Three-dimensional flow field load reconstruction method based on reinforcement learning of gated circulation unit

Abstract

The invention relates to the technical field of hydroelectric generators, in particular to a three-dimensional flow field load reconstruction method based on reinforcement learning of a gated circulation unit, which comprises the following steps of: establishing a Python-based finite element secondary development platform, and realizing finite element simulation and obtaining a finite element simulation result; fusing a plurality of simulation characteristic indexes extracted from the finite element simulation result by using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation index, and optimizing simulation parameters; constructing an initial three-dimensional flow field load reconstruction model based on reinforcement learning of a gated circulation unit, and training the initial three-dimensional flow field load reconstruction model by using the optimized simulation parameters; and based on the trained three-dimensional flow field load reconstruction model, obtaining a reconstructed three-dimensional flow field load. By means of the reconstruction method, automatic simulation of finite element software and full-automatic optimization iteration of a parameter decision and reinforcement learning reconstruction network in the three-dimensional flow field load reconstruction process of the hydroelectric generating set component can be achieved.

Description

Technical Field

[0001] The invention relates to the technical field of hydroelectric generators, and in particular to a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning. Background Art

[0002] There are currently two main ideas for the reconstruction of three-dimensional flow fields in hydroelectric generator technology: three-dimensional flow field reconstruction based on finite elements and heuristic algorithms, and physical field reconstruction and optimization based on deep learning technology.

[0003] Three-dimensional flow field load reconstruction method based on finite element and heuristic algorithm: Finite element simulation software provides the function of finite element analysis of engineering structures. Users can implement finite element analysis of engineering structures and obtain simulation results by manually inputting parameters, but there is a lack of automatic simulation process. Users can only rely on the experience of engineers to manually adjust parameters and conduct repeated experiments, so the efficiency is low.

[0004] Three-dimensional flow field reconstruction and optimization based on deep learning technology: Existing optimization tools use deep learning technology to learn historical physical field data and experimental data through methods such as feature fusion and end-to-end mapping to obtain a trained deep learning model, thereby predicting the physical field distribution under new experimental data. This type of method relies on deep learning model parameters that are mostly based on experience or require manual repeated adjustment, lacking intelligent adjustment of specific simulation models, and poor generalization ability. In addition, this type of deep learning technology lacks interpretability, which is not conducive to engineers' understanding and optimization of the model.

[0005] In general, the problems existing in the existing technical solutions are mainly: the three-dimensional flow field reconstruction method of the hydraulic generator lacks universality, the optimization process requires more manual intervention, and the efficiency is low. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, which can realize automatic simulation of finite elements and fully automatic optimization iteration of parameter decision and reinforcement learning reconstruction network in the process of three-dimensional flow field load reconstruction of hydropower unit components.

[0007] The present invention is achieved by adopting the following technical solutions:

[0008] The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning includes the following steps:

[0009] Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results;

[0010] Step S2. extracting multiple simulation feature indicators from the finite element simulation results, fusing the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimizing the simulation parameters based on the finite element simulation result evaluation indicator;

[0011] Step S3. construct an initial three-dimensional flow field load reconstruction model based on gated recurrent unit reinforcement learning, and use the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model until a trained three-dimensional flow field load reconstruction model is obtained;

[0012] Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

[0013] The step S1 specifically includes the following steps:

[0014] Step S 11 . Initialize the finite element simulation process and establish the target parameters that need to be optimized;

[0015] Step S 12 .Through the secondary development interface of the finite element simulation system, the fluid-solid coupling physical field simulation capability of the finite element simulation system is combined with the Python data calculation capability;

[0016] Step S 13 .Call the finite element simulation system through Python to perform simulation analysis and obtain the finite element simulation results.

[0017] The step S2 specifically includes the following steps:

[0018] Step S 21 . Extract multiple simulation characteristic indicators from the finite element simulation results, including extracting physical field values ​​from the finite element simulation results;

[0019] Step S 22 .Combining each simulation characteristic index into a series of comprehensive evaluation indexes;

[0020] Step S 23 . Analyze the mapping rules between each comprehensive evaluation index and the whole machine performance to determine the weight and influence of each comprehensive evaluation index, and use the multi-scale model fusion algorithm to fuse the comprehensive evaluation index to obtain the final finite element simulation result evaluation index;

[0021] Step S 24 .Based on the evaluation index of finite element simulation results, the target parameters are optimized using the Bayesian optimization algorithm.

[0022] The optimization of target parameters using the Bayesian optimization algorithm specifically includes:

[0023] Step S 241 .Build the initial dataset:

[0024] X0={x1,x2,...,x n}

[0025] Y0={y1,y2,...,y n}

[0026] where x1,x2,...,x n are the target parameters of finite element simulation, y1, y2, ..., y n It is the evaluation index of finite element simulation results;

[0027] Step S 242 .Based on the initial data set, a surrogate model is established using Gaussian process;

[0028] Step S 243 .Select the expected improvement function as the acquisition function for the next step of optimization:

[0029]

[0030] Where μ(x) is the predicted mean of the surrogate model at point x, σ(x) is the predicted standard deviation at point x; Φ is the cumulative distribution function of the standard normal distribution, φ is the probability density function of the standard normal distribution; f best is the current best objective function value;

[0031] Step S 244 .Optimize the selected acquisition function on the proxy model to find the next sampling point:

[0032]

[0033] Use the newly selected target parameter x for the finite element simulation next Conduct finite element simulation and obtain new evaluation indicators of finite element simulation results, and add them to the initial data set;

[0034] Step S 245 Update the surrogate model using new data and calculate a new posterior distribution.

[0035] Step S 246 .Repeat the above steps until the maximum number of iterations is reached.

[0036] After the optimization of the target parameters is completed, the optimized target parameters are passed to the finite element simulation system for the next modeling simulation.

[0037] The gated recurrent unit reinforcement learning includes a decision network and an evaluation network; the decision network processes the optimized simulation parameters and system status through the gated recurrent unit to generate a decision; the evaluation network evaluates the quality of the decision generated by the decision network through the gated recurrent unit, and provides rewards or penalties as feedback by comparing the difference between the output of the system status and the reference input.

[0038] Using the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model specifically refers to: evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model to obtain the evaluation results, and updating and iterating the gated recurrent unit reinforcement learning based on the evaluation results.

[0039] The evaluation of the reconstructed load results of the initial three-dimensional flow field load reconstruction model specifically refers to:

[0040] Step S 51 .Based on the initial three-dimensional flow field load reconstruction model, obtain the reconstructed three-dimensional flow field load value;

[0041] Step S 52 . Input the reconstructed three-dimensional flow field load value into the finite element simulation system for finite element simulation to obtain the simulated stress, calculate the mean square error between the simulated stress and the true stress value, and obtain the error loss1;

[0042] Step S 53 .Evaluate the reconstructed three-dimensional flow field load value based on real physical conditions to obtain the error loss2;

[0043] Step S 54 .Sum the error loss1 and the error loss2 to obtain the three-dimensional flow field load reconstruction evaluation result.

[0044] The step S 53 Specifically refers to:

[0045] Establish a solid wall boundary condition, set the velocity at the boundary to zero, and calculate the stress boundary value f of the theoretical physical field through the fluid physical parameters in the hydropower unit:

[0046]

[0047] In the formula, ρ is the density of the fluid, v is the velocity field, p is the pressure field, and μ is the viscosity coefficient of the fluid;

[0048] The error loss2 is obtained by calculating the mean square error between the reconstructed three-dimensional flow field load boundary value and the calculated theoretical physical field stress boundary value.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] 1. In the present invention, by introducing the digital twin system, a Python-based finite element secondary development platform is established, which realizes the integration of modules such as finite element simulation analysis, simulation result acquisition, parameter optimization algorithm construction, and finite element input parameter optimization. Compared with the prior art, by applying the digital twin system throughout the design process, the integrated intelligent optimization of the entire system is realized, and the integrity and synergy of the system design are improved.

[0051] Furthermore, the present invention can construct a finite element simulation result evaluation system through the digital twin system, which can more comprehensively and accurately evaluate the performance of power generation equipment and improve the accuracy of the evaluation. Based on this, the intelligent optimization of the finite element simulation input parameters of the hydropower unit components can be achieved, and it can be realized through a neural network model with embedded physical knowledge. Compared with the existing deep learning algorithm, it is more interpretable and can achieve high-precision reconstruction of the entire flow field with only less data.

[0052] Finally, based on the optimization of simulation parameters, the three-dimensional flow field reconstruction based on the finite element simulation input parameters and finite element simulation results of the hydropower unit components is realized through reinforcement learning based on gated recurrent units. Compared with conventional neural networks, the performance and efficiency of the system are significantly improved. In summary, the present invention has obvious technical advantages in the intelligent optimization of simulation parameters, the accuracy and efficiency of three-dimensional flow field reconstruction, the synergy of the system, the interpretability and adaptability of the algorithm, etc. These advantages make the process of three-dimensional flow field reconstruction of hydropower unit components more intelligent, efficient, and stable, and can better adapt to different working conditions and needs.

[0053] 2. In the present invention, the closed-loop full automation of parametric finite element modeling and simulation calculation, post-processing, and parameter correction is realized through finite element development tools such as Python and APDL. It is more efficient than the existing manual solution technology. The closed-loop fully automated finite element solution significantly improves the performance and efficiency of the system.

[0054] 3. The finite element simulation result evaluation system constructed by the present invention comprehensively considers various performance indicators of power generation equipment, improves the comprehensiveness and accuracy of the evaluation, and plays a key role in the comprehensive improvement of system performance.

[0055] 4. When evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model, the present invention evaluates the results of the initial three-dimensional flow field load reconstruction model based on finite element simulation and physical boundary conditions, which improves the comprehensiveness and accuracy of the evaluation and plays a key role in improving the convergence speed and accuracy of the algorithm. DETAILED DESCRIPTION

[0056] Example 1

[0057] As a basic implementation mode of the present invention, the present invention includes a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, comprising the following steps:

[0058] Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results.

[0059] Step S2. extract multiple simulation feature indicators from the finite element simulation results, fuse the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimize the simulation parameters based on the finite element simulation result evaluation indicator.

[0060] Step S3. Construct an initial three-dimensional flow field load reconstruction model based on gated recurrent unit reinforcement learning, and use the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model until a trained three-dimensional flow field load reconstruction model is obtained.

[0061] Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

[0062] Example 2

[0063] As a preferred embodiment of the present invention, the present invention includes a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, comprising the following steps:

[0064] Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results.

[0065] Step S2. extract multiple simulation feature indicators from the finite element simulation results, fuse the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimize the simulation parameters based on the finite element simulation result evaluation indicator.

[0066] Specifically, the step S2 specifically includes the following steps:

[0067] Step S 21 .Extract multiple simulation feature indicators from finite element simulation results, including extracting physical field values ​​from finite element simulation results.

[0068] Step S 22 .Integrate each simulation feature index into a series of comprehensive evaluation indicators.

[0069] Step S 23.Analyze the mapping rules between each comprehensive evaluation index and the performance of the whole machine to determine the weight and influence of each comprehensive evaluation index, and use the multi-scale model fusion algorithm to fuse the comprehensive evaluation indicators to obtain the final finite element simulation result evaluation index.

[0070] Step S 24 .Based on the evaluation index of finite element simulation results, the simulation parameters are optimized using the Bayesian optimization algorithm.

[0071] Step S3. Construct an initial three-dimensional flow field load reconstruction model based on gated recurrent unit reinforcement learning, and use the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model until a trained three-dimensional flow field load reconstruction model is obtained. Wherein, the gated recurrent unit reinforcement learning includes a decision network and an evaluation network. The decision network processes the optimized simulation parameters and system state through the gated recurrent unit to generate a decision. The evaluation network estimates the quality of the decision generated by the decision network through the gated recurrent unit, and provides rewards or penalties as feedback by comparing the difference between the output of the system state and the reference input.

[0072] Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

[0073] Example 3

[0074] As another preferred embodiment of the present invention, the present invention includes a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, comprising the following steps:

[0075] Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results.

[0076] Step S2. extract multiple simulation feature indicators from the finite element simulation results, fuse the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimize the simulation parameters based on the finite element simulation result evaluation indicator.

[0077] Step S3. Construct an initial three-dimensional flow field load reconstruction model based on gated recurrent unit reinforcement learning, and use the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model until a trained three-dimensional flow field load reconstruction model is obtained.

[0078] Among them, using the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model specifically refers to: evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model, obtaining the evaluation results, and updating and iterating the gated recurrent unit reinforcement learning based on the evaluation results. Specifically, evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model specifically refers to:

[0079] Step S 51 .Based on the initial three-dimensional flow field load reconstruction model, obtain the reconstructed three-dimensional flow field load value.

[0080] Step S 52 .The reconstructed three-dimensional flow field load value is input into the finite element simulation system for finite element simulation to obtain the simulated stress. The mean square error between the simulated stress and the true stress value is calculated to obtain the error loss1.

[0081] Step S 53 .Evaluate the reconstructed three-dimensional flow field load value based on real physical conditions to obtain the error loss2.

[0082] Step S 54 .Sum the error loss1 and the error loss2 to obtain the three-dimensional flow field load reconstruction evaluation result.

[0083] Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

[0084] Example 4

[0085] As the best embodiment of the present invention, the present invention includes a three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, comprising the following steps:

[0086] Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results. Specifically, the following steps are included:

[0087] Step S 11 .Initial settings are made for the finite element simulation process and target parameters that need to be optimized are established. The input parameters for the finite element simulation of the hydropower unit component model usually include basic parameters such as materials, other boundary conditions of the model, load application location, application form and application intensity and other model setting parameters. The initial settings include the hydropower unit component model, material parameters, boundary conditions and meshing method. The target parameters that need to be optimized are the load application location and intensity in the boundary conditions, and the range of the design parameters must be clearly defined.

[0088] Step S 12. Through the secondary development interface of the finite element simulation system, the fluid-solid coupling physical field simulation capability of the finite element simulation system is combined with the Python data calculation capability. Wherein, the finite element simulation system includes ansys, abaqus, comsol, etc.

[0089] Step S 13 .Call the finite element simulation system through Python to perform simulation analysis and obtain the finite element simulation results.

[0090] Step S2. Extract multiple simulation feature indicators from the finite element simulation results, fuse the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimize the target parameters in the simulation parameters based on the finite element simulation result evaluation indicator.

[0091] In view of the large-scale and complex structural characteristics of power generation equipment, it is necessary to make a comprehensive comparison between the overall physical field values ​​and the local simulated physical field values, taking into account the inherent characteristics, production capacity, economy, real-time operating status data, safety performance and operating reliability of the power generation equipment. By establishing a finite element simulation result evaluation system for hydropower unit components, a comprehensive evaluation index that can comprehensively reflect the operating status of the equipment can be constructed to reveal the mapping rules between each comprehensive evaluation index and the performance of the whole machine.

[0092] Specifically, step S2 may include the following steps:

[0093] Step S 21 .Extract multiple simulation feature indicators from finite element simulation results, including extracting physical field values ​​from finite element simulation results.

[0094] Extract key physical field values, such as temperature, stress, strain, etc., from the finite element simulation results. According to the finite element simulation results of the hydropower unit components, extract the water pressure distribution of the flow surface and the stress distribution of the bolt section at key locations. Consider the distribution of physical field values ​​under different working conditions and loads.

[0095] Analyze the overall physical field values ​​to understand the global performance of the equipment, such as the deformation of the whole machine, the overall temperature distribution, etc. Analyze the local physical field values, and pay attention to the stress concentration and temperature gradient of the key parts of the equipment. Combine the simulation results with the actual operation data, and consider the state of the equipment in real-time operation, such as load, speed, operating temperature, etc., to improve the practicality of the evaluation indicators.

[0096] Step S 22.Synthesize each simulation characteristic index into a series of comprehensive evaluation indexes. The comprehensive evaluation indexes such as health, stability, safety and economy are used to quantify the above physical field values ​​and ensure that these comprehensive evaluation indexes can objectively reflect the performance of the equipment. Select appropriate simulation characteristic indicators, such as stress level, temperature gradient, deformation, etc., to characterize the performance of the equipment in multiple dimensions. And select a fusion algorithm to integrate each simulation characteristic index into a single comprehensive evaluation index.

[0097] Step S 23 .Analyze the mapping rules between each comprehensive evaluation index and the performance of the whole machine to determine the weight and influence of each comprehensive evaluation index, and use the multi-scale model fusion algorithm to fuse the comprehensive evaluation indicators to obtain the final finite element simulation result evaluation index, thereby reducing the amount of data required for the subsequent state prediction model construction.

[0098] Step S 24 .Based on the evaluation index of finite element simulation results, the Bayesian optimization algorithm is used to optimize the target parameters. After completing the optimization of the target parameters, the optimized target parameters are passed to the finite element simulation system through the secondary development interface for the next modeling simulation.

[0099] The optimization of target parameters using the Bayesian optimization algorithm specifically includes:

[0100] Step S 241 .Select the prior distribution of Gaussian process (GP), which can represent our initial assumptions about the shape of the objective function. Based on the evaluation indicators of finite element simulation results, construct the initial data set:

[0101] X0={x1,x2,...,x n}

[0102] Y0={y1,y2,...,y n}

[0103] where x1,x2,...,x n are the target parameters of finite element simulation, y1, y2, ..., y n It is an evaluation index for finite element simulation results.

[0104] Step S 242 .Based on the initial dataset, a surrogate model is established using Gaussian process.

[0105] Step S 243 .Select the expected improvement (EI) function as the acquisition function for the next step of optimization. The acquisition function is used to guide the selection of the next sampling point. It measures the potential benefits of sampling at different locations and helps find the point with the most hope of improving the objective function:

[0106]

[0107] Where μ(x) is the predicted mean of the surrogate model at point x, σ(x) is the predicted standard deviation at point x; Φ is the cumulative distribution function of the standard normal distribution, φ is the probability density function of the standard normal distribution; f best is the current best objective function value.

[0108] Step S 244 .Optimize the selected acquisition function on the proxy model to find the next sampling point:

[0109]

[0110] Use the newly selected target parameter x for the finite element simulation next Finite element simulation is performed, and new evaluation indicators of finite element simulation results are obtained and added to the initial data set.

[0111] Step S 245 .Update the surrogate model using the new data and compute a new posterior distribution.

[0112] Step S 246 .Repeat the above steps until the maximum number of iterations is reached.

[0113] Step S3. Construct an initial three-dimensional flow field load reconstruction model based on the gated recurrent unit reinforcement learning algorithm, and use the optimized target parameters to train the initial three-dimensional flow field load reconstruction model. In the process of interaction with the simulation environment, learn how to predict the load value of the finite element simulation to minimize the simulation error until a trained three-dimensional flow field load reconstruction model is obtained.

[0114] The system environment of the reinforcement learning algorithm consists of: system state space s, system action space a, and system feedback r. The state space s of the reinforcement learning algorithm is the randomly generated system load value or the system load value iterated after several steps of iteration. The system action space a is the update direction and update step of the current step system load value. The system feedback r is a three-dimensional flow field load reconstruction evaluation system.

[0115] The reinforcement learning algorithm includes the actor and the critic, that is, the decision network and the evaluation network. The actor processes the finite element input parameters and system state through the gated recurrent unit (GRU) to generate decisions, solving the gradient vanishing problem of traditional RNN in long sequence processing. The critic evaluates the quality of these decisions through the gated recurrent unit, and provides rewards or penalties as feedback by comparing the difference between the output of the system state and the reference input. The input and output dimensions of the actor and critic networks depend on the design parameter space and the finite element simulation characteristics of the model.

[0116] During the training of the initial three-dimensional flow field load reconstruction model, the decision network and the evaluation network are iteratively updated: the decision network is updated by stochastic gradient descent using the policy gradient method, and it is ensured that the decision network can select the optimal load value to minimize the simulation error. The action behavior of the intelligent agent is corrected according to the state value obtained under any strategy. In each iteration cycle, the decision network is updated based on the current state s t , select action a t . t Interact with the proxy model and obtain feedback r(s t ) and reach the next state s t+1 The evaluation network calculates the state value V based on the feedback obtained from the current state π (s t ). By calculating the advantage function A π (s t ,a t ) to update the decision network based on the gradient descent algorithm. And repeat iteratively until the decision error rate made by the decision network is less than 5%.

[0117] Advantage function A π (s t ,a t ) is calculated as:

[0118] A π (s t ,a t )=r(s t ,a t )+γV π (s t+1 )-V π (s t )

[0119] Among them, γ is the attenuation coefficient, r(s t ,a t ) is the system feedback, V π (s t ) is the current state value, V π (s t+1 ) is the next state value.

[0120] The gradient of the decision network is calculated based on the advantage function, and the decision network is updated according to the stochastic gradient descent algorithm, thereby correcting the value judgment and strategy formulation process of the decision network on the system state. The iteration of the evaluation network needs to correct the value of the system state based on the Bellman iteration equation, calculate the parameter gradients before and after the correction, and update the evaluation network according to the stochastic gradient descent algorithm.

[0121] Among them, the gradient calculation of the decision network is as follows:

[0122]

[0123] In a given state s t Next, action a t The logarithmic probability of the corresponding strategy; A(s t ,a t ) is the advantage function.

[0124] The system state value is corrected through the Bellman iteration equation, and the state value network is updated using the Bellman iteration equation to better reflect the value of the current state of the system. The Markov decision chain is introduced, including the model state transition probability, state, action, reward and strategy. The state collection, action collection and reward collection of the agent are determined, and the total reward from the start state prediction to the final state is obtained according to the recursive relationship of the Bellman iteration equation. By correcting the state value, the decision network is guided to generate finite element parameter decisions that are more conducive to system performance.

[0125] The reward value calculation method for reinforcement learning is:

[0126]

[0127] Among them, γ is the attenuation coefficient, r t is the system feedback obtained in step t.

[0128] The iterative formula of the value function is as follows:

[0129]

[0130] Among them, γ is the attenuation coefficient, r(s t ) is system feedback.

[0131] Among them, the gated recurrent unit (GRU) is an improved recurrent neural network (RNN) structure that solves the long-term dependency problem in traditional RNN by introducing a gate mechanism. GRU is a simplified version of the long short-term memory network (LSTM), which reduces the number of parameters and performs as well as or even better than LSTM in some cases.

[0132] GRU controls the flow of information through two main gating mechanisms: reset gate and update gate.

[0133] The Reset Gate is used to control the degree of integration between the input of the current time step and the previous memory. The Reset Gate determines whether the previous state information should be forgotten. The Update Gate controls whether the information of the current time step should be retained to the next time step. It determines the degree of update of the hidden state, that is, the mixing ratio of new information and old information.

[0134] Compared with LSTM, GRU has only two gates, while LSTM has three (input gate, forget gate, output gate). Therefore, the calculation of GRU is relatively simpler.

[0135] The gating formula of GRU is:

[0136] Update Gate:

[0137] z t =σ(W z x t +U z h t-1 )

[0138] Update gate z t Control the hidden state h of the previous time step t-1 Hidden state h for the current time step t impact.

[0139] Reset the gate:

[0140] r t =σ(W r x t +U r h t-1 )

[0141] Reset Gate t Decide whether to discard the memory of the previous time step. If r t Close to 0, the memory of the previous time step will be reset.

[0142] Candidate hidden states:

[0143]

[0144] here is the candidate hidden state for the current time step. Reset gate r t Used to control whether to keep the hidden state of the previous time step.

[0145] Hide status updates:

[0146]

[0147] The final hidden state h t is updated by the gate z t The old state of controlt-1 and candidate hidden states Update gate z t It determines whether to keep old memories or update them with new information.

[0148] Among them, x t is the input of the current time step t, h t-1 is the hidden state of the previous time step t, h t is the hidden state at the current time step t, z t is the update gate, r t To reset the gate.

[0149] The advantages of GRU are: effective handling of long-term dependency problems. GRU can better handle long-distance dependencies in sequences through the gate mechanism. High computational efficiency. Due to its simple structure and fewer parameters, GRU has fast training and reasoning speed. It is also easy to train. The GRU structure is relatively simple, making it easier to perform parameter tuning and model training.

[0150] Using the optimized target parameters to train the initial three-dimensional flow field load reconstruction model specifically refers to: evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model to obtain evaluation results, and updating and iterating the gated recurrent unit reinforcement learning based on the evaluation results to make the load reconstruction results of the reinforcement learning algorithm more accurate.

[0151] Specifically, evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model specifically refers to:

[0152] Step S 51 .In each iteration cycle of the reinforcement learning algorithm, the decision network will reconstruct the three-dimensional flow field load according to the input parameters, obtain the reconstructed three-dimensional flow field load value, and store it for load reconstruction evaluation.

[0153] Step S 52 .The reconstructed three-dimensional flow field load value is input into the finite element simulation system for finite element simulation to obtain the simulated stress. The finite element simulation process is mass scaled to accelerate the simulation process. The simulated stress and the true stress value are calculated to obtain the error loss1.

[0154] Step S 53 .Evaluate the reconstructed three-dimensional flow field load value based on real physical conditions to obtain the error loss2.

[0155] In a three-dimensional flow field, the load must satisfy physical equations such as the Navier-Stokes equation as physical information boundary conditions. Through the Navier-Stokes equation, using the solid wall boundary condition, that is, the condition that the velocity of the fluid at the solid boundary is zero, the stress boundary value f of the theoretical physical field is calculated through the fluid physical parameters in the hydropower unit:

[0156]

[0157] Where ρ is the density of the fluid, v is the velocity field, p is the pressure field, and μ is the viscosity coefficient of the fluid.

[0158] The error loss2 is obtained by calculating the mean square error between the reconstructed load boundary value and the stress boundary value of the theoretical physical field.

[0159] Step S 54 .Sum the error loss1 and the error loss2 to obtain the three-dimensional flow field load reconstruction evaluation result.

[0160] Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

[0161] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical scheme and technical concept of the present invention without creative mental labor, which all fall within the scope of protection of the present invention.

Claims

1. A three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning, characterized by: The following steps are involved: Step S1. Establish a Python-based finite element secondary development platform to achieve finite element simulation initial settings, finite element simulation, and acquisition of finite element simulation results; Step S2. extracting multiple simulation feature indicators from the finite element simulation results, fusing the extracted multiple simulation feature indicators using a multi-scale model fusion algorithm to obtain a finite element simulation result evaluation indicator, and optimizing the simulation parameters based on the finite element simulation result evaluation indicator; Step S3. construct an initial three-dimensional flow field load reconstruction model based on gated recurrent unit reinforcement learning, and use the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model until a trained three-dimensional flow field load reconstruction model is obtained; Step S4. Reconstruct the three-dimensional flow field load model based on the training to obtain the reconstructed three-dimensional flow field load.

2. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 1 is characterized in that: The step S1 specifically includes the following steps: Step S 11 . Initialize the finite element simulation process and establish the target parameters that need to be optimized; Step S 12 .Through the secondary development interface of the finite element simulation system, the fluid-solid coupling physical field simulation capability of the finite element simulation system is combined with the Python data calculation capability; Step S 13 .Call the finite element simulation system through Python to perform simulation analysis and obtain the finite element simulation results.

3. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 1 is characterized in that: The step S2 specifically includes the following steps: Step S 21 . Extract multiple simulation characteristic indicators from the finite element simulation results, including extracting physical field values ​​from the finite element simulation results; Step S 22 .Combining each simulation characteristic index into a series of comprehensive evaluation indexes; Step S 23 . Analyze the mapping rules between each comprehensive evaluation index and the whole machine performance to determine the weight and influence of each comprehensive evaluation index, and use the multi-scale model fusion algorithm to fuse each comprehensive evaluation index to obtain the final finite element simulation result evaluation index; Step S 24 .Based on the evaluation index of finite element simulation results, the target parameters are optimized using the Bayesian optimization algorithm.

4. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 3 is characterized in that: The optimization of target parameters using the Bayesian optimization algorithm specifically includes: Step S 241 .Build the initial dataset: X0={x1,x2,…,x n } Y0={y1,y2,...,y n } where x1,x2,...,x n are the target parameters of finite element simulation, y1, y2, ..., y n It is the evaluation index of finite element simulation results; Step S 242 .Based on the initial data set, a surrogate model is established using Gaussian process; Step S 243 .Select the expected improvement function as the acquisition function for the next step of optimization: Where μ(x) is the predicted mean of the surrogate model at point x, σ(x) is the predicted standard deviation at point x; Φ is the cumulative distribution function of the standard normal distribution, φ is the probability density function of the standard normal distribution; f best is the current best objective function value; Step S 244 .Optimize the selected acquisition function on the proxy model to find the next sampling point: Acquisition function (x) Use the newly selected target parameter x for the finite element simulation next Conduct finite element simulation and obtain new evaluation indicators of finite element simulation results, and add them to the initial data set; Step S 245 Update the surrogate model using new data and calculate a new posterior distribution. Step S 246 .Repeat the above steps until the maximum number of iterations is reached.

5. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 3 is characterized in that: After the optimization of the target parameters is completed, the optimized target parameters are passed to the finite element simulation system for the next modeling simulation.

6. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 1 is characterized in that: The gated recurrent unit reinforcement learning includes a decision network and an evaluation network; the decision network processes the optimized simulation parameters and system status through the gated recurrent unit to generate a decision; The evaluation network evaluates the quality of the decisions generated by the decision network through gated recurrent units, and provides rewards or penalties as feedback by comparing the output of the system state with the reference input.

7. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to any one of claims 1 to 6, characterized in that: Using the optimized simulation parameters to train the initial three-dimensional flow field load reconstruction model specifically refers to: evaluating the reconstructed load results of the initial three-dimensional flow field load reconstruction model to obtain the evaluation results, and updating and iterating the gated recurrent unit reinforcement learning based on the evaluation results.

8. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 7 is characterized in that: The evaluation of the reconstructed load results of the initial three-dimensional flow field load reconstruction model specifically refers to: Step S 51 .Based on the initial three-dimensional flow field load reconstruction model, obtain the reconstructed three-dimensional flow field load value; Step S 52 . Input the reconstructed three-dimensional flow field load value into the finite element simulation system for finite element simulation to obtain the simulated stress, calculate the mean square error between the simulated stress and the true stress value, and obtain the error loss1; Step S 53 .Evaluate the reconstructed three-dimensional flow field load value based on real physical conditions to obtain the error loss2; Step S 54 .Sum the error loss1 and the error loss2 to obtain the three-dimensional flow field load reconstruction evaluation result.

9. The three-dimensional flow field load reconstruction method based on gated recurrent unit reinforcement learning according to claim 8 is characterized in that: The step S 53 Specifically refers to: Establish a solid wall boundary condition, set the velocity at the boundary to zero, and calculate the stress boundary value f of the theoretical physical field through the fluid physical parameters in the hydropower unit: In the formula, ρ is the density of the fluid, v is the velocity field, p is the pressure field, and μ is the viscosity coefficient of the fluid; The error loss2 is obtained by calculating the mean square error between the reconstructed three-dimensional flow field load boundary value and the calculated theoretical physical field stress boundary value.

Images