A Visualized Distribution Network Investment Strategy Optimization Method Incorporating Machine Learning

Abstract

This invention belongs to the field of power distribution technology, specifically disclosing a visual distribution network investment strategy optimization method combined with machine learning, including the following steps: S1, establishing an overall investment strategy evaluation system; S2, acquiring historical investment data from the power grid company and performing data preprocessing; S3, dividing each region into power supply zones according to load density, and using the TSO-XGBoost algorithm to train historical investment data for different power supply zones to generate an investment planning black box; S4, using the SHAP algorithm to analyze the generated black box and obtain a visualized investment contribution ratio; S5, combining the obtained weight ratios with the investment model to derive the optimal investment strategy that meets investment requirements. This invention employs the above-mentioned visual distribution network investment strategy optimization method combined with machine learning, using the SHAP algorithm to analyze the obtained "black box" to obtain a visualized proportion of each investment contribution. This facilitates clear guidance for power grid companies when precise investment is required.

Description

Technical Field

[0001] This invention relates to the field of power distribution technology, and in particular to a visual power distribution network investment strategy optimization method that combines machine learning. Background Technology

[0002] In recent years, to adapt to the rapid development of the economy and society, Guangdong Power Grid has invested heavily in upgrading urban and rural power grids, effectively supporting the rapid growth in electricity demand and the continuous improvement of power supply reliability and quality. However, with investment remaining at a high level, grassroots units have experienced varying degrees of problems, such as prioritizing investment over returns and project initiation over management. Therefore, establishing a precise investment control system, optimizing resource allocation, further improving investment efficiency and effectiveness, and promoting the company's high-quality development have become urgent issues to be addressed.

[0003] To improve asset operation and management, align with international standards, and achieve the optimal balance of asset risk, efficiency, and cost, Guangdong Power Grid Corporation, taking itself as an example, proposed an asset management strategy of "value for investment, full utilization of assets, monitored risks, improved performance, and excellent management." It also issued the "Guangdong Power Grid Corporation's Comprehensive Promotion Plan for Full Life Cycle Asset Management," drawing on advanced domestic and international asset management experience to achieve a qualitative leap in the company's asset management level. As the most crucial link in distribution network construction, distribution network planning should further enhance its scientific nature, fully reflecting the guiding role of planning in construction. While meeting various technical constraints, the concept of full life cycle asset management should be integrated into distribution network planning to ensure the economic efficiency of project selection and construction plans, thereby helping the company achieve its investment goals.

[0004] Currently, the company's distribution network planning approach is shifting from a problem-oriented approach to a reliability and target network structure-oriented approach. In the process of project formation and selection, it no longer focuses solely on the current state of the power grid, but rather on a more macro-level logic of overall network construction with reliability as the primary consideration. This places higher demands on precise investment and overall cost considerations for project investment. Currently, Guangdong Power Grid Company has proposed a standardized approach for the selection of distribution network planning and construction projects: generally, projects are first qualitatively categorized into five types based on investment strategies, and then project selection is based on factors such as the improvement of indicators before and after project implementation and the weight of key technical indicators. However, subjective influences cannot be completely avoided when determining indicator weights, and the planning stages and circumstances vary from place to place; this approach cannot incorporate local historical data to create site-specific investment strategies.

[0005] Therefore, to address the above problems, there is an urgent need for a visual distribution network investment strategy planning method that combines the XGBoost (ExtremeGradientBoosting) machine learning algorithm.

[0006] Approximate implementation scheme at present:

[0007] 1. Determining indicator weights based on the Analytic Hierarchy Process (AHP)

[0008] The Analytic Hierarchy Process (AHP) is applicable to problems involving multi-level indicators, complex dimensions, and many indicators that are difficult to quantify. This method is a decision-making approach based on a deep analysis of the inherent relationships between multiple objectives or multiple solutions to construct a reasonable multi-level analytical model. It then uses mathematical formulas to calculate the weights of all qualitative indicators at each level to solve complex multi-objective decision problems. AHP is used to determine the weights of an investment project's impact on investment benefits, thereby forming an investment plan.

[0009] 2. Determining index weights based on neural network models

[0010] This model can automatically identify and extract complex nonlinear relationships between investment and returns by learning from a large amount of historical investment data, making weight allocation more accurate. The neural network is highly adaptable, capable of handling high-dimensional and multi-type data, thus applicable to various decision-making scenarios. Using this method to analyze investment plans demonstrates good generalization ability. It maintains high efficiency even when dealing with large-scale investment data, thereby improving the flexibility and reliability of investment weight determination.

[0011] Disadvantages of existing technology:

[0012] 1) The weighting of indicators in the analytic hierarchy process (AHP) model is too subjective, failing to adequately consider the relationships and characteristics between data. For example, the comparison of the relative importance of different indicators. This may lead to results influenced by subjective viewpoints, thus potentially introducing uncertainty and bias. This method lacks objectivity and cannot reflect the "historical preferences" of investment strategies in different regions.

[0013] 2) The "black box" nature of neural network models makes the decision-making process opaque, making it difficult to explain the specific source and meaning of the weights, thus reducing the interpretability of the decision-making process. Furthermore, overfitting can affect the model's generalization ability, resulting in poor performance on new data. Finally, the solution of a neural network model may be a "regionally optimal solution" rather than a globally optimal solution, thus exhibiting limitations.

[0014] 3) All of the above schemes use a uniform predictive model to plan investment strategies for different regions, which fails to reflect the differences in investment. Summary of the Invention

[0015] The purpose of this invention is to provide a visual distribution network investment strategy optimization method that combines machine learning. The method uses the TSO-XGBoost algorithm to train historical investment data to obtain the "black box" of the investment strategy model. Then, the SHAP (SHapley Additive ex Planations) algorithm is used to obtain the visual index weights. Finally, the method combines the whole life cycle theory to obtain an objective investment strategy that incorporates historical preferences.

[0016] To achieve the above objectives, this invention provides a visual distribution network investment strategy optimization method combining machine learning, comprising the following steps:

[0017] S1. Establish a comprehensive investment strategy evaluation system;

[0018] S2. Obtain historical investment data from the power grid company and perform data preprocessing;

[0019] S3. Divide the power supply zones of each region according to load density, and use the TSO-XGBoost algorithm to train the historical investment data of different power supply zones to generate a black box for investment planning.

[0020] S4. Analyze the generated black box using the SHAP algorithm to obtain a visualized percentage of investment contribution.

[0021] S5. Based on the obtained weight ratios and the investment model, the optimal investment strategy that meets the investment requirements is derived.

[0022] Preferably, in S1, key investment projects and investment returns are extracted from the power grid company's investment habits, development prospects and investment needs over the years, and an indicator system for the investment evaluation model is established based on the power grid's power supply capacity, power quality, power supply reliability and economic benefits.

[0023] Preferably, in S2, the data preprocessing specifically includes:

[0024] Based on the power grid company's investment practices over the years, investment projects with the same or similar investment objectives are grouped into the same category;

[0025] Investment projects are categorized into eleven types based on their objectives: new substation outgoing lines to meet new load demand, replacement of old equipment or lines, resolution of safety hazards in medium and low voltage lines, improvement of medium voltage grid, resolution of distribution transformer overload, resolution of distribution transformer overload, resolution of low voltage in distribution areas, resolution of medium voltage line overload, resolution of medium voltage line overload, distribution automation projects, and new distribution areas to meet load demand.

[0026] Preferably, in S3, regions with similar power supply conditions are grouped into one category, and based on the differences in load density in each region, different regions are divided into six categories: A+, A, B, C, D, and E according to load density.

[0027] Preferably, the specific process of S3 is as follows:

[0028] The model was optimized using XGBoost machine learning and the Tuna Swarm Optimization Algorithm (TSO). TSO simulates two foraging behaviors of tuna swarms: spiral foraging and parabolic foraging. In TSO, each tuna represents a potential solution, and the entire swarm works together to find the optimal solution. The TSO tuna swarm optimization algorithm consists of three phases: search phase, chase phase, and attack phase.

[0029] Search phase: Individual tuna conduct random searches based on their own sensory abilities and information from the group, exploring possible solution spaces;

[0030] The pursuit phase: When an individual discovers a better solution, other individuals follow the best individual and pursue it, gradually approaching the optimal solution;

[0031] Attack phase: When an individual discovers the optimal solution, the group will concentrate its attack and eventually obtain the optimal solution;

[0032] Multiple weak learners are trained using the gradient boosting tree algorithm. Each weak learner is used to correct the error of the previous weak learner. These weak learners are combined to form a powerful model.

[0033] Given a dataset of n samples and m features, which is the historical investment dataset of an investment project. Predict investment return output using an additive model with K elements:

[0034]

[0035] Among them, f k This represents a decision tree, where K is the number of decision trees, and x... i y is the investment input, and y is the model prediction output;

[0036] The objective function is the loss function plus the model complexity, expressed as:

[0037]

[0038] Wherein, the part before the plus sign is the loss function, and the part after the plus sign is the model complexity function;

[0039] When training reaches the k-th tree, the first k-1 trees are already known. The objective function can then be rewritten as:

[0040]

[0041] Since the first k-1 terms are known, then:

[0042]

[0043] Among them, y i For the true value, f represents the predicted value of the first k-1 trees. k (x i ) represents the predicted value of the k-th tree, Ω(f) k Let ) represent the complexity of the k-th tree;

[0044] Therefore, we can find k trees that minimize the objective function.

[0045] For h types of investment benefits in the sample data, h training iterations are performed to obtain h black boxes. The input data for each training iteration is:

[0046]

[0047] Among them, S h This represents the set of investment returns in the h-th segment.

[0048] Preferably, the specific process of S4 is as follows:

[0049] In the XGBoost prediction model, for the j-th feature X in the i-th sample... ij The predictive model's predicted investment benefit for this sample is y. h The baseline for the predicted benefits of the entire model is y. base Then SHAPvalue follows the following equation:

[0050]

[0051] Among them, f h (X ij ) is X ij The SHAP value, i.e., the predicted investment benefit y of the j-th investment project in the i-th sample for investment h. h The contribution value; when f(X) ij If the value is greater than 0, it indicates that the project has a positive effect on the investment benefits; conversely, it indicates that it has a negative effect.

[0052] Quantitative analysis of the SHAP values ​​of each project:

[0053]

[0054] Where n is the number of samples, G(X) ij Let p be the investment amount of the j-th project in the i-th sample. hj This represents the average contribution per unit of investment projects of type j to the benefits of investment projects of type h.

[0055] Preferably, in S5, when conducting investment planning, the power grid company determines the investment needs for each investment benefit based on the regulations, policies, economic status, and development direction of different investment zones, and calculates the investment amount for each investment project using formula (8):

[0056]

[0057] Therefore, the present invention employs the above-mentioned method for optimizing distribution network investment strategies by combining machine learning, and the beneficial effects are as follows:

[0058] (1) This invention establishes an investment evaluation system and uses the TSO-XGBoost machine learning algorithm to train historical data to obtain a global optimal solution that satisfies historical preferences.

[0059] (2) This invention uses the SHAP algorithm to analyze the obtained "black box" and obtain a visualized percentage of each investment contribution, so as to provide clear guidance when the power grid company needs to make precise investments.

[0060] (3) Due to differences in development stage, geographical attributes, load characteristics and regional policies, the impact of investment strategies on returns varies from region to region. This invention proposes a regional division method to customize unique investment planning methods for different regions.

[0061] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0062] Figure 1 This is an overall flowchart of an embodiment of a visualization-based power distribution network investment strategy optimization method that combines machine learning according to the present invention;

[0063] Figure 2 This is a schematic diagram illustrating the construction of an investment strategy evaluation system according to an embodiment of a visual distribution network investment strategy optimization method combining machine learning, based on the present invention.

[0064] Figure 3 This is a schematic diagram illustrating the investment strategy evaluation system of an embodiment of a visual distribution network investment strategy optimization method combining machine learning according to the present invention.

[0065] Figure 4 This is a schematic diagram illustrating the principle and hyperparameter optimization process of the TSO-XGBoost algorithm, an embodiment of a visual power distribution network investment strategy optimization method combining machine learning, according to the present invention. Detailed Implementation

[0066] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0067] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0068] A "black box" model refers to a model based on statistical data and empirical summaries, typically built by analyzing and fitting existing data. Such models usually do not consider the internal physical mechanisms or processes of the system, but rather directly utilize the correlations between data for prediction or simulation. Black box models have poor interpretability and struggle to explain the underlying physical or logical mechanisms, thus potentially limiting their ability to understand system behavior and make predictions in certain situations.

[0069] This invention provides a visualization-based method for optimizing distribution network investment strategies, incorporating machine learning, to guide distribution network investment strategies. First, an investment strategy evaluation model is established to assess investment returns, taking into account the actual investment needs of the power grid company. Investment benefits such as improved distribution network reliability, increased power supply, and low-voltage benefits are used as the objective function of the investment strategy. Historical investment data of the power grid company is acquired, and development needs are predicted based on past experience, geographical attributes, load characteristics, and future regional plans, thus dividing the region into zones. The historical data of different zones is trained using the TSO-XGBoost machine learning algorithm to obtain a "black box" for calculating investment returns in each region. Simultaneously, the SHAP algorithm is used to visualize this "black box," clarifying the contribution ratio of various investments to different investment benefits. Finally, the optimal investment solution is derived from the objective function and applied to the investment planning scheme.

[0070] like Figure 1 As shown, a visualization-based power distribution network investment strategy optimization method combining machine learning includes the following steps:

[0071] S1. Establish a comprehensive investment strategy evaluation system, extracting key investment projects and investment returns from the power grid company's historical investment habits, development prospects, and investment needs, such as... Figure 2 shown. Figure 2 The leftmost column is for "Investment Projects": X1, X2, X3, ... X n The column on the right is "Investment Benefits": S1, S2, S3, ... S n This means that each investment project (X1 to X) n ) corresponds to an investment benefit (S1 to S) n This step provides a quantitative assessment of the optimization results and lays the foundation for subsequent analysis and modeling.

[0072] An indicator system for establishing an investment evaluation model is established based on the power grid's power supply capacity, power quality, power supply reliability, and economic benefits, such as... Figure 3 shown.

[0073] Power supply capacity includes capacity-to-load ratio, the proportion of expandable main transformer capacity, the proportion of line load rate, and the proportion of main transformer under heavy load.

[0074] Power quality includes voltage deviation indicators, voltage fluctuation indicators, and harmonic distortion indicators;

[0075] Power supply reliability includes average power outage frequency, average power outage time, availability factor of high-voltage equipment, and return on investment.

[0076] Economic benefits include increased power supply per unit of investment, reduced power loss per unit of investment, reduced power supply per unit of transformer capacity per unit of investment, and reduced operation and maintenance costs per unit of transformer capacity.

[0077] S2. Obtain historical investment data from the power grid company and perform data preprocessing to provide reliable feature data input for the prediction model. The specific data preprocessing steps are as follows:

[0078] Based on the power grid company's investment habits over the years, investment projects with the same or similar investment objectives are grouped into the same category to reduce the input of characteristic quantities and simplify the calculation process;

[0079] The investment projects are categorized into eleven types according to their objectives: "New substation outgoing lines to meet the power supply needs of new loads", "Replacing old or obsolete equipment or lines", "Solving safety hazards in medium and low voltage lines", "Improving the medium voltage grid", "Solving distribution transformer overload", "Solving distribution transformer overload", "Solving distribution transformer low voltage", "Solving medium voltage line overload", "Solving medium voltage line overload", "Distribution automation projects", and "New distribution areas to meet load demands".

[0080] S3. Divide the power supply into zones according to load density, and use the TSO-XGBoost algorithm to train the historical investment data of different power supply zones to generate a black box for investment planning.

[0081] Since machine learning requires a large amount of data for training, the problem is that if investment planning is carried out in a single region, the amount of data is not enough to train an accurate investment model. Therefore, in this embodiment, the concept of power supply zoning is combined to group regions with similar power supply conditions into one category to expand the amount of data, thereby obtaining a more accurate investment strategy model that is suitable for such regions.

[0082] Domestic power supply regions can be classified based on factors such as functional positioning, economic development level, electricity consumption level, load density, and user importance. Load density best reflects a region's electricity consumption level and development needs. Therefore, this embodiment classifies different regions into six categories: A+, A, B, C, D, and E, according to their load density differences. The specific classification method is shown in Table 1, and the investment focus varies across different zones.

[0083] Table 1. Classification Map of Domestic Power Supply Regions

[0084]

[0085] In Table 1, σ represents the load density of the power supply area type (unit: MW / km). 2 The power supply area is generally not less than 5 km². 2 When calculating the load density, the load of 110(66)kV dedicated lines, as well as areas without effective power supply such as high mountains, Gobi Desert, desert, water areas, and forests, should be deducted.

[0086] In this embodiment, XGBoost machine learning and the Tuna Swarm Optimization Algorithm (TSO) are used to optimize the hyperparameters of the model. The basic principle is as follows: Figure 4 As shown, we first analyze the historical investment data of different power supply zones, then assign the optimal hyperparameter combination obtained by the Tuna Swarm Optimization Algorithm (TSO) to the XGBoost prediction model, perform error analysis on the prediction results, and use SHAP to perform interpretability analysis on the prediction model.

[0087] 1. The Tuna Swarm Optimization Algorithm (TSO) is primarily used to simulate two foraging behaviors in tuna schools: spiral foraging and parabolic foraging. TSO is characterized by its strong optimization ability, fast convergence speed, and low parameter tuning requirements. Therefore, we propose to use TSO to optimize the hyperparameters of the XGBoost model. TSO is a swarm intelligence-based optimization algorithm.

[0088] In the Tuna Swarm Optimization Algorithm (TSO), each individual tuna represents a potential solution, and the entire swarm works together to find the optimal solution. The TSO tuna swarm optimization algorithm consists of three phases: the search phase, the chase phase, and the attack phase.

[0089] Search phase: Individual tuna conduct random searches based on their own sensory abilities and information from the group, exploring possible solution spaces;

[0090] The pursuit phase: When an individual discovers a better solution, other individuals follow the best individual and pursue it, gradually approaching the optimal solution;

[0091] Attack phase: When an individual discovers the optimal solution, the group will concentrate its attack and eventually obtain the optimal solution;

[0092] 2. Gradient boosting tree is an ensemble learning method that uses the gradient boosting tree algorithm to iteratively train multiple weak learners (decision trees). Each weak learner is used to correct the error of the previous weak learner. Finally, these weak learners are combined to form a powerful model.

[0093] Given a dataset of n samples and m features, which is the historical investment dataset of an investment project. Predict investment return output using an additive model with K elements:

[0094]

[0095] Among them, f k This represents a decision tree, where K is the number of decision trees, and x... i y is the investment input, and y is the model prediction output;

[0096] 3. XGBoost controls model complexity and prevents overfitting by adding regularization terms. It uses L1 and L2 regularization, and can also perform subsampling and column sampling to improve the model's generalization ability.

[0097] 4. XGBoost supports various loss functions, such as squared loss for regression problems and binary logistic loss for binary classification problems. By optimizing the loss function, XGBoost can better fit the data.

[0098] 5. The objective function is the loss function plus the model complexity, expressed as:

[0099]

[0100] Wherein, the part before the plus sign is the loss function, and the part after the plus sign is the model complexity function;

[0101] When training reaches the k-th tree, the first k-1 trees are already known. Therefore, the objective function (minimization) can be rewritten as:

[0102]

[0103] Since the first k-1 terms are known, then:

[0104]

[0105] Among them, y i For the true value, f represents the predicted value of the first k-1 trees. k (x i ) represents the predicted value of the k-th tree, Ω(f) k Let be the complexity of the k-th tree.

[0106] Therefore, we can find k trees that minimize the objective function.

[0107] For h types of investment benefits in the sample data, h training iterations are performed to obtain h black boxes. The input data for each training iteration is:

[0108]

[0109] Among them, S h This represents the set of investment returns in the h-th segment.

[0110] S4. Analyze the generated black box using the SHAP algorithm to obtain a visualized percentage of investment contribution.

[0111] In this embodiment, to obtain an interpretable optimized investment strategy model, SHAP (SHapley Additive exPlanations) is used to analyze the generated black box, obtaining the accurate contribution of different investment projects to the benefits of each investment, thus guiding the power grid company's precise investment needs. The specific process is as follows:

[0112] In the XGBoost prediction model, for the j-th feature (i.e., the investment project) in the i-th sample, X... ij The predictive model's predicted investment benefit for this sample is y. h The baseline for the predicted benefits of the entire model is y. base Then SHAPvalue follows the following equation:

[0113]

[0114] Among them, f h (X ij ) is X ij The SHAP value, i.e., the predicted investment benefit y of the j-th investment project in the i-th sample for investment h. h The contribution value; when f(X) ij If the value is greater than 0, it indicates that the project has a positive effect on the investment benefits; conversely, it indicates that it has a negative effect.

[0115] The obtained SHAP value reflects the impact of the investment project on investment benefits. To obtain a more accurate contribution, a quantitative analysis of the SHAP value of each project is conducted:

[0116]

[0117] Where n is the number of samples, G(X) ij Let p be the investment amount of the j-th project in the i-th sample. hj This represents the average contribution per unit of investment projects of type j to the benefits of investment projects of type h.

[0118] S5. Based on the obtained weight ratios and the investment model, the optimal investment strategy that meets the investment requirements is derived.

[0119] When conducting investment planning, the power grid company can determine the investment needs for each investment benefit based on factors such as the laws and policies, economic status, and development direction of different investment zones. Among these factors, the investment amount for each investment project can be calculated using formula (8).

[0120]

[0121] The method presented in this embodiment can provide investors with precise and explainable investment solutions, thereby achieving the purpose of investment planning.

[0122] Therefore, this invention employs the aforementioned visualization-based power distribution network investment strategy optimization method that combines machine learning. By combining SHAP with XGBoost, the prediction results of the XGBoost model in the investment benefit prediction task can be explained. SHAP can tell us how much each feature contributes to the prediction result of a particular sample, thereby helping investors and researchers understand the model's prediction process and providing a more accurate investment planning guidance.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual power distribution network investment strategy optimization method combining machine learning, characterized in that, The following steps are involved: S1. Establish a comprehensive investment strategy evaluation system; S2. Obtain historical investment data from the power grid company and perform data preprocessing; S3. Divide the power supply zones of each region according to load density, and use the TSO-XGBoost algorithm to train the historical investment data of different power supply zones to generate a black box for investment planning. S4. Analyze the generated black box using the SHAP algorithm to obtain a visualized percentage of investment contribution. S5. Based on the obtained weight ratios and the investment model, the optimal investment strategy that meets the investment requirements is derived. The specific process of S3 is as follows: The model was optimized using XGBoost machine learning and the Tuna Swarm Optimization Algorithm (TSO). TSO simulates two foraging behaviors of tuna swarms: spiral foraging and parabolic foraging. In TSO, each tuna represents a potential solution, and the entire swarm works together to find the optimal solution. The TSO tuna swarm optimization algorithm consists of three phases: search phase, chase phase, and attack phase. Search phase: Individual tuna conduct random searches based on their own sensory abilities and information from the group, exploring possible solution spaces; The pursuit phase: When an individual discovers a better solution, other individuals follow the best individual and pursue it, gradually approaching the optimal solution; Attack phase: When an individual discovers the optimal solution, the group will concentrate its attack and eventually obtain the optimal solution; Multiple weak learners are trained using the gradient boosting tree algorithm. Each weak learner is used to correct the error of the previous weak learner. These weak learners are combined to form a powerful model. Given that One sample and One feature is the historical investment dataset of the investment project. Using an additive model with K elements to predict investment return output: （1）； in, Representing a decision tree, For the number of decision trees, For investment input, The model predicts the output; The objective function is the loss function plus the model complexity, expressed as: （2）； Wherein, the part before the plus sign is the loss function, and the part after the plus sign is the model complexity function; When training reaches the... When planting trees, the front If all trees are known, then the objective function can be rewritten as: （3）； Because of the previous Given the terms, then: （4）； in, For the true value, For the front The predicted value of the trees, For the first The predicted value of the trees, For the first The complexity of a tree; From this, we found Find a tree that minimizes the objective function; For the sample data Investing in various ways to achieve investment benefits After this training, I received Each training iteration uses a black box, and the input data is: （5）； in, Representing the The collection of investment returns.

2. The method for optimizing a visual distribution network investment strategy by combining machine learning as described in claim 1, characterized in that, In S1, key investment projects and investment returns are extracted from the power grid company's investment habits, development prospects and investment needs over the years. An indicator system for the investment evaluation model is established based on the power grid's power supply capacity, power quality, power supply reliability and economic benefits.

3. The method for optimizing a visual distribution network investment strategy by combining machine learning as described in claim 2, characterized in that, In S2, the data preprocessing specifically involves: Based on the power grid company's investment practices over the years, investment projects with the same or similar investment objectives are grouped into the same category; Investment projects are categorized into eleven types based on their objectives: new substation outgoing lines to meet new load demand, replacement of old equipment or lines, resolution of safety hazards in medium and low voltage lines, improvement of medium voltage grid, resolution of distribution transformer overload, resolution of distribution transformer overload, resolution of low voltage in distribution areas, resolution of medium voltage line overload, resolution of medium voltage line overload, distribution automation projects, and new distribution areas to meet load demand.

4. The method for optimizing a visual distribution network investment strategy by combining machine learning as described in claim 3, characterized in that, In S3, regions with similar power supply conditions are grouped into one category. Based on the differences in load density in each region, different regions are divided into six categories: A+, A, B, C, D, and E according to load density.

5. The method for optimizing a visual distribution network investment strategy by combining machine learning according to claim 4, characterized in that, The specific process of S4 is as follows: In the XGBoost prediction model, for the th The first sample Features The prediction model predicts the investment benefits for this sample as follows: The baseline for the predicted benefits of the entire model is Then SHAPvalue follows the following equation: （6）； in, for The SHAP value, i.e. the first In the nth sample Each investment project Forecast of investment benefits The contribution value; when This indicates that the project has a positive effect on the investment benefits; conversely, it indicates that it has a negative effect. Quantitative analysis of the SHAP values ​​of each project: （7）； in, For the number of samples, For the first The first sample Investment amount of each project For the first Type of investment project The average contribution per unit of investment benefits.

6. The method for optimizing a visual distribution network investment strategy by combining machine learning as described in claim 5, characterized in that, In S5, when conducting investment planning, the power grid company determines the investment needs for each investment benefit based on the regulations, policies, economic status, and development direction of different investment zones, and calculates the investment amount for each investment project using formula (8): （8）。

Images