A method and system for investment prediction of hydropower engineering based on causal screening and joint optimization

Abstract

The application discloses a kind of causal screening and joint optimization's hydropower engineering investment prediction method and system, belong to hydropower engineering cost prediction technical field.For the feature causal structure is not clear in hydropower engineering investment prediction, dimension reduction and modeling parameter split optimization leads to low prediction accuracy under small sample high-dimensional data, poor generalization ability, the application adopts double machine learning algorithm to carry out causal inference screening key influence factor, constructs the joint prediction model of partial least square regression and support vector regression, through improving bayesian optimization to jointly optimize dimension reduction dimension and support vector regression hyperparameter, and feature standardizer, partial least square dimension reducer and support vector regression predictor are serialized and solidified into independently loadable model package.The application significantly improves prediction accuracy and robustness, realizes the engineering closed loop from parameter input to investment prediction, and is suitable for rapid investment estimation in the feasibility study stage of hydropower engineering.

Description

Technical Field

[0001] This invention relates to the field of hydropower project cost prediction technology, specifically to a method and system for predicting hydropower project investment using causal screening and joint optimization. Background Technology

[0002] Hydropower projects, as major national infrastructure projects, are characterized by large investment scale, long construction period, and complex influencing factors. Accurately predicting their total investment is of great significance for project approval decisions, fundraising, and risk control. Traditional investment estimation methods mainly include unit production capacity estimation, index estimation, and analogy methods. These methods rely heavily on expert experience, are highly subjective, and struggle to effectively quantify the coupling effects between multiple factors. Consequently, their prediction accuracy often falls short of the actual needs of engineering projects.

[0003] In recent years, machine learning methods have been gradually introduced into the field of hydropower investment forecasting, such as multiple linear regression, random forest, neural networks, and support vector regression. These methods have improved the objectivity and accuracy of forecasts to some extent through data-driven approaches. However, due to the high dimensionality, small sample size, and complex nonlinear coupling relationships between factors in hydropower engineering data, existing methods still face significant obstacles in practical applications, including feature redundancy and causal confusion, high-dimensional nonlinear modeling bottlenecks, and low efficiency in hyperparameter tuning.

[0004] The main shortcomings of existing technologies are as follows: there are a large number of initial candidate influencing factors, many of which are only statistically correlated with investment rather than having a real causal relationship. Direct modeling without screening can easily introduce noise, leading to overfitting or a decrease in generalization ability. Traditional unsupervised dimensionality reduction methods, such as principal component analysis, ignore the supervision information of the prediction target and cannot effectively retain the low-dimensional feature subspace that is most discriminative to investment. Hyperparameter tuning of the prediction model and feature dimensionality reduction are usually carried out in separate steps, making it difficult to ensure the best fit between the dimensionality-reduced features and the regression model.

[0005] The aforementioned problems collectively restrict the accuracy and robustness of hydropower project investment forecasting, and there is an urgent need for a technical solution that can screen features from a causal perspective and achieve joint optimization of dimensionality reduction and modeling parameters. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method and system for predicting hydropower project investment using causal screening and joint optimization. This aims to solve the technical problems in hydropower project investment prediction where unclear causal structures and fragmented optimization of dimensionality reduction and modeling parameters lead to low prediction accuracy and poor generalization ability with small sample high-dimensional data. To achieve the above objective, this invention provides a method for predicting hydropower project investment using causal screening and joint optimization, comprising the following steps: S1. Obtain candidate influencing factors and corresponding investment data of historical hydropower projects, use dual machine learning algorithms to perform causal inference on the candidate influencing factors, calculate the average processing effect value of each candidate influencing factor on the investment data, and screen out key influencing factors according to the preset causal effect threshold. S 2: Based on the key influencing factors and the investment data, construct a joint prediction model to be optimized, which includes partial least squares regression and support vector regression; wherein, the partial least squares regression is used to perform supervised dimensionality reduction on the key influencing factors, and the support vector regression is used to make investment predictions based on the dimensionality-reduced features. S 3: An improved Bayesian optimization algorithm is adopted to incorporate the dimensionality reduction dimension of the partial least squares regression and the hyperparameters of the support vector regression into the same parameter space for joint optimization, so as to determine the optimal parameter combination and obtain the optimized joint prediction model. S 4: Obtain the candidate influencing factor set for the hydropower project to be predicted, and based on the above... S The key impact factors identified in step 1 are used to extract corresponding data from the candidate impact factor set to form a key impact factor value set; S 5: Input the set of key influencing factor values ​​into the... S In the optimized joint prediction model obtained in the three steps, standardization, supervised dimensionality reduction, regression prediction and inverse standardization are performed sequentially by the optimized joint prediction model to output the investment prediction value of the hydropower project to be predicted.

[0007] Preferably, the S The specific steps include: Obtain candidate influencing factors and corresponding investment data for the historical hydropower projects; employ a dual machine learning algorithm to construct a first machine learning model to perform residual processing on the candidate influencing factors as processing variables, obtaining a first residual; construct a second machine learning model to perform residual processing on the investment data as the outcome variable, obtaining a second residual; and calculate the average processing effect value of each candidate influencing factor based on the first residual and the second residual, obtaining an effect value set. The candidate influencing factors whose average absolute value of the effect value is greater than the preset causal effect threshold are identified as the key influencing factors.

[0008] Preferably, the S The two steps specifically include: The key influencing factors and the investment data are respectively subjected to mean-standard deviation standardization to obtain a standardized feature matrix and a standardized target vector. Using the standardized target vector as a supervision signal, partial least squares regression is used to perform supervised dimensionality reduction on the standardized feature matrix. By iteratively extracting the latent variables that maximize the covariance between the input features and the output target, a low-dimensional feature representation is generated. Based on the low-dimensional feature representation and the standardized target vector, an initial prediction model is constructed using support vector regression with radial basis function as kernel function, and the joint prediction model to be optimized is obtained.

[0009] Preferably, the S In step 3, the parameter space for joint optimization by the improved Bayesian optimization algorithm includes: the dimensionality reduction dimension of partial least squares regression, the regularization coefficient of support vector regression, the kernel function parameter, the insensitive loss coefficient, and the kernel function type.

[0010] Preferably, the S The improved Bayesian optimization algorithm in step 3 includes the following optimization strategies: Construct multiple sets of random probe points covering the entire range of the parameter space, and merge them with the initial random exploration points to obtain an optimized starting point set; Based on the optimized starting set, a surrogate model is constructed using a Gaussian process, and the expected improved acquisition function is used for iterative optimization to obtain the optimal solution in continuous space. For discrete parameters contained in the parameter space, the optimal solution in the continuous space obtained by optimization is converted into a discrete optimal value by integer rounding or preset category mapping in order to determine the final parameter combination. Based on the final parameter combination, the optimized joint prediction model is configured and generated.

[0011] Preferably, the S In step 5, the internal processing of the optimized joint prediction model includes: The feature normalizer in the optimized joint prediction model is invoked to standardize the set of key influencing factor values, thereby obtaining a standardized feature vector. The partial least squares regression dimensionality reducer in the optimized joint prediction model is invoked to reduce the dimensionality of the standardized feature vector, thereby obtaining a low-dimensional feature representation. The support vector regression predictor in the optimized joint prediction model is invoked to perform regression prediction on the low-dimensional feature representation to obtain standardized prediction values. The inverse standardizer in the optimized joint forecasting model is invoked to perform inverse standardization on the standardized forecast value, and the investment forecast value is obtained and output.

[0012] Preferably, the method further includes a model solidification step: The SThe feature normalizer, partial least squares regression dimensionality reducer, and support vector regression predictor in the optimized joint prediction model obtained in the three steps are solidified as a whole in a serialized storage format to generate a prediction model package that can be loaded independently.

[0013] Preferably, the S The candidate impact factor set for the hydropower project to be predicted in step 4, including the type, unit of measurement, and order of the candidate impact factors, is related to the... S In step 1, the types, units of measurement, and order of candidate influencing factors for historical hydropower projects should remain consistent.

[0014] A hydropower project investment prediction system based on causal screening and joint optimization includes: The user interaction module is used to receive multiple candidate influencing factors input by the user and display the investment forecast values; The data storage module is used to persistently save the input parameters, prediction results, and timestamps for each prediction. The prediction module is connected to both the user interaction module and the data storage module. The execution prediction module contains a joint prediction model constructed and optimized by the method described above. The execution prediction module is configured to: load the joint prediction model, receive candidate influencing factors from the user interaction module, output investment prediction values ​​to the user interaction module, and send prediction records to the data storage module.

[0015] Preferably, the execution prediction module is equipped with a parameter verification unit, which is used to verify whether there are missing values ​​or values ​​exceeding the preset range of multiple candidate influencing factors input by the user before calling the joint prediction model. If the verification passes, the data is sent to the joint prediction model; if the verification fails, a prompt message is generated and returned to the user interaction module.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention uses dual machine learning causal inference to screen key influencing factors and eliminates non-causal noise at the source, overcoming the overfitting problem caused by feature redundancy in traditional methods. It employs partial least squares regression for supervised dimensionality reduction, making full use of investment label information to retain the most discriminative low-dimensional feature subspace, significantly improving the robustness of modeling under small sample high-dimensional data.

[0017] This invention improves Bayesian optimization by incorporating dimensionality reduction and support vector regression hyperparameters into the same space for joint optimization, breaking the two-step fragmented optimization model. It combines probe-point augmentation exploration with a hybrid parameter adaptive mapping mechanism to ensure optimal fit between dimensionality reduction features and the prediction model, balancing global optimization accuracy and computational efficiency. This invention integrates the normalizer, dimensionality reducer, and predictor into a single, independently loadable model package, supporting offline deployment and rapid invocation. The system integrates parameter verification and persistent storage of prediction records, forming an engineered closed loop of "input-verification-prediction-storage-display." Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 This is a system architecture diagram of the present invention.

[0020] Figure 3 This is a screenshot of the web form interface of the hydropower project investment prediction system of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The steps, modules, and logical relationships described in this embodiment constitute a complete interpretation of the technical solutions of the present invention, and those skilled in the art can reproduce the present invention based on this embodiment.

[0022] This embodiment elaborates in detail the process and technical principle of the causal screening and joint optimization method for predicting hydropower project investment according to the present invention. The complete process is as follows: Figure 1 As shown, this method constructs a high-precision prediction model and supports lightweight deployment through four stages: causal-driven screening, supervised dimensionality reduction, joint optimization, and model solidification.

[0023] The primary technical obstacle to predicting investment in hydropower projects is the large number of initial candidate influencing factors, many of which are only statistically correlated with investment rather than having a true causal relationship. If modeling is performed directly without screening, noise will be introduced and the model will overfit. To solve this problem, this invention uses a dual machine learning algorithm to perform causal inference on candidate influencing factors in order to achieve data-driven feature screening.

[0024] First, candidate influencing factors and corresponding investment data of historical hydropower projects are obtained. These candidate influencing factors typically cover multiple dimensions such as engineering economics, engineering layout, structural design, construction organization, and policy environment. Non-numerical factors are then quantified and coded.

[0025] Dual machine learning algorithm ( Double Machine Learning, DML The execution process is as follows: A first machine learning model is constructed to perform residual processing on the candidate influencing factors used as treatment variables, obtaining the first residual. The purpose of this is to eliminate the influence of confounding factors on the treatment variables. A second machine learning model is constructed to perform residual processing on the investment data used as the outcome variable, obtaining the second residual. The purpose of this is to eliminate the influence of confounding factors on the outcome variable. Based on the orthogonalization results of the first and second residuals, the average treatment effect value of each candidate influencing factor on the investment data is calculated. ATE And form an effect value set.

[0026] Obtain a preset causal effect threshold, which can be set based on the statistical distribution characteristics of the effect set or the experience of domain experts; then divide the effect set... ATE Candidate impact factors with absolute values ​​greater than the threshold are identified as key impact factors with significant causal effects and used for subsequent modeling. Through the above causal inference mechanism, this method removes non-causal noise features from the data source, significantly enhancing the model's anti-interference ability and generalization reliability in small sample high-dimensional data scenarios.

[0027] After obtaining the key influencing factors, this method constructs a joint prediction model to be optimized, which includes partial least squares regression and support vector regression, in order to solve the modeling challenges of high-dimensional, small-sample, and highly nonlinear hydropower investment data.

[0028] First, the key influencing factors and investment data are standardized using the mean-standard deviation to eliminate differences in units of measurement. The mathematical expression is as follows:

[0029] in For the characteristic matrix, For investment data vectors, and For each feature, the column mean and standard deviation, and The mean and standard deviation of the investment data are used to obtain the standardized feature matrix and the standardized target vector after standardization.

[0030] Subsequently, using the standardized target vector as the supervision signal, partial least squares regression was employed to perform supervised dimensionality reduction on the standardized feature matrix; PLSThis method iteratively extracts latent variables that maximize the covariance between input features and output target, then projects the high-dimensional original feature space onto a low-dimensional latent variable space to generate a low-dimensional feature representation. Compared with unsupervised dimensionality reduction methods such as principal component analysis, this approach yields a lower-dimensional feature representation. PLS By making full use of the labeling information of investment data, the feature subspace that is most discriminative to investment prediction is effectively preserved during the dimensionality reduction process.

[0031] Finally, a support vector regression prediction model is constructed based on the low-dimensional feature representation and the standardized target vector. SVR Radial basis functions are used as kernel functions to effectively capture the complex nonlinear relationships in hydropower investment data; thus, a joint prediction model to be optimized is obtained, in which... PLS Dimensionality reduction and SVR The various hyperparameters will be jointly optimized in subsequent steps.

[0032] Traditional modeling processes typically employ a two-step strategy: dimensionality reduction and parameter tuning followed by model parameter tuning. This fragmented optimization approach struggles to guarantee optimal fit between the reduced features and the prediction model. To overcome this deficiency, this invention utilizes an improved Bayesian optimization algorithm... PLS Dimensionality reduction and SVR Hyperparameters are incorporated into the same parameter space for integrated joint optimization.

[0033] The joint parameter space specifically includes PLS Dimensionality Reduction n components , SVR Regularization coefficient C Kernel function parameters γ, insensitive loss coefficient ε, and kernel function type kernel ; IBO The algorithm uses the negative mean squared error on the test set as the optimization objective function, constructs a surrogate model through a Gaussian process, and improves the acquisition function using the expectation. Expected Improvement, EI (This guides the search direction.)

[0034] IBO The improved optimization strategy of the algorithm is reflected in the following aspects: First, before the optimization starts, multiple sets of random probe points covering the entire parameter space are constructed and merged with the initial random exploration points to form an optimization starting point set. By enhancing the diversity of the initial exploration, the algorithm avoids getting trapped in local optima. Second, a hierarchical search mechanism is adopted to balance global exploration and local development within a limited number of iterations. Third, for discrete parameters such as kernel function type, an adaptive mapping and integer processing method for continuous-discrete mixed type parameters is designed. Numerical optimization is first performed in the continuous search space, and then the value is converted into a discrete optimal value through integer rounding or preset category mapping, thereby achieving seamless adaptation of Bayesian optimization to continuous and discrete mixed type parameters.

[0035] After iterative optimization, an optimal combination of parameters is determined. Based on this combination, an optimized joint prediction model is configured and generated. This integrated joint optimization mechanism realizes the parameter linkage optimization between the dimensionality reduction process and the prediction modeling process, ensuring that the dimensionality reduction features and prediction modeling are optimized together. SVR The optimal fit of the model effectively improves the overall modeling accuracy and robustness.

[0036] In the prediction application phase, the first step is to obtain a set of candidate influencing factors for the hydropower project to be predicted. To ensure that the data distribution is consistent with that in the training phase, the set of candidate influencing factors must match the set in the steps in terms of factor type, unit of measurement, and order. S The candidate influencing factors for historical hydropower projects remained strictly consistent.

[0037] According to the steps S 1. The key impact factors are identified, and corresponding data are extracted from the candidate impact factor set to form a key impact factor value set. If there are any missing values, the user is prompted to supplement them.

[0038] The key impact factor value set is input into the optimized joint prediction model. The model automatically executes the following processing steps: It calls the feature normalizer to standardize the key impact factor value set, obtaining standardized feature vectors; it then calls... PLS The dimensionality reducer performs dimensionality reduction on the standardized feature vectors to obtain a low-dimensional feature representation; [call / implementation] SVR The predictor performs nonlinear regression calculations on the low-dimensional feature representations to obtain standardized predicted values; finally, it calls the inverse normalizer using the formula:

[0039] The standardized forecast values ​​are converted back into actual investment forecast values ​​and then output; the entire forecasting process is highly automated and requires no human intervention.

[0040] To support lightweight deployment and rapid application in engineering fields, this method also includes a model solidification step; the feature normalizer, partial least squares regression dimensionality reducer, and support vector regression predictor in the optimized joint prediction model are solidified as a single, independently loadable prediction model package in a serialized storage format; this model package encapsulates the mean and standard deviation parameters required for standardization, PLS Dimensionally reduced loading matrix and projection matrix and SVR The model's support vectors and related parameters can be calculated by simply loading the model package during the prediction phase, without the need for online training or parameter updates, thus ensuring the stability, response efficiency, and reproducibility of the prediction service.

[0041] This method specifically limits the steps described herein. S The candidate influencing factors set for the four hydropower projects to be predicted must conform to the steps in terms of physical meaning, unit of measurement, and order of arrangement. SThe candidate influencing factors of historical hydropower projects remain completely consistent. This constraint ensures that the training data and the prediction data follow the same feature distribution, which is a prerequisite for the correct execution of the model normalizer and the effectiveness of the prediction results. It guarantees the consistency and reliability of the prediction from the perspective of the engineering process.

[0042] In summary, this embodiment fully illustrates the technical aspects and collaborative working mechanism of the causal screening and joint optimization method for predicting hydropower project investment of the present invention. This method eliminates noise factors through causal-driven screening, retains discriminative information through supervised dimensionality reduction, achieves optimal adaptation between dimensionality reduction and modeling through joint optimization, and supports engineering deployment through model solidification, thus forming a complete technical closed loop and effectively solving the long-standing technical problems in the field of hydropower project investment prediction.

[0043] To facilitate understanding, this embodiment collects the approved budget data of 48 hydropower projects during the feasibility study stage as a historical project data source to ensure the authenticity and representativeness of the data.

[0044] First, through systematic literature review and in-depth interviews with three senior hydropower engineers, a preliminary set of 60 candidate influencing factors was constructed and divided into five dimensions according to engineering attributes: engineering economics, engineering layout, structural design, construction organization, and policy environment. Non-numerical factors were quantified using ordinal coding. The actual values ​​of each factor and corresponding investment data of the above 48 hydropower stations were collected to form an initial dataset.

[0045] A dual machine learning algorithm was used to perform causal inference analysis on the initial dataset to calculate the average treatment effect of each candidate influencing factor on the investment data. ATE Specifically, a first machine learning model is constructed to perform residual processing on the candidate influencing factors used as processing variables to obtain the first residual; a second machine learning model is constructed to perform residual processing on the investment data used as the outcome variable to obtain the second residual; and the factors are calculated based on the first and second residuals. ATE The values ​​are then used to form an effect value set.

[0046] Set the causal effect threshold to 0.1 and concentrate the effect values. ATE Candidate influencing factors with absolute values ​​greater than the threshold were identified as key influencing factors with significant causal effects. After screening, 56 key influencing factors were finally retained for subsequent modeling. The causal analysis results showed that factors such as installed capacity, total construction period, rated head, and diversion method had significant causal driving effects on investment.

[0047] The mean-standard deviation standardization was performed on the above 56 key influencing factors and one-dimensional investment data, and the mathematical expression is as follows:

[0048] in For the characteristic matrix, For investment data vectors, and For each feature, the column mean and standard deviation, and The mean and standard deviation of the investment data are used; standardization is performed on all 48 samples and the resulting standardized dataset is divided into a training set with 38 samples and a test set with 10 samples in an 8:2 ratio, with the two sets physically isolated to simulate real prediction scenarios.

[0049] Using the standardized target vector as the supervision signal, partial least squares regression is used to perform supervised dimensionality reduction on the standardized feature matrix in the training set; PLS By iteratively extracting latent variables that maximize the covariance between the input features and the output target, a low-dimensional feature representation is generated, which reduces the dimensionality. n components As one of the parameters to be optimized, a support vector regression initial prediction model with radial basis function as kernel function is constructed based on the low-dimensional feature representation and the standardized target vector, forming a joint prediction model to be optimized.

[0050] An improved Bayesian optimization algorithm is used to... PLS Dimensionality reduction and SVR Hyperparameters are used for integrated joint optimization; the joint parameter space specifically includes: PLS Dimensionality Reduction n components , SVR Regularization coefficient C Kernel function parameters γ, insensitive loss coefficient ε, and kernel function type kernel .

[0051] IBO The algorithm's optimization strategy includes the following key improvements: First, 10 sets of random probe points covering the entire parameter space are constructed and merged with 5 initial random exploration points to form an optimization starting point set, enhancing the ability to explore the global optimal solution; Based on the optimization starting point set, a Gaussian process is used to construct a surrogate model and 295 sequential iterations are performed using the expected improvement acquisition function, with the exploration factor set to 0.05; For discrete parameters such as kernel function type, an adaptive mapping and integer processing method for continuous-discrete mixed-type parameters is adopted, converting the continuous space optimal solution into a discrete optimal value through integer rounding or preset category mapping.

[0052] After 300 iterations of optimization, an optimal combination of parameters was finally determined: PLS The dimensionality reduction is 12. , , The kernel function type is rbfBased on this optimal parameter combination, an optimized joint prediction model is configured and generated.

[0053] The feature normalizer in the optimized joint prediction model, PLS Dimensionality reducer and SVR The predictor is stored as a single, independently loadable prediction model package in a serialized storage format; this model package does not include online training functionality, ensuring the stability and responsiveness of the prediction service.

[0054] In the predictive application phase, a set of candidate influencing factors for the hydropower project to be predicted is obtained, with the factor types, units of measurement, and order of arrangement strictly consistent with historical data; according to the steps... S 1. The 56 identified key impact factors were extracted from the candidate impact factor set to form a key impact factor value set.

[0055] The feature normalizer in the prediction model package is called to standardize the set of key influencing factor values, resulting in a standardized feature vector; [The text then repeats the command.] PLS The dimensionality reducer compresses the 56-dimensional standardized feature vector into a 12-dimensional low-dimensional feature representation; [call / implement] SVR Predictor based on RBF The kernel function performs nonlinear regression calculations and outputs standardized predicted values; finally, the inverse normalizer is called using the formula:

[0056] The standardized forecast values ​​are converted back into actual investment forecast values.

[0057] To test the performance of the prediction model, a coefficient of determination is selected. R ² Root mean square error RMSE and percentage error MAPE The expressions for these evaluation metrics are as follows:

[0058]

[0059]

[0060] in , , These are the actual value, the predicted value, and the average value, respectively.

[0061] Table 1 shows a comparison of model performance under different dimensionality reduction strategies; (not adopted) PLS Dimensional reduction IBO - SVR Model R ² is 0.867. MAPE It is 9.915%. RMSE The value is 6.055, while the present inventionIBO - PLS - SVR Model R ² Reaching 0.999, MAPE As low as 0.471%. RMSE Only 0.293, proving PLS Supervised dimensionality reduction plays a crucial role in improving prediction accuracy.

[0062] Table 1 Comparison of model performance under different dimensionality reduction strategies and optimization algorithms

[0063] Table 2 further compares the present invention with typical methods in the prior art; the results show that the model of the present invention is significantly superior to the model in all evaluation indicators. PSO - BPNN , PSO - LSSVM and their combination PLS The variant demonstrates the advantages of improved Bayesian optimization over particle swarm optimization in terms of optimization efficiency and accuracy, and SVR Stronger adaptability to small-sample nonlinear data in hydropower projects.

[0064] Table 2 Performance Comparison of the Invention and Existing Techniques

[0065] Table 3 compares the performance of different optimization algorithms. PLS - SVR Performance in model parameter tuning; the proposed method IBO The algorithm in RMSE , MAPE , R ² It outperforms traditional Bayesian optimization, ant colony optimization, genetic algorithm, particle swarm optimization and whale optimization algorithms in all three accuracy metrics, while its training time of only 114 seconds is significantly better than other algorithms, balancing global optimization accuracy and running efficiency.

[0066] Table 3 Evaluation of the model parameter tuning effect of different optimization algorithms

[0067] Table 4 shows the generalization ability of the model of the present invention in predicting different investment categories; for the three prediction tasks of static investment, dynamic investment, and underground powerhouse investment, the model of the present invention... IBO - PLS - SVR The model significantly outperforms in all metrics. PSO - BPNN and PSO - LSSVM By comparing different algorithms, it is demonstrated that this method has good adaptability and scalability.

[0068] Table 4 Comparison of model prediction performance for different investment categories

[0069] like Figure 2 - Figure 3 As shown, the present invention also provides a hydropower project investment prediction system based on causal screening and joint optimization. This system includes a user interaction module, a data storage module, and an execution prediction module. The user interaction module receives multiple candidate influencing factors input by the user and displays the predicted investment value. The data storage module persistently stores the input parameters, prediction results, and timestamps for each prediction. The execution prediction module is connected to both the user interaction module and the data storage module. The execution prediction module contains a joint prediction model constructed and optimized according to the above method. The execution prediction module is configured to: load the joint prediction model, receive candidate influencing factors from the user interaction module, output the predicted investment value to the user interaction module, and send the prediction record to the data storage module.

[0070] The prediction module is equipped with a parameter verification unit, which is used to verify whether there are missing values ​​or values ​​exceeding the preset range of multiple candidate influencing factors input by the user before calling the joint prediction model. If the verification passes, the data is sent to the joint prediction model. If the verification fails, a prompt message is generated and returned to the user interaction module. The prediction module contains the optimized joint prediction model package constructed by the aforementioned method.

[0071] The system is based on lightweight Web The service framework enables local deployment, without relying on external networks or cloud services. Users input 56 key parameters of the hydropower project to be predicted through a web form interface. The parameter verification unit inside the system will automatically detect missing values ​​or outliers that exceed the preset range and prompt for correction. After the verification is passed, the prediction module will execute the standardization, dimensionality reduction, regression prediction and inverse standardization processes in sequence and return the investment prediction value to the user interface.

[0072] The data storage module uses an embedded relational database to automatically record the project name, all input parameters, prediction results, and operation timestamps for each prediction, forming a traceable operation log that supports historical queries and scheme comparisons. The entire system forms an engineered closed loop of "input-verification-prediction-storage-display," which is suitable for rapid investment estimation scenarios in the project feasibility study stage.

[0073] In summary, this embodiment fully verifies the feasibility, high accuracy, and practicality of the present invention in real hydropower engineering scenarios. Through a technical chain of causal inference screening, supervised dimensionality reduction, improved Bayesian joint optimization, and lightweight deployment, it effectively solves the problems of feature redundancy, high-dimensional nonlinear modeling bottlenecks, and low efficiency in hyperparameter tuning in existing technologies, demonstrating significant technological advancements and broad engineering application value. The present invention uses dual machine learning causal inference to screen key influencing factors, eliminating non-causal noise at the source and overcoming the overfitting problem caused by feature redundancy in traditional methods. It employs partial least squares regression for supervised dimensionality reduction, fully utilizing investment label information to retain the most discriminative low-dimensional feature subspace, significantly improving the robustness of modeling under small-sample high-dimensional data.

[0074] This invention improves Bayesian optimization by incorporating dimensionality reduction and support vector regression hyperparameters into the same space for joint optimization, breaking the two-step fragmented optimization mode. It combines probe point enhancement exploration and hybrid parameter adaptive mapping mechanism to ensure the best fit between dimensionality reduction features and prediction model, balancing global optimization accuracy and computational efficiency. This invention also solidifies the normalizer, dimensionality reducer and predictor into an independently loadable model package, supporting offline deployment and rapid invocation.

[0075] The above description is merely a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural or procedural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting investment in hydropower projects using causal screening and joint optimization, characterized in that, Includes the following steps: S 1. Obtain candidate influencing factors and corresponding investment data of historical hydropower projects, use dual machine learning algorithms to perform causal inference on the candidate influencing factors, calculate the average processing effect value of each candidate influencing factor on the investment data, and screen out key influencing factors according to the preset causal effect threshold. S 2: Based on the key influencing factors and the investment data, construct a joint prediction model to be optimized, which includes partial least squares regression and support vector regression; wherein, the partial least squares regression is used to perform supervised dimensionality reduction on the key influencing factors, and the support vector regression is used to make investment predictions based on the dimensionality-reduced features. S 3: An improved Bayesian optimization algorithm is adopted to incorporate the dimensionality reduction dimension of the partial least squares regression and the hyperparameters of the support vector regression into the same parameter space for joint optimization, so as to determine the optimal parameter combination and obtain the optimized joint prediction model. S 4: Obtain the candidate influencing factor set for the hydropower project to be predicted, and based on the above... S The key impact factors identified in step 1 are used to extract corresponding data from the candidate impact factor set to form a key impact factor value set; S 5: Input the set of key influencing factor values ​​into the... S In the optimized joint prediction model obtained in the three steps, standardization, supervised dimensionality reduction, regression prediction and inverse standardization are performed sequentially by the optimized joint prediction model to output the investment prediction value of the hydropower project to be predicted.

2. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that, The S The specific steps include: Obtain candidate influencing factors and corresponding investment data for the historical hydropower projects; employ a dual machine learning algorithm to construct a first machine learning model to perform residual processing on the candidate influencing factors as processing variables, obtaining a first residual; construct a second machine learning model to perform residual processing on the investment data as the outcome variable, obtaining a second residual; and calculate the average processing effect value of each candidate influencing factor based on the first residual and the second residual, obtaining an effect value set. The candidate influencing factors whose average absolute value of the effect value is greater than the preset causal effect threshold are identified as the key influencing factors.

3. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that, The S The two steps specifically include: The key influencing factors and the investment data are respectively subjected to mean-standard deviation standardization to obtain a standardized feature matrix and a standardized target vector. Using the standardized target vector as a supervision signal, partial least squares regression is used to perform supervised dimensionality reduction on the standardized feature matrix. By iteratively extracting the latent variables that maximize the covariance between the input features and the output target, a low-dimensional feature representation is generated. Based on the low-dimensional feature representation and the standardized target vector, an initial prediction model is constructed using support vector regression with radial basis function as kernel function, and the joint prediction model to be optimized is obtained.

4. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that: The S In step 3, the parameter space for joint optimization by the improved Bayesian optimization algorithm includes: the dimensionality reduction dimension of partial least squares regression, the regularization coefficient of support vector regression, the kernel function parameter, the insensitive loss coefficient, and the kernel function type.

5. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 4, characterized in that, The S The improved Bayesian optimization algorithm in step 3 includes the following optimization strategies: Construct multiple sets of random probe points covering the entire range of the parameter space, and merge them with the initial random exploration points to obtain an optimized starting point set; Based on the optimized starting set, a surrogate model is constructed using a Gaussian process, and the expected improved acquisition function is used for iterative optimization to obtain the optimal solution in continuous space. For discrete parameters contained in the parameter space, the optimal solution in the continuous space obtained by optimization is converted into a discrete optimal value by integer rounding or preset category mapping in order to determine the final parameter combination. Based on the final parameter combination, the optimized joint prediction model is configured and generated.

6. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that, The S In step 5, the internal processing of the optimized joint prediction model includes: The feature normalizer in the optimized joint prediction model is invoked to standardize the set of key influencing factor values, thereby obtaining a standardized feature vector. The partial least squares regression dimensionality reducer in the optimized joint prediction model is invoked to reduce the dimensionality of the standardized feature vector, thereby obtaining a low-dimensional feature representation. The support vector regression predictor in the optimized joint prediction model is invoked to perform regression prediction on the low-dimensional feature representation to obtain standardized prediction values. The inverse standardizer in the optimized joint forecasting model is invoked to perform inverse standardization on the standardized forecast value, and the investment forecast value is obtained and output.

7. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that, The method also includes a model solidification step: The S The feature normalizer, partial least squares regression dimensionality reducer, and support vector regression predictor in the optimized joint prediction model obtained in the three steps are solidified as a whole in a serialized storage format to generate a prediction model package that can be loaded independently.

8. The hydropower project investment prediction method based on causal screening and joint optimization according to claim 1, characterized in that, The S The candidate impact factor set for the hydropower project to be predicted in step 4, including the type, unit of measurement, and order of the candidate impact factors, is related to the... S In step 1, the types, units of measurement, and order of candidate influencing factors for historical hydropower projects should remain consistent.

9. A hydropower project investment forecasting system based on causal screening and joint optimization, characterized in that, include: The user interaction module is used to receive multiple candidate influencing factors input by the user and display the investment forecast values. The data storage module is used to persistently save the input parameters, prediction results, and timestamps for each prediction. The prediction module is connected to both the user interaction module and the data storage module. The execution prediction module contains a joint prediction model constructed and optimized according to any one of claims 1 to 8; The execution prediction module is configured to: load the joint prediction model, receive candidate influencing factors from the user interaction module, output investment prediction values ​​to the user interaction module, and send prediction records to the data storage module.

10. The hydropower project investment prediction system based on causal screening and joint optimization according to claim 9, characterized in that, The execution prediction module is equipped with a parameter verification unit, which is used to verify whether there are missing values ​​or values ​​exceeding the preset range of multiple candidate influencing factors input by the user before calling the joint prediction model. If the verification passes, the data is sent into the joint prediction model; if the verification fails, a prompt message is generated and returned to the user interaction module.

Images