[0056] ;

[0057] Physical constraint gradient:

[0058] ;

[0059] in, represents the data fitting term, represents the gradient operator, is the turbine efficiency predicted by the neural network prediction model; represents the dynamic efficiency of the steam turbine after constraint correction and the actual operating efficiency η of the steam turbine t The difference, Represents a physical constraint.

[0060] Through the above technical solution, the present invention provides a dual-drive calculation method for the dynamic efficiency of a CSP steam turbine under variable operating conditions, which has the following beneficial effects:

[0061] 1. This invention pioneers a dynamic correction model for turbine efficiency under variable operating conditions. By reconstructing the Flügel equation, it establishes a real-time coupled correction mechanism for pressure ratio and temperature. Compared with traditional fixed efficiency models, this invention reduces the error in turbine efficiency calculation.

[0062] 2. The present invention introduces a neural network optimization efficiency prediction with Carnot efficiency hard constraints, embedding the Carnot efficiency limit into the loss function as a hard constraint, ensuring that the turbine efficiency prediction results conform to the laws of thermodynamics and improving the accuracy of the real-time efficiency calculation of the turbine;

[0063] 3. The present invention establishes a dual-channel coupling mechanism: the output of the physical model not only serves as a constraint, but also participates in the construction of input features, forming a "physical mechanism-data-driven" two-way feedback, which solves the error instability problem of the physical model under extreme working conditions and improves the accuracy of the dynamic efficiency prediction of the steam turbine under extreme working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0065] Figure 1 This is a flow chart of the method for calculating the dynamic efficiency of a steam turbine in a solar thermal power station under variable operating conditions disclosed in an embodiment of the present invention.

[0066] Figure 2 This is the prediction flow chart of the neural network prediction model for turbine variable operating efficiency constraints. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0068] The present invention provides a dual-drive calculation method for dynamic efficiency of a CSP steam turbine under variable operating conditions. Figure 1 As shown, the following steps are included:

[0069] Step 1: Collect real-time meteorological data of the CSP plant and real-time data of steam turbine equipment.

[0070] Real-time meteorological data is obtained from the meteorological station of the CSP plant, including real-time normal direct solar irradiance (DNI), temperature, air pressure, humidity, wind speed, and wind direction, with a collection frequency of 1 minute.

[0071] Obtain real-time data of steam turbine equipment from the CSP power station information management system, including steam mass flow, inlet steam pressure, exhaust steam pressure, main steam temperature and other parameters.

[0072] The above parameters serve as the basic input parameters for subsequent calculations.

[0073] Step 2: Using the collected real-time data of the steam turbine equipment and the design parameters of the steam turbine, the temperature correction term is introduced by reconstructing the Flügel formula, and a new physical model for dynamic calculation of the turbine variable operating efficiency is established to obtain the turbine efficiency under actual operating conditions. The traditional Flügel formula only considers the influence of the pressure ratio on the flow rate, while the present invention, through the coupling derivation of the ideal gas state equation and the isentropic flow equation, introduces the temperature deviation (the design steam inlet temperature of the steam turbine and the actual steam inlet temperature) into the flow ratio calculation for the first time, establishes a real-time pressure-temperature coupling mechanism, and solves the industry pain point of large flow calculation errors under variable temperature conditions. The specific process is as follows:

[0074] 2.1 Reconstruction of Flügel's formula

[0075] The original mass flow formula is:

[0076] ;

[0077] in:

[0078] m is the actual steam mass flow rate of the turbine under operating conditions, in t / h. The data can be obtained from the CSP plant information management system.

[0079] m0 is the steam mass flow rate at the design operating condition of the steam turbine, in t / h, which is the nameplate parameter of the steam turbine;

[0080] P in is the actual steam inlet pressure of the steam turbine, in MPa. The data can be obtained from the CSP plant information management system.

[0081] P out is the actual exhaust pressure of the steam turbine, in MPa. The data can be obtained from the CSP plant information management system.

[0082] P 0,in is the steam turbine design inlet pressure, in MPa, the data can be obtained from the steam turbine thermal design book;

[0083] P 0,out It is the design exhaust pressure of the steam turbine, in MPa. The data can be obtained from the thermal design book of the steam turbine.

[0084] 2.2 Introducing temperature correction term

[0085] In order to solve the problem that the traditional Flügel formula ignores the influence of temperature on steam mass flow, a temperature correction term is introduced. With the help of the ideal gas state equation and the isentropic flow energy equation, a new Flügel flow formula with temperature correction is derived. Specifically:

[0086] Ideal gas state equation:

[0087] ;

[0088] Where P is pressure, V is volume, n is the number of particles, T is temperature, and R is the Boltzmann constant.

[0089] Isentropic flow energy equation:

[0090] For compressible fluids (such as steam), isentropic flow satisfies:

[0091] ;

[0092] Where h is the specific enthalpy and v is the flow rate. For an ideal gas, h=C p T(C p represents the specific heat capacity that varies with steam temperature).

[0093] Outlet flow rate v out It can be approximated as:

[0094] ;

[0095] in, is the actual operating steam inlet temperature of the turbine, It is the exhaust temperature of the steam turbine under actual operating conditions.

[0096] Calculated by isentropic relationship:

[0097] ;

[0098] Where k is the specific heat ratio (usually 1.3 for steam).

[0099] Mass flow can be defined as the product of density (ρ) and velocity (v) and area (A):

[0100] ;

[0101] ;

[0102] Among them, ρ in and ρ 0,in are the inlet densities of actual and design conditions, v out and v 0,out are the outlet flow rates of actual and design conditions, respectively.

[0103] Use the ideal gas equation of state to find the density ratio:

[0104] ;

[0105] Among them, T in is the actual steam inlet temperature of the steam turbine, in °C, and the data can be obtained from the information management system of the CSP station; T 0,inIt is the design steam inlet temperature of the steam turbine, in °C, obtained from the thermal design book of the steam turbine.

[0106] 2.3 Isentropic flow energy equation to find the flow rate ratio:

[0107] According to isentropic flow, the outlet flow rate is:

[0108] ;

[0109] in, Indicates the specific heat capacity that varies with steam temperature;

[0110] Substituting the isentropic relationship, we get:

[0111] ;

[0112] Flow rate ratio:

[0113] ;

[0114] 2.4 Combining the density and flow rate ratios, we can obtain the preliminary mass flow rate ratio:

[0115] Substituting the mass flow rate ratio, we get the formula:

[0116] ;

[0117] 2.5 The pressure term associated with the original mass flow formula and substituted into the preliminary mass flow rate formula yields the following formula:

[0118] ;

[0119] 2.6 Integrating all terms, we finally get the new Flugel flow formula:

[0120] ;

[0121] 2.7 Establishing a new physical model for dynamic calculation of turbine efficiency under variable operating conditions

[0122] Establish the turbine efficiency formula under actual operating conditions based on temperature and flow changes:

[0123] ;

[0124] in: is the actual operating efficiency of the steam turbine;

[0125] η0 is the turbine design operating efficiency, which is obtained from the turbine thermal design book;

[0126] is the flow area change index (reflecting the impact of equipment wear on efficiency), which can be obtained by the following formula:

[0127] ;

[0128] in:

[0129] P in is the actual steam inlet pressure, in MPa;

[0130] P out is the actual exhaust steam pressure, unit: MPa;

[0131] β is the entropy increase coefficient:

[0132] ;

[0133] Among them, C p Indicates the specific heat capacity that changes with steam temperature. C p =2.01+0.0012×(T in -250), unit (kJ / kg·K);

[0134] η is It is the isentropic efficiency of the steam turbine under variable working conditions. It is a function of the actual working condition inlet steam temperature of the steam turbine as the independent variable and changes with the change of the inlet steam temperature. is =0.85×(1-0.003×|T in -T0|), B is the steam gas constant (usually 0.461 kJ / kg·K);

[0135] d decreases with increasing exhaust / inlet pressure ratio, reflecting the decreasing sensitivity of efficiency to flow rate changes as the flow area increases due to wear. β is positively correlated with the specific heat capacity, which varies with steam temperature, and the turbine's variable-condition isentropic efficiency. It quantifies the rate of entropy loss when the temperature deviates from the design value.

[0136] In step three, the collected meteorological data and the actual operating efficiency of the turbine calculated by the physical model are introduced into the trained neural network prediction model of the turbine variable operating efficiency constraint to obtain the dynamic efficiency of the turbine after constraint correction.

[0137] Although step two reconstructed the dynamic correction model for steam turbine efficiency, improving the accuracy of real-time turbine efficiency calculations to a certain extent, the errors in the calculations remained unstable under extreme operating conditions. Therefore, step three introduced a constrained neural prediction network, forming a closed loop: "the physical model provides the mechanism boundaries - the neural network learns data trends - and the constraint layer corrects the mechanism deviations." This not only avoids the overfitting risk of pure data models, but also compensates for the accuracy loss caused by physical model calculations. This resolves the problem of unstable errors in the physical model under extreme operating conditions and further ensures the accuracy of the turbine dynamic efficiency calculations.

[0138] Existing neural networks mostly use fixed thresholds (such as the upper limit of Carnot efficiency), while the present invention uses a correlation model between dynamic thresholds and temperature deviations to dynamically adjust the constraint strength according to the working conditions, thereby improving the prediction accuracy while ensuring physical rationality.

[0139] The actual operating efficiency η of the steam turbine calculated by the physical model in step 2 t , as the input parameter and benchmark reference value of the constraint neural network in step 3, providing boundary constraints based on thermodynamic mechanisms for the prediction results of the neural network. Specifically, the constraint neural network uses the operating data of the target solar thermal power station for one year as the training sample. During the training process, η t Physical constraints are imposed on the prediction results of the neural network as a benchmark to ensure that they are within a reasonable thermodynamic range; by embedding the Carnot efficiency limit as a hard constraint into the model, the calculation error problem of traditional pure data-driven models that deviates from physical laws under extreme working conditions is effectively avoided. The calculation process is shown in Figure 2 .

[0140] The neural network prediction model includes an input layer, a hidden layer, and an output layer, and the hidden layer includes three layers.

[0141] The feature vector of the input layer is:

[0142] ;

[0143] Among them, DNI is the direct normal solar irradiance, and Metdata is meteorological data including temperature, air pressure, humidity, wind speed and wind direction. This information can be obtained through the meteorological observation station of the solar thermal power station. is the actual steam inlet pressure, is the actual steam inlet temperature of the steam turbine, and m is the steam mass flow rate of the steam turbine under actual operating conditions, both of which are obtained from the information management system of the CSP station. It is the actual operating efficiency of the steam turbine obtained by calculating the physical model in step 2.

[0144] Hidden layer structure:

[0145] The number of hidden layer nodes was optimized by grid search. When the number of nodes was 128→64→32, the model loss function was minimized (the validation set error was 3.2%), so this structure was adopted.

[0146] First hidden layer:

[0147] ;

[0148] ;

[0149] in, is the linear transformation output of the first hidden layer, which is the result of the first linear combination of the input features; W [1] is the weight matrix of the first hidden layer, , dimension is R 128 , where "128" is the number of nodes in the first hidden layer and "6" corresponds to the dimension of the input layer feature vector. This matrix is ​​used to perform linear transformation on the input features. Each element W xy [1] It represents the influence weight of the yth feature of the input layer on the xth node of the first hidden layer, and captures the potential correlation between different input features and turbine efficiency through learning.

[0150] b [1] is the bias vector of the first hidden layer, with dimension R 128 , each element b x [1] It is the bias term of the xth node in the first hidden layer, which is used to adjust the baseline value of the linear transformation result, compensate for the impact of the mean shift of the input features on the model, and enhance the adaptability of the network to changes in feature distribution.

[0151] ReLU is the activation function, which is used to transform the linear result z [1] Nonlinear mapping is performed to simulate the nonlinear relationship between turbine efficiency and input parameters by retaining positive values ​​and suppressing negative values, solving complex correlation problems that cannot be captured by linear models.

[0152] a [1] It is z [1] The output after ReLU activation has the same dimension as the number of nodes in the first hidden layer (128 dimensions), which is used to capture the low-order nonlinear correlations in the input features related to turbine efficiency.

[0153] Second hidden layer:

[0154] ;

[0155] ;

[0156] in, is the linear transformation output of the second hidden layer, and is the low-order nonlinear feature a of the output of the first hidden layer [1] The result of the quadratic linear combination; W [2] is the weight matrix of the second hidden layer, ; dimension is R 64 ×128, "64" represents the number of nodes in the second hidden layer, and "128" corresponds to the feature dimension of the first hidden layer output. This matrix is ​​used to perform a quadratic linear transformation on the 128-dimensional intermediate features output by the first hidden layer. By distributing weights across nodes, it fuses the synergistic effects of different low-level features to generate more abstract mid-level features.

[0157] b [2] is the bias vector of the second hidden layer, with dimension R 64 , same function as b [1] , which is used to adjust the linear transformation benchmark of the second hidden layer to ensure that the distribution of mid-order features meets the requirements of subsequent learning.

[0158] Activation function ReLU: further enhances nonlinear mapping capabilities, reduces the risk of model overfitting by screening effective mid-order features (suppressing redundant information), and lays the foundation for high-order feature extraction.

[0159] a [2] It is z [2] The output after ReLU activation has a dimension of 64, which is used to fuse low-order features and extract more abstract mid-order nonlinear features.

[0160] Third hidden layer:

[0161] ;

[0162] ;

[0163] in, is the linear transformation output of the third hidden layer, and is the intermediate feature a of the output of the second hidden layer [2] The result of three linear combinations; W [3] is the weight matrix of the third hidden layer, , dimension is R 32×64 "32" represents the number of nodes in the third hidden layer, and "64" corresponds to the feature dimension of the second hidden layer output. This matrix reduces and fuses mid-order features, focusing on key influencing factors (such as the dominant role of temperature deviation under extreme operating conditions), generating 32-dimensional high-order features that are directly related to the core variation patterns of turbine efficiency.

[0164] b [3] is the bias vector of the third hidden layer, with dimension R 32 ; Used to calibrate the benchmark value of high-order features to ensure that it matches the efficiency prediction target of the output layer in terms of numerical range.

[0165] Activation function ReLU: Through nonlinear screening, it retains the high-order features that are most critical for efficiency prediction and provides high-quality input for the linear prediction of the output layer.

[0166] a[3] It is z [3] The output after ReLU activation has a dimension of 32, which is used to further refine key features and provide high-order feature support for the output layer to predict turbine efficiency.

[0167] Output layer:

[0168] ;

[0169] Among them, W [4] is the weight matrix of the output layer, with dimension R 1×32 ;a [3] is the activation output vector of the third hidden layer, with a dimension of 32; b [4] is the bias vector of the output layer; is the turbine efficiency predicted by the neural network prediction model;

[0170] The turbine efficiency predicted by the neural network prediction model is constrained and corrected, and the dynamic efficiency of the turbine after constraint correction is obtained as follows:

[0171] ;

[0172] If the turbine output efficiency predicted by the neural network Exceeding the Carnot efficiency η Carnot , then the final efficiency of the turbine Forced to η Carnot If it is lower than 0.35 (the minimum reasonable efficiency limit of the steam turbine), it will be forcibly set to 0.35. The setting of 0.35 for the minimum efficiency of the steam turbine is based on three years of operating data from six domestic solar thermal power plants. The actual operating efficiency limit of the steam turbine is 0.32-0.38, and the average value is added with a 5% safety margin.

[0173] in, is Carnot efficiency;

[0174] ;

[0175] Among them, T cond is the saturation temperature corresponding to the condenser pressure, in °C, obtained from the following formula:

[0176] ;

[0177] Among them, P cond is the condenser pressure, which can be obtained from the CSP plant information management system.

[0178] During the training process of the neural network prediction model, the dynamic efficiency of the steam turbine after constraint correction of the neural network output is controlled by the loss function. The actual operating efficiency η of the steam turbine calculated by the physical model in step 2 t The deviation is within the set range; the risk of overfitting is reduced, so that the model can stably output reasonable results under different operating conditions.

[0179] Loss Function It is a dual-constraint loss function, consisting of a data fitting term and a physical constraint term, as follows:

[0180] ;

[0181] in, is the final turbine efficiency after the output of the i-th sample constraint neural network and the Carnot efficiency constraint; is the actual operating efficiency of the steam turbine calculated by the physical model in step 2 for the i-th sample; n is the number of training samples; is the dynamic weight coefficient:

[0182] ;

[0183] Among them, βs is the adjustment strength coefficient, the initial value is set to 5, the control sensitivity to changes in bias;

[0184] When the neural network predicts the result Compared with the calculated value η of the physical model t When the deviation is large, Increase to strengthen the physical constraint; when the deviation is small, Reduce to focus on data fitting, so that the neural network can adaptively balance the priorities of "data fitting" and "physical mechanism" according to the prediction deviation during training.

[0185] ε is the threshold value. Based on the deviation between the actual steam inlet temperature and the design temperature of the turbine, the tolerance of the physical constraint term of the loss function is dynamically adjusted. This ensures that the prediction results conform to the laws of thermodynamics while taking into account the prediction flexibility and accuracy under different operating conditions.

[0186] ;

[0187] in, is the actual steam inlet temperature of the steam turbine, in °C. The data can be obtained from the CSP plant information management system. It is the design steam inlet temperature of the steam turbine, in °C, obtained from the thermal design book of the steam turbine.

[0188] When |T in -T 0,in When ∣>50℃ (extreme temperature deviation condition), ε=0.07-0.02=0.05. The greater the temperature deviation, the stricter the constraint.

[0189] The loss function is minimized through gradient calculation. The data fitting term gradient is used to drive the neural network prediction results to approach the true efficiency value, ensuring the model's learning effect on historical operating data; the physical constraint term gradient is used to limit the deviation between the neural network prediction results and the calculated values ​​of the physical model in step 2, ensuring that the prediction results conform to the thermodynamic mechanism and avoiding physical irrationality caused by pure data drive.

[0190] Gradient of the data fitting term:

[0191] ;

[0192] Physical constraint gradient:

[0193] ;

[0194] in, represents the data fitting term, represents the gradient operator, is the turbine efficiency predicted by the neural network prediction model; represents the dynamic efficiency of the steam turbine after constraint correction and the actual operating efficiency η of the steam turbine t The difference, Represents a physical constraint.

[0195] Step 3: Final output of the turbine dynamic efficiency after constraint correction .

[0196] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dual-drive calculation method for dynamic efficiency of a CSP steam turbine under variable operating conditions, characterized in that: The steps include: Step 1: Collect real-time meteorological data of the CSP plant and real-time data of the steam turbine equipment; Step 2: Using the collected real-time data of the steam turbine equipment and the turbine design parameters, the Flügel formula is reconstructed, and a temperature correction term is introduced to establish a new physical model for dynamic calculation of the turbine's variable operating efficiency, thereby obtaining the actual operating efficiency of the turbine. Step 3: The collected real-time meteorological data and the actual operating efficiency of the steam turbine calculated by the physical model are fed into the trained neural network prediction model for the variable operating efficiency constraint of the steam turbine to obtain the dynamic efficiency of the steam turbine after constraint correction. In step 2, a new physical model for dynamic calculation of turbine efficiency under variable operating conditions is established as follows: ； in, is the actual operating efficiency of the steam turbine, η0 is the design operating efficiency of the steam turbine; m is the actual operating steam mass flow rate of the steam turbine, m0 is the design operating steam mass flow rate of the steam turbine; is the flow area variation index; is the actual operating steam inlet temperature of the turbine, is the steam inlet temperature of the steam turbine under design conditions; β is the entropy increase coefficient; The reconstructed Flügel formula is as follows: ； in, is the actual steam inlet pressure of the turbine under working conditions, is the actual exhaust pressure of the steam turbine. is the steam inlet pressure of the turbine under design conditions, is the exhaust pressure of the steam turbine under design conditions; In step 3, the neural network prediction model includes an input layer, a hidden layer and an output layer, and the hidden layer includes three layers; The eigenvector of the input layer is: ； Among them, DNI is the direct solar irradiance, and Metdata is meteorological data including temperature, air pressure, humidity, wind speed and wind direction. is the actual operating steam inlet pressure, is the actual operating steam inlet temperature of the steam turbine, m is the actual operating steam mass flow rate of the steam turbine, is the actual operating efficiency of the steam turbine; First hidden layer: ； ； in, is the linear transformation output of the first hidden layer; W [1] is the weight matrix of the first hidden layer, ; b [1] is the bias vector of the first hidden layer, with dimension R 128 ;ReLU is the activation function; a [1] It is z [1] The output after ReLU activation has a dimension of R 128 ; Second hidden layer: ； ； in, is the linear transformation output of the second hidden layer; W [2] is the weight matrix of the second hidden layer, ; b [2] is the bias vector of the second hidden layer, with dimension R 64 ;a [2] It is z [2] The output after ReLU activation has a dimension of R 64 ; Third hidden layer: ； ； in, is the linear transformation output of the third hidden layer; W [3] is the weight matrix of the third hidden layer, ; b [3] is the bias vector of the third hidden layer, with dimension R 32 ;a [3] It is z [3] The output after ReLU activation has a dimension of R 32 ; The output layer: ； Among them, W [4] is the weight matrix of the output layer, with dimension R 1×32 ;a [3] is the activation output vector of the third hidden layer, with a dimension of 32; b [4] is the bias vector of the output layer; is the turbine efficiency predicted by the neural network prediction model; The turbine efficiency predicted by the neural network prediction model is constrained and corrected, and the dynamic efficiency of the turbine after constraint correction is obtained as follows: ； in, is Carnot efficiency; ； Among them, T cond is the saturation temperature corresponding to the condenser pressure, obtained by the following formula: ； Among them, P cond is the condenser pressure.

2. A dual-drive calculation method for dynamic efficiency of a CSP steam turbine under variable operating conditions according to claim 1, characterized in that: In step 1, real-time meteorological data is obtained from the CSP plant meteorological station, including real-time normal solar direct irradiance, temperature, air pressure, humidity, wind speed, and wind direction, with a collection frequency of 1 minute per time; real-time data of the steam turbine equipment is obtained from the CSP plant information management system, including steam mass flow, steam inlet pressure, exhaust pressure, and main steam temperature.

3. The dual-drive calculation method for dynamic efficiency of a CSP steam turbine under variable operating conditions according to claim 1 is characterized in that: In step 3, during the training of the neural network prediction model, the dynamic efficiency of the steam turbine after constraint correction of the neural network output is controlled by the loss function. The actual operating efficiency η of the steam turbine calculated by the physical model in step 2 t The deviation is within the set range; Loss Function as follows: ； in, is the final turbine efficiency of the i-th sample after being output by the neural network and passing the Carnot efficiency constraint; is the actual operating efficiency of the steam turbine calculated by the physical model in step 2 for the i-th sample; n is the number of training samples; is the dynamic weight coefficient: ； Among them, βs is the adjustment intensity coefficient, and the initial value is set to 5; ε is the threshold: ； in, is the actual operating steam inlet temperature of the turbine, It is the steam inlet temperature of the turbine under design operating conditions.

4. A dual-drive calculation method for dynamic efficiency of a CSP steam turbine under variable operating conditions according to claim 3, characterized in that: The loss function is minimized through gradient calculation. The data fitting term gradient is used to drive the neural network prediction results to approach the true efficiency value. The physical constraint term gradient is used to limit the deviation between the neural network prediction results and the values ​​calculated by the physical model in step 2. Gradient of the data fitting term: ； Physical constraint gradient: ； in, represents the data fitting term, represents the gradient operator, is the turbine efficiency predicted by the neural network prediction model; represents the dynamic efficiency of the steam turbine after constraint correction and the actual operating efficiency η of the steam turbine t The difference, Represents a physical constraint.

Images