Method for identifying shafting parameters of water-turbine generator set

Abstract

A water-turbine generator set shaft system parameter identification method comprises the following steps: constructing a water-turbine generator set shaft system equation, and carrying out nondimensionalization on the water-turbine generator set shaft system equation; constructing a neural network NN part of the adaptive staged physical information neural network by adopting a full-connection neural network framework; the method comprises the following steps: selecting structural parameters of a hydro-generator needing to be predicted, and then initializing the structural parameters within a range; taking the shafting equation as physical driving information of the physical information neural network, and taking the data set in the step 1 as data driving information of the physical information neural network; based on the physical driving information and the data driving information in the step 4, constructing a multi-target adaptive loss function; and training the loss function by adopting a staged training strategy, so that the loss function is gradually reduced to obtain a structure parameter prediction value of the neural network. According to the method, the physical rule is combined with data driving, and the physical equation is used as a constraint condition to be fused into the loss function of the neural network, so that the interpretability and generalization ability of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of shaft system analysis, and in particular relates to a method for identifying shaft system parameters of a hydro-generator set. Background Art

[0002] Hydroelectric generators are the core equipment of hydropower stations, converting the mechanical energy of water into electrical energy. Their operational stability directly impacts the economic benefits of hydropower stations. However, due to factors such as installation techniques, hydraulics, and mechanics, the shafting parameters of hydroelectric generators can deviate from acceptable ranges. This can lead to problems such as abnormal vibration, collision, and friction in the shafting system. To ensure the safe operation of hydroelectric generators, it is essential to know the specific values of their shafting parameters. Current estimation methods for these parameters suffer from insufficient accuracy, especially when noisy data is present, resulting in poor prediction results.

[0003] In recent years, with the rapid development of computer technology and artificial intelligence, neural networks have achieved remarkable results in data regression and nonlinear inversion. However, methods that rely solely on data-driven methods often suffer from insufficient generalization capabilities, resulting in poor performance when processing unseen data. Summary of the Invention

[0004] The present invention provides a method for identifying parameters of a shaft system of a hydro-generator set to solve the problems raised in the above-mentioned background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for identifying shaft parameters of a hydro-generator set comprises the following steps: Step 1: Construct the shaft system equation of the hydro-generator set and make it dimensionless. Use the fourth-order Runge-Kutta algorithm to solve the dimensionless shaft system equation to obtain the data set. Step 2: Use the fully connected neural network framework to build the neural network NN part of the adaptive staged physical information neural network and determine the input unit and output unit of the neural network; Step 3: Select the structural parameters of the hydro-generator to be predicted, set the value range of the structural parameters based on expert experience, and then initialize the structural parameters within the value range; Step 4: Use the shaft system equation as the physical driving information of the physical information neural network, and define the physical loss based on the physical driving information; use the data set in step 1 as the data driving information of the physical information neural network, and define the data loss based on the data driving information; Step 5: Based on the physical driving information and data driving information in step 4, a multi-objective adaptive loss function is constructed to optimize the performance of the adaptive staged physical information neural network. Step 6: Use a phased training strategy to train the loss function so that the loss function gradually decreases and the predicted values of the structural parameters of the neural network are obtained.

[0006] Furthermore, in step 1, the shaft system equation of the hydro-generator set is constructed as follows: the shaft system is simplified to a double-disc torsional vibration system consisting of a rotor and a runner, and the external forces are considered to be unbalanced magnetic pull and oil film force. The shaft system equation of the hydro-generator is expressed as: ; in, are the masses of the generator rotor and turbine runner respectively, are the damping at the rotor and bearing respectively, are the rotor axis displacement, are the axial displacement of the runner, are the stiffness of the rotor and bearing respectively, Rotor Unbalanced magnetic pull in the direction of Wheel The oil film force in the direction of are the mass eccentricities of the rotor and runner, is the angular velocity, is the rotation angle of the rotor and the runner.

[0007] Furthermore, in step 1, the dimensionless processing of the shaft system equation of the hydro-generator set is constructed, specifically: ; Where t is time, τ is dimensionless time, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, are the rotor axis displacement, are the axial displacement of the runner, are the damping at the rotor and bearing respectively, are the stiffness of the rotor and bearing respectively, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is The dimensionless displacement in the direction, are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, respectively.

[0008] Furthermore, in step 1, the dimensionless axis system equation is as follows: ; in, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, Rotor Unbalanced magnetic pull in the direction of Wheel The oil film force in the direction of are the mass eccentricities of the rotor and runner respectively; 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is The dimensionless displacement in the direction, are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, respectively.

[0009] Furthermore, in step 2, the input of the neural network is normalized using maximum and minimum normalization. ; in, is the original eigenvalue, min is the minimum value of the original eigenvalue, and max is the maximum value of the original eigenvalue.

[0010] Furthermore, in step 4, the physical loss is defined as follows based on the physical driving information: ; in, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, For the rotor Unbalanced magnetic pull in the direction of is the mass eccentricity of the rotor; are the rotor axis displacement, are the axial displacement of the runner, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is Dimensionless displacement in direction; are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, is the number of matching points, is the dimensionless time point of the collocation point.

[0011] Furthermore, in step 4, the data loss is the mean square error between the predicted values of the neural network and the true values of the rotor and wheel, which is defined as follows: ; in, is the number of training data points, is the dimensionless time point of the training point, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is Dimensionless displacement in the direction.

[0012] Furthermore, in step five, the adaptive loss function is as follows: ; in, and are the adaptive loss coefficients for data loss term and physical loss term, respectively. is the data loss term, is the physical loss term, are the neural network parameters.

[0013] Furthermore, in step 5, "building a multi-objective adaptive loss function based on the physical driving information and data driving information in step 4", the following steps are specifically included: For the physical loss term, negative log-likelihood is used as the loss function, which is in the form of: ; Data loss term is output The Gaussian probability model of is expressed as follows: ; in, 、 Represent the output of physical loss term and data loss term respectively, is the predicted value of the neural network, and are the adaptive loss coefficients for data loss term and physical loss term, respectively. are neural network parameters, is the physical loss item; The adaptive loss function is obtained by combining the loss function and the joint probability model.

[0014] Furthermore, in step six, a phased training strategy is adopted, specifically including: in the first phase of training, the parameters of the neural network and the unknown parameters are trained synchronously using the adaptive weight method; in the second phase of training, the parameters of the neural network are frozen, and only the optimizer L-BFGS-B is used to optimize the physical loss term.

[0015] The present invention can achieve the following beneficial effects: By integrating physical laws with data-driven approaches, this invention uses physical information neural networks to accurately estimate the shaft system parameters of hydro-generator sets, significantly improving prediction accuracy and enhancing model generalization capabilities, thus ensuring the safe and stable operation of core equipment in hydropower stations. It has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 This is a schematic diagram of the adaptive phased physical information neural network framework proposed in the present invention; Figure 2 The figure is a schematic diagram of the shaft system structure of the hydro-generator set of the present invention.

[0017] In the accompanying drawings, the components represented by the reference numerals are as follows: 1. Upper guide bearing; 2. Generator rotor; 3. Lower guide bearing; 4. Flange; 5. Water guide bearing; 6. Turbine runner. DETAILED DESCRIPTION

[0018] To facilitate understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The accompanying drawings provide embodiments of the present application. However, the present application may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0019] This application proposes a method for identifying the shaft parameters of a hydro-generator set. Figure 1 As shown, the specific steps include: Step 1: Construct the shaft system equation of the hydro-generator set and make it dimensionless. Use the fourth-order Runge-Kutta algorithm to solve the dimensionless shaft system equation to obtain the data set.

[0020] The specific construction of the shaft system equation of the hydro-generator set is as follows: the shaft system is simplified into a double-disc torsional vibration system consisting of the generator rotor 2 and the turbine runner 6. The simplified model is as follows: Figure 2 As shown, Figure 2 From top to bottom, there are upper guide bearing 1, generator rotor 2, lower guide bearing 3, flange 4, water guide bearing 5 and turbine runner 6. Considering the unbalanced magnetic pull and oil film force as external forces, the shaft system equation of the turbine generator is expressed as: (1); in, are the masses of the generator rotor and turbine runner respectively, are the damping at the rotor and bearing respectively, are the rotor axis displacement, are the axial displacement of the runner, are the stiffness of the rotor and bearing respectively, Rotor Unbalanced magnetic pull in the direction of Wheel The oil film force in the direction of are the mass eccentricities of the rotor and runner, is the angular velocity, is the rotation angle of the rotor and the runner.

[0021] The dimensionless processing of the shaft system equation of the hydro-generator set is constructed as follows: (2); Where t is time, τ is dimensionless time, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, are the rotor axis displacement, are the axial displacement of the runner, are the damping at the rotor and bearing respectively, are the stiffness of the rotor and bearing respectively, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is The dimensionless displacement in the direction, are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, respectively.

[0022] The dimensionless axis system equation is as follows: (3); The fourth-order Runge-Kutta algorithm is used to solve the shaft equation of the above hydro-generator to obtain the training data set. =[0,3], , intercept the data of [1, 2] as the training set.

[0023] Step 2: Use the artificial neural network ANN framework to build the neural network NN part of the adaptive staged physical information neural network AST-PINN, and determine the input unit and output unit of the neural network.

[0024] The input layer has an input unit of ,in, is the dimensionless observation time point, used for data loss term; is the dimensionless collocation time point, used for the physical loss term. The hidden layer is set to 4, with each layer containing 200 neural units. To avoid gradient saturation, the activation function is the hyperbolic tangent function tanh. The output layer has 4 neural units corresponding to the predicted axial and radial displacements of the generator rotor and the axial and radial displacements of the turbine runner, respectively. The neural network weights are initialized using the Xavier method.

[0025] The input of the neural network is normalized using maximum and minimum normalization, that is, (4); in, is the original eigenvalue, min is the minimum value of the original eigenvalue, and max is the maximum value of the original eigenvalue. Through this formula, each eigenvalue in the original data will be scaled to the range between -1 and 1.

[0026] Step 3: Select the structural parameters of the hydro-generator to be predicted, set the value range of the structural parameters based on expert experience, and then initialize the structural parameters within the range.

[0027] In the shaft system equations, dimensionless stiffness and damping coefficients Set the parameters to be identified as {0.107, 0.124, 1.53, 2.041}, and the initial training values as {0.3, 0.3, 0.5, 0.5}.

[0028] Step 4: Use the axis system equation as the physical driving information of the physical information neural network, and use the data set in step 1 as the data driving information of the physical information neural network.

[0029] Physical loss is defined based on physical driver information as follows: (5); in, is the dimensionless time point of the collocation point; is the number of matching points, totaling 1000 points; the explanations of other parameters are the same as above.

[0030] The data-driven information is the mean square error between the predicted values of the neural network rotor and the true value of the wheel. The data loss is defined as follows: (6); in, is the number of training data points, a total of 20 points; is the dimensionless time point of the training point; the interpretation of other parameters is the same as above.

[0031] Step 5: Based on the physical driving information and data driving information in step 4, a multi-objective adaptive loss function is constructed to optimize the performance of the adaptive staged physical information neural network AST-PINN.

[0032] The adaptive loss function is shown as follows: (7); in, and are the adaptive loss coefficients for data loss term and physical loss term, respectively. is the data loss term, For physical loss, add and To avoid excessive drop in the adaptive loss coefficient.

[0033] The specific derivation process of the adaptive loss function is: Assume that the likelihood distribution of the neural network output is a mean , the variance is For the physical loss term, we output Gaussian distribution With an uncertainty parameter .

[0034] During the training process of a neural network, uncertain parameters can be optimized using the maximum likelihood function. Since the goal is to minimize the objective function, the negative log-likelihood can be used as the loss function, which has the following form: (8); Data loss term is output The Gaussian probability model of is expressed as follows: (9); Combining the above equations (8) and (9), we get the total loss function equation (7). The relevant parameters in (7), (8), and (9) are consistent with the above, and the remaining parameters are common knowledge disclosed by the Gaussian formula. Specifically, 、 Represent the output of physical loss term and data loss term respectively, is the predicted value of the neural network, are the neural network parameters.

[0035] Step 6: Use a phased training strategy to train the loss function so that the loss function gradually decreases and the predicted values of the structural parameters of the neural network are obtained.

[0036] A phased training strategy is adopted. In the first phase of training, the parameters of the neural network and the unknown parameters are trained synchronously using the adaptive weight method. The optimizer is Adam, the learning rate is 0.001, and the training is repeated 9000 times. In the second phase of training, the parameters of the neural network are frozen, and only the optimizer L-BFGS-B is used to optimize the physical loss term (5), and the training is repeated 500 times.

[0037] The adaptive phased physical information neural network AST-PINN proposed in this invention, The relative error is 1.869%; for The relative error is 4.839%; for The relative error is 0%; for The relative error is 0.098%.

[0038] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for identifying shaft parameters of a hydro-generator set, characterized in that: The following steps are involved: Step 1: Construct the shaft system equation of the hydro-generator set and make it dimensionless. Use the fourth-order Runge-Kutta algorithm to solve the dimensionless shaft system equation to obtain the data set. Step 2: Use the fully connected neural network framework to build the neural network NN part of the adaptive staged physical information neural network and determine the input unit and output unit of the neural network; Step 3: Select the structural parameters of the hydro-generator to be predicted, set the value range of the structural parameters based on expert experience, and then initialize the structural parameters within the value range; Step 4: Use the shaft system equation as the physical driving information of the physical information neural network, and define the physical loss based on the physical driving information; use the data set in step 1 as the data driving information of the physical information neural network, and define the data loss based on the data driving information; Step 5: Based on the physical driving information and data driving information in step 4, a multi-objective adaptive loss function is constructed to optimize the performance of the adaptive staged physical information neural network. Step 6: Use a phased training strategy to train the loss function so that the loss function gradually decreases and the predicted values of the structural parameters of the neural network are obtained.

2. The method for identifying shaft parameters of a hydro-generator set according to claim 1, characterized in that: In step 1, the shaft system equation of the hydro-generator set is constructed as follows: the shaft system is simplified to a double-disc torsional vibration system consisting of a rotor and a runner, and the external forces are considered to be unbalanced magnetic pull and oil film force. The shaft system equation of the hydro-generator is expressed as: ； in, are the masses of the generator rotor and turbine runner respectively, are the damping at the rotor and bearing respectively, are the rotor axis displacement, are the axial displacement of the runner, are the stiffness of the rotor and bearing respectively, Rotor Unbalanced magnetic pull in the direction of Wheel The oil film force in the direction are the mass eccentricities of the rotor and runner, is the angular velocity, is the rotation angle of the rotor and the runner.

3. The method for identifying shaft parameters of a hydro-generator set according to claim 2, characterized in that: In step 1, the dimensionless processing of the shaft system equation of the hydro-generator set is constructed, specifically: ； Where t is time, τ is dimensionless time, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, are the rotor axis displacement, are the axial displacement of the runner, are the damping at the rotor and bearing respectively, are the stiffness of the rotor and bearing respectively, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is The dimensionless displacement in the direction, are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, respectively.

4. A method for identifying shaft parameters of a hydro-generator set according to claim 3, characterized in that: In step 1, the dimensionless axis system equation is as follows: ； in, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, Rotor Unbalanced magnetic pull in the direction of Wheel The oil film force in the direction are the mass eccentricities of the rotor and runner respectively; 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is The dimensionless displacement in the direction, are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, respectively.

5. The method for identifying shaft parameters of a hydro-generator set according to claim 1, characterized in that: In step 2, the input of the neural network is normalized using the maximum and minimum normalization. ； in, is the original eigenvalue, min is the minimum value of the original eigenvalue, and max is the maximum value of the original eigenvalue.

6. The method for identifying shaft parameters of a hydro-generator set according to claim 1, characterized in that: In step 4, the physical loss is defined based on the physical driver information as follows: ； in, is the angular velocity, is the characteristic length, are the masses of the generator rotor and turbine runner respectively, For the rotor Unbalanced magnetic pull in the direction of is the mass eccentricity of the rotor; are the rotor axis displacement, are the axial displacement of the runner, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is Dimensionless displacement in direction; are the dimensionless damping at the rotor and bearing, respectively, are the dimensionless stiffness of the rotor and bearing, is the number of matching points, is the dimensionless time point of the collocation point.

7. The method for identifying shaft parameters of a hydro-generator set according to claim 6, characterized in that: In step 4, the data loss is the mean square error between the predicted values of the neural network and the true values of the rotor and wheel, which is defined as follows: ； in, is the number of training data points, is the dimensionless time point of the training point, 、 The rotor is The dimensionless displacement in the direction, 、 The wheel is Dimensionless displacement in the direction.

8. The method for identifying shaft parameters of a hydro-generator set according to claim 7, characterized in that: In step 5, the adaptive loss function is as follows: ； in, and are the adaptive loss coefficients for data loss term and physical loss term, is the data loss term, is the physical loss term, are the neural network parameters.

9. The method for identifying shaft parameters of a hydro-generator set according to claim 8, characterized in that: In step 5, "building a multi-objective adaptive loss function based on the physical driving information and data driving information in step 4", the following steps are specifically included: For the physical loss term, negative log-likelihood is used as the loss function, which is in the form of: ； Data loss term is output The Gaussian probability model of is expressed as follows: ； in, 、 Represent the output of physical loss term and data loss term respectively, is the predicted value of the neural network, and are the adaptive loss coefficients for data loss term and physical loss term, are neural network parameters, is the physical loss item; The adaptive loss function is obtained by combining the loss function and the joint probability model.

10. The method for identifying shaft parameters of a hydro-generator set according to claim 1, characterized in that: In step six, a phased training strategy is adopted, specifically including: in the first phase of training, the parameters of the neural network and the unknown parameters are trained synchronously using the adaptive weight method; in the second phase of training, the parameters of the neural network are frozen, and only the optimizer L-BFGS-B is used to optimize the physical loss term.

Images