Intelligent water conservancy design simulation system based on digital twinning

Abstract

This invention relates to the field of computer-aided engineering technology and discloses an intelligent hydraulic design simulation system based on digital twins, comprising: a parametric geometric feature extraction module, a reduced-order model database, a tangent space linear extrapolation module, and a flow field reconstruction module. This invention utilizes the differential properties of physical equations to construct a locally linearized mapping, reducing the dimensionality of nonlinear flow field solutions to deterministic matrix-vector multiplication operations, replacing traditional statistical regression and iterative solutions. This invention ensures that the reconstruction results evolve along the tangent direction of the physical gradient, avoiding non-physical guesses of statistical models in sparse sample regions, and achieving high-fidelity flow field reconstruction that combines real-time performance with physical evolution conservation.

Description

Technical Field

[0001] This invention relates to an intelligent water conservancy design simulation system based on digital twins, belonging to the field of computer-aided engineering technology. Background Technology

[0002] In the current design of complex curved surfaces in large-scale water conservancy projects, conventional techniques rely on computer-aided design to construct parametric geometric models and computational fluid dynamics to solve the Navier-Stokes equations to evaluate flow capacity or energy dissipation efficiency. However, due to the inherent heterogeneity in data topology between computer-aided design, which uses continuous boundary representation, and engineering simulation, which relies on discrete meshes, geometric fine-tuning at the design end causes the simulation end to discard the original mesh and re-execute the geometric cleanup, mesh generation, and equation iterative solution process. This destructive reconstruction process results in a huge time lag between physical performance feedback and geometric design operations, thus severing the real-time mapping relationship between shape and flow.

[0003] Besides the problem of interaction between geometric data and mesh topology, existing simulation calculation strategies lack flexibility and reverse guidance capabilities. For example, Chinese invention patent CN111666668B discloses a method for determining parameters in CFD-DEM numerical simulation of gravel movement in a pipeline containing sand. It determines contact parameters and coupling exchange terms through interactive verification of physical experiments and numerical simulation, and establishes a forward solution benchmark for multiphase flow. However, the core technology relies on unidirectional full solution of physical control equations. Faced with high-frequency geometric iterations at the design end, it cannot analyze the differential gradient information of the flow field state evolving with geometric deformation. Small changes in curvature or boundary require restarting the iteration calculation of the underlying equations. It is impossible to directly deduce the flow field distribution under new working conditions without calling the solver. The serial mode of one change and one calculation makes it difficult for designers to use existing data to mine the sensitivity correlation between geometric control points and hydraulic performance. The optimization design of complex curved surfaces remains in the stage of blind trial and error based on experience.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a real-time simulation mechanism that follows the physical evolution gradient, maintains the integrity of the flow field, and provides reverse design guidance, while avoiding the expensive overhead of full computational fluid dynamics solution. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: An intelligent water conservancy design simulation system based on digital twins, comprising:

[0006] The parametric geometric feature extraction module is used to obtain the geometric control point coordinate vectors of the design object and convert them into geometric feature descriptors based on preset topological mapping rules.

[0007] The reduced-order model database stores pre-computed sets of orthogonal basis functions for the flow field and discrete anchor point datasets. The orthogonal basis function set contains multiple linearly independent spatial mode vectors for the flow field. The discrete anchor point dataset contains reference mode weight coefficient vectors corresponding to several reference geometric states, as well as modal sensitivity Jacobian matrices corresponding to each reference geometric state. The modal sensitivity Jacobian matrix is ​​used to characterize the rate of change of the partial derivatives of the flow field mode weight coefficients with respect to the geometric feature descriptors.

[0008] The tangent space linear extrapolation module is used to calculate the geometric deviation vector between the real-time input geometric feature descriptor and the selected reference geometric state in the discrete anchor point dataset, and to perform matrix-vector multiplication based on the modal sensitivity Jacobian matrix and the geometric deviation vector to generate the modal coefficient increment vector. The tangent space linear extrapolation module linearly superimposes the modal coefficient increment vector with the reference modal weight coefficient vector to determine the target modal weight coefficient for the current working condition.

[0009] The flow field reconstruction module is used to perform a weighted summation operation on the target modal weight coefficients and the orthogonal basis function set of the flow field, generate reconstructed flow field data, and map it onto the surface of the design object.

[0010] Preferably, the system further includes: an inverse sensitivity analysis module, used to calculate the gradient vector of the preset hydraulic performance index function relative to the geometric feature descriptor based on the modal sensitivity Jacobian matrix and using the chain rule after the target modal weight coefficients are output by the tangent space linear extrapolation module; and an optimization guidance display module, used to back-map the gradient vector to the coordinate space of the geometric control points and generate vector identifiers in the design view that indicate the optimized displacement direction of the geometric control points; the vector identifiers are used to characterize the influence weight and improvement direction of each geometric control point on the hydraulic performance index.

[0011] Preferably, the tangent space linear derivation module uses the following formula to perform matrix-vector multiplication operations: ,in, For target modal weighting coefficients, As the baseline modal weighting coefficient, Here is the modal sensitivity Jacobian matrix. The geometric deviation vector is used; matrix-vector multiplication is based on the linear approximation of the manifold tangent plane to determine the evolution of the flow field modes as geometric parameters change.

[0012] Preferably, the system further includes: a geometric orthogonality verification module, used to store the reference orthogonal basis matrix of the geometric feature space and calculate the orthogonal residual components of the real-time input geometric feature descriptors outside the subspace spanned by the reference orthogonal basis matrix; a deduction effectiveness control module, used to monitor the energy ratio of the orthogonal residual components in real time; when the energy ratio exceeds a preset threshold, it is determined that the current design object is outside the effective computational domain, the output of the flow field reconstruction module is blocked, and a trigger signal is generated to call the background numerical calculation unit to perform the full physical equation solution.

[0013] Preferably, the system further includes: a sensor data interface for acquiring real-time monitoring data from physical sensors arranged at specific spatial coordinates of the physical hydraulic engineering project; an observation matrix construction unit for extracting corresponding basis function components from a reduced-order model database based on the spatial coordinates of the physical sensors, constructing a sparse observation matrix, wherein the sparse observation matrix defines the linear relationship between modal coefficients and sensor monitoring data; and a flow field state inversion module for constructing an overdetermined linear equation system based on the sparse observation matrix and real-time monitoring data, and solving the linear equation system using the least squares method to determine the measured modal weight coefficients characterizing the current flow field state of the physical engineering project.

[0014] Preferably, the system further includes: an incremental subspace update module, used to calculate the orthogonal projection residual vector of the calibrated flow field data on the orthogonal basis function set when the calibrated flow field data is acquired; and a basis function library expansion unit, used to perform Schmitt orthogonalization and normalization processing on the orthogonal projection residual vector when the magnitude of the orthogonal projection residual vector exceeds a preset update threshold, generate new incremental orthogonal basis functions, and append and store them in the reduced-order model database.

[0015] Preferably, the flow field orthogonal basis function set is a set of standard orthogonal bases extracted by performing intrinsic orthogonal decomposition or singular value decomposition on the flow field snapshot matrix covering the design domain; the modal sensitivity Jacobian matrix is ​​a sensitivity matrix pre-calculated by solving the adjoint equation of the fluid dynamics control equation or by using the central difference method.

[0016] Preferably, the tangent space linear extrapolation module is also used to select multiple neighboring reference geometric states in the discrete anchor point dataset that are closest to the Euclidean distance of the geometric feature descriptor of the real-time input; perform local tangent space extrapolation based on the modal sensitivity Jacobian matrix corresponding to each neighboring reference geometric state respectively; and perform weighted averaging on multiple extrapolation results based on the reciprocal of the Euclidean distance to determine the target modal weight coefficient.

[0017] Preferably, the parametric geometric feature extraction module includes a topology mapping unit, which is used to map design objects with different mesh topologies into a vector space with a uniform dimension; the geometric feature descriptor is a shape deformation feature vector after removing rigid body displacement and rotation components.

[0018] Preferably, the flow field reconstruction module includes a three-dimensional scalar field visualization unit; the three-dimensional scalar field visualization unit is used to map the pressure field data, velocity field data and turbulent kinetic energy data in the reconstructed flow field data into different color and transparency attributes, and overlay them on the surface of the three-dimensional geometric model of the design object to form a visualization model of the global flow field distribution.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In intelligent water conservancy design, the Jacobian matrix, which represents the rate of change of modal coefficients with geometric features, is pre-calculated and stored offline, and a linear extrapolation mechanism based on the tangent space of the manifold is constructed. The Jacobian matrix carries the differential information of the physical equations, and the real-time extrapolation process is forced to proceed along the tangent direction of physical evolution, which is different from statistical interpolation or black-box regression that depends on the location of sample points. This avoids numerical oscillations or non-physical smoothing phenomena caused by the lack of physical constraints in sparse sample regions by conventional reduced-order models. It ensures that when the geometric parameters change continuously, the generated flow field data always maintains the first-order approximation accuracy of the solution of the Navier-Stokes equations in mathematical structure, and preserves the integrity of key high-frequency physical features such as shock waves and vortex shedding in the millisecond-level response.

[0021] 2. Utilizing the analytical differentiability of the non-intrusive reduced-order mapping model itself, a gradient backpropagation mechanism for flow field performance indicators is established; the partial derivative vector of the target hydraulic performance with respect to each geometric control point is directly calculated using the chain rule, mapping the numerical sensitivity back to the three-dimensional geometric topological space; the experience-dependent blind trial-and-error iteration is transformed into deterministic navigation with the gradient vector indicating the optimization direction, intuitively identifying the geometric root causes of local flow deterioration, and achieving complex surface morphology correction and directional optimization without calling the adjoint equation solver.

[0022] 3. By constructing a dual orthogonal projection verification mechanism based on geometric features and flow field modes, it has the ability to adaptively identify and expand knowledge for unknown working conditions; when the geometric shape or flow field features input in real time exceed the scope of the subspace spanned by the current basis function library, it calculates the orthogonal residual components to quantify the cognitive bias and triggers a targeted incremental update process; it uses the Schmitt orthogonalization principle to peel off the independent features of new working conditions and transform them into new orthogonal basis vectors to be added to the model library; the incremental update strategy avoids the catastrophic forgetting problem caused by traditional deep learning models when absorbing new knowledge, and ensures that the accuracy of expressing complex irregular boundaries or special flow states is continuously improved with the accumulation of actual working conditions throughout the entire life cycle of the simulation system. Attached Figure Description

[0023] Figure 1 This is a flowchart of the real-time flow field reconstruction calculation based on tangent space linear extrapolation of the present invention;

[0024] Figure 2This is a comparison chart showing the trend of reconstruction error with sampling density under different geometric deformation rates according to the present invention;

[0025] Figure 3 This is a diagram illustrating the overall architecture of the digital twin system that integrates physical monitoring and real-time simulation engines, as presented in this invention. Detailed Implementation

[0026] This specific embodiment is only used to illustrate the present invention and is not intended to limit the scope of the present invention. Those skilled in the art should understand that modifications or equivalent substitutions can be made to this specific embodiment without departing from the spirit and scope of the present invention, and such modifications and substitutions all fall within the protection scope of the claims of the present invention.

[0027] This invention provides an intelligent hydraulic design simulation system based on digital twins, comprising a computational architecture including a parametric geometry analysis module, a flow field orthogonal basis library storage unit, a manifold projection coefficient inference engine, and a flow field linear reconstruction renderer. The parametric geometry analysis module listens for geometric change events in the computer-aided design environment via an application programming interface or data bus, and extracts the coordinate vectors of the geometric control points of the current hydraulic structures. To address the topological heterogeneity between the computer-aided design model and the computer-aided engineering mesh, this module incorporates topological mapping rules to convert control points of non-uniform rational B-spline surfaces with different vertex numbers and connectivity relationships into standardized geometric feature descriptors with fixed dimensions. ;in, , To determine the dimension of the geometric feature descriptor, an isomorphic mapping procedure based on isoparametric resampling is constructed, with a predefined standardized parameter domain. and Fixed topology grid point set For any input non-uniform rational B-spline (NURBS) surface solid, perform arc length parameterization normalization processing to calculate... Physical coordinate vectors on the surface corresponding to each node ,Will Flatten the control points according to row priority, perform a weighted Prouk analysis, and use the sensitivity of each control point to hydraulic performance as the weight matrix. Solve for the minimum objective function The translation vector T and rotation matrix R, after removing rigid body motion components, result in a constant output dimension. Shape feature descriptor This ensures the basis consistency of the input vector space.

[0028] The flow field orthogonal basis library storage unit is used to store the pre-calculated set of flow field orthogonal basis functions. And anchor point data for several reference geometric states, and a set of orthogonal basis functions for the flow field. The flow field snapshot matrix is ​​obtained by performing full computational fluid dynamics calculations on discrete sampling points covering the design parameter space. and the matrix The first part extracted by performing truncated singular value decomposition or eigenorthogonal decomposition There are principal mode vectors; where each basis function is... All dimensions are A spatial vector is used to characterize an independent spatial distribution pattern of the flow field at a specific energy level, and satisfies the orthogonality condition. Anchor point data contains reference geometric feature descriptors The corresponding baseline mode weight coefficient vector and modal sensitivity Jacobian matrix The modal sensitivity Jacobian matrix Each element in Used to quantify the The weighting coefficient of the first flow field mode varies with the weighting coefficient of the second flow field mode. The partial derivatives of the geometric parameters change, i.e. This matrix is ​​obtained by solving the adjoint equation in the offline stage or by using the central difference method, and is used to carry the tangent space gradient information of the flow field evolution with geometric deformation.

[0029] A manifold projection coefficient inference engine is used to perform online inference processes. This inference engine employs a linear extrapolation mechanism based on the manifold tangent space to compute the geometric feature descriptors input in real time. Nearest neighbor reference anchor selected in the memory cell Geometric deviation vector between ,in Perform matrix-vector multiplication operations. Calculate the modal coefficient increment caused by geometric changes. ; through the linear superposition formula Determine the target modal weighting coefficients for the current operating condition. This calculation process is based on linear algebra operations, and its computational complexity is O(n log n). , and flow field grid size This is irrelevant, allowing the system to respond to continuous changes in design parameters at a fixed time tick; the flow field linear reconstruction renderer is used to perform physics generation tasks, and this renderer performs matrix operations. This process maps low-dimensional modal coefficients back to a high-dimensional physical grid space, generating reconstructed flow field data that includes pressure, velocity, or turbulent kinetic energy distributions. The reconstructed flow field data is mapped onto the geometric surface of the design object and rendered as cloud maps or streamlines using a 3D visualization engine. The system also includes a geometric sensitivity inverse analysis module, which utilizes the chain rule to map the gradient of preset hydraulic performance indicators with respect to modal coefficients. With modal sensitivity Jacobian matrix Multiply to calculate the gradient vector of the performance index relative to the geometric control points. ,in The gradient vector is mapped back to the design view and displayed as a vector arrow superimposed on each geometric control point to indicate the direction and magnitude of movement that each control point should make to improve hydraulic performance.

[0030] The system further includes a geometric manifold orthogonality verification module for monitoring the credibility of the reduced-order model. This module calculates the input geometric feature descriptors in real time. Orthogonal residual components outside the geometric subspace spanned by the training samples Specifically, the system pre-constructs orthogonal basis matrices of geometric features. And calculate the projection residual energy ratio ,in When the ratio When the preset safety threshold is exceeded, the system determines that the current design deviates from the effective computation domain of the model and triggers the inference confidence circuit breaker controller, enabling the instruction-level computation graph collapse mechanism based on the graphics pipeline, and allocating an independent circuit breaker status word in the GPU Unified Buffer Object (UBO). The verification unit utilizes the Computation Unified Device Architecture (CUDA) kernel to compute the residuals in parallel. ,once Exceeding the threshold and performing an atomic write operation will Setting, hardware-level conditional branch instructions are implanted during the vertex processing stage: reading If the logic is true, then the clipping space coordinates of the current vertex will be... Writing a Not-Indexed (NaN) pixel causes the rasterization unit to discard pixels during the primitive assembly stage, physically preventing invalid pixels from being written to the frame buffer, thus preventing artifact misdirection, blocking the flow field reconstruction output and generating a warning signal. Simultaneously, a trigger command is generated to invoke the background computing unit to perform full computational fluid dynamics calculations. The system also includes an incremental subspace update module to process new data generated from the full computation. This module uses an incremental singular value decomposition algorithm to calculate new flow field data within the existing orthogonal basis library. orthogonal projection residual vector on If vector Length of the module If the update threshold is exceeded, then the vector... Perform normalization and use it as a new basis function. Add to the base library.

[0031] Example 1: This example is applied to the shape optimization design of the reverse arc section of a high-head dam spillway. When the system faces this condition, designers need to fine-tune the radius of curvature and the starting angle of the reverse arc section to eliminate the risk of cavitation erosion that may be caused by high-speed water flow in local areas. This requires that the negative pressure on the wall be kept above the critical cavitation pressure threshold. When designers drag the control points of the reverse arc section using computer-aided design software, the parametric geometry analysis module reads the operation in real time through the data interface and maps the resulting geometric topology changes into standardized geometric feature descriptors. The system calls the pre-stored set of orthogonal basis functions for the flow field in the reduced-order model database. and discrete anchor point dataset, orthogonal basis function set of flow field Including the front The first-order principal mode vectors, these mode vectors are obtained by considering the design space of the spillway. The data were obtained by performing intrinsic orthogonal decomposition on full computational fluid dynamics flow field snapshots under different working conditions.

[0032] The tangent space linear inference module uses Euclidean distance to locate the current geometric feature descriptor in the discrete anchor point dataset. Find the nearest neighbor reference anchor point and obtain the modal sensitivity Jacobian matrix corresponding to that anchor point. The matrix The local linear response rate of the flow field modal coefficients as a function of geometric parameters is quantified, and the geometric deviation vector of the system is calculated. And perform matrix-vector multiplication operations. The modal coefficient increments are determined by utilizing the differential properties of the physical equations embodied in the Jacobian matrix. This ensures that the derivation process proceeds along the tangent space direction of the fluid dynamics equation solutions, guaranteeing that the generated flow field solution satisfies the mass and momentum conservation constraints in its mathematical structure even when geometric parameters undergo nonlinear deformation. The flow field reconstruction module then updates the target modal weight coefficients and the orthogonal basis function set of the flow field. A linear superposition operation is performed to generate reconstructed flow field data containing the global pressure distribution, which is then mapped back to the three-dimensional geometric surface of the spillway. When the reconstruction results show that a negative pressure zone below a preset threshold appears in a local area of ​​the reverse arc segment, the inverse sensitivity analysis module is activated. This module reuses the modal sensitivity Jacobian matrix. The gradient vector of the minimum pressure index relative to the coordinates of each geometric control point is calculated using the chain rule. The system converts the gradient vector into a visualized 3D vector arrow, which is overlaid on the corresponding geometric control point. This vector arrow points in the direction of the fastest increase in pressure gradient, and its length is proportional to the sensitivity value. Designers, guided by this vector arrow, fine-tune the control point position along the arrow's direction to eliminate negative pressure zones without repeated trial and error. During the simulation, the geometric orthogonality verification module synchronously monitors the real-time input geometric feature descriptors. Compared to the projected residuals of the reference orthogonal basis matrix, if the designer modifies the inverse arc segment into an irregular structure that does not appear in the training sample space, the energy proportion of the orthogonal residual components will exceed [a certain percentage]. The deduction effectiveness control module blocks the output of the flow field reconstruction module and triggers a request to solve all physical equations in the background.

[0033] Example 2: This example uses a comparative verification experiment based on real engineering parameters to quantitatively demonstrate the effectiveness of the constructed intelligent hydraulic design simulation system in resolving the contradiction between high-precision flow field reconstruction and real-time response delay, and to verify its stability under large geometric deformation conditions. The test platform uses a workstation configured with an Intel Xeon Gold 6248R processor (48 cores), 256GB of memory and an NVIDIA A100 GPU. The data source is taken from the real design dataset of a large hydropower station spillway project, which contains 500 sets of high-precision three-dimensional flow field snapshots under different shape parameters. These snapshots are obtained by solving the unsteady Reynolds-averaged Navier-Stokes equations using the commercial computational fluid dynamics software ANSYS Fluent, and serve as the full true value benchmark for the experiment.

[0034] The core parameter in this experiment is the number of orthogonal basis cutoff modes. With anchor sampling density ,parameter The energy percentage that determines the flow field reconstruction directly affects the physical fidelity; parameters The effective radius of the linear extrapolation of the tangent space directly affects the extrapolation accuracy under large geometric deformations. The experimental design did not use a single empirical value, but instead constructed a gradient test matrix: Set to four levels: [10, 30, 50, 80]. The dataset was set to three levels: sparse, medium, and dense. The experimental procedure was as follows: 20% of the full ground truth dataset was randomly selected as the test set, and the remaining 80% was used to construct a reduced-order model database. For each sample in the test set, the flow field prediction was performed using the system of this invention (tangent space linear extrapolation) and the traditional radial basis function interpolation method, respectively. Two key indicators were examined: one was the relative root mean square error of the entire field, used to characterize the reconstruction accuracy of the physical field; the other was the time taken for a single inference, used to characterize the real-time interactive performance. The experimental results showed performance differences in... Under the configuration, the traditional radial basis function interpolation method rapidly increases the flow field error to over 15% when dealing with working conditions where the geometric deformation exceeds 5%, and non-physical pressure oscillations occur in the near-wall region. In contrast, the system of the present invention, with the tangential guidance provided by the Jacobian matrix, can stably control the error within 3.2% under the same deformation amplitude, as shown in Table 1.

[0035] Table 1: Performance Comparison Data of Different Inference Methods

[0036]

[0037] As can be seen from the data in Table 1, the system of the present invention not only achieves an order-of-magnitude improvement in accuracy, but more importantly, it shows an advantage in the qualitative indicator of whether there are non-physical oscillations. This is because the tangent space extrapolation mechanism forces the flow field solution to evolve along the tangent plane of the solution manifold of the physical equation, thereby embedding physical conservation laws in the mathematical structure, rather than just statistical fitting of discrete data points. In addition, the extrapolation time is stable at 8 milliseconds, which meets the real-time interactive requirements of designers (usually requiring a latency of less than 30 milliseconds). Further boundary tests show that when the geometric deformation rate exceeds 15%, the orthogonal residual energy monitoring module of the present invention accurately triggers the circuit breaker mechanism to prevent potentially large error output.

[0038] Example 3: This example combines Figures 1 to 3 This document describes an intelligent water conservancy design simulation system based on digital twins, such as... Figure 1 As shown, the core calculation process begins with the coordinate vector of the geometric control points of the design object at the input end. This vector is then transformed into a standardized geometric feature descriptor by the parametric geometric feature extraction module. The descriptor then enters the tangent space linear extrapolation module and retrieves the discrete anchor point dataset stored in the reduced-order model database, which contains the baseline modal weight coefficients and the modal sensitivity Jacobian matrix. The modal coefficient increment is generated by calculating the product of the geometric deviation and the Jacobian matrix and superimposed on the baseline coefficients, thereby outputting the target modal weight coefficients for the current working condition. Subsequently, the flow field reconstruction module receives these coefficients and performs a weighted summation operation in conjunction with the orthogonal basis function set of the flow field retrieved from the database, ultimately generating reconstructed flow field data mapped onto the surface of the design object.

[0039] like Figure 2 As shown, this line graph displays the relative root mean square error (RMSE) across the entire field, expressed in percentage, as the sampling density at anchor points increases. The trend relationship is divided into three levels: sparse, medium, and dense. The figure includes three error curves corresponding to small deformation (3%), medium deformation (7%), and large deformation (12%), respectively. The large deformation curve marked with a triangle shows that its error decreases as the sampling density increases from sparse to dense, while the small deformation curve marked with a circle maintains a low error level at all densities. Figure 3As shown, the system architecture includes a top-level designer interaction terminal, which houses a computer-aided design software host environment and 3D field rendering and interaction plugins. Geometric parameters and reconstructed field data are transmitted via TCP / IP network protocol. The left side of the system connects to the physical engineering site's physical hydraulic engineering and physical sensor array for monitoring pressure and flow velocity. Real-time monitoring data is transmitted back through sensor data interfaces. The core layer is a digital twin computing server, equipped with a multi-core processor and graphics accelerator card. It integrates a real-time intelligent inference engine, including a tangent space linear inference unit, a flow field reconstruction and mapping unit, and a geometric orthogonality verification unit, as well as a background computing and update service, including a full physical equation solving unit and an incremental subspace update unit. It is connected to the underlying model storage system via a high-speed data bus. This storage system is responsible for managing the flow field orthogonal basis function set, modal sensitivity matrix, and discrete anchor point dataset.

[0040] Example 4: This example provides a systematic technical implementation path and parameter calibration procedure for two key functional units: the incremental subspace update module and the inference confidence circuit breaker controller. For the implementation of the incremental subspace update module, the system adopts an energy criterion based on orthogonal projection residuals to drive the dynamic expansion of the basis functions. When the system receives the new flow field snapshot vector generated by the full computational fluid dynamics calculation, When this happens, it is projected onto the current flow field orthogonal basis function set. Calculate the projection components in the spanned subspace. The system calculates the orthogonal residual vector of the snapshot outside the subspace. and the corresponding residual energy ratio ,in , To avoid redundant expansion of the base library due to numerical noise, the system sets an energy threshold based on the truncation error tolerance. The determination of this threshold follows the procedure below: During the offline training phase, the full training set is calculated before truncation. The average reconstruction error energy ratio after the first mode and will Set as to Between, when real-time calculation When the system determines that the current snapshot contains new physical features that cannot be expressed by the existing base library, it then adjusts the residual vector accordingly. Perform Schmitt orthogonalization and normalization operations to generate new basis functions. And added to In the middle, incremental updates of the base library dimension are completed.

[0041] Regarding the threshold setting for the circuit breaker controller based on inference confidence, this embodiment employs a statistical calibration method based on geometric-physical mapping consistency. This controller relies on geometric feature descriptors. In geometrically orthogonal basis matrix External projection residual energy ratio To determine the safe circuit breaker threshold During the offline phase, the system performs the following calibration process: constructing a test set containing extrapolated samples of the design space boundary; and calculating the geometric residual energy ratio for each sample in the test set. and the corresponding flow field prediction error This is defined as the relative error between the predicted solution and the full true solution; and it is achieved through fitting... and Find the correlation curve between them Exceeding the allowable error in engineering, such as The corresponding geometric residual critical value, and the critical value of the time. Set the circuit breaker threshold for online operation Through the integration of the above mechanisms, the system can automatically identify potential accuracy risks and trigger self-evolution when design parameters exceed the original training space. For example, in the scenario of optimizing the reverse arc section of a spillway, when designers try a new type of discontinuous curvature design, the geometric residual energy ratio... leap to (Exceeding the calibrated threshold) The system blocks real-time output and calls the background CFD solution; after the solution is completed, the calculated flow field residual energy ratio for (Exceeded update threshold) The system extracts new vortex structure modes and updates the baseline library. In similar design simulations, the system uses the updated baseline library to control the flow field prediction error within a certain range. Within.

[0042] Example 5: This example provides a standardized offline calibration and data filling procedure for the constructed intelligent hydraulic design simulation system, ensuring the system's versatility and high fidelity when facing different types of hydraulic structures, different Reynolds number ranges, and different geometric complexities. Regarding the construction of the orthogonal basis function set for the flow field, the procedure defines the boundaries of the sample space and the sampling strategy. For any specific type of hydraulic structure, the procedure requires determining its design parameter space. ,in To determine the number of geometric control parameters, the Latin hypercube sampling method is used. Internal generation not less than For each initial sample point, the procedure performs a full-scale solution of the three-dimensional unsteady Reynolds-averaged Navier-Stokes equations to obtain high-resolution flow field snapshot data. During the solution process, the flow field residual convergence curve is monitored. When the residuals of both the momentum equation and the continuity equation are below a certain value... And the fluctuation range of the key physical quantities is smaller than Only when all valid snapshots have been collected can the snapshot be considered valid. After all valid snapshots have been collected, the procedure performs intrinsic orthogonal decomposition and determines the validity based on the cumulative energy percentage being no less than [amount missing]. Standard cutoff selection before The principal modal vectors serve as the orthogonal basis function set for this type of building. .

[0043] Regarding the calculation of the modal sensitivity Jacobian matrix, the specification defines a numerical calculation standard based on the complex step-difference method. To avoid the sensitivity of step size selection to truncation and rounding errors in the traditional finite difference method, the specification requires that the design parameters be included. Extend to the complex field and apply an imaginary perturbation. ,in Set as On an order of magnitude, by solving the flow field equations in the complex domain and using the formula By directly extracting the partial derivatives of the modal coefficients with respect to geometric parameters, this procedure not only eliminates the dependence on the difference step size but also improves the accuracy of derivative calculation to the machine precision level, ultimately yielding the Jacobian matrix. It needs to be checked by singular value decomposition; if its condition number exceeds... If the parameter space near the anchor point is determined to have geometric topological singularity, it is necessary to encrypt the sampling in this region and regenerate the anchor point data until the Jacobian matrix of all anchor points satisfies the numerical stability requirement.

[0044] Example 6: This example provides a set of procedures for real-time monitoring of online simulation errors and adaptive model calibration for the constructed intelligent hydraulic design simulation system. This ensures that the system maintains a preset accuracy baseline during actual deployment, even when faced with accumulating data deviations and environmental disturbances. For real-time quantification of online simulation errors, the procedure defines a monitoring method based on the conservation of flow field residual energy. The system performs real-time monitoring at each time step... The current flow field state vector is output by the flow field reconstruction renderer. The system will Substitute into the discretized Navier-Stokes equations In the process of calculating the unbalanced residual vector ,in To compress high-dimensional residual vectors into scalar indices that can be determined in real time, the system calculates the physical residual energy norm. ,in The procedure sets a dynamic threshold. Its initial value is determined by the average residual level during the offline verification phase, when continuous Each time step was monitored When the system determines that the current reduced-order model has undergone physical drift, that is, although the model prediction results are continuous on the geometric manifold, they have deviated from the physical conservation laws.

[0045] For the execution of adaptive model calibration, the procedure employs an online correction strategy based on local subspace rotation. Once a physical drift is detected, the system suspends the current real-time inference task and starts from the most recent... Extract flow field snapshot sequences from historical time steps to construct a local snapshot matrix. The system Perform incremental singular value decomposition and extract the previous... Individual principal correction mode The system calculates the corrected mode and the current base library mode. The principal angle is used to determine the rotation matrix of the basis function space. Finally, the base library update operation is performed. This realigns the basis function subspace to the physical evolution direction of the current flow field. This calibration process is performed in parallel in the background and takes no more than [time missing]. Milliseconds; after calibration, the system automatically switches to the new base library. Continue the simulation.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent hydraulic design simulation system based on digital twins, characterized in that, include: The parametric geometric feature extraction module is used to obtain the geometric control point coordinate vectors of the design object and convert them into geometric feature descriptors based on preset topological mapping rules. The reduced-order model database stores pre-computed sets of orthogonal basis functions for the flow field and discrete anchor point datasets. The orthogonal basis function set contains multiple linearly independent spatial mode vectors for the flow field. The discrete anchor point dataset contains reference mode weight coefficient vectors corresponding to several reference geometric states, as well as modal sensitivity Jacobian matrices corresponding to each reference geometric state. The modal sensitivity Jacobian matrix is ​​used to characterize the rate of change of the partial derivatives of the flow field mode weight coefficients with respect to the geometric feature descriptors. The tangent space linear extrapolation module is used to calculate the geometric deviation vector between the real-time input geometric feature descriptor and the selected reference geometric state in the discrete anchor point dataset, and to perform matrix-vector multiplication based on the modal sensitivity Jacobian matrix and the geometric deviation vector to generate the modal coefficient increment vector. The tangent space linear extrapolation module linearly superimposes the modal coefficient increment vector with the reference modal weight coefficient vector to determine the target modal weight coefficient for the current working condition. The flow field reconstruction module is used to perform a weighted summation operation on the target modal weight coefficients and the orthogonal basis function set of the flow field, generate reconstructed flow field data, and map it onto the surface of the design object.

2. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The system also includes: an inverse sensitivity analysis module, used to calculate the gradient vector of the preset hydraulic performance index function relative to the geometric feature descriptor based on the modal sensitivity Jacobian matrix and using the chain rule after the target modal weight coefficients are output by the tangent space linear extrapolation module; and an optimization guidance display module, used to back-map the gradient vector to the coordinate space of the geometric control points and generate vector identifiers in the design view that indicate the direction of the optimized displacement of the geometric control points; the vector identifiers are used to characterize the influence weight of each geometric control point on the hydraulic performance index and the direction of improvement.

3. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The tangent space linear derivation module uses the following formula to perform matrix-vector multiplication operations: ,in, For target modal weighting coefficients, As the baseline modal weighting coefficient, Here is the modal sensitivity Jacobian matrix. The geometric deviation vector is used; matrix-vector multiplication is based on the linear approximation of the manifold tangent plane to determine the evolution of the flow field modes as geometric parameters change.

4. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The system also includes: a geometric orthogonality verification module, used to store the reference orthogonal basis matrix of the geometric feature space and calculate the orthogonal residual components of the geometric feature descriptors input in real time outside the subspace spanned by the reference orthogonal basis matrix; a deduction effectiveness control module, used to monitor the energy ratio of the orthogonal residual components in real time; when the energy ratio exceeds a preset threshold, it is determined that the current design object is outside the effective computational domain, the output of the flow field reconstruction module is blocked, and a trigger signal is generated to call the background numerical calculation unit to perform the full physical equation solution.

5. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The system also includes: a sensor data interface for acquiring real-time monitoring data from physical sensors deployed at specific spatial coordinates of the physical hydraulic engineering project; an observation matrix construction unit for extracting corresponding basis function components from a reduced-order model database based on the spatial coordinates of the physical sensors, constructing a sparse observation matrix, which defines the linear relationship between modal coefficients and sensor monitoring data; and a flow field state inversion module for constructing an overdetermined linear equation system based on the sparse observation matrix and real-time monitoring data, solving the linear equation system using the least squares method, and determining the measured modal weight coefficients characterizing the current flow field state of the physical engineering project.

6. The intelligent water conservancy design simulation system based on digital twins according to claim 5, characterized in that, The system also includes: an incremental subspace update module, used to calculate the orthogonal projection residual vector of the calibrated flow field data on the orthogonal basis function set when the calibrated flow field data is acquired; and a basis function library expansion unit, used to perform Schmitt orthogonalization and normalization on the orthogonal projection residual vector when the magnitude of the orthogonal projection residual vector exceeds a preset update threshold, generate new incremental orthogonal basis functions, and append and store them to the reduced-order model database.

7. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The flow field orthogonal basis function set is a set of standard orthogonal bases extracted by performing eigenorthogonal decomposition or singular value decomposition on the flow field snapshot matrix covering the design domain; the modal sensitivity Jacobian matrix is ​​a sensitivity matrix obtained by solving the adjoint equation of the fluid dynamics control equation or by pre-calculating it using the central difference method.

8. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The tangent space linear extrapolation module is also used to select multiple neighboring reference geometric states in the discrete anchor point dataset that are closest to the Euclidean distance of the geometric feature descriptor of the real-time input; perform local tangent space extrapolation based on the modal sensitivity Jacobian matrix corresponding to each neighboring reference geometric state, and perform weighted averaging on multiple extrapolation results based on the reciprocal of the Euclidean distance to determine the target modal weight coefficients.

9. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The parametric geometric feature extraction module includes a topology mapping unit, which is used to map design objects with different mesh topologies into a vector space with a uniform dimension. The geometric feature descriptor is the shape deformation feature vector after removing the rigid body displacement and rotation components.

10. The intelligent water conservancy design simulation system based on digital twins according to claim 1, characterized in that, The flow field reconstruction module includes a three-dimensional scalar field visualization unit. This three-dimensional scalar field visualization unit is used to map the pressure field data, velocity field data and turbulent kinetic energy data in the reconstructed flow field data into different color and transparency attributes, and overlay them on the surface of the three-dimensional geometric model of the design object to form a visualization model of the global flow field distribution.

Images