Battery life prediction method and device based on life prediction model

Abstract

The invention provides a battery life prediction method and device based on a life prediction model. The method comprises the following steps: acquiring a time sequence characteristic data set of a to-be-predicted battery; screening the time sequence feature data set to obtain a first feature set; the first feature set comprises a plurality of key degradation features, and the key degradation features are used for representing core features of the to-be-predicted battery in the degradation process; performing feature interaction processing on the time sequence feature data set by using a multi-rank tensor decomposition structure through L1 regularization and L2 regularization strategies to obtain a second feature set; the second feature set comprises nonlinear interaction features among multiple physical quantities in the time sequence feature data set; fusing the first feature set and the second feature set to obtain a fused feature set; inputting the fused feature set into a life prediction model to obtain the remaining service life of the to-be-predicted battery; the life prediction model is obtained through training according to a historical fusion feature set. According to the embodiment of the invention, the prediction precision of predicting the remaining service life of the battery can be improved.

Description

Technical Field

[0001] The present application relates to the field of battery technology, and in particular to a battery life prediction method and device based on a life prediction model. Background Art

[0002] As global energy transformation and sustainable development demands grow, accurate prediction of the remaining useful life (RUL) of lithium-ion batteries, core energy storage components, is crucial for optimizing battery management systems, extending equipment life, improving system safety, and reducing maintenance costs. Existing battery life prediction methods primarily rely on data-driven models.

[0003] These data-driven models are trained on large amounts of historical data, aiming to establish a mapping relationship between degradation characteristics and RUL. However, the battery degradation process is affected by multiple factors, including complex electrochemical reactions, thermodynamic behavior, and external operating conditions such as temperature and charge and discharge rates. The dynamic coupling relationships between battery characteristics are complex and changeable, and data-driven models struggle to fully capture these dynamic coupling relationships. This results in limitations in data-driven models when identifying key degradation characteristics, which in turn limits the improvement of prediction accuracy and makes it difficult to meet the demand for high-precision predictions in practical applications. Summary of the Invention

[0004] The present application provides a battery life prediction method and device based on a life prediction model, which can improve the prediction accuracy of the remaining service life of the battery.

[0005] In a first aspect, an embodiment of the present application provides a battery life prediction method based on a life prediction model, the method comprising:

[0006] Acquire a time series feature data set of a battery to be predicted, where the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data;

[0007] The time series feature data set is screened to obtain a first feature set; the first feature set includes multiple key degradation features, which are used to characterize the core features of the battery to be predicted during the degradation process;

[0008] Through L1 regularization and L2 regularization strategies, the multi-rank tensor decomposition structure is used to perform feature interaction processing on the time series feature dataset to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature dataset;

[0009] Fusing the first feature set and the second feature set to obtain a fused feature set;

[0010] The fused feature set is input into the life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fused feature set.

[0011] In a second aspect, the present application provides a battery life prediction device based on a life prediction model, the device comprising:

[0012] An acquisition module, configured to acquire a time series feature data set of a battery to be predicted, wherein the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data;

[0013] a screening module, configured to screen the time series feature data set to obtain a first feature set; the first feature set includes a plurality of key degradation features, the key degradation features being used to characterize core features of the battery to be predicted during the degradation process;

[0014] a processing module, configured to perform feature interaction processing on the time series feature dataset using a multi-rank tensor decomposition structure through L1 regularization and L2 regularization strategies to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature dataset;

[0015] a fusion module, configured to fuse the first feature set and the second feature set to obtain a fused feature set;

[0016] The determination module is used to input the fusion feature set into a life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fusion feature set.

[0017] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising: a processor and a memory storing computer program instructions;

[0018] When the processor executes the computer program instructions, it implements the battery life prediction method based on the life prediction model as in any one of the embodiments of the first aspect.

[0019] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, a battery life prediction method based on a life prediction model as in any one of the embodiments in the first aspect is implemented.

[0020] In a fifth aspect, an embodiment of the present application provides a computer program product. When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes a battery life prediction method based on a life prediction model as in any one of the embodiments in the first aspect above.

[0021] In a battery life prediction method and device based on a life prediction model provided in an embodiment of the present application, a time series feature dataset of the battery to be predicted is first obtained, covering key time series data such as voltage, temperature, current and capacity decay. Next, the time series feature dataset is screened to obtain a first feature set. This process retains key degradation features closely related to the battery degradation process and removes redundant and irrelevant information, thereby reducing the data dimension and reducing the computational complexity. Furthermore, using a multi-rank tensor decomposition structure and combining L1 and L2 regularization strategies, feature interaction processing is performed on the time series feature dataset to obtain a second feature set. This step can capture the complex dynamic coupling relationship between battery features and extract nonlinear interaction features between multiple physical quantities. In this way, the shortcomings of traditional data-driven models in feature interaction mining are compensated, enabling the life prediction model to understand the battery degradation mechanism more comprehensively and accurately. Finally, by fusing the first feature set and the second feature set to form a fused feature set, and inputting it into the life prediction model trained based on historical data, an accurate prediction of the remaining service life of the battery is achieved. The fusion feature set integrates key degradation features and nonlinear interaction features, enabling the life prediction model to better fit the battery degradation trajectory, thereby significantly improving the prediction accuracy of the remaining battery service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0023] Figure 1 1 is a flow chart of a battery life prediction method using a life prediction model provided in an embodiment of the present application;

[0024] Figure 2 It is a heat map of the correlation between the features and RUL provided in the embodiments of the present application;

[0025] Figure 3 is a feature importance distribution diagram provided by an embodiment of the present application;

[0026] Figure 4 This is one of the curves showing the changing trend of average temperature and capacity of different batteries during the charge and discharge cycle provided in the embodiments of the present application;

[0027] Figure 5 This is the second graph showing the changing trend of average temperature and capacity of different batteries during cyclic charge and discharge provided by the embodiments of the present application;

[0028] Figure 6is a scatter plot of the predicted values ​​and true values ​​of the lifespan prediction model provided in the embodiment of the present application;

[0029] Figure 7 is a distribution diagram of the prediction residual relative to the predicted value of the life prediction model provided in the embodiment of the present application;

[0030] Figure 8 is a probability distribution diagram of the absolute value of the prediction error of the life prediction model provided in the embodiment of the present application;

[0031] Figure 9 This is a comparison curve between the actual value and the predicted value of the life prediction model provided by the embodiment of the present application;

[0032] Figure 10 It is a structural diagram of a battery life prediction device of a life prediction model provided in an embodiment of the present application;

[0033] Figure 11 It is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features therein can be combined with each other in the absence of conflict.

[0035] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0036] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0037] The degradation process of lithium batteries is affected by complex electrochemical reactions, thermodynamic behavior, and external operating conditions such as temperature and charge and discharge rates. This makes the processing of high-dimensional time series features complex and the dynamic coupling relationship between features difficult to effectively capture. In addition, existing methods are limited in their effectiveness in modeling the nonlinear interactions between multiple physical quantities such as voltage, temperature, and internal resistance, making it difficult to fully explore the key information in the degradation pattern. In small sample scenarios, the model is prone to insufficient generalization ability, and the prediction accuracy and robustness are difficult to meet the needs of practical applications. At the same time, the processing of high-dimensional feature space increases computational complexity, affecting the engineering practicality of the model.

[0038] With the transformation of the global energy structure and the urgent need for sustainable development, the importance of lithium-ion batteries as core energy storage components for electric vehicles (EVs), renewable energy storage systems, and portable electronic devices has become increasingly prominent. Lithium batteries, with their advantages such as high energy density, long cycle life, and low self-discharge rate, have become a key component in promoting the development of clean energy technologies. However, batteries inevitably experience capacity fade and performance degradation during use. This process is affected by complex electrochemical reactions, thermodynamic behavior, and external operating conditions such as temperature and charge and discharge rates. Accurately predicting the remaining service life of lithium batteries is crucial for optimizing the battery management system (BMS), extending the equipment operating cycle, and improving system safety. In addition, accurate RUL prediction can effectively reduce maintenance costs, reduce the risk of unexpected failures, and provide an important reference for battery recycling and cascade utilization.

[0039] Traditional battery life prediction methods mainly rely on analysis based on physical or electrochemical models, such as the Equivalent Circuit Model (ECM) or the Single Particle Model (SPM) [6]. These methods can provide high prediction accuracy under specific conditions by describing the physical and chemical processes inside the battery. However, due to the complexity of the lithium battery aging mechanism and the diversity of actual operating conditions, physical models have significant limitations in generalization and computational efficiency. In addition, physical models usually require detailed parameter calibration and a large amount of experimental data support, which is often difficult to achieve in actual engineering applications.

[0040] In recent years, data-driven machine learning methods have attracted widespread attention due to their ability to mine nonlinear degradation patterns from battery monitoring data. These methods construct a mapping relationship between degradation features and RUL by analyzing time series data such as voltage, current, and temperature, without the need for in-depth understanding of the physical and chemical mechanisms inside the battery. Among them, the Gradient Boosting Decision Tree (GBDT) algorithm, especially the extreme gradient boosting algorithm, has performed well in RUL prediction tasks due to its efficient feature processing capabilities and ability to model nonlinear relationships. However, existing data-driven methods still face several challenges. First, battery monitoring data usually contains high-dimensional time series features, and directly inputting them into the model may increase computational complexity and reduce model performance. Second, the dynamic coupling relationship between features is difficult to fully capture through traditional feature engineering methods. Finally, conventional machine learning models are prone to overfitting or insufficient generalization when processing small sample data sets or noise interference.

[0041] To address these challenges, existing research has attempted to improve model performance by introducing feature selection, attention mechanisms, and optimization algorithms. Recursive Feature Elimination (RFE) achieves dimensionality reduction by iteratively removing low-importance features, but often ignores interactions between features, resulting in the loss of critical degradation information. Attention mechanisms enhance the model's ability to focus on critical temporal information by assigning dynamic weights to different features. However, modeling cross-dimensional feature interactions remains insufficient when dealing with complex couplings between multiple physical quantities such as voltage, temperature, and internal resistance. Furthermore, optimization algorithms based on swarm intelligence, such as genetic algorithms (GAs), particle swarm optimization (PSOs), and the binary firefly algorithm (BFAs), achieve global optimization by simulating the behavior of biological swarms and demonstrate strong search capabilities in feature selection. However, the convergence speed and stability of these algorithms are easily affected by the initial parameter settings, and they may become trapped in local optima in high-dimensional feature spaces. Therefore, developing a hybrid modeling framework that can simultaneously achieve efficient feature selection, capture complex feature interactions and improve prediction accuracy has become an urgent need in the field of lithium battery RUL prediction.

[0042] Therefore, existing technologies have certain shortcomings in predicting the remaining useful life of lithium-ion batteries. Methods based on physical or electrochemical models, such as equivalent circuit models or single-particle models, rely on detailed parameter calibration and large amounts of experimental data. They are computationally complex and have limited generalization capabilities, making them difficult to adapt to diverse operating conditions. Although data-driven machine learning methods can mine nonlinear degradation patterns, processing high-dimensional time series features increases the computational burden, and feature selection methods often ignore interactions between features, resulting in the loss of critical degradation information. Traditional attention mechanisms are limited in their effectiveness when modeling the nonlinear coupling between multiple physical quantities such as voltage, temperature, and internal resistance, and are prone to overfitting in small sample scenarios, resulting in insufficient prediction accuracy and robustness. Optimization algorithms such as genetic algorithms or particle swarm optimization are prone to falling into local optimality in high-dimensional feature spaces, resulting in inefficient feature screening.

[0043] In order to solve the problems existing in the related art, the embodiments of the present application provide a battery life prediction method and device based on a life prediction model.

[0044] The following first introduces the battery life prediction method based on the life prediction model provided in the embodiment of the present application. Figure 1 As shown, the method specifically includes the following steps:

[0045] S100 , obtaining a time series feature data set of a battery to be predicted, where the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data.

[0046] Optionally, in the embodiments of the present application, the battery to be predicted may be a lithium-ion battery whose remaining useful life (RUL) needs to be assessed. Such batteries can degrade over time due to factors such as electrochemical reactions and thermodynamic losses. Therefore, it is necessary to collect operating status data (such as voltage and temperature) to predict the remaining useful life cycle, providing a basis for maintenance strategies and cascade utilization decisions in the battery management system.

[0047] A time series feature dataset is a collection of battery operating status data recorded in chronological order. It consists of multi-dimensional time series data such as voltage, temperature, and current. This dataset is indexed by time, and each time point corresponds to a set of feature values ​​(such as the voltage, temperature, and current values ​​at a certain moment), forming a dynamic sequence that reflects the battery degradation process.

[0048] Voltage time series data is a serial record of the terminal voltage changes over time during the battery's charge and discharge process, reflecting the dynamic process of the electrochemical reaction inside the battery, such as the voltage rise rate during charging and the platform voltage offset during discharge.

[0049] Temperature time series data is a sequence of monitoring changes in battery surface or internal temperature over time. Temperature is a key external factor affecting battery degradation: high temperatures accelerate electrolyte decomposition and electrode material aging, while low temperatures reduce ion migration, thereby affecting capacity release.

[0050] Current time series data records the changes in battery charge and discharge current over time, reflecting the battery's workload (such as charging current and discharge pulse intensity). Current directly affects the battery's polarization and heat generation. High current conditions can exacerbate electrode material loss and accelerate capacity decay.

[0051] Capacity decay time series data is a sequential record of the battery's actual usable capacity decreasing over time (or cycle count). A battery's lifespan is generally considered to have ended when its capacity drops below 80% of its nominal value, making the capacity decay rate a key factor in predicting remaining useful life. This data, presented as a "cycle count-capacity" or "time-capacity" curve, directly reflects the final outcome of battery degradation.

[0052] Optionally, in a feasible implementation of the present application, first, the operating data of the battery to be predicted can be collected in real time through a battery management system (BMS) or an external sensor, wherein the voltage timing data is recorded by a voltage sensor at a fixed sampling frequency (such as 100Hz) during the charging and discharging process, the temperature timing data is monitored by a pre-embedded NTC thermistor to monitor the temperature changes of the battery surface or the tab, and the current timing data can use a Hall current sensor to capture the changes in the size and direction of the charging and discharging current.

[0053] Capacity decay timing data can be obtained in two ways: one is to perform full charge and discharge tests regularly to calculate the ratio of actual discharge capacity to nominal capacity; the other is to estimate the dynamic capacity decay value in real time based on current integration (Coulomb counting method) combined with voltage correction algorithm.

[0054] The collected data undergoes preprocessing steps: denoising and filtering (e.g., Kalman filtering) are performed on the voltage, temperature, and current data to remove outliers; data with different sampling frequencies are time-aligned (e.g., resampling to 1Hz); and capacity decay data is smoothed to eliminate test errors. Finally, the processed voltage, temperature, current, and capacity decay data are integrated into a time series to form a time series feature dataset containing multi-dimensional features, providing standardized input for subsequent feature screening.

[0055] S200 , screening the time series feature data set to obtain a first feature set; the first feature set includes a plurality of key degradation features, and the key degradation features are used to characterize core features of the battery to be predicted during the degradation process.

[0056] Optionally, in an embodiment of the present application, the key degradation feature is the core indicator in the first feature set used to directly characterize the decline in battery performance, reflecting the degradation mechanism of the battery in the process of electrochemical reaction, thermodynamic behavior, etc. The key degradation feature has the following characteristics: strong correlation with RUL (such as the fluctuation amplitude of the voltage statistic is significantly negatively correlated with the capacity decay), high sensitivity to the degradation process (such as the linear coefficient p1-3 of the capacity decay model can reflect the aging rate), and clear physical meaning (such as the internal resistance increment can be directly related to the electrode material loss). The role of the key degradation feature is to provide the degradation feature that best reflects the health status of the battery for life prediction, such as the peak-to-valley difference in the voltage time series data, the extreme value change rate of the temperature time series data, etc., which are all key features that can quantify the degree of battery degradation.

[0057] Optionally, in a feasible implementation of the present application, first, BFA is combined with the eXtreme Gradient Boosting (XGBoost) dynamic fitness function to intelligently screen out key degradation features that are highly correlated with the remaining battery life from high-dimensional time series feature data.

[0058] Specifically, we first set the number of firefly individuals to 25 (for illustrative purposes only, subject to adjustment), and the maximum number of evolutionary generations to 50 (for illustrative purposes only, subject to adjustment). Each firefly individual represents a feature subset, and its position is represented by a binary vector with a dimension equal to the total number of features. A value of 1 indicates that the corresponding feature is included in the subset, and a value of 0 indicates that it is excluded.

[0059] The inverse of the root mean square error (RMSE) of the XGBoost model on the validation set was then used as the fitness function. A smaller RMSE indicates a higher fitness value, indicating a stronger correlation between the feature subset and the remaining useful life.

[0060] In each iteration, the Cartesian distance and attractive force between individual fireflies are calculated. The positions of the individual fireflies are updated based on the attractive force and a preset step-size scaling factor. A sigmoid function is applied to the updated position values, followed by thresholding to convert them into binary decisions and generate a new feature subset. For each updated feature subset, the fitness value is recalculated to assess the predictive performance of the new feature subset.

[0061] After 50 generations of evolution, the algorithm terminates. The feature subset corresponding to the individual with the highest fitness value among all firefly individuals is selected as the first feature set. This feature subset can include key degradation features such as voltage statistics and capacity decay parameters, which can effectively characterize the core characteristics of the battery during the degradation process, while reducing data dimensions, improving computational efficiency, and providing a solid foundation for subsequent life prediction. In this way, the S200 effectively filters out key features from high-dimensional time series feature data, providing a strong guarantee for improving the accuracy of battery remaining service life prediction.

[0062] S300, through L1 regularization and L2 regularization strategies, using a multi-rank tensor decomposition structure, perform feature interaction processing on the time series feature data set to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature data set.

[0063] Optionally, in an embodiment of the present application, L1 and L2 regularization are two regularization methods used in machine learning to prevent overfitting. L1 regularization (Lasso regression) adds the sum of the absolute values ​​of the weight coefficients to the loss function, forcing some weight coefficients to be zero, thereby achieving feature selection and model sparsification; L2 regularization (Ridge regression) adds the sum of the squares of the weight coefficients to make the weight coefficients as a whole tend to be smooth but not zero. In battery life prediction, the combination of these two strategies can capture the interaction of features such as voltage and temperature while avoiding overfitting of the model to noisy data and improving the generalization ability of the model.

[0064] Multi-rank tensor decomposition is a high-dimensional data processing method that represents raw feature data (such as multidimensional time series data of voltage, temperature, and current) as the product of multiple low-rank tensors. For example, a three-dimensional tensor (such as time × feature × battery) is decomposed into the product of multiple factor matrices, each representing a potential dimension (such as degradation pattern, time trend, etc.). This decomposition can explicitly capture high-order interactions between features, such as the three-dimensional coupling relationship between voltage, temperature, and capacity decay, breaking through the limitation of traditional methods that can only process two-dimensional data.

[0065] Multi-physics nonlinear interactions refer to the complex coupling between different physical quantities, such as voltage, temperature, and current, which cannot be expressed through linear combinations. Examples include the temperature-voltage synergy effect: the nonlinear relationship between the rate of voltage plateau drop and capacity decay at high temperatures; charge-discharge rate-internal resistance coupling: the impact of internal resistance growth on voltage fluctuations during high-current charging; and temperature gradient-capacity retention: the impact of internal battery temperature nonuniformity on long-term capacity retention. These interactions are captured through nonlinear transformations via tensor decomposition.

[0066] Optionally, in a feasible implementation method of the present application, the implementation process of S300 is to reshape the original voltage, temperature, current and other time series feature data into a three-dimensional data cube, in which the three dimensions represent time, feature type and number of battery samples respectively. Then, this three-dimensional cube is decomposed into the sum of multiple rank tensors through decomposition, and each rank tensor corresponds to a potential interaction pattern, such as voltage-temperature synergy effect, current-capacity attenuation relationship, etc. During the decomposition process, L1 and L2 regularization strategies are applied simultaneously. L1 regularization forces some factors to be zero to achieve feature sparsification, and L2 regularization prevents overfitting and improves the generalization ability of the model. By adjusting the number of ranks of the decomposition, interaction patterns of different complexities can be captured.

[0067] After decomposition, high-dimensional interaction features are generated by calculating nonlinear combinations between factor matrices. These interaction features incorporate complex nonlinear relationships between the original features, such as the impact of voltage fluctuations on capacity at high temperatures. Dimensionality reduction techniques are then used to remove redundant information, yielding a second feature set with a dimension approximately 30%-50% of the original features. This feature set includes third-order interaction terms such as "temperature × voltage × current," as well as time-series dynamic features such as the coupling of temperature change rate and voltage platform offset, significantly improving the ability to characterize battery degradation mechanisms.

[0068] S400: Fusing the first feature set and the second feature set to obtain a fused feature set.

[0069] Optionally, in a feasible implementation of the present application, the first feature set and the second feature set can first be standardized to unify the numerical range of each feature to the interval [0, 1] to eliminate dimensional differences. Subsequently, a feature stacking strategy is adopted to splice the standardized first feature set and the second feature set by column to form an initial fused feature set with the dimension of the sum of the two feature sets.

[0070] The final generated fusion feature set has both physical interpretability (such as the capacity decay parameters of the first feature set) and data-driven interactive characteristics (such as the high-order tensor factors of the second feature set), providing comprehensive and concise input for subsequent life prediction models.

[0071] S500, inputting the fused feature set into a life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fused feature set.

[0072] Optionally, in an embodiment of the present application, the lifespan prediction model is a machine learning model trained based on a historical fusion feature set to predict the remaining useful life of a battery. This lifespan prediction model integrates the first feature set and the second feature set, and predicts the degradation trajectory of a new battery by learning the mapping relationship between features in historical data and battery lifespan.

[0073] In a battery life prediction method based on a life prediction model provided in an embodiment of the present application, a time series feature data set of the battery to be predicted is first obtained, covering key time series data such as voltage, temperature, current and capacity decay. Then, the time series feature data set is screened to obtain a first feature set. This process retains the key degradation features closely related to the battery degradation process and removes redundant and irrelevant information, thereby reducing the data dimension and reducing the computational complexity. Furthermore, the multi-rank tensor decomposition structure is used, and combined with the L1 and L2 regularization strategies, the time series feature data set is subjected to feature interaction processing to obtain a second feature set. This step can capture the complex dynamic coupling relationship between battery features and extract the nonlinear interaction features between multiple physical quantities. In this way, the shortcomings of the traditional data-driven model in feature interaction mining are compensated, so that the life prediction model can understand the degradation mechanism of the battery more comprehensively and accurately. Finally, by fusing the first feature set and the second feature set to form a fused feature set, and inputting it into the life prediction model trained based on historical data, an accurate prediction of the remaining service life of the battery is achieved. The fusion feature set integrates key degradation features and nonlinear interaction features, enabling the life prediction model to better fit the battery degradation trajectory, thereby significantly improving the prediction accuracy of the remaining battery service life.

[0074] In one embodiment, the filtering of the time series feature data set to obtain the first feature set includes:

[0075] Randomly generate a preset number of feature subsets based on the time series feature data set; one feature subset corresponds to one firefly individual;

[0076] Based on each firefly individual, the firefly position of each firefly individual is updated through evolutionary iteration. After the evolutionary iteration update is completed, the firefly individual with the maximum fitness value is selected from the iteratively updated firefly individuals; the fitness value is used to characterize the prediction performance of the feature subset corresponding to the firefly individual for the remaining battery life;

[0077] The feature subset corresponding to the firefly individual with the maximum fitness value is determined as the first feature set.

[0078] Optionally, in an embodiment of the present application, the firefly individual is the basic unit in the binary firefly algorithm, and each individual corresponds to a candidate solution for a feature subset. In the battery life prediction scenario, a firefly individual represents a set of feature combinations selected from the original time series feature data set (such as voltage, temperature, current and other features). For example, if the original feature set contains 100 features, each firefly individual can be represented as a 100-dimensional binary vector, and the position with a value of "1" in the vector indicates that the corresponding feature is selected, and the position with a value of "0" indicates that the feature is eliminated. By randomly generating a preset number of firefly individuals, the algorithm can cover different feature selection possibilities and provide a variety of initial solutions for subsequent optimization.

[0079] The fitness value is a quantitative indicator for evaluating the quality of individual fireflies (i.e., feature subsets), and is directly related to the feature subset's predictive performance for the remaining battery life. In BFA, the fitness value can be defined as the inverse of the RMSE of the XGBoost model on the validation set. The smaller the RMSE, the more accurate the information contained in the feature subset is in predicting the RUL, and the higher the corresponding fitness value. During the evolutionary iteration, the algorithm guides the population to evolve towards a more optimal feature combination by comparing the fitness values ​​of individual fireflies, and ultimately selects the individual with the largest fitness value as the first feature set, ensuring that the selected features contribute the most to the RUL prediction.

[0080] In these optional embodiments, a binary firefly algorithm simulates firefly swarm behavior, randomly generates feature subsets, iteratively optimizes them, and uses fitness to evaluate the performance of feature combinations in predicting battery remaining life. This algorithm automatically and efficiently selects key degradation features from a large number of time series features, avoiding the subjectivity and blindness of manual screening. This effectively reduces data dimensionality, improves feature screening efficiency and model prediction accuracy, and enhances model generalization.

[0081] In one embodiment, each evolutionary iterative update includes:

[0082] Get the Cartesian distance and attraction between individual fireflies;

[0083] Performing a position update operation on each firefly individual to obtain an updated feature subset corresponding to each firefly individual; wherein the position update operation includes: updating the firefly position of each firefly individual according to the Cartesian distance and attraction between each firefly individual to obtain an updated firefly position; applying a normalization function to the continuous position values ​​in the updated firefly position and converting them into binary decisions through a thresholding function; and determining the updated feature subset corresponding to the firefly individual based on the binary decision;

[0084] For the updated feature subsets corresponding to each firefly individual, a fitness value calculation operation is performed respectively to obtain the fitness value corresponding to each updated feature subset.

[0085] Optionally, in an embodiment of the present application, the Cartesian distance is a Euclidean distance that measures the difference between two firefly individuals (feature subsets). Attraction is defined based on the Cartesian distance. The closer the Cartesian distance, the stronger the attraction, guiding the fireflies to move towards the better individuals. For example, if two individuals have the same values ​​on 80% of the feature dimensions, the Cartesian distance is small, and attraction makes them more likely to converge in evolution. The firefly position is a continuous vector, and each dimension corresponds to the selection probability of a feature.

[0086] Optionally, in one specific implementation of the present application, like many nature-inspired optimization algorithms, the firefly algorithm draws on observed behaviors in the natural world to inform heuristic approaches to solving complex optimization problems, ranging from continuous to discrete domains. Within the realm of feature selection, each firefly represents a unique combination of selected features from a dataset, typically represented as a binary vector where a single bit indicates the inclusion or exclusion of a specific feature. The "luminosity" or "brightness" of a firefly is determined based on the effectiveness of its feature subset, which is evaluated using predefined criteria such as classification accuracy or root mean squared error. As the iterative process progresses, the fireflies traverse the space of feature subsets, systematically attracting configurations with excellent evaluation scores. At the same time, they introduce random perturbations to ensure comprehensive exploration of the search space. This interplay of deterministic attraction and probabilistic exploration can result in a nuanced search strategy, allowing BFA to identify optimal or near-optimal feature subsets, thereby enhancing prediction accuracy while reducing dimensionality.

[0087] Among them, the main algorithm equation of the solution vector (set position update method) of the firefly algorithm is expressed as:

[0088]

[0089] in, is the position vector of the i-th / j-th firefly in the t-th generation (corresponding to the encoding of the feature subset); is the updated position vector of the i-th firefly in the t+1 generation (new feature subset candidate); α is the scaling factor that controls the random walk step length, is a random number vector drawn from a Gaussian distribution at each iteration, γ is a scale-dependent parameter that controls the visibility of fireflies, and β0 is the attraction constant when the distance between two fireflies is zero. It is worth noting that the distance r between the i-th firefly and the j-th firefly is ijis defined as their Cartesian distance to eliminate any ambiguity in higher dimensions. In addition, BFA is designed for continuous optimization problems. The feature selection problem is expressed as an array of Boolean values, so a binary version of BFA is required, which is implemented by incorporating a threshold-based activation process, where the continuous position values ​​of the fireflies are rounded and compared with a threshold to make a binary decision on feature selection. The movement of fireflies towards brighter individuals is adjusted by an attraction factor β, which decreases with increasing distance and is further affected by a randomization term α. The attraction formula of BFA is as follows:

[0090]

[0091] Among them, β min This binary adaptation enables the algorithm to perform feature selection tasks by interpreting each bit as the presence or absence of a feature, rather than its position in a continuous space.

[0092] Therefore, BFA's nonlinear attraction mechanism allows short-range attraction to be more effective than long-range attraction, naturally dividing the firefly swarm into various subswarms, each of which is likely to converge around a local optimum. Within these subswarms, the emergence of a global solution is theoretically guaranteed. Furthermore, given its ability to operate with multiple swarms, BFA is well-suited to addressing nonlinear, multimodal optimization challenges.

[0093] In order to optimize the feature space of the battery remaining useful life prediction model, the binary firefly algorithm is used to intelligently screen the battery degradation features, and the correlation between the features and RUL is deeply explored through visual analysis.

[0094] like Figure 2 As shown, Figure 2 This is a heat map of the correlation between features and remaining useful life (RUL). The vertical axis represents different features (such as p1-3, V8-1, etc.), the horizontal axis "lifetime" represents the remaining battery life, and the color bar on the right shows the correlation coefficient range (-1 to 1). The values ​​reflect the linear correlation between the features and RUL: "lifetime" has a correlation coefficient of 1.00 (perfectly positive correlation); p1-3, p7-3, etc. are positively correlated (the larger the value, the longer the RUL tends to be); V5-1, V5-2, etc. are negatively correlated (the larger the value, the shorter the RUL tends to be), which can help screen key features that affect battery life.

[0095] like Figure 2The heat map of the correlation between features and RUL shown clearly demonstrates the statistical correlation between the feature subset optimized by BFA and the remaining service life of the battery. The results show that the voltage-related features V5-1 and V5-2 exhibit the strongest negative correlation, indicating that the increase in voltage statistical features may be closely related to the decrease in RUL, reflecting the significant degradation of electrochemical performance during battery aging. In contrast, the linear coefficients p1-3 and exponential coefficients p7-3 of the capacity decay model show a significant positive correlation, indicating that batteries with slower linear rates of capacity decay or better exponential model parameters may have longer life. This differentiated correlation pattern not only reveals the different effects of voltage and capacity characteristics on RUL prediction, but also further reflects the complex coupling relationship between electrochemical properties and thermodynamic behavior during battery aging.

[0096] like Figure 3 As shown, this is a "feature importance" bar chart, which is used to show the contribution of each feature to the battery remaining useful life (RUL) prediction model. The horizontal axis "AbsoluteCorrelation" represents the feature importance value, and the vertical axis represents different features (such as V5-1, V5-2, etc.). The longer the bar, the greater the impact of the feature on the model prediction results. Among them, features such as V5-1, V5-2, and V6-2 are of high importance and are key factors in predicting battery life; the short tail features (such as V2-2 and V5-3) have little contribution to the prediction and can assist in screening key features and optimizing the model.

[0097] Figure 3 The feature importance distribution bar chart shown further quantifies the contribution of each feature to RUL prediction. The results show that the voltage-related features V5-1 and V5-2 have the highest absolute correlation coefficients, showing the strongest predictive ability, highlighting the dominant role of voltage statistical features in RUL prediction. The importance of the linear coefficient p1-3 of the capacity decay model ranks fourth, indicating that it has certain predictive value in the capacity decay trend. In addition, temperature-related features also show a certain correlation, suggesting that the impact of thermal management on battery life cannot be ignored. The BFA algorithm effectively eliminates low-importance features with absolute correlation coefficients less than 0.1 through intelligent screening. These features include some statistical skewness indicators. These features are often retained in traditional methods, but their actual contribution to RUL prediction is limited. The simplification of the feature space after screening helps to improve the computational efficiency and prediction accuracy of the model, while avoiding the risk of overfitting, and providing a more reliable feature basis for subsequent model optimization.

[0098] Alternatively, in another optional implementation of the present application, first, the original features are encoded into a binary vector (such as 100 dimensions, with each bit corresponding to the selection of a feature), and 25 firefly individuals are randomly generated as the initial feature subset. Then, iterative optimization is performed: the Euclidean distance between individuals is calculated (if two individuals have the same value on 80% of the features, the distance is small), and the attraction is calculated based on formula (2) (the closer the distance, the stronger the attraction). Next, the position is updated according to formula (1) (random perturbations are introduced to avoid local optimality), and the Sigmoid function is applied to the updated continuous position value to map it to the [0,1] interval, and converted to a binary decision by a threshold of 0.5 (such as the feature corresponding to the value 0.8 is selected), and a new feature subset is generated. After each iteration, the subset performance is evaluated using the XGBoost model (with the inverse of RMSE as the fitness value). After 50 generations of iteration, the feature subset corresponding to the individual with the highest fitness is selected as the first feature set.

[0099] In these optional embodiments, the attraction and repulsion mechanism between individual fireflies is utilized to automatically explore the optimal feature combination, avoiding the blindness of manual screening; through iterative updates and random perturbations, local search and global exploration are balanced, and the complex nonlinear relationship between features is effectively handled; the fitness value is directly related to the model performance, ensuring that the selected feature subset contributes most to the prediction of the remaining battery life, improving prediction accuracy and computational efficiency, and is suitable for rapid dimensionality reduction and feature extraction of high-dimensional time series data.

[0100] In one embodiment, the L1 regularization and L2 regularization strategies are used to perform feature interaction processing on the time series feature dataset using a multi-rank tensor decomposition structure to obtain a second feature set, including:

[0101] Inputting the temporal feature dataset into a tensor product attention network, wherein the tensor product attention network includes a multi-rank tensor decomposition structure, and the multi-rank tensor decomposition structure is used to map the feature dataset to a multi-rank tensor space; L1 regularization and L2 regularization strategies are used to prevent overfitting during the training process of the tensor product attention network;

[0102] Decomposing the time series feature dataset using the multi-rank tensor decomposition structure to obtain tensor product decomposition results corresponding to the query matrix, the key matrix, and the value matrix respectively;

[0103] Calculating, based on the tensor product decomposition result, an attention score between multiple physical quantities in the time series feature dataset, wherein the attention score is used to characterize the importance and relevance of each physical quantity in the battery degradation process;

[0104] According to the attention score, nonlinear interaction features between multiple physical quantities in the temporal feature dataset are determined, and the nonlinear interaction features are determined as the second feature set.

[0105] Tensor Product Attention Network (TPA) is a feature interaction modeling method based on multi-rank tensor decomposition. Its core idea is to decompose the query (Q), key (K), and value (V) matrices of the traditional attention mechanism into paired subspace combinations through low-rank projection. TPA maps high-dimensional features to tensor spaces of different ranks by introducing a learnable projection matrix group, thereby explicitly capturing cross-dimensional nonlinear relationships. Compared with standard dot product attention, TPA shows three advantages in battery RUL prediction tasks: (1) multi-rank decomposition alleviates the computational burden of direct interaction of high-dimensional features; (2) the joint regularization strategy (L1 / L2) suppresses overfitting in small sample scenarios; (3) the interpretability of feature-level attention weights provides a new perspective for degradation mechanism analysis.

[0106] make Denote t=1,...,T as the latent vector corresponding to a sequence of length t. A typical multi-head attention block has h heads, each of size d h , satisfying d model =h×d h Standard attention projects the entire sequence into three tensors Among them, Represents the slice of the tth token.

[0107] Instead of forming the query, key, or value for each head through a single linear map, TPA transforms each Q t ,K t ,V t Linear mapping, which is reduced to the sum of tensor products, has rank R q , R k and R v Specifically, for each token t, we define

[0108]

[0109] in Therefore, for the query slice, each tensor product Likewise, a similar definition applies to the key K t Sum value piece V t .

[0110] Each factor in the tensor product depends on the labeled hidden state x t , the query can be written as (6), and the same is true for keys and values.

[0111]

[0112] Merge the hierarchical indices into one output dimension. For queries,

[0113]

[0114] Then reshape it into and Sum R q , and The yield scaling,

[0115]

[0116] Repeat for all markers and reconstruct A similar procedure can be applied to obtain R k and R v After factorizing Q, K, V, multi-head attention behaves like Transformers. For each head i∈{1,...,h},

[0117]

[0118] in It is a slice along the head size, and the connection of these h heads along the last dimension is obtained A tensor that is represented by the output weight matrix Projection back

[0119] TPA(Q,K,V)=Concat(head1,...,head h )W O (10)

[0120] For parameter initialization, use Xavier initialization to initialize the weight matrix Specifically, each entry of the weight matrix is ​​drawn from a uniform distribution bounded by where n in and n out are the input and output dimensions of the corresponding weight matrices. This initialization strategy helps maintain the variance of activations and gradients throughout the network.

[0121] Optionally, in a feasible implementation of the present application, first, the original time series features are reshaped into a three-dimensional tensor Where I is the time dimension, J is the feature type, and K is the number of samples. Subsequently, multi-rank tensor decomposition is used to decompose the tensor into the product of three factor matrices: query, key, and value. Each matrix corresponds to the potential representation of a different physical quantity. During the decomposition process, L1 regularization and L2 regularization are used to prevent overfitting. After decomposition, the attention score is calculated. Finally, the feature combination after attention weighting (such as temperature-voltage synergy effect, current-capacity attenuation relationship) is used as the second feature set.

[0122] In these optional embodiments, the time series features are mapped to the multi-rank tensor space through the tensor product attention network and multi-rank tensor decomposition, combined with L1 / L2 regularization to prevent overfitting, and the query, key, and value matrix decomposition results are obtained. The attention score is calculated to characterize the importance and relevance of the physical quantity, thereby determining the nonlinear interaction features as the second feature set, which can capture high-order nonlinear relationships and improve the feature explanatory power and model generalization.

[0123] In one embodiment, the lifespan prediction model is an XGBoost ensemble model;

[0124] Inputting the fused feature set into a life prediction model to obtain the remaining service life of the battery to be predicted includes:

[0125] The fused feature set is input into the XGBoost integrated model, and the fused feature set is processed by the XGBoost integrated model to obtain the remaining service life of the battery to be predicted; wherein, the XGBoost integrated model optimizes the objective function of the XGBoost integrated model through a second-order Taylor expansion, and introduces L1 regularization and L2 regularization terms to suppress overfitting.

[0126] Optionally, in a feasible implementation of the present application, the XGBoost algorithm represents an advanced form of integrated machine learning, which uses gradient boosting technology to optimize decision tree models and demonstrates excellent ability in identifying complex patterns and interactions in data sets. The typical input of the XGBoost algorithm includes a matrix of eigenvectors selected by BFA and their corresponding target values. The output is a robust prediction model that provides an accurate estimate of the target variable based on the input features. Therefore, XGBoost uses an additive training strategy to sequentially improve the model by optimizing the objective function. Then the tth objective function of the model can be expressed as,

[0127]

[0128] Where l is the loss term of the tth round, and Ω is the regularization term of the model, as shown below,

[0129]

[0130] γ and λ are both custom parameters, and T is the number of leaves, which affects the balance between model fitting and model complexity. In addition, in order to provide more accurate prediction results, XGBoost performs a second-order Taylor expansion on Equation (11),

[0131]

[0132] in is the first-order derivative, is the second-order derivative, and then by substituting, from obj * The loss function score and wj represented by * The weight solution represented by is:

[0133]

[0134] XGBoost is an efficient gradient boosting framework that is particularly well-suited for smaller datasets with nonlinear and complex patterns. It sequentially builds a set of decision trees, each correcting the errors of the previous trees. This iterative approach improves prediction accuracy and reveals subtle trends in the data. Furthermore, XGBoost incorporates L1 (Lasso) and L2 (Ridge) regularization techniques to prevent overfitting, helping to identify underlying patterns and their cumulative effects, even when these dominant trends are not apparent on the surface.

[0135] In these optional embodiments, the objective function is optimized using a second-order Taylor expansion, which can fit the residual more accurately and improve prediction accuracy compared to traditional GBDT; the L1 / L2 regularization term is introduced to suppress overfitting and enhance the generalization ability of the model; high-dimensional and nonlinear features are automatically processed to effectively capture the complex interactions between multiple physical quantities; parallel computing is supported to improve training efficiency; the importance of features can be quantified to provide a basis for battery degradation mechanism analysis.

[0136] In one embodiment, the lifespan prediction model is pre-trained in the following manner:

[0137] Obtaining a training sample set; the training sample set includes historical time series characteristics of the battery and the actual remaining service life corresponding to the battery;

[0138] Performing feature screening on the historical time series features to obtain a first historical feature set;

[0139] By using L1 regularization and L2 regularization strategies and a multi-rank tensor decomposition structure, feature interaction processing is performed on the historical time series features to obtain a second historical feature set;

[0140] fusing the first historical feature set and the second historical feature set to obtain a historical fused feature set;

[0141] Based on the historical fusion feature set, prediction is performed using the life prediction model to obtain a predicted remaining service life;

[0142] Based on the actual remaining service life and the predicted remaining service life, a first loss function is calculated, and the life prediction model is trained based on the first loss function.

[0143] Optionally, in one feasible implementation of the present application, when training a life prediction model, a training sample set containing historical battery time series features and corresponding actual remaining service life is first obtained. For example, the features and life labels of multiple battery groups at different cycle times are collected from a relevant data set. Next, the historical time series features are screened, and a binary firefly algorithm is used to randomly generate several feature subsets, each corresponding to a firefly individual. By iteratively updating the firefly positions and combining fitness value evaluation, a first historical feature set that can characterize the core characteristics of battery degradation is selected.

[0144] Then, the historical time series features are processed using a multi-rank tensor decomposition structure, mapping the feature dataset into a multi-rank tensor space. L1 and L2 regularization strategies are employed during training to prevent overfitting. After decomposing the query, key, and value matrices, attention scores are calculated between the various physical quantities to identify the nonlinear interaction characteristics between the multiple physical quantities, forming a second historical feature set. The first and second historical feature sets are then fused, and a dimensionality reduction process is performed to obtain a fused historical feature set.

[0145] Finally, the historical fusion feature set is input into the XGBoost integrated model for prediction to obtain the predicted remaining service life. The loss function is then calculated based on the actual remaining service life and the predicted results. The model is trained by optimizing the loss function. During this period, the second-order Taylor expansion is used to optimize the objective function, and the L1 and L2 regularization terms are used to suppress overfitting. After adjusting the model hyperparameters and other operations, the model has the ability to accurately predict the remaining service life of the battery.

[0146] In these optional embodiments, the battery life prediction performance is improved through systematic feature engineering and model training: the binary firefly algorithm is used to screen key degradation features, combined with tensor decomposition to capture the nonlinear interaction of multiple physical quantities, L1 / L2 regularization is used to prevent overfitting, XGBoost integrated model training is used after feature fusion, the objective function is optimized through second-order Taylor expansion, and the dimension is reduced after feature fusion to support real-time high-precision prediction.

[0147] It should be noted that the various optional implementation methods introduced in the embodiments of the present application can be implemented in combination with each other or separately if they do not conflict with each other, and the embodiments of the present application do not limit this.

[0148] To facilitate understanding of the battery life prediction method based on the life prediction model provided in the above embodiment, the battery life prediction method based on the life prediction model is described below using a specific scenario embodiment.

[0149] Optionally, in an embodiment of the present application, the present application proposes a hybrid model (i.e., a life prediction model) that integrates a binary firefly algorithm, a tensor product attention mechanism, and an extreme gradient boosting algorithm for accurate prediction of the remaining service life of lithium-ion batteries. The technical solution includes the following core contents: First, BFA is used for feature selection. By simulating the intelligent behavior of biological groups and combining the dynamic fitness function based on XGBoost, in the optimization process of 25 firefly individuals and 50 generations of evolution, key degradation features such as voltage statistics and capacity decay parameters are intelligently screened from high-dimensional time series features, effectively reducing the dimension and retaining the core information. Secondly, a new TPA network is designed to map features to a low-rank subspace through multi-rank tensor decomposition, explicitly capturing the nonlinear interaction between multiple physical quantities such as voltage, temperature, and internal resistance, and introducing L1-L2 joint regularization to enhance the generalization ability in small sample scenarios. The high-level features generated by TPA are fused with the original features to form a comprehensive feature set. Finally, an XGBoost integrated prediction model is constructed, and the second-order Taylor expansion and regularization optimization are used to accurately fit complex degradation trajectories. Experimental verification shows that the model has passed the five-fold cross validation on the MIT lithium battery dataset, with an average RMSE of 0.0247 and R 2 The RUL prediction performance is 0.9559, demonstrating high accuracy and robustness. This application significantly improves RUL prediction performance through an end-to-end optimization framework that integrates feature selection, nonlinear interaction modeling, and integrated prediction, providing an efficient solution for battery health management and energy storage system optimization.

[0150] Specifically, by combining BFA with XGBoost's dynamic fitness function, key degradation features are intelligently screened, reducing the risk of dimensional disasters and improving computational efficiency. TPA introduces multi-rank tensor decomposition and L1-L2 joint regularization to explicitly capture nonlinear interactions between features and enhance the model's ability to model electrochemical and thermodynamic dynamics. XGBoost ensemble learning optimizes the fitting accuracy of complex degradation trajectories. This invention aims to overcome the shortcomings of high-dimensional feature processing, feature interaction modeling, and generalization capabilities for small sample scenarios, providing an efficient and scalable solution for battery health management and energy storage system optimization.

[0151] Optionally, in an embodiment of the present application, a hybrid model (i.e., life prediction model) integrating binary firefly algorithm, tensor product attention mechanism and XGBoost is proposed for accurate prediction of the remaining service life of lithium-ion batteries. It mainly includes: (1) using BFA combined with XGBoost dynamic fitness function, through 25 firefly individuals and 50 generations of evolution, intelligently screening key degradation features such as voltage statistics and capacity attenuation parameters from high-dimensional time series features, reducing dimensions and improving computational efficiency; (2) designing a new TPA network, explicitly capturing the nonlinear interactions between multiple physical quantities such as voltage, temperature, and internal resistance through multi-rank tensor decomposition, and combining L1-L2 joint regularization to enhance the generalization ability in small sample scenarios; (3) fusing the high-level features generated by TPA with the original features, inputting them into the XGBoost integrated model, and optimizing the complex degradation trajectory fitting using the second-order Taylor expansion, significantly improving the prediction accuracy (RMSE reached 0.0247, R 2 is 0.9559).

[0152] This application introduces BFA combined with the XGBoost dynamic fitness function, and intelligently screens key degradation features such as voltage statistics and capacity attenuation parameters through 25 firefly individuals and 50 generations of evolution, significantly reducing feature dimensions and improving computational efficiency. It optimizes the TPA design and uses multi-rank tensor decomposition to explicitly capture the nonlinear interactions between multiple physical quantities such as voltage, temperature, and internal resistance. It combines L1-L2 joint regularization to enhance the generalization ability of small sample scenarios, making it more efficient and stable than the standard TPA. It fuses the high-level features generated by TPA with the original features screened by BFA, inputs them into the XGBoost integrated model, and uses the second-order Taylor expansion to optimize the fitting of complex degradation trajectories, significantly improving the prediction accuracy (RMSE reaches 0.0247, R 2 The experimental results show that the RMSE and R of this application are better than those of the best existing technology in the five-fold cross validation. 2 All of these have been improved accordingly, with narrower error distribution and stronger model stability, providing a more accurate and robust RUL prediction solution for battery health management and energy storage system optimization.

[0153] This study used MIT lithium-ion battery cycle life test data to construct training and test datasets, which is the largest public dataset available for long-term battery degradation research to date. A total of 124 randomly selected APR18650M1A lithium iron phosphate (LFP) / graphite cells (nominal capacity 1.1Ah, from A123 Systems) were cycled in an environmental chamber (4°C) with various charge profiles and constant discharge profiles (30C-2.0V). The cells were charged from 0% to 80% state of charge, with 72 different charging strategies for each cell, and then charged from 100% state of charge to 1% in 80C constant current / constant voltage mode. The cutoff voltages for the cycling tests were 3.6V and 2V, respectively. The voltage, current, and battery temperature were continuously measured and recorded during the cycling process.

[0154] like Figure 4 As shown, Figure 4 This is a "Battery Capacity Trend" line chart, with cycle number on the horizontal axis and discharge capacity (in Ah) on the vertical axis. It shows the capacity decay of four battery groups: Battery 1, 6, 11, and 16, as the cycle number increases. Different colored curves represent different batteries, reflecting the differences in individual battery life and performance degradation characteristics.

[0155] like Figure 5 As shown, Figure 5 This is a line chart of "Battery Temperature Change," with cycle number plotted on the horizontal axis and average temperature (°C) plotted on the vertical axis. It shows the temperature evolution of Battery 1, 6, 11, and 16 as the cycle number increases. Different colored curves correspond to different batteries, reflecting individual differences in thermal characteristics during cycling. This helps analyze battery thermal management and degradation mechanisms.

[0156] Figure 4 Four sets of representative battery data were selected and distinguished by different colors, so that the temperature evolution of each battery can be compared intuitively. The temperature curve can reflect the thermal management performance of the battery during operation. Abnormal temperature fluctuations or continuous temperature rise often indicate battery aging or potential failure, which is one of the important indicators for evaluating the health of the battery. The service life of a battery is defined as the number of charge and discharge cycles that the battery can run before its maximum available capacity drops below 80% of its nominal value. Figure 5As shown in the figure, four sets of battery data are also selected to clearly demonstrate the capacity decay trajectories of different batteries. Capacity decay is a direct manifestation of battery aging. The change in the slope of the curve reflects the stage characteristics of the decay rate, such as the initial stable period and the later accelerated decay period. This is a key reference for predicting the remaining battery life. Together, the two curves form the core visualization analysis tool for battery health status.

[0157] The performance of the proposed hybrid model that integrates the binary firefly algorithm, tensor product attention mechanism, and XGBoost was evaluated using a five-fold cross-validation strategy. The root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R) were used to evaluate the performance of the model. 2 ), which is used to measure the overall regression ability and fitting level of the model. RMSE is an indicator to measure the average amplitude of the prediction error. The smaller the value, the smaller the deviation between the predicted value and the actual value, as shown in formula (16). MAPE measures the proportion of the prediction error to the true value, as shown in formula (17).

[0158]

[0159] Among them, n represents the total number of test samples, y i represents the true value of the i-th sample, that is, the actual battery life percentage, represents the life span value predicted by the model for the i-th sample,

[0160] R 2 It is an indicator of the goodness of fit of the regression model, measuring the ability of the predicted values ​​to explain the actual observed values.

[0161] Its value range is (-∞,1]. The closer it is to 1, the stronger the explanatory power of the model, as shown in formula (18).

[0162]

[0163] The results on the test set show that the RMSE of the lifespan prediction model is 0.0247, R 2 The average lifespan prediction model is 0.9559 and the average cost per watt-hour is 5.68%. These indicators show that the lifetime prediction model can predict the remaining service life of lithium batteries with high accuracy and effectively capture complex nonlinear degradation patterns.

[0164] like Figure 6 As shown, Figure 6This is a scatter plot of "Predicted Values ​​vs. True Values." The horizontal axis shows true values ​​(%), and the vertical axis shows predicted values ​​(%). The blue dots represent the model's predictions, and the black dashed line "Ideal" represents the ideal prediction line (where the predicted value is exactly equal to the true value). The closer the dots are to the dashed line, the more accurate the prediction. This provides a visual assessment of model performance and helps determine the accuracy of models in prediction tasks such as battery life.

[0165] like Figure 7 As shown, Figure 7 This is the "Residual Plot," used to evaluate model prediction error. The horizontal axis shows the predicted values ​​(%), and the vertical axis shows the residuals (%, the difference between the true value and the predicted value). The orange dots represent the residuals for each sample, and the black dotted line "Zero Error" represents the ideal line where the residual is zero. The residuals are randomly distributed above and below the dotted line.

[0166] like Figure 8 As shown, Figure 8 This is a histogram of the "Absolute Forecast Error Distribution." The horizontal axis shows the absolute error (AbsoluteError, %), which is the absolute value of the difference between the predicted value and the true value, and the vertical axis shows the frequency (the number of samples with the corresponding error). The blue bar chart shows the proportion of samples in different error ranges, and the blue curve is the error distribution trend line. This allows you to intuitively determine the model's prediction accuracy and error distribution characteristics, assisting in evaluating forecast performance.

[0167] like Figure 9 As shown, Figure 9 This is a line chart comparing true values ​​to predicted values. The horizontal axis is the sample index (SampleIndex), and the vertical axis is the life percentage (LifePercent). The solid blue dots represent the true values, and the dashed orange squares represent the predicted values. This allows you to visually observe the fluctuation trends between the predicted and true values ​​for each sample group, helping you determine the model's performance in prediction tasks like battery life.

[0168] Figure 6 The displayed scatter plot of predicted values ​​and true values ​​shows that the data points are closely distributed around the ideal line with small deviation, verifying the model's excellent ability in tracking the capacity decay trajectory. Figure 7 The residual plot further shows that the residuals are randomly distributed around zero without obvious systematic deviations, showing the robustness of the model to different degradation modes. Figure 8As shown in the figure, most prediction errors are concentrated below 0.05, with a narrow distribution, highlighting the consistency of the model prediction. The training and validation mean square errors of the five-fold validation converged stably after about 30 cycles, and the gap between the training and validation losses was very small, indicating that the L1-L2 joint regularization effectively prevented overfitting. The statistical analysis of the cross-validation results showed that the RMSE and R 2 The standard deviations of the two models are lower than 0.0015 and 0.008, respectively, confirming the generalization ability of the model on different data subsets. TPA significantly enhances XGBoost's ability to model the dynamic interaction of electrochemistry and thermodynamics by extracting high-level features from voltage, current, and temperature time series data. Figure 9 The comparison curve between the true value and the predicted value further demonstrates the model's ability to accurately track the degradation trajectory on 100 test samples.

[0169] In order to verify the superiority of the proposed BFA-TPA-XGBoost model, it was compared with a variety of advanced battery RUL prediction methods, including decision tree (DT), random forest (RF), support vector machine (SVM) and Gaussian process regression (GPR) in battery life prediction. For support vector machine, the Gaussian function was selected as the kernel function. For Gaussian process regression, the exponential kernel function was used. The same training and test data were used to evaluate the performance of these methods, and 5-fold cross-validation techniques were used to optimize the model parameters during training. Table 1 shows the prediction performance of all methods. From RMSE, MAPE, R 2 The proposed BFA-TPA-XGBoost method has the best prediction performance, with RMSE of 0.0247, MAPE of 5.68%, and R 2 As high as 0.9559, significantly better than other benchmark models. In comparison, although decision trees have advantages such as strong interpretability, they are prone to overfitting and poor generalization in battery nonlinear degradation modeling; although random forests have certain anti-overfitting capabilities, they are difficult to fully capture time series degradation information; SVM and GPR perform unstable in small sample scenarios, especially GPR, which has a significantly increased computational cost when processing high-dimensional features and noisy data, affecting the practicality of the model. Relatively speaking, the error distribution of BFA-TPA-XGBoost is narrower and concentrated in the low error area.

[0170] The advantages of the BFA-TPA-XGBoost model in error distribution, robustness, and computational efficiency are mainly attributed to three aspects. First, BFA realizes the intelligent screening of key degenerate features and significantly compresses the feature dimensions. Second, TPA effectively models the complex interactions between features, enhancing the model's temporal sensitivity and physical consistency. Third, XGBoost, as an integrated model with strong generalization capabilities, further improves prediction accuracy and stability.

[0171] Table 1 Performance comparison of different models

[0172]

[0173] Figure 10 A schematic structural diagram of a battery life prediction device based on a life prediction model provided in another embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0174] Reference Figure 10 , a battery life prediction device based on a life prediction model may include:

[0175] An acquisition module 1001 is configured to acquire a time series feature data set of a battery to be predicted, wherein the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data;

[0176] A screening module 1002 is configured to screen the time series feature data set to obtain a first feature set; the first feature set includes a plurality of key degradation features, and the key degradation features are used to characterize the core features of the battery to be predicted during the degradation process;

[0177] The processing module 1003 is configured to perform feature interaction processing on the time series feature dataset using an L1 regularization and an L2 regularization strategy and a multi-rank tensor decomposition structure to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature dataset;

[0178] A fusion module 1004 is configured to fuse the first feature set and the second feature set to obtain a fused feature set;

[0179] The determination module 1005 is configured to input the fused feature set into a life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fused feature set.

[0180] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application, and are devices corresponding to the above-mentioned method. All implementation methods in the above-mentioned method embodiment are applicable to the embodiment of the device. Its specific functions and the technical effects brought about can be found in the method embodiment part, which will not be repeated here.

[0181] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0182] Figure 11 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application is shown.

[0183] The device may include a processor 1101 and a memory 1102 storing program instructions.

[0184] When the processor 1101 executes the program, the steps in any of the above method embodiments are implemented.

[0185] For example, the program can be divided into one or more modules / units, one or more modules / units are stored in the memory 1102 and executed by the processor 1101 to complete the present application. One or more modules / units can be a series of program instruction segments that can perform specific functions, and the instruction segments are used to describe the execution process of the program in the device.

[0186] Specifically, the processor 1101 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0187] Memory 1102 may include a large capacity memory for data or instructions. By way of example and not limitation, memory 1102 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 1102 may include removable or non-removable (or fixed) media. Where appropriate, memory 1102 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, memory 1102 is a non-volatile solid-state memory.

[0188] The memory may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical or other physical / tangible memory storage devices. Thus, generally, the memory includes one or more tangible (non-transitory) readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to an aspect of the present disclosure.

[0189] The processor 1101 implements any one of the methods in the above embodiments by reading and executing program instructions stored in the memory 1102 .

[0190] In one example, the electronic device may further include a communication interface 1103 and a bus 1110. The processor 1101, the memory 1102, and the communication interface 1103 are connected via the bus 1110 and communicate with each other.

[0191] The communication interface 1103 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0192] Bus 1110 includes hardware, software or both, and the components of online data flow metering equipment are coupled to each other. For example, but not limitation, bus can include accelerated graphics port (AGP) or other graphics bus, enhanced industry standard architecture (EISA) bus, front side bus (FSB), hypertransport (HT) interconnection, industry standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 1110 can include one or more buses. Although the present application embodiment describes and shows a specific bus, the application considers any suitable bus or interconnection.

[0193] In addition, in combination with the methods in the above embodiments, embodiments of the present application may provide a storage medium for implementation. The storage medium stores program instructions; when the program instructions are executed by a processor, any one of the methods in the above embodiments is implemented.

[0194] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned method embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0195] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0196] An embodiment of the present application provides a computer program product, which is stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-mentioned method embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0197] It should be understood that the present application is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present application is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present application.

[0198] The functional modules shown in the above block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), suitable firmware, a plug-in unit, a function card or the like. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium that can store or transmit information. The example of a machine-readable medium includes an electronic circuit, a semiconductor memory device, a ROM, a flash memory, an erasable ROM (EROM), a floppy disk, a CD-ROM, an optical disk, a hard disk, an optical fiber medium, a radio frequency (RF) link, or the like. The code segment can be downloaded via a computer grid such as the Internet, an intranet, etc.

[0199] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps. In other words, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0200] Aspects of the present disclosure have been described above with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed via the processor of the computer or other programmable data processing device enable the implementation of the function / action specified in one or more boxes of the flowchart and / or block diagram. This processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit. It is also understood that each box in the block diagram and / or the flowchart and the combination of the boxes in the block diagram and / or the flowchart can also be implemented by the dedicated hardware that performs the specified function or action, or can be implemented by the combination of dedicated hardware and computer instructions.

[0201] The above is only a specific implementation method of the present application. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the scope of protection of the present application is not limited to this. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the scope of protection of this application.

Claims

1. A battery life prediction method based on a life prediction model, characterized in that: The method comprises: Acquire a time series feature data set of a battery to be predicted, where the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data; The time series feature data set is screened to obtain a first feature set; the first feature set includes a plurality of key degradation features, and the key degradation features are used to characterize core features of the battery to be predicted during the degradation process; By using L1 regularization and L2 regularization strategies and a multi-rank tensor decomposition structure, feature interaction processing is performed on the time series feature dataset to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature dataset; Fusing the first feature set and the second feature set to obtain a fused feature set; The fused feature set is input into a life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fused feature set.

2. The method according to claim 1, characterized in that The filtering of the time series feature data set to obtain a first feature set includes: Randomly generate a preset number of feature subsets based on the time series feature data set; one feature subset corresponds to one firefly individual; Based on each firefly individual, the firefly position of each firefly individual is updated through evolutionary iteration. After the evolutionary iteration update is completed, the firefly individual with the maximum fitness value is selected from the iteratively updated firefly individuals; the fitness value is used to characterize the prediction performance of the feature subset corresponding to the firefly individual for the remaining battery life; The feature subset corresponding to the firefly individual with the maximum fitness value is determined as the first feature set.

3. The method according to claim 2, characterized in that Each evolutionary iterative update includes: Get the Cartesian distance and attraction between individual fireflies; Performing a position update operation on each firefly individual to obtain an updated feature subset corresponding to each firefly individual; wherein the position update operation includes: updating the firefly position of each firefly individual according to the Cartesian distance and attraction between each firefly individual to obtain an updated firefly position; applying a normalization function to the continuous position values ​​in the updated firefly position and converting them into binary decisions through a thresholding function; and determining the updated feature subset corresponding to the firefly individual based on the binary decision; For the updated feature subsets corresponding to each firefly individual, a fitness value calculation operation is performed respectively to obtain the fitness value corresponding to each updated feature subset.

4. The method according to claim 1, wherein The L1 regularization and L2 regularization strategies are used to utilize a multi-rank tensor decomposition structure to perform feature interaction processing on the time series feature dataset to obtain a second feature set, including: Inputting the temporal feature dataset into a tensor product attention network, wherein the tensor product attention network includes a multi-rank tensor decomposition structure, and the multi-rank tensor decomposition structure is used to map the feature dataset to a multi-rank tensor space; L1 regularization and L2 regularization strategies are used to prevent overfitting during the training process of the tensor product attention network; Decomposing the time series feature dataset using the multi-rank tensor decomposition structure to obtain tensor product decomposition results corresponding to the query matrix, the key matrix, and the value matrix respectively; Calculating, based on the tensor product decomposition result, an attention score between multiple physical quantities in the time series feature dataset, wherein the attention score is used to characterize the importance and relevance of each physical quantity in the battery degradation process; According to the attention score, nonlinear interaction features between multiple physical quantities in the temporal feature dataset are determined, and the nonlinear interaction features are determined as the second feature set.

5. The method according to claim 1, wherein The lifespan prediction model is an XGBoost integrated model; Inputting the fused feature set into a life prediction model to obtain the remaining service life of the battery to be predicted includes: The fused feature set is input into the XGBoost integrated model, and the fused feature set is processed by the XGBoost integrated model to obtain the remaining service life of the battery to be predicted; wherein, the XGBoost integrated model optimizes the objective function of the XGBoost integrated model through a second-order Taylor expansion, and introduces L1 regularization and L2 regularization terms to suppress overfitting.

6. The method according to claim 1, characterized in that The lifespan prediction model is pre-trained in the following way: Obtaining a training sample set; the training sample set includes historical time series characteristics of the battery and the actual remaining service life corresponding to the battery; Performing feature screening on the historical time series features to obtain a first historical feature set; By using L1 regularization and L2 regularization strategies and a multi-rank tensor decomposition structure, feature interaction processing is performed on the historical time series features to obtain a second historical feature set; fusing the first historical feature set and the second historical feature set to obtain a historical fused feature set; Based on the historical fusion feature set, prediction is performed using the life prediction model to obtain a predicted remaining service life; Based on the actual remaining service life and the predicted remaining service life, a first loss function is calculated, and the life prediction model is trained based on the first loss function.

7. A battery life prediction device based on a life prediction model, characterized in that: The device comprises: An acquisition module, configured to acquire a time series feature data set of a battery to be predicted, wherein the time series feature data set includes at least one of voltage time series data, temperature time series data, current time series data, and capacity decay time series data; a screening module, configured to screen the time series feature data set to obtain a first feature set; the first feature set includes a plurality of key degradation features, the key degradation features being used to characterize core features of the battery to be predicted during the degradation process; a processing module, configured to perform feature interaction processing on the time series feature dataset using a multi-rank tensor decomposition structure through L1 regularization and L2 regularization strategies to obtain a second feature set; the second feature set includes nonlinear interaction features between multiple physical quantities in the time series feature dataset; a fusion module, configured to fuse the first feature set and the second feature set to obtain a fused feature set; The determination module is used to input the fusion feature set into a life prediction model to obtain the remaining service life of the battery to be predicted; the life prediction model is trained based on the historical fusion feature set.

8. An electronic device, characterized in that: The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the battery life prediction method based on the life prediction model according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the battery life prediction method based on the life prediction model according to any one of claims 1 to 6 is implemented.

10. A computer program product, characterized in that When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the battery life prediction method based on the life prediction model as described in any one of claims 1 to 6.

Images