Power grid technical improvement evaluation and optimization method based on deep learning

Abstract

The invention discloses a power grid technical improvement evaluation and optimization method based on deep learning, and relates to the technical field of power grid technical improvement, and the method comprises the following steps: collecting the historical data of a power grid technical improvement project, carrying out the feature extraction and standardization processing of a candidate project, and obtaining a standardized feature vector; inputting the standardized feature vector into a deep learning model, predicting an effect index and an implementation risk of each candidate item, and structuring to obtain a prediction result; calculating a precondition according to a prediction result, and sorting and screening the candidate items by adopting a generative intelligent genetic algorithm to obtain an optimal scheme set; and screening an optimal investment configuration scheme set based on the optimal scheme set and the candidate project sorting result. The method is used for solving the problems that the benefit of a technical improvement scheme is difficult to quantify and an optimization mechanism lacks objectivity.

Description

Technical Field

[0001] This invention relates to the field of power grid upgrading technology, and more specifically, to a deep learning-based method for power grid upgrading evaluation and optimization. Background Technology

[0002] The quantitative correlation between input and output in technological upgrading projects is insufficient, making it difficult to scientifically assess the actual benefits of these projects. There is a lack of a mapping system between problem-solving and key performance indicators for individual projects, hindering the precise measurement of investment benefits. Furthermore, the lack of input-output linkage analysis means that investment benefit evaluation results are not sufficiently linked to subsequent project investment decisions and regional-level investment allocation optimization, failing to achieve a closed-loop evaluation mechanism. The static nature of investment allocation decisions prevents adjustments and optimizations based on actual circumstances. Some regions, due to their large historical investment scale, maintain a high proportion of investment allocation, but their responsiveness to the actual needs of other regions needs improvement. Investment allocation decisions do not fully consider input-output evaluation results, often focusing too much on the scale of capital investment while neglecting the output benefits of different regions. This exhibits a "heavy on input, light on benefits" phenomenon, resulting in the underutilization of investment return potential in some high-efficiency regions and affecting the scientific and rational nature of overall investment decisions. Current selection mechanisms rely heavily on subjective judgment, lacking unified selection standards and a systematic evaluation system, which interferes with the scientific and rational nature of decision-making. Different projects have varying implementation conditions, benefit outputs, time cycles, and risk levels.

[0003] To improve the scientific nature of the selection and decision-making process for technological upgrading projects, it is urgent to build a more systematic and quantitative selection mechanism. This mechanism should be based on project characteristics and input-output benefits, and combined with dynamic adjustments and quantitative evaluation tools to enhance the scientific nature of the selection and decision-making process and ensure the maximum investment effectiveness of technological upgrading projects.

[0004] The above-disclosed technical solutions have at least the following technical problems: Existing methods are insufficient in characterizing the relationship between investment and effectiveness in technological upgrading projects, making it difficult to scientifically assess the actual benefits of individual projects. The input-output characteristics of different projects have not formed a unified quantitative evaluation system; current project selection mechanisms rely heavily on expert experience and subjective judgment, lacking a systematic mapping relationship, resulting in a lack of objective basis for benefit assessment and prioritization of different types of projects; currently, there is a lack of deep learning-based predictive models to quantify the effectiveness and risks of candidate projects, and a lack of optimization algorithms that combine predictive results, making it difficult to achieve optimal investment allocation under multiple constraints such as budget, category, region, and risk, thus failing to maximize overall benefits.

[0005] To address the above problems, this invention proposes a solution. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a deep learning-based method for evaluating and optimizing power grid technological upgrades. By constructing a quantitative mapping relationship and combining deep learning prediction and intelligent optimization algorithms, the method achieves a scientific evaluation of the effectiveness and risks of candidate projects and a dynamic optimization of investment allocation, thereby solving the problems of difficulty in quantifying the benefits of technological upgrade schemes and the lack of objectivity in the selection mechanism.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for evaluating and optimizing power grid technological upgrades includes the following steps: In a preferred embodiment, the historical data of the power grid upgrading project includes basic project attribute data, project problem and demand data, project implementation content data, project results and acceptance data, project risk and deviation data, and project related assets and operation data.

[0008] In a preferred embodiment, the feature extraction specifically involves: extracting project attribute features, problem type features, and performance indicator features from historical data of power grid technical transformation projects; establishing a global attribute feature dictionary, a global problem dictionary, and a global indicator dictionary based on the project attribute features, problem type features, and performance indicator features; and establishing attribute feature vectors, problem feature vectors, and indicator feature vectors based on the global attribute feature dictionary, global problem dictionary, and global indicator dictionary.

[0009] In a preferred embodiment, the standardized feature vector is obtained by concatenating all standardized feature vectors according to the problem point-attribute feature-performance indicator.

[0010] In a preferred embodiment, the step of inputting standardized feature vectors into a deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtaining the prediction results after structuring, specifically involves: configuring the hierarchy and parameters of the deep learning model; training the deep learning model based on historical data of power grid technical transformation projects; constructing a test set to evaluate the model accuracy and selecting qualified deep learning models; and using the qualified deep learning models to output and structure the prediction results for the candidate projects that need to be evaluated.

[0011] In a preferred embodiment, the generative intelligent genetic algorithm specifically comprises: converting candidate items into chromosomes and intelligently initializing the population; evaluating and ranking the candidate items through a fitness function and preconditions; and obtaining a set of preferred schemes based on the initialized population and the evaluation results.

[0012] In a preferred embodiment, the genetics specifically involves: using an adaptive tournament selection method to screen for high-quality parent chromosomes; using single-point crossover to combine the high-quality parent chromosomes to generate offspring chromosomes; and performing random mutations on the offspring chromosomes to obtain a set of preferred schemes.

[0013] In a preferred embodiment, the adaptive tournament selection method filters parent chromosomes according to the tournament size defined by the iteration stage, and retains parent chromosomes with high fitness to directly enter the offspring population.

[0014] In a preferred embodiment, the step of selecting the optimal investment allocation scheme set based on the preferred scheme set and the ranking results of candidate projects specifically involves: ranking the evaluation results of candidate projects in the preferred scheme set, outputting the top-ranked project schemes, and the set of project schemes is the optimal investment allocation scheme set.

[0015] An apparatus for evaluating and optimizing power grid technological upgrading based on deep learning includes: a collection and processing module for collecting historical data of power grid technological upgrading projects, extracting and standardizing features from candidate projects to obtain standardized feature vectors; an input-output module for inputting the standardized feature vectors into a deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtaining the prediction results in a structured manner; a calculation and screening module for calculating preconditions based on the prediction results, using a generative intelligent genetic algorithm to rank and screen candidate projects to obtain a set of preferred solutions; and a final output module for screening the set of optimal investment allocation solutions based on the set of preferred solutions and the ranking results of candidate projects.

[0016] The technical effects and advantages of the deep learning-based power grid upgrading evaluation and optimization method of this invention are as follows: This invention constructs a unified feature vector for model calculation by extracting and standardizing features from candidate projects, ensuring comparability and data consistency among different projects. It utilizes deep learning models to predict project performance indicators and implementation risks, improving the objectivity and accuracy of the evaluation. Finally, it combines generative intelligent genetic algorithms for multi-constraint optimization, outputting a set of investment allocation schemes. This achieves intelligent evaluation of power grid upgrading projects and dynamic optimization of investment allocation, effectively solving the problems of difficulty in quantifying the benefits of upgrading schemes and the lack of objectivity in the selection mechanism. Attached Figure Description

[0017] Figure 1 A schematic diagram of the power grid technical transformation evaluation and optimization method based on deep learning provided in an embodiment of the present invention.

[0018] Figure 2 A schematic diagram of the device structure for the deep learning-based power grid technical transformation evaluation and optimization method provided in an embodiment of the present invention. Detailed Implementation

[0019] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 This invention presents a deep learning-based method for evaluating and optimizing power grid technological upgrades, comprising the following steps: S1. Collect historical data of power grid technical transformation projects, extract features and standardize candidate projects to obtain standardized feature vectors; S2, input the standardized feature vector into the deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtain the prediction results in a structured manner; S3. Calculate the preconditions based on the prediction results, and use a generative intelligent genetic algorithm to sort and screen the candidate projects to obtain the set of preferred solutions; S4. Based on the preferred solution set and the ranking results of candidate projects, select the optimal investment allocation solution set.

[0021] This invention constructs a unified feature vector for model calculation by extracting and standardizing features from candidate projects, ensuring comparability and data consistency among different projects. It utilizes deep learning models to predict project performance indicators and implementation risks, improving the objectivity and accuracy of the evaluation. Finally, it combines generative intelligent genetic algorithms for multi-constraint optimization, outputting a set of investment allocation schemes. This achieves intelligent evaluation of power grid upgrading projects and dynamic optimization of investment allocation, effectively solving the problems of difficulty in quantifying the benefits of upgrading schemes and the lack of objectivity in the selection mechanism.

[0022] S1. Collect historical data of power grid technical renovation projects, perform keyword clustering analysis on project information, collect historical data of power grid technical renovation projects, extract features and standardize candidate projects to obtain standardized feature vectors.

[0023] S11, In this embodiment, the historical data of the power grid upgrading project specifically includes basic project attribute data, project problem and demand data, project implementation content data, project results and acceptance data, project risk and deviation data, and project related asset and operation data, specifically: The basic attribute data of the project includes: Project identification information: Project ID (project number), project name, project approval number, implementing unit, city / region.

[0024] Specialty and type information: Major, voltage level, project level, and technical renovation type.

[0025] Time and scale information: Year of project initiation, planned implementation period, actual implementation period, planned investment amount, actual investment amount, and source of funding.

[0026] The project issues and requirements data include: Core problem description: Problem type, specific text of the problem, scope of the problem's impact, and urgency of the problem.

[0027] Data supporting project approval: policy basis for project approval, equipment status basis, and source of demand.

[0028] The project implementation data includes: Target and scope of renovation: Name of equipment to be renovated, equipment model and specifications, and quantity to be renovated.

[0029] Technical solution data: core modification process, type of technology adopted, and supporting measures.

[0030] The project results and acceptance data include: Performance indicators can be categorized into three types: safety indicators, efficiency indicators, and economic indicators. For example, safety indicators include: N-1 pass rate, fault tripping rate, and average power outage time per user; efficiency indicators include: power supply reliability rate, voltage qualification rate, and overall line loss rate.

[0031] Acceptance results data: acceptance conclusion, acceptance time, acceptance report summary, and explanation of items that did not meet the standards.

[0032] The project risk and deviation data include: Investment deviation data: amount of investment deviation, investment deviation rate, and reasons for deviation.

[0033] Project schedule deviation data: number of days of project schedule deviation, project schedule deviation rate, and reasons for delays.

[0034] Risk record data: risk events, risk losses, risk response measures and their effects during implementation.

[0035] The project's associated assets and operational data include: Basic asset data: service life of equipment before renovation, original value of assets, net value of assets, identification of over-aged assets, and asset condition rate.

[0036] Operational status data: average annual failure rate of equipment before the upgrade, average number of maintenance visits per year, equipment load rate, and historical fault records.

[0037] Regional correlation data: electricity load growth rate, regional GDP growth rate, and frequency of natural disasters in the region where the project is located.

[0038] S12, the feature extraction specifically includes: S121, by extracting project attribute characteristics, problem type characteristics, and effectiveness indicator characteristics from historical data of power grid technical renovation projects, specifically: Project attribute characteristics are derived directly from basic project attribute data, such as voltage level, investment amount, and construction period; problem type characteristics are derived from project problem and requirement data, such as tower corrosion and severe icing; and performance indicator characteristics are derived from project performance and acceptance data, such as a 10% increase in the N-1 pass rate.

[0039] S122, based on project attribute characteristics, problem type characteristics, and performance indicator characteristics, establish a global attribute characteristic dictionary, a global problem dictionary, and a global indicator dictionary, specifically as follows: Establish a global attribute feature dictionary , where n is the number of all project attribute features, where ~ Define project attributes and characteristics; establish a global issue dictionary. , among them ~ Each of the z problem type features is represented; a global indicator dictionary is established. , among them ~ Each of the d performance indicators represents a characteristic of a performance indicator.

[0040] S123, based on the global attribute feature dictionary, global problem dictionary, and global indicator dictionary, establish attribute feature vectors, problem feature vectors, and indicator feature vectors, specifically as follows: The attribute feature vector is Each vector represents a value corresponding to a project attribute, with a one-to-one correspondence to w. For example, the global attribute feature dictionary w = [voltage level, investment amount, construction period], and the attribute feature vector m = [220, 100000, 24]. Each problem encountered in the project is recorded as 1, otherwise 0, resulting in the problem feature vector. , ~ Corresponding to the global problem dictionary ~ For example: global problem dictionary If the problem with project A is "severe icing", then the feature vector corresponding to project A is: The expected improvement effect of each project is directly converted into a numerical value to obtain the indicator feature vector. Indicator Feature Vector Each vector value corresponds to an indicator in the global indicator dictionary. For example, if the global indicator dictionary is [line tripping rate, N-1 pass rate, ...], and the indicator for project A is to increase the N-1 pass rate by 10%, then the corresponding indicator feature vector is [0, 0.10, ...].

[0041] S13, the standardized feature vector obtained after the standardization process is as follows: Choose a suitable standardization method based on the data range of different characteristics, for example: Since the investment amount has a large range, min-max normalization is used, and the calculation formula is as follows:

[0042] In the formula, This is the normalized investment amount. This represents the current investment amount for the project. This represents the minimum investment amount across all projects. This represents the maximum value within the range of investment amounts for all projects.

[0043] Construction periods are typically measured in months, and the range of values ​​is relatively small. Z-score standardization is used, and the calculation formula is as follows:

[0044] In the formula, For the standardized construction period, For the current project schedule, This is the average duration of all projects. The standard deviation is calculated from the duration of the current project and the average duration of all projects.

[0045] The range of effectiveness improvement rate values ​​is very small, so they can be directly converted into decimals. If the numerical distribution of similar indicators differs significantly, Z-score standardization can be used for further processing.

[0046] Categorical variables such as voltage levels are discrete. To avoid misinterpreting voltage values ​​directly as numerical values, one-hot encoding can be used. For example, voltage levels of 110kV, 220kV, and 500kV can be converted into three-dimensional vectors. , , .

[0047] It should be noted that, in the categorical variables, the length of each vector is not necessarily 1, but the length is fixed. For example, the voltage level corresponds to the following attribute feature vector: Then the standardized It is fixed as a three-dimensional vector with a fixed length of 3 bits.

[0048] S14, the standardized feature vector is obtained by concatenating all standardized feature vectors according to the problem point-attribute feature-performance indicator, specifically: The global problem dictionary is The global attribute feature dictionary is The global indicator dictionary is [trip rate, N-1 pass rate].

[0049] Example project A, the problem is: "Severe icing" -> The attribute characteristics are: investment amount (assuming it is 0.5 after standardization): 500 million -> Construction period (assuming standardization is 1.0): 32 months -> Voltage level: 220kV The performance indicator is: tripping rate decrease of 5% -> N-1 improves by 10% -> ; The standardized feature vector is .

[0050] S2, input the standardized feature vector into the deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtain the prediction results in a structured manner; In this embodiment, the deep learning model uses a fully connected neural network as its core model.

[0051] S21, configure the layers and parameters of the deep learning model, specifically as follows: The model hierarchy and parameter configuration, taking the 9-dimensional standardized feature vector of example project A in S14 as an example, are designed according to a fixed process of "input layer -> hidden layer -> output layer", and the parameters of each layer can be reused: Input layer: The number of neurons is the same as the dimension of the S2 normalized feature vector. The example dimension is 9. If the feature dimension changes in the actual project, the number will be adjusted accordingly. Its function is to directly receive the feature vector without additional processing.

[0052] Hidden layer 1: Set 64 neurons, use ReLU activation function, retain only positive information in feature vectors, and extract complex relationships between features, such as the potential relationship between "investment amount" and "operation and maintenance cost reduction rate".

[0053] Hidden layer 2: Set 32 ​​neurons, also using the ReLU activation function to further compress redundant features, optimize model learning efficiency, and balance prediction accuracy and training speed.

[0054] Output layer: Eight neurons are set up, corresponding to eight prediction indicators: five performance indicators and three risk indicators, as shown in Table 2. If *a* key indicators need to be predicted, then *a* output layer neurons are set up. The activation function is selected according to the indicator type. Performance indicators, such as "change in line tripping rate" and "change in N-1 pass rate", can be directly output as continuous values ​​(in percentage or decimal form that match the indicator) using "linear activation". Risk probability, such as "probability of technical failure": use "sigmoid activation" to compress the output to between 0 and 1.

[0055] Table 2

[0056] S22, based on historical data from power grid upgrade projects, trains a deep learning model, specifically as follows: Input data processing refers to processing the historical data of each historical power grid technical transformation project in accordance with the feature extraction and standardization process of S1. Specifically, this involves: extracting project attribute features, problem type features, and performance indicator features; standardizing different types of features separately; and concatenating all standardized features to form the input feature vector of the historical project.

[0057] Output data processing refers to extracting the true index values ​​from the input feature vector and normalizing them, specifically: Extracting the true value: Extract the actual results of the historical project from the project performance and acceptance data, project risk and deviation data of S1. For example, if the actual line tripping rate of a certain historical project decreased by 5.2% and the investment deviation rate was 7.8%, it can be used as the learning target of the model, that is, the output true value. Output metric normalization: Due to the large differences in magnitude between different metrics, such as the maintenance cost reduction rate of 4.9% and the technical failure probability of 0.02, all output metrics need to be processed using min-max normalization to compress the metric values ​​to between 0 and 1, avoiding the model's bias towards learning large numerical values. During the operation, the minimum value (y_min) and maximum value (y_max) of each metric need to be recorded to form a normalization parameter table. In subsequent predictions, this table needs to be used to restore the normalized values ​​of the model output to the true metric values.

[0058] The processed input-output training pairs are divided in a 7:2:1 ratio to ensure that the model can fully learn and validate its generalization ability. Training set: 7%, used for the model to learn the correlation between feature vectors and true indicator values; Validation set: accounting for 2%, used to adjust model parameters and prevent the model from only learning the training data and not understanding new data; Test set: accounting for 1%, used to finally evaluate the prediction accuracy of the model and determine whether the model is usable.

[0059] The model training process uses the deep learning framework PyTorch to build and train the model, as shown in Table 3: Table 3

[0060] The training operation procedure is as follows: Loading data: Import the divided training and validation sets into the deep learning framework PyTorch, and the framework will automatically load the data in batches; Initialize the model: Configure the model in the framework according to the S211 architecture: Input layer -> Hidden layer 1 -> Hidden layer 2 -> Output layer; Start training: In each round of training, the framework updates the model parameters with the training set data and calculates the training error. In each round of validation, the model error is evaluated with the validation set data. If the validation error does not decrease for three consecutive rounds, the early stopping mechanism is triggered, and training is stopped. During training, the deep learning model with the smallest validation error is saved.

[0061] S23, Construct a test set, evaluate the model accuracy, and select qualified deep learning models, specifically: The test set is the collection of historical projects used to test the model.

[0062] Load the test set data and the optimal model saved in S22; use the model to predict the indicator value of each item in the test set, and evaluate it using the mean absolute error (MAE). The mean absolute error is the average of the absolute deviations between the predicted value and the actual indicator value. The smaller the mean absolute error, the more accurate the model prediction. The MAE qualification standard is fixed; meeting it is considered qualified. The initial MAE qualification standard is 0.5%, which is adjusted according to the actual accuracy requirements. For example, if the predicted indicator type is a performance indicator, the MAE qualification standard is 0.5%, the actual tripping rate is 5.2%, the predicted value is 5.0%, and the MAE is 0.2%, which is less than the MAE qualification standard, so it is qualified. If the MAE of all metrics meets the qualification criteria, the model is usable. If not, the dataset needs to be re-examined and retrained to obtain a qualified deep learning model.

[0063] S24, using a qualified deep learning model, outputs and structures the prediction results for the candidate items to be evaluated: Following the method in S14, the basic information of candidate projects is converted into standardized feature vectors; the qualified deep learning model from S23 is imported, and the standardized feature vectors of the candidate projects are input into the model to obtain the normalized predicted values ​​output by the model; inverse normalization is then performed to obtain the true values: using the normalization parameter table, the normalized predicted values ​​are restored to the true indicator values. For example, the normalized predicted value of the tripping decrease rate of candidate project A is 0.48. If the indicator y_min=0% and y_max=10%, This means the line tripping rate decreased by 5.2%.

[0064] The prediction results for each candidate item are structured into an n-dimensional vector, which is the final prediction result. Here, n represents one item ID plus n-1 indicator values. For example, the structured prediction result for candidate item B is... Vector index rules: 0=Project ID, 1=Line tripping reduction rate, 2=N-1 pass rate, 3=Power supply reliability improvement rate, 4=Operation and maintenance cost reduction rate, 5=Intelligent inspection coverage improvement rate, 6=Investment deviation rate, 7=Construction period deviation rate, 8=Technical failure probability.

[0065] S3. Calculate the prerequisites based on the prediction results, and use a generative intelligent genetic algorithm to sort and screen candidate projects to obtain a set of preferred solutions.

[0066] In this embodiment, the step of using a generative intelligent genetic algorithm to solve the problem and output the optimal investment allocation scheme set is as follows: S31, calculate the preconditions based on the prediction results, specifically: Based on the prediction results, the total benefit score and the total risk score are calculated. The total benefit score is calculated as follows: The total benefit score is calculated as follows: Total safety benefit score × 0.4 + Total economic benefit score × 0.2 + Total efficiency benefit score × 0.1. The weights can be dynamically adjusted according to the actual situation.

[0067] For example, in Project C, if the predicted N-1 pass rate increases by 9.5%, the line tripping rate decreases by 5.2%, the maintenance cost decreases by 5.1%, and the intelligent inspection coverage increases by 6.3%, then the safety benefit score = 9.5% + 5.2% = 14.7%, the economic benefit score = 5.1%, the efficiency benefit score = 6.3%, and the total benefit score = 14.7% × 0.4 + 5.1% × 0.2 + 6.3% × 0.1 = 5.88% + 1.02% + 0.63% = 0.0753.

[0068] The total risk score is calculated as follows: Total Risk Score = Investment Risk Score × 0.6 + Technology Risk Score × 0.4 For example, in Project C, the predicted investment deviation rate is 8.0% and the probability of technical failure is 2.1%. Therefore, the investment risk score is 0.080, the technical risk score is 0.021, and the total risk score is 0.080×0.6+0.021×0.4=0.048+0.0084=0.048084.

[0069] The hard constraints include total investment budget constraints, project category constraints, regional balance constraints, and risk ceiling constraints, specifically: The total investment budget constraint is that the sum of the actual investment amounts of all selected projects ≤ the preset total budget B. The actual investment is obtained by inversely normalizing the standardized investment amount of S2: Actual investment = Standardized investment amount × (Historical maximum investment - Historical minimum investment) + Historical minimum investment. For example, if the standardized investment amount of a project in S2 is 0.5, and the historical investment max = 10 million yuan and min = 1 million yuan, then the actual investment = 0.5 × (1000 - 100) + 100 = 5.5 million yuan; if the total budget B = 50 million yuan, the sum of the actual investments of the selected projects must ≤ 50 million yuan.

[0070] The project category constraints are as follows: the number of projects for improving safety and stability must be ≥2, and the number of projects for improving intelligence level must be ≥1. This constraint can be adjusted according to the actual project situation.

[0071] The regional balance constraint requires that each city / region has at least one selected project to avoid concentrating resources in a single region. The region to which the project belongs is based on the project attribute characteristics of S2 - implementation region, such as the eastern district of a city.

[0072] The risk ceiling constraint is that the total technical failure probability of all selected projects is ≤5%. This threshold can be adjusted according to the actual situation to control the overall technical risk. The total technical failure probability is the sum of the predicted technical failure probabilities of each selected project. The projects are implemented independently, and the probabilities are superimposed.

[0073] S32, the generative intelligent genetic algorithm is specifically as follows: S321 converts candidate items into chromosomes and intelligently initializes the population, specifically as follows: Using binary encoding, one chromosome corresponds to one project configuration scheme; Chromosome length is the total number of candidate items. For example, if there are 10 candidate items, the chromosome length is 10. In a chromosome, each gene position has a value of 0 or 1: 1 indicates that the item is selected, and 0 indicates that the item is not selected; Example: Candidate items 1 (safety), 2 (intelligentization), 3 (economic) ... 10 (disaster resistance), chromosome [1,1,0,0,1,0,0,0,0,0] indicates that items 1, 2, and 5 are selected, and the rest are not selected.

[0074] The population is intelligently initialized by generating an initial population in batches and with constraint verification. For example, the population size is set to 50, which means 50 feasible schemes initially.

[0075] The first batch consists of 20 plans: priority will be given to meeting the safety and intelligent constraints. Each plan must include at least 2 safety-related projects and 1 intelligent project. After calculating the actual investment, if the budget is exceeded, the project with the largest investment and the lowest total benefit score will be removed until the budget is met. If the investment does not reach the lower limit of the budget, such as a total budget of 50 million and the current plan is only 20 million, then the economic category project with the smallest investment and the highest total benefit score will be added.

[0076] The second batch consists of 20 plans: priority is given to meeting regional balance constraints, with each plan covering all cities and prefectures. Budget and category constraints are also verified, and projects are adjusted if they are not met.

[0077] The third batch consists of 10 plans: projects are randomly selected, but after generation, all constraints must be verified for each plan. If a constraint is violated, the plan will be adjusted by removing high-investment, low-efficiency projects and adding low-cost, compliant projects to ensure that each plan is feasible.

[0078] Assume that after intelligent initialization, the initial population is 30.

[0079] S322 evaluates and ranks candidate projects based on a fitness function and preconditions, specifically as follows: The fitness function is the core basis for the algorithm to select a scheme. The formula is: Fitness = Total Benefit Score × 0.7 - Total Risk Score × 0.3 In the formula, 0.7 is the benefit weight and 0.3 is the risk weight, which can be adjusted according to the actual situation; If a proposal violates any constraint, such as exceeding the budget or insufficient safety-related items, its fitness will be set to 0 and the proposal will be eliminated. A higher fitness value indicates a better solution. Example: A solution has a total benefit score of 24.49%, a total risk score of 16.01%, and satisfies all constraints. If the plan exceeds the budget, the fitness is 0.

[0080] S33, based on the initial population and evaluation results, a set of optimal schemes is obtained through genetics, specifically: The genetic operation includes three steps: selection, crossover, and mutation. Each step incorporates constraint verification and intelligent adjustment to avoid invalid solutions.

[0081] S331 uses an adaptive tournament selection method to screen for high-quality parent chromosomes, specifically: S3311, the adaptive tournament selection method filters parent chromosomes according to the tournament size defined in the iteration stage, and retains parent chromosomes with high fitness to directly enter the offspring population, specifically as follows: In the first 30% of iterations, the tournament size is 3 to retain more diversity and avoid premature convergence; in the last 70% of iterations, the tournament size is 7 to increase selection pressure and make it easier for high-fit individuals to be selected as parents.

[0082] Individuals in the top 10% of fitness in each generation do not participate in crossover mutation and directly enter the next generation population. If the best individual in a certain generation improves by less than 0.1% compared to the previous generation, the elite retention ratio is forcibly increased to 20% to accelerate the accumulation of high-quality genes. The aforementioned 30%, 70%, tournament scale, etc., are adjusted according to actual needs. The top 10% of fitness must be an even number, and if they are not even, they are forcibly converted to an even number by adding or subtracting one.

[0083] For example: The first 10 iterations: 3 schemes are randomly selected from the current population (30 schemes), which means the tournament size is 3; The parent generation can be selected from the three sets with the highest fitness to participate in subsequent breeding. Repeat the above operation 10 times to obtain 10 parent generation schemes.

[0084] The last 20 iterations: 7 schemes are randomly selected from the current population (30 schemes), which means the tournament size is 7; Select the one with the highest fitness among these 7 sets as the parent generation to participate in subsequent breeding; Repeat the above operation 20 times to obtain 20 parent generation schemes.

[0085] The total parent generation scheme is the sum of the parent generation schemes from the two iterations.

[0086] Elite retention: The two (3-1) parent schemes with the highest fitness are retained from the total parent schemes and do not participate in crossover mutations, directly entering the next generation population.

[0087] S332 uses single-point crossover to combine high-quality parent chromosomes to generate offspring chromosomes. The steps are as follows: The 28 sets (30-2) of parent generations were randomly divided into 14 pairs, with 2 sets of parent generations in each pair, such as parent generation 1 and parent generation 2. For each pair of parents, one crossover point is randomly selected. For example, if the chromosome length is 10, the crossover point = 4, that is, the first 4 positions are the head and the last 6 positions are the tail. Exchange tail genes: Parent 1's head + Parent 2's tail -> Offspring 1; Parent 2's head + Parent 1's tail -> Offspring 2. Offspring constraint verification and adjustment: If the offspring exceeds the budget, remove the project with the largest investment and the lowest total benefit score; if the offspring category does not meet the standard, supplement the corresponding category project with the smallest investment. For example, if the safety category is insufficient, supplement the safety category project with the lowest investment, until all constraints are satisfied. Example: Parent 1 chromosome [1,0,1,0,1,0,0,1,0,0] (total investment 20.5 million), Parent 2 chromosome [0,1,1,0,0,1,0,0,1,0] (total investment 19.8 million), crossover point = 4, offspring 1 is [1,0,1,0,0,1,0,0,1,0], calculate the total investment = 5.5 million + 6 million + 4.5 million + 5.2 million = 21.2 million ≤ 50 million, there are 2 safety category projects and 1 intelligent category project, so the constraints are satisfied.

[0088] Assume that all offspring meet the constraint condition, i.e., 28 offspring chromosomes.

[0089] S333, random mutation of the offspring chromosomes to obtain the optimal scheme set, the steps are as follows: For each of the 28 offspring, the mutation probability of each offspring chromosome segment was set to 5%. If a mutation occurs, randomly select one gene locus on the offspring chromosome and flip its value from 0 to 1 or 1 to 0. If the mutation exceeds the budget or the category does not meet the requirements, the gene locus for mutation is selected again. If the original mutation exceeds the budget, the gene locus of the newly added project is flipped back to 0 until the solution is feasible. Example: Offspring chromosome 1 [1,0,1,0,0,1,0,0,1,0], randomly select gene locus 8, currently = 0, corresponding to project 8, investment 5 million, flip -> 1, the mutated chromosome [1,0,1,0,0,1,0,1,1,0], total investment = 550 + 600 + 450 + 500 + 520 = 26.2 million ≤ 50 million, then the constraint is satisfied.

[0090] Assume that all conditions are met after the mutation, that is, there are 28 offspring chromosomes after the mutation.

[0091] Set a termination condition; if the condition is met, the iteration stops. Specifically: Termination condition 1: Iteration count reached: The preset maximum number of iterations is 30 generations, which can be adjusted according to actual needs; Termination condition 2: Stable fitness: The optimal solution has been stabilized and there is no better solution if the fitness change is ≤0.1% in the population with the highest fitness over 10 consecutive generations. The 10 consecutive generations and fitness change ≤0.1% may be adjusted according to the actual situation.

[0092] Assuming termination condition 1 is met, the last generation population consists of the two parent schemes with the highest fitness retained in S3131 plus the 28 mutated offspring chromosomes in S3133, for a total of 30 chromosomes, which is the optimal scheme set.

[0093] S4 outputs the optimal investment allocation scheme set by selecting the best scheme set.

[0094] In this embodiment, the evaluation results of candidate projects in the preferred scheme set are sorted, and the top-ranked project schemes are output. The set of project schemes is the optimal investment allocation scheme set, specifically: After the iteration terminates, the top 5 fitness scores are selected from the last generation of the population. These top 5 scores are reference values, and the selected schemes are retained according to actual needs to obtain the optimal investment allocation scheme set. Each scheme must include the following information: Selected projects: Specify the project ID and its category, such as "Project 1 - Security", "Project 5 - Intelligentization", "Project 8 - Economy"; Key indicators: Total actual investment amount, such as 20.5 million yuan; total benefit score, such as 24.49%; total risk score, such as 16.01%. Applicable scenarios: Describe based on regional needs and business priorities, such as "applicable to the eastern region with high security requirements and limited budget" or "applicable to the western region with priority on intelligent upgrades and sufficient budget".

[0095] Example 2, Figure 2 An apparatus for a method of dividing a power distribution network into grids is presented, comprising a collection and processing module, an input and output module, a calculation and filtering module, and a final output module. The data collection and processing module is used to collect historical data of power grid technical renovation projects, extract features and standardize candidate projects to obtain standardized feature vectors. The input / output module is used to input standardized feature vectors into the deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtain the prediction results in a structured manner. The calculation and screening module is used to calculate the preconditions based on the prediction results, and to sort and screen the candidate projects using a generative intelligent genetic algorithm to obtain a set of preferred solutions. The final output module is used to select the optimal investment allocation scheme set based on the preferred scheme set and the candidate project ranking results.

[0096] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0097] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0098] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0099] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0100] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0101] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for evaluating and optimizing power grid technological upgrades, characterized in that, Includes the following steps: Historical data on power grid technical upgrade projects are collected, and candidate projects are subjected to feature extraction and standardization to obtain standardized feature vectors. Standardized feature vectors are input into a deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and the prediction results are obtained in a structured manner. Based on the prediction results, the prerequisites are calculated, and the candidate projects are sorted and screened using a generative intelligent genetic algorithm to obtain a set of preferred solutions. Based on the preferred solution set and the ranking results of candidate projects, the optimal investment allocation solution set is selected.

2. The deep learning-based power grid upgrading evaluation and optimization method according to claim 1, characterized in that, The historical data of the power grid technical upgrade project includes basic project attribute data, project problem and demand data, project implementation content data, project results and acceptance data, project risk and deviation data, and project related assets and operation data.

3. The deep learning-based power grid upgrading evaluation and optimization method according to claim 2, characterized in that, The feature extraction specifically includes: By using historical data of power grid technical upgrade projects, we can extract project attribute characteristics, problem type characteristics, and effectiveness indicator characteristics. Establish a global attribute feature dictionary, a global problem dictionary, and a global indicator dictionary based on project attribute characteristics, problem type characteristics, and performance indicator characteristics; Attribute feature vectors, problem feature vectors, and indicator feature vectors are created based on the global attribute feature dictionary, global problem dictionary, and global indicator dictionary.

4. The deep learning-based power grid upgrading evaluation and optimization method according to claim 3, characterized in that, The standardized feature vector is obtained by concatenating all standardized feature vectors according to the problem point, attribute feature, and performance indicator.

5. The deep learning-based power grid upgrading evaluation and optimization method according to claim 4, characterized in that, The process involves inputting standardized feature vectors into a deep learning model to predict the performance indicators and implementation risks of each candidate project. The resulting predictions are then structured to obtain the final predictions. Configure the hierarchy and parameters of the deep learning model; A deep learning model was trained based on historical data from power grid upgrade projects. Build a test set, evaluate the model accuracy, and select qualified deep learning models; A qualified deep learning model is used to output and structure the prediction results for the candidate items that need to be evaluated.

6. The deep learning-based power grid upgrading evaluation and optimization method according to claim 5, characterized in that, The generative intelligent genetic algorithm is specifically as follows: Convert candidate items into chromosomes and intelligently initialize the population; Candidate projects are evaluated and ranked using a fitness function and preconditions. Based on the initial population and evaluation results, the optimal scheme set is obtained through genetics.

7. The deep learning-based power grid upgrading evaluation and optimization method according to claim 6, characterized in that, The specific genetics referred to are: Adaptive tournament selection was used to screen for high-quality paternal chromosomes. High-quality parent chromosomes are combined to generate offspring chromosomes through single-point crossover. Random mutations were performed on the offspring chromosomes to obtain a set of optimal schemes.

8. The deep learning-based power grid upgrading evaluation and optimization method according to claim 7, characterized in that, The adaptive tournament selection method selects parent chromosomes based on the tournament size defined in the iteration stage, and retains parent chromosomes with high fitness to directly enter the offspring population.

9. The power grid technological upgrading evaluation and optimization method based on deep learning according to claim 8, characterized in that, The process of selecting the optimal investment allocation scheme set based on the preferred scheme set and the candidate project ranking results is as follows: The evaluation results of candidate projects in the preferred solution set are sorted, and the top-ranked project solutions are output. The set of project solutions is the optimal investment allocation solution set.

10. An apparatus for evaluating and optimizing power grid technological upgrading based on any one of claims 1-9, comprising: The data collection and processing module is used to collect historical data of power grid technical renovation projects, extract features and standardize candidate projects to obtain standardized feature vectors. The input / output module is used to input standardized feature vectors into the deep learning model to predict the effectiveness indicators and implementation risks of each candidate project, and obtain the prediction results in a structured manner. The calculation and screening module is used to calculate the preconditions based on the prediction results, and to sort and screen the candidate projects using a generative intelligent genetic algorithm to obtain a set of preferred solutions. The final output module is used to select the optimal investment allocation scheme set based on the preferred scheme set and the candidate project ranking results.

Images