New energy charging station asset credibility evaluation optimization method based on machine learning

Abstract

The invention discloses a new energy charging station asset credibility evaluation optimization method based on machine learning, and the method comprises the following steps: S1, collecting multi-source asset data, accessing an IOT platform, and generating a credible data chain; s2, calculating a confidence score, and constructing a graph structure asset network; s3, constructing a graph attention network model, and generating asset node embedding representation; s4, applying a wild horse group optimization algorithm to jointly optimize model parameters; and S5, training by using the optimized model, and outputting an asset evaluation result. S6, generating an abstract and performing hash encryption, binding a data chain and writing the data chain into an ant chain; and S7, continuously collecting and comparing new data, and triggering fine adjustment and updating of the model. According to the invention, accurate assessment and credible management of assets in a new energy charging station are realized, and the intelligence of asset operation state sensing and the credibility of data results are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and particularly relates to a new energy charging station asset credible evaluation optimization method based on machine learning. BACKGROUND

[0002] With the rapid development of the new energy industry, charging stations, as an important part of new energy infrastructure, play a key role in electric vehicle energy supply, energy scheduling and urban energy network construction. New energy charging stations are usually composed of multiple types, multiple manufacturers and multiple batches of power asset equipment, such as charging piles, battery packs, distribution boxes, communication modules and cooling devices. These assets have problems such as heterogeneous running state, non-uniform data format and non-fixed maintenance period in the long-term operation process, which brings great challenges to the operation and maintenance management and value assessment of charging stations.

[0003] In the prior art, the evaluation of new energy charging station assets mostly relies on static data statistical analysis methods or scoring systems based on experience rules. For example, a rule base is established by analyzing asset running time, usage frequency, historical fault records and other indicators, and then asset health scoring is performed according to preset weights. Although this method has certain reference significance in early system construction, it lacks comprehensive modeling capability for multi-dimensional factors such as asset coordination relationship, real-time state dynamic change and data credibility, resulting in inaccurate evaluation results and inability to adapt to the current multi-source heterogeneous and complex state asset operation environment.

[0004] With the development of machine learning technology, some research has begun to introduce deep learning models to predict and evaluate asset operation state, especially in handling time series data. However, most methods rely on traditional sequence modeling methods or fully connected neural network architectures, which are difficult to effectively express the topology and collaborative behavior between asset devices. In recent years, graph neural networks (GNN) have shown excellent structure modeling capability in the fields of transportation, finance and recommendation systems, and have gradually been concerned in the field of new energy charging, but there are still problems such as insufficient generalization ability, missing edge weight features and weak credible data integration.

[0005] Based on the development of graph neural networks, graph attention network (GAT) is a modeling method based on dynamic weighting of node relationships. By introducing an attention mechanism, the importance of adjacent nodes is automatically learned, and compared with traditional graph convolution methods, it has stronger structural adaptability and interpretability. However, existing research still has the following problems in the application of GAT to charging asset evaluation: first, the collaborative relationship between nodes is only represented by the adjacency matrix, lacking the ability to model edge features such as asset call frequency and shared lines, and unable to reflect the true dependence strength between assets; second, no trusted mechanism is introduced in the node feature fusion process, and the quality differences between different data sources are not effectively utilized, affecting the stability of model prediction; third, the attention mechanism may over-aggregate or overfit for high-dimensional graph structures, lacking sparse regulation means, increasing model redundancy and computational complexity.

[0006] At the level of optimization algorithm, the selection of graph model structure parameters and training hyperparameters has a significant impact on model performance. Traditional methods such as grid search and Bayesian optimization are difficult to balance search efficiency and global optimality in large-scale parameter space. Some methods try to introduce swarm intelligence optimization mechanisms such as particle swarm and genetic algorithm to adjust hyperparameters, but these methods are often limited to global exploration stage, lack robust local jumping mechanism and credible guidance strategy, and are prone to fall into local optimum.

[0007] In addition, current asset evaluation results are mostly stored in databases or local caches at the data management level, lacking a unified trusted mechanism, and there are problems such as tamperable evaluation data, difficult to trace, and unverifiable, especially when third-party auditing or on-chain value measurement of asset evaluation conclusions is required. Some existing research has begun to explore the combination of blockchain and asset data evidence, but lacks standardization and integrity support in terms of data structure organization, summary generation method, and model process parameter mapping on the chain, resulting in fragmented trusted chain information and lack of practical deployment feasibility.

[0008] Therefore, how to provide a new energy charging station asset trusted evaluation optimization method based on machine learning is a problem that needs to be solved by those skilled in the art. SUMMARY

[0009] One purpose of the present invention is to propose a trusted evaluation optimization method for new energy charging station assets based on machine learning. The present invention fully integrates graph attention network modeling, wild horse swarm optimization algorithm parameter adjustment mechanism, trusted data chain construction and blockchain evidence storage technology, and systematically realizes the whole process of trusted modeling, feature aggregation, state prediction and result on-chain of multi-source heterogeneous assets in new energy charging stations. By introducing optimizable structural parameters such as edge weight guidance coefficient, node credibility fusion weight and attention sparsity rate threshold, a trusted-driven graph attention network model structure is constructed; the structural parameters and training hyperparameters of the model are jointly optimized by the improved wild horse swarm optimization algorithm; and the tamper-proof on-chain evidence of asset evaluation is realized by the trusted data chain and the asset node evaluation result summary hash binding mechanism. This method has the advantages of high credibility of evaluation results, strong model generalization ability, high optimization efficiency and good data traceability.

[0010] According to an embodiment of the present invention, a method for optimizing trustworthy assessment of new energy charging station assets based on machine learning includes the following steps:

[0011] S1. Collect multi-source asset data of assets and equipment in new energy charging stations, connect the multi-source asset data to the Ant IOT platform through the DTU module, process it using the trusted SDK, and generate a trusted data chain;

[0012] S2. Based on the trusted data chain, calculate the confidence score and integrate it with the asset status, operation and maintenance records, and location information to form node features. Use assets as nodes and construct edges based on collaborative behaviors to form a graph-structured asset network and adjacency relationship representation.

[0013] S3. Based on the graph structure asset network and adjacency relationship representation, a graph attention network model is constructed to generate an embedded representation of the asset node;

[0014] S4. Apply the wild horse swarm optimization algorithm to jointly optimize the structural parameters and training hyperparameters of the graph attention network model;

[0015] S5. Use the optimized graph attention network model for training, aggregate and predict the features of asset nodes based on the trusted data chain, and output the asset evaluation results;

[0016] S6. Generate a summary of the node features, embedded representation of the asset node, asset assessment results, and parameter configuration. After hash encryption, it is associated and bound to the trusted data chain of the corresponding multi-source asset data and written into the Ant Chain to form a trusted chain certificate.

[0017] S7. Continuously collect new asset data and build a trusted data chain, compare it with the evaluation results of assets already on the chain, and automatically trigger the Wild Horse Swarm Optimization algorithm to fine-tune the parameters of the graph attention network model when the deviation between the prediction and feedback exceeds the set threshold.

[0018] Optionally, the multi-source asset data of the asset device specifically includes operating status, operation and maintenance records, geographic location and device identification information, which are used to construct node features and support asset evaluation modeling of the graph attention network.

[0019] Optionally, the processing by using the trusted SDK specifically includes encryption signature, hash digest and timestamp annotation on the multi-source asset data, which are used to generate a trusted data chain of the multi-source asset data that is traceable and verifiable.

[0020] Optionally, the S2 specifically includes:

[0021] S21, extracting, from the generated trusted data chain, a data credibility index C of each asset device at a time point t t , a data integrity ratio I t , a number of successful uploads N t , a maximum number of historical uploads N max , and an asset evaluation prediction error E t ;

[0022] S22, calculating a confidence score D of the asset node t :

[0023]

[0024] wherein α, β, γ, δ are confidence weighting coefficients;

[0025] S23, collecting an operating state vector, an operation and maintenance record vector and a location information vector of each asset, and splicing the calculated confidence score D t into a node feature vector X t ;

[0026] S24, constructing an edge set E according to asset collaborative behavior data, defining an edge e ij = 1 if the number of collaborative uses F ij of the asset device i and the asset device j within a given time window is greater than or equal to θ, and 0 otherwise, wherein θ is a set collaborative threshold;

[0027] S25, constructing a graph structure asset network G = (V, E) based on the node set V = {v1, v2, …, vn} n} and the edge set E, taking each asset device as a node in the graph, vn n representing the nth asset device, and deriving an adjacency matrix A ∈ {0, 1} n×n , wherein A ij = 1 represents that the node i and the node j have a collaborative relationship;

[0028] S26, outputting a node feature matrix X ∈ Rn×d A, where d is the dimension number of each node feature, R is the real number set, as the input of the graph attention network model.

[0029] Optionally, the S3 specifically includes:

[0030] S31, receiving the generated graph structure asset network and the adjacent relationship representation, extracting the node feature matrix and the adjacent matrix as the input basis of the graph attention network model;

[0031] S32, setting the structure parameters and training hyperparameters of the graph attention network model, including the number of network layers, embedding dimension, number of attention heads, and activation function type, while introducing the following newly added structure parameters and training hyperparameters: edge weight guide coefficient λ e , node credibility fusion weight μ c , attention sparsity threshold τ a ;

[0032] S33, using the derived adjacent matrix and edge feature information, combining the set edge weight guide coefficient λ e , fusing edge features and adjacent node features to guide the attention weight calculation process, and constructing a feature aggregation mechanism with edge attribute perception;

[0033] S34, in the node feature construction process, introducing the node credibility fusion weight μ c to control the proportion of node confidence information in the overall feature representation;

[0034] S35, after completing the attention weight calculation, according to the set attention sparsity threshold τ a , filtering the attention connections higher than the threshold in each node, limiting the feature aggregation range, and completing the attention pruning operation;

[0035] S36, completing the weighted aggregation of adjacent features of each node, generating the embedding representation of the asset node, and outputting all node embedding vectors.

[0036] Optionally, the S4 specifically includes:

[0037] S41, setting the initial parameters of the mustang group optimization algorithm, including the population size P, the maximum number of iterations T, and initializing the position vector H k of each mustang individual, where each mustang individual represents three parameters to be optimized in the graph attention network model: edge weight guide coefficient λ e , node credibility fusion weight μ c , and attention sparsity threshold τ a ;

[0038] S42, apply the parameter set of each mustang individual to the graph attention network model respectively to train and predict the asset nodes and calculate the mean absolute error and count the sparsity Sparsity(A (τ) ) after attention pruning, construct the fitness function F(H k ):

[0039]

[0040] wherein Y represents the actual asset evaluation label, represents the predicted asset evaluation result output by the current graph attention network model, ω1 and ω2 are weight coefficients, A (τ) is the attention weight matrix obtained under the pruning threshold τ;

[0041] S43, in each sub-group, according to the fitness function value F(H k ) of the mustang individual, the node confidence aggregation average value the confidence consistency index ρ k and the attention weight standard deviation σ k , calculate the main horse score value S k :

[0042]

[0043] wherein γ1, γ2 and γ3 are weight coefficients, and the individual with the highest score in each sub-group is selected as the main horse according to the main horse score value ranking;

[0044] S44, for non-main horse individuals, update the standard direction according to the main horse guiding strategy to generate candidate positions

[0045]

[0046] wherein is the position vector of the kth mustang individual in the tth iteration, represents the parameter vector position selected as the main horse in the tth round, D t is the direction disturbance vector, r1 is the main horse guiding coefficient, and r2 is the disturbance factor;

[0047] S45, if the fitness improvement amplitude of the mustang individual in the last r rounds is less than the threshold or the confidence consistency is lower than the threshold, trigger the historical jump mechanism to generate a jump position:

[0048]

[0049] wherein η is the jump step coefficient, is the jump position vector of the kth mustang individual, is the parameter position of the kth mustang individual in the t-1th round, is the parameter position of the kth mustang individual before the rth round;

[0050] S46, judge whether the current mustang individual satisfies the jump trigger condition, if yes, set the jump position vector as the next round iteration position Otherwise, set the candidate position as the next round iteration position:

[0051]

[0052] S47, reapply the updated position of the mustang individual to the graph attention network model, perform a new round of training and evaluation, and recalculate F(H k );

[0053] S48, sort all mustang individuals according to the fitness function value F(H k ), record the mustang individual with the minimum fitness as the current optimal solution H * , and update the historical trajectory; if a mustang individual is better than the rest of the subpopulation in fitness and confidence consistency, trigger the mustang individual to migrate across the subpopulation and replace the target subpopulation leader;

[0054] S49, judge whether the termination condition is met: if the fitness function value of the optimal mustang individual changes less than a preset threshold ∈ in continuous q iterations, or the number of iterations reaches the maximum number T, terminate the optimization;

[0055] S410, output the current optimal parameter combination as the structural parameters and training hyperparameters of the graph attention network model.

[0056] Optionally, the S5 specifically comprises:

[0057] S51, receive the graph attention network model parameters optimized by the mustang population optimization algorithm, and load them into the graph attention network model structure;

[0058] S52, input the node feature data and adjacent relationship representation from the trusted data chain into the graph attention network model as input data, and initialize the graph attention network model training process;

[0059] S53, during the graph attention network model training process, perform attention weight calculation and adjacent feature aggregation between nodes according to the optimized parameter configuration;

[0060] S54, in the feature aggregation process, introduce credibility scoring and attention pruning strategy to complete the trusted fusion and sparse aggregation of asset node information;

[0061] S55, after the aggregation of the graph attention network model is completed, a prediction task of asset node state is performed, and evaluation result data of each asset node is generated;

[0062] S56, the asset evaluation result output by the graph attention network model is associated with the corresponding asset node to form a node-level evaluation output set.

[0063] Optionally, the graph attention network model parameter optimized by the wild horse swarm optimization algorithm specifically includes an edge weight guide coefficient, a node credibility fusion weight and an attention sparsity threshold, which is used to guide the weighted aggregation process of adjacent features in the graph attention mechanism, and improve the accuracy and credibility of the asset node evaluation result.

[0064] Optionally, the S6 specifically includes:

[0065] S61, the node features, the embedded representation of the asset node, the evaluation result of the asset and the parameter configuration are uniformly arranged to construct a data structure for abstract generation;

[0066] S62, the data structure constructed is encoded to generate intermediate abstract information for abstract generation;

[0067] S63, the intermediate abstract information is standardized in format to comprehensively generate a data abstract for trusted binding;

[0068] S64, the generated data abstract is subjected to hash encryption processing to form an unforgeable abstract identifier for on-chain binding;

[0069] S65, the encrypted abstract identifier is associated and bound with the trusted data chain of the corresponding multi-source asset data to form an asset node-level binding structure;

[0070] S66, the asset node-level binding structure is written into the Ant Chain to form a trusted chain storage result of the asset node level.

[0071] The beneficial effects of the present application are:

[0072] Based on the full analysis of the technical difficulties of complex structure, heterogeneous data and lack of credibility in new energy charging station asset evaluation, the present application proposes a complete solution integrating graph structure modeling, optimization algorithm driving, trusted data binding and chain storage, effectively overcomes many deficiencies in the prior art, and significantly improves the accuracy, credibility and intelligent level of asset evaluation.

[0073] By constructing a graph attention network model introducing edge attribute and credibility mechanism, the application can establish a dynamic collaborative relationship expression framework between asset devices, realize joint modeling of node state, edge dependence strength and data confidence, and has more expression ability and structural adaptability compared with traditional evaluation models. By introducing edge weight guide coefficient, node credibility fusion weight and attention sparsity threshold and other structural hyperparameters, the model can be structurally fine-tuned according to the actual operation characteristics of the asset and the data quality, and the learning ability and aggregation accuracy of the model for complex asset graph structure are improved.

[0074] Meanwhile, the application designs an improved wild horse optimization algorithm with a credible consistency master horse scoring mechanism and a historical jump local search strategy, and constructs an efficient optimization framework suitable for graph model parameter adjustment. The framework can simultaneously consider prediction accuracy, model sparsity and credible score, and realize stronger global search ability and stable local convergence effect in the search space, effectively solving the problems of difficult parameter adjustment and slow convergence of traditional methods.

[0075] In terms of data credibility management, the application constructs an evaluation result summary hash encryption and chain binding mechanism based on the credible data chain generated by multi-source asset data, ensures the traceability, verifiability and non-tamperability of the whole process of asset evaluation data from input features, model parameters and prediction results to chain records. Through the combination with Ant Chain, the asset node level credible on-chain is realized, which provides a reliable data basis for subsequent evaluation audit, value mapping and collaborative scheduling.

[0076] In summary, the application not only improves the structural modeling ability and prediction accuracy of the asset evaluation model, but also establishes an evaluation closed-loop system with optimization driving, credibility enhancement and chain verification, has the comprehensive advantages of high credibility, high intelligence, high security and high application value, and is suitable for new energy charging station asset digital operation and value management and other key scenes, and has good industrial promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0077] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0078] Figure 1 The flowchart of the new energy charging station asset credible evaluation optimization method based on machine learning proposed by the application;

[0079] Figure 2 The execution flowchart of the wild horse optimization algorithm of the new energy charging station asset credible evaluation optimization method based on machine learning proposed by the application. DETAILED DESCRIPTION

[0080] The application will be described in further detail below with reference to the drawings. These drawings are simplified schematic diagrams and only show the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0081] Reference Figure 1 And Figure 2 A new energy charging station asset credible evaluation optimization method based on machine learning includes the following steps:

[0082] S1, collect multi-source asset data of asset equipment in a new energy charging station, and access the multi-source asset data to an Ant IOT platform through a DTU module, process the multi-source asset data using a credible SDK, and generate a credible data chain;

[0083] S2, based on the credible data chain, calculate a confidence score, integrate asset status, operation and maintenance records, and location information into node features; treat the asset as a node, construct edges according to collaborative behavior, form a graph structure asset network and an adjacency relationship representation;

[0084] S3, based on the graph structure asset network and the adjacency relationship representation, construct a graph attention network model to generate an embedded representation of the asset node;

[0085] S4, apply the wild horse swarm optimization algorithm to jointly optimize the structure parameters and training hyperparameters of the graph attention network model;

[0086] S5, use the optimized graph attention network model for training, based on the credible data chain, aggregate and predict the features of the asset node, and output the evaluation result of the asset;

[0087] S6, integrate the node features, the embedded representation of the asset node, the evaluation result of the asset, and the parameter configuration to generate a summary, and after being hashed and encrypted, the summary is associated and bound with the corresponding credible data chain of the multi-source asset data, and is written into the Ant chain to form a credible chain record;

[0088] S7, continuously collect new asset data and construct a credible data chain, compare with the evaluation result of the asset that has been chained, and when the prediction and feedback deviation exceeds the set threshold, automatically trigger the wild horse swarm optimization algorithm to fine-tune the parameters of the graph attention network model.

[0089] The application realizes the trusted modeling and information fusion of multi-source assets of new energy charging stations by constructing a graph structure asset network and introducing a trusted data chain, effectively improving the integrity and authenticity of asset evaluation. The graph attention network is used for feature aggregation and embedded representation of asset nodes, combined with edge weight guidance, trust fusion and sparse regulation mechanism, which enhances the modeling ability of the model for complex collaborative relationship. The wild horse group optimization algorithm is introduced to jointly optimize the structure parameters and training hyperparameters of the graph model, so that the model has stronger generalization ability and convergence performance while ensuring accuracy. The asset evaluation results are bound to the trusted data chain through summary generation and hash encryption mode, and are stored on the chain for evidence, ensuring that the evaluation data is tamper-proof and the whole process is traceable. The model has a dynamic learning mechanism during operation, which can automatically trigger optimization when the deviation is exceeded, forming a closed-loop adaptive updating capability. The overall scheme has the advantages of high trustworthiness of evaluation results, efficient model optimization, precise structure expression and traceable evaluation data, and is suitable for key scenarios such as trusted operation and maintenance and value analysis of new energy charging station assets.

[0090] In the embodiment, the multi-source asset data of the asset device specifically includes operating status, operation and maintenance records, geographic location and device identification information, which are used to construct node features and support asset evaluation modeling of the graph attention network.

[0091] In the embodiment, the processing using the trusted SDK specifically includes encrypting and signing, hashing and summarizing, and timestamping the multi-source asset data, which is used to generate a trusted data chain of traceable and verifiable multi-source asset data.

[0092] In the embodiment, S2 specifically includes:

[0093] S21, from the generated trusted data chain, extracting the data trustworthiness index C t , data integrity ratio I t , number of successful uploads N t , maximum number of historical uploads N max , and asset evaluation prediction error E t of each asset device at time point t;

[0094] S22, calculating the confidence score D t of the asset node:

[0095]

[0096] Wherein, α, β, γ, δ are confidence weighting coefficients;

[0097] S23, collecting the operating state vector, operation and maintenance record vector and location information vector of each asset, and splicing the calculated confidence score D t into a node feature vector Xt ;

[0098] S24, constructing an edge set E according to asset collaborative behavior data, if asset device i and asset device j are used collaboratively for a number of times F ij ≥ θ in a given time window, defining an edge e ij = 1, otherwise 0, wherein θ is a set collaborative threshold;

[0099] S25, constructing a graph structure asset network G = (V, E) based on a node set V = {v1, v2, …, vn} and an edge set E, taking each asset device as a node in the graph, vn representing the nth asset device, and deriving an adjacency matrix A ∈ {0, 1} n , wherein A n = 1 indicates that node i and node j have a collaborative relationship; n×n ij

[0100] S26, outputting a node feature matrix X ∈ R n×d and an adjacency matrix A, wherein d is the dimension number of each node feature, and R is a real number set, as an input of a graph attention network model.

[0101] The application constructs a node confidence score mechanism with a trusted enhancement capability by extracting and quantifying key indicators in a trusted data chain, effectively making up for the lack of traditional graph models in perceiving node data quality. By combining the running state, operation and maintenance records and location information of the asset device with the confidence score to form a node feature vector, the attributes of the asset in the time, space and trust dimensions are comprehensively reflected. Further, the edge set is dynamically constructed through asset collaborative behavior data, ensuring that the graph structure asset network can truly express the business association and physical collaborative relationship between devices. In the graph modeling process, the node set and the edge set jointly define the graph structure asset network, and the derivation of the adjacency matrix enables the model to efficiently obtain the adjacency relationship, providing a clear topological basis for subsequent feature aggregation of the graph attention network. The finally output node feature matrix and adjacency matrix as the model input not only have information richness and structure expressiveness, but also ensure the trustworthiness and explainability of the graph model training input. The method improves the accuracy of asset state modeling, the rationality of graph structure expression and the credibility of subsequent prediction results, enhancing the adaptability and intelligent level of the new energy charging station asset evaluation system in complex environments.

[0102] In the embodiment, the S3 specifically comprises:

[0103] S31, receiving the generated graph structure asset network and adjacency relationship representation, extracting the node feature matrix and the adjacency matrix as the input basis of the graph attention network model;

[0104] ​​S32, set the structure parameters and training hyperparameters of the graph attention network model, including the number of network layers, embedding dimension, number of attention heads, and activation function type, while introducing the following newly added structure parameters and training hyperparameters: edge weight guide coefficient λ e , node credibility fusion weight μ c , attention sparsity threshold τ a ;

[0105] S33, using the derived adjacency matrix and edge feature information, combining the set edge weight guide coefficient λ e , the edge features and adjacent node features are fused to guide the attention weight calculation process, and a feature aggregation mechanism with edge attribute perception is constructed;

[0106] S34, in the node feature construction process, introduce node credibility fusion weight μ c , control the proportion of node confidence information in the overall feature representation;

[0107] S35, after completing the attention weight calculation, according to the set attention sparsity threshold τ a , filter the attention connections higher than the threshold in each node, limit the feature aggregation range, and complete the attention pruning operation;

[0108] S36, complete the weighted aggregation of each node adjacency feature, generate the embedding representation of the asset node, and output all node embedding vectors.

[0109] The application realizes accurate representation and reliable modeling of new energy charging station asset nodes by constructing a graph attention network model with structure enhancement capability. First, the model takes the node feature matrix and adjacency matrix as input to ensure the synchronous fusion of structure information and attribute information. Second, by setting the number of network layers, embedding dimension, number of attention heads and other basic parameters, and introducing edge weight guide coefficient, node credibility fusion weight and attention sparsity threshold, the model's adaptability to edge features and node confidence is enhanced. Using the edge weight guide mechanism, the control ability of adjacent features on the aggregation direction and intensity is effectively improved. The node credibility fusion mechanism ensures the controllability of the influence of low-quality data on model training. The attention pruning mechanism reduces redundant connections while ensuring the expression ability of the model, improving the training efficiency and model generalization ability. The finally output asset node embedding representation has rich context semantics and reliable structure characteristics, providing high-quality input for subsequent asset state prediction and evaluation. This method significantly improves the model performance while maintaining the integrity of the structure expression, and has the beneficial effects of strong interpretability, high flexibility and good adaptability.

[0110] In the embodiment, the S4 specifically includes:

[0111] S41. Set the initial parameters of the wild horse group optimization algorithm, including the population size P, the maximum number of iterations T, and initialize the position vector H of each wild horse individual. k , where each wild horse individual represents three parameters to be optimized in the graph attention network model: edge weight guidance coefficient λ e , node credibility fusion weight μ c , attention sparseness rate threshold τ a ;

[0112] S42. Apply the parameter combination of each wild horse individual to the graph attention network model, train and predict the asset nodes, and calculate the mean absolute error And count the sparsity after attention pruning (A (τ) ), construct the fitness function F(H k ):

[0113]

[0114] Among them, Y represents the actual asset evaluation label, represents the predicted asset evaluation result output by the current graph attention network model, ω1 and ω2 are weight coefficients, and A (τ) is the attention weight matrix obtained under the pruning threshold of τ;

[0115] S43. In each subgroup, according to the fitness function value F(H k ), node confidence aggregate average Confidence consistency index ρ k and the standard deviation of attention weights σ k , calculate the main horse score S k :

[0116]

[0117] Among them, γ1, γ2 and γ3 are weight coefficients, and the individual with the highest score in each subgroup is selected as the main horse according to the ranking of the main horse score;

[0118] S44: For non-master horse individuals, perform standard direction updates based on the master horse guidance strategy to generate candidate positions

[0119]

[0120] in, is the position vector of the kth wild horse individual in the tth round iteration, Indicates the parameter vector position of the horse selected as the main horse in round t, D t is the direction disturbance vector, r1 is the main horse guidance coefficient, and r2 is the disturbance factor;

[0121] S45, if the fitness improvement of the wild horse individual in the last r rounds is less than a threshold value, or the confidence consistency is lower than a threshold, triggering the historical jump mechanism to generate a jump position:

[0122]

[0123] wherein η is a jump step coefficient, is the jump position vector of the kth wild horse individual, is the parameter position of the kth wild horse individual in the t-1th round, is the parameter position of the kth wild horse individual before the rth round;

[0124] S46, judging whether the current wild horse individual satisfies the jump trigger condition, if yes, taking the jump position vector as the next round iteration position Otherwise, taking the candidate position as the next round iteration position:

[0125]

[0126] S47, reapplying the updated position of the wild horse individual to the graph attention network model, performing a new round of training and evaluation, and recalculating F(H k );

[0127] S48, sorting all wild horse individuals according to the fitness function value F(H k ), recording the wild horse individual with the minimum fitness as the current optimal solution H * , and updating the historical trajectory; if a wild horse individual is superior to the rest of the subgroups in terms of fitness and confidence consistency, triggering the wild horse individual to migrate across groups and replace the target sub-group leader;

[0128] S49, judging whether the termination condition is satisfied: if the fitness function value of the optimal wild horse individual changes less than a preset threshold ∈ in continuous q iterations, or the iteration number reaches the maximum number T, the optimization is terminated;

[0129] S410, outputting the current optimal parameter combination as the structural parameters and training hyperparameters of the graph attention network model.

[0130] The wild horse group optimization algorithm with structural improvement is introduced in the graph attention network model parameter optimization process, which effectively improves the precision and efficiency of model parameter search. By taking the edge weight guide coefficient, node credibility fusion weight and attention sparsity threshold as the core parameters to be optimized, and constructing a double-objective fitness function combining prediction error and sparsity, the optimization process considers both evaluation accuracy and structural simplicity. In the main horse selection mechanism, the node confidence aggregation mean, confidence consistency and attention weight standard deviation are introduced as scoring factors, so as to avoid the deviation selection problem caused by excessive dependence on a single error indicator, and improve the credibility and stability of the population dominant direction. At the same time, the invention sets up a jump triggering mechanism, when the fitness improves slowly or the confidence fluctuates abnormally, the parameter jump operation based on historical trajectory is executed, which effectively jumps out of the local optimal trap. Through the combination of main horse guiding strategy and jump updating, the dynamic migration and collaborative optimization of wild horse group are realized, and an optimization process with high adaptability, high convergence speed and high stability is constructed. The optimal parameter combination output finally can significantly improve the prediction performance, structural sparsity and credible expression ability of the graph attention model, and provides stable and efficient optimization support for asset evaluation model.

[0131] In the embodiment, the S5 specifically includes:

[0132] S51, receiving the graph attention network model parameters optimized by the wild horse group optimization algorithm, and loading into the graph attention network model structure;

[0133] S52, taking the node feature data and adjacent relationship from the trusted data chain as the input data of the graph attention network model, and initializing the graph attention network model training process;

[0134] S53, in the graph attention network model training process, the attention weight calculation and adjacent feature aggregation between nodes are performed according to the optimized parameters;

[0135] S54, in the feature aggregation process, the credibility score and attention pruning strategy are introduced to complete the credible fusion and sparse aggregation of asset node information;

[0136] S55, after the graph attention network model aggregation is completed, the prediction task of asset node state is executed, and the evaluation result data of each asset node is generated;

[0137] S56, associating the asset evaluation result output by the graph attention network model with the corresponding asset node to form a node-level evaluation output set.

[0138] This method achieves high-quality feature aggregation and prediction capabilities at the asset node level by loading the optimal parameter combination obtained from the Wild Horse Swarm Optimization algorithm into the graph attention network model. This process fully utilizes the optimized edge weight guidance coefficient, node credibility fusion weight, and attention sparsity threshold to ensure that the model can both highlight key adjacent information during feature aggregation and control the interference of low-credibility features on the results, effectively enhancing the model's stability and anti-interference capabilities. During training, combined with trusted data chain input, the model accurately allocates attention between nodes and fuses adjacent features, achieving dual representation of structure and semantics. Furthermore, by introducing a sparsification mechanism to prune attention weights, the model's computational efficiency and structural simplicity are improved in complex graph structures. During the prediction phase, the model outputs an independent evaluation result for each asset node, accurately binding it to the node information to form a clearly structured evaluation output set. This method offers the advantages of high evaluation result accuracy, strong feature aggregation credibility, and well-regulated structural output, providing stable and reliable intelligent support for asset operating status perception and value assessment at new energy charging stations.

[0139] In this embodiment, the graph attention network model parameters optimized by the wild horse swarm optimization algorithm specifically include the edge weight guidance coefficient, the node credibility fusion weight and the attention sparsity rate threshold, which are used to guide the weighted aggregation process of adjacent features in the graph attention mechanism and improve the accuracy and credibility of the asset node evaluation results.

[0140] In this embodiment, S6 specifically includes:

[0141] S61. Unify and organize the node features, embedded representations of asset nodes, asset evaluation results, and parameter configurations to construct a data structure for summary generation.

[0142] S62, encoding the constructed data structure to generate intermediate summary information for summary generation;

[0143] S63. Normalize the format of the intermediate summary information and comprehensively generate a data summary for trusted binding;

[0144] S64. Perform hash encryption on the generated data digest to form an unalterable digest identifier for on-chain binding;

[0145] S65. Associate and bind the encrypted summary identifier with the corresponding trusted data chain of multi-source asset data to form an asset node-level binding structure;

[0146] S66. Write the asset node-level binding structure into Ant Chain to form a trusted chain evidence storage result at the asset node level.

[0147] The application realizes the two-way trusted transmission of asset evaluation results in the data layer and the chain layer by constructing an information summary based on the fusion of model output and asset data, forming a complete, structured and trusted binding process. First, the system unifies the input features, embedded representation, evaluation results and model parameter configuration of the asset node, constructs an abstract data structure, and ensures that the on-chain record has context integrity and source verifiability. Then, standard abstract information is generated through encoding and standardization processing, and a hash encryption operation is performed to ensure the uniqueness and tamper resistance of the abstract identifier. The encrypted abstract is associated with the trusted data chain of the original asset data, effectively establishing a trusted mapping relationship between the asset evaluation behavior and the data ontology. Finally, the binding structure is written into Ant Chain, realizing the trusted chain storage of asset node level, and making the evaluation process have the ability of full-process traceability, result verification and data audit. The mechanism of the application significantly enhances the trustworthiness and supervisability of asset evaluation results, providing a safe and transparent foundation for asset management, operation audit and intelligent decision-making of new energy charging stations.

[0148] Embodiment 1

[0149] In order to verify the feasibility of the application in implementation, the application is applied to a large new energy charging station, which deploys a total of 120 sets of asset equipment, covering direct current fast charging piles, alternating current slow charging piles, battery energy storage cabinets, power distribution control cabinets and edge sensing equipment, etc., with a running time span of more than 3 years. Due to the diversity of equipment types, dynamic fluctuations in state and complex data sources, traditional asset evaluation methods face the following problems in actual operation: the evaluation accuracy is not high, and the cooperative relationship between devices cannot be effectively expressed; some low-quality data misleads model training, and the evaluation result is not reliable enough; there is a lack of unified storage method for evaluation data, making it difficult to support cross-department and cross-platform data audit and traceability.

[0150] Therefore, the project team introduced the "new energy charging station asset trusted evaluation optimization method based on machine learning" proposed by the application for system replacement and pilot deployment. First, the DTU device is used to collect all kinds of asset operation data, fault records, maintenance logs and location states, and the trusted data chain generated by the Ant IOT platform is combined to extract key factors such as data integrity, upload frequency and confidence index of assets at different time points for constructing node features.

[0151] Then, a graph structure asset network is constructed, and the cooperative behavior of each asset node in a given time window is mapped as an edge in the graph, and the graph attention network model is trained. In the model construction, key parameters such as edge weight guide coefficient, node trust degree fusion weight and attention sparsity threshold are introduced, and the wild horse group optimization algorithm is used for joint optimization, so that the model has the ability of self-adaptive adjustment and trusted feature enhancement in structure.

[0152] After 27 rounds of optimization iterations, the evaluation model reached a state of convergence. Compared with the original traditional evaluation model of this station, the method of the application achieved significant improvement in multiple core indicators. The model prediction mean absolute error decreased from 0.121 to 0.083, the attention sparsity rate increased from 28.3% to 51.2%, the correlation of node prediction results and actual feedback data increased to 0.87; the model's identification rate for high-risk assets increased from 76.4% to 92.1%, providing more accurate reference for subsequent operation and maintenance strategies. At the same time, the newly added confidence fusion mechanism enables the model to have sensitive control ability for data quality differences, achieving a 23.6% increase in confidence score perception on the same batch of data. The optimized model also reduces the number of convergence rounds from 42 to 27, saving about 35% of training time.

[0153] After the evaluation results are generated, the system structures the embedded features, prediction results and model parameters of each asset node, generates summary information, binds it with the original asset data chain after hash processing, and successfully writes it into Ant Chain for notarization, ensuring the authenticity and verifiability of the evaluation results. The deployment stage system records show that the asset evaluation summary generation success rate is 100%, and the data chain association integrity rate reaches 99.4%, which is significantly better than the traditional system's problem of notarization and chain fragmentation.

[0154] Table 1 Performance comparison between the method of the application and the traditional evaluation method

[0155]

[0156] In terms of evaluation mean absolute error (MAE), the error of the traditional evaluation method is 0.121, while the error of the method of the application is reduced to 0.083, with a significant improvement in accuracy, indicating that the application has advantages in model construction, feature fusion and training strategy, and can more accurately reflect the true state of new energy charging station assets. In terms of attention sparsity rate, the traditional model is 28.3%, while the application improves to 51.2%, which means that the application effectively enhances the selectivity of the attention mechanism and the simplicity of the model structure by introducing the attention sparsity rate threshold, reduces redundant connections, and improves operational efficiency and interpretability.

[0157] In terms of the correlation of prediction results and feedback data, the traditional method is 0.69, while the method of the application reaches 0.87, indicating that the application captures the asset operation trend more accurately, and the prediction results are highly consistent with the actual feedback, which is conducive to building a trusted and reliable intelligent operation and maintenance system. In terms of high-risk asset identification rate, the model of the application achieves an identification rate of 92.1%, which is 15.7% higher than the traditional method, which is of great significance for early intervention of potential risk assets and reduction of equipment failure rate.

[0158] In terms of the number of model convergence rounds, traditional methods require 42 iterations to reach convergence, while this method only requires 27 rounds, improving optimization efficiency by approximately 35%. This significantly reduces the time required for model training and is particularly suitable for real-time evaluation scenarios deployed on edge devices or resource-constrained platforms. Furthermore, this method introduces a confidence-aware mechanism, enabling the model to perceive and adjust the quality of input data, achieving a 23.6% confidence score improvement, minimizing the impact of low-quality data on evaluation results during aggregation.

[0159] In terms of trusted chain-related indicators, the present invention achieved a 100% success rate for writing assessment result summaries onto the chain, whereas traditional methods were unable to support trusted chain-up operations. Furthermore, the present invention achieved a 99.4% data chain association integrity rate, significantly improving the traceability and supervisory visualization capabilities of assessment data, and addressing the shortcomings of traditional methods in terms of data transparency and tamper resistance.

[0160] Overall, the method of the present invention performs superiorly in multiple dimensions such as model structure design, optimization strategy, prediction capability and data credibility assurance, verifying its practical application value in the asset assessment scenario of new energy charging stations, and has significant engineering adaptability and technological advancement.

[0161] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A new energy charging station asset trustworthiness assessment optimization method based on machine learning, characterized by: The steps include: S1. Collect multi-source asset data of assets and equipment in new energy charging stations, connect the multi-source asset data to the Ant IOT platform through the DTU module, process it using the trusted SDK, and generate a trusted data chain; S2. Based on the trusted data chain, calculate the confidence score and integrate it with the asset status, operation and maintenance records, and location information to form node features. Use assets as nodes and construct edges based on collaborative behaviors to form a graph-structured asset network and adjacency relationship representation. S3. Based on the graph structure asset network and adjacency relationship representation, a graph attention network model is constructed to generate an embedded representation of the asset node; S4. Apply the wild horse swarm optimization algorithm to jointly optimize the structural parameters and training hyperparameters of the graph attention network model; S5. Use the optimized graph attention network model for training, aggregate and predict the features of asset nodes based on the trusted data chain, and output the asset evaluation results; S6. Generate a summary of the node features, embedded representation of the asset node, asset assessment results, and parameter configuration. After hash encryption, it is associated and bound to the trusted data chain of the corresponding multi-source asset data and written into the Ant Chain to form a trusted chain certificate. S7. Continuously collect new asset data and build a trusted data chain, compare it with the evaluation results of assets already on the chain, and automatically trigger the Wild Horse Swarm Optimization algorithm to fine-tune the parameters of the graph attention network model when the deviation between the prediction and feedback exceeds the set threshold.

2. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The multi-source asset data of the asset equipment specifically includes operating status, operation and maintenance records, geographic location and equipment identification information, which are used to construct node features and support asset evaluation modeling of the graph attention network.

3. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The processing using the trusted SDK specifically includes encrypting and signing the multi-source asset data, hashing the summary, and timestamping the data to generate a trusted data chain of traceable and verifiable multi-source asset data.

4. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The S2 specifically includes: S21. Extract the data credibility index C of each asset device at time point t from the generated trusted data chain. t , Data Integrity Ratio I t , Number of successful uploads N t 、The maximum number of uploads in history N max , and asset valuation prediction error E t ; S22. Calculate the confidence score D of the asset node t : Among them, α, β, γ, δ are confidence weighting coefficients; S23, collect the operating status vector, operation and maintenance record vector and location information vector of each asset, and compare them with the calculated confidence score D t Spliced ​​into node feature vector X t ; S24. Construct an edge set E based on the asset collaborative behavior data. If asset equipment i and asset equipment j are used collaboratively for the number of times F in a given time window, ij ≥θ, then define the edge e ij =1, otherwise 0, where θ is the set collaboration threshold; S25, based on the node set V={v1,v2,…,v n } and the edge set E to build a graph structure asset network G = (V, E), taking each asset device as a node in the graph, v n Represents the nth asset device and derives the adjacency matrix A∈{0,1} n×n , where A ij =1 indicates that there is a collaborative relationship between node i and node j; S26, output node feature matrix X∈R n×d and the adjacency matrix A, where d is the number of dimensions of each node feature and R is a set of real numbers, which serves as the input of the graph attention network model.

5. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Receive the generated graph structure asset network and adjacency relationship representation, extract the node feature matrix and adjacency matrix as the input basis of the graph attention network model; S32. Set the structural parameters and training hyperparameters of the graph attention network model, including the number of network layers, embedding dimension, number of attention heads, and activation function type. At the same time, introduce the following new structural parameters and training hyperparameters: edge weight guidance coefficient λ e , node credibility fusion weight μ c , attention sparseness rate threshold τ a ; S33, using the derived adjacency matrix and edge feature information, combined with the set edge weight guidance coefficient λ e , the edge features are fused with the adjacent node features, guiding the attention weight calculation process and building an edge attribute-aware feature aggregation mechanism; S34. In the process of node feature construction, the node credibility fusion weight μ is introduced c , controls the proportion of node confidence information in the overall feature representation; S35. After completing the attention weight calculation, according to the set attention sparse rate threshold τ a , filter the attention connections above the threshold in each node, limit the feature aggregation range, and complete the attention pruning operation; S36. Complete the weighted aggregation of the adjacency features of each node, generate the embedded representation of the asset node, and output the embedding vectors of all nodes.

6. The method for optimizing trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Set the initial parameters of the wild horse group optimization algorithm, including the population size P, the maximum number of iterations T, and initialize the position vector H of each wild horse individual. k , where each wild horse individual represents three parameters to be optimized in the graph attention network model: edge weight guidance coefficient λ e , node credibility fusion weight μ c , attention sparseness rate threshold τ a ; S42. Apply the parameter combination of each wild horse individual to the graph attention network model, train and predict the asset nodes, and calculate the mean absolute error And count the sparsity after attention pruning (A (τ) ), construct the fitness function F(H k ): Among them, Y represents the actual asset evaluation label, represents the predicted asset evaluation result output by the current graph attention network model, ω1 and ω2 are weight coefficients, and A (τ) is the attention weight matrix obtained under the pruning threshold of τ; S43. In each subgroup, according to the fitness function value F(H k ), node confidence aggregate average Confidence consistency index ρ k and the standard deviation of attention weights σ k , calculate the main horse score S k : Among them, γ1, γ2 and γ3 are weight coefficients, and the individual with the highest score in each subgroup is selected as the main horse according to the ranking of the main horse score; S44: For non-master horse individuals, perform standard direction updates based on the master horse guidance strategy to generate candidate positions in, is the position vector of the kth wild horse individual in the tth round iteration, Indicates the parameter vector position of the horse selected as the main horse in round t, D t is the direction disturbance vector, r1 is the main horse guidance coefficient, and r2 is the disturbance factor; S45. If the fitness improvement of the wild horse individual in the last r rounds is less than the threshold, or the confidence consistency is lower than the threshold, the historical jump mechanism is triggered and the jump position is generated: Among them, η is the jump step coefficient, is the jumping position vector of the kth wild horse individual, is the parameter position of the kth wild horse individual in the t-1th round, is the parameter position of the kth wild horse individual before round r; S46, determine whether the current wild horse individual meets the jump trigger condition, if so, set the jump position vector As the next iteration position Otherwise, the candidate position As the next iteration position: S47, reapply the updated position of the wild horse individual to the graph attention network model, perform a new round of training and evaluation, and recalculate F(H k ); S48, according to the fitness function value F(H k ) Sort all wild horse individuals and record the wild horse individual with the minimum fitness as the current optimal solution H * , update the historical trajectory; if a wild horse individual is better than the rest of the subgroup in terms of fitness and confidence consistency, it will trigger the cross-group migration of the wild horse individual and replace the main horse of the target subgroup; S49, judging whether the termination condition is met: if the fitness function value of the optimal wild horse individual changes less than a preset threshold ∈ in consecutive q iterations, or the number of iterations reaches the maximum number T, then the optimization is terminated; S410: Output the current optimal parameter combination As the structural parameters and training hyperparameters of the graph attention network model.

7. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The S5 specifically includes: S51, receiving the graph attention network model parameters optimized by the wild horse swarm optimization algorithm and loading them into the graph attention network model structure; S52. Use the node feature data and adjacency relationship representation from the trusted data chain as input data for the graph attention network model, and initialize the graph attention network model training process; S53, during the graph attention network model training process, performing attention weight calculation and adjacent feature aggregation between nodes according to the optimized parameter configuration; S54. In the feature aggregation process, the credibility score and attention pruning strategy are introduced to complete the credible fusion and sparse aggregation of asset node information; S55. After the graph attention network model aggregation is completed, the asset node status prediction task is performed to generate evaluation result data for each asset node; S56. Associate the asset evaluation results output by the graph attention network model with the corresponding asset nodes to form a node-level evaluation output set.

8. The method for optimizing the trustworthy assessment of new energy charging station assets based on machine learning according to claim 7 is characterized in that: The graph attention network model parameters optimized by the wild horse swarm optimization algorithm specifically include the edge weight guidance coefficient, the node credibility fusion weight and the attention sparsity rate threshold, which are used to guide the weighted aggregation process of adjacent features in the graph attention mechanism and improve the accuracy and credibility of the asset node evaluation results.

9. The method for optimizing trustworthy assessment of new energy charging station assets based on machine learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Unify and organize the node features, embedded representations of asset nodes, asset evaluation results, and parameter configurations to construct a data structure for summary generation. S62, encoding the constructed data structure to generate intermediate summary information for summary generation; S63. Normalize the format of the intermediate summary information and comprehensively generate a data summary for trusted binding; S64. Perform hash encryption on the generated data digest to form an unalterable digest identifier for on-chain binding; S65. Associate and bind the encrypted summary identifier with the corresponding trusted data chain of multi-source asset data to form an asset node-level binding structure; S66. Write the asset node-level binding structure into Ant Chain to form a trusted chain evidence storage result at the asset node level.

Images