Power grid investment optimization method and system based on multi-head neural network and uncertainty quantification

Abstract

The invention discloses a power grid investment optimization method and system based on a multi-head neural network and uncertainty quantification, and relates to the technical field of power system planning and decision making. The method comprises the following steps: firstly, constructing a unified data interface to access multi-source heterogeneous data of a power grid planning project, carrying out anomaly detection through an isolated forest algorithm, and filling a missing value by adopting XGBoost regression prediction; the numerical features are standardized by adopting quantile transformation, and the category features are vectorized by adopting CatBoost coding. On this basis, a multi-head output deep learning model based on a multi-layer perceptron is constructed, and an attention mechanism and a Monte Carlo Dropout method are integrated in the network. And finally, on the basis of a comprehensive evaluation model fusing fuzzy evaluation and grey correlation analysis, the optimal investment portfolio is optimized and solved under the budget constraint. According to the method, accurate prediction and uncertainty quantification of power grid planning benefits are realized, and scientificity and reliability of investment decision making are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system planning and decision-making technology, and in particular to a power grid investment optimization method and system based on multi-head neural networks and uncertainty quantification. Background Technology

[0002] Power grid benefit assessment is a core component of power system planning and investment decision-making, traditionally relying primarily on qualitative analysis based on expert experience and quantitative analysis using linear statistical models. Expert-based methods employ techniques such as the Delphi method and the analytic hierarchy process (AHP) for comprehensive evaluation, while statistical models utilize multiple regression and path analysis to establish mathematical relationships between influencing factors and benefit indicators. These methods have developed into a relatively complete process system through long-term power grid planning practice.

[0003] However, existing technologies have significant limitations: expert experience methods are highly subjective in their evaluation results, making it difficult to establish precise quantitative relationships and reveal the interaction mechanisms between influencing factors; traditional statistical models are based on linear assumptions, making it difficult to accurately characterize the nonlinear dynamic characteristics that are prevalent in power grid systems, and the prediction accuracy is significantly reduced in new scenarios such as high-proportion renewable energy access; at the same time, existing methods have insufficient ability to fuse and process multi-source heterogeneous data, and cannot effectively utilize multimodal information such as structured numerical data, unstructured text, geospatial data, and real-time monitoring data.

[0004] With the deepening of the construction of new power systems and the continuous reform of the power market, the decision-making environment for power grid planning is becoming increasingly complex, requiring simultaneous consideration of new factors such as clean energy consumption, extreme weather response, and digital transformation. Traditional methods are no longer sufficient to meet practical needs in processing massive amounts of multi-source heterogeneous data, capturing complex nonlinear relationships, and providing reliable uncertainty quantification, resulting in a lack of scientific basis for investment decisions, waste of resources, and low investment efficiency. Summary of the Invention

[0005] One of the objectives of this invention is to provide a power grid investment optimization method based on multi-head neural networks and uncertainty quantification, so as to solve the shortcomings of traditional methods in data processing, model adaptability and uncertainty quantification, and improve the scientificity and reliability of power grid planning and investment decisions.

[0006] To address the aforementioned technical problems, this invention adopts the following technical solution: a power grid investment optimization method based on multi-head neural networks and uncertainty quantification, comprising the following steps:

[0007] S1. Data Processing: Access multi-source heterogeneous data through a unified interface, and sequentially perform anomaly detection, missing value imputation, feature standardization and encoding, and automatically filter out key feature subsets using a differentiable feature selection network;

[0008] S2. Model Building and Prediction: Construct a multi-head output neural network based on a multilayer perceptron (MLP), integrating the attention mechanism and the Monte Carlo Dropout method;

[0009] S3. Evaluation and Optimization: The comprehensive score of the project is calculated by integrating fuzzy evaluation and grey relational analysis, and the optimal portfolio is solved based on evolutionary algorithm under budget and benefit constraints.

[0010] Preferably, step S1 includes:

[0011] S11. Establish a unified data interface layer to access structured tabular data, text reports, GIS geospatial data, and time-series monitoring data in parallel.

[0012] S12. Use the isolated forest algorithm for outlier detection and set an adaptive threshold to identify outliers in the data.

[0013] S13. Use a gradient boosting tree-based prediction model to intelligently fill in missing values;

[0014] S14. Numerical features are standardized using quantile transformation, and high cardinality categorical features are vectorized using supervised coding techniques.

[0015] S15. End-to-end feature selection is achieved through a differentiable feature selection network, and its calculation formula is as follows:

[0016] ;

[0017] in, For the input feature vector, For learnable feature weight parameters, For the Sigmoid function, This indicates element-wise multiplication.

[0018] More preferably, in step S14, the formula for the quantile transformation is:

[0019] ;

[0020] in, Let be the empirical distribution function. It is the quantile function of the standard normal distribution.

[0021] More preferably, in step S14, the method of vectorizing high cardinality category features using supervised coding is as follows:

[0022] CatBoost encoding is used to convert high-cardinality categorical features into supervised numerical representations. The dataset is sorted in chronological or random order, and for each sample, statistics are calculated using only the preceding samples. For the i-th sample of a categorical feature, its encoded value is:

[0023] ;

[0024] Where, sum_of_previous_targets is the sum of the target values ​​of this class in the previous samples; count_of_previous_occurrences is the number of times this class appears in the previous samples; prior is the global target mean; and smoothing is the smoothing coefficient, which controls the weight of the prior.

[0025] Then, K-fold cross-validation is performed, dividing the data into K folds, and the encoding of each fold is calculated using data from the other K-1 folds. This effectively avoids data leakage and ensures the reliability of the encoding.

[0026] More preferably, step S2 includes:

[0027] S21. Introducing a multi-head output structure: After the shared hidden layer, three independent multilayer perceptron-based sub-networks are connected in parallel as output heads, which are used for regression prediction of operational efficiency, financial efficiency and social efficiency indicators, respectively.

[0028] S22. At the end of the shared hidden layer, an attention network layer is embedded to dynamically learn and weight the importance of different input features to the prediction results.

[0029] S23. Apply the Monte Carlo Dropout method to all hidden layers of the network to transform the multilayer perceptron-based network into a Bayesian approximation network, and output the probability distribution of the predicted value through multiple forward propagations.

[0030] S24. Many indicators for power grid planning benefit prediction are based on continuous numerical predictions, employing the Huber loss function to ensure the stability and robustness of the prediction results.

[0031] ;

[0032] in, Actual measured values ​​representing operational, financial, or social benefits; This represents the model's predicted value; It is a threshold parameter used to control the behavior of the loss function when the prediction error is less than or equal to When the error is greater than 1, the loss function is expressed as squared loss; when the error is greater than 1, the loss function is expressed as squared loss. When the loss is linear, it exhibits a linear loss pattern.

[0033] More preferably, step S3 includes:

[0034] S31. Establish a comprehensive evaluation model that integrates fuzzy evaluation and grey relational analysis, and calculate the comprehensive score of the project:

[0035] ;

[0036] in, For fuzzy membership degree scoring, The grey relational degree between the project and the ideal solution;

[0037] S32. Design an improved genetic algorithm that integrates a local search algorithm and an adaptive genetic algorithm, and introduces a hill-climbing method for local fine-tuning. Under the conditions of satisfying the total investment budget and individual benefit constraints, solve for the optimal investment portfolio. The improved fitness function is:

[0038] ;

[0039] in, The total benefit score;

[0040] Score the budget utilization rate;

[0041] Score for diversity;

[0042] This is a budget penalty item.

[0043] More preferably, in step S31:

[0044] The fuzzy evaluation includes the following steps:

[0045] Constructing a benefit indicator matrix:

[0046] ;

[0047] Among them, the sub-indicators of each benefit are:

[0048] ;

[0049] Fuzzy membership degree calculation uses the trapezoidal membership function:

[0050] ;

[0051] When calculating operational benefits, a, b, c, and d are taken as 0.6, 0.7, 0.9, and 1.0, respectively; when calculating financial benefits, a, b, c, and d are taken as 0.08, 0.12, 0.20, and 0.25, respectively; when calculating social benefits, a, b, c, and d are taken as 0.5, 0.7, 0.9, and 1.0, respectively.

[0052] The grey relational analysis includes the following steps:

[0053] The historical best values ​​for each indicator are used as a reference sequence:

[0054] ;

[0055] Perform correlation coefficient calculation:

[0056] ;

[0057] in, This indicates a double minimum operation. This indicates a double maximum operation. ρ is the resolution coefficient, which is set to 0.5.

[0058] More preferably, in step S31, the comprehensive evaluation model integrating fuzzy evaluation and grey relational analysis is as follows:

[0059] ;

[0060] in, For fuzzy evaluation scores; The score is the grey relational score; .

[0061] In addition, the present invention also provides a power grid investment optimization system based on multi-head neural networks and uncertainty quantification, which includes:

[0062] The data processing module is configured to access multi-source heterogeneous data through a unified interface, perform anomaly detection, missing value imputation, feature standardization and encoding in sequence, and automatically filter out key feature subsets using a differentiable feature selection network.

[0063] The model building and prediction module is configured to build a multi-head output neural network based on a multilayer perceptron, integrating the attention mechanism and the Monte Carlo Dropout method.

[0064] The evaluation and optimization module is configured to calculate the comprehensive score of the project by integrating fuzzy evaluation and grey relational analysis, and to solve for the optimal investment portfolio under budget and benefit constraints based on an evolutionary algorithm.

[0065] The system operates using the method described above.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] 1. The constructed intelligent data processing pipeline realizes the automated fusion of multi-source heterogeneous data, and improves the accuracy of prediction results by optimizing the quality of the dataset.

[0068] 2. By using a multi-head output neural network architecture based on MLP and combining it with an attention mechanism, this invention can accurately capture the complex nonlinear relationship between power grid planning benefits and various influencing factors.

[0069] 3. Introduce the Monte Carlo Dropout method to quantify uncertainty and provide decision-makers with a complete risk analysis.

[0070] 4. Fuzzy grey comprehensive evaluation is used to further improve the prediction results to be closer to expert judgment. Attached Figure Description

[0071] Figure 1 This is a schematic diagram illustrating the overall framework of the principle of this invention;

[0072] Figure 2 This is a schematic diagram of the network structure of the multilayer perceptron in the embodiment;

[0073] Figure 3 This is a schematic diagram of the neuron model in the embodiment. Detailed Implementation

[0074] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0075] The principle framework of the power grid investment optimization system based on multi-head neural networks and uncertainty quantification in this embodiment is as follows: Figure 1 As shown, the system adopts a hierarchical design, realizing fully automated decision-making from multi-source heterogeneous data access (layer 1), through intelligent data processing (layer 2) and deep learning prediction (layer 3), to comprehensive evaluation and optimization (layer 4). It innovatively introduces path analysis prior knowledge fusion and Monte Carlo Dropout uncertainty quantification mechanisms.

[0076] The power grid investment optimization method based on multi-head neural networks and uncertainty quantification for implementing the above system specifically includes the following steps:

[0077] S1. Intelligent Data Processing: Accesses multi-source heterogeneous data through a unified interface, performs anomaly detection, missing value imputation, feature standardization and encoding in sequence, and automatically filters out key feature subsets using a differentiable feature selection network.

[0078] Specifically:

[0079] S11. Anomaly detection is performed using the Isolation Forest algorithm. The anomaly scores are calculated as follows:

[0080] ;

[0081] Where h(x) is the path length of sample x in the isolated tree, E(h(x)) is the expected path length in all trees, and c(n) is the average path length of the binary search tree.

[0082] Based on business needs or experience, a threshold ε for anomaly scores is set (e.g., ε = 0.6). In this invention, we set this threshold by combining grid search with business verification. All samples with s(x) > ε are marked as outliers, and their indices are output for subsequent data repair or removal operations.

[0083] S12. Using features with missing values ​​as new prediction targets and other complete features as input, train an XGBoost regression prediction model to predict missing values.

[0084] Determine which features in the dataset have missing values ​​and divide the dataset into a training set and a validation set. The training set consists of samples with non-missing features, and the validation set consists of samples with missing features.

[0085] The XGBoost regressor is an efficient and flexible model that handles categorical features well and fully utilizes all information in the dataset. Its core components are the loss function L and the regularization Ω.

[0086] ;

[0087] in, T is the number of leaf nodes, w is the leaf weight, and γ and λ are regularization parameters.

[0088] The trained XGBoost model is used to predict the prediction set, and the prediction results of the missing values ​​are used to fill in the corresponding actual positions in the original dataset.

[0089] S13. To further improve data quality and reduce the impact of outliers on the system, feature encoding and transformation are performed on the original data to convert it into a format suitable for machine learning model processing.

[0090] Mapping using the quantile function of the target distribution, the core formula for quantile transformation is:

[0091] in, Let be the empirical distribution function. It is the quantile function of the standard normal distribution.

[0092] S14. CatBoost encoding is used to convert high cardinality class features into supervised numerical representations, avoiding the dimensionality explosion problem of one-hot encoding.

[0093] Sort the dataset in chronological or random order and, for each sample, calculate the statistic using only the "first" samples.

[0094] For the i-th sample of the categorical feature, its encoded value is:

[0095] ;

[0096] Where: sum_of_previous_targets is the sum of the target values ​​of this class in the previous samples; count_of_previous_occurrences is the number of times this class appears in the previous samples; prior: global target mean (smoothing prior); smoothing: smoothing coefficient, which controls the weight of the prior.

[0097] K-fold cross-validation divides the data into K folds, and uses the data from the other K-1 folds to calculate the code for each fold. This effectively avoids data leakage and ensures the reliability of the code.

[0098] S15. Differentiable Feature Selection Network Implementation: After feature encoding, a differentiable feature selection network is introduced to automatically filter feature vectors. The core formula of this network is:

[0099] ;

[0100] in, For the input feature vector, For learnable feature weight parameters, For the Sigmoid function, This indicates element-wise multiplication.

[0101] S2. Deep Learning Model Construction and Prediction: Construct a multi-head output neural network based on a multilayer perceptron (MLP), integrating attention mechanism and Monte Carlo Dropout method.

[0102] Specifically:

[0103] S21, integrated attention mechanism, Monte Carlo Dropout method and multilayer perceptron (MLP) model.

[0104] A multilayer perceptron is a simple feedback-forward neural network. In a neural network, it typically contains one input layer, one output layer, and multiple hidden layers, such as... Figure 2As shown, the input layer is only responsible for receiving input information from the network model. It typically uses the feature vectors of the processed data samples as input, which are then further processed by multiple hidden layers and output layers. Finally, the output layer outputs the experimental results. The number of nodes and layers in the hidden layers can be adjusted according to the actual experimental situation. The multilayer perceptron learns the mapping relationship between input information and output results. To make the model's predictions more accurate, the corresponding parameter weights and bias values ​​are continuously updated and adjusted during model training. Activation functions are typically added between each layer in the network to introduce non-linear factors.

[0105] The basic unit of a multilayer perceptron is called a neuron, and the model of a neuron is as follows: Figure 3 As shown, a neuron receives information from n other neurons as input signals, and then weights the received input signals according to their weights. If the stimulus intensity is higher than a threshold, the final output needs to be further processed by an activation function and obtained through a nonlinear transformation. The neuron model can be represented as:

[0106] ;

[0107] In the formula, x i Indicates input data, w ij Represents the weight, θ j Represents the threshold, y i represents the output, and f represents the activation function.

[0108] Before using a multilayer perceptron (MLP) to perform tasks such as prediction and regression, the network model needs to be trained. This involves adjusting the weights of the parameters in the network model using existing data samples. The network model is considered well-trained when the difference between its output and the labeled values ​​of the data samples is very small. Typically, MLPs require multiple training rounds, each with forward and backward propagation, and parameter adjustments are made using gradient descent.

[0109] Forward propagation involves using the output signal of a neuron in the previous layer as the input to the next layer in the network model, calculating the output of the next layer, and so on, until the information is passed to the output layer. Let l represent the number of layers in the neural network, and a (l) W represents the output of the neural layer. (l) With b (l) f represents the weights and biases between layer l-1 and layer l. l Let l represent the activation function of layer l. Then the forward propagation process of the network can be represented as:

[0110] ;

[0111] In the formula, the weight W (l)These are the parameters that the network needs to learn.

[0112] Neural networks continuously adjust their parameters to minimize the loss function, which evaluates the difference between the network's predictions and the actual values. Huber loss is commonly used in this approach.

[0113] ;

[0114] Backpropagation transmits the error back to each layer of the network and updates the weight parameters through gradient descent until the error value is lower than the predetermined standard in the network.

[0115] Attention mechanisms involve filtering out a small amount of important information from a large amount of data, allowing the model to focus on this crucial information while ignoring the majority of less important information. Using attention mechanisms, weights can be calculated for data signals, enabling the machine to select the more important signals for learning.

[0116] The basic principle of attention mechanisms is to imagine the components in the original data (Source) as a series of <key, value> pairs. Given a target value element (Query), the weight coefficient for each key's value is obtained by calculating the similarity or relevance between the Query and each key. This weighted sum is then used to obtain the final attention value. Therefore, the essence of attention mechanisms is to sum the values ​​of elements in the original data and calculate the weight coefficients for the corresponding values ​​using the Query and Key. The basic idea of ​​attention mechanisms can be summarized by the following formula:

[0117] ;

[0118] Note the specific calculation process of the mechanism, which can be summarized into three stages. The first stage is calculating the Query and Key. i The similarity and relevance are determined by calculating the vector dot product between the two pairs. The formula is as follows:

[0119] .

[0120] The second stage involves normalizing the similarity scores obtained in the first stage. Due to the different calculation methods, the numerical range of the similarity scores varies. Therefore, the similarity scores obtained in the first stage are transformed using the SoftMax calculation method. The SoftMax method constrains the calculated similarity scores to a probability distribution where the sum of the weights of all elements is 1. Furthermore, the inherent mechanism of SoftMax further emphasizes the weights of important elements. Weight coefficient 'a' i The formula is as follows:

[0121] ;

[0122] In the third stage, based on the weighting coefficient a i The Value is weighted to obtain the final attention value, as shown in the following formula:

[0123] ;

[0124] An attention mechanism is introduced after the last hidden layer of the MLP, forming an MLP network with an incorporated attention mechanism. The first three layers of this network are MLP hidden layers, the fourth layer introduces an attention layer, and the fifth layer is the output layer. Dropout layers are inserted after each hidden layer and after each attention layer of the model, maintaining Dropout activation and performing multiple random forward propagations.

[0125] S3. Comprehensive Evaluation and Optimization: The comprehensive score of the project is calculated by integrating fuzzy evaluation and grey relational analysis, and the optimal investment portfolio is solved based on evolutionary algorithms under budget and benefit constraints.

[0126] Specifically:

[0127] S31. Perform fuzzy comprehensive evaluation:

[0128] Constructing a benefit indicator matrix:

[0129] ;

[0130] Sub-indicators of each benefit:

[0131] ;

[0132] Fuzzy membership degree calculation uses the trapezoidal membership function:

[0133] ;

[0134] When calculating operational benefits, a, b, c, and d are taken as 0.6, 0.7, 0.9, and 1.0, respectively; when calculating financial benefits, a, b, c, and d are taken as 0.08, 0.12, 0.20, and 0.25, respectively; and when calculating social benefits, a, b, c, and d are taken as 0.5, 0.7, 0.9, and 1.0, respectively.

[0135] S32. Perform grey relational analysis:

[0136] The historical best values ​​for each indicator are used as a reference sequence:

[0137] ;

[0138] The formula for calculating the correlation coefficient is as follows:

[0139] ;

[0140] in, ρ is the resolution coefficient, taken as 0.5.

[0141] S33. A comprehensive evaluation is conducted by combining fuzzy evaluation and grey relational analysis. The formula is as follows:

[0142] ;

[0143] in: For fuzzy evaluation scores; The score is the grey relational score; .

[0144] Calculate the overall score of the project:

[0145] ;

[0146] in, For fuzzy membership degree scoring, The grey relational degree between the project and the ideal solution.

[0147] S34. An improved genetic algorithm specifically designed for the power grid planning portfolio optimization problem is used to optimize the portfolio. The improved genetic algorithm integrates a local search algorithm and an adaptive genetic algorithm, and introduces a hill-climbing method for local fine-tuning. Under the conditions of satisfying the total investment budget and individual benefit constraints, it solves for the optimal portfolio. The improved fitness function is:

[0148] ;

[0149] in, The total benefit score;

[0150] Score the budget utilization rate;

[0151] Score for diversity;

[0152] This is a budget penalty item.

[0153] Furthermore, to ensure the feasibility and efficiency of the optimization process, the specific computational details are expanded. The following formula forms the basis for the implementation of the genetic operator and is closely integrated with the aforementioned comprehensive evaluation model:

[0154] 1. Initialization.

[0155] Set the generation counter t=0, set the maximum generation Tmax=500, and use a heuristic strategy to generate N=100 individuals as the initial population P(0).

[0156] 2. Individual evaluation.

[0157] Calculate the fitness of each individual in the population P(t):

[0158] in, For diversity entropy; The selection ratio for the project is λ=100, which is the penalty coefficient.

[0159] 3. Selection operation.

[0160] Apply the improved selection operator to the population:

[0161] Elite Selection: Choose the one with the highest adaptability. Each individual directly enters the next generation.

[0162] Tournament selection: For the remaining 90 positions, an adaptive tournament selection is used; 3 individuals are randomly selected from the population as participants;

[0163] Selection probability: .

[0164] 4. Cross operation.

[0165] Apply the improved crossover operator to the selected pairs of individuals:

[0166] Crossover probability:

[0167] ;

[0168] Crossover method (equal probability selection):

[0169] Uniform crossover: ,in ;

[0170] Two-point intersection: random selection ,exchange Gene segments within an interval.

[0171] 5. Mutation operation.

[0172] Apply the improved mutation operator to the population:

[0173] Adaptive variability rate:

[0174] ;

[0175] Mutation strategy:

[0176] Smart replacement mutation (30% probability):

[0177] , ;

[0178] , ;

[0179] if and ;

[0180] Then perform the replacement: .

[0181] Bit flip mutation: flip each gene bit with probability pm.

[0182] 6. Repair calculation.

[0183] Implement remedial strategies for individuals who violate budget constraints:

[0184] While :

[0185] Sort by ri in ascending order

[0186] for j=1 to ∣O∣:

[0187] xO[j]=0

[0188] if break.

[0189] 7. Local search.

[0190] Perform a hill-climbing local search on the individual with a probability of 0.2:

[0191] ;

[0192] in, .

[0193] 8. Population renewal.

[0194] After selection, crossover, mutation, repair and local search operations, the next generation population P(t+1) is obtained from the population P(t). Its fitness value is calculated and sorted.

[0195] The algorithm is iterated continuously, and the optimal result is obtained when the termination condition is met: t≥Tmax=500, or the optimal fitness improvement is less than 0.1% for 10 consecutive generations.

[0196] Experimental results.

[0197] Table 1. Operational Benefit Forecast and Uncertainty Quantification

[0198]

[0199] This table presents the predicted operational benefits and uncertainty quantification results for eight power grid planning projects. Operational benefit indicators include project load rates, with actual values ​​ranging from 0.85 to 0.97. The predicted mean is highly close to the actual value (e.g., the actual value for the Hegang 500 kV transmission and transformation project is 0.94, and the predicted mean is 0.92). The standard deviation is controlled within the range of 0.02-0.05, indicating high prediction accuracy. The 95% confidence interval is narrow (the difference between the upper and lower limits is 0.08-0.12). For example, the interval for the Daqing Linyuan Thermal Power Plant 220 kV transmission project is [0.91, 0.99], covering the actual value of 0.97, demonstrating that the model can reliably quantify the predicted uncertainty and provide risk reference for decision-making.

[0200] Table 2 Financial Benefit Forecasts and Uncertainty Quantification

[0201]

[0202] This table presents the financial benefit forecast results, focusing on the accuracy of the project investment forecast (unit: RMB 10,000). The actual investment decision differs slightly from the forecasted final investment (e.g., the Heihe Anbei 500 kV transmission and transformation project: actual RMB 474,177,900, forecast RMB 472,000,000), but the standard deviation is relatively large (RMB 12,000,000-42,000,000), reflecting the high volatility of the investment data. The 95% confidence interval is moderately wide (e.g., the Ning'an Biomass Power Plant 110 kV transmission project: interval [RMB 2,560,000, 3,540,000]), with actual values ​​all falling within the interval, verifying the model's effectiveness in quantifying financial risk, especially suitable for budget control of large-scale investment projects.

[0203] Table 3 Social Benefit Prediction and Uncertainty Quantification

[0204]

[0205] This table assesses the predictive performance of social benefits. The actual social benefit scores range from 0.78 to 0.92, while the predicted mean is slightly lower than the actual value (e.g., the actual score for the Heihe Anbei 500 kV transmission and transformation project is 0.92, and the predicted score is 0.89). However, the standard deviation is small (0.03-0.06), indicating good predictive stability. The 95% confidence interval (e.g., the interval [0.82, 0.94] for the Heilongjiang Huangou pumped storage power station project) effectively captures the actual values, highlighting the model's advantage in quantifying the uncertainty of social benefits and helping to balance policy orientation and empirical analysis.

[0206] Table 4. Fuzzy Grey Comprehensive Evaluation Score (Top 15 Items)

[0207]

[0208] This table ranks the top 15 projects based on their comprehensive evaluation scores, incorporating both fuzzy membership scores and grey relational analysis. The Heihe Anbei 500kV transmission and transformation project ranks first with a comprehensive score of 0.91. Its high fuzzy score (0.92) and grey relational analysis (0.90) indicate that the project is close to an ideal solution in terms of operation, financial benefits, and social impact. The scores range from 0.55 (ranked 15th) to 0.91, showing a clear gradient and reflecting the differences between projects. The combination of fuzzy and grey scores (each weighted 50%) enhances the objectivity of the evaluation and provides a scientific basis for prioritizing investment.

[0209] Table 5. Portfolio Optimization Results (Budget Constraint: 3 billion RMB)

[0210]

[0211] The portfolio optimization results table shows that, under a budget constraint of 3 billion yuan, the improved genetic algorithm successfully selected the top 6 high-scoring projects (such as the Heihe Anbei and Hegang projects) to form the optimal portfolio, with a total investment of 2,542,571,100 yuan and a total comprehensive score of 5.02. Although some subsequent projects scored relatively high (such as the Qianjin project ranked 7th with a score of 0.76), they were excluded because their inclusion would lead to budget overruns. This demonstrates the algorithm's effective balance between maximizing benefits and project diversity under strict budget control.

[0212] To facilitate understanding by those skilled in the art of the improvements of this invention over the prior art, some of the accompanying drawings and descriptions have been simplified. The above embodiments are preferred implementations of this invention. In addition, this invention can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this invention.

Claims

1. A power grid investment optimization method based on a multi-head neural network and uncertainty quantification, characterized in that, The method comprises the following steps: S1, data processing: accessing multi-source heterogeneous data through a unified interface, sequentially performing anomaly detection, missing value filling, feature standardization and coding, and automatically screening out a key feature subset by using a differentiable feature selection network; S2, model construction and prediction: constructing a multi-head output neural network based on a multilayer perceptron, integrating an attention mechanism and a Monte Carlo Dropout method; S3, evaluation and optimization: calculating a comprehensive score of a project by fusing fuzzy evaluation and grey correlation analysis, and solving an optimal investment portfolio under the constraints of budget and benefit based on an evolutionary algorithm.

2. The power grid investment optimization method based on multi-head neural network and uncertainty quantification according to claim 1, characterized in that, Step S1 comprises: S11, establishing a unified data interface layer and accessing structured table data, text reports, GIS geographic spatial data and time series monitoring data in parallel; S12, detecting outliers by using an isolation forest algorithm, and setting an adaptive threshold to identify data outliers; S13, intelligently filling missing values by using a prediction model based on a gradient boosting tree; S14, standardizing numerical features by using quantile transformation, and vectorizing high-base category features by using supervised coding technology; S15, realizing end-to-end feature screening by using a differentiable feature selection network, and the calculation formula is: ； wherein, is an input feature vector, is a learnable feature weight parameter, is a Sigmoid function, denotes element-wise multiplication.

3. The power grid investment optimization method based on multi-head neural network and uncertainty quantification of claim 2, wherein: In step S14, the formula of the quantile transformation is: ； wherein is an empirical distribution function, is a quantile function of the standard normal distribution.

4. The power grid investment optimization method based on multi-head neural network and uncertainty quantification of claim 3, wherein: In step S14, the way of vectorizing high-base category features by using supervised coding technology is: The high-base category features are converted into supervised numerical representation by using CatBoost coding, the data set is sorted in time sequence or random order, and for each sample, only the previous samples are used to calculate the statistics, and for the i-th sample of the category feature, the encoding value is: ； Wherein, sum_of_previous_targets is the sum of target values of the category in the previous samples; count_of_previous_occurrences is the number of occurrences of the category in the previous samples; prior is the global target mean; smoothing is a smoothing coefficient, which controls the weight of the prior; Then, K-fold cross-validation is performed, the data is divided into K folds, and for each fold, the data of other K-1 folds is used to calculate the encoding, which can effectively avoid data leakage and ensure the reliability of the encoding.

5. The power grid investment optimization method based on multi-head neural network and uncertainty quantification according to claim 4, characterized in that, Step S2 comprises: S21, introducing a multi-head output structure: connecting three independent multilayer perceptron subnetworks in parallel as output heads after the shared hidden layer, which are respectively used for regression prediction of operation benefit, financial benefit and social benefit indicators; S22, embedding an attention network layer at the end of the shared hidden layer, dynamically learning and weighting the importance of different input features to the prediction result; S23, applying the Monte Carlo Dropout method to all hidden layers of the network, converting the multilayer perceptron into a Bayesian approximation network, and outputting the probability distribution of the prediction value through multiple forward propagation; S24, the power grid planning benefit prediction multiple indicators are continuous numerical prediction, and the Huber loss function is used to ensure the stability and robustness of the prediction result: ； wherein, represents an actual measured value of operational or financial or social benefit; represents a predicted value of the model; is a threshold parameter that controls the behavior of the loss function, which behaves as a squared loss when the prediction error is less than or equal to and behaves as a linear loss when the error is greater than .

6. The power grid investment optimization method based on multi-head neural network and uncertainty quantification of claim 5, wherein, Step S3 comprises: S31, a comprehensive evaluation model of fusion fuzzy evaluation and grey correlation analysis is established, and a project comprehensive score is calculated: ； wherein, is the fuzzy membership score, is the grey correlation degree of the project with the ideal scheme; S32, a genetic algorithm is designed to integrate local search algorithm and adaptive genetic algorithm, and a hill climbing method is introduced for local fine tuning, under the conditions of meeting the total investment budget and single benefit constraints, an optimal investment portfolio is solved, and the fitness function is improved as: ； wherein, Total Benefit Score; is the budget utilization score; For diversity score; For budget penalty item.

7. The power grid investment optimization method based on multi-head neural network and uncertainty quantification of claim 6, wherein, In step S31: The fuzzy evaluation includes the following steps: An efficiency index matrix is constructed: ； Wherein, each efficiency sub-index: ； The fuzzy membership degree is calculated by using a trapezoidal membership function: ； When calculating the operation efficiency, a, b, c, and d are 0.6, 0.7, 0.9, and 1.0 respectively; when calculating the financial benefit, a, b, c, and d are 0.08, 0.12, 0.20, and 0.25 respectively; when calculating the social benefit, a, b, c, and d are 0.5, 0.7, 0.9, and 1.0 respectively; The grey correlation analysis includes the following steps: The historical optimal value of each index is taken as the reference sequence: ； The correlation coefficient is calculated: ； wherein denotes a double minimum operation, denotes a double maximum operation, ; p is a resolution factor, taken as 0.

5.

8. The power grid investment optimization method based on multi-head neural network and uncertainty quantification of claim 7, wherein: In step S31, the comprehensive evaluation model of fusion fuzzy evaluation and grey correlation analysis is: ； wherein, is the fuzzy evaluation score; is the grey correlation degree score; .

9. A power grid investment optimization system based on multi-headed neural networks and uncertainty quantification, characterized in that, Including: The data processing module is configured to access multi-source heterogeneous data through a unified interface, sequentially perform anomaly detection, missing value filling, feature standardization and coding, and automatically select a key feature subset by using a differentiable feature selection network; The model construction and prediction module is configured to construct a multi-head output neural network based on a multi-layer perceptron, integrate an attention mechanism and a Monte Carlo Dropout method; The evaluation and optimization module is configured to calculate a project comprehensive score by fusing fuzzy evaluation and grey correlation analysis, and solve an optimal investment portfolio under the constraints of budget and benefit based on an evolutionary algorithm.

10. The power grid investment optimization system based on multi-headed neural network and uncertainty quantification of claim 9, wherein: The method of any one of claims 1-8 is used.

Images