Intelligent Optimization Carbon Assessment Method and Platform for Green Design Parameters of Insulators

By combining a carbon emission factor library and a proxy model, the inefficiency and parameter conflict problems of existing carbon assessment methods are solved, achieving efficient and accurate carbon assessment and synchronization of multi-node optimization, while ensuring the balance of electrical performance.

CN123115901APending Publication Date: 2026-11-17JIANGXI SHUNSEN ELECTRIC POWER EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611339260.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing carbon assessment methods rely on manual, phased calculations, which are inefficient and fail to reflect the time-varying characteristics of power grid emission factors in the power supply area. Furthermore, in distributed, multi-node parallel optimization scenarios, the update rhythm and magnitude of parameters are inconsistent, which can easily lead to parameter coverage conflicts and make it difficult to guarantee synchronization consistency.

Method used

The carbon emission factor library is organized by three levels of indexing: material type, manufacturing process, and power supply area. Combined with carbon emission proxy model and electrical performance proxy model, the evolutionary state matrix is ​​constructed through iterative optimization using robust carbon index and feasibility probability constraints. A classification model is used to handle parameter conflicts, and mechanism accounting is performed to update the model when the uncertainty exceeds the threshold.

Benefits of technology

It improves the accuracy and risk controllability of carbon assessment, shortens the design optimization time, enhances the synchronous stability and electrical performance of multi-node optimization, and realizes the traceability of carbon assessment results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115901A_ABST
    Figure CN123115901A_ABST
Patent Text Reader

Abstract

This application relates to the field of green design technology for power equipment, and discloses an intelligent optimization carbon assessment method and platform for green design parameters of insulators. The method includes: obtaining green design parameters of the insulators, querying a carbon emission factor library to obtain initial carbon assessment values; training a carbon emission proxy model and an electrical performance proxy model, outputting the carbon emission prediction mean, carbon emission prediction uncertainty, and feasibility probability; constructing an evolutionary state matrix and extracting the material ratio update frequency level and structural dimension modification range level through a classification model; clustering candidate design parameter features when the level product exceeds a conflict threshold, adjusting the level, and obtaining the target adjustment variables for component parameters; performing mechanism verification on the target candidate parameters and incrementally updating the carbon emission proxy model when the model uncertainty component exceeds a threshold; synchronizing the green design parameters of each node and outputting the carbon assessment results. This application improves the accuracy of insulator carbon emission assessment and the stability of multi-node optimization synchronization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of green design technology for power equipment, and in particular to a method and platform for intelligent optimization of carbon assessment of green design parameters for insulators. Background Technology

[0002] Design parameters such as the outer diameter of the insulator's sheds, creepage distance, core rod size, and material ratio determine not only electrical performance but also carbon emission levels at each stage of raw material acquisition, manufacturing, and end-of-life recycling. Green design requires accurate assessment of carbon emissions and optimization of design parameters during the design phase. Existing carbon assessment methods rely on manual, phased calculations and static emission factors, resulting in low calculation efficiency and an inability to reflect the time-varying characteristics of emission factors in the power grid of the power supply area. Existing design optimization methods are mostly single-node serial optimizations, requiring repeated calls to mechanism calculations or electric field simulations for candidate parameter evaluation, which is computationally time-consuming. In distributed, multi-node parallel optimization scenarios, the update rhythm and magnitude of parameters at each node are inconsistent, easily leading to parameter coverage conflicts and making it difficult to guarantee synchronization consistency. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides an intelligent optimization carbon assessment method and platform for green design parameters of insulators, which is used to accurately assess carbon emissions during the green design process of insulators and maintain the consistency and stability of parameter synchronization in multi-node parallel optimization scenarios.

[0004] Firstly, this application provides a method for intelligent optimization of carbon assessment of green design parameters for insulators, the method comprising: Obtain green design parameters for insulators, query the carbon emission factor library, and calculate carbon emissions by life cycle stage to obtain an initial carbon assessment value. Using green design parameters as input and initial carbon assessment values ​​and electrical performance indicators as labels, we train carbon emission proxy models and electrical performance proxy models, and output the carbon emission prediction mean, carbon emission prediction uncertainty and feasibility probability. The search space for green design parameters is divided into distributed nodes. The robust carbon index is used as the fitness and the feasibility probability is used as the constraint to iteratively optimize and construct an evolutionary state matrix. When the elements of the evolutionary state matrix exceed the preset state threshold, the update frequency and update magnitude of the node data are obtained, input into the pre-trained classification model, and the update frequency level of the material ratio and the modification magnitude level of the structural size are extracted. If they do not exceed the threshold, the iteration continues. When the product of the material ratio update frequency level and the structural dimension modification range level exceeds the preset conflict threshold, the clustering of candidate design parameter features yields the component feature target cluster. The material ratio update frequency level and the structural dimension modification range level are adjusted according to the cluster center, and the component parameter target adjustment variable is obtained. If it does not exceed the threshold, the variable from the previous round is used. When the model uncertainty component in the carbon emission prediction uncertainty exceeds the preset uncertainty threshold, the carbon emission mechanistic calculation is performed by decomposing the carbon emission by life cycle stage and the carbon emission proxy model is updated. The green design parameters of each node are updated by adjusting the component parameter target variables to complete the synchronization and output the carbon assessment results.

[0005] Secondly, this application provides an intelligent carbon assessment platform for optimizing green design parameters of insulators, the platform comprising: The parameter acquisition and carbon accounting module is used to acquire green design parameters for insulators, query the carbon emission factor library, and calculate the initial carbon assessment value by life cycle stage. The surrogate model training module is used to train the carbon emission surrogate model and the electrical performance surrogate model, and outputs the carbon emission prediction mean, carbon emission prediction uncertainty and feasibility probability. The distributed optimization module is used to divide the search space into distributed nodes for iterative optimization, construct an evolutionary state matrix, and extract the material ratio update frequency level and structural size modification level through a classification model. The conflict resolution module is used to cluster candidate design parameter features when the product of gears exceeds the conflict threshold, adjust the gears, and obtain the target adjustment variable of component parameters with corrected weights. The feedback update module is used to perform mechanism verification of the target candidate parameters and incrementally update the carbon emission proxy model when the model uncertainty components exceed the threshold. The synchronization and evaluation output module is used to synchronize the green design parameters of each node and output the carbon assessment results.

[0006] Compared with the prior art, the beneficial effects of this application are at least as follows: (1) The carbon emission factor library is organized by three levels of index: material type, manufacturing process and power supply area. The power grid emission factor is dynamically updated according to the power supply area time series power grid data. The initial carbon assessment value corresponds to the actual power supply structure during the manufacturing period, which improves the accuracy of carbon assessment. (2) The carbon emission proxy model switches between Gaussian process regression and random forest regression according to the number of training samples. The carbon emission prediction uncertainty distinguishes between the model uncertainty component and the factor propagation component. When the model uncertainty component exceeds the threshold, the incremental update is fed back by mechanism accounting, which improves the accuracy and risk controllability of carbon emission prediction. The evaluation of candidate parameters has been reduced from minutes or more of repeated calls to mechanism accounting or electric field simulation to milliseconds of computation by the proxy model, which shortens the evaluation time of design optimization. (3) The evolution state matrix monitors the synchronization state with dimensionless relative deviation. Synchronization conflicts are resolved by classification model level adjustment and clustering, which improves the stability of multi-node optimization synchronization. (4) The robust carbon index is combined with the feasibility probability constraint. Before output, the recommended parameters are back-substituted into the electrical performance proxy model for verification, taking into account both carbon emission optimization and electrical performance compliance. The carbon assessment results are traceable. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart of a method for intelligent optimization of carbon assessment of green design parameters for insulators, as described in an embodiment of this application. Figure 2 This is a schematic diagram of the training and gear mapping process of the classification model in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of an intelligent optimization carbon assessment platform for green design parameters of insulators, as described in an embodiment of this application. Detailed Implementation

[0009] This application provides a method and platform for intelligent optimization of carbon assessment of green design parameters for insulators. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, platform, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0010] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent optimization carbon assessment method for green design parameters of insulators in this application includes: Step S1: Obtain the green design parameters for insulators, query the carbon emission factor library, and calculate the carbon emissions according to the life cycle stage to obtain the initial carbon assessment value.

[0011] Furthermore, green design parameters include the outer diameter of the umbrella skirt, creepage distance, core rod size, and material composition; a carbon emission factor database can be consulted, including: The system queries a carbon emission factor database organized by material type, manufacturing process, and power supply area index to obtain the material carbon emission factor corresponding to the material ratio, the manufacturing carbon emission factor corresponding to the insulator forming process, and the power grid emission factor corresponding to the power supply area of ​​the manufacturing plant. The material carbon emission factor, manufacturing carbon emission factor, and power grid emission factor all have an uncertainty range. The power grid emission factor is dynamically updated based on the time-series power grid data of the power supply area.

[0012] Furthermore, carbon emissions are broken down and calculated according to life cycle stages to obtain an initial carbon assessment value, including: The life cycle includes at least the raw material acquisition stage, the manufacturing stage, and the end-of-life recycling stage. The carbon emissions of the raw material acquisition stage are obtained by summing the products of the amount of each component material used and its corresponding carbon emission factor. The mass of each component of the insulator and the net weight of the product are estimated from the outer diameter of the skirts, creepage distance, core rod size, and material ratio. The process fuel heat consumption is calculated based on the product's net weight and unit mass heat consumption coefficient. The electricity consumption is calculated based on the process hours and equipment rated power. The carbon emissions of the manufacturing stage are obtained by multiplying the process fuel heat consumption and the manufacturing carbon emission factor, and then adding the electricity consumption and the grid emission factor. The carbon deduction for the end-of-life recycling stage is calculated based on the recovery rate of recyclable components and the recycling deduction factor. The algebraic sum of the carbon emissions of the raw material acquisition stage, the manufacturing stage, and the carbon deduction is determined as the initial carbon assessment value.

[0013] Specifically, the green design parameters for insulators are entered into the product data management system by the design team or exported from the 3D model. These parameters include the outer diameter of the sheds, creepage distance, core rod size, and material ratio. The outer diameter of the sheds refers to the maximum diameter of the outer edge of the sheds, measured in mm. The creepage distance refers to the shortest discharge path length measured along the surface of the insulator sheds between the metal fittings at both ends. The core rod size refers to the diameter of the fiberglass core rod that bears the mechanical load. The material ratio refers to the mass ratio between the high-temperature vulcanized silicone rubber matrix, aluminum hydroxide filler, and silica functional additives. The above four types of parameters are arranged in a fixed order to form a design parameter vector. Therefore, the design parameter vector includes at least the umbrella skirt outer diameter, creepage distance, core rod size component, and material ratio component. The material ratio component records the matrix, filler, and additive sub-components in parallel. The vector dimension is determined by the actual dimension after each component is expanded in parallel. The dimension of subsequent steps such as clustering, distance, and sensitivity calculation are all based on this actual dimension. This vector serves as both the retrieval basis for the three-level index of the carbon emission factor library and as the input quantity jointly referenced in subsequent calculations and optimizations.

[0014] A carbon emission factor database is pre-established and stored on the evaluation platform. Each record in the database is a triple consisting of an index key, a factor value, and an uncertainty range. The carbon emission factor refers to the carbon dioxide equivalent converted from a unit of activity: the material carbon emission factor is expressed in kilograms of carbon dioxide equivalent per kilogram of material (kgCO2e / kg), the manufacturing carbon emission factor is expressed in kilograms of carbon dioxide equivalent per megajoule of process fuel (kgCO2e / MJ), and the grid emission factor is expressed in kilograms of carbon dioxide equivalent per kilowatt-hour (0.55 kgCO2e / kWh). The index key is organized in three levels: material type, manufacturing process, and power supply region. Searches are performed level by level: the first level uses the name of each component in the material composition to find the corresponding material carbon emission factor; the second level uses the molding process used for the insulator to find the corresponding manufacturing carbon emission factor for the process fuel; and the third level uses the power supply region of the manufacturing plant to find the grid emission factor. The order of the three levels corresponds one-to-one with the index hierarchy. The result of the previous level limits the value range of the subsequent level's search, thus avoiding cross-database mismatches. The uncertainty interval is the mean of the statistical sample of the factor plus or minus two standard deviations. The standard deviation is calculated from the factor sample set. The upper and lower bounds of the interval are written into the database along with the factor value as input for subsequent uncertainty propagation calculations.

[0015] The grid emission factor is dynamically updated based on the time-series grid data of the power supply area: the average emission factor of the regional grid for the month is calculated based on the monthly power generation structure of the area and the carbon content per unit calorific value of each fuel type. The calculation formula is: Average emission factor of the regional grid for the month = Sum of the products of monthly power generation and emission factor of each fuel type ÷ Total monthly power generation of the area. The emission factor of each fuel type is converted from its carbon content per unit calorific value and oxidation rate. The old value in the database is replaced monthly, and its uncertainty range is updated synchronously. After the annual data is released, the monthly series is corrected based on the annual average. This ensures that the grid emission factor hit by the third-level index corresponds to the actual power supply structure during the manufacturing period.

[0016] After factor matching is completed, the calculation is decomposed according to the life cycle stage. Raw material acquisition stage: First, the amount of each component is estimated from the structural dimensions. The umbrella skirt profile is taken from the pre-stored standard umbrella template. The number of umbrellas and the spacing between umbrellas are allocated according to the creepage distance of the template unit. The generatrix equation is retrieved with the template. The volume of the umbrella skirt is obtained by rotating the volume integral of the outer diameter of the umbrella skirt and the creepage distance according to the generatrix equation of the template. The volume of the mandrel is obtained by the mandrel size. The two are multiplied by the corresponding material density to obtain the mass of the silicone rubber part and the mandrel part. Then, the mass of the silicone rubber part is decomposed into the mass of matrix, filler and additives according to the material ratio. The carbon emission factor of each component is multiplied by each component and summed to obtain the carbon emission of the raw material acquisition stage. Manufacturing Stage: Energy consumption in the manufacturing stage is calculated separately for two paths: process fuel heat consumption and electricity consumption. Process fuel heat consumption is calculated by multiplying the net weight of the product by the heat consumption coefficient per unit mass, which is obtained by regression analysis of energy consumption statistics for similar processes. Electricity consumption is calculated by multiplying the process hours by the rated power of the equipment. The carbon emissions on the fuel side are obtained by multiplying the process fuel heat consumption by the manufacturing carbon emission factor, and the carbon emissions on the electricity side are obtained by multiplying the electricity consumption by the grid emission factor. The two are added together to obtain the carbon emissions in the manufacturing stage. The two paths correspond to the units of measurement of the manufacturing carbon emission factor and the grid emission factor, respectively, and are dimensionlessly consistent. Waste Recycling Stage: The carbon deduction is obtained by multiplying the mass of the recyclable component by its recovery rate, and then by the recycling deduction factor. The recycling deduction factor is the difference between the virgin material factor and the recycled material factor for that component. Finally, the carbon emissions from the raw material acquisition stage, the manufacturing stage, and the waste recycling stage are algebraically summed according to a convention of positive, positive, and negative signs to obtain the initial carbon assessment value. This value is used as a training label and input into the training phase of the surrogate model.

[0017] Step S2: Using green design parameters as input and initial carbon assessment values ​​and electrical performance indicators as labels, train the carbon emission proxy model and the electrical performance proxy model, and output the carbon emission prediction mean, carbon emission prediction uncertainty and feasibility probability.

[0018] Furthermore, training the carbon emission proxy model and the electrical performance proxy model includes: Training samples are generated using Latin hypercube sampling within the allowable range of each component of the green design parameters. Among the electrical performance indicators, the creepage distance label is calculated analytically from the green design parameters, while the pollution flashover withstand voltage label is obtained from experimental data or electric field simulation. When the number of training samples does not exceed the preset sample threshold, a Gaussian process regression model is used to train the carbon emission surrogate model, and the model uncertainty component in the carbon emission prediction uncertainty is determined by the square root of its output posterior variance. When the number of training samples exceeds the preset sample threshold, a random forest regression model is used to train the carbon emission surrogate model, and the model uncertainty component in the carbon emission prediction uncertainty is determined by the sample standard deviation between the outputs of each regression tree. The carbon emission prediction uncertainty also includes a factor propagation component, which is calculated by taking the square root of the sum of the squares of the products of the partial derivatives of various carbon emission factors and the factor standard deviations. The formula is linear with respect to the factors, and the partial derivatives correspond to the activity quantity. The electrical performance surrogate model is trained using electrical performance indicators as labels, and the feasibility probability is calculated by the predicted mean and predicted variance of its output, used to characterize the probability that the candidate design parameters meet the rated electrical performance requirements.

[0019] In this embodiment, the surrogate model refers to the machine learning regression fitting of the mapping relationship between the design parameter vector and the target quantity. After fitting, a single matrix operation at the millisecond level replaces the electrical simulation calculations required for each candidate parameter evaluation, and simultaneously outputs the carbon emission prediction mean, carbon emission prediction uncertainty, and feasibility probability. The carbon emission surrogate model also serves as the carrier for uncertainty quantification and backfeeding updates, which is a prerequisite for high-frequency evaluation of candidate parameters in subsequent iterative optimization. Training samples are generated by Latin hypercube sampling within the allowable value range of each component of the green design parameters, with a sample size of 500 to 2,000. Latin hypercube sampling divides each dimension's value range into several sub-intervals and samples once in each sub-interval, so that the samples are evenly distributed in the design parameter space. For each sample, the life cycle decomposition calculation in step S1 is called to obtain the initial carbon assessment value, which serves as the training label for the carbon emission surrogate model; at the same time, the creepage distance and pollution flashover withstand voltage are calculated, which serve as the training labels for the electrical performance surrogate model. The creepage distance is calculated analytically by dividing the creepage distance of the sample by the system's highest operating voltage, directly derived from the sample parameters. The pollution flashover withstand voltage cannot be analytically determined from the parameters; its label is obtained from artificial pollution test data, or by substituting the sample parameters into a pre-built finite element simulation template for the umbrella skirt electric field to solve in batches, recording the withstand voltage value for each sample. The simulation and test data are merged to form the label for this dimension. Before training, the input vector and label are normalized separately, i.e., each component is subtracted from the sample mean and divided by the sample standard deviation, to eliminate the influence of dimensional differences on the kernel function measurement.

[0020] The carbon emission proxy model selects a regression model based on the relationship between the number of training samples and a preset sample threshold, with the preset sample threshold set to 2,000. When the number of training samples does not exceed the preset sample threshold, Gaussian process regression is used for training. A Gaussian process refers to a set of random functions defined on the input space, where the function values ​​at any finite number of input points follow a joint Gaussian distribution. The task of regression is to find the posterior distribution of the target quantity at new input points after observing the training samples. The kernel function is the radial basis function, and the hyperparameters include the length scale of each dimension and the noise variance, which are solved by maximizing the marginal likelihood function of the training data. During inference, the carbon emission prediction mean at new input points is given by a linear combination of training labels weighted by the kernel function. The model uncertainty component in the carbon emission prediction uncertainty is the square root of the posterior variance. Both vary with the kernel distance between the query point and the training samples. The posterior variance of query points far from the dense training sample area increases, and the model uncertainty component increases accordingly. The factor propagation component is calculated analytically using the first-order sensitivity of the kernel formula, and its sum with the model uncertainty component constitutes the carbon emission prediction uncertainty. When the number of training samples exceeds a preset threshold, the computational cost of Gaussian process regression increases exponentially with the number of training samples, making training and inference unacceptably time-consuming. In this case, a random forest regression model is used to train the carbon emission surrogate model. The random forest performs bootstrap resampling on the training set and trains multiple regression trees. Each tree recursively partitions the input space at each node according to the splitting criterion of maximizing variance reduction. The carbon emission prediction mean is the arithmetic mean of the outputs of all trees, and the carbon emission prediction uncertainty is the sample standard deviation between the outputs of each regression tree. As an alternative, a quantile regression forest is used, recording the label values ​​at the leaf nodes and outputting the 10th and 90th quantiles. The difference between the two quantiles is used as the carbon emission prediction uncertainty.

[0021] The electrical performance proxy model and the carbon emission proxy model share the same training sample set, trained using creepage distance and flashover withstand voltage as labels. The model structure and training method are identical, outputting the predicted mean and variance of electrical performance. The feasibility probability is calculated as follows: The value of any electrical performance index at a given candidate parameter is considered a normally distributed random variable with the predicted mean as its expected value and the square root of the predicted variance as its standard deviation. When a lower limit is set for this index, the feasibility probability equals the probability that the random variable is greater than the lower limit, calculated using the quotient of the standard normal distribution function. The quotient is the predicted mean minus the lower limit, divided by the predicted standard deviation. When an upper limit is set, the probability that the random variable is less than the upper limit is taken. The feasibility probabilities of each electrical performance index are multiplied to obtain the feasibility probability of the candidate design parameter. When the predicted standard deviation of a certain index is 0, the feasibility probability of that index is either 1 or 0 depending on whether the predicted mean meets the rated value, to avoid division by zero errors in the calculation. The obtained carbon emission prediction mean, carbon emission prediction uncertainty (including model uncertainty component and factor propagation component), and feasibility probability are output along with the query point, serving as the fitness and constraint inputs for the iterative optimization in step S3.

[0022] Step S3: Divide the search space of green design parameters into distributed nodes, and iterate to optimize using robust carbon index as fitness and feasibility probability as constraints. Construct an evolutionary state matrix based on the average carbon emission prediction of each node and the number of iterations. When the elements of the evolutionary state matrix exceed the preset state threshold, obtain the node data update frequency and update magnitude, input them into the classification model pre-trained based on historical synchronization cycle data, and extract the material ratio update frequency level and structural size modification magnitude level. If they do not exceed the threshold, continue iterating.

[0023] In the evolutionary state matrix, rows correspond to distributed nodes, columns correspond to iteration periods, and matrix elements are the ratio of the absolute value of the deviation of the carbon emission prediction mean of the best individual in the current period relative to the carbon emission prediction mean of all nodes in the same generation to the carbon emission prediction mean of all nodes in the same generation.

[0024] Furthermore, iterative optimization is performed with feasibility probability as a constraint, including: When the feasibility probability is less than a preset probability threshold, a penalty function is applied to the fitness. The penalty term of the penalty function is positively correlated with the degree to which the feasibility probability deviates from the probability threshold. The fitness corrected after applying the penalty function is equal to the sum of the robust carbon index and the penalty term, so that candidate design parameters with insufficient electrical performance are eliminated in the iteration. The feasibility probability of the candidate design parameter is obtained by multiplying the feasibility probabilities of each electrical performance index. When the prediction standard deviation of any electrical performance index is 0, the feasibility probability of that index is taken as 1 or 0 depending on whether the prediction mean meets the rated value. The electrical performance index includes at least creepage distance and flashover withstand voltage. The creepage distance is an analytically determined quantity, and its prediction variance does not participate in the normality assumption calculation. The feasibility probability is directly taken as 1 or 0 depending on whether the prediction mean meets the rated value.

[0025] Furthermore, the robust carbon index equals the mean carbon emission forecast plus the product of the risk aversion coefficient and the uncertainty of the carbon emission forecast, specifically: Robust carbon index = mean carbon emission prediction + risk aversion coefficient × carbon emission prediction uncertainty; carbon emission prediction uncertainty is the sum of model uncertainty component and factor propagation component. The model uncertainty component is determined by the square root of the posterior variance of the model output when using a Gaussian process regression model, and by the sample standard deviation between the outputs of each regression tree when using a random forest regression model. The factor propagation component is obtained by analyzing the carbon emission factor with attached uncertainty interval according to the first-order sensitivity of the accounting formula. The risk aversion coefficient is set according to the confidence requirement of carbon assessment. The higher the confidence requirement, the larger the risk aversion coefficient, causing the optimization result to shift towards the design parameter region with low uncertainty.

[0026] In this implementation, the evolutionary state matrix is ​​constructed at the end of each iteration cycle, with rows corresponding to distributed nodes and columns corresponding to iteration cycles. The matrix elements are defined by relative deviation: the average carbon emission prediction of the best individual at that node in the current period is subtracted from the average carbon emission prediction of all nodes in the same generation, the absolute value of the difference is divided by the average carbon emission prediction of all nodes in the same generation, and the quotient is a dimensionless quantity. This relative deviation characterizes the degree to which the carbon emission level of a single node deviates from the global average, decoupling it from the absolute magnitude of carbon emissions: the average carbon emission predictions of different evaluation objects or different iteration stages may differ by several times. If the dimensional carbon emission prediction average is directly used as the matrix element, the same state threshold cannot be uniformly set across magnitudes. After the relative deviation is dimensionless, the state threshold maintains the same meaning across all nodes throughout the entire iteration process. The state threshold is set to 0.2, with a range of 0.1 to 0.3; each element is compared, and nodes whose elements exceed the state threshold are determined to be out-of-synchrony nodes and proceed to the parameter extraction stage.

[0027] The classification model is characterized by the frequency and magnitude of node data updates. (See also...) Figure 2Node data update frequency refers to the average number of times a node submits design parameter updates to the aggregation end within a preset sliding window length, measured in times per iteration cycle. Historical frequency features from classification training and grade execution are both statistically analyzed using the same sliding window diameter, with 0.5 times per iteration cycle corresponding to one submission every two cycles. Node data update magnitude refers to the normalized Euclidean distance between two adjacent submitted design parameter vectors. Features are first normalized using minimum-maximum scaling, linearly scaling both the frequency and distance to the 0-1 range to eliminate the dimensional differences between them; otherwise, the distance feature, with its larger numerical range, would dominate the classification boundary. Training samples are taken from frequency and amplitude pairs recorded in historical synchronization cycles. Conflict category labels are back-labeled based on whether a parameter overlay event occurred within that cycle, using synchronization logs: each synchronization log records the node submission timestamp, the submitted design parameter vector, and whether a subsequent parameter overlay event occurred. Back-labeling involves statistically analyzing frequency and amplitude pairs for each historical synchronization cycle and checking if an overlay event occurred within that cycle. If it did, it is labeled as a conflict category; otherwise, it is labeled as a normal category. Conflict categories are further divided into four classes based on their frequency and amplitude relative to the median of historical conflict samples. Normal categories are not further subdivided. The classification model outputs the conflict category along with the amplitude-dominant component group. The amplitude-dominant component group is determined by the larger of the structural dimension update amplitude and the material ratio update amplitude. If they are equal, the structural dimension becomes the dominant component, used for gear adjustment direction distribution. The classification model uses a support vector machine or decision tree, with input being a normalized two-dimensional feature vector and output being the conflict category. Before the number of historical synchronization cycle samples reaches a preset number, each node executes at the default low-frequency, low-amplitude gear. Once the preset number of samples is accumulated, the classification model is activated.

[0028] The classification output is a conflict category and cannot be directly used as a parameter tuning parameter. Therefore, each conflict category has a pre-stored corresponding gear adjustment rule. The gear adjustment rule records the adjustment direction and adjustment level of the material ratio update frequency gear and the structural size modification range gear relative to the currently executed gear. The material ratio update frequency gear takes three levels: 2 times, 1 time, and 0.5 times per iteration cycle. The structural size modification range gear takes three levels: 5%, 2.5%, and 1%. The mapping table is set according to synchronization experience: high-frequency, high-amplitude classes are due to complex causes of conflict, so the highest level is maintained until the product of the levels exceeds the conflict threshold, i.e., the material ratio update frequency is set to 2 times per iteration cycle and the structural size modification amplitude is set to 5%, and then proceed to step S4 clustering resolution; the cause of high-frequency, low-amplitude classes is excessively high submission frequency, so the material ratio update frequency is reduced by one level from the current level and the structural size modification amplitude is set to 1%; the cause of low-frequency, high-amplitude classes is excessively high modification amplitude, so the structural size modification amplitude is reduced by one level from the current level and the material ratio update frequency is set to 0.5 times per iteration cycle; low-frequency, low-amplitude classes are occasional coverage under the existing rhythm, so the material ratio update frequency is set to 0.5 times per iteration cycle and the structural size modification amplitude is set to 1%, and a random delay that does not cross the current iteration cycle is added for staggered submission; normal classes do not go through the mapping table and the current control parameters remain unchanged. The downgraded levels are distributed in two groups based on magnitude: when structural dimensions are dominant, the structural dimension modification magnitude level is downgraded first; when material proportions are dominant, the material proportion update frequency level is downgraded first. If they are equal, structural dimensions are prioritized. If the priority downgraded level is already at a boundary level, the other level is downgraded by one level. If both levels are at boundary levels, staggered submissions are used as a fallback. The classified output, after looking up the mapping table, obtains the directional material proportion update frequency and structural dimension modification magnitude of the node, which serves as the input for conflict judgment and level selection in step S4. The frequency level is used to determine the submission rhythm. The magnitude level, after being converted into a normalized proportional quantity, serves as the upper limit constraint for the synchronous update step size of the structural dimension components. The step size is based on the structural dimension modification magnitude value obtained from clustering, and is truncated when it exceeds the value. The step size of the material proportion components is based on the material proportion update rate obtained from clustering, and is truncated by the preset upper limit of the proportion step size.

[0029] Step S4: When the product of the material ratio update frequency level and the structural size modification range level exceeds the preset conflict threshold, cluster the candidate design parameter features to obtain the component feature target cluster. Adjust the material ratio update frequency level and the structural size modification range level according to the cluster center, and use the deviation of the carbon emission prediction mean of each node as the correction weight to obtain the component parameter target adjustment variable. If it does not exceed the threshold, the variable from the previous round is used.

[0030] Furthermore, obtain the target adjustment variables for the component parameters, including: The cluster centers of the target clusters of component features are extracted, and the Euclidean distance between the centers is calculated to obtain the center point distance matrix. The center point distance matrix is ​​a symmetric square matrix with the number of cluster centers as its order, the diagonal elements are 0, and the off-diagonal elements are the Euclidean distances between pairwise cluster centers. When the maximum distance value exceeds a preset distance threshold, the two corresponding center points are taken as target center points. The material ratio update rate is calculated based on the material ratio data of the target center points, and the structural size modification magnitude is obtained based on the standard deviation of the corresponding structural size data. The material ratio update rate is the Euclidean distance of the material ratio components of the two target center points in the normalized space, which represents the degree of divergence of the optimal trends that trigger the conflict in the material ratio direction. The structural size modification magnitude is the normalized value of each component of the structural size of the samples within the clusters where the two target center points are located. The maximum value of the standard deviation represents the range of structural size dispersion explored within the optimal trend neighborhood, with the amplitude value being a proportional quantity with a dimension of 1. Candidate adjustment variables are calculated decoupled by component groups: the material ratio candidate adjustment component is the material ratio update rate, and the structural size candidate adjustment component is the structural size modification amplitude value. The two component groups are based on the statistics of their corresponding component group directions and are not multiplied together. The deviation of the carbon emission prediction mean of each node relative to the mean of all nodes is normalized to the correction weight. The material ratio candidate adjustment component is multiplied by the correction weight to obtain the material ratio component of the component parameter target adjustment variable, and the structural size candidate adjustment component is multiplied by the correction weight to obtain the structural size component of the component parameter target adjustment variable. The component parameter target adjustment variable is composed of the material ratio component and the structural size component.

[0031] In this embodiment, the material ratio update rate is taken as the Euclidean distance between the material ratio components of the two target center points in the normalized space. The larger the distance, the greater the divergence of the optimal trends that trigger the conflict in the material ratio direction, and the more the material ratio components need to be adjusted to converge in a consistent direction. The structural size modification amplitude is taken as the maximum value of the normalized standard deviation of each component of the structural size of the samples within the cluster where the two target center points are located. The larger the standard deviation, the wider the range of structural size dispersion explored in the neighborhood of the optimal trend. Updating the structural size components within a step size not exceeding this dispersion range can avoid jumping out of the verified region in one step. The resulting amplitude value is a proportional quantity with a dimension of 1. The step size of the material ratio component is based on the material ratio update rate, and a preset upper limit for the ratio step size is set. The preset upper limit for the ratio step size is set according to the normalized proportional quantity of the allowable fluctuation range of the material ratio, which is 0.05. The step size of the structural size component is based on the structural size modification amplitude value and the upper limit is the normalized proportional quantity converted by the structural size modification amplitude level. When the benchmark exceeds the upper limit, it is truncated to the upper limit. The product measurement calibrates in conjunction with the conflict threshold. Material ratio updates are measured per iteration cycle, and structural dimension modifications are expressed as percentages (e.g., 5% is represented as 5). The product is obtained by directly multiplying the frequency and amplitude values. A conflict threshold of three values ​​is assigned to this threshold, ranging from 1 to 5: the product for high-frequency, high-amplitude categories is 2 × 5 = 10, exceeding the threshold and triggering conflict resolution; the products for high-frequency, low-amplitude categories and low-frequency, high-amplitude categories are 2 × 1 = 2 and 0.5 × 5 = 2.5 respectively, while the product for low-frequency, low-amplitude categories is 0.5 × 1 = 0.5, all within the threshold, and the previous round's component parameter target adjustment variables are used. In the first iteration, there are no variables from the previous round, and the component parameter target adjustment variables are set to 0, meaning each node maintains its initial control parameters. Variables from the previous round are only available from the second iteration onwards.

[0032] The clustering objects are the top five individuals in the current population with the lowest modified fitness, representing the nodes that triggered the conflict. These are the design parameter vectors corresponding to the five individuals with the lowest modified fitness. The dimension of these vectors is consistent with the green design parameters, representing the actual dimension after each component and its sub-components are expanded. The top five individuals in fitness represent the optimal trend of the current search direction for that node, and they are used as the cluster centers, representing the direction of parameter adjustment. If all individuals in the current population are used, a large number of redundant samples with mediocre fitness will flatten the differences between clusters, dilute the cluster boundaries, and cause the cluster centers to deviate from the optimal trend, resulting in distortion. Clustering uses the K-means algorithm: the number of clusters is incremented sequentially from 2. When the number of clusters is 1, the distance matrix between the center points has no off-diagonal elements, and the two target center points cannot be obtained. Therefore, the lower limit of the number of clusters is set to 2. The optimal trend without divergence is identified by the distance threshold criterion. The criterion for stopping the increase is that the decrease in the sum of squares within the clusters as the number of clusters increases is less than a preset ratio of 10%. When the decrease in the number of clusters increases from k to k+1, k is taken as the number of clusters in this clustering. If the criterion for stopping the increase is still not met when the calculation reaches the upper limit, the upper limit is taken. The upper limit of the number of clusters is the smaller value of the number of nodes that trigger conflict and 5, and is not less than 2. When the number of nodes that trigger conflict is 1, 2 is taken. The distance threshold is executed as 50% of the diagonal length of the aforementioned normalized feature space. When the maximum distance value does not exceed the preset distance threshold, it is determined that the optimal trend is without divergence. The candidate adjustment variables are not recalculated, and the target adjustment variables of the component parameters in the previous round are used. When there are no variables from the previous round in the first round, 0 is taken.

[0033] The Euclidean distance between centers is calculated in the feature space after normalizing each dimension of the cluster features to the 0-1 range. The dimension of the cluster features is determined by the actual dimension after expanding each component and subcomponent of the design parameter vector. The normalization method is linear scaling: for each dimension, the maximum and minimum values ​​of the triggering conflict samples in that dimension are taken, and the values ​​are mapped to 0-1 using minimum-maximum scaling. After normalization, each dimension contributes to the distance in the same order of magnitude. If calculated directly using the original dimensions, the outer diameter of the umbrella skirt can reach hundreds of millimeters, and the difference contribution will overwhelm the contribution of the material ratio, which is on the order of 0-1. The actual distance between centers degenerates into the distance of the single dimension of the umbrella skirt outer diameter, and the distance threshold cannot be uniformly set in the feature space of the actual dimension after expansion. The Euclidean distance between each cluster center is calculated pairwise, forming a symmetric distance matrix with the number of cluster centers as the order. The diagonal elements are 0, and the off-diagonal elements are the distances between the corresponding two centers. The maximum distance value is obtained by traversing this distance matrix and taking the largest off-diagonal element. The distance threshold is set to 50% of the diagonal length of the normalized feature space, and the value is pre-configured according to the actual dimension.

[0034] The correction weights are normalized by dividing the absolute value of the deviation by the sum of all absolute values ​​of deviation. First, the absolute value of the deviation of the carbon emission prediction mean of each node relative to the overall carbon emission prediction mean of all nodes is calculated. Then, the absolute value of the deviation of a single node is divided by the sum of the absolute values ​​of deviation of all nodes. When this sum is 0, the correction weight for each node is the reciprocal of the number of nodes. The resulting correction weights are between 0 and 1, and the sum of all node deviations is 1. This normalization method ensures that the total weight sum remains constant and is independent of the number of nodes. Therefore, the magnitude of the target adjustment variable for the component parameters obtained by multiplying the candidate adjustment variable by the correction weight is controllable. If normalization is performed using the maximum value, the sum of the weights varies with the number of nodes; for the same degree of separation, the more nodes there are, the larger the sum of the weights, and the magnitude of the target adjustment variable obtained by multiplication becomes uncontrollable. The further the carbon emission deviation of a node is from the global mean, the more advanced or lagging the optimization process, resulting in a larger correction weight and a larger adjustment amount, which is positively correlated with the degree of deviation.

[0035] Step S5: When the model uncertainty component in the carbon emission prediction uncertainty exceeds the preset uncertainty threshold, call the life cycle stage decomposition calculation to perform mechanism calculation on carbon emissions and backfeed to update the carbon emission proxy model.

[0036] Furthermore, the carbon emission proxy model is updated by invoking the life cycle stage-based decomposition and accounting to perform mechanistic accounting of carbon emissions and re-feeding back the model, including: Select target candidate parameters with the largest carbon emission prediction uncertainty and a preset number from the candidate design parameters in the current iteration; perform mechanism verification on the carbon emission of each target candidate parameter; add the target candidate parameters and their corresponding mechanism verification values ​​to the training set, and incrementally update the carbon emission surrogate model until the model uncertainty component in the global carbon emission prediction uncertainty is lower than the preset convergence threshold.

[0037] Specifically, the triggering and convergence criteria take the model uncertainty component from the carbon emission prediction uncertainty. In the case of Gaussian process regression, the model uncertainty component is taken as the square root of the posterior variance, i.e., the posterior standard deviation, which is determined by the kernel function and the location of the training samples and shrinks as the number of training samples increases. In the case of random forest, the inter-tree standard deviation of the predicted values ​​of each tree is taken, and the denser the training samples, the smaller the inter-tree divergence. The model uncertainty component and the component obtained by factor uncertainty propagation have different uses: the propagated component characterizes the impact of the fluctuation of the carbon emission factor itself on carbon emission prediction, and its magnitude is determined by the width of the uncertainty interval of the factor library. Refeeding and sample addition can only correct the deviation of the surrogate model relative to the mechanism calculation, and cannot compress the factor interval width; if the triggering and convergence criteria mistakenly use the uncertainty containing the propagated component, the convergence condition will never be met during refeeding. The uncertainty threshold is set relative to the carbon emission prediction mean, taking 10% of the mean. That is, when the ratio of the model uncertainty component to the carbon emission prediction mean is greater than 0.1, the backfeed is triggered. The convergence criterion is also taken from the model uncertainty component in the carbon emission prediction uncertainty, and the convergence threshold is taken as 5% of the average value of the carbon emission prediction mean of all sampling points.

[0038] The selection pool for target candidate parameters comprises all individuals from each distributed node population in the current iteration. After merging, they are uniformly sorted according to the model uncertainty components from largest to smallest, and the top five are selected as target candidate parameters. Five is the preset number, ranging from 3 to 8. The convergence criterion is global. Only refeeding back individuals belonging to the trigger node cannot reduce the uncertainty of other subspaces. The effect of refeeding on convergence will become ineffective as the number of nodes increases. Therefore, the selection pool must be merged across the entire domain. Each target candidate parameter is calculated by decomposing it according to its life cycle stage. The algebraic sum of the three-stage formulas for raw material acquisition, manufacturing, and scrapping / recycling is used to obtain the mechanistic calculation value of the parameter.

[0039] Incremental updates are implemented according to the surrogate model type. Gaussian process regression is updated by expanding the kernel matrix in blocks: the original kernel matrix remains unchanged, only the kernel function values ​​between the new samples and the original samples are calculated to expand the matrix rows and columns. The trigonometric equations are solved through incremental updates using the Cholesky factor, ensuring that hyperparameters remain unchanged during the incremental process, avoiding full retraining. Random forest is updated by expanding the tree pool: in each round, self-sampling is performed on the entire set after merging the original training set and the new samples. An equal number of trees as the target candidate parameters are added to the tree pool. When the tree pool size exceeds the upper limit of 800 trees, the tree with the worst out-of-bag error is eliminated. Out-of-bag error refers to the prediction error on samples not selected by self-sampling. If the incremental method is not limited and full retraining is performed in each round, the refeeding time for a scale of 500 trees is close to the retraining cost, losing the low-cost advantage of active learning. After the incremental update, the output of the surrogate model is still the carbon emission prediction mean and model uncertainty components, used for optimization and triggering criteria in the next iteration cycle. When the training samples exceed the preset sample threshold due to the accumulation of backfilling, switch to random forest and retrain with all training samples; tree pool elimination takes effect when the size reaches the upper limit; in the first round, i.e., when there is a conflict, only the conflict judgment and cluster analysis are completed in that round, and the component parameter target adjustment variable is still 0. From the next round onwards, the newly calculated variable will be used.

[0040] The global model uncertainty statistic is calculated using a sampling approximation: 500 sampling points are generated within the search space using Latin hypercube sampling, and each point is input into the updated carbon emission surrogate model. The 95th percentile of the model uncertainty component at each sampling point is taken as the global model uncertainty statistic. The 95th percentile is a preset high quantile value; using the high quantile instead of the mean is to avoid a few low-uncertainty regions lowering the statistic and masking high-risk regions. The backfeeding ends when the global model uncertainty statistic falls below the convergence threshold, and synchronous updates begin; if it still exceeds the threshold, the screening, calculation, and incremental updates are repeated. The maximum number of backfeeding attempts in a single optimization process is 10. If convergence is not achieved after reaching the upper limit, optimization continues using the current carbon emission surrogate model, and the uncertainty is marked as not converged in the carbon assessment results, serving as a second exit insurance in addition to limiting the model uncertainty components as a trigger criterion.

[0041] Step S6: Update the green design parameters of each node with the component parameter target adjustment variables to complete the synchronization and output the carbon assessment results.

[0042] Furthermore, the output carbon assessment results include: Based on the candidate design parameters retained at the end of the iteration, a Pareto front is plotted between the carbon emission prediction mean and the feasibility probability; the sensitivity of each green design parameter to the carbon emission prediction mean is calculated, and the parameter contribution ranking is generated; based on the Pareto front, the parameter contribution ranking, and the carbon emission prediction mean, carbon emission prediction uncertainty, and feasibility probability of the candidate design parameters with the best robust carbon index, a carbon assessment result report is generated.

[0043] In this implementation, synchronous updates are performed in the normalized parameter space. The update direction is the normalized unit direction of the vector pointing from the current best individual of the node to the globally best individual recorded after synchronization in the previous iteration. If the two coincide, the direction update is skipped in this iteration. The step size is the target adjustment variable of the component parameters of the node. The updated parameter is equal to the parameter of the current best individual plus the product of the unit direction and the step size. The operation is reverse-mapped back to the physical value after the normalized space is completed. The step size is a dimensionless proportional quantity. Directly adding it to dimensional parameters would introduce dimensional confusion, so it must be normalized before the operation. After the update, the parameter value range is checked component by component. Parameters exceeding the value range of the component are truncated to the interval boundary to prevent synchronous updates from violating the search space constraints. The truncated node is reintroduced with individuals inside the interval in the initial population of the next iteration to avoid the population stagnation at the edge. Nodes whose frequency levels are adjusted submit updates according to the submission rhythm corresponding to the adjusted frequency level, with 0.5 times corresponding to once every two cycles. The synchronization update step size is based on the clustering amplitude value, and the amplitude level is converted into a normalized proportional value and then constrained by an upper limit. After each node completes its update, it broadcasts the updated parameters along with the iteration cycle number to all nodes. Synchronization is considered complete when all nodes reach a consensus on the parameter version recorded for the same iteration cycle number. Nodes with inconsistent versions replay the update based on the latest cycle number version to ensure that the parameter version used for matrix statistics and conflict judgment in subsequent iteration cycles is consistent.

[0044] Iteration termination is triggered by two criteria: a maximum of 200 iterations, ranging from 100 to 500; or, if the improvement in the global optimal robust carbon index is less than one-thousandth for twenty consecutive iterations, stagnation is determined, whichever comes first. The retained candidate design parameters refer to the union of the last generation of individuals in each distributed node at termination and the historical elite archive throughout the optimization process. The historical elite archive records the current best individual for each node generation by generation, and the union, after deduplication, forms the input pool for drawing the Pareto front. Relying solely on the last generation of individuals may miss excellent solutions that appear midway through the iteration but are replaced by new individuals; the elite archive compensates for this omission.

[0045] Sensitivity is calculated using the first-order exponent of variance decomposition. Each green design parameter is first normalized to the 0-1 range. One thousand samples are generated in the normalized search space using Latin hypercube sampling. These samples are then input into the carbon emission surrogate model to obtain one thousand carbon emission prediction mean values. The first-order exponent is the ratio of the conditional expected variance of that dimension to the total variance. Since all exponents are estimated using the same set of samples, the caliber is consistent, and the ranking results are comparable. Each component of the design parameter vector and each sub-component of the material mix ratio is assigned an exponent, and the parameters are ranked from largest to smallest based on their contribution.

[0046] The optimal candidate design parameters for robust carbon index are selected by applying a coefficient of 1.28 corresponding to a 90% confidence level requirement: for each individual in the input pool, the minimum sum of the carbon emission prediction mean, the product of 1.28, and the carbon emission prediction uncertainty is taken, where the carbon emission prediction uncertainty is the sum of the model uncertainty component and the factor propagation component, consistent with the robust carbon index. This optimal individual is used to perform electrical performance proxy model verification. If the feasibility probability is not lower than the probability threshold, it is designated as the recommended parameter. The probability threshold is the same as the preset probability threshold used in the penalty function, ranging from 90% to 98%. If the condition is not met, the next best individual is selected for verification along the Pareto front towards the side with higher feasibility probability, until a qualified individual is found. The penalty function only guarantees the search tendency rather than achieving the target point by point. Back-substitution verification is the final checkpoint for the electrical qualification of the recommended parameters. The carbon assessment report includes a Pareto frontier data table, parameter contribution ranking, and the values ​​of four components among the recommended parameters: skirt outer diameter, creepage distance, core rod size, and material ratio. The material ratio component specifies the ratio of each sub-component of matrix, filler, and additives, as well as their carbon emission prediction mean, carbon emission prediction uncertainty, and feasibility probability. If the active learning reaches the reinjection limit but still fails to converge, it also indicates that the uncertainty has not converged. Additionally, a snapshot of the power grid emission factors in the power supply area is attached for traceability of the results.

[0047] The numerical implementation example uses composite insulators for 110kV AC lines as the evaluation object. The design parameter vector is taken as follows: shed outer diameter 120mm to 180mm, creepage distance 2400mm to 3600mm, core rod diameter 24mm to 32mm, and material composition components within allowable ranges. 1000 training samples are generated using Latin hypercube sampling. Example of factor library hit values: silicone rubber material factor 3.2kgCO2e / kg, aluminum hydroxide filler 0.6kgCO2e / kg, silica functional additive factor 3. The carbon emission factors were calculated as follows: 0.0 kg CO2e / kg, FRP core rod factor 4.5 kg CO2e / kg, primary metal accessory factor 2.1 kg CO2e / kg (regenerated factor 0.5 kg CO2e / kg), manufacturing carbon emission factor 0.07 kg CO2e / MJ, and grid emission factor 0.55 kg CO2e / kWh, with an uncertainty range of ±0.04. The calculation coefficients were taken as follows: unit mass heat consumption coefficient 4.0 MJ / kg, process time 2.5 h, equipment rated power 12 kW, and metal accessory recovery rate 90%. Initial carbon assessment values ​​for each sample were obtained through three-stage calculations, and a surrogate model was trained. Eight distributed nodes were used, with a state threshold of 0.2, a conflict threshold of 3, an uncertainty threshold of 10% of the carbon emission prediction mean, a convergence threshold of 5%, a probability threshold of 95%, and a risk aversion coefficient of 1.28. The initial design uses an outer diameter of 150mm for the umbrella skirt, a creepage distance of 3150mm, a core rod diameter of 30mm, and a material ratio of 60:30:10 by weight. This translates to a base weight of 2.88kg, filler weight of 1.44kg, additive weight of 0.48kg, core rod weight of 1.98kg, and metal accessories weight of 1.80kg, resulting in a net product weight of 8.58kg. The recommended parameters are an outer diameter of 138mm for the umbrella skirt, a creepage distance of 3150mm, a core rod diameter of 28mm, and the same material ratio. This translates to a base weight of 2.34kg, filler weight of 1.17kg, additive weight of 0.39kg, core rod weight of 1.72kg, and accessories weight of 1.80kg, resulting in a net product weight of 7.42kg.Three-stage accounting: Initial design carbon emissions during raw material acquisition = 2.88 × 3.2 + 1.44 × 0.6 + 0.48 × 3.0 + 1.98 × 4.5 = 9.22 + 0.86 + 1.44 + 8.91 = 20.43 kg CO2e; Carbon emissions during manufacturing = 8.58 × 4.0 × 0.07 + 2.5 × 12 × 0.55 = 2.40 + 16.50 = 18.90 kg CO2e; Carbon offset from waste recovery = 1.80 × 90% × (2.1 − 0.5) = 2.59 kg CO2e; Initial carbon assessment value = 20.43 + 18.90 − 2.59 = 3 6.74 kg CO2e; The recommended parameters for carbon emissions during the raw material acquisition stage are 2.34 × 3.2 + 1.17 × 0.6 + 0.39 × 3.0 + 1.72 × 4.5 = 7.49 + 0.70 + 1.17 + 7.74 = 17.10 kg CO2e, and the carbon emissions during the manufacturing stage are 7.42 × 4.0 × 0.07 + 30 × 0.55 = 2.08 + 16.50 = 18.58 kg CO2e. The carbon credit is the same as before, 2.59 kg CO2e. The carbon assessment value is 17.10 + 18.58 - 2.59 = 33.09 kg CO2e, which is a decrease of 9.95% compared to the initial design. The iteration reached 132 generations, triggering the stagnation criterion and terminating. The global best robust carbon index improvement was less than 0.1% for 20 consecutive generations, and the Pareto front contained 32 non-dominated solutions. The recommended parameter creepage distance = 3150 ÷ ​​126 = 25.0 mm / kV is not lower than the lower limit of the rated value. The feasibility probability of the pollution flashover withstand voltage is 95.2%, and the feasibility probability is not lower than 95%. The parameters with the largest contribution are the umbrella skirt outer diameter and the core rod diameter, with first-order sensitivity indices of 0.29 and 0.26, respectively, which is consistent with the recommended path of carbon reduction by size reduction with unchanged proportions.

[0048] The above describes a method for intelligent optimization of carbon assessment of green design parameters for insulators in embodiments of this application. The following describes a platform for intelligent optimization of carbon assessment of green design parameters for insulators in embodiments of this application. Please refer to [link / reference]. Figure 3 One embodiment of the intelligent optimization carbon assessment platform for green design parameters of insulators in this application includes: The parameter acquisition and carbon accounting module is used to obtain green design parameters for insulators, query the carbon emission factor library, and calculate the initial carbon assessment value by decomposing it according to the life cycle stage.

[0049] The surrogate model training module is used to train carbon emission surrogate models and electrical performance surrogate models, and outputs the carbon emission prediction mean, carbon emission prediction uncertainty, and feasibility probability.

[0050] The distributed optimization module is used to divide the search space into distributed nodes for iterative optimization, construct an evolutionary state matrix, and extract the material ratio update frequency level and structural size modification level through a classification model.

[0051] The conflict resolution module is used to cluster candidate design parameter features when the product of gears exceeds the conflict threshold, adjust the gears, and obtain the target adjustment variable of component parameters with corrected weights.

[0052] The feedback update module is used to perform mechanism verification of the target candidate parameters and incrementally update the carbon emission proxy model when the model uncertainty components exceed the threshold.

[0053] The synchronization and evaluation output module is used to synchronize the green design parameters of each node and output the carbon assessment results.

[0054] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the platform and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent optimization of carbon assessment of green design parameters for insulators, characterized in that, The method includes: Obtain green design parameters for insulators, query the carbon emission factor library, and calculate carbon emissions by life cycle stage to obtain an initial carbon assessment value. Using green design parameters as input and initial carbon assessment values ​​and electrical performance indicators as labels, we train carbon emission proxy models and electrical performance proxy models, and output the carbon emission prediction mean, carbon emission prediction uncertainty and feasibility probability. The search space for green design parameters is divided into distributed nodes. The robust carbon index is used as the fitness and the feasibility probability is used as the constraint to iteratively optimize and construct an evolutionary state matrix. When the elements of the evolutionary state matrix exceed the preset state threshold, the update frequency and update magnitude of the node data are obtained, input into the pre-trained classification model, and the update frequency level of the material ratio and the modification magnitude level of the structural size are extracted. If they do not exceed the threshold, the iteration continues. When the product of the material ratio update frequency level and the structural dimension modification range level exceeds the preset conflict threshold, the clustering of candidate design parameter features yields the component feature target cluster. The material ratio update frequency level and the structural dimension modification range level are adjusted according to the cluster center, and the component parameter target adjustment variable is obtained. If it does not exceed the threshold, the variable from the previous round is used. When the model uncertainty component in the carbon emission prediction uncertainty exceeds the preset uncertainty threshold, the carbon emission mechanistic calculation is performed by decomposing the carbon emission by life cycle stage and the carbon emission proxy model is updated. The green design parameters of each node are updated by adjusting the component parameter target variables to complete the synchronization and output the carbon assessment results.

2. The method according to claim 1, characterized in that, The green design parameters include at least the outer diameter of the umbrella skirt, creepage distance, core rod size, and material ratio. The carbon emission factor database includes: Query the carbon emission factor database organized by material type, manufacturing process, and power supply area to obtain material carbon emission factors, manufacturing carbon emission factors, and power grid emission factors; The material carbon emission factor, the manufacturing carbon emission factor, and the power grid emission factor all have an uncertainty range. The power grid emission factor is dynamically updated based on the time-series power grid data of the power supply area.

3. The method according to claim 2, characterized in that, The initial carbon assessment value is obtained by breaking down and calculating carbon emissions according to life cycle stages, including: The life cycle stage includes at least the raw material acquisition stage, the manufacturing stage, and the end-of-life recycling stage; the carbon emissions of the raw material acquisition stage are obtained by summing the products of the amount of each component material and the corresponding material carbon emission factor. The process fuel heat consumption and power consumption are calculated from the outer diameter of the umbrella skirt, the creepage distance, and the size of the mandrel, respectively. The carbon emissions during the manufacturing stage are obtained by multiplying the process fuel heat consumption by the manufacturing carbon emission factor and adding the power consumption by the grid emission factor. The carbon offset during the scrapping and recycling stage is calculated based on the recovery rate of recyclable components and the recycling offset factor. The algebraic sum of the carbon emissions during the raw material acquisition stage, the carbon emissions during the manufacturing stage, and the carbon credit is determined as the initial carbon assessment value.

4. The method according to claim 1, characterized in that, Training carbon emission proxy models and electrical performance proxy models includes: Training samples are generated by Latin hypercube sampling within the allowable range of each component of the green design parameters; among the electrical performance indicators, the creepage distance label is obtained by analytical calculation of the green design parameters, and the pollution flashover withstand voltage label is obtained by experimental data or electric field simulation. When the number of training samples does not exceed the preset sample threshold, the carbon emission surrogate model is trained using a Gaussian process regression model, and the model uncertainty component in the carbon emission prediction uncertainty is determined by the square root of the posterior variance of the model output. When the number of training samples exceeds the preset sample threshold, the carbon emission surrogate model is trained using a random forest regression model, and the model uncertainty component in the carbon emission prediction uncertainty is determined by the sample standard deviation between the outputs of each regression tree. The electrical performance surrogate model is trained using electrical performance indicators as labels, and the feasibility probability is calculated based on the predicted mean and predicted variance of the model output.

5. The method according to claim 4, characterized in that, Iterative optimization with feasibility probability as a constraint includes: When the feasibility probability is less than a preset probability threshold, a penalty function is applied to the fitness. The penalty term of the penalty function is positively correlated with the degree to which the feasibility probability deviates from the probability threshold. The fitness corrected after applying the penalty function is equal to the sum of the robust carbon index and the penalty term, so that candidate design parameters with insufficient electrical performance are reduced in weight and eliminated in the iteration. The feasibility probability of the candidate design parameter is obtained by multiplying the feasibility probabilities of each electrical performance index. The electrical performance index includes at least creepage distance and flashover withstand voltage.

6. The method according to claim 1, characterized in that, The robust carbon index is equal to the carbon emission prediction mean plus the product of the risk aversion coefficient and the carbon emission prediction uncertainty, specifically: The carbon emission prediction uncertainty is the sum of the model uncertainty component and the factor propagation component. The model uncertainty component is determined by the square root of the posterior variance of the model output when using a Gaussian process regression model, and by the sample standard deviation between the outputs of each regression tree when using a random forest regression model. The factor propagation component is obtained by the carbon emission factor with the attached uncertainty interval through analytical propagation of the first-order sensitivity of the accounting formula. The risk aversion coefficient is set according to the confidence level requirements of the carbon assessment.

7. The method according to claim 1, characterized in that, Obtain the target adjustment variables for component parameters, including: Extract the cluster centers of the target clusters of component features and calculate the Euclidean distance between the centers to obtain the center point distance matrix; when the maximum distance value exceeds the preset distance threshold, take the two corresponding center points as the target center points; The material ratio update rate is calculated based on the material ratio data of the target center point, and the structural dimension modification range is obtained based on the standard deviation of the corresponding structural dimension data. The candidate adjustment variables are calculated in decoupled form by component group. The material ratio candidate adjustment component is the material ratio update rate, and the structural dimension candidate adjustment component is the structural dimension modification range. The deviation of the carbon emission prediction mean of each node relative to the mean of all nodes is normalized into the correction weight. The candidate adjustment components for material ratio and structural dimensions are multiplied by the correction weights to determine the material ratio and structural dimension components of the target adjustment variables for component parameters.

8. The method according to claim 1, characterized in that, The process involves decomposing and calculating carbon emissions by lifecycle stage, performing mechanistic accounting, and then updating the carbon emission proxy model, including: Select target candidate parameters with the largest carbon emission prediction uncertainty and a preset number from the candidate design parameters in the current iteration; perform mechanism verification on the carbon emission of each target candidate parameter; The target candidate parameters and their corresponding mechanism calculation values ​​are added to the training set, and the carbon emission surrogate model is updated incrementally until the model uncertainty component in the global carbon emission prediction uncertainty is lower than the preset convergence threshold.

9. The method according to claim 1, characterized in that, Output carbon assessment results, including: Based on the candidate design parameters retained at the end of the iteration, plot the Pareto front between the carbon emission prediction mean and the feasibility probability; calculate the sensitivity of each green design parameter to the carbon emission prediction mean, and generate a parameter contribution ranking. A carbon assessment report is generated based on the Pareto frontier, parameter contribution ranking, and the mean carbon emission prediction, carbon emission prediction uncertainty, and feasibility probability of the candidate design parameters with the best robust carbon index.

10. A smart carbon assessment platform for green design parameters of insulators, used to implement the smart carbon assessment method for green design parameters of insulators as described in any one of claims 1-9, characterized in that, The platform includes: The parameter acquisition and carbon accounting module is used to acquire green design parameters for insulators, query the carbon emission factor library, and calculate carbon emissions according to life cycle stages to obtain an initial carbon assessment value. The proxy model training module is used to train carbon emission proxy models and electrical performance proxy models with green design parameters as input, initial carbon assessment values ​​and electrical performance indicators as labels, and outputs the carbon emission prediction mean, carbon emission prediction uncertainty and feasibility probability. The distributed optimization module is used to divide the search space of green design parameters into distributed nodes, and iteratively optimizes the parameters with robust carbon index as fitness and feasibility probability as constraints to construct an evolutionary state matrix. When the elements of the evolutionary state matrix exceed the preset state threshold, the update frequency and update magnitude of the node data are obtained, input into the pre-trained classification model, and the update frequency level of material ratio and the modification magnitude level of structural size are extracted. If they do not exceed the threshold, the iteration continues. The conflict resolution module is used to cluster candidate design parameter features to obtain component feature target clusters when the product of the material ratio update frequency level and the structural size modification range level exceeds the preset conflict threshold. The module adjusts the material ratio update frequency level and the structural size modification range level according to the cluster center and obtains the component parameter target adjustment variable. If the conflict threshold is not exceeded, the variable from the previous round is used. The feedback update module is used to call the life cycle stage decomposition accounting to perform mechanism accounting on carbon emissions and feedback update the carbon emission proxy model when the model uncertainty component in the carbon emission prediction uncertainty exceeds the preset uncertainty threshold. The synchronization and evaluation output module is used to update the green design parameters of each node with the component parameter target adjustment variables to complete the synchronization and output the carbon assessment results.