An optimization algorithm for inversion of service stress cloud map of hydropower unit bolts

Abstract

This invention discloses an optimization algorithm for stress cloud map inversion of hydropower unit bolts during service, relating to the field of hydropower unit bolt stress measurement technology. The algorithm includes the following steps: S1: Finite element simulation model construction, including the following steps: S2: Data cleaning of the physical field simulation data and field measurement data obtained in step S1 to obtain a physical field dataset; S3: Stress cloud map inversion model construction; S4: Optimal parameter search. This invention overcomes the shortcomings of existing technologies in actual environment modeling and structural optimization, providing a new technical approach for the maintenance and performance improvement of hydropower unit bolts. It achieves high-precision simulation in actual operating environments, overcoming the shortcomings of existing technologies in digital modeling. Based on theoretical physical constraints, it can construct a reinforcement learning algorithm to invert the stress distribution of bolts during long-term service, providing comprehensive information for structural health monitoring, while existing technologies have certain limitations in this regard.

Description

Technical Field

[0001] This invention relates to the field of stress calculation technology for hydropower unit bolts, specifically to an optimization algorithm for inverting stress cloud diagrams of hydropower unit bolts during service. Background Technology

[0002] In fault diagnosis of hydropower unit equipment, existing technologies mainly focus on structural mechanics and materials engineering, as well as bolt stress analysis. The following are the characteristics of current technical solutions for bolt fault diagnosis in hydropower units:

[0003] (1) In terms of structural mechanics and materials engineering, existing technologies mainly model and analyze the bolt structure of hydropower units through the principles of structural mechanics and materials engineering; focusing on the static analysis of the structure and considering the basic mechanical properties of materials such as strength and stiffness.

[0004] (2) In terms of stress analysis of bolts, the stress analysis of bolts is mainly based on the principle of statics, focusing on the stress state of bolts under stress conditions; and calculating the stress distribution under stress conditions based on the material and geometry of the bolt.

[0005] (3) Regarding finite element simulation technology: Commonly used technical solutions employ finite element simulation technology, but simplifications are often made during modeling, which may not accurately reflect the actual working environment. While considering the overall structure, finite element simulation presents certain challenges in the detailed modeling of local areas.

[0006] (4) Regarding digital twin technology: Most existing technologies do not fully utilize digital twin technology, lack accurate digital simulation of the actual operating environment, and fail to fully consider the impact of actual operating data on the structure and bolt stress. Summary of the Invention

[0007] This invention aims to provide an inversion optimization algorithm for stress cloud diagrams of hydropower unit bolts during service. Through the design of a finite element simulation model of the hydropower unit bolts, the actual operating state of the components is accurately depicted. High-quality data is obtained by cleaning finite element simulation data and actual measurement data for training and testing the prediction model. An embedded physical knowledge neural network is constructed to achieve optimization of finite element input parameters and rapid reconstruction of the three-dimensional flow field. An evaluator network and a participant network based on gradient descent iterative updates are constructed using finite element evaluation indicators to achieve parameter updates of the embedded physical knowledge neural network and iterative optimization of the finite element simulation input parameters of the hydropower unit component model.

[0008] To achieve the above-mentioned objectives, the technical solution of the present invention is as follows:

[0009] An optimization algorithm for inverting stress cloud diagrams of bolts in hydropower units includes the following steps:

[0010] S1: Finite element simulation model construction, including the following steps:

[0011] S1.1: Determine the basic structure of the component:

[0012] The hydropower unit is decomposed into a fluid computational domain and a structural computational domain. Three-dimensional modeling software is used to perform three-dimensional modeling of the fluid computational domain and the top cover bolt structural computational domain of the hydropower unit.

[0013] The fluid computation domain includes the volute, fixed guide vanes and movable guide vane area, impeller area and tailrace pipe; the structural computation domain includes the outer top cover flange, inner top cover and bolts.

[0014] S1.2: Finite element simulation model establishment and calculation: After meshing the three-dimensional models of the fluid calculation domain and the top cover bolt structure calculation domain in step S1.1, a finite element simulation model is established. In the finite element simulation model, the contact constraints of each component and the boundary conditions of the fluid calculation domain are set, and multiple sets of different working condition loads are input into the finite element simulation model to establish a hydropower unit finite element simulation model. Fluid-solid coupling simulation calculations are carried out to obtain physical field simulation data under different working condition loads.

[0015] The meshing is performed by using unstructured tetrahedral meshes to divide the entire flow channel of the hydropower unit, with the meshes for the top cover flow surface area, impeller, and guide vanes being fined.

[0016] The boundary conditions of the fluid computational domain include the volute inlet (using pressure inlet boundary conditions, given the total working head of the hydropower unit), the wall (using no-slip wall boundary conditions), the tailrace outlet (using mass flow boundary conditions), and the runner region (since the runner is a rotating component, this region uses the rotating reference coordinate system method, given the speed of the runner, and uses a frozen rotor at the dynamic-static interface).

[0017] The contact constraints are the horizontal contact between the inner and outer top cover flanges, the horizontal contact between the bolt head and the inner top cover flange, and the contact between the bolt bottom thread and the inner top cover threaded hole.

[0018] Different working condition loads include: multiple sets of working condition loads with different load locations, different load intensities, and different load application methods.

[0019] The fluid-structure interaction (FSI) simulation calculation includes applying the gravity load of the turbine based on the gravitational acceleration, applying the preload load of the bolts and locking the bolt preload force, applying water pressure to the top cover and water guide bearing seat, that is, the water pressure load calculated by the external fluid is transferred to the top cover and water guide bearing seat through the fluid-structure interaction surface, and solving the fluid-structure interaction equation through fluid-structure interaction simulation according to the working load and boundary conditions to obtain the physical field simulation data under different working conditions.

[0020] The establishment of the finite element simulation model also includes mesh independence verification. Mesh independence verification is carried out by increasing the number of meshes on the existing finite element simulation model. When the change in the physical field does not exceed 5% as the number of meshes increases, the verification is completed. When the change in the physical field exceeds 5% as the number of meshes increases, the number of meshes is increased again for independence verification until the verification is completed.

[0021] S2: Perform data cleaning on the physical field simulation data and field measurement data obtained in step S1.2 to obtain the physical field dataset;

[0022] Data cleaning includes the following steps:

[0023] S2.1: Handling Missing Values: Detect missing values ​​in physical field simulation data and field measurement data, and handle missing values ​​by filling in missing values ​​or deleting rows or columns containing missing values.

[0024] S2.2: Handling outliers: Detect outliers in physical field simulation data and field measurement data, and handle outliers using methods including substitution, deletion, data normalization, and data verification.

[0025] S2.3: Convert data types: Based on the analysis objectives or modeling requirements, determine the data type required for each field, and adjust the types of physical field simulation data and field measurement data accordingly, including the conversion of numerical data, categorical data, etc.

[0026] S2.4: Standardized Data: The physical field simulation data and field measurement data are standardized to ensure that each set of data has the same dimensions and distribution, matching the input dimensions of the reinforcement learning algorithm, and thus obtaining a physical field dataset; this facilitates the construction of the subsequent service stress cloud map inversion optimization algorithm.

[0027] S3: Use the physical field dataset obtained in step S2 to establish a stress cloud map inversion model for hydropower unit bolt reinforcement learning, and iteratively update the stress cloud map inversion model for hydropower unit bolt reinforcement learning. The effect of finite element stress cloud map inversion is verified by comparing the running results of actual working conditions and finite element simulation with the inversion results.

[0028] The establishment of the reinforcement learning stress cloud map inversion model for hydropower unit bolts includes: constructing a Markov decision chain model from the physical field dataset obtained in step S2, establishing a reinforcement learning model, applying the reinforcement learning model to solve the Markov decision chain model, and coupling the reinforcement learning model with the Markov decision chain model to obtain the reinforcement learning stress cloud map inversion model for hydropower unit bolts.

[0029] The finite element input parameters of the physical field dataset are used as the input of the reinforcement learning stress cloud map inversion model of hydropower unit bolts to invert the stress cloud map of hydropower unit bolts. The stress cloud map of hydropower unit bolts is the stress value of each grid point after the finite element simulation model is meshed, and the stress value of the boundary of the stress cloud map is the boundary stress value.

[0030] Markov decision chain models include states s t Action a t Markov decision chain model feedback r ( s t );

[0031] The state s t The stress inversion results for initializing the Markov decision chain model are used in each action a. t The subsequent stress inversion results are s t+1 Action a t The step size for changing the stress magnitude is used to change the state of each iteration. s t ;

[0032] Using the Navier-Stokes equations and the solid-wall boundary condition (i.e., the fluid velocity is zero at the solid boundary), the stress boundary values ​​of the theoretical physical field are calculated from the fluid physical parameters in the hydroelectric generator unit. The general form of the Navier-Stokes equations is:

[0033]

[0034] Where ρ is the fluid density, v is the velocity field, p is the pressure field, μ is the fluid viscosity coefficient, and f is the external force, such as gravity.

[0035] For the bolts of the hydroelectric generator unit, establish a solid wall boundary condition, set the velocity at the boundary to zero, solve the equation, and obtain the current state. s t The stress boundary value;

[0036] Markov decision chain model feedback For the current state s t The mean square error between the stress value and the stress value in the physical field dataset, plus the difference between the calculated theoretical stress contour map boundary value and the boundary stress value inverted from the stress contour map inversion model, is used to measure the current predicted stress, i.e., state s. t The difference between the actual dataset and the real dataset; where the iterative formula for the state of the Markov decision chain model in each cycle is as follows:

[0037] .

[0038] The reinforcement learning model includes a policy network I (participant) and a value network I (evaluator).

[0039] The policy network I is responsible for generating decisions, that is, deciding the direction of change of the finite element input parameters (whether to increase or decrease); the value network I is responsible for evaluating the quality of the decisions output by the policy network I, that is, calculating the error based on the predicted value of the finite element simulation results and the actual value of the simulation results in the dataset. The smaller the error, the better the decision.

[0040] The policy network I consists of a radial basis function neural network, including an input layer, two hidden layers and an output layer, which generates corresponding control policies by gradually accumulating system experience.

[0041] The value network I part consists of an input layer, a hidden layer, and an output layer. It evaluates the performance of the current finite element parameter decision by comparing the difference between the physical field simulation data of the finite element simulation results and the actual test data. It generates a reward (error decreases) or a penalty (error increases) as the feedback value for adaptive learning based on the error change trend, and introduces a long-term cost function as the evaluation criterion.

[0042] The input and output dimensions (data format) of the policy network I and the value network I are determined by the number of input parameters of the finite element simulation and the number of nodes in the extracted finite element simulation physics field.

[0043] The iterative update includes: updating the policy network I using the stochastic gradient descent algorithm, correcting the state value of the Markov decision chain model using the Bellman iterative equation, and performing simulation applications and verification to achieve the inversion prediction of the stress cloud map.

[0044] The update of policy network I using stochastic gradient descent algorithm includes updating policy network I using the gradient method based on embedded physical knowledge policy network I through stochastic gradient descent (how to update is in the next paragraph), and ensuring that the finite element parameters generated by policy network I can maximize the improvement of the evaluation index of Markov decision chain model through the evaluation of value network I.

[0045] The stochastic gradient descent algorithm is used to train the objective function of policy network I, measure the performance of the current policy, and guide the updating of network parameters. The stochastic gradient descent algorithm consists of two parts: the first part is the loss function composed of the difference between the finite element simulation result and the actual measurement value corresponding to the sensor, and the second part is the difference between the field boundary conditions predicted by policy network I and the ideal boundary conditions. The predicted boundary conditions are the velocity and pressure on the fluid boundary in the simulation result, and the ideal boundary conditions are the velocity and pressure on the fluid boundary that conform to the basic equation of fluid-structure interaction mentioned above.

[0046] The update of policy network I and value network I using the stochastic gradient descent algorithm includes: first, randomly generating the initial state of the target Markov decision chain model or the current state s after iterative updates. t The policy π is generated based on the policy network I and initialized, according to the current state s of the Markov decision chain model. t To determine the action to be performed, a t , will action a t Interact with the Markov decision chain model and transition to a new state s according to the iterative formula of the Markov decision chain model's state. t+1 Obtain feedback from the Markov decision chain model Value Network I estimates the value of actions through feedback from a Markov decision chain model. , The expression is:

[0047]

[0048] Where S represents the state of the Markov decision chain model. G t Feedback for Markov decision chain model The expected value of the mathematical expression;

[0049] Each state-action group (s) can be calculated using the value function. t a t The advantage function A π :

[0050]

[0051] The gradient of policy network I is calculated based on the advantage function, and then policy network I is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network I on the state of the Markov decision chain model. The iteration of value network I requires the correction of the state value of the Markov decision chain model based on the Bellman iteration equation, and the calculation of the parameter gradients before and after the correction, and the updating of value network I according to the stochastic gradient descent algorithm.

[0052] The gradient of policy network I is calculated as follows:

[0053]

[0054] To define the network parameters of policy network I, In a given state s t Next, action a t The log probability of the corresponding strategy; This is the dominant function.

[0055] The iterative formula for the value function is as follows:

[0056]

[0057] in, The attenuation coefficient is... Feedback for Markov decision chain models.

[0058] The simulation application and verification include: the reinforcement learning algorithm interacts with the Markov decision chain model in each iteration cycle, and the policy network I is based on the state of the Markov decision chain model. s t Generate decision action a t The decision action a t To make decisions about the direction of change of finite element input parameters; the Markov decision chain model is based on decision action a. t Feedback is generated, and the state of the Markov decision chain model changes from... s t Transformation to s t+1 Value network I is responsible for evaluating the quality of the decisions output by policy network I. This evaluation is based on feedback from a Markov decision chain model. In other words, the difference between the predicted value of the finite element simulation result and the actual value of the simulation result in the dataset, plus the difference between the theoretical stress cloud map boundary value calculated based on the physical formula and the inverted stress boundary value, the sum of the two differences is the error. The smaller the error, the better the decision. When the error is greater than 5%, the stress cloud map inversion optimization algorithm iteration continues. When the error is less than 5%, the stress cloud map inversion optimization algorithm iteration is completed.

[0059] S4: Optimal parameter search, including using the stress cloud map inversion model of hydropower unit bolt reinforcement learning obtained in step S3, establishing a hydropower unit bolt reinforcement learning working condition parameter search model, performing optimal working condition parameter search, verifying and visualizing the search results, and obtaining the hydropower unit bolt service stress cloud map inversion optimization algorithm.

[0060] Step S4 includes the following steps:

[0061] S4.1: Utilizing the reinforcement learning stress cloud map inversion model of hydropower unit bolts, a search model for hydropower unit operating parameters is constructed, including:

[0062] By obtaining the operating environment of the hydropower unit under different operating conditions through historical data or measurements, an appropriate boundary type is selected, and the value of each boundary condition is determined according to the measurement data or standard specifications of the operating conditions. Based on performance indicators or energy efficiency, the objective function in the optimization task is selected, and the stress indicators in the simulation results are mapped to the objective function to ensure that the objective function can truly reflect the stress state of the equipment components and is closely related to the actual operating conditions. Based on the actual situation, the range of each action variable and the constraints of water pressure distribution in the environment are set to construct the hydropower unit operating condition parameter search model.

[0063] Boundary types include fixed boundary conditions and symmetric boundary conditions.

[0064] S4.2: In the hydropower unit operating condition parameter search model obtained in step S4.1, a reinforcement learning environment oriented towards the operating condition design parameters is constructed using the maximum entropy reinforcement learning algorithm, including:

[0065] S4.21: Construct a maximum entropy reinforcement learning algorithm, and introduce an adjustable entropy term into the maximum entropy reinforcement learning algorithm. ;

[0066] The maximum entropy reinforcement learning algorithm consists of a value network II and a policy network II. In each iteration, the policy network II learns from the state of the reinforcement learning environment. s t Formulate actions for it a t And obtain the reward function from the reinforcement learning environment, state s t Transition to new status s t+1 Value Network II evaluates the sum of the reward function and the policy entropy. To assess the current state of the reinforcement learning environment s t Conduct value assessment; based on advantage function The gradient of policy network II is calculated, and policy network II is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network II regarding the system state. The iteration of value network II requires correction of system state value based on Bellman iteration equation, and calculation of parameter gradients before and after correction, and updating value network II according to stochastic gradient descent algorithm. To define the network parameters of Policy Network II, In a given state s t Next, action The log probability of the corresponding strategy; The dominant function;

[0067] S4.22: In the reinforcement learning working condition parameter search model of hydropower unit bolts, the control variables in the finite element simulation environment are analyzed. By mapping these control variables to actions in reinforcement learning, the action space of reinforcement learning is determined, so that the policy network II can adjust these variables for optimization under different working conditions.

[0068] S4.23: Determine the state space in the reinforcement learning environment. Considering the state changes caused by actions, the policy features can be defined as the stress distribution of the bolts. The inherent features that do not change in the environment are the size of the constructed three-dimensional hydroelectric generator model and the fluid-structure interaction calculation form.

[0069] S4.24: Determine the environmental constraints and design the reinforcement learning reward function based on the objective function; the reward function guides the policy network II to learn and select the optimal operating parameters by providing positive feedback for actions that approach the optimal parameters, ensuring that the reward function reflects the optimization direction of the objective function;

[0070] S4.3: Obtaining Optimal Operating Parameters: In a reinforcement learning environment, the operating parameter search model utilizes the maximum entropy reinforcement learning algorithm to perform online search and optimization of operating parameters. It iterates through the optimal reinforcement learning strategy based on maximum entropy reinforcement learning, and searches for the optimal operating parameters based on the optimal strategy; including:

[0071] Based on the reinforcement learning environment established in S4.2, the reinforcement learning environment is solved using the maximum entropy reinforcement learning algorithm;

[0072] From cumulative rewards and policy entropy The two components form the evaluation function of the value function, and the optimal strategy is calculated through the evaluation function of the value function. The optimal strategy is then imported into the strategy network II to output the optimal operating parameters.

[0073] The evaluation function of the value function is:

[0074]

[0075] in, State Action The group has a probability distribution of Expected reward value at that time; s t and a t These represent the state and action at time t, respectively. In the state s t Next action a t The reward value obtained; For state st The entropy value of the time strategy π; α Temperature parameters determine its ability to explore different strategies; π For control strategies; π * This is the optimal control strategy.

[0076] S4.4: Verify the optimal operating parameters obtained from the search, including:

[0077] Finite element simulations were performed using the optimal working parameters obtained from the search. The differences between the physical field simulation data in the finite element simulation results and the physical field simulation data output by the working parameter search model were compared to verify the feasibility and effectiveness of the search algorithm.

[0078] Multiple parameter points are generated near the optimal search parameters using a sampling algorithm. The optimal working condition parameters obtained from the search, along with the sampling points near them and the corresponding finite element simulation data, are then expanded into a new training set for online iterative updates of the reinforcement learning stress cloud map inversion model.

[0079] The optimal operating parameters were verified on actual equipment, and the actual working performance was observed and measured. The finite element simulation model, stress cloud diagram inversion model, and operating parameter search model were integrated into a digital twin model. The operating parameters before and after optimization were applied to the actual equipment and the finite element simulation model, respectively, and the comparison results before and after optimization were observed to confirm the effectiveness of the optimal operating parameters obtained in the digital twin model.

[0080] S4.5: Display the parameter search results, including:

[0081] The parameter search and iteration process is visualized using icons, including the time and accuracy of each iteration. When the optimal parameters obtained by the search coincide with the physical field under actual working conditions with an accuracy of more than 90%, the iterative training stops, and the inversion optimization algorithm for the service stress cloud map of hydropower unit bolts is obtained.

[0082] The service stress cloud map inversion optimization algorithm also includes verification and error analysis of the physical field simulation results in step S1.2. This verification and error analysis includes: verifying the finite element simulation model in step S1.2; comparing the error between the physical field simulation data predicted by the finite element simulation model and the actual measured values; when the error is greater than 10%, temporarily retaining the set of finite element simulation data; using the finite element simulation data that meets the error conditions to establish an inversion optimization algorithm model; optimizing the load parameters of the current finite element simulation using the inversion optimization algorithm; updating the finite element simulation until the error is less than 10% and the error conditions are met; and adding the set of finite element simulation data to the training set of the inversion optimization algorithm model.

[0083] The field measurement data includes stress in the actual environment. Stress measurement in the actual environment uses a hollow cylindrical force sensor, which includes a built-in elastic element, resistance strain gauge, temperature compensation circuit, and conversion circuit. The sensor measures the average stress on the actual bolt. A single sensor is installed between the top cover bolt and the inner top cover. The field test conditions are the entire start-up and shutdown process of the unit under high head, from shutdown to startup, changing the opening degree, and finally shutting down. The bolt stress distribution measured under actual conditions is compared and analyzed with the finite element simulation results.

[0084] The actual measured value is the average stress on the bolt measured by the hollow cylindrical force sensor; the field test conditions are the entire start-up and shutdown process of the unit under high head, from shutdown to startup, changing the opening degree to finally shutting down.

[0085] The beneficial effects of this invention are:

[0086] 1. The stress cloud map inversion optimization algorithm for hydropower unit bolts in this invention constructs a stress cloud map inversion model based on finite element stress simulation results for different operating conditions of hydropower units, and searches for the optimal operating parameters of hydropower unit bolts based on the constructed stress cloud map inversion model. Compared with other operating parameter search algorithms, this algorithm utilizes a finite element simulation database to establish a stress cloud map inversion model. During the search process, the search algorithm does not need to interact with the lengthy finite element simulation, but interacts with the highly efficient stress cloud map inversion model, greatly improving search efficiency and saving hardware resources such as computing power and time resources. Furthermore, the accuracy of the model is strictly controlled during model training, therefore the operating parameter search results obtained by this algorithm are accurate and reliable.

[0087] 1. The finite element simulation model of hydropower unit bolts in this invention establishes an accurate prediction model of the normal state of power generation equipment components for different operating conditions and their corresponding physical field simulation data, in order to measure the physical field simulation data under various operating conditions. By utilizing the difference between the model's predicted values ​​and the actual measured values, the accuracy of the prediction model is evaluated, accurately depicting the operating state of the power generation equipment components, realizing equipment state simulation, and providing simulation data for cloud map inversion optimization.

[0088] 2. The efficient data cleaning and preprocessing method in this invention develops a high-quality and effective data cleaning method for local finite element simulation test data and actual measurement data of power generation equipment components to ensure high-quality data output and establish highly robust joint data for use as the training and testing basis for core algorithm components.

[0089] 3. This invention employs a reinforcement learning algorithm based on policy gradient optimization and Bellman iteration equations, and guides the convergence of reinforcement learning based on embedded physical knowledge, making the autonomous decision-making of Markov decision chain model parameters more intelligent and accurate. The reinforcement learning algorithm enhances the learning and adaptability of the Markov decision chain model, playing a positive role in improving its performance.

[0090] 4. Optimal Operating Condition Parameter Search Algorithm in this Invention: For prediction models of physical field simulation data under different operating conditions, an automatic optimal operating condition parameter search algorithm is developed using a deep reinforcement learning framework based on intelligent technology. A reasonable objective function for updating the optimal strategy is constructed to achieve a reasonable exploration of the optimal parameters of power generation equipment components, thereby improving the algorithm's ability to explore prediction models. Attached Figure Description

[0091] Figure 1 This is a schematic diagram of the optimization algorithm for inverting stress cloud diagrams of hydropower unit bolts in service, as presented in this invention. Detailed Implementation

[0092] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0093] Example 1

[0094] like Figure 1 As shown in the figure, this embodiment provides an inversion optimization algorithm for stress cloud map of hydropower unit bolts during service, including the following steps:

[0095] S1: Finite element simulation model construction, including the following steps:

[0096] S1.1: Determine the basic structure of the components: Decompose the hydropower unit into a fluid calculation domain and a structural calculation domain, and use 3D modeling software to perform 3D modeling of the fluid calculation domain and the top cover bolt structural calculation domain of the hydropower unit;

[0097] S1.2: Finite element simulation model establishment and calculation: After meshing the three-dimensional models of the fluid calculation domain and the top cover bolt structure calculation domain in step S1.1, a finite element simulation model is established. In the finite element simulation model, the contact constraints of each component and the boundary conditions of the fluid calculation domain are set, and multiple sets of different working condition loads are input into the finite element simulation model to establish a hydropower unit finite element simulation model. Fluid-solid coupling simulation calculation is carried out to obtain physical field simulation data under different working condition loads.

[0098] S2: Perform data cleaning on the physical field simulation data and field measurement data obtained in step S1.2 to obtain the physical field dataset;

[0099] S3: Construction of stress cloud map inversion model: Using the physical field dataset obtained in step S2, a stress cloud map inversion model for hydropower unit bolt reinforcement learning is established, and the stress cloud map inversion model for hydropower unit bolt reinforcement learning is iteratively updated. The effect of finite element stress cloud map inversion is verified by comparing the running results of actual working conditions and finite element simulation with the inversion results.

[0100] S4: Optimal parameter search, including using the stress cloud map inversion model of hydropower unit bolt reinforcement learning obtained in step S3, establishing a hydropower unit bolt reinforcement learning working condition parameter search model, performing optimal working condition parameter search, verifying and visualizing the search results, and obtaining the best working condition parameters for hydropower unit bolts.

[0101] The stress cloud map inversion optimization algorithm for hydropower unit bolts in this embodiment constructs a stress cloud map inversion model based on finite element stress simulation results for different operating conditions of hydropower units. Then, based on the constructed stress cloud map inversion model, it searches for the optimal operating parameters for the hydropower unit bolts. Compared with other operating parameter search algorithms, this algorithm utilizes a finite element simulation database to build the stress cloud map inversion model. During the search process, the search algorithm does not need to interact with the lengthy finite element simulation but interacts with the highly efficient stress cloud map inversion model, greatly improving search efficiency and saving hardware resources such as computing power and time. Furthermore, the accuracy of the model is strictly controlled during model training, therefore, the operating parameter search results obtained by this algorithm are accurate and reliable.

[0102] Example 2

[0103] Compared with Example 1, the difference in this embodiment is that, in this embodiment, step S1 includes the following steps:

[0104] S1.1: Determine the basic structure of the component:

[0105] The hydropower unit is decomposed into a fluid computational domain and a structural computational domain. Three-dimensional modeling software is used to perform three-dimensional modeling of the fluid computational domain and the top cover bolt structural computational domain of the hydropower unit.

[0106] The fluid computation domain includes the volute, fixed guide vanes and movable guide vane area, impeller area and tailrace pipe; the structural computation domain includes the outer top cover flange, inner top cover and bolts.

[0107] S1.2: Finite element simulation model establishment and calculation: After meshing the three-dimensional models of the fluid calculation domain and the top cover bolt structure calculation domain in step S1.1, a finite element simulation model is established. In the finite element simulation model, the contact constraints of each component and the boundary conditions of the fluid calculation domain are set, and multiple sets of different working condition loads are input into the finite element simulation model to establish a hydropower unit finite element simulation model. Fluid-solid coupling simulation calculations are carried out to obtain physical field simulation data under different working condition loads.

[0108] The meshing is performed by using unstructured tetrahedral meshes to divide the entire flow channel of the hydropower unit, with the meshes for the top cover flow surface area, impeller, and guide vanes being fined.

[0109] The boundary conditions of the fluid computational domain include the volute inlet (using pressure inlet boundary conditions, given the total working head of the hydropower unit), the wall (using no-slip wall boundary conditions), the tailrace outlet (using mass flow boundary conditions), and the runner region (since the runner is a rotating component, this region uses the rotating reference coordinate system method, given the speed of the runner, and uses a frozen rotor at the dynamic-static interface).

[0110] The contact constraints are the horizontal contact between the inner and outer top cover flanges, the horizontal contact between the bolt head and the inner top cover flange, and the contact between the bolt bottom thread and the inner top cover threaded hole.

[0111] Different working condition loads include: multiple sets of working condition loads with different load locations, different load intensities, and different load application methods.

[0112] The fluid-structure interaction (FSI) simulation calculation includes applying the gravity load of the turbine based on the gravitational acceleration, applying the preload load of the bolts and locking the bolt preload force, applying water pressure to the top cover and water guide bearing seat, that is, the water pressure load calculated by the external fluid is transferred to the top cover and water guide bearing seat through the fluid-structure interaction surface, and solving the fluid-structure interaction equation through fluid-structure interaction simulation according to the working load and boundary conditions to obtain the physical field simulation data under different working conditions.

[0113] The establishment of the finite element simulation model also includes mesh independence verification. Mesh independence verification involves increasing the number of meshes on the existing finite element simulation model. Verification is complete when the change in the physical field does not exceed 5% as the number of meshes increases. If the change in the physical field exceeds 5% as the number of meshes increases, the number of meshes is increased again for independence verification until verification is complete. The remaining steps are the same as in Example 1.

[0114] In this embodiment, the finite element simulation model of the hydropower unit bolts establishes an accurate prediction model of the normal state of the power generation equipment components for different operating conditions and their corresponding physical field simulation data, in order to measure the physical field simulation data under various operating conditions. By utilizing the difference between the model's predicted values ​​and the actual measured values, the accuracy of the prediction model is evaluated, accurately depicting the operating state of the power generation equipment components, realizing equipment state simulation, and providing simulation data for cloud map inversion optimization.

[0115] Example 3

[0116] Compared with Example 1, the difference in this embodiment is that the data cleaning in step S2 in this embodiment includes the following steps:

[0117] S2.1: Handling Missing Values: Detect missing values ​​in physical field simulation data and field measurement data, and handle missing values ​​by filling in missing values ​​or deleting rows or columns containing missing values.

[0118] S2.2: Handling outliers: Detect outliers in physical field simulation data and field measurement data, and handle outliers using methods including substitution, deletion, data normalization, and data verification.

[0119] S2.3: Convert data types: Based on the analysis objectives or modeling requirements, determine the data type required for each field, and adjust the types of physical field simulation data and field measurement data accordingly, including the conversion of numerical data, categorical data, etc.

[0120] S2.4: Data Standardization: Standardize the physical field simulation data and field measurement data to ensure that each data set has the same dimensions and distribution, matching the input dimensions of the reinforcement learning algorithm, and obtain a physical field dataset; this facilitates the subsequent construction of the service stress cloud map inversion optimization algorithm. The remaining steps are the same as in Example 1.

[0121] This embodiment presents a highly efficient data cleaning and preprocessing method. Targeting local finite element simulation test data and actual measurement data of power generation equipment components, it develops a high-quality and effective data cleaning method to ensure high-quality data output and establish highly robust joint data for use as the training and testing basis for core algorithm components.

[0122] Example 4

[0123] Compared with Example 1, the difference in this embodiment is that step S3 in this embodiment includes the following steps:

[0124] The establishment of the reinforcement learning stress cloud map inversion model for hydropower unit bolts includes: constructing a Markov decision chain model from the physical field dataset obtained in step S2, establishing a reinforcement learning model, applying the reinforcement learning model to solve the Markov decision chain model, and coupling the reinforcement learning model with the Markov decision chain model to obtain the reinforcement learning stress cloud map inversion model for hydropower unit bolts.

[0125] The finite element input parameters of the physical field dataset are used as the input of the reinforcement learning stress cloud map inversion model of hydropower unit bolts to invert the stress cloud map of hydropower unit bolts. The stress cloud map of hydropower unit bolts is the stress value of each grid point after the finite element simulation model is meshed, and the stress value of the boundary of the stress cloud map is the boundary stress value.

[0126] Markov decision chain models include states s t Action a t Markov decision chain model feedbackr ( s t );

[0127] The state s t The stress inversion results for initializing the Markov decision chain model are used in each action a. t The subsequent stress inversion results are s t+1 Action a t The step size for changing the stress magnitude is used to change the state of each iteration. s t；

[0128] Using the Navier-Stokes equations and the solid-wall boundary condition (i.e., the fluid velocity is zero at the solid boundary), the stress boundary values ​​of the theoretical physical field are calculated from the fluid physical parameters in the hydroelectric generator unit. The general form of the Navier-Stokes equations is:

[0129]

[0130] Where ρ is the fluid density, v is the velocity field, p is the pressure field, μ is the fluid viscosity coefficient, and f is the external force, such as gravity.

[0131] For the bolts of the hydroelectric generator unit, establish a solid wall boundary condition, set the velocity at the boundary to zero, solve the equation, and obtain the current state. The stress boundary value;

[0132] Markov decision chain model feedback For the current state s t The mean square error between the stress value and the stress value in the physical field dataset, plus the difference between the calculated theoretical stress contour map boundary value and the boundary stress value inverted from the stress contour map inversion model, is used to measure the current predicted stress, i.e., state s. t The difference between the actual dataset and the real dataset; where the iterative formula for the state of the Markov decision chain model in each cycle is as follows:

[0133] .

[0134] The reinforcement learning model includes a policy network I (participant) and a value network I (evaluator).

[0135] The policy network I is responsible for generating decisions, that is, deciding the direction of change of the finite element input parameters (whether to increase or decrease); the value network I is responsible for evaluating the quality of the decisions output by the policy network I, that is, calculating the error based on the predicted value of the finite element simulation results and the actual value of the simulation results in the dataset. The smaller the error, the better the decision.

[0136] The policy network I consists of a radial basis function neural network, including an input layer, two hidden layers and an output layer, which generates corresponding control policies by gradually accumulating system experience.

[0137] The value network I part consists of an input layer, a hidden layer, and an output layer. It evaluates the performance of the current finite element parameter decision by comparing the difference between the physical field simulation data of the finite element simulation results and the actual test data. It generates a reward (error decreases) or a penalty (error increases) as the feedback value for adaptive learning based on the error change trend, and introduces a long-term cost function as the evaluation criterion.

[0138] The input and output dimensions (data format) of the policy network I and the value network I are determined by the number of input parameters of the finite element simulation and the number of nodes in the extracted finite element simulation physics field.

[0139] The iterative update includes: updating the policy network I using the stochastic gradient descent algorithm, correcting the state value of the Markov decision chain model using the Bellman iterative equation, and performing simulation applications and verification to achieve the inversion prediction of the stress cloud map.

[0140] The update of policy network I using stochastic gradient descent algorithm includes updating policy network I using the gradient method based on embedded physical knowledge policy network I through stochastic gradient descent (how to update is in the next paragraph), and ensuring that the finite element parameters generated by policy network I can maximize the improvement of the evaluation index of Markov decision chain model through the evaluation of value network I.

[0141] The stochastic gradient descent algorithm is used to train the objective function of policy network I, measure the performance of the current policy, and guide the updating of network parameters. The stochastic gradient descent algorithm consists of two parts: the first part is the loss function composed of the difference between the finite element simulation result and the actual measurement value corresponding to the sensor, and the second part is the difference between the field boundary conditions predicted by policy network I and the ideal boundary conditions. The predicted boundary conditions are the velocity and pressure on the fluid boundary in the simulation result, and the ideal boundary conditions are the velocity and pressure on the fluid boundary that conform to the basic equation of fluid-structure interaction mentioned above.

[0142] The update of policy network I and value network I using the stochastic gradient descent algorithm includes: first, randomly generating the initial state of the target Markov decision chain model or the current state s after iterative updates. t The policy π is generated based on the policy network I and initialized, according to the current state s of the Markov decision chain model. t To determine the action to be performed, a t , will action a tInteract with the Markov decision chain model and transition to a new state s according to the iterative formula of the Markov decision chain model's state. t+1 Obtain feedback from the Markov decision chain model Value Network I estimates the value of actions through feedback from a Markov decision chain model. , The expression is:

[0143]

[0144] Where S represents the state of the Markov decision chain model. G t Feedback for Markov decision chain model The expected value of the mathematical expression;

[0145] Each state-action group (s) can be calculated using the value function. t a t The advantage function A π :

[0146]

[0147] The gradient of policy network I is calculated based on the advantage function, and then policy network I is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network I on the state of the Markov decision chain model. The iteration of value network I requires the correction of the state value of the Markov decision chain model based on the Bellman iteration equation, and the calculation of the parameter gradients before and after the correction, and the updating of value network I according to the stochastic gradient descent algorithm.

[0148] The gradient of policy network I is calculated as follows:

[0149]

[0150] To define the network parameters of policy network I, In a given state s t Next, action a t The log probability of the corresponding strategy; This is the dominant function.

[0151] The iterative formula for the value function is as follows:

[0152]

[0153] in, The attenuation coefficient is... Feedback for Markov decision chain models.

[0154] The simulation application and verification include: the reinforcement learning algorithm interacts with the Markov decision chain model in each iteration cycle, and the policy network I is based on the state of the Markov decision chain model. s t Generate decision action a t The decision action a t To make decisions about the direction of change of finite element input parameters; the Markov decision chain model is based on decision action a. t Feedback is generated, and the state of the Markov decision chain model changes from... s t Transformation to s t+1 Value network I is responsible for evaluating the quality of the decisions output by policy network I. This evaluation is based on feedback from a Markov decision chain model. In other words, the difference between the predicted value from the finite element simulation and the actual value from the simulation in the dataset, plus the difference between the theoretical stress contour map boundary value calculated based on the physical formula and the inverted stress boundary value, is the error. The smaller the error, the better the decision. When the error is greater than 5%, the stress contour map inversion optimization algorithm iteration continues; when the error is less than 5%, the stress contour map inversion optimization algorithm iteration is complete. The remaining steps are the same as in Example 1.

[0155] In this embodiment, a reinforcement learning algorithm based on policy gradient optimization and Bellman iteration equations is employed. Embedded physics knowledge guides the convergence of the reinforcement learning process, making the system's parameter autonomous decision-making more intelligent and accurate. The reinforcement learning algorithm enhances the system's learning and adaptive capabilities, positively impacting system performance.

[0156] Example 5

[0157] Compared with Example 1, the difference in this embodiment is that step S4 in this embodiment includes the following steps:

[0158] S4.1: Utilizing the reinforcement learning stress cloud map inversion model of hydropower unit bolts, a search model for hydropower unit operating parameters is constructed, including:

[0159] By obtaining the operating environment of the hydropower unit under different operating conditions through historical data or measurements, an appropriate boundary type is selected, and the value of each boundary condition is determined according to the measurement data or standard specifications of the operating conditions. Based on performance indicators or energy efficiency, the objective function in the optimization task is selected, and the stress indicators in the simulation results are mapped to the objective function to ensure that the objective function can truly reflect the stress state of the equipment components and is closely related to the actual operating conditions. Based on the actual situation, the range of each action variable and the constraints of water pressure distribution in the environment are set to construct the hydropower unit operating condition parameter search model.

[0160] Boundary types include fixed boundary conditions and symmetric boundary conditions.

[0161] S4.2: In the hydropower unit operating condition parameter search model obtained in step S4.1, a reinforcement learning environment oriented towards the operating condition design parameters is constructed using the maximum entropy reinforcement learning algorithm, including:

[0162] S4.21: Construct a maximum entropy reinforcement learning algorithm, and introduce an adjustable entropy term into the maximum entropy reinforcement learning algorithm. ;

[0163] The maximum entropy reinforcement learning algorithm consists of a value network II and a policy network II. In each iteration, the policy network II learns from the state of the reinforcement learning environment. s t Formulate actions for it a t And obtain the reward function from the reinforcement learning environment, state s t Transition to new status s t+1 Value Network II evaluates the sum of the reward function and the policy entropy. To assess the current state of the reinforcement learning environment s t Conduct value assessment; based on advantage function The gradient of policy network II is calculated, and policy network II is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network II regarding the system state. The iteration of value network II requires correction of system state value based on Bellman iteration equation, and calculation of parameter gradients before and after correction, and updating value network II according to stochastic gradient descent algorithm. To define the network parameters of Policy Network II, In a given state s t Next, action The log probability of the corresponding strategy; The dominant function;

[0164] S4.22: In the reinforcement learning working condition parameter search model of hydropower unit bolts, the control variables in the finite element simulation environment are analyzed. By mapping these control variables to actions in reinforcement learning, the action space of reinforcement learning is determined, so that the policy network II can adjust these variables for optimization under different working conditions.

[0165] S4.23: Determine the state space in the reinforcement learning environment. Considering the state changes caused by actions, the policy features can be defined as the stress distribution of the bolts. The inherent features that do not change in the environment are the size of the constructed three-dimensional hydroelectric generator model and the fluid-structure interaction calculation form.

[0166] S4.24: Determine the environmental constraints and design the reinforcement learning reward function based on the objective function. The reward function guides the policy network II to learn and select the optimal operating parameters by providing positive feedback for actions that approach the optimal parameters, ensuring that the reward function reflects the optimization direction of the objective function.

[0167] S4.3: Obtaining Optimal Operating Parameters: In a reinforcement learning environment, the operating parameter search model utilizes the maximum entropy reinforcement learning algorithm to perform online search and optimization of operating parameters. It iterates through the optimal reinforcement learning strategy based on maximum entropy reinforcement learning, and searches for the optimal operating parameters based on the optimal strategy; including:

[0168] Based on the reinforcement learning environment established in S4.2, the reinforcement learning environment is solved using the maximum entropy reinforcement learning algorithm;

[0169] From cumulative rewards and policy entropy The two components form the evaluation function of the value function, and the optimal strategy is calculated through the evaluation function of the value function. The optimal strategy is then imported into the strategy network II to output the optimal operating parameters.

[0170] The evaluation function of the value function is:

[0171]

[0172] in, State Action The group has a probability distribution of Expected reward value at that time; s t and a t These represent the state and action at time t, respectively. In the state s t Next action a t The reward value obtained; For state s t The entropy value of the time strategy π; α Temperature parameters determine its ability to explore different strategies; π For control strategies; π * This is the optimal control strategy.

[0173] S4.4 Verify the optimal operating parameters obtained from the search, including:

[0174] Finite element simulations were performed using the optimal working parameters obtained from the search. The differences between the physical field simulation data in the finite element simulation results and the physical field simulation data output by the working parameter search model were compared to verify the feasibility and effectiveness of the search algorithm.

[0175] Multiple parameter points are generated near the optimal search parameters using a sampling algorithm. The optimal working condition parameters obtained from the search, along with the sampling points near them and the corresponding finite element simulation data, are then expanded into a new training set for online iterative updates of the reinforcement learning stress cloud map inversion model.

[0176] The optimal operating parameters were verified on actual equipment, and the actual working performance was observed and measured. The finite element simulation model, stress cloud diagram inversion model, and operating parameter search model were integrated into a digital twin model. The operating parameters before and after optimization were applied to the actual equipment and the finite element simulation model, respectively, and the comparison results before and after optimization were observed to confirm the effectiveness of the optimal operating parameters obtained in the digital twin model.

[0177] S4.5 Display the parameter search results, including:

[0178] The parameter search and iteration process is visualized using icons, including the time and accuracy of each iteration. When the optimal parameters obtained by the search achieve a coincidence accuracy of over 90% with the physical field under actual working conditions, the iterative training stops, resulting in the optimization algorithm for inverting the stress cloud map of the hydropower unit bolts. The remaining steps are the same as in Example 1.

[0179] The optimal operating condition parameter search algorithm in this embodiment is developed using a deep reinforcement learning framework based on intelligent technology to automatically search for optimal operating condition parameters for prediction models of physical field simulation data under different operating conditions. A reasonable objective function for updating the optimal strategy is constructed to achieve a reasonable exploration of the optimal parameters of power generation equipment components, thereby improving the algorithm's ability to explore prediction models.

[0180] Example 6

[0181] Compared with Example 1, the difference in this embodiment is that, in this embodiment, the service stress cloud map inversion optimization algorithm further includes verification and error analysis of the physical field simulation results in step S1.2. The verification and error analysis of the physical field simulation results includes: verifying the finite element simulation model in step S1.2, comparing the error between the physical field simulation data predicted by the finite element simulation model and the actual measured values; when the error is greater than 10%, temporarily retaining the set of finite element simulation data, and using the finite element simulation data that meets the error conditions to establish an inversion optimization algorithm model, and optimizing the working condition parameter load in the current finite element simulation through the inversion optimization algorithm, and updating the finite element simulation until the error is less than 10% and the error conditions are met; adding the set of finite element simulation data to the training set of the inversion optimization algorithm model;

[0182] The field measurement data includes stress in the actual environment. Stress measurement in the actual environment uses a hollow cylindrical force sensor, which includes a built-in elastic element, resistance strain gauge, temperature compensation circuit, and conversion circuit. The sensor measures the average stress on the actual bolt. A single sensor is installed between the top cover bolt and the inner top cover. The field test conditions are the entire start-up and shutdown process of the unit under high head, from shutdown to startup, changing the opening degree, and finally shutting down. The bolt stress distribution measured under actual conditions is compared and analyzed with the finite element simulation results.

[0183] The actual measured value is the average stress on the bolt measured by the hollow cylindrical force sensor; the field test conditions are the entire start-up and shutdown process of the unit under high head, from shutdown to startup, changing the opening degree to finally shutting down.

[0184] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An optimization algorithm for inverting stress cloud diagrams of bolts in hydropower units, characterized in that: Includes the following steps: S1: Finite element simulation model construction, including the following steps: S1.1: Determine the basic structure of the components: Decompose the hydropower unit into a fluid calculation domain and a structural calculation domain, and use 3D modeling software to perform 3D modeling of the fluid calculation domain and the top cover bolt structural calculation domain of the hydropower unit; the fluid calculation domain includes the volute, the fixed guide vane and the movable guide vane area, the impeller area and the tailrace pipe; the structural calculation domain includes the outer top cover flange, the inner top cover and the bolts; S1.2: Finite element simulation model establishment and calculation: After meshing the three-dimensional models of the fluid calculation domain and structural calculation domain in step S1.1, a finite element simulation model is established. In the finite element simulation model, the contact constraints of each component and the boundary conditions of the fluid calculation domain are set, and multiple sets of different working condition loads are input into the finite element simulation model to establish a hydropower unit finite element simulation model. Fluid-solid coupling simulation calculations are carried out to obtain physical field simulation data under different working condition loads. S2: Perform data cleaning on the physical field simulation data and field measurement data obtained in step S1.2 to obtain the physical field dataset; S3: Use the physical field dataset obtained in step S2 to establish a stress cloud map inversion model for hydropower unit bolt reinforcement learning, and iteratively update the stress cloud map inversion model for hydropower unit bolt reinforcement learning. The effect of finite element stress cloud map inversion is verified by comparing the running results of actual working conditions and finite element simulation with the inversion results. S4: Optimal parameter search, including using the stress cloud map inversion model of hydropower unit bolts obtained in step S3 to establish a hydropower unit bolt reinforcement learning working condition parameter search model, performing optimal working condition parameter search, and verifying and visualizing the search results to obtain the optimal working condition parameters for hydropower unit bolts; step S4 specifically includes the following steps: S4.1: Using the stress cloud map inversion model of the hydropower unit bolt reinforcement learning obtained in step S3, construct the hydropower unit operating condition parameter search model; S4.2: In the hydropower unit operating condition parameter search model obtained in step S4.1, the maximum entropy reinforcement learning algorithm is used to construct a reinforcement learning environment for the operating condition design parameters. S4.3: Obtaining the optimal operating parameters: In the reinforcement learning environment, the operating parameter search model uses the maximum entropy reinforcement learning algorithm to complete the online search and optimization of the operating parameters. It iterates the optimal reinforcement learning strategy based on the maximum entropy reinforcement learning and searches for the optimal operating parameters based on the optimal strategy. S4.4: Verify the optimal operating parameters obtained in step S4.3; S4.5: Display the parameter search results, including: the parameter search and iteration process is visualized in the form of icons, including the time and accuracy of each iteration. When the optimal parameters obtained by the search coincide with the physical field under actual working conditions with an accuracy of more than 90%, the iteration training stops, and a service stress cloud map inversion optimization algorithm for hydropower unit bolts is obtained.

2. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: The service stress cloud map inversion optimization algorithm also includes verification and error analysis of the physical field simulation results in step S1.

2. The verification and error analysis of the physical field simulation results includes: verifying the finite element simulation model in step S1.2, comparing the error between the physical field simulation data predicted by the finite element simulation model and the actual measured value; when the error is greater than 10%, temporarily retaining the set of finite element simulation data, and using the finite element simulation data that meets the error condition to establish an inversion optimization algorithm model, and optimizing the working condition parameter load in the current finite element simulation through the inversion optimization algorithm, and updating the finite element simulation until the error is less than 10% and the error condition is met; and adding the set of finite element simulation data to the training set of the inversion optimization algorithm model.

3. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: In step S1.2, the mesh division is performed by using unstructured tetrahedral meshes to divide the entire flow channel of the hydropower unit, and the meshes of the top cover flow surface area, impeller and guide vanes are densified.

4. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: In step S1.2, the boundary conditions of the fluid computation domain include the volute inlet, wall, tailrace outlet, and impeller area. The contact constraints are the horizontal contact between the inner and outer top cover flanges, the horizontal contact between the bolt head and the inner top cover flange, and the contact between the bolt bottom thread and the inner top cover threaded hole.

5. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: In step S1.2, the different working condition loads include: multiple sets of working condition loads with different load positions, different load intensities, and different load application methods.

6. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: In step S1.2, the fluid-structure coupling simulation calculation includes applying the gravity load of the turbine according to the gravitational acceleration, applying the preload load of the bolts and locking the preload force of the bolts, applying water pressure to the top cover and water guide bearing seat, that is, the water pressure load calculated by the external fluid is transmitted to the top cover and water guide bearing seat through the fluid-structure coupling surface, and solving the fluid-structure coupling equation through fluid-structure coupling simulation according to the working load and boundary conditions to obtain the physical field simulation data under different working conditions.

7. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: In step S1.2, the model establishment also includes mesh independence verification. Mesh independence verification is performed by increasing the number of meshes on the basis of the existing finite element simulation model. When the change in the physical field does not exceed 5% as the number of meshes increases, the verification is completed. When the change in the physical field exceeds 5% as the number of meshes increases, the number of meshes is increased again for independence verification until the verification is completed.

8. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: The data cleaning in step S2 includes the following steps: S2.1: Handling missing values: Detect missing values ​​in physical field simulation data and field measurement data, and handle missing values ​​by filling in missing values ​​or deleting rows or columns containing missing values; S2.2: Handling outliers: Detect outliers in physical field simulation data and field measurement data, and handle outliers using methods including substitution, deletion, data normalization, and data verification. S2.3: Convert data types: Determine the required data type for each field based on the analysis objectives or modeling requirements, and adjust the data types in the physical field simulation data and field measurement data accordingly; S2.4: Standardized Data: Standardize the physical field simulation data and field measurement data to make each set of data have the same dimensions and distribution, and obtain the physical field dataset.

9. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: The step S3, establishing the reinforcement learning stress cloud map inversion model for hydropower unit bolts, includes: constructing a Markov decision chain model from the physical field dataset obtained in step S2, establishing a reinforcement learning model, applying the reinforcement learning model to solve the Markov decision chain model, and coupling the reinforcement learning model with the Markov decision chain model to obtain the reinforcement learning stress cloud map inversion model for hydropower unit bolts; using the finite element input parameters of the physical field dataset as input to the reinforcement learning stress cloud map inversion model for hydropower unit bolts, and inverting the stress cloud map of the hydropower unit bolts, wherein the stress cloud map of the hydropower unit bolts is the stress value of each grid point after meshing of the finite element simulation model, and the stress value of the boundary of the stress cloud map is the boundary stress value; the Markov decision chain model includes states. s t Action a t Markov decision chain model feedback r ( s t ); where state s t The stress inversion results for initializing the Markov decision chain model are used in each action a. t The subsequent stress inversion result is s t+1 Action a t The step size for changing the stress magnitude is used to change the state s of each iteration. t By applying the Navier-Stokes equations and solid-wall boundary conditions, the stress boundary values ​​of the theoretical physical field are calculated using fluid physical parameters within the hydroelectric generator unit; a Markov decision chain model is used for feedback. r ( s t ) represents the current state s t The mean square error between the stress value and the stress value in the physical field dataset, plus the sum of the difference between the calculated theoretical stress contour map boundary value and the boundary stress value inverted from the stress contour map inversion model, is used to measure the current predicted stress, i.e., state s. t The difference between the actual dataset and the real dataset; where the iterative formula for the state of the Markov decision chain model in each cycle is as follows: 。 10. The inversion optimization algorithm for stress cloud map of hydropower unit bolts according to claim 9, characterized in that: In step S3, the reinforcement learning model includes a policy network I and a value network I.

11. The inversion optimization algorithm for stress cloud map of hydropower unit bolts according to claim 10, characterized in that: The strategy network I is responsible for generating decisions, which are decisions about the direction of change of the finite element input parameters; Value Network I is responsible for evaluating the quality of the decisions output by Policy Network I. The evaluation is based on the error calculated between the predicted values ​​from the finite element simulation results and the actual values ​​from the simulation results in the dataset. The smaller the error, the better the decision.

12. The inversion optimization algorithm for stress cloud map of hydropower unit bolts according to claim 10, characterized in that: Policy Network I consists of a radial basis function neural network, including an input layer, two hidden layers and an output layer, which generates corresponding control policies by gradually accumulating system experience.

13. The inversion optimization algorithm for stress cloud map of service bolts in hydropower units according to claim 10, characterized in that: The value network I part consists of an input layer, a hidden layer, and an output layer. It evaluates the performance of the current finite element parameter decision by comparing the difference between the physical field simulation data of the finite element simulation results and the actual test data. It generates rewards or penalties as feedback values ​​for adaptive learning based on the error change trend, and introduces a long-term cost function as an evaluation criterion.

14. The inversion optimization algorithm for stress cloud map of hydropower unit bolts according to claim 10, characterized in that: The iterative update in step S3 includes: updating the policy network I using the stochastic gradient descent algorithm, correcting the state value of the Markov decision chain model using the Bellman iterative equation, and performing simulation application and verification to achieve inversion prediction of the stress cloud map.

15. The inversion optimization algorithm for stress cloud diagram of service bolts in hydropower units according to claim 14, characterized in that: The step of updating the policy network I using the stochastic gradient descent algorithm includes updating the policy network I using the gradient method based on embedded physical knowledge in the policy network I through stochastic gradient descent, and ensuring that the finite element parameters generated by the policy network I can maximize the improvement of the evaluation index of the Markov decision chain model through the evaluation of the value network I.

16. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 15, characterized in that: The stochastic gradient descent algorithm is used to train the objective function of policy network I, measure the performance of the current policy, and guide the updating of network parameters. The stochastic gradient descent algorithm consists of two parts: the first part is the loss function formed by the difference between the finite element simulation result corresponding to the sensor and the actual measurement value; the second part is the difference between the field boundary conditions predicted by the policy network I and the ideal boundary conditions. The predicted boundary conditions are the velocity and pressure on the fluid boundary in the simulation result, and the ideal boundary conditions are the velocity and pressure on the fluid boundary that conform to the basic equation of fluid-structure interaction mentioned above.

17. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 16, characterized in that: The step of updating the policy network I and value network I using the stochastic gradient descent algorithm includes: first, randomly generating the initial state of the target Markov decision chain model or the current state after iterative updates. s t Based on the policy π generated by initializing the policy network I, and according to the current state of the Markov decision chain model... s t To decide on the action to be performed a t , will the action a t Interact with the Markov decision chain model and transition to a new state according to the iterative formula of the Markov decision chain model's state. Obtain feedback from the Markov decision chain model Value Network I estimates the value of actions through feedback from a Markov decision chain model. , The expression is: in, In strategy The mathematical expectation is given by S, where S is the state of the Markov decision chain model. Each state-action group (s) can be calculated using the value function. t a t The advantage function Its definition is as follows: The gradient of policy network I is calculated based on the advantage function, and then policy network I is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network I on the state of the Markov decision chain model. The iteration of value network I requires the correction of the state value of the Markov decision chain model based on the Bellman iteration equation, and the calculation of the parameter gradients before and after the correction, and the updating of value network I according to the stochastic gradient descent algorithm. The gradient of policy network I is calculated as follows: in, To define the network parameters of policy network I, In a given state s t Next, action a t The log probability of the corresponding strategy; The dominant function; The iterative formula for the value function is as follows: in, The attenuation coefficient is... Feedback for Markov decision chain models.

18. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 14, characterized in that: The simulation application and verification include: the reinforcement learning algorithm interacts with the Markov decision chain model in each iteration cycle, and the policy network I is based on the state of the Markov decision chain model. s t Generate decision action a t The decision action a t To make decisions about the direction of change of finite element input parameters; the Markov decision chain model is based on decision action a. t Feedback is generated, and the state of the Markov decision chain model changes from... s t Transformation to s t+1 Value network I is responsible for evaluating the quality of the decisions output by policy network I. This evaluation is based on feedback from a Markov decision chain model. In other words, the difference between the predicted value of the finite element simulation result and the actual value of the simulation result in the dataset, plus the difference between the theoretical stress cloud map boundary value calculated based on the physical formula and the inverted stress boundary value, the sum of the two differences is the error. The smaller the error, the better the decision. When the error is greater than 5%, the stress cloud map inversion optimization algorithm iteration continues. When the error is less than 5%, the stress cloud map inversion optimization algorithm iteration is completed.

19. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: Step S4.1 includes: obtaining the operating environment of the hydropower unit under different operating conditions through historical data or measurements; selecting appropriate boundary types, including fixed boundary conditions and symmetrical boundary conditions; determining the value of each boundary condition based on the measurement data or standard specifications of the operating conditions; selecting the objective function in the optimization task based on performance indicators or energy efficiency; mapping the stress indicators in the simulation results to the objective function to ensure that the objective function can truly reflect the stress state of the equipment components and is closely related to the actual operating conditions; and setting the range of each action variable and the constraints of water pressure distribution in the environment based on the actual situation to construct a hydropower unit operating condition parameter search model.

20. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: Step S4.2 includes: S4.21: Construct a maximum entropy reinforcement learning algorithm, and introduce an adjustable entropy term into the maximum entropy reinforcement learning algorithm. ; The maximum entropy reinforcement learning algorithm consists of a value network II and a policy network II. In each iteration, the policy network II learns from the state of the reinforcement learning environment. s t Formulate actions for it a t And obtain the reward function from the reinforcement learning environment, state s t Transition to new status s t+1 Value Network II evaluates the sum of the reward function and the policy entropy. To assess the current state of the reinforcement learning environment s t Conduct a value assessment; among which, Temperature parameter; based on dominance function The gradient of policy network II is calculated, and policy network II is updated according to the stochastic gradient descent algorithm. This corrects the value judgment and policy formulation process of policy network II regarding the system state. The iteration of value network II requires correction of system state value based on Bellman iteration equation, and calculation of parameter gradients before and after correction, and updating value network II according to stochastic gradient descent algorithm. To define the network parameters of Policy Network II, In a given state s t Next, action The log probability of the corresponding strategy; For state-action pairs under policy 𝜋 The dominant function; S4.22: In the reinforcement learning working condition parameter search model of hydropower unit bolts, the control variables in the finite element simulation environment are analyzed. By mapping these control variables to actions in reinforcement learning, the action space of reinforcement learning is determined, so that the policy network II can adjust these variables for optimization under different working conditions. S4.23: Determine the state space in the reinforcement learning environment. Considering the state changes caused by actions, the policy features can be defined as the stress distribution of the bolts. The inherent features that do not change in the environment are the size of the constructed three-dimensional hydroelectric generator model and the fluid-structure interaction calculation form. S4.24: Determine the environmental constraints and design the reinforcement learning reward function based on the objective function. The reward function guides the policy network II to learn and select the optimal operating parameters by providing positive feedback for actions that approach the optimal parameters, ensuring that the reward function reflects the optimization direction of the objective function.

21. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: Step S4.3 includes: based on the reinforcement learning environment established in S4.2, solving the reinforcement learning environment using the maximum entropy reinforcement learning algorithm; From cumulative rewards and policy entropy The two components form the evaluation function of the value function, and the optimal strategy is calculated through the evaluation function of the value function. The optimal strategy is then imported into the strategy network II to output the optimal operating parameters. The evaluation function of the value function is: in, State Action The group has a probability distribution of Expected reward value at that time; s t and a t These represent the state and action at time t, respectively. In the state s t Next action a t The reward value obtained; For state s t The entropy value of the time strategy π; α For temperature parameters; π For control strategies; π * This is the optimal control strategy.

22. The inversion optimization algorithm for stress cloud diagram of hydropower unit bolts according to claim 1, characterized in that: Step S4.4 includes: Finite element simulations were performed using the optimal working parameters obtained from the search. The differences between the physical field simulation data in the finite element simulation results and the physical field simulation data output by the working parameter search model were compared to verify the feasibility and effectiveness of the search algorithm. Multiple parameter points are generated near the optimal search parameters using a sampling algorithm. The optimal working condition parameters obtained from the search, along with the sampling points near them and the corresponding finite element simulation data, are then expanded into a new training set for online iterative updates of the reinforcement learning stress cloud map inversion model. The optimal operating parameters were verified on actual equipment, and the actual working performance was observed and measured. The finite element simulation model, stress cloud diagram inversion model, and operating parameter search model were integrated into a digital twin model. The operating parameters before and after optimization were applied to the actual equipment and the finite element simulation model, respectively, and the comparison results before and after optimization were observed to confirm the effectiveness of the optimal operating parameters obtained in the digital twin model.

Images