Method for predicting health state of energy storage battery

Abstract

The invention designs a method for predicting the health state of an energy storage battery, and the method comprises the steps: S1, carrying out the collection and monitoring of the charging and discharging process of an operation battery, and collecting and storing the charging and discharging capacity data of each cycle of the battery; s2, preprocessing the collected discharge capacity data of the batteries, calculating battery health state data, constructing a time sequence sample by using a sliding window mechanism, and dividing the battery health state data of each battery into an input sequence and an output sequence; s3, constructing a hierarchical neural network architecture and training the hierarchical neural network architecture; s4, performing Bayesian optimization on the hierarchical neural network architecture based on an Optuna framework; and S5, predicting the energy storage battery by using the optimized neural network architecture to obtain the health state of the battery. Through the above design, the prediction precision of the battery health state can be improved, and the stability and robustness of overall prediction are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion power batteries, specifically relating to a method for predicting the health status of energy storage batteries. Background Technology

[0002] Common methods for predicting battery state of health (SOH) mainly fall into three categories: model-based methods, traditional machine learning-based methods, and deep learning-based methods. Model-based methods rely on precise modeling of the battery's internal mechanisms, requiring complex mathematical descriptions and a large number of experimental parameters, making them difficult to generalize across different operating conditions or batches of batteries. Traditional machine learning methods, such as Support Vector Machines (SVM) and Gaussian Process Regression (GPR), have weak robustness to interference and limited ability to fit complex nonlinear degradation processes. Deep learning methods are highly sensitive to hyperparameter selection and can automatically extract time-series features, capturing the dynamic changes in battery degradation.

[0003] However, statistical learning models such as support vector machines and dual Kalman filtering algorithms have high requirements for battery models, do not consider external environmental bias noise, and are prone to noise accumulation under certain conditions, resulting in weak anti-interference ability.

[0004] RNN-based prediction schemes can improve the generalization ability of battery models, often outperforming statistical learning models. The state of equilibrium (SOH) of a battery is a slowly decaying process, its changes depending on the long-term usage history throughout the battery's lifespan (e.g., cumulative charge / discharge cycles, depth, temperature history, etc.). Traditional RNNs suffer from vanishing / exploding gradient problems, making it difficult to effectively capture these long-term dependencies. While schemes based solely on long short-term memory networks can improve the model's long-term dependencies, they cannot reasonably distinguish the importance of information across different time spans and struggle to understand the deep features inherent in the original data.

[0005] Therefore, the model's performance is highly sensitive to hyperparameters (such as hidden layer size, learning rate, number of hidden layers, time step, etc.). Traditional manual or grid search hyperparameter tuning is inefficient and makes it difficult to guarantee finding the optimal parameter combination, affecting the model's generalization ability on different battery samples. Summary of the Invention

[0006] To address the challenges in battery SOH prediction, such as high noise in raw data, insufficient modeling of long-term dependencies, inadequate acquisition of deep correlation information in battery cycling data, unreasonable allocation of key information and weights, and difficulty in efficiently optimizing model hyperparameters, this application designs a method for predicting the health status of energy storage batteries. This method aims to improve prediction accuracy and model robustness by fully exploring the deep features of battery cycling time series and allocating attention to the correlations across different time spans.

[0007] A method for predicting the health status of energy storage batteries includes the following steps: Step S1: Collect and monitor the charging and discharging process of the battery, and collect and save the charging and discharging capacity data of the battery for each cycle; Step S2: Preprocess the collected battery discharge capacity data, calculate the battery health status data, and construct a time series sample using a sliding window mechanism to divide the battery health status data of each battery into an input sequence and an output sequence. Step S3: Construct a hierarchical neural network architecture and train it; Step S4: Perform Bayesian optimization on the hierarchical neural network architecture based on the Optuna framework to automatically search and adjust key network parameters so that the model can achieve the best performance under different datasets and tasks. Step S5: Use the optimized neural network architecture to predict the energy storage battery and obtain its battery health status.

[0008] Preferably, in step S2, during the preprocessing stage of the battery discharge capacity data, the original capacity data file of the battery is imported, and the ratio of the current capacity to the initial capacity is calculated and converted into health status data to characterize the battery performance degradation in a standardized form.

[0009] Preferably, in step S2, the method for preprocessing the collected battery discharge capacity data includes: Step S21: Use enhanced data noise reduction processing on the collected battery discharge capacity data to remove noise and retain the decay trend; Step S22: Initially eliminate spike noise through median filtering, and then perform deep smoothing by multiple Savitzky-Golay smoothing filters, dynamically adjusting the window size to adapt to curves of different lengths and avoid trend distortion. Step S23: Next, calculate the State of Health (SOH) data of the battery after noise reduction based on the battery discharge capacity data; calculate the ratio of the discharge capacity of each battery cycle to the initial rated capacity of the battery to reflect the degree to which the battery performance is maintained relative to the initial state. The calculation formula is: SOH = Current capacity / Initial capacity * 100%; Step S24: Perform global normalization on the SOH sequences of all battery samples using MinMaxScaler, and linearly map them to... Within this range, the model can better learn the relative change patterns of different samples; Step S25: Construct time series samples using a sliding window mechanism; divide the SOH data of each battery into input and output sequences to construct training and test datasets.

[0010] Preferably, in step S25, the input sequence consists of historical SOH data with a length equal to the input step size; the output sequence is the SOH prediction target with the corresponding future output step size.

[0011] Preferably, in step S3, the hierarchical neural network architecture is constructed by sequentially connecting a convolutional neural network, a long short-term memory network, a multi-head attention mechanism, and an output layer.

[0012] Preferably, in step S3, the training method for the hierarchical neural network architecture includes: For the SOH sequence prediction task, the mean squared error loss function is used as the optimization objective to measure the difference between the model's predicted values ​​and the true values, thereby guiding the update of network parameters. To further prevent the model from overfitting on the training set while ensuring the efficiency of the training process, an early stopping mechanism is introduced. By monitoring the changing trend of training loss or verification loss, training is automatically terminated when no significant improvement is observed for several consecutive rounds.

[0013] Preferably, in step S4, the objective function for Bayesian optimization of the hierarchical neural network architecture based on the Optuna framework is the average value of the model across all test samples. Fractions enable high accuracy and strong generalization ability in SOH sequence prediction.

[0014] Preferably, in step S4, the optimal combination of hyperparameters of the CLMA model is automatically searched using the Bayesian optimization algorithm of the Optuna framework. The key parameters include time step, hidden layer dimension, number of attention heads, and learning rate. After obtaining the optimal hyperparameters, model shaping and prediction are performed. The CLMA model was reinitialized with the optimized parameters and trained using the entire training dataset. After training, predictions were made using the test dataset, and the prediction results were denormalized to restore the original SOH dimensions.

[0015] The advantages and effects of this application are as follows: (1) The method for predicting the health status of energy storage batteries designed in this application combines enhanced data denoising processing and CLMA hybrid network structure. The designed hierarchical neural network architecture model can fully eliminate noise interference in the original capacity or battery health status SOH data, and can fully explore the multi-level features of the battery SOH time series. (2) The method for predicting the health status of energy storage batteries designed in this application uses a convolutional neural network (CNN), which extracts local features of the input sequence through nonlinear transformation to enhance the feature representation capability; among them, the long short-term memory network (LSTM) can capture long-term dependencies and cyclic degradation trends in the sequence; while the multi-head attention mechanism weights the importance of different time steps in the historical sequence, so that the model can comprehensively consider the influence of each stage when predicting the future SOH decay trajectory, thereby significantly improving the accuracy of SOH prediction.

[0016] (3) Regarding data robustness, the method for predicting the health status of energy storage batteries designed in this application uses mean value filtering to effectively suppress spike noise, while multiple Savitzky-Golay smoothing processes deeply remove high-frequency fluctuations while preserving the true degradation trend of the battery, making the input sequence smoother and cleaner. With this high-quality data input, the model can learn the battery degradation law under reduced noise interference, thereby enhancing the overall prediction stability and robustness, and maintaining good performance even in the face of measurement errors or abnormal cyclic data.

[0017] (4) Regarding the efficiency of hyperparameter optimization, the method for predicting the health status of energy storage batteries designed in this application adopts the Optuna framework for Bayesian optimization, replacing the traditional manual parameter tuning or grid search method. Through an automated and intelligent search strategy, it can quickly find the optimal combination in the high-dimensional hyperparameter space, including key parameters such as time step, batch size, hidden layer dimension, number of attention heads, and learning rate, thereby ensuring the generalization ability and prediction performance of the model on different battery samples, while significantly saving experimental time and labor costs.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0019] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0021] Figure 1 A flowchart of a method for predicting the health status of energy storage batteries designed for this application; Figure 2 Comparison chart of SOH degradation trajectory prediction for battery B0005; Figure 3 Comparison chart of SOH degradation trajectory prediction for battery B0006; Figure 4 A comparison chart showing the predicted SOH degradation trajectory of battery B0007. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0023] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0025] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0026] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0027] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0028] Example 1: This invention proposes a neural network scheme for battery state of health (SOH) estimation. Its components include a convolutional neural network (CNN) layer for feature extraction, a long short-term memory (LSTM) layer for capturing the time dependence and long-term degradation trends in battery cycling data, a multi-head attention layer for attention allocation between short-term and long-term information, and enhanced data denoising and Bayesian hyperparameter search. The core of this approach lies in fully mining the deep features of battery cycling time series data and allocating attention to the correlations across different time spans to improve prediction accuracy and model robustness.

[0029] Specific methods include: 1. Data Collection In a laboratory environment, each battery undergoes standardized cycle discharge testing. The charging and discharging process is run under controlled conditions of constant temperature and humidity using precision testing equipment. During each cycle, key operating parameters are monitored and recorded in real time, with a focus on collecting and storing charge and discharge capacity data for each cycle to reflect the performance degradation patterns of the battery over long-term use. This discharge capacity data will serve as the core input features for subsequent State of Health (SOH) estimation models, used to analyze battery degradation trends and predict remaining life.

[0030] 2. Data Preprocessing First, enhanced data denoising (using the `enhanced_denoising` function) was applied to the battery discharge capacity data collected in the experiment to remove noise and preserve the decay trend. Median filtering (kernel_size=5) was used to initially eliminate spike noise, followed by multiple Savitzky-Golay smoothing filters for deep smoothing. The window size was dynamically adjusted to accommodate curves of different lengths, avoiding trend distortion. During processing, iterative optimization (default 50 iterations) was employed. Initially, a larger window was used to smooth the overall trend, while later the window was gradually reduced to optimize local details, thus preserving the true decay characteristics to the greatest extent possible while removing noise. Finally, the denoised data was saved to provide high-quality data for subsequent model training.

[0031] Next, the State of Discharge (SOH) is calculated based on the noise-reduced battery discharge capacity data. Specifically, the discharge capacity of each cycle is compared to the battery's initial rated capacity to reflect the degree to which the battery performance is maintained relative to its initial state. The calculation formula is as follows:

[0032] Subsequently, to eliminate the dimensional differences between different battery samples and improve the training stability and convergence speed of the model, the SOH sequences of all battery samples were globally normalized using MinMaxScaler, linearly mapping them to... Within this range, the model can better learn the relative change patterns of different samples.

[0033] Finally, a sliding window mechanism is used to construct time series samples. The SOH data for each battery is divided into an input sequence and an output sequence. The input sequence consists of historical SOH data of length *timestep*, providing contextual information for the model; the output sequence is the SOH prediction target for the corresponding future output step size. By dividing the data in this way, the model can learn the mapping relationship between historical degradation patterns and future health status in the time series, thereby achieving accurate prediction of future battery performance changes.

[0034] 3. Model Building This invention proposes a hierarchical neural network architecture (called CLMA, or Convolutional-LSTM-Multihead Attention hybrid network) to fully utilize feature representations at different levels in time series data to achieve high-precision prediction of SOH. The overall network consists of a convolutional neural network (CNN), a long short-term memory network (LSTM), a multi-head attention mechanism, and an output layer (Linear Layer) connected in series. The functions of each module are described below: A Convolutional Neural Network (CNN) consists of multiple convolutional and pooling layers. Its feature extraction capability is jointly controlled by key parameters such as kernel size, number of filters, pooling window size, and stride. The original time-series input undergoes local perception and feature mapping through convolutional layers and the ReLU activation function, followed by feature dimensionality reduction through pooling layers, ultimately outputting a high-dimensional representation with temporal local patterns. In this way, CNNs can perform nonlinear feature transformation and feature extraction on the time series before it enters the LSTM, effectively capturing local dependencies and providing more discriminative feature representations for subsequent temporal modeling.

[0035] Long Short-Term Memory (LSTM) networks are used to capture dynamic dependencies and long-term degradation trends in SOH time series. Key parameters include the hidden layer dimension (hidden_size), which controls the capacity of information representation between time steps, and the number of stacked LSTM layers (num_layers), which enhances the model's ability to model complex temporal dependencies. The hidden state sequence output by the LSTM provides rich temporal feature representations for subsequent attention mechanisms.

[0036] Multi-Head Attention takes the output hidden state sequence of an LSTM as input. This module uses multiple parallel attention heads (controlled by the parameter `num_heads`) to perform weighted calculations on different positions of the input sequence, allowing the model to simultaneously learn dependencies between different time steps across multiple representation subspaces. This mechanism effectively alleviates the problem of LSTM's insufficient ability to capture long-range dependencies in long-sequence prediction, enabling the model to dynamically focus on feature changes at key moments when dealing with long-term degradation processes.

[0037] The output layer (Linear Layer) receives the output sequence from the multi-head attention mechanism and extracts the hidden state vector of its last time step as the final global temporal representation. This layer compresses the high-dimensional features of the hidden_size dimension to the target prediction dimension through a fully connected mapping, thereby generating the SOH prediction sequence for several future steps.

[0038] 4. Model Training First, the core metrics and loss function for training were defined. For the SOH sequence prediction task, the mean squared error loss function (MSELoss) was used as the optimization objective to measure the difference between the model's predicted values ​​and the true values, thereby guiding the update of network parameters. To improve training efficiency and stability, the AdamW optimizer was selected. This optimizer combines Adam's adaptive gradient advantage with the weight decay strategy, which helps to alleviate overfitting and maintain parameter regularization.

[0039] To further prevent overfitting on the training set while ensuring efficient training, an early stopping mechanism is introduced. This mechanism monitors the training loss or the trend of the validation loss and automatically terminates training if no significant improvement is observed for several consecutive rounds (controlled by the patience parameter). The minimum improvement threshold is controlled by the delta parameter, ensuring that the model only considers training to have improved when the loss decreases beyond the minimum improvement threshold, thus avoiding premature termination of training due to noise fluctuations.

[0040] 5. Model parameter tuning In terms of hyperparameter optimization, this model uses Bayesian optimization based on the Optuna framework to automatically search for and adjust key network parameters, enabling the model to achieve optimal performance on different datasets and tasks.

[0041] The optimized objective function is the average of the model across all test samples. The score, an metric that comprehensively reflects the degree of fit between predicted and actual values ​​and the model's interpretability, is used to analyze the search space, which encompasses several key hyperparameters of the model structure and training process, including timestep, batch size, and model tuning. In terms of hyperparameter optimization, this model uses Bayesian optimization based on the Optuna framework to automatically search for and adjust key network parameters, enabling the model to achieve optimal performance on different datasets and tasks.

[0042] The optimization objective function includes the model's average batch size across all test samples, the LSTM hidden layer dimension, the number of multi-head attention heads, the number of LSTM layers, the learning rate, the number of training epochs, the early stopping tolerance, the minimum improvement threshold (delta), the kernel size, the number of filters, the pooling window size, and the stride. Through systematic Bayesian optimization, the model can quickly converge to the optimal solution in the high-dimensional hyperparameter space, thereby achieving high accuracy and strong generalization ability in SOH sequence prediction.

[0043] For the results of this invention applied to the dataset BatteryAgingARC-FY08Q4 published by NASA, please refer to [the documentation / reference needed]. Figures 2-4 , Figure 2 This is a comparison chart showing the predicted SOH (Surface Offset Hull) degradation trajectory of battery B0005. Figure 3 This is a comparison chart showing the predicted SOH (State of Health) degradation trajectory of battery B0006. Figure 4 This is a comparison chart of the predicted SOH degradation trajectory for battery B0007; the solid blue line represents the actual SOH curve of each battery group, the dashed red line represents the predicted value of the model, and the dashed gray line represents the boundary between the training set and the test set.

[0044] Depend on Figure 2 As can be seen, after battery B0005 entered the test set phase, its State of Harmony (SOH) exhibited a non-linear decreasing trend, accompanied by multiple capacity rebounds due to resting (i.e., local peaks on the curve). Nevertheless, the red predicted curve closely followed the changing trend of the blue true curve, not only maintaining consistency in the overall decreasing slope but also keenly capturing the capacity rebound phenomenon, indicating that the model has extremely high prediction accuracy for battery B0005.

[0045] Depend on Figure 3 It is evident that the degradation rate of battery B0006 is relatively fast, yet the model still successfully captures the overall degradation direction of SOH and key local abrupt change points (such as the violent fluctuations near the 90th cycle). After the 100th cycle, the predicted curve and the actual curve once again highly overlap, demonstrating that the model has good self-correction ability and robustness in long-term prediction.

[0046] Depend on Figure 4As can be seen, the prediction performance for battery B0007 is the best. Whether in the stable degradation period or during the period of sharp capacity recovery fluctuations (e.g., around cycles 90 and 150), the red prediction curve almost completely overlaps with the blue actual curve. This further verifies that the method described in this invention has excellent generalization ability and stability on different battery cells.

[0047] Furthermore, in the data loading and SOH conversion stage, the original capacity data file is imported, and the ratio of the current capacity to the initial capacity is calculated to convert it into state of health (SOH,%) data, which characterizes battery performance degradation in a standardized form. Subsequently, the enhanced noise reduction preprocessing stage begins. For each battery sample (config.sample_list) specified in the configuration file, data cleaning is performed sequentially. Specifically, this includes using median filtering to remove spike noise and iterating through multiple Savitzky-Golay smoothing filters to optimize curve smoothness, thereby effectively suppressing high-frequency noise and preserving the true degradation trend. After processing, the denoised SOH data is used to calculate the corresponding capacity value and saved to the denoised_SOH.xlsx file.

[0048] Furthermore, in the sequence construction and normalization stage, the denoised SOH data is loaded as the model input, and MinMaxScaler is used to normalize the data of all battery samples to eliminate dimensional differences and improve training efficiency. According to the set timestep and prediction step size, the normalized data is divided into input sequence (X) and output sequence (Y) using a sliding window mechanism to construct training and testing datasets.

[0049] Furthermore, in the hyperparameter optimization stage, the average coefficient of determination is defined as the objective function, and the Bayesian optimization algorithm of the Optuna framework is used to automatically search for the optimal combination of hyperparameters of the CLMA model, including key parameters such as time step, hidden layer dimension, number of attention heads, and learning rate, so as to obtain the optimal model structure configuration.

[0050] After obtaining the optimal hyperparameters, the model is finalized and used for prediction. The CLMA model is reinitialized using the optimized parameters and trained using the entire training set. After training, predictions are made using the test set, and the prediction results are denormalized to restore the original SOH dimensions.

[0051] Furthermore, in the results evaluation and visualization stage, multiple performance evaluation indicators are calculated, including mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R^2), to comprehensively evaluate the model's predictive performance. Simultaneously, a comparison graph of the predicted SOH curve and the actual degradation trajectory is plotted to visually demonstrate the model's fitting effect on changes in battery health state, thereby verifying the high accuracy and reliability of the proposed method.

[0052] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter alterations to these embodiments within the spirit and principles of the present invention, achieved through conventional substitutions or by achieving the same function without departing from the principles and spirit of the present invention, fall within the scope of protection of the present invention.

Claims

1. A method for predicting the health status of energy storage batteries, characterized in that, Includes the following steps: Step S1: Collect and monitor the charging and discharging process of the battery, and collect and save the charging and discharging capacity data of the battery for each cycle; Step S2: Preprocess the collected battery discharge capacity data, calculate the battery health status data, and construct a time series sample using a sliding window mechanism to divide the battery health status data of each battery into an input sequence and an output sequence. Step S3: Construct a hierarchical neural network architecture and train it; Step S4: Perform Bayesian optimization on the hierarchical neural network architecture based on the Optuna framework to automatically search and adjust key network parameters so that the model can achieve the best performance under different datasets and tasks. Step S5: Use the optimized neural network architecture to predict the energy storage battery and obtain its battery health status.

2. The method for predicting the health status of energy storage batteries according to claim 1, characterized in that, In step S2, during the preprocessing stage of the battery's discharge capacity data, the original capacity data file of the battery is imported, and the ratio of the current capacity to the initial capacity is calculated and converted into health status data to characterize the battery performance degradation in a standardized form.

3. The method for predicting the health status of energy storage batteries according to claim 1, characterized in that, In step S2, the method for preprocessing the collected battery discharge capacity data includes: Step S21: Use enhanced data noise reduction processing on the collected battery discharge capacity data to remove noise and retain the decay trend; Step S22: Initially eliminate spike noise through median filtering, and then perform deep smoothing by multiple Savitzky-Golay smoothing filters, dynamically adjusting the window size to adapt to curves of different lengths and avoid trend distortion. Step S23: Next, calculate the State of Health (SOH) data of the battery after noise reduction based on the battery discharge capacity data; calculate the ratio of the discharge capacity of each battery cycle to the initial rated capacity of the battery to reflect the degree to which the battery performance is maintained relative to the initial state. The calculation formula is: SOH = Current capacity / Initial capacity * 100%; Step S24: Perform global normalization on the SOH sequences of all battery samples using MinMaxScaler, and linearly map them to... Within this range, the model can better learn the relative change patterns of different samples; Step S25: Construct time series samples using a sliding window mechanism; divide the SOH data of each battery into input and output sequences to construct training and test datasets.

4. The method for predicting the health status of energy storage batteries according to claim 3, characterized in that, In step S25, the input sequence consists of historical SOH data with a length equal to the input step size; the output sequence is the SOH prediction target with the corresponding future output step size.

5. The method for predicting the health status of energy storage batteries according to claim 1, characterized in that, In step S3, the construction of the hierarchical neural network architecture consists of a convolutional neural network, a long short-term memory network, a multi-head attention mechanism, and an output layer connected in series.

6. The method for predicting the health status of an energy storage battery according to claim 5, characterized in that, In step S3, the training method for the hierarchical neural network architecture includes: For the SOH sequence prediction task, the mean squared error loss function is used as the optimization objective to measure the difference between the model's predicted values ​​and the true values, thereby guiding the update of network parameters. To further prevent the model from overfitting on the training set while ensuring the efficiency of the training process, an early stopping mechanism is introduced. By monitoring the changing trend of training loss or verification loss, training is automatically terminated when no significant improvement is observed for several consecutive rounds.

7. The method for predicting the health status of energy storage batteries according to claim 1, characterized in that, In step S4, the objective function for Bayesian optimization of the hierarchical neural network architecture based on the Optuna framework is the average value of the model across all test samples. Fractions enable high accuracy and strong generalization ability in SOH sequence prediction.

8. The method for predicting the health status of energy storage batteries according to claim 7, characterized in that, In step S4, the Bayesian optimization algorithm of the Optuna framework is used to automatically search for the optimal combination of hyperparameters of the CLMA model. The key parameters include time step, hidden layer dimension, number of attention heads, and learning rate. After obtaining the optimal hyperparameters, model shaping and prediction are performed. The CLMA model was reinitialized with the optimized parameters and trained using the entire training dataset. After training, predictions were made using the test dataset, and the prediction results were denormalized to restore the original SOH dimensions.

Images