Method and system for constructing multi-source data comprehensive safety evaluation model of energy storage battery system

Abstract

The invention discloses a method and a system for constructing a multi-source data comprehensive safety evaluation model of an energy storage battery system, and the method comprises the steps: calculating the component score accumulated contribution degree and comprehensive component score of each index parameter by adopting a multivariate statistical analysis method; obtaining the risk score of each index parameter, calculating the cloud probability feature value of each risk level based on a cloud model, and forming a risk cloud probability data set; according to a Transform architecture based on a multi-head attention mechanism, an evaluation model is constructed; adjusting the number of heads in a multi-head attention mechanism according to the accumulated contribution degree of the component scores, and training the model by taking the comprehensive component scores and the risk cloud probability data set as an input data set and an output data set of the evaluation model; optimizing hyper-parameters of the trained evaluation model by adopting a subtraction optimizer algorithm improved based on golden sine by taking F1 score optimization as an optimization target; and the evaluation model after hyper-parameter optimization is used as a multi-source data comprehensive safety evaluation model of the energy storage battery system, so that the accuracy and the reliability of an evaluation result are improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology for energy storage battery systems. Specifically, it relates to a method and system for constructing a comprehensive safety evaluation model for multi-source data of energy storage battery systems based on a multi-head attention mechanism. Background Technology

[0002] With the rapid development of energy storage technology, energy storage battery systems are playing an increasingly important role in energy storage and conversion. However, the safety of energy storage battery systems has always been a key factor restricting their widespread application. Safety hazards such as battery thermal runaway, overcharging and over-discharging, and electrical faults can lead to serious accidents such as fires and explosions. To ensure the safe operation of energy storage battery systems, a comprehensive and accurate safety assessment is necessary.

[0003] In existing technologies, energy storage battery safety evaluation methods are often based on a single type of data, such as relying solely on electrical parameters like battery voltage and current, which fails to fully consider the complex factors involved in battery operation. Furthermore, these methods struggle to effectively handle the complex relationships between multi-source data, resulting in limited accuracy and reliability of the evaluation results. In recent years, machine learning technology has been increasingly applied to the field of energy storage battery safety evaluation; however, existing machine learning models are insufficient in fusing multi-source data and mining deep data features, failing to meet the growing demand for high-precision safety evaluation of energy storage battery systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for constructing a comprehensive safety evaluation model for energy storage battery systems based on multi-source data using a multi-head attention mechanism. This addresses the limitations of traditional energy storage battery safety evaluation methods, which rely on single data sources and struggle to handle complex correlations between multiple data sources, as well as the deficiencies of existing machine learning models in fusing multi-source data and mining deep features. The goal is to achieve a comprehensive and accurate safety evaluation of energy storage battery systems, thereby improving the accuracy and reliability of the evaluation results.

[0005] The present invention adopts the following technical solution.

[0006] This invention proposes a method for constructing a multi-source data-integrated safety evaluation model for energy storage battery systems, comprising:

[0007] Obtain the various index parameters of the energy storage battery; use multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component scores of each index parameter after normalization.

[0008] Obtain risk scores for various parameters of energy storage batteries, calculate cloud probability feature values ​​for each risk level based on the cloud model, and construct a risk cloud probability dataset.

[0009] An evaluation model is constructed based on the Transformer architecture with a multi-head attention mechanism. The number of heads in the multi-head attention mechanism is dynamically adjusted according to the cumulative contribution of the component scores. The comprehensive component scores are used as the input dataset of the evaluation model, and the risk cloud probability dataset is used as the output dataset of the evaluation model.

[0010] With the optimal F1 score as the optimization objective, the hyperparameters of the trained evaluation model are optimized using a subtraction optimizer algorithm based on the golden sine.

[0011] The hyperparameter-optimized evaluation model serves as a comprehensive safety evaluation model for multi-source data of energy storage battery systems.

[0012] The parameters of energy storage batteries include: voltage, temperature, internal resistance, capacity, state of charge, state of health, state of energy, and state of power.

[0013] After normalizing the various indicator parameters, the standardized indicator parameters are obtained as follows:

[0014]

[0015] In the formula, x ij Let X be the j-th index parameter. j The i-th set of measurements, For x ij Standardized values, and s j All are the j-th indicator parameter X j The corresponding intermediate variable, n, is the j-th indicator parameter X. j The number of measurement groups, p is the number of index parameters.

[0016] Calculate the covariance matrix of the standardized index parameters, and solve the covariance matrix to obtain the unit eigenvectors and eigenvalues;

[0017] Using multivariate statistical analysis, and utilizing unit eigenvectors and eigenvalues, the cumulative contribution of the component scores and the comprehensive component scores of the standardized indicator parameters are calculated, including:

[0018] The component scores are determined using the following relationship:

[0019]

[0020] In the formula, z ij For the score of the i-th component of the j-th indicator parameter, a kj Let a be the j-th unit eigenvector. j The kth element, Let X be the k-th index parameter. k The i-th standardized value;

[0021] The overall component scores are as follows:

[0022]

[0023] In the formula, Z represents the comprehensive component score. λ represents the contribution of a component to the score. k This is the kth eigenvalue.

[0024] Under risk level r, the j-th risk constraint range is constructed using the maximum and minimum risk scores of the j-th indicator parameter of the energy storage battery; the j-th set of overall characteristic parameters of the cloud model is calculated using the j-th risk constraint range, as follows:

[0025]

[0026] In the formula, E x,j E n,j H e,j Z′ represents the expectation, entropy, and hyperentropy of the j-th group of the cloud model, respectively. j,min Z′ j,max These are the minimum and maximum risk scores for the j-th indicator parameter of the energy storage battery, respectively.

[0027] Under risk level r, the weighted sum of the overall feature parameters of the p groups of the cloud model constitutes the cloud probability feature value for risk level r. The following relationship must be satisfied:

[0028]

[0029] In the formula, α j The component score contribution of the j-th index parameter of the energy storage battery is used as the weighting coefficient for calculating the weighted sum.

[0030] satisfy and When the threshold is reached, the number of heads in the multi-head attention mechanism is set to q, where Q is the threshold.

[0031] The hyperparameters used to evaluate the model include: the number of heads q in the multi-head attention mechanism, the key and query dimensions d, and the hidden layer dimension D of the feedforward neural network. ff Model training learning rate L R Number of training input samples b, number of training rounds N a Regularized Dropout probability D R We employ a subtraction optimizer algorithm based on the golden sine wave to find the optimal combination of hyperparameters for the F1 score through iterative calculation.

[0032] In the position update step of the subtraction optimizer based on the golden sine improvement, the search agent position update formula is:

[0033]

[0034] In the formula, X is the update position for the i-th search agent. zi Let ωt be the position of the i-th search agent, and sin(ωt) be the golden sine function. t is the current iteration number. Let N be the m-dimensional vector of the i-th search agent, whose norm ranges from [0,1] and follows a normal distribution. υ For special operators, it means:

[0035]

[0036] in, It is an m-dimensional vector whose components are random numbers generated from the set {1,2}. The operation "*" represents the Hadamard product of two vectors. F(A) and F(B) are the objective function values ​​of search agents A and B, respectively. sign(·) is the sign function.

[0037] Objective function for accepting the new position:

[0038]

[0039] Among them, F i new and F i These are search proxies and X zi The objective function value.

[0040] This invention also proposes a system for constructing a multi-source data integrated safety evaluation model for energy storage battery systems, comprising: an input data processing module, an output data processing module, a model building module, and a model training module;

[0041] The input data processing module is used to acquire various index parameters of the energy storage battery; it uses multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component score of each index parameter after normalization; the comprehensive component score is the input dataset of the evaluation model.

[0042] The output data processing module is used to obtain the risk scores of various indicators of energy storage batteries, calculate the cloud probability feature values ​​of each risk level based on the cloud model, and form a risk cloud probability dataset; the risk cloud probability dataset is the output dataset of the evaluation model.

[0043] The model training module is used to construct an evaluation model based on a Transformer architecture with a multi-head attention mechanism; dynamically adjust the number of heads in the multi-head attention mechanism according to the cumulative contribution of component scores; train the model using the input and output datasets; optimize the hyperparameters of the trained evaluation model with the optimal F1 score as the optimization objective, using a subtraction optimizer algorithm based on a golden sine improvement; the hyperparameter-optimized evaluation model serves as a comprehensive safety evaluation model for multi-source data of energy storage battery systems.

[0044] The present invention is also a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0045] The present invention is also a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0046] The beneficial effects of this invention, compared with the prior art, include at least the following: This invention comprehensively considers the complex factors and complex correlations between data in multi-source data of energy storage battery systems. Through normalization processing, calculation of covariance matrices, and multivariate statistical analysis, it mines the inherent connections between data, integrates and classifies data to generate a decoupled input dataset, providing a high-quality data foundation for accurate evaluation and improving evaluation accuracy. This invention employs a multi-head attention mechanism, focusing on different feature subspaces of data in parallel, enabling it to capture key information from multiple perspectives and accurately integrate data. Simultaneously, it adaptively adjusts the number of heads based on the cumulative contribution of component scores, dynamically optimizing according to data characteristics, enhancing the ability to process complex data, improving the model's accuracy in capturing complex correlations between multi-source data, and thus improving the reliability of safety evaluation. This invention uses a subtraction optimizer (SABO) algorithm based on a golden sine improvement to optimize hyperparameters, using the F1 score as a comprehensive evaluation index, balancing global search and local exploitation, avoiding getting trapped in local optima, finding the optimal hyperparameter combination, enhancing model performance and reliability, and making safety evaluation more accurate and efficient. Attached Figure Description

[0047] Figure 1 This is a flowchart of a method for constructing a multi-source data integrated safety evaluation model for energy storage battery systems based on a multi-head attention mechanism, as proposed in this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0049] This invention proposes a method for constructing a multi-source data-based integrated safety evaluation model for energy storage battery systems based on a multi-head attention mechanism, such as... Figure 1 As shown, it includes the following steps:

[0050] Step 1: Obtain the various index parameters of the energy storage battery; use multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component scores of the normalized index parameters.

[0051] Specifically, the parameters of energy storage batteries include, but are not limited to: voltage, temperature, internal resistance, capacity, state of charge, state of health, state of energy, and state of power.

[0052] Traditional methods often rely on only a single type of data (such as electrical parameters like voltage and current) without considering the comprehensive correlation of multi-source data (eight types of characteristic parameters including voltage, temperature, internal resistance, capacity, SOC, SOH, SOE, and SOP). Therefore, this invention proposes multi-source data fusion processing to integrate multi-dimensional heterogeneous data of energy storage batteries, covering the electrical, physical, and state parameters of battery operation, forming a complete data characterization system.

[0053] Common data preprocessing methods include the maximum-minimum method and standard deviation normalization method, but they do not consider the correlation between the feature parameters, leading to an imbalance of feature weights during model training. This invention applies a multivariate statistical analysis method after normalizing the standard deviation of the index parameters to decouple the various index parameters of the energy storage battery. The data after data processing are component scores, which have mutually independent characteristics.

[0054] Specifically, data preprocessing is performed on each indicator parameter, including:

[0055] 1) Normalize the parameters of each indicator to obtain the standardized parameters of each indicator;

[0056] The formula for normalizing the various indicator parameters is as follows:

[0057]

[0058] In the formula, x ij Let X be the j-th index parameter. j The i-th set of measurements, For x ij Standardized values, and s j All are the j-th indicator parameter X j The corresponding intermediate variable, n, is the j-th indicator parameter X. j The number of measurement groups, p is the number of index parameters.

[0059] Normalization eliminates dimensional differences, unifying characteristic parameters of different dimensions to the same scale, making the covariance matrix equivalent to the correlation coefficient matrix, which facilitates subsequent multivariate statistical analysis.

[0060] Calculate the covariance matrix of the standardized index parameters, and solve the covariance matrix to obtain the unit eigenvectors and eigenvalues.

[0061] 3) Using multivariate statistical analysis, the cumulative contribution of component scores and comprehensive component scores of standardized indicator parameters are calculated by utilizing unit eigenvectors and eigenvalues. This enables the fusion and classification of standardized indicator parameters, generating the input dataset Z for model training.

[0062] Component scores are determined using unit eigenvectors and standardized index parameters, with the following relationship:

[0063]

[0064] In the formula, z ij For the score of the i-th component of the j-th indicator parameter, a kj Let a be the j-th unit eigenvector. j The kth element, Let X be the k-th index parameter. k The i-th standardized value;

[0065] The composite component score was calculated using multivariate statistical analysis, as follows:

[0066]

[0067] In the formula, Z represents the comprehensive component score. λ represents the contribution of a component to the score. k This is the kth eigenvalue.

[0068] If the safety of energy storage batteries is evaluated based on a single type of data or normalized multidimensional data, it is impossible to fully consider complex factors and process the complex relationships between data. Therefore, this invention obtains multi-source characteristic parameters such as voltage, temperature, and internal resistance, and after normalization processing, calculation of covariance matrix and multivariate statistical analysis, it explores the inherent relationships of the data, integrates and classifies the data to generate a decoupled input dataset, provides a high-quality data foundation for accurate evaluation, and improves the accuracy of evaluation.

[0069] Traditional methods use coupled data for modeling, resulting in high model complexity and low computational efficiency due to the interconnected information between data points. This invention calculates the covariance matrix of standardized index parameters, which quantifies the linear correlation between features. By solving for the eigenvalues ​​and unit eigenvectors of the covariance matrix, the decoupled comprehensive component scores are extracted and used as the training input dataset for a multi-source data comprehensive evaluation model of energy storage battery systems based on a multi-head attention mechanism. Through a combination of improvements in multi-source data fusion, normalization, and multivariate statistical analysis, a high-dimensional, decoupled input dataset is constructed, solving the core problems of traditional methods such as "partial data, inconsistent scale, and ambiguous correlations." This lays the foundation for accurate extraction of deep features using subsequent multi-head attention mechanisms and Transformer models, ultimately achieving comprehensiveness, accuracy, and efficiency in energy storage battery safety evaluation.

[0070] Step 2: Obtain the risk scores of various indicators of the energy storage battery, calculate the cloud probability feature values ​​of each risk level based on the cloud model, and construct a risk cloud probability dataset.

[0071] Specifically, step 2 includes:

[0072] Step 2.1: Obtain the risk scores of various parameters of the energy storage battery;

[0073] The eight characteristic parameters of energy storage batteries (voltage, temperature, internal resistance, etc.) may have conflicting trends (such as SOC rising but SOH falling) or deviate from the threshold (such as normal voltage but excessive temperature). Expert scoring methods should be used to comprehensively judge the risk level and analyze the inconsistency of multi-source data.

[0074] Multiple experts were invited to score the risk level of typical multi-source data combination scenarios. In this example, the risk level was divided into 5 levels, and each scenario corresponded to a set of feature parameters. The scoring results were discretized into fuzzy language values. In this example, "almost no risk" corresponds to an expert score of 0-2; "slight risk" corresponds to an expert score of 2-4; "moderate risk" corresponds to an expert score of 4-6; "medium risk" corresponds to an expert score of 6-8; and "high risk" corresponds to an expert score of 8-10, thus forming the original evaluation dataset.

[0075] Traditional safety assessments rely on single data points or fixed thresholds (e.g., determining risk solely based on voltage deviation from a threshold), failing to integrate expert experience with the complex correlations between multi-source data. This invention establishes a multi-dimensional expert scoring system based on the consistency / conflict analysis of eight characteristic parameters of energy storage batteries (e.g., when voltage is normal but temperature is abnormal, experts need to comprehensively assess the risk level). Through expert evaluation of inconsistencies in multi-source data (e.g., risk level determination when SOC and SOH trends diverge), a risk assessment standard closer to actual operating scenarios is generated. Since safety assessments are subjective, the industry typically uses expert scoring methods to quantify these subjective factors.

[0076] Step 2.2: Under risk level r, construct the j-th risk constraint range using the maximum and minimum risk scores of the j-th indicator parameter of the energy storage battery; calculate the j-th set of overall characteristic parameters of the cloud model using the j-th risk constraint range, as follows:

[0077]

[0078] In the formula, E x,j E n,j H e,j Z′ represents the expectation, entropy, and hyperentropy of the j-th group of the cloud model, respectively. j,min Z′ j,max These are the minimum and maximum risk scores for the j-th indicator parameter of the energy storage battery, respectively.

[0079] Traditional methods often use deterministic thresholds to classify risk levels (e.g., "voltage > 3.7V is high risk"), which cannot quantify data volatility and the ambiguity of expert judgment. Therefore, this invention adopts a risk probability analysis method based on cloud models. It introduces cloud model theory, and for each risk level, it statistically analyzes all corresponding expert scores, transforming these scores into a risk cloud probability distribution. The inverse cloud algorithm is then used to fit three characteristic parameters—expectation, entropy, and hyperentropy—to characterize the ambiguity and randomness of risk levels based on the cloud model.

[0080] Step 2.3: Under risk level r, the weighted sum of the overall feature parameters of the p groups of the cloud model constitutes the cloud probability feature value for risk level r. The following relationship must be satisfied:

[0081]

[0082]

[0083] In the formula, α j The contribution of the component score of the j-th index parameter of the energy storage battery is used as the weighting coefficient for calculating the weighted sum;

[0084] Step 2.4: Construct a risk cloud probability dataset using the cloud probability feature values ​​of each risk level.

[0085] During model training, traditional methods output data labels as discrete levels divided by fixed thresholds, such as "high / medium / low risk," which do not reflect the dynamic risk evolution of real-time data. This invention constructs a dynamic output dataset by combining expert scoring with a cloud model-based probability distribution analysis method. By using the risk cloud probability dataset as the model output training set, each sample corresponds to a continuous probability distribution (e.g., the probability of "high" risk level is 75%, "medium" is 20%, and "low" is 5%), rather than a single label, making the model output closer to the dynamic risk characteristics of energy storage systems.

[0086] This invention combines a multi-source data expert scoring mechanism with a risk cloud probabilistic model, solving the problems of traditional evaluation methods such as "separation of experience and data, insufficient handling of uncertainty, and rigid output." It constructs a risk assessment standard system that integrates domain knowledge, quantifies fuzziness, and supports dynamic decision-making. This improvement upgrades the model output from "simple classification" to "probabilistic risk prediction," significantly enhancing the scientific rigor, flexibility, and engineering practicality of energy storage battery safety evaluation, especially suitable for complex energy storage system scenarios with multi-parameter coupling and variable operating environments. The risk cloud probabilistic evaluation method transforms experts' fuzzy judgments on inconsistencies in multi-source data into a computable probability distribution through a cloud model.

[0087] Step 3: Construct an evaluation model based on the Transformer architecture using a multi-head attention mechanism; dynamically adjust the number of heads in the multi-head attention mechanism according to the cumulative contribution of component scores; use the comprehensive component scores as the input dataset for the evaluation model and the risk cloud probability dataset as the output dataset for the evaluation model; and train the model.

[0088] Traditional machine learning models (such as random forests and backpropagation neural networks) use fixed structures to process multi-source data, making it difficult to capture long-distance dependencies and hierarchical associations between features (such as the time-series correlation between temperature anomalies and subsequent internal resistance growth). Therefore, this invention introduces a Transformer architecture with a multi-head attention mechanism to process different feature subspaces of multi-source data in parallel, enhancing the model's ability to capture complex coupling relationships through multi-view feature extraction. Furthermore, by combining residual connections and layer normalization, it avoids the gradient vanishing problem in deep network training, supporting the construction of deeper models to mine deeper features of the data.

[0089] Specifically, satisfy and In this case, the number of heads in the multi-head attention mechanism is set to q, where Q is a threshold, and in this embodiment, Q = 0.8.

[0090] This invention proposes a method to determine the appropriate number of multi-head attention heads for the current data characteristics based on the cumulative contribution of component scores and its relationship with a threshold. Traditional attention models typically set the number of heads empirically, without establishing a correlation with the feature complexity of the input data, leading to insufficient or redundant heads. Therefore, the number of heads is dynamically adjusted based on the cumulative contribution of component scores calculated in step 1.

[0091] This invention features targeted designs for the model's input and output data. Traditional methods directly use raw data or simply processed labels (such as discrete risk levels), failing to fully utilize the dimensionality reduction features and probabilistic risk descriptions of multi-source data. The input data of this invention uses the component scores generated in step 1 as input, instead of the original 8-dimensional features, decoupling data features while retaining key information and solving the core problems of "partial data, scale confusion, and ambiguous correlations." The output data uses the risk cloud probability vector generated in step 2 as a supervision signal, instead of a single discrete label, forcing the model to learn the probabilistic mapping relationship between data and risk levels. Step 3, through innovative Transformer architecture, dynamic head number adjustment, and a combination of probabilistic data-driven improvements, solves the core problems of traditional models: "weak multi-source correlation modeling ability, rigid structure, and coarse output."

[0092] The constructed evaluation model includes: an input layer, a multi-head attention mechanism, residual connections and layer normalization, a feedforward neural network, further residual connections and layer normalization, and an output layer. The input layer receives a preprocessed input dataset Z; the multi-head attention mechanism transforms the preprocessed multi-source data into a matrix H suitable for model processing; the gradient vanishing problem is avoided through residual connections and layer normalization operations; the feedforward neural network further processes the features extracted by the multi-head attention mechanism; after further residual connections and layer normalization, the output of the current layer's Transformer is input to the next layer; the output layer determines the risk level of the energy storage system through linear transformation and the softmax function.

[0093] The evaluation model uses the comprehensive component score Z as the input dataset and compares it with the trainable parameter matrix W. q W k W v Multiply to generate multi-head attention matrices Q, K, and V; rearrange Q, K, and V according to the number of heads q in the multi-head attention; perform a scaled dot product on the rearranged Q and K, and use it as a scoring function to calculate the attention score; obtain the attention distribution through a probability distribution transformation function; multiply the attention distribution with V to obtain the attention output of each head; finally, concatenate the outputs of q attention heads and multiply them with the trainable parameter matrix W0 to obtain the final multi-head attention output H.

[0094] Existing machine learning models lack the ability to integrate multi-source data and mine deep features. This invention employs a multi-head attention mechanism, focusing on different feature subspaces of the data in parallel. Compared to traditional single-attention mechanisms, it can capture key information from multiple perspectives and accurately integrate data. Simultaneously, it adaptively adjusts the number of heads based on the cumulative contribution of component scores, dynamically optimizing according to data features. This enhances the model's ability to process complex data, improves its accuracy in capturing complex correlations between multi-source data, and ultimately improves the reliability of security assessments.

[0095] Step 4: With the optimal F1 score as the optimization objective, the hyperparameters of the trained evaluation model are optimized using a subtraction optimizer algorithm based on the golden sine.

[0096] The hyperparameters used to evaluate the model include: the number of heads q in the multi-head attention mechanism, the key and query dimensions d, and the hidden layer dimension D of the feedforward neural network. ff Model training learning rate L R Number of training input samples b, number of training rounds N a Regularized Dropout probability D R We employ a subtraction optimizer algorithm based on the golden sine wave to find the optimal combination of hyperparameters for the F1 score through iterative calculation.

[0097] The subtraction optimizer based on the golden sine curve is a data optimization method, an improvement on gradient optimization. Common gradient optimization methods use maximum gradient descent, which is prone to getting trapped in local optima. To address this issue, this invention applies the golden spiral to optimize the gradient optimization method. The golden spiral has the advantage of broad coverage, preventing data from getting trapped in local optima.

[0098] In the position update step of the subtraction optimizer based on the golden sine improvement, the search agent position update formula is:

[0099]

[0100] In the formula, X is the update position for the i-th search agent. zi Let ωt be the position of the i-th search agent, and sin(ωt) be the golden sine function. t is the current iteration number. Let N be the m-dimensional vector of the i-th search agent, whose norm ranges from [0,1] and follows a normal distribution. υ For special operators, it means:

[0101]

[0102] in, It is an m-dimensional vector whose elements are random numbers generated from the set {1,2}. The operation "*" represents the Hadamard product of two vectors (i.e., the components of the resulting vector are formed by multiplying the corresponding components of two given vectors). F(A) and F(B) are the objective function values ​​of search agents A and B, respectively, and sign(·) is the sign function.

[0103] Objective function for accepting the new position:

[0104]

[0105] Among them, F i new and F i These are search proxies and X zi The objective function value.

[0106] This invention constructs an evaluation model based on the Transformer architecture, utilizing residual connections and layer normalization to avoid gradient vanishing, and employing a feedforward neural network to deeply process features, thereby enhancing the model's expressive power. It uses a subtraction optimizer (SABO) algorithm based on a golden sine wave to optimize hyperparameters, using the F-score as a comprehensive evaluation metric. This balances global search and local exploitation, avoiding getting trapped in local optima, finding the optimal hyperparameter combination, enhancing model performance and reliability, and making safety evaluation more accurate and efficient.

[0107] Step 5: The hyperparameter-optimized evaluation model serves as the comprehensive safety evaluation model for multi-source data of the energy storage battery system.

[0108] This invention also proposes a system for constructing a multi-source data integrated safety evaluation model for energy storage battery systems, comprising: an input data processing module, an output data processing module, a model building module, and a model training module;

[0109] The input data processing module is used to acquire various index parameters of the energy storage battery; it uses multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component score of each index parameter after normalization; the comprehensive component score is the input dataset of the evaluation model.

[0110] The output data processing module is used to obtain the risk scores of various indicators of energy storage batteries, calculate the cloud probability feature values ​​of each risk level based on the cloud model, and form a risk cloud probability dataset; the risk cloud probability dataset is the output dataset of the evaluation model.

[0111] The model training module is used to construct an evaluation model based on a Transformer architecture with a multi-head attention mechanism; dynamically adjust the number of heads in the multi-head attention mechanism according to the cumulative contribution of component scores; train the model using the input and output datasets; optimize the hyperparameters of the trained evaluation model with the optimal F1 score as the optimization objective, using a subtraction optimizer algorithm based on a golden sine improvement; the hyperparameter-optimized evaluation model serves as a comprehensive safety evaluation model for multi-source data of energy storage battery systems.

[0112] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0113] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0114] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0115] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for constructing a multi-source data comprehensive safety evaluation model of an energy storage battery system, characterized in that, include: Obtain the various index parameters of the energy storage battery; use multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component scores of each index parameter after normalization. Obtain risk scores for various parameters of energy storage batteries, calculate cloud probability feature values ​​for each risk level based on the cloud model, and construct a risk cloud probability dataset. An evaluation model is constructed based on the Transformer architecture with a multi-head attention mechanism. The number of heads in the multi-head attention mechanism is dynamically adjusted according to the cumulative contribution of the component scores. The comprehensive component scores are used as the input dataset of the evaluation model, and the risk cloud probability dataset is used as the output dataset of the evaluation model. With the optimal F1 score as the optimization objective, the hyperparameters of the trained evaluation model are optimized using a subtraction optimizer algorithm based on the golden sine. The hyperparameter-optimized evaluation model serves as a comprehensive safety evaluation model for multi-source data of energy storage battery systems.

2. The method for constructing a multi-source data integrated safety evaluation model for an energy storage battery system according to claim 1, characterized in that, The parameters of energy storage batteries include: voltage, temperature, internal resistance, capacity, state of charge, state of health, state of energy, and state of power. After normalizing the various indicator parameters, the standardized indicator parameters are obtained as follows: wherein x ij is the jth index variable X j , i is the ith measurement value of the jth index variable X is the standardized value of x ij , s and s j are the corresponding intermediate variables of the jth index variable X j , n is the number of measurement groups of the jth index variable X j , and p is the number of index variables.

3. The method for constructing a multi-source data integrated safety evaluation model for energy storage battery systems according to claim 2, characterized in that, Calculate the covariance matrix of the standardized index parameters, and solve the covariance matrix to obtain the unit eigenvectors and eigenvalues; Using multivariate statistical analysis, and utilizing unit eigenvectors and eigenvalues, the cumulative contribution of the component scores and the comprehensive component scores of the standardized indicator parameters are calculated, including: The component scores are determined using the following relationship: wherein z ij is the score of the i-th component of the j-th index variable, a kj is the k-th element of the j-th unit eigenvector a j , is the i-th set of normalized values of the k-th index variable X k . The overall component scores are as follows: where Z is the composite component score, is the contribution of the component score, λ k is the kth eigenvalue.

4. The method for constructing a multi-source data integrated safety evaluation model for an energy storage battery system according to claim 1, characterized in that, Under risk level r, the j-th risk constraint range is constructed using the maximum and minimum risk scores of the j-th indicator parameter of the energy storage battery; the j-th set of overall characteristic parameters of the cloud model is calculated using the j-th risk constraint range, as follows: In the formula, E x,j , E n,j , H e,j are respectively the expectation, entropy and hyper entropy of the jth group of cloud models, Z′ j,min , Z′ j,max are respectively the minimum and maximum of the risk score of the jth index parameter of the energy storage battery. The weighted sum of the p groups of overall characteristic parameters of the cloud model under the risk level r constitutes the cloud probability characteristic value of the risk level r satisfies the following relationship: In the formula, α j is the component score contribution degree of the jth index parameter of the energy storage battery, which is used as a weight coefficient for calculating the weighted sum.

5. The method for constructing a multi-source data integrated safety evaluation model for an energy storage battery system according to claim 3, characterized in that, satisfy and if q < Q, then set the number of heads in the multi-head attention mechanism to q, where Q is a threshold value.

6. The method for constructing a multi-source data integrated safety evaluation model for an energy storage battery system according to claim 1, characterized in that, The evaluation model super parameters include: the number of heads q, the key and query dimension d, and the feedforward neural network hidden layer dimension D in the multi-head attention mechanism ff , the model training learning rate L R , the number of training input samples b, and the number of training rounds N a , the regularization Dropout probability D R An improved subtraction optimizer algorithm based on the golden sine is used to find the optimal super parameter combination that maximizes the F1 score through iterative calculation.

7. The method for constructing a multi-source data integrated safety evaluation model for an energy storage battery system according to claim 6, characterized in that, In the position update step of the subtraction optimizer based on the golden sine improvement, the search agent position update formula is: wherein, Xi is the updated position of the ith search agent, X zi Xi is the position of the ith search agent, sin(ωt) is a golden sine function, t is the current iteration number, Xi is the ith search agent m-dimensional vector, whose norm is a normal distribution with a value range of [0, 1], N is the number of search agents, υ is a special operator, which means: in, It is an m-dimensional vector whose components are random numbers generated from the set {1,2}. The operation "*" represents the Hadamard product of two vectors. F(A) and F(B) are the objective function values ​​of search agents A and B, respectively. sign(·) is the sign function. Objective function for accepting the new position: Among them, F i new and F i These are search proxies and X zi The objective function value.

8. A system for constructing a multi-source data integrated safety evaluation model for energy storage battery systems, characterized in that, include: Input data processing module, output data processing module, model training module; The input data processing module is used to acquire various index parameters of the energy storage battery; it uses multivariate statistical analysis to calculate the cumulative contribution of the component scores and the comprehensive component scores of each index parameter after normalization; the comprehensive component scores are the input dataset of the evaluation model. The output data processing module is used to obtain the risk scores of various indicators of energy storage batteries, calculate the cloud probability feature values ​​of each risk level based on the cloud model, and form a risk cloud probability dataset. The risk cloud probability dataset is the output dataset for evaluating the model; The model training module is used to construct an evaluation model based on a Transformer architecture with a multi-head attention mechanism; dynamically adjust the number of heads in the multi-head attention mechanism according to the cumulative contribution of component scores; train the model using the input and output datasets; optimize the hyperparameters of the trained evaluation model with the optimal F1 score as the optimization objective, using a subtraction optimizer algorithm based on a golden sine improvement; the hyperparameter-optimized evaluation model serves as a comprehensive safety evaluation model for multi-source data of energy storage battery systems.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.

Images