Lithium iron phosphate battery remaining service life prediction method based on composite optimization algorithm

Abstract

The invention provides a lithium iron phosphate battery remaining service life prediction method based on a composite optimization algorithm, relates to the technical field of lithium iron phosphate battery service life management and control, and solves the problems of insufficient precision, poor robustness and limited multi-working-condition adaptability in an existing method. The method comprises the following steps: acquiring life cycle parameter data of a battery and forming a data set, constructing a battery degradation model based on an EKF technology, inputting the data set, performing global search optimization through a GWO algorithm, and jointly optimizing parameters of the battery degradation model under a Bayesian optimization framework in combination with a GA algorithm and a DRL; state parameter data in the battery using process are collected in real time, the current battery degradation state is calculated in an iterative updating optimization mode, and therefore the time point when the lithium iron phosphate battery reaches the preset failure threshold value is predicted, and the remaining service life of the lithium iron phosphate battery is predicted. According to the method, the modeling precision of the degradation process of the lithium iron phosphate battery and the RUL prediction accuracy can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of lithium iron phosphate battery life management technology, specifically to a method for predicting the remaining lifespan of lithium iron phosphate batteries based on a composite optimization algorithm. Background Technology

[0002] The development and application of lithium-ion battery technology has become a core support for the new energy industry, especially in the fields of electric vehicles and large-scale energy storage systems. With the continuous iteration of battery material systems, lithium iron phosphate systems have gained widespread application due to their excellent thermal stability and long cycle life. However, the performance degradation mechanism of batteries under complex operating conditions still restricts the reliability and economy of the system. How to accurately assess the battery health status and predict the remaining service life has become a technical bottleneck that urgently needs to be overcome.

[0003] In battery lifecycle management, capacity decay and internal resistance growth constitute the main degradation characteristics. This degradation process is influenced by multi-physics coupling effects and exhibits a significant nonlinear evolution law. When batteries experience different charge / discharge rates, temperature fluctuations, and changes in storage conditions, microscopic damage such as deterioration of the active material structure and thickening of the solid electrolyte interface film gradually accumulates, ultimately leading to a decrease in usable capacity and a decline in kinetic performance. Existing research shows that it is difficult to accurately capture the degradation trajectory under complex operating conditions using only a single electrochemical parameter or equivalent circuit model.

[0004] Traditional battery life prediction methods primarily rely on empirical formula fitting or mechanistic modeling. The former depends on extrapolating trends from historical data, making it prone to significant biases when faced with sudden changes in operating conditions; while the latter, although reflecting the essence of electrochemical reactions, involves cumbersome parameter identification and consumes substantial computational resources. As battery system integration scales up, these methods face dual challenges in terms of real-time performance and universality. In recent years, data-driven prediction techniques based on support vector machines and recurrent neural networks have emerged, but their generalization ability is limited by the completeness of training samples and they lack robustness to abnormal operating conditions.

[0005] The integration of optimization algorithms and intelligent learning techniques offers new insights into solving these problems. Some studies have attempted to improve the parameter optimization efficiency of neural networks using particle swarm optimization, while others have utilized genetic algorithms to enhance the adaptive capabilities of degradation models. However, single optimization strategies are prone to getting trapped in local optima and struggle to balance global search speed with convergence speed. While deep reinforcement learning demonstrates advantages in dynamic decision-making, its reward function design and environment interaction mechanisms have yet to provide an effective solution for addressing battery degradation characteristics. Existing technologies indicate that constructing a prediction framework that combines accuracy and stability remains a key challenge for those skilled in the art.

[0006] Current battery management systems (BMS) place higher demands on predictive models, requiring them to adapt to the characteristic evolution patterns of different aging stages while ensuring real-time response. Existing predictive methods often suffer from error accumulation when processing multi-scale degradation data due to rigid model structures or insufficient feature extraction. Particularly in scenarios with increasing battery pack inconsistencies, a mature technical solution for the synergistic optimization of individual differences and system-level prediction has yet to be developed. This necessitates the industry's urgent exploration of novel predictive architectures that integrate multimodal data and possess dynamic adaptability to address the complex and ever-changing usage environments in real-world applications. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing battery remaining useful life (RUL) prediction methods, such as insufficient accuracy, poor robustness, and limited adaptability to various operating conditions, when dealing with complex nonlinear degradation processes. Therefore, this invention proposes a lithium iron phosphate battery RUL prediction method based on a composite optimization algorithm. By integrating a multimodal optimization strategy and a dynamic feature learning mechanism, this invention overcomes the limitations of traditional single-method approaches, such as sensitivity to complex operating condition disturbances, one-sided feature extraction, and error accumulation effects. This effectively improves the modeling accuracy of the lithium iron phosphate battery degradation process and the accuracy of its RUL prediction.

[0008] The present invention employs the following technical solutions to achieve its objective: A method for predicting the remaining service life of lithium iron phosphate batteries based on a composite optimization algorithm includes the following steps: S1. Obtain and preprocess the life cycle parameter data of lithium iron phosphate batteries to form a multivariate time series dataset; S2. Based on the extended Kalman filter (EKF) technique, construct a battery degradation model and set initial parameters; S3. Using a multivariate time series dataset as input data, the Grey Wolf Optimization (GWO) algorithm is used for global search optimization. At the same time, the Genetic Algorithm (GA) and Deep Reinforcement Learning (DRL) are combined to jointly optimize the parameters of the battery degradation model under the Bayesian optimization framework. S4. Real-time acquisition of state parameter data of lithium iron phosphate battery during use, input into battery degradation model, and calculation of current battery degradation state through iterative update optimization, thereby predicting the time point when the lithium iron phosphate battery reaches the preset failure threshold and realizing the prediction of its remaining service life.

[0009] Preferably, after step S4, the method further includes: S5. Obtain the actual failure time of the lithium iron phosphate battery. Based on the previously predicted remaining service life, form a prediction error index. Process the prediction error index through deep reinforcement learning (DRL) to further optimize and update the parameters of the battery degradation model.

[0010] Preferably, after step S5, the method further includes: S6. Obtain the predicted remaining lifespan of the lithium iron phosphate battery after each optimization of the battery degradation model, present it to the outside world through the battery management system (BMS), and adjust the battery usage strategy according to the predicted results under the preset optimization objectives.

[0011] Specifically, in step S1, the charging and discharging current of the lithium iron phosphate battery is obtained through a sensor installed inside the battery pack. Charging and discharging voltage Internal resistance The system records the battery cycle count, charge / discharge SOC depth, and operating environment parameters, along with temperature, and stores the data in chronological order. During the data acquisition and recording process, wavelet transform or sliding window value filter is used to remove high-frequency noise and outliers from the data, and the data is normalized. For missing points in the data, time series interpolation or weighted estimation based on adjacent time steps is used to fill in the missing points. In this way, the acquisition and preprocessing of lithium iron phosphate battery life cycle parameter data are completed, forming a multivariate time series dataset.

[0012] Furthermore, in step S2, the battery degradation model is an extended Kalman filter (EKF) model; when constructing the EKF model, the key state variables of battery degradation are first selected. It is caused by the battery's internal resistance and capacity Composition, that is subscript Representing the One time step; The state transition function of the EKF model is then defined as a nonlinear function based on the battery physical degradation mechanism, and its form is: ,in To input control variables, This is process noise; Further define the measurement function of the EKF model, and establish the corresponding measurement equation based on the sensor observation process, which takes the form of: ,in For observation purposes, To observe noise; Finally, based on these two types of functions and The corresponding state transition matrix is ​​constructed. and measurement matrix These two types of matrices are used for prediction and updating of the EKF model during the iterative process, thereby completing the construction of the EKF model; After construction, the initial state estimates of the EKF model are set based on the multivariate time series dataset. State covariance matrix Process noise covariance and observation noise covariance , which serves as the initial condition for the EKF model during the iterative process.

[0013] Preferably, in step S3, the Gray Wolf Optimization Algorithm (GWO) is used for global search optimization to adjust the process noise covariance in the EKF model. Observation noise covariance And the parameters in the state transition function; the optimization objective is to minimize the prediction error of the remaining lifetime; When optimizing using GWO, GWO algorithm instances are run in parallel in multiple independent search subspaces to ensure the diversity of the solution space. For multiple candidate solutions obtained by the GWO algorithm, the merits of each candidate solution are evaluated by defining the fitness function as the root mean square error of the remaining lifetime prediction error.

[0014] Preferably, in step S3, the genetic algorithm GA performs selection, crossover, and mutation operations on the population to explore the global regions of multiple solution spaces in the GWO algorithm; the deep reinforcement learning DRL dynamically adjusts the probability distribution of genetic operators in the genetic algorithm GA through feedback from its agent learning environment, thereby achieving adaptive adjustment of the optimization process; the parameters in the EKF model are optimized under the alternating action of the genetic algorithm GA and the deep reinforcement learning DRL.

[0015] Preferably, in step S3, the parameter tuning process of the EKF model is regarded as solving a black-box function optimization problem through the Bayesian optimization framework. In the process of solving this optimization problem, Gaussian process regression (GPR) is used to construct a performance prediction surrogate model to describe the probability distribution of the parameter space of the EKF model and its remaining useful life prediction accuracy, which serves as the basis for optimization.

[0016] Preferably, in step S3, the parameter optimization objectives of the EKF model include minimizing the remaining lifespan prediction error, maximizing battery lifespan, and improving charge and discharge efficiency. When performing the multi-objective optimization process, a multi-objective evolutionary algorithm based on Pareto optimal solution is determined. Through Pareto front search, a set of non-dominated solutions is obtained for model decision-making, ensuring the performance trade-off between different optimization objectives.

[0017] Specifically, in step S4, the iterative update and optimization process of the EKF model includes two parts: prediction and update. In the prediction section, the state prediction equation and the covariance matrix prediction equation are as follows:

[0018]

[0019] In the equation, Represents time step Prior state estimation, which is based on time steps The posterior state estimate is derived from the input control variables, i.e. This is the state transition function. For time step Posterior state estimation, For time step Input control variables; Represents time step The prior state covariance matrix is ​​used to represent the uncertainty of the state estimation; Represents the time step The calculated state transition matrix; Represents time step The posterior state covariance matrix; Here, denoted as process noise covariance, is used to represent the impact of process noise on state estimation. In the update section, the Kalman gain calculation formula and the state update equation are as follows:

[0020]

[0021] In the equation, Represents time step The calculated Kalman gain is used to balance the weights of the predicted and observed values; For time step The calculated measurement matrix; The observation noise covariance is used to represent the impact of observation noise on the measured value. Represents time step The posterior state estimate, which combines the predicted values ​​and new observations, is derived as follows: For time step Observations For measurement functions; According to time step Posterior state estimation The battery degradation state represented by the value can predict the time point when the lithium iron phosphate battery reaches the preset failure threshold. The prediction process adopts Monte Carlo sampling or sliding window method, combined with the parameter uncertainty of EKF model, and performs multiple rounds of iteration. Finally, the probability distribution and confidence interval of the remaining service life are output to complete the prediction process.

[0022] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention employs an Extended Kalman Filter (EKF) model for dynamic estimation of the degradation state of lithium iron phosphate batteries, addressing the nonlinearity issues inherent in the degradation process. Simultaneously, by combining the Grey Wolf Optimization (GWO) algorithm and the Genetic Algorithm (GA), global and local optimizations are achieved, precisely adjusting model parameters and reducing the prediction error of the Residual Limiting (RUL). The Deep Reinforcement Learning (DRL) and Bayesian optimization techniques applied in the method adaptively adjust the model, thereby improving the robustness and accuracy of the prediction system. Furthermore, multi-objective optimization techniques can simultaneously optimize multiple indicators such as battery life, charging efficiency, and degradation rate, ensuring optimal battery performance under various operating conditions. Through real-time data acquisition and intelligent decision support, this invention provides accurate RUL prediction results for battery management systems, extending battery life and reducing maintenance costs.

[0023] This invention significantly improves the prediction accuracy and robustness of RUL (Range Limiting) for lithium iron phosphate (LFP) batteries through a combination of various optimization algorithms. The method accurately estimates the battery's degradation state, and based on the prediction results, battery performance can be optimized under different operating conditions, thereby extending battery life and improving charge / discharge efficiency. Adjustments to battery management strategies based on this invention can reduce battery maintenance costs, enhance battery safety, and ultimately provide reliable health and safety management support for electric vehicles and energy storage systems. Attached Figure Description

[0024] The present invention will be further described in detail with reference to the following figures, specifically including one figure as follows: Figure 1 This is a schematic diagram illustrating the overall process of the method of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] A method for predicting the remaining lifespan of lithium iron phosphate batteries based on a composite optimization algorithm. Figure 1 This document provides a brief overview of the overall process of the method, which can be viewed concurrently. The key steps of the method can be summarized as follows: S1. Obtain and preprocess the life cycle parameter data of lithium iron phosphate batteries to form a multivariate time series dataset; S2. Based on the extended Kalman filter (EKF) technique, construct a battery degradation model and set initial parameters; S3. Using a multivariate time series dataset as input data, the Grey Wolf Optimization (GWO) algorithm is used for global search optimization. At the same time, the Genetic Algorithm (GA) and Deep Reinforcement Learning (DRL) are combined to jointly optimize the parameters of the battery degradation model under the Bayesian optimization framework. S4. Real-time acquisition of state parameter data of lithium iron phosphate battery during use, input into battery degradation model, and calculation of current battery degradation state through iterative update optimization, thereby predicting the time point when the lithium iron phosphate battery reaches the preset failure threshold and realizing the prediction of its remaining service life.

[0028] This implementation method will describe the details of each step in the order described above.

[0029] First, battery data needs to be collected and preprocessed. The collection and preprocessing methods are the same for both data used for model training and data used as the basis for prediction in practical applications. This implementation uses a high-precision sensor installed within the battery pack to acquire the charging and discharging current of the lithium iron phosphate battery. Charging and discharging voltage Internal resistance For key parameters such as temperature, the sampling frequency can be set to a range of 1Hz to 10Hz to balance data real-time performance and computational resource consumption. Simultaneously, it is necessary to record multi-dimensional information such as battery cycle count, charge / discharge state of charge (SOC) depth, and operating environment parameters, and store all the collected data in chronological order.

[0030] Preprocessing is required during data collection and recording: wavelet transform or sliding window midpoint filter is used to remove high-frequency noise and outliers from the data; the data is normalized using min-max standardization or Z-score standardization to ensure that parameters of different dimensions are on a uniform scale; for missing points in the data, time series interpolation or weighted estimation based on adjacent time steps is used to fill in the missing points, thereby avoiding data gaps from affecting model training or data application. Time series interpolation can use mature methods such as linear interpolation or spline interpolation, and there are no restrictions here.

[0031] Step S2 in this embodiment is a key step in constructing the battery degradation model. The battery degradation model is an Extended Kalman Filter (EKF) model; when constructing the EKF model, the key state variables of battery degradation are first selected. It is caused by the battery's internal resistance and capacity Composition, that is subscript Representing the Each time step is used to distinguish between different scenarios during the iteration process.

[0032] The state transition function of the EKF model is then defined as a nonlinear function based on the battery physical degradation mechanism, and its form is: ,in Input control variables, such as the previously collected charging and discharging current. ; As process noise, this implementation assumes that it follows a zero-mean Gaussian distribution.

[0033] Further define the measurement function of the EKF model, and establish the corresponding measurement equation based on the sensor observation process, which takes the form of: ,in For observation purposes, To observe noise.

[0034] Finally, based on these two types of functions and The corresponding state transition matrix is ​​constructed. and measurement matrix These two types of matrices are used for prediction and updating of the EKF model during the iterative process, thereby completing the construction of the EKF model.

[0035] In this embodiment, after the model is constructed, the initial state estimate of the EKF model is set based on the historical data in the multivariate time series dataset and the preliminary results after experimental calibration. State covariance matrix Process noise covariance and observation noise covariance , which serves as the initial condition for the EKF model during the iterative process.

[0036] After the above process, the EKF model is completed, but it still needs to be optimized by combining and applying various optimization algorithms. Specifically, step S3 is used to make the model achieve better prediction results for battery degradation state.

[0037] In this embodiment, the Grey Wolf Optimization Algorithm (GWO) is used for global search optimization to adjust the process noise covariance in the EKF model. Observation noise covariance The parameters in the state transition function are also considered; the optimization objective is to minimize the prediction error of the remaining lifetime.

[0038] When optimizing using GWO, GWO algorithm instances are run in parallel in multiple independent search subspaces to ensure the diversity of the solution space. For multiple candidate solutions obtained by the GWO algorithm, the merits of each candidate solution are evaluated by defining the fitness function as the root mean square error of the remaining lifetime prediction error.

[0039] To improve the estimation accuracy and prediction stability of the EKF model parameters, this design employs a combination of the Grey Wolf Optimization Algorithm (GWO) and various intelligent optimization methods. The process first utilizes GWO's global search capability to find optimal or near-optimal solutions in a high-dimensional parameter space. GWO is a biomimetic intelligent optimization algorithm inspired by the social hierarchy and cooperative behavior of grey wolf packs during hunting. By simulating this behavioral mechanism, the algorithm can effectively approximate the optimal parameter combination in a complex solution space. In this process, the optimization objective is to minimize the remaining lifetime prediction error; therefore, the fitness function is defined as the root mean square error between the predicted result and the actual remaining lifetime, serving as a criterion for evaluating candidate quality.

[0040] Similarly, in step S3, the genetic algorithm GA performs selection, crossover, and mutation operations on the population to explore the global regions of multiple solution spaces in the GWO algorithm; the deep reinforcement learning DRL dynamically adjusts the probability distribution of genetic operators in the genetic algorithm GA through feedback from its agent learning environment, thereby achieving adaptive adjustment of the optimization process; the parameters in the EKF model are optimized under the alternating action of the genetic algorithm GA and the deep reinforcement learning DRL, thereby improving the convergence speed and the diversity of solutions.

[0041] The introduction of Genetic Algorithm (GA) and Deep Reinforcement Learning (DRL) further improves the efficiency and accuracy of EKF model parameter optimization. GA can systematically explore multiple solution spaces involved in the gray wolf optimization process. By simulating selection, crossover, and mutation operations in biological evolution, GA can continuously approach better solution regions while maintaining population diversity. This mechanism effectively compensates for the search blind spots that may exist in single optimization methods, enhancing the global optimization capability of the overall optimization process. Building on this, Deep Reinforcement Learning (DRL) is used to construct an intelligent optimization environment with adaptive adjustment capabilities. The DRL agent dynamically adjusts the application probability distribution of various genetic operators in GA through perception and feedback learning of the current optimization state. This allows the optimization process to automatically adjust its strategy based on the position and trend of the current solution, thereby improving convergence speed and optimization quality.

[0042] This implementation also employs a Bayesian optimization framework, treating the parameter tuning process of the EKF model as solving a black-box function optimization problem. In solving this optimization problem, Gaussian process regression (GPR) is used to construct a performance prediction surrogate model, describing the probability distribution of the EKF model's parameter space and its remaining lifespan prediction accuracy, serving as the basis for optimization. Furthermore, this process also guides parameter search by collecting performance evaluation samples, calculating the posterior distribution and expected improvement indicators, and capturing the nonlinearity and uncertainty in the battery degradation process.

[0043] The Bayesian optimization framework embeds the entire optimization process, transforming the EKF model parameter tuning problem into a typical black-box function optimization task. This framework continuously accumulates historical optimization data and utilizes Gaussian Process Regression (GPR) to build a performance prediction surrogate model, describing the probability distribution of model prediction performance under different parameter combinations. This surrogate model not only reflects the prediction accuracy at each point in the parameter space but also quantifies its uncertainty, providing a statistical basis for subsequent optimization decisions. Based on this model, the Bayesian optimization strategy can intelligently select the next most promising parameter combination for evaluation, thus efficiently guiding the entire optimization process towards the optimal parameter configuration. This approach significantly improves the automation and robustness of parameter tuning, giving the EKF model stronger generalization ability and higher prediction accuracy.

[0044] In this embodiment, the parameter optimization objectives of the EKF model include minimizing the remaining lifespan prediction error, maximizing battery lifespan, and improving charge and discharge efficiency. When performing the multi-objective optimization process, a multi-objective evolutionary algorithm based on Pareto optimal solution is determined. Through Pareto front search, a set of non-dominated solutions are obtained for model decision-making, ensuring the performance trade-off between different optimization objectives.

[0045] The parameter optimization of the EKF model does not merely pursue the optimal result of a single objective, but comprehensively considers multiple interrelated and potentially conflicting optimization objectives. These objectives include minimizing the prediction error of remaining battery life, extending the overall battery life, and improving the charging and discharging efficiency of the battery during use. To achieve simultaneous optimization of these multiple performance aspects, the method here also introduces a multi-objective evolutionary algorithm based on Pareto optimal solutions to replace the traditional single-objective optimization approach.

[0046] This multi-objective optimization process employs the Pareto front search mechanism to find a set of non-dominated solutions in a complex parameter space—that is, a set of solutions that cannot further improve another objective without sacrificing one. This set of solutions provides the model with multiple feasible parameter configuration schemes, enabling the system to flexibly select and make decisions based on the needs of real-world application scenarios. This approach not only enhances the model's adaptability and robustness but also effectively ensures the balance and coordination between different optimization objectives, thereby strengthening the practicality and engineering application value of the prediction results.

[0047] In step S4 of this embodiment, the iterative update and optimization process of the EKF model includes two parts: prediction and update. In the prediction section, the state prediction equation and the covariance matrix prediction equation are as follows:

[0048]

[0049] In the equation, Represents time step Prior state estimation, which is based on time steps The posterior state estimate is derived from the input control variables, i.e. This is the state transition function. For time step Posterior state estimation, For time step Input control variables; Represents time step The prior state covariance matrix is ​​used to represent the uncertainty of the state estimation; Represents the time step The calculated state transition matrix; Represents time step The posterior state covariance matrix; Here, denoted as process noise covariance, is used to represent the impact of process noise on state estimation. In the update section, the Kalman gain calculation formula and the state update equation are as follows:

[0050]

[0051] In the equation, Represents time step The calculated Kalman gain is used to balance the weights of the predicted and observed values; For time step The calculated measurement matrix; The observation noise covariance is used to represent the impact of observation noise on the measured value. Represents time step The posterior state estimate, which combines the predicted values ​​and new observations, is derived as follows: For time step Observations For measurement functions; According to time step Posterior state estimation The battery degradation state represented by the value can predict the time point when the lithium iron phosphate battery reaches the preset failure threshold. The prediction process adopts Monte Carlo sampling or sliding window method, combined with the parameter uncertainty of EKF model, and performs multiple rounds of iteration. Finally, the probability distribution and confidence interval of the remaining service life are output to complete the prediction process.

[0052] As a preferred embodiment, such as Figure 1 As shown, after step S4, the method further includes: S5. Obtain the actual failure time of the lithium iron phosphate battery. Based on the previously predicted remaining service life, form a prediction error index. Process the prediction error index through deep reinforcement learning (DRL) to further optimize and update the parameters of the battery degradation model.

[0053] Prediction error metrics can include absolute error, mean squared error, etc. These metrics are fed back to the Deep Reinforcement Learning (DRL), where the policy gradient algorithm can fine-tune the parameters of the EKF model online based on the metric information, that is, dynamically adjust the state transition matrix according to the error feedback. Process noise covariance and observation noise covariance , and Also in matrix form, parameters are further corrected through gradient descent or Bayesian update strategies. The agent in Deep Reinforcement Learning (DRL) receives prediction error metrics as reward feedback, and optimizes the model's adaptability through policy updates, improving robustness under environmental changes or abnormal operating conditions. This approach enables the EKF model to synchronize with the battery degradation process, improving future prediction accuracy.

[0054] As a preferred embodiment, such as Figure 1 As shown, after step S5, the method further includes: S6. Obtain the predicted remaining lifespan of the lithium iron phosphate battery after each optimization of the battery degradation model, present it to the outside world through the battery management system (BMS), and adjust the battery usage strategy according to the predicted results under the preset optimization objectives.

[0055] After the Battery Management System (BMS) presents the remaining lifespan prediction results in real time, maintenance personnel can further assess the health status of the lithium iron phosphate batteries based on this. Furthermore, based on the remaining lifespan prediction results, the BMS can dynamically adjust parameters such as charging current limits and depth of charge / discharge to optimize battery efficiency and slow down capacity degradation. Combining the optimal parameter combination output from the multi-objective optimization approach, the model balances battery lifespan, charging efficiency, and degradation rate, effectively achieving multi-dimensional optimization of battery management to meet different operating conditions and requirements.

Claims

1. A method for predicting the remaining service life of lithium iron phosphate batteries based on a composite optimization algorithm, characterized in that, The steps include: S1. Obtain and preprocess the life cycle parameter data of lithium iron phosphate batteries to form a multivariate time series dataset; S2. Based on the extended Kalman filter (EKF) technique, construct a battery degradation model and set initial parameters; S3. Using a multivariate time series dataset as input data, the Grey Wolf Optimization (GWO) algorithm is used for global search optimization. At the same time, the Genetic Algorithm (GA) and Deep Reinforcement Learning (DRL) are combined to jointly optimize the parameters of the battery degradation model under the Bayesian optimization framework. S4. Real-time acquisition of state parameter data of lithium iron phosphate battery during use, input into battery degradation model, and calculation of current battery degradation state through iterative update optimization, thereby predicting the time point when the lithium iron phosphate battery reaches the preset failure threshold and realizing the prediction of its remaining service life.

2. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 1, characterized in that, After step S4, the method further includes: S5. Obtain the actual failure time of the lithium iron phosphate battery. Based on the previously predicted remaining service life, form a prediction error index. Process the prediction error index through deep reinforcement learning (DRL) to further optimize and update the parameters of the battery degradation model.

3. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 2, characterized in that, After step S5, the method further includes: S6. Obtain the predicted remaining lifespan of the lithium iron phosphate battery after each optimization of the battery degradation model, present it to the outside world through the battery management system (BMS), and adjust the battery usage strategy according to the predicted results under the preset optimization objectives.

4. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 1, characterized in that: In step S1, the charging and discharging current of the lithium iron phosphate battery is obtained through a sensor installed inside the battery pack. Charging and discharging voltage Internal resistance The system records the battery cycle count, charge / discharge SOC depth, and operating environment parameters, along with temperature, and stores the data in chronological order. During the data acquisition and recording process, wavelet transform or sliding window value filter is used to remove high-frequency noise and outliers from the data, and the data is normalized. For missing points in the data, time series interpolation or weighted estimation based on adjacent time steps is used to fill in the missing points. In this way, the acquisition and preprocessing of lithium iron phosphate battery life cycle parameter data are completed, forming a multivariate time series dataset.

5. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 1, characterized in that: In step S2, the battery degradation model is an Extended Kalman Filter (EKF) model; when constructing the EKF model, the key state variables of battery degradation are first selected. It is caused by the battery's internal resistance and capacity Composition, that is subscript Representing the One time step; The state transition function of the EKF model is then defined as a nonlinear function based on the battery physical degradation mechanism, and its form is: ,in To input control variables, This is process noise; Further define the measurement function of the EKF model, and establish the corresponding measurement equation based on the sensor observation process, which takes the form of: ,in For observation purposes, To observe noise; Finally, based on these two types of functions and The corresponding state transition matrix is ​​constructed. and measurement matrix These two types of matrices are used for prediction and updating of the EKF model during the iterative process, thereby completing the construction of the EKF model; After construction, the initial state estimates of the EKF model are set based on the multivariate time series dataset. State covariance matrix Process noise covariance and observation noise covariance , which serves as the initial condition for the EKF model during the iterative process.

6. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 5, characterized in that: In step S3, the Grey Wolf Optimization (GWO) algorithm is used for global search optimization to adjust the process noise covariance in the EKF model. Observation noise covariance And the parameters in the state transition function; The optimization objective is to minimize the prediction error of the remaining useful life; When optimizing using GWO, GWO algorithm instances are run in parallel in multiple independent search subspaces to ensure the diversity of the solution space. For multiple candidate solutions obtained by the GWO algorithm, the merits of each candidate solution are evaluated by defining the fitness function as the root mean square error of the remaining lifetime prediction error.

7. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 5, characterized in that: In step S3, the genetic algorithm GA performs selection, crossover, and mutation operations on the population to explore the global regions of multiple solution spaces in the GWO algorithm; the deep reinforcement learning DRL dynamically adjusts the probability distribution of genetic operators in the genetic algorithm GA through feedback from its agent learning environment, thereby achieving adaptive adjustment of the optimization process; the parameters in the EKF model are optimized under the alternating action of the genetic algorithm GA and the deep reinforcement learning DRL.

8. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 5, characterized in that: In step S3, the parameter tuning process of the EKF model is regarded as solving a black-box function optimization problem through the Bayesian optimization framework. In the process of solving this optimization problem, Gaussian process regression (GPR) is used to construct a performance prediction surrogate model to describe the probability distribution of the parameter space of the EKF model and its remaining useful life prediction accuracy, which serves as the basis for optimization.

9. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 5, characterized in that: In step S3, the parameter optimization objectives of the EKF model include minimizing the remaining lifespan prediction error, maximizing battery lifespan, and improving charge and discharge efficiency. When performing the multi-objective optimization process, a multi-objective evolutionary algorithm based on Pareto optimal solution is determined. Through Pareto front search, a set of non-dominated solutions is obtained for model decision-making, ensuring the performance trade-off between different optimization objectives.

10. The method for predicting the remaining service life of a lithium iron phosphate battery according to claim 5, characterized in that: In step S4, the iterative update and optimization process of the EKF model includes two parts: prediction and update. In the prediction section, the state prediction equation and the covariance matrix prediction equation are as follows: In the equation, Represents time step Prior state estimation, which is based on time steps The posterior state estimate is derived from the input control variables, i.e. This is the state transition function. For time step Posterior state estimation, For time step Input control variables; Represents time step The prior state covariance matrix is ​​used to represent the uncertainty of the state estimation; Represents the time step The calculated state transition matrix; Represents time step The posterior state covariance matrix; Here, denoted as process noise covariance, is used to represent the impact of process noise on state estimation. In the update section, the Kalman gain calculation formula and the state update equation are as follows: In the equation, Represents time step The calculated Kalman gain is used to balance the weights of the predicted and observed values; For time step The calculated measurement matrix; The observation noise covariance is used to represent the impact of observation noise on the measured value. Represents time step The posterior state estimate, which combines the predicted values ​​and new observations, is derived as follows: For time step Observations For measurement functions; According to time step Posterior state estimation The battery degradation state represented by the value can predict the time point when the lithium iron phosphate battery reaches the preset failure threshold. The prediction process adopts Monte Carlo sampling or sliding window method, combined with the parameter uncertainty of EKF model, and performs multiple rounds of iteration. Finally, the probability distribution and confidence interval of the remaining service life are output to complete the prediction process.

Images