Direct air cooling unit axial flow fan group wind field reconstruction method based on physical information neural network

Abstract

The invention relates to the field of cold end optimization control systems, in particular to a method for reconstructing a wind field of an axial flow fan group of a direct air cooling unit based on a physical information neural network. Comprising the following steps: S1, arranging a plurality of wind speed and pressure sensors near an axial flow fan group of the direct air cooling unit, and collecting sample data when axial flow fans rotate at a high speed by taking wind speeds and pressures in all directions of a wind field of the axial flow fan group of the unit as target variables; s2, performing data preprocessing on the sample data; s3, determining a differential equation of fluid motion satisfied by the axial flow fan group area, and constructing a physical information neural network of the axial flow fan group; s4, training a physical information neural network model of the axial flow fan group by using the preprocessed data obtained in the S2; and S5, reconstructing a wind field of the axial flow fan group of the direct air cooling unit according to the trained physical information neural network model of the axial flow fan group in the step S4. According to the method, the wind field data of the axial flow fan group of the direct air cooling unit is quickly and accurately predicted by using a physical information neural network method.

Description

Technical Field

[0001] The invention relates to the field of cold end optimization control systems, and in particular to a wind field reconstruction method for an axial flow fan group of a direct air cooling unit based on a physical information neural network. Background Art

[0002] The wind field of the axial flow fan is an important feature of the direct air-cooled condenser. The high-speed rotating fan blades pressurize the ambient air and blow it to the surface of the condenser, thereby condensing the exhaust steam that has done work in the turbine. However, the air pressurized by the high-speed rotating fan blades is easily affected by environmental factors (wind direction, wind speed, etc.), which reduces the actual amount of air involved in condensation, thereby reducing the efficiency of condensation. In order to maintain a stable back pressure, it is necessary to increase the fan speed, which in turn increases power consumption and reduces the economic benefits of the power plant.

[0003] At the same time, the air that has participated in the heat exchange is difficult to be discharged in time due to the influence of environmental wind and buildings, and it is easy to be sucked into the fan again to participate in the heat exchange, which will lead to a further decrease in condensation efficiency and further affect the economic benefits of the power plant. In summary, the establishment of the wind field of the axial flow fan group is particularly important. By establishing the wind field of the axial flow fan group, we can understand the air flow near the axial flow fan group, thereby adjusting the speed and operation mode of the axial flow fan or installing wind shields, thereby changing the flow field near the axial flow fan, increasing the air involved in the heat exchange and enhancing the economic benefits. However, the efficient prediction of the wind field of the axial flow fan group of direct air-cooled units has always been a technical problem that needs to be solved urgently.

[0004] In order to explore the changing law of the wind field of the axial flow fan group of the direct air-cooled unit and its influence mechanism on the condenser heat exchange, domestic and foreign scientific researchers have carried out a lot of research, mainly including experimental methods and numerical simulation methods.

[0005] Although the experimental method can reflect the actual working conditions of the axial flow fan group, due to cost issues, the number of sensors arranged near the axial flow fan group is limited, making it difficult to accurately describe the overall shape of the wind field of the axial flow fan group and the details of the local vortex shape. The numerical simulation method is a relatively popular method recently (FLUENT, CFD), but it has very high requirements on the mesh quality. When you want to simulate the turbulent field generated by the high-speed rotation of the axial flow fan, the required pre-processing, calculation time and post-processing of massive data are very amazing.

[0006] A physical information neural network is a neural network that imposes physical information constraints during the training process. Compared with traditional neural networks, physical information neural networks can obtain stronger generalization capabilities with fewer data samples. At the same time, the physical information neural network method can make full use of data information to solve problems such as high-dimensional nonlinearity and difficulty in establishing physical models. To this end, the present invention proposes a wind field reconstruction method for an axial flow fan group based on a physical information neural network. Summary of the invention

[0007] In order to solve the problems existing in the prior art, the present invention provides a method for reconstructing the wind field of an axial flow fan group of a direct air-cooled unit based on a physical information neural network. The method uses the measurement data of wind speed and pressure sensors, combines neural network technology and the Navier-Stokes equations to construct a substitution model, and can fully mine the hidden physical information under the condition of incomplete data information. By minimizing the loss function, the mapping relationship between each output physical quantity and the model input is obtained, and the wind field of the axial flow fan group of the direct air-cooled unit is quickly and accurately predicted using the monitoring data of limited sensors.

[0008] The present invention adopts the following technical scheme: a method for reconstructing the wind field of an axial flow fan group of a direct air cooling unit based on a physical information neural network, comprising: S1: Multiple wind speed and pressure sensors are arranged near the axial flow fan group of the direct air cooling unit, and the wind speed and pressure in each direction of the wind field of the axial flow fan group of the unit are used as target variables to collect sample data when the axial flow fan rotates at high speed; S2: Preprocess the sample data, construct a training data matrix, and divide the data matrix into a training set and a test set; S3: Determine the differential equation of fluid motion satisfied by the axial flow fan group area, and construct the physical information neural network of the axial flow fan group; S4: Use the preprocessed data obtained in S2 to train the physical information neural network model of the axial flow fan group, and optimize the back propagation gradient during the training process; minimize the loss function composed of the physical loss term and the data loss term to obtain the mapping relationship between the model input and each output physical quantity, and quickly predict the wind field parameters of the axial flow fan; input the test set into the trained model to evaluate the accuracy and generalization of the model; S5: Reconstruct the wind field of the axial flow fan group of the direct air cooling unit according to the physical information neural network model of the axial flow fan group trained in S4.

[0009] In some embodiments, step S1 includes: S11: Multiple wind speed sensor measuring points are arranged near the axial flow fan group of the unit to collect wind speed measurement values ​​in various directions at time and space coordinates; S12: multiple pressure sensor measuring points arranged near the axial flow fan group of the unit are used to collect pressure measurement values ​​at the time and space coordinates; S13: Read the ambient temperature at the corresponding sampling time , Ambient wind speed and turbine exhaust steam Measurement data of S14: According to the acquisition time corresponding to the corresponding measuring point coordinates and the ambient temperature at the corresponding sampling time , Ambient wind speed and turbine exhaust steam The measured data are combined with the wind speed and pressure data in each direction measured by the sensors near the corresponding axial flow fan group of the unit to construct a data set composed of sampling time t, spatial coordinates x, y, z of the measuring point, and ambient temperature. , Ambient wind speed , Turbine exhaust steam flow , lateral speed u , longitudinal speed v , vertical speed w ,pressure p The sample data composed of .

[0010] In some embodiments, step S2 includes: S21: Use the max-min normalization method to map the data measured by the sensor to [0,1]; S22: The preprocessed data is combined into a [ N d ,11], where N d Represents the total number of measurement points in the measurement data matrix. The i-th row of the measurement data matrix A includes the time-space coordinate points , environment and working condition information And the flow field information corresponding to this position ; in, is the time information corresponding to the i-th row of data, is the corresponding horizontal axis coordinate, is the corresponding vertical axis coordinate, is the corresponding vertical axis coordinate, is the ambient temperature at the corresponding moment, is the ambient wind speed at the corresponding moment, is the steam exhaustion of the steam turbine at the corresponding moment; is the lateral velocity corresponding to the space-time coordinate point, is the corresponding longitudinal velocity, is the corresponding vertical velocity, is the corresponding pressure information; S23: Divide the measurement data matrix A into a training set and a test set.

[0011] In some embodiments, step S3 includes: S31: Building a fully connected neural network: S32: In the fully connected neural network, partial differential equations reflecting the flow field laws are embedded to learn the fluid flow laws in the wind field.

[0012] In some embodiments, in step S31, the fully connected neural network is represented in the following recursive form: in, is the input layer of the neural network; t is the time information; x, y, z represent the spatial coordinates; Indicates the ambient temperature, Indicates the ambient wind speed, Indicates the exhaust steam flow of the steam turbine; The neural network i layer; and Represent the weight and bias of the network respectively; The neural network i- 1st floor; is the activation function; l is the number of hidden layers of the neural network; is the output layer; and denote the output layer weights and biases, For the l -1st floor; u is the lateral velocity information, v is the longitudinal velocity information, w is the vertical velocity information, p For pressure information.

[0013] In some embodiments, in step S32, the partial differential equation reflecting the flow field law is: in is the partial differential symbol; t is time; x, y, z represent spatial coordinates; u is the lateral velocity, v is the longitudinal velocity, w is the vertical velocity; is the air density; P For pressure; is the kinematic viscosity of air.

[0014] In some embodiments, step S4 includes: S41: Construct a total loss function consisting of a physics loss term and a data loss term; S42: Use the Adam optimization function to update the training parameters during the back-propagation process.

[0015] S43: Use the test set constructed in S2 to input the trained model to evaluate the accuracy and generalization of the model.

[0016] In step S41, the total loss function includes: in, Loss represents the total loss function, , Represents the numerical loss , partial differential equation loss The weight coefficient of and are data sets for numerical and partial differential equations, respectively. is the predicted value of the lateral velocity; is the predicted value of longitudinal velocity; is the predicted value of vertical velocity; is the predicted value of pressure information; For the The time information corresponding to the row measurement data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate; is the measured value of the lateral velocity; is the measured value of longitudinal velocity; is the measured value of vertical velocity; is the measured value of pressure information; is the residual of the continuity equation; are the residuals of the three terms of the Navier-Stokes equations; For the i Time information corresponding to the row coordinate data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate.

[0017] 9. The method for reconstructing the wind field of an axial flow fan group of a direct air cooling unit based on a physical information neural network according to claim 7, characterized in that the step S42 comprises: Use the Adam optimization function to back-propagate and update the training parameters W ,b , get the best training parameters , Minimize the loss function consisting of the physics loss term and the data loss term: in, W is the weight; b For bias.

[0018] 10. The method for reconstructing the wind field of an axial flow fan group of a direct air cooling unit based on a physical information neural network according to claim 1, characterized in that step S5 comprises: S51: Based on the network trained in step S4, the wind field of the axial flow fan group of the direct air cooling unit is reconstructed, and the horizontal, longitudinal and vertical speeds output by the network are converted into u, v, w Synthesize the final wind speed vector; S52: Based on the trained axial flow fan group physical information neural network model, the physical information neural network model of the axial flow fan group is trained, adjusted and updated offline using the data collected by the sensor in real time, and the error between the wind speed and wind pressure of the inverted wind field and the wind speed and wind pressure of the actual measuring point is calculated online. , , and , and set the error threshold according to actual project needs :If the prediction error calculated in real time , , and Both are less than the threshold , that is, the error is within an acceptable range, then continue to use the current physical neural network model; if the prediction error calculated in real time is , , and One or more of the , that is, the error exceeds the acceptable range, the offline trained axial flow fan group physical information neural network model is used to re-invert the wind field.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a wind field reconstruction method for axial flow fan groups of direct air-cooling units based on physical information neural networks. By introducing the NS equation, a deep neural network model with an embedded physical information loss function is constructed. This method can fully mine the hidden physical information under limited data samples, realize the wind field reconstruction and prediction of axial flow fan groups of direct air-cooling units, and provide a new idea for the rapid and accurate prediction of the wind field of axial flow fan groups of direct air-cooling units.

[0020] The present invention uses a physical information neural network method to quickly and accurately predict the wind field data of the axial flow fan group of the direct air-cooled unit. Compared with traditional numerical simulation, experiments, and machine learning methods, it can save a lot of computing resources and experimental costs, and reflect the flow field laws more completely and accurately, providing technical guidance for subsequent cold end optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is an overall flow chart of a wind field reconstruction method of an axial flow fan group of a direct air cooling unit based on a physical information neural network of the present invention; Figure 2 It is a framework diagram of the physical information neural network model of the present invention; Figure 3 This is a diagram of the sensor arrangement of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0023] like Figure 1 It is an overall flow chart of the method embodiment of the present invention, comprising the following steps: S1: Multiple wind speed and pressure sensors are arranged near the axial flow fan group of the direct air cooling unit. The wind speed and pressure in each direction of the wind field of the axial flow fan group of the unit are used as target variables to collect sample data when the axial flow fan rotates at high speed.

[0024] S2: Preprocess the sample data and construct a training data matrix, using 80% of the data as the training set and the remaining 20% ​​as the test set.

[0025] S3: Determine the differential equations of fluid motion satisfied by the axial fan group area, and construct a neural network model embedded with a physical prior loss function.

[0026] S4: Use the preprocessed data obtained in S2 to train the axial flow fan group physical information neural network model, and optimize the back propagation gradient during the training process; minimize the loss function composed of the physical loss term and the data loss term to obtain the mapping relationship between the model input and each output physical quantity, and quickly predict the axial flow fan wind field parameters. Input the test set into the trained model to evaluate the accuracy and generalization of the model.

[0027] S5: Based on the physical neural network model trained in S4, the wind field of the axial flow fan group of the direct air-cooling unit is reconstructed.

[0028] In this embodiment, the specific process of step S1 includes: S11, multiple wind speed sensor measuring points arranged near the axial flow fan group of the unit collect wind speed measurement values ​​in various directions at the time and space coordinates. This embodiment uses a thermal sensitive anemometer.

[0029] S12, multiple pressure sensor measuring points arranged near the axial flow fan group of the unit are collected to obtain pressure measurement values ​​at the time and space coordinates. In this embodiment, a high-precision differential pressure gauge is used to obtain the pressure.

[0030] Specifically, the thermal anemometer and high-precision differential pressure gauge are arranged in each cooling unit of the condenser (i.e. Figure 3 In the center of the axial flow fans (C11-C88), a total of 64 are arranged.

[0031] S13, in order to adapt to the dynamic changes in the load of the direct air-cooling unit and the interference of environmental factors, read the ambient temperature at the corresponding sampling time from the power plant data storage server , Ambient wind speed and turbine exhaust steam Taking it as part of the network input enables the network to have the ability to describe wind fields under multiple working conditions and different environmental influences.

[0032] S14. Collect the corresponding measuring point coordinates, the corresponding collection time, and the ambient temperature at the sampling time. , Ambient wind speed and turbine exhaust steam The sample data is constructed by combining the measurement data of the corresponding unit's axial flow fan group with the wind speed and pressure data in all directions measured by sensors near the unit's axial flow fan group.

[0033] In this embodiment, the specific process of step S2 includes: S21: In order to avoid the difficulty of adjusting the neural network training parameters caused by the order of magnitude difference between different physical quantities, and the significant impact of this difference on the size of the back-propagation gradient, we use the max-min normalization method to non-dimensionalize physical quantities such as wind speed and pressure; mapping the physical quantities measured by the sensor (wind speed and pressure in all directions, etc.) to [0,1] can improve the performance and convergence speed of the model. The formula is as follows: (1) in, is the normalized physical quantity, x are the measured physical quantity data, and are the maximum and minimum values ​​in the measured data sets of each physical quantity, respectively.

[0034] S22: The preprocessed low-resolution flow field information is combined into a [ N d ,11], where N d represents the total number of measurement points in the measurement data matrix. The i-th (1≤i≤ N d ) line includes the space-time coordinate point , environment and working condition information And the flow field information corresponding to this position .

[0035] in, is the time information corresponding to the i-th row of data, is the corresponding horizontal axis coordinate, is the corresponding vertical axis coordinate, is the corresponding vertical axis coordinate, is the ambient temperature at the corresponding moment, is the ambient wind speed at the corresponding moment, is the exhaust steam volume of the steam turbine at the corresponding moment. is the lateral velocity corresponding to the space-time coordinate point, is the corresponding longitudinal velocity, is the corresponding vertical velocity, is the corresponding pressure information; S23: Use 80% of the measurement data matrix A as the training set and the remaining 20% ​​as the test set.

[0036] In this embodiment, the specific process of step S3 includes: S31: Construct a fully connected neural network, set the number of layers, initial conditions and hyperparameters according to actual needs so that the neural network has the ability to accurately approximate complex nonlinear partial differential equation systems. In this embodiment, the neural network has 5 hidden layers, 8 neurons in each layer, and then randomly initialized.

[0037] Specifically, the fully connected neural network described in S31 is expressed in the following recursive form: (2) in, is the input layer of the neural network; t For time information; x, y, z Represents spatial coordinates; Indicates the ambient temperature, Indicates the ambient wind speed, Indicates the exhaust steam flow of the turbine. The neural networki layer; W i and b i Represent the weight and bias of the network respectively; The neural network i- 1st floor; is the activation function; l is the number of hidden layers of the neural network; is the output layer; W l and b l denote the output layer weights and biases, For the l -1st floor; u is the lateral velocity information, v is the longitudinal velocity information, w is the vertical velocity information, p is the pressure information; in this embodiment, the ReLU activation function is selected.

[0038] S32: Compared with the traditional data-driven model, this embodiment embeds partial differential equations that reflect the laws of the flow field in the neural network structure to learn the laws of fluid flow in the wind field; it can effectively avoid the problems of overfitting and slow convergence rate of the model.

[0039] Specifically, in S32, the Navier Stokes equation and the continuity equation are embedded in the prediction model as physical constraints by using automatic differentiation technology. Compared with traditional manual calculations, automatic differentiation has higher accuracy and efficiency. Compared with numerical differentiation, automatic differentiation can avoid truncation errors and rounding errors. The process of obtaining the Navier Stokes equation and the continuity equation is as follows: (3) in is the partial differential symbol; t is time; x, y, z Represents spatial coordinates; u is the lateral velocity, v is the longitudinal velocity, w is the vertical velocity; is the air density; take 1.29 kg / m 3 . P For pressure; is the kinematic viscosity of air; the viscosity of air depends mainly on temperature. In this example, at 15°C, the viscosity of air is , the viscosity of air at other temperatures can be obtained by looking up the table.

[0040] In this embodiment, the specific process of step S4 includes: S41: Construct a loss function consisting of a physics loss term and a data loss term.

[0041] Specifically, the loss function constructed in this embodiment includes the wind speed and pressure data measured by sensors near the axial flow fan group and the errors of the wind speed and pressure data obtained by the physical information neural network, that is, the numerical loss. The result of substituting the wind speed and pressure calculated by the physical information neural network into the Navier-Stokes equation and the continuity equation is the partial differential equation loss; the specific definition is as follows: (4) (5) (6) (7) in, Loss Represents the total loss function of the physical information neural network, , Represents the numerical loss , partial differential equation loss The weight coefficient, in this embodiment, Take 0.001, is 1; and are data sets for numerical and partial differential equations, respectively. is the predicted value of the lateral velocity; is the predicted value of longitudinal velocity; is the predicted value of vertical velocity; is the predicted value of pressure information; is the time information corresponding to the i-th row of measurement data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate; is the measured value of the lateral velocity; is the measured value of longitudinal velocity; is the measured value of vertical velocity; It is the measured value of pressure information. is the first residual equation in equation (7); is the second residual equation in equation (7); is the third residual equation in equation (7); is the fourth residual equation in equation (7); is the time information corresponding to the i-th row of coordinate data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate.

[0042] S42: Use the Adam optimization function to update the training parameters during the back-propagation process.

[0043] Use the Adam optimization function to back-propagate and update the training parameters W , b , get the best training parameters , Minimize the loss function consisting of the physics loss term and the data loss term: (8) in, W is the weight; b The maximum number of Epochs in this example is 40000, and the threshold Take 2%.

[0044] S43: Use the test set constructed in S2 to input the model trained in S4 to evaluate the accuracy and generalization of the model.

[0045] Specifically, once the deep neural network has completed the training process, it requires relatively little computing time and resources for practical application. Therefore, the prediction model can establish an accurate mapping relationship between the model input and each output physical quantity by fitting the limited and sparse measurement data and combining the NS equation. In this way, the model can quickly predict the wind field of the axial flow fan group.

[0046] In this embodiment, the specific process of step S5 includes: S51: Based on the network trained in S4, the wind field of the axial flow fan group of the direct air-cooling unit is reconstructed, and the horizontal, longitudinal and vertical speeds output by the network are transformed into u , v , w Resulting in the final wind speed vector.

[0047] S52: In order to maximize economic benefits while retaining the data information that has the greatest impact on the wind farm, the 28 wind speed and wind pressure sensors installed on the outermost cooling unit of the condenser must be retained, because the outermost cooling unit is most affected by changes in wind direction and environmental factors. The internal cooling units are surrounded by external cooling units, and the impact caused by changes in wind direction and environmental factors can be ignored. Therefore, in this embodiment, the sensors of the C33-C66 cooling units are removed, and the sensors of the remaining cooling units are retained.

[0048] S53: Based on the trained physical information neural network model, the model is trained, adjusted and updated offline using the data collected by the sensor in real time after processing in step S52, and the error between the wind speed and wind pressure of the inverted wind field and the wind speed and wind pressure of the actual measuring point is calculated online. , , and , this embodiment sets the error threshold 5%: If the prediction error calculated in real time is , , and Both are less than the threshold , that is, the error is within an acceptable range, then continue to use the current physical neural network model; if the prediction error calculated in real time is , , and One or more of the , that is, if the error exceeds the acceptable range, the offline trained physical information neural network model is used to re-invert the wind field.

Claims

1. A method for reconstructing the wind field of an axial flow fan group of a direct air cooling unit based on a physical information neural network, characterized in that: include: S1: Multiple wind speed and pressure sensors are arranged near the axial flow fan group of the direct air cooling unit, and the wind speed and pressure in each direction of the wind field of the axial flow fan group of the unit are used as target variables to collect sample data when the axial flow fan rotates at high speed; S2: Preprocess the sample data, construct a training data matrix, and divide the data matrix into a training set and a test set; S3: Determine the differential equation of fluid motion satisfied by the axial flow fan group area, and construct the physical information neural network of the axial flow fan group; S4: Use the preprocessed data obtained in S2 to train the physical information neural network model of the axial flow fan group, and optimize the back propagation gradient during the training process; minimize the loss function composed of the physical loss term and the data loss term to obtain the mapping relationship between the model input and each output physical quantity, and quickly predict the wind field parameters of the axial flow fan; input the test set into the trained model to evaluate the accuracy and generalization of the model; S5: Reconstruct the wind field of the axial flow fan group of the direct air cooling unit according to the physical information neural network model of the axial flow fan group trained in S4.

2. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 1 is characterized in that: The step S1 comprises: S11: Multiple wind speed sensor measuring points are arranged near the axial flow fan group of the unit to collect wind speed measurement values ​​in various directions at time and space coordinates; S12: multiple pressure sensor measuring points arranged near the axial flow fan group of the unit are used to collect pressure measurement values ​​at the time and space coordinates; S13: Read the ambient temperature at the corresponding sampling time , Ambient wind speed and turbine exhaust steam Measurement data; S14: According to the acquisition time corresponding to the corresponding measuring point coordinates and the ambient temperature at the corresponding sampling time , Ambient wind speed and turbine exhaust steam The measured data are combined with the wind speed and pressure data in each direction measured by the sensors near the corresponding axial flow fan group of the unit to construct a data set composed of sampling time t, spatial coordinates x, y, z of the measuring point, and ambient temperature. , Ambient wind speed , Turbine exhaust steam flow , lateral speed u , longitudinal speed v , vertical speed w ,pressure p The sample data composed of .

3. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 1 is characterized in that: The step S2 comprises: S21: Use the max-min normalization method to map the data measured by the sensor to [0,1]; S22: The preprocessed data is combined into a [ N d ,11], where N d Represents the total number of measurement points in the measurement data matrix. The i-th row of the measurement data matrix A includes the time-space coordinate points , environment and working condition information And the flow field information corresponding to this position ; in, is the time information corresponding to the i-th row of data, is the corresponding horizontal axis coordinate, is the corresponding vertical axis coordinate, is the corresponding vertical axis coordinate, is the ambient temperature at the corresponding moment, is the ambient wind speed at the corresponding moment, is the steam exhaustion of the steam turbine at the corresponding moment; is the lateral velocity corresponding to the space-time coordinate point, is the corresponding longitudinal velocity, is the corresponding vertical velocity, is the corresponding pressure information; S23: Divide the measurement data matrix A into a training set and a test set.

4. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 1 is characterized in that: The step S3 comprises: S31: Building a fully connected neural network: S32: In the fully connected neural network, partial differential equations reflecting the flow field laws are embedded to learn the fluid flow laws in the wind field.

5. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 4 is characterized in that: In step S31, the fully connected neural network is represented in the following recursive form: in, is the input layer of the neural network; t is the time information; x, y, z represent the spatial coordinates; Indicates the ambient temperature, Indicates the ambient wind speed, Indicates the exhaust steam flow of the steam turbine; The neural network i layer; and Represent the weight and bias of the network respectively; The neural network i- 1st floor; is the activation function; l is the number of hidden layers of the neural network; is the output layer; and denote the output layer weights and biases, For the l -1st floor; u is the lateral velocity information, v is the longitudinal velocity information, w is the vertical velocity information, p For pressure information.

6. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 4 is characterized in that: In step S32, the partial differential equation reflecting the flow field law is: in is the partial differential symbol; t is time; x, y, z represent spatial coordinates; u is the lateral velocity, v is the longitudinal velocity, w is the vertical velocity; is the air density; P For pressure; is the kinematic viscosity of air.

7. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 1 is characterized in that: The step S4 comprises: S41: Construct a total loss function consisting of a physics loss term and a data loss term; S42: Use the Adam optimization function to update the training parameters during the back-propagation process. S43: Use the test set constructed in S2 to input the trained model to evaluate the accuracy and generalization of the model.

8. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 7 is characterized in that: In step S41, the total loss function includes: in, Loss represents the total loss function, , Represents the numerical loss , partial differential equation loss The weight coefficient of and are data sets for numerical and partial differential equations, respectively. is the predicted value of the lateral velocity; is the predicted value of longitudinal velocity; is the predicted value of vertical velocity; is the predicted value of pressure information; For the The time information corresponding to the row measurement data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate; is the measured value of the lateral velocity; is the measured value of longitudinal velocity; is the measured value of vertical velocity; is the measured value of pressure information; is the residual of the continuity equation; are the residuals of the three terms of the Navier-Stokes equations; For the i Time information corresponding to the row coordinate data; is the corresponding horizontal axis coordinate; is the corresponding vertical axis coordinate; is the corresponding vertical axis coordinate.

9. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 7 is characterized in that: The step S42 comprises: Use the Adam optimization function to back-propagate and update the training parameters W , b , get the best training parameters , Minimize the loss function consisting of the physics loss term and the data loss term: in, W is the weight; b For bias.

10. The wind field reconstruction method of the axial flow fan group of the direct air cooling unit based on the physical information neural network according to claim 1 is characterized in that: The step S5 comprises: S51: Based on the network trained in step S4, the wind field of the axial flow fan group of the direct air cooling unit is reconstructed, and the horizontal, longitudinal and vertical speeds output by the network are converted into u, v, w Synthesize the final wind speed vector; S52: Based on the trained axial flow fan group physical information neural network model, the physical information neural network model of the axial flow fan group is trained, adjusted and updated offline using the data collected by the sensor in real time, and the error between the wind speed and wind pressure of the inverted wind field and the wind speed and wind pressure of the actual measuring point is calculated online. , , and , and set the error threshold according to actual project needs :If the prediction error calculated in real time , , and Both are less than the threshold , that is, the error is within an acceptable range, then continue to use the current physical neural network model; if the prediction error calculated in real time is , , and One or more of the , that is, the error exceeds the acceptable range, the offline trained axial flow fan group physical information neural network model is used to re-invert the wind field.

Images