A method and system for deriving engineering structural parameters for budget-constrained design

Abstract

This invention discloses a method and system for deriving engineering structural parameters for cost-constrained design, relating to the fields of engineering design and engineering cost technology. The method includes: constructing a cost prediction model; generating an initial design influence parameter set through an external large language model; obtaining the predicted engineering cost value through the cost prediction model; iteratively optimizing the initial design influence parameter set using a covariance matrix adaptive evolutionary strategy algorithm combined with a rule engine to obtain an optimized design influence parameter set that meets the target engineering cost constraint; optionally performing design scheme analysis to identify unreasonable parameters and iteratively optimize them; and outputting the optimal engineering design influence parameter set and a design scheme report. This invention, through the assistance of a large language model in generating initial parameters and the evolutionary algorithm-driven reverse deduction, achieves automatic optimization of design influence parameters under target cost constraints, improving design efficiency and parameter rationality, and is applicable to cost-constrained design scenarios such as underground powerhouses of pumped storage power stations.

Description

Technical Field

[0001] This invention relates to the fields of engineering design and engineering cost technology, specifically to a method and system for deriving engineering structural parameters for budget-constrained design. Background Technology

[0002] Engineering cost limit design is a core aspect of engineering construction project control. The rationality of its design influencing parameters directly determines the effectiveness of project cost control, construction quality, and operational efficiency, and is of great significance to project initiation decisions, fundraising, and risk control.

[0003] Traditional cost-limited design methods rely heavily on designers' experience, combining historical project comparisons and cost quota calculations to determine design impact parameters, followed by manual cost calculation and adjustments. This approach has significant limitations: over-reliance on experience leads to strong subjectivity, making it difficult to quantify the coupling effects of multiple parameters; insufficient use of historical project data easily results in a disconnect between design impact parameters and target costs; parameter adjustments are cumbersome and inefficient, lacking systematic optimization logic and making it difficult to quickly find the optimal parameter combination.

[0004] In recent years, artificial intelligence technology has been gradually penetrating the field of engineering cost-limited design. Among them, intelligent computing methods have provided technical support for the construction of high-precision cost prediction models, while the progress of artificial intelligence technology in information retrieval and semantic understanding has opened up new avenues for in-depth mining of historical project data and intelligent evaluation of design schemes. The combination of the two has improved the scientificity and efficiency of cost-limited design to a certain extent.

[0005] However, in practical engineering applications, existing technical solutions still have the following prominent problems: First, the use of historical project data is inefficient, relying heavily on manual or simple keyword screening. It does not fully leverage the semantic understanding and information retrieval advantages of artificial intelligence technology, making it difficult to quickly discover matching similar project experiences and generate reasonable initial design impact parameters, resulting in low design efficiency, high optimization costs, and a tendency to produce deviations. Second, cost control and design impact parameter optimization are disconnected from each other and lack an effective feedback mechanism. They cannot drive dynamic parameter optimization with the target cost as the core, which easily leads to the problem of design impact parameters being out of touch with the target cost. Third, the level of intelligence and engineering is insufficient. The design relies too much on human experience and has a long cycle. Related research is mostly at the theoretical level and lacks deployable end-to-end systems and management functions such as parameter verification and process recording, resulting in poor practicality. Fourth, there is a lack of deep collaboration among various intelligent technologies. Functional modules such as cost prediction, parameter optimization, and scheme evaluation often operate in isolation and are simply superimposed, making it difficult to form an organically collaborative whole, which affects the overall effect and parameter accuracy of the budget design.

[0006] Therefore, there is an urgent need to develop a cost-limited design method and system that integrates the advantages of intelligent retrieval, intelligent computing and intelligent evaluation, can efficiently derive the optimal engineering design influence parameter set and support engineering deployment, and solve core problems such as inefficient data utilization, insufficient technology integration, poor cost design coordination and lack of system integration, so as to meet the needs of actual engineering applications. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for deriving engineering structural parameters for budget-constrained design, addressing problems such as inefficient use of historical project data in current engineering design, insufficient integration of large language models and optimization algorithms, poor coordination between design and cost, and a lack of deployable closed-loop budget-constrained design systems. To achieve the above objective, the technical solution is as follows: A method for deriving engineering structural parameters for budget-constrained design according to this invention includes the following steps: S1, Build and train the cost prediction model to form a cost prediction model package that can be directly loaded; S2. Based on the target project cost and known design influence parameters, an initial design influence parameter set is generated. The initial design influence parameter set is then used to predict the cost through the cost prediction model to obtain the project cost prediction value. The project cost prediction value is then compared with the target project cost. If the comparison result does not meet the preset error requirement, the initial design influence parameter set is adjusted and the cost prediction is re-executed. The comparison and adjustment are repeated until the preset error requirement is met to obtain the optimized design influence parameter set. S3: The user selects whether to perform design scheme analysis. If the user selects to perform the design scheme analysis, the optimized design influence parameter group is evaluated and the evaluation result is output to determine whether there are any unreasonable parameters. If there are unreasonable parameters, the user's selection instruction determines whether to perform iterative optimization on the corresponding design influence parameters. If so, the iterative optimization is completed and a new optimized design influence parameter group is obtained, and then the process proceeds to step S4. If there are no unreasonable parameters, the process proceeds directly to step S4. If the user selects not to perform the design scheme analysis or determines not to perform iterative optimization, the process proceeds directly to step S4. S4. Output the optimized design influence parameter set obtained in step S2 or S3 as the optimal engineering design influence parameter set that satisfies the target cost constraint; if the design scheme analysis is performed, a design scheme report can be output.

[0008] Preferably, the construction process of the cost prediction model of S1 is as follows: S11, construct a set of engineering cost influencing parameters, quantify the non-numerical influencing parameters in the set of engineering cost influencing parameters by ordinal encoding, and form an original feature set; collect historical engineering cost data, extract the values ​​of each influencing parameter in the original feature set, and extract the target cost data corresponding to the values ​​of each influencing parameter from the historical engineering cost data to construct an initial dataset; S12, The causal forest algorithm is used to perform causal inference analysis on the initial dataset, and key influencing parameters are determined accordingly; S13, the key impact parameters are standardized using a feature standardizer to obtain standardized key impact parameters; the target cost data corresponding to the key impact parameters are standardized using an engineering cost standardizer to obtain standardized target cost; the standardized key impact parameters and the standardized target cost together form a standardized dataset; S14, the standardized dataset is divided proportionally to obtain a training set and a test set; S15, A partial least squares regression dimensionality reducer is used to perform supervised dimensionality reduction on the standardized key influence parameters within the training set, with the standardized target cost as the supervised signal, to generate low-dimensional features; based on the low-dimensional features and the standardized target cost within the training set, a support vector regression prediction model is built. S16, using the negative mean square error corresponding to the test set as the optimization objective function, the optimization algorithm is used to jointly tune the support vector regression prediction model and the partial least squares regression dimensionality reducer to determine the optimal parameter combination; S17, combine the training set and the test set to form a full training set, retrain the support vector regression prediction model based on the optimal parameter combination, reconstruct the partial least squares regression dimensionality reducer and the support vector regression prediction model, and obtain four types of components: feature standardizer, engineering cost standardizer, partial least squares regression dimensionality reducer and support vector regression prediction model that match the optimal parameter combination. S18 performs serialization and solidification processing on the four types of components that match the optimal parameter combination, forming a cost prediction model package that can be directly loaded.

[0009] Preferably, the set of influencing parameters is divided into material price category, regional geology category, plant layout category and installed parameters category according to engineering attributes.

[0010] Preferably, in step S1, the training set and test set obtained by dividing the standardized dataset are allocated in an 8:2 ratio; the optimization algorithm is any one of Bayesian optimization algorithm, whale optimization algorithm, genetic algorithm, and particle swarm optimization algorithm.

[0011] Preferably, step S2 includes the following: S21, The target cost text data is transformed into vector data using an open-source model, and the vector data, the original target cost text, and the metadata used to describe the text attributes are stored together in the vector database. S22, receive the target engineering value and known design impact parameters input from the outside, call the external large language model through the application programming interface, retrieve the historical project data in the vector database, and generate the initial design impact parameter group; S23, the initial design influence parameter group is input into the cost prediction model package. After the feature standardization operation is completed by the feature standardizer, it is sequentially input into the partial least squares regression dimensionality reducer and the support vector regression prediction model to obtain the standardized prediction value. After the standardized prediction value is inversely transformed, the project cost prediction value is output. S24, calculate the relative error between the predicted project cost and the target project cost, and determine whether the relative error meets the preset limiting conditions; if it does, retain the initial design influence parameter group; if it does not, use the covariance matrix adaptive evolution strategy algorithm, in conjunction with the rule engine, to iteratively optimize the initial design influence parameter group until the relative error meets the preset limiting conditions, and output the optimized design influence parameter group, which includes unknown design influence parameters.

[0012] Preferably, the preset limiting condition is that the relative error between the predicted engineering cost and the target engineering cost is less than 5%; the iterative optimization stopping condition is that the number of iterations reaches 100 and there is a set of design influence parameters that meet the relative error requirement.

[0013] A system for deriving engineering structural parameters for cost-limited design, comprising: The user interaction layer is used to receive the target project cost and known design impact parameters, and output the optimal engineering design impact parameter set and design scheme report that satisfy the target cost constraint. The service scheduling layer is equipped with business scheduling routes to realize data interaction and command forwarding between the user interaction layer and the core function layer. The core functional layer communicates with the service scheduling layer and consists of a data receiving module, a large language model calling module, a cost prediction calculation module, a parameter iteration optimization module, and a parameter parsing and processing module, arranged sequentially according to the data flow. Each module establishes a data connection with the other through standardized interfaces. The data storage layer communicates with the core functional layer. The data storage layer has a built-in cost prediction model package, vector database, historical project data and relational database, providing data and model support for the core functional layer.

[0014] Preferably, the service scheduling route includes: The cost prediction route is used to receive multi-dimensional engineering feature parameters from the user interaction layer, call the locally loaded cost prediction model package through the cost prediction calculation module to perform cost prediction, and return the numerical results of the corresponding design influence parameter group. The parameter optimization route is used to receive the target cost constraints and known design influence parameter boundary conditions of the user interaction layer, trigger the covariance matrix adaptive evolution strategy algorithm through the parameter iteration optimization module to iteratively solve the appropriate design influence parameter set, and output the design influence parameter set to the parameter parsing route; The parameter parsing route is used to retrieve historical project data from the data storage layer and connect to the external large language model interface through the large language model calling module. The parameter parsing processing module performs parameter parsing on the generated design influence parameter group and generates corresponding text data, which is then output to the user interaction layer.

[0015] Preferably, the cost prediction model includes an integrated feature normalizer, an engineering cost normalizer, a partial least squares regression dimensionality reducer, and a support vector regression prediction model; the vector database stores vector data, original text, and metadata used to describe text attributes.

[0016] Preferably, the data receiving module is used to receive the target engineering value and known design influence parameters input by the user, and to perform a rationality check on the input parameters and output a verification prompt message; The large language model calling module connects to an external large language model to retrieve data and generate the initial design influence parameter set; The cost prediction calculation module loads the cost prediction model package and sequentially performs feature standardization, partial least squares regression dimensionality reduction, support vector regression prediction and numerical inverse transformation on the initial design influence parameter group, and outputs the predicted engineering cost value. The parameter iteration optimization module is equipped with a covariance matrix adaptive evolution strategy algorithm and a rule engine to complete parameter iteration calculations; The parameter parsing and processing module calls an external large language model, and evaluates the optimized design influence parameter group based on the vector database retrieval results and preset evaluation criteria, and generates text data.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The method for deriving structural parameters for cost-limited design of this invention forms a closed-loop cost-limited design mechanism guided by the target cost through a complete process: constructing a cost prediction model, generating an initial design influence parameter set, iteratively optimizing until the target cost constraint is met, and evaluating and re-optimizing the optimization results. This method changes the traditional model of repeatedly adjusting parameters based on manual experience, making the cost-limited design process verifiable and traceable. By bringing cost control forward to the parameter design stage, the impact of parameter combinations on cost can be quantitatively assessed in the early stages of design, effectively avoiding the disconnect between design and cost. At the same time, the system automatically records process data at each stage of the entire process, supporting design backtracking and result reproduction, and possessing good engineering deployability and practical value.

[0018] This invention employs a causal forest algorithm to screen key influencing parameters, effectively filtering out noise interference from non-critical factors and improving the representativeness of the model's input features. By jointly tuning the partial least squares regression dimensionality reduction and support vector regression prediction models, it eliminates multicollinearity while preserving the main information of the original features, significantly improving the accuracy and generalization performance of cost prediction. It leverages historical project data retrieved from a vector database and generates initial design influencing parameter sets through an external large language model, fully utilizing historical project experience to ensure the initial solution is reasonable and engineering-acceptable, greatly reducing the computational cost of subsequent iterative parameter finding. It uses a covariance matrix adaptive evolution strategy algorithm combined with a rule engine for iterative optimization, efficiently solving multi-parameter coupled nonlinear optimization problems. Simultaneously, the rule engine limits parameter ranges and verifies encoding mappings to ensure that generated parameters conform to engineering logic and design specifications. Parameter analysis and design scheme evaluation are conducted through an external large language model. Utilizing the semantic understanding and knowledge reasoning capabilities of the large language model, it assists designers in identifying parameter anomalies and making adjustment decisions, reducing the risk of design deviations. The system's modules are decoupled through standardized interfaces, supporting local deployment and offline operation, meeting the data security requirements of classified projects, and possessing good compatibility and scalability. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the main logic of the method for deriving engineering structure parameters in this embodiment of the invention. Figure 2 This is the main logic flowchart of the engineering structure parameter derivation method S2 in this embodiment of the invention; Figure 3 This is a user interface diagram of the user interaction layer of the engineering structure parameter derivation system in this embodiment of the invention; Figure 4 This is a diagram of the large language model configuration interface of the engineering structure parameter derivation system in this embodiment of the invention; Figure 5 This is a diagram showing the output interface of the engineering cost prediction value of the engineering structure parameter derivation system in this embodiment of the invention. Figure 6 This is a diagram showing the output interface of the design scheme report of the engineering structure parameter derivation system in this embodiment of the invention. Detailed Implementation

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] This embodiment uses the underground powerhouse of a pumped storage power station as an example for illustration. All data are from the finalized budget of a real project, and are intended to provide a more detailed explanation of the present invention. However, the present invention is not limited to the following embodiment.

[0022] This embodiment provides a method and system for deriving engineering structural parameters for budget-constrained design. It is used to derive and optimize the set of engineering design influence parameters in reverse under the constraint of target cost, and output the optimal set of engineering design influence parameters that meets the budget requirements.

[0023] The method for deriving engineering structural parameters in this embodiment is as follows: Figures 1-2 As shown, the main steps include: S1. Construct and train the cost prediction model to form a directly loadable cost prediction model package; the construction process mainly includes the following: S11. Construct a set of parameters affecting engineering cost. The set of parameters affecting engineering cost is divided into material price, regional geology, plant layout, and installed parameters according to engineering attributes. The non-numerical parameters in the set of parameters affecting engineering cost are quantified by ordinal coding to form an original feature set. Collect historical engineering cost data, extract the values ​​of each parameter in the original feature set, and extract the corresponding target cost data from the historical engineering cost data to construct the initial dataset.

[0024] S12 uses the causal forest algorithm to conduct causal inference analysis on the initial dataset and thereby determine the key influencing parameters. The key influencing parameters include steel bar price, cement price, sand unit price, stone unit price, location, factory type, factory burial form, factory length, factory width, factory height, installed capacity, number of installed units, and surrounding rock type.

[0025] S13, using a feature standardizer, the key influencing parameters are standardized using the mean-standard deviation method to obtain standardized key influencing parameters; using an engineering cost standardizer, the target cost data corresponding to the key influencing parameters are standardized using the mean-standard deviation method to obtain standardized target costs; the standardized dataset consists of the standardized key influencing parameters and the standardized target costs; the expression for the mean-standard deviation method is:

[0026]

[0027] in, To standardize key influencing parameters, The key influencing parameter feature matrix, The column mean of the key influencing parameter characteristics; The standard deviation of the key influencing parameter characteristics; To standardize the target cost, The target cost vector; The column mean of the target construction cost; The standard deviation of the target cost is listed.

[0028] S14 divides the standardized dataset into a training set and a test set in an 8:2 ratio; the training set is used for model training, and the test set is used for performance evaluation during the hyperparameter optimization process. The two are physically isolated to simulate real prediction scenarios.

[0029] S15 employs a partial least squares regression dimensionality reducer to perform supervised dimensionality reduction on standardized key influencing parameters in the training set, using the standardized target cost as the supervised signal, to generate low-dimensional features. The partial least squares regression dimensionality reducer iteratively extracts latent variables that maximize the covariance between the input features and the output target, using the standardized target cost as the supervised signal, to generate low-dimensional feature representations. The dimensionality reduction dimension is not used as a preset value but is used as a parameter to be optimized in subsequent parameter tuning. Based on the low-dimensional features and the standardized target cost, a support vector regression prediction model is constructed. The support vector regression prediction model uses radial basis functions as kernel functions to capture nonlinear relationships in the data.

[0030] S16 uses the negative mean square error of the test set as the optimization objective function and employs an optimization algorithm to jointly tune the support vector regression prediction model and the partial least squares regression dimensionality reducer to determine the optimal parameter combination. The optimization algorithm is one of Bayesian optimization algorithm (BO), whale optimization algorithm (WOA), genetic algorithm (GA), or particle swarm optimization algorithm (PSO).

[0031] The parameters to be optimized include: the regularization coefficient C(0.1, 100) of the support vector regression (SVR) prediction model, and the kernel function parameters. (0.001,1), insensitive loss (0.01,1), kernel function type (linear, poly, rbf, sigmoid), kernel function degree (2, 5), and the dimensionality reduction dimension n (2, 8) of the partial least squares regression dimensionality reducer.

[0032] In this embodiment, a Bayesian optimization algorithm is used to jointly tune the model. The optimization objective is the negative mean squared error (-MSE) on the test set, and the parameters to be optimized include: the SVR regularization coefficient C (0.1, 100), kernel function parameters, etc. (0.001,1), insensitive loss (0.01,1), kernel function type (linear,poly,rbf,sigmoid), kernel function degree (2,5), and PLS dimension reduction n (2,8).

[0033] To better evaluate the performance of the prediction model, the coefficient of determination (R²), root mean square error (RMSE), and percentage error (MAPE) were selected as evaluation metrics. Case studies show that the SVR model of this invention has high prediction accuracy, and all metrics outperform mainstream algorithms such as Extreme Gradient Boosting (XGBoost), Extreme Learning Machine (ELM), Lightweight Gradient Boosting (LightGBM), Backpropagation Neural Network (BPNN), and Support Vector Machine (SVM). The comparison results are shown in Table 1 below. The input variables for all models in the table are principal components after PLS dimensionality reduction, and all parameters were optimized using a Bayesian optimization algorithm.

[0034] Table 1 Comparison of Predictive Performance of Different Machine Learning Algorithms

[0035] Finally, a set of optimal parameter combinations was determined through Bayesian optimization, including the optimal dimensionality reduction dimension of PLS ​​and other SVR hyperparameters.

[0036] S17. The training set and the test set are combined to form a full training set. The support vector regression prediction model is retrained using the optimal parameter combination. The partial least squares regression dimensionality reducer and the support vector regression prediction model are reconstructed, resulting in four types of components: feature normalizer (scaler_X), engineering cost normalizer (scaler_Y), partial least squares regression dimensionality reducer (PLS), and support vector regression prediction model (SVR) with the optimal parameter combination.

[0037] S18 serializes and solidifies four types of components—feature standardizer, engineering cost standardizer, partial least squares regression dimensionality reducer, and support vector regression prediction model—to form a cost prediction model package that can be directly loaded. The cost prediction model package is organized in dictionary format, stored in joblib format, and loaded once when the system starts, ensuring the consistency and reproducibility of the prediction process.

[0038] S2, based on the target project cost and known design influence parameters, an initial design influence parameter set is generated. The cost prediction model is then used to predict the cost of this initial design influence parameter set, obtaining a project cost prediction value. This predicted cost value is compared with the target project cost. If the comparison result does not meet the preset error requirement, the initial design influence parameter set is adjusted, and the cost prediction is re-executed. This comparison and adjustment are repeated until the preset error requirement is met, resulting in an optimized design influence parameter set. Specifically, this includes: S21 uses the open-source model all-MiniLM-L6-v2 to transform the target cost text data into vector data, and stores the vector data, the original target cost text, and metadata used to describe the text attributes into a vector database; the vector database uses the Chroma database.

[0039] S22, the user inputs the target engineering value and known design impact parameters through the user interaction layer. The system performs a rationality check on the input parameters and outputs warning information for parameters that exceed the preset reasonable range. The system calls an external large language model through the application programming interface to retrieve historical project data in the vector database and generate the initial design impact parameter set. The external large language models include the GPT series, Tongyi Qianwen series, DeepSeek series, and Claude series, which are accessed through API interfaces and the vendor's native SDK.

[0040] S23. Input the initial design influence parameter group into the cost prediction model package, complete the feature standardization through the feature standardizer, and then input it into the partial least squares regression dimensionality reducer and support vector regression prediction model in sequence to obtain the standardized prediction value. After inverse transformation, output the engineering cost prediction value.

[0041] S24, calculate the relative error between the predicted project cost and the target project cost, and determine whether the relative error meets the preset limiting conditions; the preset limiting condition is that the relative error is less than 5%; if it meets the conditions, retain the initial design influence parameter set; if it does not meet the conditions, use the covariance matrix adaptive evolution strategy algorithm, in conjunction with the rule engine, to iteratively optimize the initial design influence parameter set until the relative error meets the preset limiting conditions, and output the optimized design influence parameter set, which includes unknown design influence parameters; the iterative optimization stopping condition of the covariance matrix adaptive evolution strategy algorithm is that the algorithm loop reaches 100 times and there is a design influence parameter set that meets the error requirements; the rule engine is used to limit the parameter value range, verify the ordinal encoding mapping relationship, and ensure that the parameters conform to the engineering logic.

[0042] In this embodiment, the stopping condition for iterative optimization of the initial design influence parameter group is that the number of iterations reaches 100 and there is data with a relative error of less than 5%. If there is no data with error meeting the requirement, the design influence parameters are re-initialized and the optimization steps are repeated. The optimization algorithm uses the CMA-ES algorithm, which, combined with a rule engine, can significantly improve the parameter finding efficiency.

[0043] The rule engine is designed to ensure that the generated parameters are reasonable and conform to engineering logic. This includes limiting the range of generated parameters and ensuring correct ordinal encoding mapping. For example, the region of a parameter is defined as follows: 1 represents a Class I region / economically developed, with convenient transportation and good construction conditions; 2 represents a Class II region / relatively developed, with relatively convenient transportation and good construction conditions; 3 represents a Class III region; 4 represents a Class IV region; 5 represents a Class V region; 6 represents a Class II / Class VI region in Tibet; 7 represents a Class III region in Tibet; and 8 represents a Class IV region in Tibet. During the optimization process, the region can only be an integer from 1 to 8.

[0044] After the initial design influence parameter set is optimized, the system will output a design result that meets the requirements and has the smallest relative error, such as... Figure 5 As shown, it includes known design influence parameter information, relative error, optimized design influence parameter group, and optimization process log, etc.

[0045] S3: The user selects whether to perform the design scheme analysis. If the user selects to perform the design scheme analysis, the optimized design influence parameter group is evaluated and the evaluation result is output to determine whether there are any unreasonable parameters. If there are unreasonable parameters, the user determines whether to perform iterative optimization on the corresponding design influence parameters according to the user's selection instruction. If so, the new optimized design influence parameter group is updated after the iterative optimization is completed, and then the process proceeds to step S4. If there are no unreasonable parameters, the process proceeds directly to step S4. If the user selects not to perform the design scheme analysis or determines not to perform iterative optimization, the process proceeds directly to step S4.

[0046] In this embodiment, an external large language model is invoked to evaluate the optimized design impact parameter group based on the vector database retrieval results and preset evaluation criteria, and generate text data. The external large language model uses a server push event protocol to transmit text information. If the evaluation result indicates that the parameter is abnormal, the user can choose to iteratively optimize the corresponding parameter. The iterative optimization repeats steps S23 and S24 until the iteration number requirement is met, and the optimized design impact parameter group that meets the error requirement and has the smallest error is output.

[0047] The prompts from the external large language model must be limited to national and industry standards for evaluation, including the "Design Code for Underground Powerhouse of Hydropower Station" (NB / T 35090-2016), the "Design Code for Loads of Hydraulic Structures" (SL / T744-2016), the "Design Code for Hydraulic Concrete Structures" (SL / T 191-2025), and the "Engineering Rock Mass Classification Standard" (GB / T50218-2014). The evaluation dimensions include technical rationality, economic rationality, risk control, design experience, and structural mechanics verification. If the evaluation results indicate abnormal parameters, the user can choose to iteratively optimize the corresponding parameters. The iterative optimization repeats steps S23 and S24 until the iteration number requirement is met, and the optimized design influence parameter set that meets the error requirements and has the smallest error is output.

[0048] S4. Output the optimized design influence parameter set obtained in step S2 or S3 as the optimal engineering design influence parameter set that satisfies the target cost constraint; if the design scheme analysis is performed, a design scheme report can be output. The design scheme report includes the final optimal engineering design influence parameter set and related data information such as known design influence parameters.

[0049] The system can save the final optimal engineering design influence parameter group data to a relational database, or export it as an Excel, PDF, or Word file. The relational database uses an SQLite database to store data such as project identifier, design influence parameters, cost values, and timestamps.

[0050] The engineering structural parameter derivation system of this embodiment, such as Figures 3-6 As shown, where Figure 4 Includes, Figure 4 (a) Screenshot of the interface for selecting the large language model. Figure 4 (b) is the state diagram for when selected, including the user interaction layer, service scheduling layer, core function layer, and data storage layer: The user interaction layer is used to receive the target project cost and known design influence parameters, and output the optimal set of engineering design influence parameters and design scheme report that meets the target cost constraints.

[0051] The service scheduling layer is configured with service scheduling routes to enable data interaction and command forwarding between the user interaction layer and the core function layer; the service scheduling routes include: The cost prediction route is used to receive multi-dimensional engineering feature parameters from the user interaction layer, call the locally loaded cost prediction model package through the cost prediction calculation module to perform cost prediction, and return the numerical results of the corresponding design influence parameter group. The parameter optimization route is used to receive the target cost constraints and known design influence parameter boundary conditions of the user interaction layer, trigger the covariance matrix adaptive evolution strategy algorithm through the parameter iteration optimization module to iteratively solve the appropriate design influence parameter set, and output the design influence parameter set to the parameter parsing route; The parameter parsing routing is used to retrieve historical project data from the data storage layer and connect to the external large language model interface through the large language model calling module. The generated design influence parameter group is evaluated by the parameter parsing processing module and corresponding text data is generated and output to the user interaction layer.

[0052] The core functional layer communicates with the service scheduling layer and is configured sequentially according to data flow: a data receiving module, a large language model invocation module, a cost prediction calculation module, a parameter iteration optimization module, and a parameter parsing and processing module. Each module establishes a data connection through a standardized interface. The data receiving module is used to receive the target engineering value and known design influence parameters input by the user, and to perform a rationality check on the input parameters and output a verification prompt message. The large language model calling module connects to external large language models to retrieve data and search vector databases to generate initial design influence parameter sets. The cost prediction calculation module loads the cost prediction model package and sequentially performs feature standardization, partial least squares regression dimensionality reduction, support vector regression prediction and numerical inverse transformation on the initial design influence parameter group, and outputs the predicted engineering cost value. The parameter iteration optimization module has a built-in covariance matrix adaptive evolution strategy algorithm and rule engine. Based on preset error conditions and loop termination conditions, it iteratively calculates the initial design influence parameter set and outputs the optimized design influence parameter set. The parameter parsing and processing module calls an external large language model, and evaluates the optimized design impact parameter group based on the results of vector database retrieval and preset evaluation criteria, and generates text data.

[0053] The data storage layer communicates with the core functional layer and includes a built-in cost prediction model package, vector database, historical project data, and relational database, providing data and model support for the core functional layer. The cost prediction model package integrates a feature normalizer, an engineering cost normalizer, a partial least squares regression dimensionality reducer, and a support vector regression prediction model. The vector database stores vector data, original text, and metadata used to describe text attributes. The relational database stores project identifiers, design impact parameters, cost values, timestamps, etc.

[0054] The system adopts a local deployment mode, builds a web service based on the Flask framework, listens on port localhost:5000, and all calculations are completed on the local terminal or internal server; it supports cross-domain resource sharing configuration and is compatible with mainstream browsers; model updates adopt an offline mode, and after accumulating sufficient new project data, a new version of the model package is retrained to replace the old version.

[0055] This embodiment uses a causal forest algorithm to screen key influencing parameters, and PLS supervised dimensionality reduction combined with optimization algorithms to jointly tune parameters, thus constructing a high-precision cost prediction model. It generates an initial design influencing parameter set by retrieving data from a vector database, and iteratively optimizes the parameters using the CMA-ES algorithm combined with a rule engine, achieving parameter reverse inference guided by the target cost. It uses an external large language model to realize parameter evaluation and report generation. The system is closed-loop, traceable, and easy to deploy throughout the entire process, effectively improving the efficiency and accuracy of engineering budget design.

[0056] The above are merely embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All modifications, equivalent substitutions, and improvements made within the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for deriving engineering structural parameters for budget-constrained design, characterized in that, Includes the following steps: S1, Build and train the cost prediction model to form a cost prediction model package that can be directly loaded; The construction process of the cost prediction model in step S1 is as follows: S11, construct a set of engineering cost influencing parameters. The set of influencing parameters is divided into material price category, regional geology category, plant layout category, and installed equipment parameter category according to engineering attributes. The non-numerical influencing parameters in the set of engineering cost influencing parameters are quantified by ordinal encoding to form an original feature set. Historical engineering cost data is collected, and the values ​​of each influencing parameter in the original feature set are extracted. At the same time, the target cost data corresponding to the values ​​of each influencing parameter are extracted from the historical engineering cost data to construct an initial dataset. S12, The causal forest algorithm is used to perform causal inference analysis on the initial dataset, and key influencing parameters are determined accordingly; S13, the key impact parameters are standardized using a feature standardizer to obtain standardized key impact parameters; the target cost data corresponding to the key impact parameters are standardized using an engineering cost standardizer to obtain standardized target cost; the standardized key impact parameters and the standardized target cost together form a standardized dataset; S14, the standardized dataset is divided proportionally to obtain a training set and a test set; S15, A partial least squares regression dimensionality reducer is used to perform supervised dimensionality reduction on the standardized key influence parameters within the training set, with the standardized target cost as the supervised signal, to generate low-dimensional features; based on the low-dimensional features and the standardized target cost within the training set, a support vector regression prediction model is built. S16, using the negative mean square error corresponding to the test set as the optimization objective function, the optimization algorithm is used to jointly tune the support vector regression prediction model and the partial least squares regression dimensionality reducer to determine the optimal parameter combination; S17, combine the training set and the test set to form a full training set, retrain the support vector regression prediction model based on the optimal parameter combination, reconstruct the partial least squares regression dimensionality reducer and the support vector regression prediction model, and obtain four types of components: feature standardizer, engineering cost standardizer, partial least squares regression dimensionality reducer and support vector regression prediction model that match the optimal parameter combination. S18, perform serialization and solidification processing on the four types of components that match the optimal parameter combination to form a cost prediction model package that can be directly loaded; S2. Based on the target project cost and known design influence parameters, an initial design influence parameter set is generated. The cost prediction model is used to predict the cost of the initial design influence parameter set to obtain the project cost prediction value. The project cost prediction value is compared with the target project cost. If the comparison result does not meet the preset error requirement, the covariance matrix adaptive evolution strategy algorithm is used to adjust the initial design influence parameter set and the cost prediction is re-executed. The comparison and adjustment are repeated until the preset error requirement is met to obtain the optimized design influence parameter set. S3, the user chooses whether to perform the design scheme analysis. If the user chooses to perform the design scheme analysis, the optimized design influence parameter group is evaluated and the evaluation results are output to determine whether there are any unreasonable parameters. If there are unreasonable parameters, it is determined whether to iteratively optimize the corresponding design impact parameters according to the user's selection instructions; if so, the covariance matrix adaptive evolution strategy algorithm is adopted in conjunction with the rule engine to iteratively optimize the corresponding design impact parameters. After the iterative optimization is completed, a new optimized design impact parameter group is obtained, and then the process proceeds to step S4; if there are no unreasonable parameters, the process proceeds directly to step S4. If the user chooses not to perform the design scheme analysis or determines not to perform iterative optimization, then proceed directly to step S4; the rule engine is used to limit the value range of parameters and verify the ordinal encoding mapping relationship; S4. Output the optimized design influence parameter set obtained in step S2 or S3 as the optimal engineering design influence parameter set that satisfies the target cost constraint.

2. The method according to claim 1, characterized in that, In step S1, the training set and test set obtained by dividing the standardized dataset are allocated in an 8:2 ratio; the optimization algorithm is any one of Bayesian optimization algorithm, whale optimization algorithm, genetic algorithm, and particle swarm optimization algorithm.

3. The method according to claim 1, characterized in that, The specific content of step S2 includes: S21, The target cost text data is transformed into vector data using an open-source model, and the vector data, the original target cost text, and the metadata used to describe the text attributes are stored together in the vector database. S22, receive the target engineering value and known design impact parameters input from the outside, call the external large language model through the application programming interface, retrieve the historical project data in the vector database, and generate the initial design impact parameter group; S23, the initial design influence parameter group is input into the cost prediction model package. After the feature standardization operation is completed by the feature standardizer, it is sequentially input into the partial least squares regression dimensionality reducer and the support vector regression prediction model to obtain the standardized prediction value. The standardized prediction value is then inversely transformed and the project cost prediction value is output. S24, calculate the relative error between the predicted project cost and the target project cost, and determine whether the relative error meets the preset limiting conditions; if it does, use the initial design influence parameter set as the optimized design influence parameter set; if it does not, use the covariance matrix adaptive evolution strategy algorithm in conjunction with the rule engine to iteratively optimize the initial design influence parameter set until the relative error meets the preset limiting conditions, and output the optimized design influence parameter set, which includes unknown design influence parameters.

4. The method according to claim 3, characterized in that, The preset limiting condition is that the relative error between the predicted project cost and the target project cost is less than 5%; the iterative optimization stopping condition is that the number of iterations reaches 100, or there is a set of design influence parameters that meet the relative error requirement.

5. A system for deriving engineering structural parameters for budget-constrained design, characterized in that, The system performs the method according to any one of claims 1-4, the system comprising: The user interaction layer is used to receive the target project cost and known design impact parameters, and output the optimal engineering design impact parameter set and design scheme report that satisfy the target cost constraint. The service scheduling layer is equipped with business scheduling routes to realize data interaction and command forwarding between the user interaction layer and the core function layer. The core functional layer communicates with the service scheduling layer and consists of a data receiving module, a large language model calling module, a cost prediction calculation module, a parameter iteration optimization module, and a parameter parsing and processing module, arranged sequentially according to the data flow. Each module establishes a data connection with the other through standardized interfaces. The data storage layer communicates with the core functional layer. The data storage layer has a built-in cost prediction model package, vector database, historical project data and relational database, providing data and model support for the core functional layer.

6. The system according to claim 5, characterized in that, The service scheduling route includes: The cost prediction route is used to receive multi-dimensional engineering feature parameters from the user interaction layer, call the locally loaded cost prediction model package through the cost prediction calculation module to perform cost prediction, and return the numerical results of the corresponding design influence parameter group. The parameter optimization route is used to receive the target cost constraints and known design influence parameter boundary conditions of the user interaction layer, trigger the covariance matrix adaptive evolution strategy algorithm through the parameter iteration optimization module to iteratively solve the appropriate design influence parameter set, and output the design influence parameter set to the parameter parsing route; The parameter parsing route is used to retrieve historical project data from the data storage layer and connect to the external large language model interface through the large language model calling module. The parameter parsing processing module performs parameter parsing on the generated design influence parameter group and generates corresponding text data, which is then output to the user interaction layer.

7. The system according to claim 5, characterized in that, The cost prediction model includes an integrated feature normalizer, an engineering cost normalizer, a partial least squares regression dimensionality reducer, and a support vector regression prediction model; the vector database stores vector data, original text, and metadata used to describe text attributes.

8. The system according to claim 5, characterized in that, The data receiving module is used to receive the target engineering value and known design impact parameters input by the user, and to perform a rationality check on the input parameters and output a verification prompt message. The large language model calling module connects to an external large language model to retrieve data and generate the initial design influence parameter set; The cost prediction calculation module loads the cost prediction model package and sequentially performs feature standardization, partial least squares regression dimensionality reduction, support vector regression prediction and numerical inverse transformation on the initial design influence parameter group, and outputs the predicted engineering cost value. The parameter iteration optimization module is equipped with a covariance matrix adaptive evolution strategy algorithm and a rule engine to complete parameter iteration calculations; The parameter parsing and processing module calls an external large language model, and evaluates the optimized design influence parameter group based on the vector database retrieval results and preset evaluation criteria, and generates text data.

Images