A water bucket intelligent optimization method and device

Abstract

The application provides a water bucket intelligent optimization method and device, and belongs to the field of water bucket optimization. The method solves the problems of poor water bucket optimization efficiency, precision and engineering practicability of the existing mode. A water bucket intelligent optimization method comprises the following steps: obtaining a real CFD sample set containing multiple groups of water bucket geometric parameters and corresponding runner hydraulic efficiency; constructing a literature distillation knowledge base; constructing a water bucket geometric parameter physical feasible region; sampling to generate candidate samples in the physical feasible region, and screening to obtain effective candidate samples; predicting the effective candidate samples by using a multi-teacher integrated model, and removing abnormal samples; building a mechanism perception residual proxy model, adjusting the model by combining pseudo-label pre-training with a real CFD sample set, and adopting residual truncation calibration; performing risk stratification and diversity deduplication on the candidate samples, and outputting a CFD simulation verification optimal sample batch. The method is mainly used for water bucket optimization.

Description

Technical Field

[0001] This invention belongs to the field of water bucket optimization, and in particular relates to a water bucket intelligent optimization method and device. Background Technology

[0002] Bucket turbine units are widely used in high-head, low-flow hydropower generation scenarios. As the core flow-through component of the runner, the geometric parameters of the bucket directly affect the runner's hydraulic efficiency, unit operation stability, and overall power generation performance. Precisely optimizing the bucket's geometric structure is key to reducing hydraulic losses and improving the unit's energy utilization rate.

[0003] Current optimization of water bucket structures largely relies on CFD simulation iterations, which iteratively select the optimal parameters by simulating different geometric parameter schemes one by one. This approach has significant drawbacks. On the one hand, CFD simulations consume a lot of computing power and have long cycles, resulting in low efficiency in large-scale parameter optimization, difficulty in covering the complete parameter domain, and a tendency to miss optimal structural schemes. On the other hand, traditional optimization relies solely on pure simulation data iteration without considering engineering constraints such as hydraulic mechanisms, geometric assembly, and operational risks. Parameter optimization lacks clear boundaries, easily generating a large number of physically infeasible and engineering-unsuitable invalid samples, unnecessarily increasing simulation costs. Furthermore, traditional surrogate models have insufficient ability to fit the physical mechanisms of complex flow fields in water buckets, exhibiting poor prediction accuracy and stability, and are unable to effectively identify abnormal samples, failing to meet the requirements of high-precision structural optimization, thus limiting the performance optimization effect of water buckets.

[0004] Existing data-driven optimization schemes rely solely on simulation sample modeling and lack support from domain theoretical knowledge and engineering constraints. They suffer from problems such as ambiguous definition of the physical feasible domain of parameters and lack of sample selection and risk verification mechanisms. They generally suffer from poor sample validity, weak model generalization ability, and insufficient reliability of optimization results. They cannot balance the efficiency, accuracy, and engineering practicality of bucket turbine optimization, thus restricting the performance upgrade and engineering application of bucket turbine units. Summary of the Invention

[0005] In view of this, the present invention aims to propose a water bucket intelligent optimization method and device to solve the problems of poor efficiency, accuracy and engineering practicality of existing water bucket optimization methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: According to a first aspect of the present invention, a water bucket intelligent optimization method is provided, comprising the following steps:

[0007] Obtain a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies;

[0008] Based on publicly available literature in the field, a literature distillation knowledge base containing geometric constraints, manifold constraints, and risk stratification rules is constructed through large-scale model knowledge distillation.

[0009] By combining real CFD sample sets with a literature distillation knowledge base, a physical feasible domain for the geometric parameters of the water bucket is constructed.

[0010] Candidate samples are generated by sampling within the physically feasible domain, and valid candidate samples are obtained by constraint verification and screening through literature distillation knowledge base;

[0011] A multi-teacher ensemble model is used to predict effective candidate samples, and a weighted pseudo-label is generated by combining the credibility index. Abnormal samples are removed based on the model prediction discrepancies.

[0012] A mechanism-aware residual proxy model was built. The model training was completed by combining pseudo-label pre-training with adjustment of real CFD sample sets, and residual truncation calibration was adopted.

[0013] Based on the trained model, a constrained active acquisition function is constructed to perform risk stratification and diversity deduplication on candidate samples, and CFD simulation is used to verify the optimal sample batch.

[0014] Furthermore, the literature distillation knowledge base includes variable semantics, key geometric factors, proportional constraints, physical feasibility rules, and / or risk stratification rules.

[0015] Furthermore, the constraints for constructing the physical feasible region include variable positivity constraints, gap geometry constraints, dimensionless geometric scaling constraints, and real sample manifold distance constraints.

[0016] Furthermore, the dimensionless geometric proportion constraint constructs a valid interval based on the quantiles, interquartile ranges, and extrapolation expansion coefficients of the real CFD samples, and combines engineering conditions for truncation correction; wherein, the manifold distance constraint generates manifold confidence that is negatively correlated with distance based on the standardized least-norm distance between the candidate sample and the real CFD sample, and quantifies the sample distribution fit.

[0017] Furthermore, during the multi-teacher model prediction, dynamic adaptive model weights are generated by integrating the model's own reliability with the candidate's geometric and physical credibility features. These weighted integrations yield pseudo-labels, and high-uncertainty pseudo-label samples are eliminated by quantifying model prediction discrepancies.

[0018] Furthermore, the mechanism-aware residual proxy model uses a physically inspired baseline model or a traditional model baseline as the main trend, and compensates for prediction errors by fitting the residual learning model; the pseudo-label pre-training and real CFD sample adjustment respectively adopt a confidence-weighted loss function and a real sample loss function, which can be superimposed with a ranking enhancement loss function.

[0019] Furthermore, the constrained active acquisition function integrates the predicted mean, uncertainty exploration term, teacher disagreement penalty term, manifold distance penalty term, and constraint violation penalty term, sets a truncation upper limit for model uncertainty to avoid invalid extrapolation; similar samples are eliminated by standardizing the geometric distance of candidate samples, and the optimal sample batch is generated by combining batch diversity and acquisition score for CFD simulation verification.

[0020] According to a second aspect of the present invention, a water bucket intelligent optimization device is provided, comprising:

[0021] The acquisition module is used to acquire a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies;

[0022] The knowledge base construction module is used to construct a literature distillation knowledge base based on publicly available literature in the domain, which includes geometric constraints, manifold constraints, and risk stratification rules through large model knowledge distillation.

[0023] The physical feasible region construction module is used to construct the physical feasible region of the water bucket geometry parameters by combining real CFD sample sets and literature distillation knowledge base;

[0024] The filtering module is used to sample and generate candidate samples within the physical feasible domain, and then filter them to obtain valid candidate samples through literature distillation knowledge base constraint verification.

[0025] The prediction module is used to predict effective candidate samples using a multi-teacher ensemble model, generate weighted pseudo-labels by combining the credibility index, and remove abnormal samples based on the model prediction discrepancies.

[0026] The model training module is used to build a mechanism-aware residual proxy model. The model training is completed by combining pseudo-label pre-training with adjustment of real CFD sample sets, and residual truncation calibration is adopted.

[0027] The output module is used to construct a constrained active acquisition function based on the trained model, perform risk stratification and diversity deduplication on candidate samples, and output the optimal sample batch for CFD simulation verification.

[0028] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the water bucket intelligent optimization method as described above.

[0029] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent optimization method for water buckets.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. This method combines real CFD simulation samples with knowledge from relevant literature to construct the physical feasible region of the bucket geometry parameters. It relies on geometric constraints, manifold constraints, and risk stratification rules to complete sample verification and screening, effectively avoiding the problems of unbounded parameter values ​​and poor sample rationality in traditional optimization processes. It significantly reduces invalid samples that violate hydraulic mechanisms and engineering realities, significantly reduces CFD simulation calculation costs, and improves overall optimization efficiency.

[0032] 2. This method employs a multi-teacher ensemble model combined with a reliability index to generate pseudo-labels and removes outlier samples, effectively improving the shortcomings of single-model predictions, such as large bias and insufficient stability, and enhancing the accuracy and reliability of sample labels. By combining this with a mechanism-aware residual surrogate model to complete training and residual truncation calibration, the model's ability to fit the complex hydraulic characteristics of the internal flow field of the water bucket is strengthened. This addresses the drawbacks of traditional surrogate models, such as low physical mechanism fit and weak generalization ability, effectively improving the accuracy of parameter performance prediction.

[0033] 3. This method achieves sample risk stratification and diversity deduplication through a constrained active acquisition function, which can accurately screen out the optimal simulation samples with high value and high reliability. It realizes efficient and high-precision optimization of the water bucket geometric parameters, effectively taps the performance improvement potential of the water bucket structure, and takes into account the theoretical rationality and engineering practicality of the water bucket optimization design. It has good engineering promotion and application value. Attached Figure Description

[0034] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0035] Figure 1 This is a flowchart of the intelligent optimization method for water buckets according to the present invention;

[0036] Figure 2 This is a structural block diagram of the intelligent water bucket optimization device described in this invention;

[0037] Figure 3 This is a structural block diagram of the electronic device described in this invention. Detailed Implementation

[0038] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0039] The following description, with reference to the accompanying drawings, describes the water bucket optimization method, apparatus, electronic device, and storage medium according to embodiments of the present invention.

[0040] Specifically, Figure 1 This is a flowchart illustrating a water bucket intelligent optimization method provided in an embodiment of the present invention.

[0041] like Figure 1 As shown, the intelligent optimization method for water buckets includes the following steps:

[0042] S1. Obtain a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies. Specifically, the input data is a small sample of real CFD (Computational Fluid Dynamics) data, where each sample includes bucket geometry parameters and the corresponding hydraulic efficiency:

[0043]

[0044] in, Represents the real CFD sample set; N represents the number of real CFD samples; i represents the sample index; x i Represents the geometric parameter vector of the i-th water bucket; η i This represents the turbine hydraulic efficiency corresponding to the i-th sample.

[0045]

[0046] Where B is the width of the bucket, H is the depth of the inner surface of the bucket, L is the length of the bucket, R is the size of the notch, α is the inlet angle of the water divider, β is the outlet angle, w1 and w2 are the guide line inclination angles, and η is the hydraulic efficiency of the runner.

[0047] S2, based on publicly available literature in the field, constructs a literature distillation knowledge base containing geometric constraints, manifold constraints, and risk stratification rules through large-scale model knowledge distillation. Specifically, literature on bucket optimization for impulse turbines, CFD numerical simulations, geometric design experience, and active safety / constraint learning is input into the large model to perform literature extraction and knowledge distillation, including variable semantics, key geometric factors, proportional constraints, physical feasibility rules, and risk stratification rules. Literature Distillation Knowledge Base The extraction method is as follows:

[0048]

[0049] in, It uses positive and geometric legality constraints to prevent non-physical candidates such as negative dimensions, negative angles, and gaps exceeding the limit; It is a geometric proportional constraint that constrains proportional relationships such as B / H, L / B, R / B, α / β, and w1 / w2. Represents the manifold constraint of the real samples, controlling the distance from candidate points to the distribution of real CFD samples; Risk stratification rules are used to differentiate between robust candidates, medium-to-high-risk candidates, high-risk extrapolation candidates, and extreme audit candidates. It should be noted that the large model does not generate CFD labels, predict hydraulic efficiency, or replace the surrogate model. The constructed literature distillation knowledge base integrates multi-dimensional refined constraint rules. Through positive and geometric validity constraints, it completely eliminates non-physical parameter schemes such as negative dimensions, negative angles, and gap out-of-bounds errors, ensuring the physical validity of candidate samples from the source. Fixed geometric ratio constraints standardize the proportional relationships of key water tank structures, conforming to conventional engineering design configurations and avoiding hydraulic distortion problems caused by structural proportion imbalances. Simultaneously, relying on real sample manifold constraints, it strictly controls the data distribution distance between candidate samples and real CFD samples, ensuring that optimized samples closely match the characteristics of the real flow field distribution. Combined with risk stratification rules to perform gradient hierarchical division of candidate samples, it can accurately identify samples with different optimization risk levels, achieving differentiated screening and control. Unlike conventional large-model-assisted optimization schemes, this method strictly defines the application boundaries of the large model, relying solely on it to extract domain knowledge and construct rules. It does not generate simulation labels, participate in hydraulic efficiency prediction, or replace the proxy model to complete performance fitting work. This avoids the inherent defects of large models, such as illusory predictions and large fitting biases. The entire process uses real CFD sample data as the core and physical constraint rules as the boundaries. While making full use of prior domain knowledge, it ensures the authenticity, objectivity, and engineering credibility of the overall optimization results.

[0050] S3. Combining the real CFD sample set and the literature distillation knowledge base, the physical feasible region of the conch geometry parameters is constructed. Specifically, based on the distribution of real CFD samples and the literature distillation knowledge base, the physical feasible region of candidate conch geometries is constructed. The physical feasible region includes variable positivity constraints, gap geometry constraints, dimensionless scaling constraints, and real sample manifold distance constraints. The formulas for variable positivity constraints and gap geometry constraints are as follows:

[0051]

[0052] The first set of inequalities represents positive variable constraints, while the second set represents gap geometric constraints. If R has different physical meanings in specific engineering designs, 0 < R < B can be replaced with the corresponding legal engineering range.

[0053] Dimensionless geometric proportion characteristic formula as follows:

[0054]

[0055] Here, x represents a candidate geometric sample. Proportional features can constrain the relationships between candidate geometric combinations, avoiding situations where individual variables are within range but the overall geometric proportions are distorted. Furthermore, each sample (e.g., the j-th proportional feature) is required to... It conforms to the proportional constraints based on quantiles and IQR (Interquartile Range):

[0056]

[0057] Among them, Q low,j and Q high,j Let IQR represent the lower and upper quantiles of the j-th proportionality feature in the real CFD sample, respectively. j λ represents the interquartile range of the j-th proportional feature; λ is the extrapolation expansion coefficient, used to control the extent to which the feasible region is expanded relative to the true sample distribution. If a proportional feature has positive or upper bound requirements, it should be further truncated to the engineering feasible range based on this interval. Its manifold distance constraint is the standardized minimum distance from the candidate point to the true sample manifold:

[0058]

[0059] Where d(x) represents the standardized minimum distance from candidate point x to the real sample manifold; x i represents the geometric vector of the i-th real CFD sample; s represents the scale vector of each geometric variable, which can be the standard deviation, IQR, or engineering allowable scale; ||·||2 represents the L2 norm.

[0060] Manifold reliability is defined based on distance:

[0061]

[0062] Where q(x) represents the manifold confidence of the candidate point, and the larger the value, the closer the candidate point is to the true CFD sample distribution; τ d This is a manifold distance scale hyperparameter used to control the rate at which confidence decays with distance.

[0063] By integrating the distribution characteristics of real CFD sample data with prior knowledge from literature distillation, a multi-dimensional and quantifiable physical feasible domain for the bucket geometry parameters is established. Precise and compliant definition of the parameter space is achieved through multi-level structured constraints. By setting positive variable constraints and gap geometry constraints, the physical value range of each structural parameter is strictly limited in the form of quantitative inequalities. The legal value range can be adaptively adjusted according to the actual engineering definition of key parameters such as the bucket radius, fundamentally eliminating invalid configurations that violate physical construction laws, such as negative parameters and structural out-of-bounds errors, thus ensuring the basic structural legitimacy of all candidate geometric samples.

[0064] Meanwhile, the introduction of dimensionless proportional characteristic constraints and a quantile-IQR-based interval narrowing mechanism overcomes the limitations of traditional single-variable interval constraints. It provides overall coupled constraints for multi-dimensional geometric proportions, effectively addressing the problem of distorted overall configuration proportions and poor structural matching despite compliant individual parameter values. The constraint interval is constructed based on the quantile distribution and interquartile range of real samples, and the feasible region expansion range can be controllably adjusted through extrapolation expansion coefficients, balancing the exploratory nature and stability of sample optimization. Furthermore, adaptive interval truncation can be achieved based on the engineering boundary characteristics of various proportional features, further aligning with actual design specifications.

[0065] Furthermore, a quantitative evaluation system based on standardized minimum manifold distance is constructed to accurately characterize the degree of fit between candidate geometric samples and real CFD sample data manifolds. The quality of candidate samples is quantified using manifold reliability metrics. This constraint method uses the distribution patterns of real samples as a benchmark, relies on scale vector normalization of geometric variable deviations to accurately screen candidate configurations that fit the characteristics of the real flow field distribution, and quantifies sample reliability through a distance decay mechanism, effectively suppressing the mechanistic distortion problem caused by excessive parameter extrapolation. This feasible region construction method achieves a deep integration of empirical rules, engineering boundaries, and real data distribution, ensuring that the optimization parameter space possesses physical completeness, engineering standardization, and data authenticity. This significantly improves the overall quality of subsequent surrogate model training samples, providing solid data and mechanistic support for high-precision, highly robust structural optimization.

[0066] S4 involves sampling and generating candidate samples within the physically feasible region, which are then filtered and validated using a literature distillation knowledge base constraint to obtain valid candidate samples. Specifically, candidate generation methods may include Latin hypercube sampling, real sample neighborhood perturbation, trust-region local expansion sampling, physical boundary sampling, table generation model candidates, and a pre-built small-sample augmentation candidate pool.

[0067]

[0068] in, Indicates the generated candidate bucket geometry parameters; p gen (x) represents the candidate geometric generation distribution; ∈ K litThis indicates that candidates must meet the physical feasibility, proportion, and risk constraints given in the literature distillation knowledge base. Within the physically feasible domain, a multi-strategy fusion candidate sample generation mechanism is employed, integrating Latin hypercube sampling, real sample neighborhood perturbation, trust region local expansion sampling, physical boundary sampling and table generation models, and small-sample enhanced candidate pooling. This approach balances the sampling requirements of global parameter space exploration, refined local optimum mining, and effective boundary region coverage. Compared to the shortcomings of traditional single-sampling methods, such as uneven sampling coverage, weak local optimization ability, missing boundary samples, and poor generalization in small-sample scenarios, this method's multi-sampling mode achieves uniform traversal of the global parameter space, refined densification of high-quality sample neighborhoods, and effective placement of points in physical boundary regions. This greatly enriches the diversity and spatial coverage of candidate samples, effectively avoiding getting trapped in local optima during the optimization process. All generated candidate samples are uniformly subjected to full-dimensional constraints and verification using the literature distillation knowledge base. This ensures that all output candidate samples strictly conform to physical feasibility rules, dimensionless proportional ratio specifications, and risk stratification constraint systems. This guarantees that all valid candidate samples participating in model training and simulation verification fall within a physically reasonable, compliant, and risk-controllable optimization space. This screening-before-modeling mechanism eliminates invalid samples with abnormal mechanisms, distorted configurations, and high extrapolation risks from the source of the samples. It completely avoids the sample redundancy, simulation waste, and model contamination problems caused by traditional random sampling and blind sampling, further improving the overall purity and effective utilization rate of the optimized samples. This lays a high-quality sample foundation for subsequent high-precision surrogate model training and efficient and accurate optimization.

[0069] S5 employs a multi-teacher ensemble model to predict valid candidate samples, generates weighted pseudo-labels based on a credibility index, and removes outlier samples based on model prediction discrepancies. Specifically, for candidate geometries filtered by constraints, the multi-teacher model is invoked for prediction. Teacher models may include TabPFNv2 regression model, GBM, Random Forest, ExtraTrees, KRR, GPR, physics-inspired baseline models, and mechanism-aware residual models. The m-th teacher model predicts the following output:

[0070]

[0071] in, This indicates that the m-th teacher model is a candidate Efficiency prediction; f m Let m represent the m-th teacher model; M represents the total number of teacher models. Candidate-related teacher weights.

[0072]

[0073] in, Represents the normalized teacher weights; r mIndicates the reliability of the m-th teacher in the in-process verification; Indicates the confidence level of the candidate manifold; Indicates the candidate's physical feasibility score. Reliability-weighted pseudo-label. :

[0074]

[0075] Simultaneously calculate teacher disagreements:

[0076]

[0077] Where, r m Let q(x) represent the reliability of the teacher model in the trade-off validation, q(x) represent the confidence of the candidate manifold, p(x) represent the physical feasibility, and σ represent the reliability of the teacher model in the trade-off validation. T (x) represents teacher model divergence. If teacher divergence exceeds a threshold, the weight of the candidate pseudo-label is reduced or it is removed. Pseudo-labels are only used for pre-training or low-weighted auxiliary learning and are not used as true CFD labels. By integrating the fitting advantages and generalization characteristics of different models, the shortcomings of single machine learning models, such as limited fitting dimensions, fixed biases, and insufficient adaptability to complex hydraulic mapping relationships, are overcome, significantly improving the compatibility and robustness of the overall prediction system. Abandoning the fixed-weight ensemble prediction method, a dynamic normalized weight mechanism is constructed that integrates the reliability of the model itself, the credibility of the sample manifold, and the physical feasibility score. Prediction weights are allocated according to the differences in the compromise validation performance of each teacher model. Combined with the fit of sample data distribution and physical compliance, weighted pseudo-labels are accurately generated, enabling the pseudo-label results to objectively balance the model fitting accuracy and the physical rationality of the samples. This effectively solves the problem of limited prediction accuracy caused by the average weighting and lack of distinction between superior and inferior models in traditional ensemble prediction.

[0078] We introduce a teacher disagreement quantification index based on model reliability weighting. By accurately identifying abnormal candidate samples with poor prediction consistency and high uncertainty through the dispersion of multi-model prediction results, we adopt weight suppression or direct elimination strategies for high disagreement samples that exceed the threshold. We filter out high uncertainty samples from both data and prediction dimensions, effectively avoiding label noise and training bias caused by single-model prediction errors and multi-model fitting conflicts, and ensuring the cleanliness and credibility of pre-training data.

[0079] By strictly defining the application boundaries of pseudo-labels, using them only for model pre-training and low-weighted assisted learning, and not replacing real CFD sample labels as training ground values, the technical mechanism avoids the fitting distortion and mechanism deviation risks inherent in pure data pseudo-labels. Pseudo-label-assisted pre-training can effectively expand the model's prior learning information and optimize the model's initial parameters. Then, fine-tuning can be completed based on real CFD samples, balancing model training efficiency with the absolute reliability of the predicted ground values.

[0080] S6. A mechanism-aware residual surrogate model is built. Model training is completed through pseudo-label pre-training combined with adjustments to a real CFD sample set, and residual truncation calibration is employed. Specifically, a mechanism-aware residual surrogate model is constructed, using a physically inspired baseline or a traditional model baseline as the main trend, and a deep residual network or residual head is used to learn the residual error. Mechanism-aware residual surrogate model.

[0081]

[0082] Where η phys (x) represents the physical-inspired baseline or the baseline prediction of the traditional model; Δ represents the residual learning model with parameter θ; φ(x) represents the input features consisting of the original geometric variables, dimensionless proportional features, and literature distillation constraint features.

[0083] Training is divided into pseudo-label pre-training and fine-tuning phases. In pre-training, candidate pseudo-labels and their trustworthy pseudo-labels are generated, and the loss is:

[0084]

[0085] Among them, L pre This represents the loss from pseudo-label pre-training; This indicates the j-th candidate to be generated; This indicates its pseudo-label; ω j This indicates the credibility weight of the candidate.

[0086] Correspondingly, the real CFD fine-tuning loss L real for:

[0087]

[0088] Where, x i and η i These represent the geometric parameters and efficiency labels of real CFD samples. Additionally, a ranking-enhanced loss is added.

[0089]

[0090] Among them, L rank λ represents the optional sorting enhancement loss, used to improve the ability to efficiently sort candidate results. rank λ represents the weight of the ranking loss. When ranking enhancement is not used, λ rank It can be set to 0.

[0091] After training is complete, the residual calibration output is... :

[0092]

[0093] Where γ represents the residual calibration coefficient; c represents the residual cutoff threshold; clip(z, -c, c) means restricting the residual z to the range of [-c, c] to reduce the risk of numerical instability when extrapolating small samples.

[0094] Model evaluation was performed using locked-partition cross-validation. All candidate generation, pseudo-label construction, standardization, teacher training, and surrogate model training were completed within the training fold, while the testing fold was used only for extrapolation evaluation of real CFD samples.

[0095]

[0096] Among them, R 2 The coefficient of determination is used to evaluate the overall predictive fit. Let be the out-of-bounds predicted value for the i-th sample; is the true mean efficiency; n is the number of evaluation samples; RMSE is used to evaluate the prediction error scale; MAE is used to evaluate the mean absolute prediction bias. Spearman is used to evaluate ranking relevance; Hit@k (hitting ability of the top k high-efficiency candidates) and Top-koverlap are used to evaluate the ability to identify high-efficiency candidates; Prediction hash, Fixed seed, and Locked folds are used for experimental reproducibility auditing.

[0097] The mechanism-aware residual surrogate model employs a hierarchical fitting architecture with a physical baseline guiding the trend and a residual network compensating for errors, distinguishing it from the black-box fitting mode of traditional pure data-driven models. The model uses a physically inspired baseline to construct the main predicted trend of hydraulic efficiency, ensuring that the basic prediction results align with the inherent mechanisms of fluid mechanics. Simultaneously, a residual learning module specifically fits the refined flow field errors and structural coupling deviations that the physical baseline cannot cover. Furthermore, it integrates original geometric variables, dimensionless proportional features, and literature distillation constraint features to construct a multi-dimensional input system, enabling the model to possess physical mechanism awareness. This effectively addresses the problems of missing mechanisms and the disconnect between fitting logic and hydraulic laws in traditional surrogate models, significantly improving the model's accuracy in representing the mapping relationship between complex funnel configurations and hydraulic efficiency.

[0098] A two-stage hierarchical training strategy of pseudo-label pre-training and real-sample fine-tuning is adopted, coupled with a multi-dimensional composite loss function to constrain model convergence. In the pre-training stage, a pseudo-label loss function with confidence weights is used to quickly optimize the initial parameters of the model using a massive number of high-confidence candidate samples, fully exploring the implicit patterns in the parameter space and compensating for the shortcomings of insufficient real CFD sample quantity and inadequate model convergence under small sample conditions. In the fine-tuning stage, real CFD sample loss is used to accurately correct the parameters, thoroughly correcting the subtle fitting biases caused by pseudo-labels. Simultaneously, a configurable ranking enhancement loss function is introduced to strengthen the model's ability to rank and identify high-efficiency, high-quality candidate samples, balancing the model's fitting accuracy with the targeted optimization, effectively solving the problems of limited convergence and weak ability to identify high-quality samples in traditional single-loss training models.

[0099] Meanwhile, this method introduces a residual truncation calibration mechanism, which constrains the model residual output by using residual calibration coefficients and a fixed truncation threshold. This limits the residual fluctuation range within a reasonable range, effectively suppressing residual surges and prediction instability in parameter extrapolation, boundary samples, and sparse regions of small samples. It avoids the defects of traditional data models that are prone to extrapolation divergence and extreme value prediction distortion, and significantly improves the stability and robustness of the model in full parameter space prediction.

[0100] A locked-partition cross-validation system is employed to quantitatively evaluate model performance. Training and testing folds are strictly separated, and all data processing and modeling workflows, including candidate generation, pseudo-label construction, and model training, are completed within the training fold. Extrapolation performance is evaluated only through independent testing folds. Multi-dimensional metrics, including R², RMSE, MAE, Spearman correlation coefficient, Hit@k, and Top-k overlap, are used to comprehensively quantify the model's fitting accuracy, error level, ranking ability, and high-quality sample selection capability. Furthermore, relying on a fixed random seed, prediction hashing, and locked-partition partitioning mechanism, the modeling process is traceable, reproducible, and auditable, effectively avoiding model evaluation distortion caused by random experimental errors. This ensures the objectivity and fairness of model performance evaluation and further enhances the engineering reliability and technical standardization of this optimization method.

[0101] S7 constructs a constrained active acquisition function based on the trained model, performs risk stratification and diversity deduplication on candidate samples, and outputs the optimal sample batch for CFD simulation verification. Specifically, the constrained active acquisition function is constructed based on the prediction mean, fold uncertainty, teacher divergence, manifold distance, and physical constraint penalty of the mechanism-aware surrogate model.

[0102] Constrained active data acquisition function:

[0103]

[0104] Where a(x) represents the constrained active acquisition function; μ(x) represents the predicted mean of candidate x; σ(x) represents the prediction uncertainty of the fold model; σ max κ represents the upper bound of uncertainty cutoff, used to prevent meaningless high UCBs (Upper Confidence Bounds, i.e., upper confidence bounds or optimistic confidence bounds) arising purely from huge uncertainties; κ represents the uncertainty exploration coefficient; d(x) represents the manifold distance; σ T (x) represents teacher disagreement; V(x) represents penalty for violating physical or literature constraints; λ d , λ T and λ v The weights are respectively for manifold distance penalty, teacher disagreement penalty, and constraint violation penalty.

[0105] Based on the acquisition function and risk rules, four candidate categories are output. List A consists of robust CFD priority verification candidates, typically with high manifold confidence and low uncertainty; List B consists of medium-to-high risk target candidates, usually supported by UCB to achieve a lower target efficiency threshold; List C consists of high-risk extrapolation hypothesis candidates, physically feasible but with more pronounced extrapolation; List D consists of extreme audit candidates, used only as stress test or appendix candidates, not as the main conclusion. To avoid overly similar recommended candidates, batch diversity selection is introduced. This step, among candidates satisfying physical constraints and risk stratification rules, comprehensively considers acquisition scores and geometric diversity to generate the next round of CFD verification batches. The standardized geometric distance between candidate i and candidate j is used to determine the next round of CFD verification batches. :

[0106]

[0107] Where s represents the geometric variable scale vector. If If the scores are below the set threshold, the two candidates are considered too similar, and the candidate with the higher collection score or the more reasonable risk level should be retained in the batch selection.

[0108] The final selected candidate batch for the next round of CFD validation is labeled B*.

[0109]

[0110] Where B represents the candidate batch set; Div(B) represents the measure of geometric diversity within a batch; λ div This represents the diversity weight. The final output is the next round of CFD validation batch, which includes several robust candidates, medium-to-high risk candidates, high-risk extrapolation candidates, and optional limit audit candidates.

[0111] The final risk stratification candidate output table is as follows:

[0112] List A Robust CFD priority validation candidates Low to medium risk; high manifold confidence; low uncertainty; suitable for priority CFD verification. List B High-risk breakthrough candidate UCB supports the 86% target; the mean may not reach the target; suitable for medium- to high-risk validation. List C High-risk extrapolation hypothesis candidates Physically feasible but more obvious by extrapolation; used to explore potentially efficient regions. List D Extreme audit candidate This is only intended as a candidate for stress testing or as an appendix, and is not considered as the main conclusion.

[0113] This method, based on a trained mechanism-aware residual surrogate model, constructs a multi-dimensionally coupled constrained active acquisition function, overcoming the limitations of traditional single upper confidence bound acquisition strategies that only consider the prediction mean and uncertainty. This acquisition function integrates multiple quantification factors, including the model prediction mean, folded uncertainty, multi-teacher prediction divergence, manifold distance, and physical constraint penalties. It avoids ineffective exploration caused by excessive uncertainty by using an upper uncertainty truncation limit, and simultaneously configures multi-dimensional weight coefficients to achieve a dynamic balance between exploration and utilization. By introducing manifold distance penalties, teacher divergence penalties, and constraint violation penalties, the method accurately suppresses the acquisition weights of candidate samples with high extrapolation bias, high prediction divergence, and violations of physical rules. This ensures that the sample acquisition process is always anchored within the physically feasible region and the model's high-confidence prediction interval, eliminating the selection of low-value, high-risk simulation samples at the sampling decision level and significantly improving the effective value density of CFD simulation batches.

[0114] Based on refined risk stratification rules, candidate samples are divided into four tiered levels: robust priority verification candidates, medium-to-high risk target candidates, high-risk extrapolation candidates, and extreme audit candidates. This achieves a differentiated and structured sample screening and verification system. Through this hierarchical classification mechanism, high-precision, highly reliable, and robust optimization samples are retained for main performance iteration and improvement. At the same time, risky candidates with appropriate extrapolation are included to expand parameter exploration boundaries. Furthermore, extreme audit candidates are used to complete stress testing of the solution. This balances the stability of optimization iteration with exploratory innovation, addressing the industry pain points of traditional active sampling, which is either overly conservative and unable to escape local optima, or blindly extrapolating and resulting in uncontrollable optimization outcomes.

[0115] A batch geometric diversity constraint screening mechanism is introduced. By quantifying the configurational differences among candidate samples through standardized geometric distance, redundant candidates with similarity exceeding a threshold are selectively eliminated. This effectively avoids the problems of sample clustering, severe configurational homogenization, and high overlap in the effective exploration space that exist in traditional batch sampling. By balancing the optimality of the acquisition score with the overall geometrical diversity of the batch, the mechanism maximizes the coverage parameter optimization space while ensuring excellent comprehensive performance of candidate samples in a single batch, thereby improving the information gain efficiency of a single round of CFD simulation iteration.

[0116] In summary, this method, through a collaborative mechanism of risk-layered management, multi-factor constraint data collection, and diversity deduplication, intelligently selects the optimal CFD verification batch that balances high performance potential, high reliability, and high spatial coverage. This achieves controllability, precision, and efficiency in the bucket parameter optimization iteration process. While significantly reducing the consumption of ineffective simulation computing power and iteration cycle, it continuously expands the optimization search space and steadily explores the upper limit of structural performance, further improving the optimality and engineering adaptability of the bucket geometry parameter optimization results.

[0117] Next, refer to the appendix. Figure 2 This invention describes a water bucket intelligent optimization device proposed according to an embodiment of the present invention.

[0118] Figure 2 This is a block diagram of the intelligent water bucket optimization device according to an embodiment of the present invention.

[0119] like Figure 2 As shown, the intelligent water bucket optimization device 100 includes:

[0120] The acquisition module 101 is used to acquire a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies;

[0121] The knowledge base construction module 102 is used to construct a literature distillation knowledge base based on open literature in the domain, which includes geometric constraints, manifold constraints and risk stratification rules through large model knowledge distillation.

[0122] The physical feasible region construction module 103 is used to construct the physical feasible region of the water bucket geometric parameters by combining the real CFD sample set and the literature distillation knowledge base.

[0123] The filtering module 104 is used to sample and generate candidate samples within the physical feasible domain, and then filter them to obtain valid candidate samples through literature distillation knowledge base constraint verification.

[0124] Prediction module 105 is used to predict effective candidate samples using a multi-teacher ensemble model, generate weighted pseudo-labels by combining credibility index, and remove abnormal samples based on model prediction discrepancies.

[0125] The model training module 106 is used to build a mechanism-aware residual proxy model. The model training is completed by combining pseudo-label pre-training with adjustment of real CFD sample sets, and residual truncation calibration is adopted.

[0126] Output module 107 is used to construct a constrained active acquisition function based on the trained model, perform risk stratification and diversity deduplication on candidate samples, and output the optimal sample batch for CFD simulation verification.

[0127] It should be noted that the foregoing explanation of the embodiment of the intelligent optimization method for water buckets also applies to the intelligent optimization device for water buckets in this embodiment, and will not be repeated here.

[0128] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. The electronic device may include:

[0129] The memory 201, the processor 202, and the computer program stored on the memory 201 and capable of running on the processor 202.

[0130] When the processor 202 executes the program, it implements the intelligent optimization method for water buckets provided in the above embodiments.

[0131] Furthermore, electronic devices also include:

[0132] Communication interface 203 is used for communication between memory 201 and processor 202.

[0133] The memory 201 is used to store computer programs that can run on the processor 202.

[0134] The memory 201 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0135] If the memory 201, processor 202, and communication interface 203 are implemented independently, then the communication interface 203, memory 201, and processor 202 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0136] Optionally, in a specific implementation, if the memory 201, processor 202, and communication interface 203 are integrated on a single chip, then the memory 201, processor 202, and communication interface 203 can communicate with each other through an internal interface.

[0137] The processor 202 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0138] This invention also provides a computer program product, which, when executed by a processor, implements the above-described intelligent optimization method for water buckets.

[0139] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent water bucket optimization method.

[0140] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0141] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0142] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0143] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0144] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0145] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0146] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0147] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A water bucket intelligent optimization method, characterized in that, Includes the following steps: Obtain a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies; Based on publicly available literature in the field, a literature distillation knowledge base containing geometric constraints, manifold constraints, and risk stratification rules is constructed through large-scale model knowledge distillation. By combining real CFD sample sets with a literature distillation knowledge base, a physical feasible domain for the geometric parameters of the water bucket is constructed. Candidate samples are generated by sampling within the physically feasible domain, and valid candidate samples are obtained by constraint verification and screening through literature distillation knowledge base; A multi-teacher ensemble model is used to predict effective candidate samples, and a weighted pseudo-label is generated by combining the credibility index. Abnormal samples are removed based on the model prediction discrepancies. A mechanism-aware residual proxy model was built. The model training was completed by combining pseudo-label pre-training with adjustment of real CFD sample sets, and residual truncation calibration was adopted. Based on the trained model, a constrained active acquisition function is constructed to perform risk stratification and diversity deduplication on candidate samples, and CFD simulation is used to verify the optimal sample batch.

2. The method according to claim 1, characterized in that: The literature distillation knowledge base includes variable semantics, key geometric factors, proportional constraints, physical feasibility rules, and / or risk stratification rules.

3. The method according to claim 1, characterized in that: The constraints for constructing the physical feasible region include variable positivity constraints, gap geometry constraints, dimensionless geometric scaling constraints, and real sample manifold distance constraints.

4. The method according to claim 3, characterized in that: The dimensionless geometric proportion constraint constructs a valid interval based on the quantiles, interquartile ranges, and extrapolation expansion coefficients of the real CFD samples, and combines engineering conditions for truncation correction; wherein, the manifold distance constraint is based on the standardized least-norm distance between the candidate sample and the real CFD sample, generating a manifold confidence that is negatively correlated with the distance, and quantifying the sample distribution fit.

5. The method according to claim 1, characterized in that: When the multi-teacher model makes predictions, it integrates the model's own reliability with the candidate's geometric and physical credibility features to generate dynamic adaptive model weights, which are then weighted and integrated to obtain pseudo-labels. High-uncertainty pseudo-label samples are then eliminated by quantifying the model's prediction discrepancies.

6. The method according to claim 1, characterized in that: The mechanism-aware residual proxy model uses a physically inspired baseline model or a traditional model baseline as the main trend, and compensates for prediction errors by fitting the residual learning model; the pseudo-label pre-training and real CFD sample adjustment respectively adopt the confidence-weighted loss function and the real sample loss function, which can be superimposed with the ranking enhancement loss function.

7. The method according to claim 1, characterized in that: The constrained active acquisition function integrates the predicted mean, uncertainty exploration term, teacher disagreement penalty term, manifold distance penalty term, and constraint violation penalty term, and sets a truncation upper limit for model uncertainty to avoid invalid extrapolation; similar samples are eliminated by standardizing the geometric distance of candidate samples, and the optimal sample batch is generated by combining batch diversity and acquisition score.

8. A water bucket intelligent optimization device, characterized in that, include: The acquisition module is used to acquire a real CFD sample set containing multiple sets of bucket geometry parameters and corresponding runner hydraulic efficiencies; The knowledge base construction module is used to construct a literature distillation knowledge base based on publicly available literature in the domain, which includes geometric constraints, manifold constraints, and risk stratification rules through large model knowledge distillation. The physical feasible region construction module is used to construct the physical feasible region of the water bucket geometry parameters by combining real CFD sample sets and literature distillation knowledge base; The filtering module is used to sample and generate candidate samples within the physical feasible domain, and then filter them to obtain valid candidate samples through literature distillation knowledge base constraint verification. The prediction module is used to predict effective candidate samples using a multi-teacher ensemble model, generate weighted pseudo-labels by combining the credibility index, and remove abnormal samples based on the model prediction discrepancies. The model training module is used to build a mechanism-aware residual proxy model. The model training is completed by combining pseudo-label pre-training with adjustment of real CFD sample sets, and residual truncation calibration is adopted. The output module is used to construct a constrained active acquisition function based on the trained model, perform risk stratification and diversity deduplication on candidate samples, and output the optimal sample batch for CFD simulation verification.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method as claimed in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, the computer program being used to cause the computer to perform the method as described in any one of claims 1-7.

Images