Energy storage equipment safety prediction management system based on historical data

Abstract

The invention discloses an energy storage equipment safety prediction management system based on historical data, and the system comprises an equipment data sensing module which is used for designing an electric energy data collection device on a branch between internal battery clusters of any energy storage equipment for collecting the electric energy data of the battery clusters; the battery state analysis module performs standardization processing on the electric energy data, and analyzes the electric energy data by using a pre-trained battery wear prediction neural network to obtain a battery wear state prediction result; and the load balancing management module dynamically selects the batteries to charge and discharge according to the wear state prediction result of each battery in the energy storage equipment. Intelligent and fine management of the energy storage equipment can be realized, the safety and stability of the energy storage system are improved, the operation and maintenance cost is reduced, and the service life of the battery is prolonged. Through the comprehensive management system, the problems of inaccurate data acquisition, untimely state analysis, unreasonable load distribution and the like in traditional energy storage equipment management are solved.

Description

Technical Field

[0001] This invention belongs to the field of energy storage management technology, specifically relating to a safety prediction and management system for energy storage devices based on historical data. Background Technology

[0002] As the global energy structure transitions towards renewable energy, energy storage devices are playing an increasingly important role in power system frequency regulation, peak shaving and valley filling, and emergency backup. However, energy storage batteries (such as lithium-ion batteries and flow batteries) are prone to capacity decay, increased internal resistance, and thermal runaway during long-term charge-discharge cycles, seriously threatening system safety and economic efficiency. Statistics show that in 2022, battery aging accounted for 63% of all energy storage system failures globally, resulting in direct economic losses exceeding $1.2 billion. Traditional energy storage device management methods mainly rely on the following technical approaches: Passive management based on threshold alarms: alarms are triggered by preset fixed thresholds for parameters such as voltage and temperature, but this cannot predict the gradual aging of the battery, and the threshold setting depends on empirical values, making it difficult to adapt to dynamic operating conditions.

[0003] Offline laboratory test modeling: Empirical life models (such as Arrhenius equations and power law models) are established through accelerated aging experiments. However, laboratory conditions differ significantly from the field operating environment, and the extrapolation error of the model can reach more than 30%.

[0004] Simple balancing strategy: adopts round-robin charging and discharging or static priority allocation, without considering the differences in the health status between batteries, which leads to some batteries being overloaded and aging faster.

[0005] The above method has three major flaws: Data dimension missing: Traditional voltage / temperature monitoring does not capture the energy characteristics of the battery cluster branch level (such as ripple coefficient and harmonic distortion rate), resulting in the missed detection of early aging characteristics; Analysis time lag: Periodic inspections based on weekly / monthly units cannot detect changes in battery status in real time, making it difficult to adjust management strategies in a timely manner; Insufficient intelligent decision-making: Static load allocation strategies exacerbate uneven losses among battery packs, shortening the overall lifespan of the system. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a safety prediction and management system for energy storage devices based on historical data, comprising: The device data sensing module is used to design power data acquisition devices on the branches between the battery clusters inside any energy storage device, and to collect power data of the battery clusters through the power data acquisition devices. The battery state analysis module is used to standardize the electrical energy data and analyze the electrical energy data using a pre-trained battery wear prediction neural network to obtain the battery wear state prediction results. The load balancing management module is used to dynamically select batteries for charging and discharging based on the predicted wear status of each battery in the energy storage device.

[0008] Preferably, the device data sensing module includes: The data acquisition unit is used to acquire the electrical energy data of each battery cell during the charging and discharging process at a high frequency through data acquisition devices pre-arranged on the branches between the battery clusters. The electrical energy data includes voltage data and current data. The data verification and synchronization unit is used to verify the collected power data in real time and ensure the consistency of data collected from multiple branches in the time dimension through a timestamp synchronization mechanism. The data preprocessing unit is used to perform data denoising, data cleaning, data smoothing, and normalization on the power data.

[0009] Preferably, the battery state analysis module performs the following operations: The power data is combined with the SOC data of the corresponding battery cell, and the SOC change rate data vector is determined based on the charging and discharging trend. The feature input of the model is generated by combining the power data feature mapping. Using the feature input and a pre-trained neural network model, a battery wear prediction neural network is used to predict the battery wear state. The battery wear prediction neural network is constructed in the following way: A massive amount of SOC data, energy data, and corresponding battery wear ratings for the same type of battery cell are pre-acquired. The massive amount of data is then standardized to obtain multiple sets of data mapped from SOC change rate data vectors to energy data features. These multiple sets of data are then divided into training and testing sets according to a preset ratio. Construct a neural network model with a cross-modal attention mechanism, set a time window length T in the input layer, and determine that the input tensor includes the electrical feature matrix. and SOC dynamic vector matrix The electrical characteristic matrix includes voltage and current. Time-series data of the electrical characteristics of the dimension, ,in, The SOC dynamic vector matrix represents the electrical characteristics at time step t and includes SOC data. Characteristics of the first derivative, ,in Represents the dynamic characteristics of SOC at time step t; Construct 1D depthwise separable convolutions in the feature extraction layer:

[0010] in, Indicates the first Layer depth can separate the output features of convolution. For activation function, This is a one-dimensional depthwise convolution operation. For the first The input features of the layer The pointwise convolution kernel parameter matrix, Input the number of channels. Number of output channels The depthwise convolution kernel parameter matrix, This is represented as the kernel width. For the first Layer bias term; Introduce a cross-modal attention mechanism in the feature fusion layer:

[0011]

[0012]

[0013] in, Output characteristics that represent electrical characteristics The output features represent the dynamic features of SOC. These represent the query matrix, key matrix, and value matrix, respectively. , , Represents the weight matrix. This represents the activation function used to transform attention scores into a probability distribution. Scaling factor The characteristics after fusion This represents the attention mechanism function. Indicates the feature fusion function; Prediction is performed in the fully connected layer:

[0014] in, Indicates the prediction result. This represents the output layer weight matrix. Indicates the output layer bias term; The neural network model is trained using training and testing sets. During training, the model is evaluated and optimized using a constructed loss function and optimizer, where the loss function is expressed as:

[0015] in For loss function, For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. Select the Adam optimizer to update parameters:

[0016] in, Indicates the first Step-by-step update of model parameters, Indicates the learning rate. This represents the gradient of the loss function with respect to the parameters; The training continues until the neural network model meets the preset requirements, resulting in a trained battery wear prediction neural network.

[0017] Preferably, the hidden layers of the battery wear prediction neural network embed an ARMA model based on time series analysis:

[0018] in, This represents the battery wear condition assessment result obtained after the battery cell has undergone j charge-discharge cycles. Represents the autoregressive parameters. Represents the moving average parameter. This represents the impact factor in the battery wear condition assessment process after the battery cell has undergone j charge-discharge cycles.

[0019] in, This represents the maximum remaining stored energy of the battery cell after j charge-discharge cycles. This represents the maximum remaining energy storage capacity that the battery will lose after the predicted number of additional charging cycles. This indicates the preset correction factor. This indicates the charge / discharge wear coefficient of the battery cell; when When the value is 0, the ARMA model is determined to be an nth-order autoregressive model, and the model variation is:

[0020] when When the value is 0, the ARMA model is determined to be an m-order moving average model, and the model variation is:

[0021] For the battery wear state assessment sequence generated by the model, the stationarity of the sequence is tested using the ADF test based on the observation results in the charge-discharge cycle. For non-stationary battery wear state assessment sequences, the difference is processed before testing. The model is optimized using the minimum information criterion method, and the autoregressive parameters and moving average parameters in the optimized model are estimated using the maximum likelihood estimation method. The model's fitness is determined by whether the residual sequence of the detection model is a white noise sequence. If it is fit, the inverse function prediction method is used to predict the model and obtain the battery wear status assessment result after the battery cell has undergone the predicted number of charge-discharge cycles.

[0022] Preferably, the feature input generation process includes: For any set of electrical energy data, determine the corresponding battery cell and the historical SOC data corresponding to that battery cell; Align the power data with the historical SOC data corresponding to each battery cell according to the time series. For historical SOC data, it is processed into a continuous SOC line chart based on time series changes, multiple inflection points in the line chart are marked, and the inflection points are traversed and judged to determine multiple vertices and valleys in the inflection points. Based on the SOC values ​​corresponding to multiple vertices and troughs in the SOC line graph, and according to the preset battery charge-discharge cycle judgment criteria, the SOC line graph is divided into stages to obtain multiple upward trend stages and multiple downward trend stages, and multiple charge-discharge cycle stages are also divided. For each charge-discharge cycle stage, multiple sub-stages are obtained by extracting it according to the multiple inflection points that exist within it; The charging and discharging rate of each sub-stage is used as a data vector. At the same time, the electrical energy features in the corresponding electrical energy data of that sub-stage are extracted, and a mapping relationship between the data vector and the electrical energy features is established to generate feature input.

[0023] Preferably, the preset battery charge / discharge cycle judgment criteria include: Sort the vertices and valleys in chronological order to obtain the vertex / valley sequence; For each point in the vertex / valley sequence, if the SOC value of the point is greater than the preset first threshold, it is determined to be the discharge start point; if the SOC value of the point is less than the preset second threshold, it is determined to be the charging start point. A charging phase is defined as the interval between a charging start point and the nearest discharging start point in the vertex / valley sequence, and a discharging phase is defined as the interval between a discharging start point and the nearest charging start point in the vertex / valley sequence. A charging phase and its adjacent corresponding discharging phase are considered as a charge-discharge cycle phase.

[0024] Preferably, the load balancing management module performs the following operations: Based on the predicted wear status of each battery in the energy storage device, each battery is arranged in order of increasing wear degree; Using the charge / discharge workload and total charge / discharge rate corresponding to the upcoming charge / discharge event as a benchmark, and with the goal of minimizing the total wear of all selected batteries for charge / discharge operations, a certain number of batteries are dynamically selected for charge / discharge operations. The number of batteries selected is determined based on the total charge / discharge rate and the rated power of the batteries in the charge / discharge event.

[0025] The present invention has achieved at least the following beneficial effects: 1. This invention enables intelligent and refined management of energy storage devices, improves the safety and stability of energy storage systems, reduces operation and maintenance costs, and extends battery life. Through a comprehensive management system, it solves problems such as inaccurate data collection, untimely status analysis, and unreasonable load allocation that exist in traditional energy storage device management.

[0026] 2. It can achieve high-precision prediction of battery wear status, thus providing accurate decision-making basis for the load balancing management module. Through comprehensive analysis methods, it can improve the intelligence level of battery management, extend battery life, reduce maintenance costs, and enhance the overall performance and reliability of the energy storage system.

[0027] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the structure of an energy storage device safety prediction and management system based on historical data in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the device data sensing module in an embodiment of the present invention; Figure 3 A flowchart illustrating the steps involved in generating feature inputs in this embodiment of the invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] This invention provides a safety prediction and management system for energy storage devices based on historical data, referring to... Figure 1 ,include: The device data sensing module is used to design power data acquisition devices on the branches between the battery clusters inside any energy storage device, and to collect power data of the battery clusters through the power data acquisition devices. The battery state analysis module is used to standardize the electrical energy data and analyze the electrical energy data using a pre-trained battery wear prediction neural network to obtain the battery wear state prediction results. The load balancing management module is used to dynamically select batteries for charging and discharging based on the predicted wear status of each battery in the energy storage device.

[0031] The working principle and beneficial effects of the above technical solution are as follows: By designing an energy data acquisition device on the branch lines between battery clusters within the energy storage device through a data sensing module, the energy data of the battery clusters can be collected in real time and accurately, including key parameters such as voltage and current. This enables comprehensive monitoring of the battery status. The battery status analysis module standardizes the collected energy data, eliminating dimensional and statistical differences between different data sources, making the data more suitable for the input requirements of the neural network model. A pre-trained battery wear prediction neural network analyzes the processed energy data, efficiently extracting feature information and predicting the battery wear status, reducing errors and delays that may occur in traditional prediction methods. The load balancing management module dynamically selects batteries for charging and discharging based on the predicted wear status of each battery, achieving intelligent management and optimized scheduling of the battery pack. This dynamic selection mechanism effectively avoids overuse or idleness of individual batteries, extending battery life and improving the efficiency and reliability of the entire energy storage system. This technical solution enables intelligent and refined management of energy storage devices, improves the safety and stability of the energy storage system, reduces operation and maintenance costs, and extends battery life. This comprehensive management system solves the problems of inaccurate data collection, untimely status analysis, and unreasonable load distribution that exist in the management of traditional energy storage equipment.

[0032] In a preferred embodiment, refer to Figure 2 The device data sensing module includes: The data acquisition unit is used to acquire the electrical energy data of each battery cell during the charging and discharging process at a high frequency through data acquisition devices pre-arranged on the branches between the battery clusters. The electrical energy data includes voltage data and current data. The data verification and synchronization unit is used to verify the collected power data in real time and ensure the consistency of data collected from multiple branches in the time dimension through a timestamp synchronization mechanism. The data preprocessing unit is used to perform data denoising, data cleaning, data smoothing, and normalization on the power data.

[0033] The working principle and beneficial effects of the above technical solution are as follows: The data acquisition unit collects electrical energy data, including voltage and current data, from each battery cell during charging and discharging at a high frequency, thereby achieving real-time and accurate monitoring of the battery status. High-frequency acquisition captures subtle changes in the battery under dynamic operating conditions. The data verification and synchronization unit verifies the collected electrical energy data in real time, using a timestamp synchronization mechanism to ensure consistency of data collected from multiple branches in the time dimension, improving data reliability while ensuring comparability and correlation between data from different branches. The data preprocessing unit performs noise reduction, cleaning, smoothing, and normalization on the electrical energy data, effectively eliminating outliers and noise interference, making the data smoother and more stable. Normalization transforms data of different dimensions and ranges to a unified scale, improving data comparability and model convergence speed. Through this technical solution, the equipment data sensing module can provide high-quality, high-reliability electrical energy data, providing a foundation for accurate analysis and prediction of battery wear prediction neural networks. The refined data processing flow not only improves the performance of the entire energy storage equipment safety prediction and management system, but also enhances the system's stability and prediction accuracy, thereby achieving efficient management and safe operation of energy storage equipment.

[0034] In a preferred embodiment, the battery state analysis module performs the following operations: The power data is combined with the SOC data of the corresponding battery cell, and the SOC change rate data vector is determined based on the charging and discharging trend. The feature input of the model is generated by combining the power data feature mapping. Using the feature input and a pre-trained neural network model, a battery wear prediction neural network is used to predict the battery wear state. The battery wear prediction neural network is constructed in the following way: A massive amount of SOC data, energy data, and corresponding battery wear ratings for the same type of battery cell are pre-acquired. The massive amount of data is then standardized to obtain multiple sets of data mapped from SOC change rate data vectors to energy data features. These multiple sets of data are then divided into training and testing sets according to a preset ratio. Construct a neural network model with a cross-modal attention mechanism, set a time window length T in the input layer, and determine that the input tensor includes the electrical feature matrix. and SOC dynamic vector matrix The electrical characteristic matrix includes voltage and current. Time-series data of the electrical characteristics of the dimension, ,in, The SOC dynamic vector matrix represents the electrical characteristics at time step t and includes SOC data. Characteristics of the first derivative, ,in Represents the dynamic characteristics of SOC at time step t; Construct 1D depthwise separable convolutions in the feature extraction layer:

[0035] in, Indicates the first Layer depth can separate the output features of convolution. For activation function, This is a one-dimensional depthwise convolution operation. For the first The input features of the layer The pointwise convolution kernel parameter matrix, Input the number of channels. Number of output channels The depthwise convolution kernel parameter matrix, This is represented as the kernel width. For the first Layer bias term; Introduce a cross-modal attention mechanism in the feature fusion layer:

[0036]

[0037]

[0038] in, Output characteristics that represent electrical characteristics The output features represent the dynamic features of SOC. These represent the query matrix, key matrix, and value matrix, respectively. , , Represents the weight matrix. This represents the activation function used to transform attention scores into a probability distribution. Scaling factor The characteristics after fusion This represents the attention mechanism function. Indicates the feature fusion function; Prediction is performed in the fully connected layer:

[0039] in, Indicates the prediction result. This represents the output layer weight matrix. Indicates the output layer bias term; The neural network model is trained using training and testing sets. During training, the model is evaluated and optimized using a constructed loss function and optimizer, where the loss function is expressed as:

[0040] in For loss function, For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. Select the Adam optimizer to update parameters:

[0041] in, Indicates the first Step-by-step update of model parameters, Indicates the learning rate. This represents the gradient of the loss function with respect to the parameters; The training continues until the neural network model meets the preset requirements, resulting in a trained battery wear prediction neural network.

[0042] The working principle and beneficial effects of the above technical solution are as follows: By combining electrical energy data with the SOC data of battery cells, the SOC change rate data vector is determined based on the charging and discharging trend, and the feature input of the model is generated using electrical energy data feature mapping. This allows for a comprehensive capture of the dynamic characteristics of the battery during the charging and discharging process. This feature generation method not only integrates electrical energy data and SOC data, but also enhances the expressive power of the data through feature mapping, providing a rich feature foundation for subsequent battery wear prediction. By pre-acquiring massive amounts of SOC data, electrical energy data, and their corresponding battery wear ratings for the same type of battery cells, and standardizing these data, high-quality training and testing sets are constructed. Large-scale data preprocessing and standardization operations ensure the consistency and reliability of the training data. By setting a time window length T at the input layer and determining that the input tensor includes an electrical feature matrix and a SOC dynamic vector matrix, both time-series data and dynamic features can be considered simultaneously. The electrical feature matrix covers time-series data of key electrical features such as voltage and current, while the SOC dynamic vector matrix contains the derivative features of the SOC data, enabling the model to simultaneously capture static electrical features and dynamic SOC changes, thereby providing a more comprehensive understanding of the battery state. The 1D depthwise separable convolution built in the feature extraction layer reduces computation and enhances the model's ability to capture local features through channel-wise convolution operations and activation functions. The combination of depthwise and pointwise convolution kernels allows the model to maintain efficient feature extraction and representation while reducing computational complexity. The cross-modal attention mechanism introduced in the feature fusion layer dynamically adjusts the weights of different modal features through the interaction of query, key, and value matrices. This attention mechanism automatically focuses on key features, ignoring irrelevant or redundant information, thereby improving the efficiency and accuracy of feature fusion. Activation functions and scaling factors optimize the calculation of attention scores, making the feature fusion process more stable and reliable. During prediction in the fully connected layer, the fused features are mapped to the prediction results through a linear combination of the output layer weight matrix and bias terms. This directly converts features into interpretable predicted values, giving the model's output a clear physical meaning. During training, the constructed loss function and Adam optimizer are used for evaluation and optimization. The loss function provides optimization direction for the model by calculating the difference between predicted and true values. The Adam optimizer accelerates the model convergence process and improves training efficiency through adaptive learning rate and gradient updates. This technical solution enables the battery state analysis module to achieve high-precision prediction of battery wear status, thus providing accurate decision-making basis for the load balancing management module. This comprehensive analysis method not only improves the intelligence level of battery management but also extends battery life, reduces maintenance costs, and enhances the overall performance and reliability of the energy storage system.

[0043] In a preferred embodiment, the hidden layer of the battery wear prediction neural network is embedded with an ARMA model based on time series analysis:

[0044] in, This represents the battery wear condition assessment result obtained after the battery cell has undergone j charge-discharge cycles. Represents the autoregressive parameters. Represents the moving average parameter. This represents the impact factor in the battery wear condition assessment process after the battery cell has undergone j charge-discharge cycles.

[0045] in, This represents the maximum remaining stored energy of the battery cell after j charge-discharge cycles. This represents the maximum remaining energy storage capacity that the battery will lose after the predicted number of additional charging cycles. This indicates the preset correction factor. This indicates the charge / discharge wear coefficient of the battery cell; when When the value is 0, the ARMA model is determined to be an nth-order autoregressive model, and the model variation is:

[0046] when When the value is 0, the ARMA model is determined to be an m-order moving average model, and the model variation is:

[0047] For the battery wear state assessment sequence generated by the model, the stationarity of the sequence is tested using the ADF test based on the observation results in the charge-discharge cycle. For non-stationary battery wear state assessment sequences, the difference is processed before testing. The model is optimized using the minimum information criterion method, and the autoregressive parameters and moving average parameters in the optimized model are estimated using the maximum likelihood estimation method. The model's fitness is determined by whether the residual sequence of the detection model is a white noise sequence. If it is fit, the inverse function prediction method is used to predict the model and obtain the battery wear status assessment result after the battery cell has undergone the predicted number of charge-discharge cycles.

[0048] The working principle and beneficial effects of the above technical solution are as follows: By embedding the ARMA model into the hidden layer of the battery wear prediction neural network, the advantages of time series analysis can be combined to efficiently predict the battery wear state. The ARMA model captures the time series characteristics of the battery wear state through autoregressive parameters and moving average parameters, thereby achieving dynamic modeling of the battery wear state. In this way, the ARMA model can effectively capture the time series characteristics of the battery wear state, providing richer feature information for the neural network. During the model construction process, the stationarity test of the battery wear state assessment sequence is performed using the ADF test to ensure that the model is suitable for stationary time series. For non-stationary sequences, differencing is used to make them stationary. The minimum information criterion method is used to optimize the model to ensure the optimal selection of model parameters. The maximum likelihood estimation method is used to estimate the autoregressive parameters and moving average parameters to further improve the prediction accuracy of the model. The model's suitability is judged by detecting whether the residual sequence of the model is a white noise sequence. When the model is suitable, the inverse function prediction method is used to predict the model, thereby obtaining the wear state assessment result of the battery cell after the predicted number of charge-discharge cycles. This neural network approach, which combines the ARMA model, can improve the accuracy of battery wear prediction, enhance the stability and reliability of the model, and provide support for the safety management and optimization of energy storage devices.

[0049] In a preferred embodiment, refer to Figure 3 The process of generating feature inputs includes: For any set of electrical energy data, determine the corresponding battery cell and the historical SOC data corresponding to that battery cell; Align the power data with the historical SOC data corresponding to each battery cell according to the time series. For historical SOC data, it is processed into a continuous SOC line chart based on time series changes, multiple inflection points in the line chart are marked, and the inflection points are traversed and judged to determine multiple vertices and valleys in the inflection points. Based on the SOC values ​​corresponding to multiple vertices and troughs in the SOC line graph, and according to the preset battery charge-discharge cycle judgment criteria, the SOC line graph is divided into stages to obtain multiple upward trend stages and multiple downward trend stages, and multiple charge-discharge cycle stages are also divided. For each charge-discharge cycle stage, multiple sub-stages are obtained by extracting it according to the multiple inflection points that exist within it; The charging and discharging rate of each sub-stage is used as a data vector. At the same time, the electrical energy features in the corresponding electrical energy data of that sub-stage are extracted, and a mapping relationship between the data vector and the electrical energy features is established to generate feature input.

[0050] The working principle and beneficial effects of the above technical solution are as follows: By aligning the electrical energy data with the historical SOC data of the corresponding battery cells over time, data synchronization is ensured, providing an accurate time reference for subsequent analysis, improving data correlation, and providing a foundation for feature extraction. By processing the historical SOC data into a continuous SOC line graph and marking multiple inflection points, the dynamic changes of the battery during charging and discharging can be clearly displayed. Traversing and judging the inflection points further identifies vertices and troughs, facilitating the understanding of the battery's charging and discharging behavior. Based on preset battery charging and discharging cycle judgment criteria, the SOC line graph is divided into stages, accurately identifying multiple rising and falling trend stages, thus dividing the complete charging and discharging cycle stages. This division method helps to understand the battery's charging and discharging patterns, providing a clear stage division for subsequent feature extraction. Each charging and discharging cycle stage is truncated according to its internal inflection points, resulting in multiple sub-stages, making the analysis more detailed. The charging and discharging rate of each sub-stage is used as a data vector, and the corresponding electrical energy features are extracted from the electrical energy data, establishing a mapping relationship between the data vector and the electrical energy features, thereby generating feature input. This feature input contains key information about charge and discharge rates and integrates electrical energy characteristics, providing a rich feature base for battery wear prediction. Through this technical solution, the feature input generation process can capture the key dynamic characteristics of the battery during charge and discharge, providing high-quality input data for the battery wear prediction model. This refined feature extraction method not only improves the model's prediction accuracy but also enhances its sensitivity to changes in battery state, thereby achieving accurate assessment and prediction of battery wear status.

[0051] In a preferred embodiment, the preset battery charge / discharge cycle judgment criteria include: Sort the vertices and valleys in chronological order to obtain the vertex / valley sequence; For each point in the vertex / valley sequence, if the SOC value of the point is greater than the preset first threshold, it is determined to be the discharge start point; if the SOC value of the point is less than the preset second threshold, it is determined to be the charging start point. A charging phase is defined as the interval between a charging start point and the nearest discharging start point in the vertex / valley sequence, and a discharging phase is defined as the interval between a discharging start point and the nearest charging start point in the vertex / valley sequence. A charging phase and its adjacent corresponding discharging phase are considered as a charge-discharge cycle phase.

[0052] The working principle and beneficial effects of the above technical solution are as follows: By arranging the vertices and valleys in chronological order to form a vertex / valley sequence, the key turning points in the battery's charging and discharging process can be clearly displayed, providing a clear time series basis for subsequent stage division. This sorting method helps identify key events in the charging and discharging process. By judging the SOC value of each point in the vertex / valley sequence and using preset first and second thresholds to determine the discharge start and charging start, the start and end of the charging and discharging process can be accurately identified. The threshold-based judgment method is not only simple and efficient, but also effectively reduces misjudgments and ensures the accuracy of stage division. By defining the charging stage as from the charging start to the nearest discharge start, and the discharging stage as from the discharge start to the nearest charging start, the battery's charging and discharging process can be clearly divided. This division method conforms to the actual working process of the battery and provides clear boundaries for the definition of subsequent charge-discharge cycle stages. By combining a charging stage and its adjacent discharge stage into a charge-discharge cycle stage, a complete charge-discharge cycle of the battery can be described. This covers all dynamic changes of the battery in a complete cycle, providing accurate cycle division for subsequent battery wear prediction. This technical solution enables precise division of battery charge-discharge cycles, providing high-quality phased data for battery wear prediction models, improving the model's sensitivity to changes in battery state, and enhancing the model's prediction accuracy and reliability, thereby achieving accurate assessment and prediction of battery wear state.

[0053] In a preferred embodiment, the load balancing management module performs the following operations: Based on the predicted wear status of each battery in the energy storage device, each battery is arranged in order of increasing wear degree; Using the charge / discharge workload and total charge / discharge rate corresponding to the upcoming charge / discharge event as a benchmark, and with the goal of minimizing the total wear of all selected batteries for charge / discharge operations, a certain number of batteries are dynamically selected for charge / discharge operations. The number of batteries selected is determined based on the total charge / discharge rate and the rated power of the batteries in the charge / discharge event.

[0054] The working principle and beneficial effects of the above technical solution are as follows: By arranging each battery in ascending order of wear level, the health status of the batteries can be intuitively displayed, providing a clear priority for subsequent battery selection. This sorting method helps identify which batteries are more suitable for charging and discharging, providing a foundation for achieving the optimization goal. Based on the charging and discharging workload and total charging and discharging rate corresponding to the charging and discharging event, combined with the optimization goal—minimizing the sum of wear levels of all selected batteries—a certain number of batteries are dynamically selected for charging and discharging. The selection strategy considers the needs of the current charging and discharging task, while also taking into account the long-term health status of the batteries, avoiding overuse of individual batteries. The number of batteries selected is determined by calculating the relationship between the total charging and discharging rate and the rated power of the battery, ensuring the efficiency and reliability of the charging and discharging operation. This not only ensures the smooth completion of the charging and discharging task but also optimizes the overall performance of the battery pack. Through this technical solution, the load balancing management module can realize intelligent management and optimized scheduling of the battery pack. This dynamic selection mechanism based on wear status not only extends the service life of the batteries but also improves the overall efficiency and reliability of the energy storage system, thereby achieving safe, stable, and efficient operation of the energy storage equipment.

[0055] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A safety prediction and management system for energy storage devices based on historical data, characterized in that, include: The device data sensing module is used to collect the power data of any energy storage device by designing power data acquisition devices on the branches between the battery clusters inside the device. The battery state analysis module is used to standardize the electrical energy data and analyze the electrical energy data using a pre-trained battery wear prediction neural network to obtain the battery wear state prediction results. The load balancing management module is used to dynamically select batteries for charging and discharging based on the predicted wear status of each battery in the energy storage device.

2. The energy storage device safety prediction and management system based on historical data according to claim 1, characterized in that, The device data sensing module includes: The data acquisition unit is used to acquire the electrical energy data of each battery cell during the charging and discharging process at a high frequency through data acquisition devices pre-arranged on the branches between the battery clusters. The electrical energy data includes voltage data and current data. The data verification and synchronization unit is used to verify the collected power data in real time and ensure the consistency of data collected from multiple branches in the time dimension through a timestamp synchronization mechanism. The data preprocessing unit is used to perform data denoising, data cleaning, data smoothing, and normalization on the power data.

3. The energy storage device safety prediction and management system based on historical data according to claim 1, characterized in that, The battery status analysis module performs the following operations: The power data is combined with the SOC data of the corresponding battery cell, and the SOC change rate data vector is determined based on the charging and discharging trend. The feature input of the model is generated by combining the power data feature mapping. Using the feature input and a pre-trained neural network model, a battery wear prediction neural network is used to predict the battery wear state. The battery wear prediction neural network is constructed in the following way: A massive amount of SOC data, energy data, and corresponding battery wear ratings for the same type of battery cell are pre-acquired. The massive amount of data is then standardized to obtain multiple sets of data mapped from SOC change rate data vectors to energy data features. These multiple sets of data are then divided into training and testing sets according to a preset ratio. Construct a neural network model with a cross-modal attention mechanism, set a time window length T in the input layer, and determine that the input tensor includes the electrical feature matrix. and SOC dynamic vector matrix The electrical characteristic matrix includes voltage and current. Time-series data of the electrical characteristics of the dimension, ,in, The SOC dynamic vector matrix represents the electrical characteristics at time step t and includes SOC data. Characteristics of the first derivative, ,in Represents the dynamic characteristics of SOC at time step t; Construct 1D depthwise separable convolutions in the feature extraction layer: in, Indicates the first Layer depth can separate the output features of convolution. For activation function, This is a one-dimensional depthwise convolution operation. For the first The input features of the layer The pointwise convolution kernel parameter matrix, Input the number of channels. Number of output channels The depthwise convolution kernel parameter matrix, This is represented as the kernel width. For the first Layer bias term; Introduce a cross-modal attention mechanism in the feature fusion layer: in, Output characteristics that represent electrical characteristics The output features represent the dynamic features of SOC. These represent the query matrix, key matrix, and value matrix, respectively. , , Represents the weight matrix. This represents the activation function used to transform attention scores into a probability distribution. Scaling factor The characteristics after fusion This represents the attention mechanism function. Indicates the feature fusion function; Prediction is performed in the fully connected layer: in, Indicates the prediction result. This represents the output layer weight matrix. Indicates the output layer bias term; The neural network model is trained using training and testing sets. During training, the model is evaluated and optimized using a constructed loss function and optimizer, where the loss function is expressed as: in For loss function, For the sample size, Let be the predicted value for the i-th sample. This represents the true value of the i-th sample. Select the Adam optimizer to update parameters: in, Indicates the first Step-by-step update of model parameters, Indicates the learning rate. This represents the gradient of the loss function with respect to the parameters; The training continues until the neural network model meets the preset requirements, resulting in a trained battery wear prediction neural network.

4. The energy storage device safety prediction and management system based on historical data according to claim 3, characterized in that, The hidden layers of the battery wear prediction neural network embed an ARMA model based on time series analysis: in, This represents the battery wear condition assessment result obtained after the battery cell has undergone j charge-discharge cycles. Represents the autoregressive parameters. Represents the moving average parameter. This represents the impact factor in the battery wear condition assessment process after the battery cell has undergone j charge-discharge cycles. in, This represents the maximum remaining stored energy of the battery cell after j charge-discharge cycles. This represents the maximum remaining energy storage capacity that the battery will lose after the predicted number of additional charging cycles. This indicates the preset correction factor. This indicates the charge / discharge wear coefficient of the battery cell; when When the value is 0, the ARMA model is determined to be an nth-order autoregressive model, and the model variation is: when When the value is 0, the ARMA model is determined to be an m-order moving average model, and the model variation is: For the battery wear state assessment sequence generated by the model, the stationarity of the sequence is tested using the ADF test based on the observation results in the charge-discharge cycle. For non-stationary battery wear state assessment sequences, the difference is processed before testing. The model is optimized using the minimum information criterion method, and the autoregressive parameters and moving average parameters in the optimized model are estimated using the maximum likelihood estimation method. The model's fitness is determined by whether the residual sequence of the detection model is a white noise sequence. If it is fit, the inverse function prediction method is used to predict the model and obtain the battery wear status assessment result after the battery cell has undergone the predicted number of charge-discharge cycles.

5. The energy storage device safety prediction and management system based on historical data according to claim 3, characterized in that, The process of generating feature inputs includes: For any set of electrical energy data, determine the corresponding battery cell and the historical SOC data corresponding to that battery cell; Align the power data with the historical SOC data corresponding to each battery cell according to the time series. For historical SOC data, it is processed into a continuous SOC line chart based on time series changes, multiple inflection points in the line chart are marked, and the inflection points are traversed and judged to determine multiple vertices and valleys in the inflection points. Based on the SOC values ​​corresponding to multiple vertices and troughs in the SOC line graph, and according to the preset battery charge-discharge cycle judgment criteria, the SOC line graph is divided into stages to obtain multiple upward trend stages and multiple downward trend stages, and multiple charge-discharge cycle stages are also divided. For each charge-discharge cycle stage, multiple sub-stages are obtained by extracting it according to the multiple inflection points that exist within it; The charging and discharging rate of each sub-stage is used as a data vector. At the same time, the electrical energy features in the corresponding electrical energy data of that sub-stage are extracted, and a mapping relationship between the data vector and the electrical energy features is established to generate feature input.

6. The energy storage device safety prediction and management system based on historical data according to claim 5, characterized in that, The preset battery charge / discharge cycle judgment criteria include: Sort the vertices and valleys in chronological order to obtain the vertex / valley sequence; For each point in the vertex / valley sequence, if the SOC value of the point is greater than the preset first threshold, it is determined to be the discharge start point; if the SOC value of the point is less than the preset second threshold, it is determined to be the charging start point. A charging phase is defined as the interval between a charging start point and the nearest discharging start point in the vertex / valley sequence, and a discharging phase is defined as the interval between a discharging start point and the nearest charging start point in the vertex / valley sequence. A charging phase and its adjacent corresponding discharging phase are considered as a charge-discharge cycle phase.

7. The energy storage equipment safety prediction and management system based on historical data according to claim 1, characterized in that, The load balancing management module performs the following operations: Based on the predicted wear status of each battery in the energy storage device, each battery is arranged in order of increasing wear degree; Using the charge / discharge workload and total charge / discharge rate corresponding to the upcoming charge / discharge event as a benchmark, and with the goal of minimizing the total wear of all selected batteries for charge / discharge operations, a certain number of batteries are dynamically selected for charge / discharge operations. The number of batteries selected is determined based on the total charge / discharge rate and the rated power of the batteries in the charge / discharge event.

Images