Power distribution terminal storage battery health state evaluation method based on multi-parameter fusion

Abstract

The invention relates to a power distribution terminal storage battery health state evaluation method based on multi-parameter fusion, and belongs to the technical field of power equipment operation and maintenance. Comprising the following steps: acquiring discharge depth and discharge times of deep discharge by collecting discharge data of a power distribution terminal storage battery in a frequent power failure scene, and obtaining a discharge characteristic data set; analyzing the discharge characteristic data set to obtain a voltage recovery curve after deep discharge, and determining dynamic recovery characteristics according to the voltage recovery curve; if the voltage recovery time of the dynamic recovery feature exceeds a preset threshold value, extracting corresponding discharge depth and discharge times from the discharge feature data set, and obtaining an internal resistance parameter change trend; analyzing discharge data in a frequent power failure scene through the retrained health state evaluation model, determining an alternating priority, and generating a battery alternating plan for operation and maintenance management optimization; accurate evaluation of the health state of the storage battery and operation and maintenance strategy optimization are realized.

Description

TECHNICAL FIELD

[0001] The application relates to a power distribution terminal battery health state evaluation method based on multi-parameter fusion and belongs to the technical field of power equipment operation and maintenance. BACKGROUND

[0002] In a power distribution system, a battery is an important power supply guarantee for a power distribution terminal such as a feeder terminal unit (FTU) and a power distribution terminal unit (DTU), and the health state thereof is directly related to the stability and reliability of power grid operation. With the development of smart grids, accurate evaluation of the health state of the battery has become a key to improving the level of power distribution automation and reducing operation and maintenance costs.

[0003] However, existing evaluation methods mostly rely on a single parameter such as internal resistance or terminal voltage, and it is difficult to comprehensively reflect the dynamic performance of the battery under complex working conditions, especially after frequent power outages cause multiple deep discharges, and there is still a deficiency in capturing the performance degradation characteristics of the battery.

[0004] These methods often ignore the dynamic changes of the discharge characteristics and voltage recovery curves, resulting in inaccurate evaluation results and difficulty in meeting actual operation and maintenance requirements. In the frequent power outage scenario, the power distribution terminal battery is repeatedly deeply discharged due to continuous faults, and the dynamic characteristics such as discharge depth and voltage recovery time will change significantly, and these changes are closely related to the remaining life of the battery.

[0005] The core challenges of the current technology are: 1) how to accurately capture the correlation changes between the discharge depth and the voltage recovery time, 2) how to effectively integrate these dynamic characteristics with static parameters such as internal resistance and terminal voltage, and 3) how to quantify the degradation trend of the battery health state through multi-parameter analysis. These challenges are due to the complexity of battery performance under dynamic working conditions, and the unsolved problems make it difficult for the operation and maintenance department to develop a scientific battery replacement plan, increasing the risk of equipment failure.

[0006] Therefore, how to accurately capture the dynamic correlation between the discharge depth and the voltage recovery time through a monitoring device after multiple deep discharges caused by frequent power outages, and integrate the dynamic and static characteristics based on a multi-parameter fusion algorithm to accurately evaluate the remaining life of the power distribution terminal battery has become a key problem for optimizing operation and maintenance strategies. SUMMARY

[0007] According to the problems described in the background, the application aims to solve the problems in the above-mentioned by providing a power distribution terminal battery health state evaluation method based on multi-parameter fusion.

[0008] To achieve the above-mentioned purpose, the application provides the following technical scheme: a power distribution terminal battery health state evaluation method based on multi-parameter fusion, comprising the following steps:

[0009] (1) By collecting the discharge data of the distribution terminal battery under frequent power outage scenarios, the discharge depth and discharge times of deep discharge are obtained to obtain the discharge characteristic data set;

[0010] (2) Analyze the discharge characteristic data set to obtain the voltage recovery curve after deep discharge, and determine the dynamic recovery characteristics based on the voltage recovery curve;

[0011] (3) If the voltage recovery time of the dynamic recovery feature exceeds the preset threshold, the corresponding discharge depth and discharge times are extracted from the discharge feature data set to obtain the change trend of the internal resistance parameter;

[0012] (4) Integrate the dynamic recovery characteristics and the change trend of the internal resistance parameters to construct a multi-parameter data set. The multi-parameter data set includes discharge depth, voltage recovery time, internal resistance value and terminal voltage fluctuation. The multi-parameter data set is standardized to obtain a comprehensive feature matrix;

[0013] (5) The comprehensive feature matrix is ​​weighted by a multi-parameter fusion algorithm to identify the nonlinear relationship between discharge depth and voltage recovery time, and the feature weights and correlation coefficients obtained by the internal resistance are quantified. The health status feature vector is obtained based on the feature weights and correlation coefficients, and a health status assessment model is constructed.

[0014] (6) Extract the weight of the internal resistance feature from the health state feature vector, calculate the degree of deviation between the voltage recovery time and the standard recovery time, and combine the degree of deviation and the internal resistance weight to predict the remaining battery life and obtain the life prediction value;

[0015] (7) Analyze the correlation between discharge depth and voltage recovery time, adjust the feature weights in the multi-parameter fusion algorithm according to the life prediction value, optimize the correlation analysis, and retrain the health status assessment model through the adjusted algorithm;

[0016] (8) The retrained health status assessment model is used to analyze the discharge data under frequent power outage scenarios, determine the rotation priority, and generate a battery rotation plan.

[0017] Preferably, the step (1) comprises the following steps:

[0018] (1.1) collecting voltage sampling point data during the discharge process, and processing the sampling point data using a median filter method to obtain initial discharge characteristic data;

[0019] (1.2) monitoring the discharge current according to the initial discharge characteristic data, and calculating the remaining capacity percentage by multiplying the current sampling value by the time to obtain the discharge number characteristic data;

[0020] (1.3) Based on the discharge number characteristic data and the discharge cut-off voltage data, the average load power is calculated using a sliding time window to obtain discharge energy characteristic data;

[0021] (1.4) classifying the discharge conditions according to the discharge energy characteristic data to obtain a data set of corresponding relationships between discharge depth and discharge times;

[0022] (1.5) Using the support vector regression method to model the corresponding relationship dataset, outputting a discharge feature dataset.

[0023] Preferably, the step (2) comprises the following steps:

[0024] (2.1) collecting a voltage sampling point sequence from the end of discharge to the recovery process through the distribution terminal, and using data smoothing to eliminate noise in the voltage sampling point sequence to obtain voltage time series characteristic data;

[0025] (2.2) calculating the voltage rise per unit time based on the voltage time series characteristic data, obtaining the remaining battery charge change value through the state of charge calculation unit, and obtaining voltage recovery characteristic data;

[0026] (2.3) collecting internal resistance sampling values ​​for the voltage recovery characteristic data, and performing three-layer decomposition and reconstruction of the internal resistance sampling values ​​using wavelet transform to obtain battery polarization characteristic data;

[0027] (2.4) Performing piecewise linear fitting based on the battery polarization characteristic data to obtain a voltage recovery curve after deep discharge, and determining dynamic recovery characteristics based on the voltage recovery curve.

[0028] Preferably, the step (3) comprises the following steps:

[0029] (3.1) Obtain the discharge depth value and the cumulative number of discharges in the corresponding time period based on the discharge feature data set to obtain the discharge feature sequence;

[0030] (3.2) performing linear temperature compensation on the internal resistance measurement value according to the ambient temperature data at the corresponding moment of the discharge characteristic sequence to obtain compensated internal resistance data;

[0031] (3.3) establishing a random forest regressor for the compensated internal resistance data, constructing a feature matrix using the discharge depth value and the cumulative number of discharges, and training an internal resistance predictor in combination with the state of charge change to obtain a predicted internal resistance value;

[0032] (3.4) A time series is constructed for the predicted internal resistance value, and the internal resistance variation curve is fitted using the least squares method.

[0033] Preferably, the step (4) comprises the following steps:

[0034] (4.1) Obtaining the voltage recovery time data and internal resistance parameter change trend curve in the dynamic recovery characteristics, and obtaining the initial multi-parameter data based on the discharge depth record and terminal voltage fluctuation measurement value;

[0035] (4.2) calculating the upper and lower quartiles using the interquartile range method based on the initial multi-parameter data, and correcting the abnormal data using the median substitution method to obtain the corrected multi-parameter data;

[0036] (4.3) performing zero-mean normalization processing on the corrected multi-parameter data, and performing multi-scale decomposition on the terminal voltage fluctuation record using discrete wavelet transform to obtain filtered voltage data;

[0037] (4.4) The principal component analysis method is used to extract the characteristic vector based on the filtered voltage data and the standardized parameter data, and the comprehensive characteristic matrix is ​​obtained through orthogonal transformation matrix operation.

[0038] Preferably, the step (5) comprises the following steps:

[0039] (5.1) Weight distribution is performed on the comprehensive feature matrix according to the fusion calculation unit to obtain the initial value of the parameter weight;

[0040] (5.2) Using the sliding window method to quantify the correlation between the internal resistance and the discharge depth based on the initial value of the parameter weight, the feature weight is optimized by the gradient descent method to obtain the feature weight coefficient;

[0041] (5.3) A combined feature function is constructed based on the feature weighting coefficients, a support vector regressor is used to fit the nonlinear relationship between discharge depth and recovery time, a three-layer perceptron network is constructed, and the network parameters are trained through a back propagation algorithm to obtain a health status assessment model.

[0042] Preferably, the step (6) comprises the following steps:

[0043] (6.1) Smoothing the internal resistance data using a sliding average method for the health state feature vector, calculating the internal resistance change rate based on the smoothed data, and obtaining the internal resistance feature weight;

[0044] (6.2) obtaining a time difference using a difference calculation method based on the internal resistance characteristic weight and the measured recovery time, calculating the recovery time deviation using a cumulative deviation function, and obtaining time deviation data;

[0045] (6.3) Using the cycle number and internal resistance change rate to construct a time series feature for the time deviation data, and training a capacity prediction model using a long short-term memory network to obtain a capacity decay curve;

[0046] (6.4) Extracting life characteristics using a rolling time window based on the capacity decay curve, establishing a remaining life prediction model using a grey prediction method, and obtaining life prediction data;

[0047] (6.5) Compensating the predicted lifespan using a temperature coefficient, establishing a mapping relationship between temperature and lifespan using an exponential function, and obtaining a predicted lifespan value for a baseline operating condition;

[0048] (6.6) Taking the standard recovery time as the benchmark, the difference between the voltage recovery time and the standard recovery time is normalized to obtain a standardized deviation metric. Combined with the weight of the internal resistance feature, the weighted fusion of the deviation metric and the internal resistance weight is calculated to generate a health degradation score. The health degradation score is used as the input of the life prediction model, and a preliminary life estimate is output. The remaining life is calculated to generate a life prediction value that represents the health status of the battery.

[0049] Preferably, the step (7) comprises the following steps:

[0050] (7.1) The correlation coefficient is calculated using the Pearson algorithm based on the discharge depth data and the voltage recovery time series, and the characteristic correlation curve is obtained by polynomial fitting;

[0051] (7.2) constructing a weight optimization function for the feature correlation curve, using the gradient descent method to calculate the weight update amount, and obtaining the updated feature weight through the dynamic learning rate;

[0052] (7.3) constructing a parameter mapping relationship using a recursive neural network based on the updated feature weights, and selecting a hidden layer structure through cross-validation to obtain a feature combination function;

[0053] (7.4) The mean square error is used as the loss function for the feature combination function, and the network parameters are trained through error back propagation to obtain the retrained health status assessment model.

[0054] Preferably, the step (8) comprises the following steps:

[0055] (8.1) Use a sliding time window to obtain time series data on discharge depth and discharge times, calculate the capacity loss rate using the capacity decay curve, and obtain the battery health score;

[0056] (8.2) constructing a state transition matrix for the battery health score, calculating the probability distribution of the normal state, the over-discharge state, and the deep discharge state through the state transition matrix, and obtaining a rotation index value;

[0057] (8.3) Feature classification of load power, discharge depth, and discharge duration is performed based on the rotation index value, and classification boundaries are calculated using a support vector machine to obtain rotation priority data;

[0058] (8.4) Genetic encoding is performed on the rotation priority data, and a crossover mutation operation is performed on the rotation interval through the genetic encoding to obtain a battery rotation plan.

[0059] Preferably, the step (6.6) comprises the following steps:

[0060] (6.6.1) Obtaining a measured voltage recovery time based on the measured operating parameters, and subtracting a reference recovery time from the measured voltage recovery time to obtain a time difference sequence;

[0061] (6.6.2) Performing maximum and minimum normalization on the time difference sequence, and obtaining a battery degradation score by linearly combining the deviation metric value and the internal resistance characteristic weight;

[0062] (6.6.3) Using the battery degradation score to construct a long short-term memory network, and training the long short-term memory network using a sliding time window to obtain a preliminary lifespan estimate;

[0063] (6.6.4) Calculate the cumulative deviation between the preliminary life estimate and the actual life, and use an exponential weighting method to compensate for the cumulative deviation to generate a life prediction value that represents the health status of the battery.

[0064] The beneficial effects of the present invention are:

[0065] 1. In step (1) of the present invention, multi-level data collection and analysis are used to generate a discharge feature dataset containing discharge depth, discharge frequency, and energy consumption, fully capturing the dynamic behavior of the battery under frequent power outages. Compared with traditional single-parameter analysis, this method improves the comprehensiveness and accuracy of feature extraction, laying a solid foundation for health status assessment.

[0066] 2. Compared with the traditional method that relies only on the analysis of voltage curves, step (2) of this method significantly improves the accuracy and prediction ability of feature extraction by combining wavelet transform with long short-term memory network. The generation of dynamic recovery features provides a key basis for health status assessment and ensures that the assessment model can adapt to complex working conditions under frequent power outages.

[0067] 3. Compared to traditional methods that rely solely on a single internal resistance measurement, step (3) of this method fully captures the nonlinear characteristics of internal resistance as it changes with discharge depth and cycle number through random forest regression and piecewise fitting. The generation of internal resistance variation trends provides a key basis for evaluating battery health status, effectively supporting operation and maintenance decisions and reducing the risk of equipment failure due to battery degradation.

[0068] 4. This method improves data quality and feature expression capabilities through interquartile range anomaly detection, wavelet transform denoising, and principal component analysis dimensionality reduction in step (4). The comprehensive feature matrix not only retains the key characteristics of the original data, but also optimizes the computational complexity through dimensionality reduction, providing strong support for building a high-precision health status assessment model.

[0069] 5. Step (5) of this method significantly improves the model's fitting accuracy and generalization capability by combining radial basis kernel functions, support vector regression, and neural networks. The dynamic update mechanism further enhances the model's adaptability to new operating conditions and ensures the reliability of the evaluation under frequent power outages. Feature sensitivity analysis provides operators with clear monitoring priorities and reduces the risk of evaluation errors caused by improper parameter selection.

[0070] 6. Based on step (6), the lifespan prediction value generated through multi-level feature extraction and model training integrates multi-dimensional information such as internal resistance change, recovery time deviation, and temperature influence. Compared with traditional methods that rely only on a single parameter, the combination of long-short-term memory network and gray prediction improves the prediction accuracy. Deviation compensation and confidence interval analysis further optimize the reliability of the results, provide precise guidance for the maintenance strategy of distribution terminal batteries, and reduce the equipment risk caused by misjudgment of lifespan.

[0071] 7. Compared with the traditional static weight method, step (7) of this method realizes dynamic adjustment of feature weights and optimization of model structure through adaptive gradient descent and recursive neural network. Exponential smoothing and dynamic calibration mechanism further enhance the stability of the model in real-time applications and reduce the risk of misjudgment.

[0072] 8. Through multi-level discharge data analysis and optimization algorithms, the generated rotation plan effectively balances battery health and power supply reliability. Compared with traditional fixed-cycle rotation, step (8) of this method accurately divides the rotation priority through hidden Markov model and support vector machine, and genetic algorithm further optimizes the timing arrangement. The online correction mechanism enhances the dynamic adaptability of the plan and ensures stable operation in frequent power outage scenarios. The monitoring results show that the health score dispersion of the battery pack is reduced after rotation, and the overall life is effectively extended, providing strong support for the reliable operation and maintenance of the distribution terminal. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is a flow chart of the method steps of the present invention; DETAILED DESCRIPTION

[0074] The embodiments of the present invention are further described below with reference to the accompanying drawings:

[0075] Example 1

[0076] likeFigure 1 As shown, the present invention provides a method for evaluating the health status of batteries in distribution terminals based on multi-parameter fusion, comprising the following steps:

[0077] (1) By collecting the discharge data of the distribution terminal battery under frequent power outage scenarios, the discharge depth and discharge times of deep discharge are obtained to obtain the discharge characteristic data set;

[0078] The step (1) comprises the following steps:

[0079] (1.1) collecting voltage sampling point data during the discharge process, and processing the sampling point data using a median filter method to obtain initial discharge characteristic data;

[0080] The data is collected through the built-in battery management unit of the power distribution terminal, and the discharge curve is drawn after the sampled data is processed for abnormal values. The discharge depth value is calculated using the curve slope change rate;

[0081] (1.2) monitoring the discharge current according to the initial discharge characteristic data, and calculating the remaining capacity percentage by multiplying the current sampling value by the time to obtain the discharge number characteristic data;

[0082] Wherein, when the remaining capacity is lower than the discharge depth threshold, the number of power outages is recorded;

[0083] (1.3) Based on the discharge number characteristic data and the discharge cut-off voltage data, the average load power is calculated using a sliding time window to obtain discharge energy characteristic data;

[0084] Among them, the discharge energy characteristic data is obtained by multiplying the load power and the discharge duration to obtain the total energy consumption of a single discharge, forming the discharge energy characteristic data. The time window method is used to extract the time series characteristics of the discharge process, and the time series correlation matrix is ​​established in combination with the discharge number characteristic data;

[0085] (1.4) classifying the discharge conditions according to the discharge energy characteristic data to obtain a data set of corresponding relationships between discharge depth and discharge times;

[0086] Among them, the density clustering method is used to classify the discharge conditions;

[0087] (1.5) Using the support vector regression method to model the corresponding relationship dataset, outputting a discharge feature dataset.

[0088] In step (1), the distribution terminal has a built-in battery management system to monitor the battery's operating status during a power outage. When a power outage event is triggered, the system automatically starts data acquisition and records key parameters during the discharge process. The depth of discharge and the number of discharges are core indicators of the battery's operating status and are directly related to the battery's performance degradation trend.

[0089] In step (1.1), the battery management unit collects voltage sampling data at a fixed frequency to generate a high-resolution voltage time series. To eliminate noise interference, the sampled data is processed using a median filter, with a sliding window size of 5 to 9 data points. The median within the window is calculated to replace outliers and ensure data smoothness. Based on the processed voltage data, the slope change of the discharge curve is calculated to determine the discharge depth value and generate initial discharge characteristic data. Slope analysis can identify the initial stable stage of discharge and the rapid decline stage in the middle and late stages of discharge.

[0090] In step (1.2), the battery management unit monitors the discharge current in real time, recording the current sampling value and discharge duration. By integrating the current and time, it calculates the amount of discharged energy and, combined with the rated capacity, determines the remaining capacity percentage. When the remaining capacity falls below the preset depth of discharge threshold, a deep discharge event is recorded, generating characteristic discharge count data.

[0091] In step (1.3), the average load power is calculated using the sliding time window method in combination with the discharge cut-off voltage and load power data recorded by the battery management unit. The mean load power within the window is calculated, and the total energy consumption of a single discharge is obtained by multiplying the power by the discharge duration.

[0092] In step (1.4), a time series analysis method is used to extract the voltage change rate and current fluctuation characteristics during the discharge process. Combined with the discharge energy characteristic data, a time series correlation matrix reflecting the discharge pattern is constructed. A density clustering algorithm is used to classify the matrix, setting the cluster radius to 0.8 and the minimum number of samples to 5 to identify the distribution of operating conditions at different discharge depths. This generates a dataset of the corresponding relationship between discharge depth and discharge number.

[0093] In step (1.5), based on the corresponding relationship dataset between discharge depth and discharge number, the support vector regression method is used to establish a prediction model, the Gaussian kernel function is selected, the penalty factor is set to 1.5, the parameters are optimized through cross-validation, and the discharge characteristic dataset is output. The model can predict the cycle life performance under different discharge depths.

[0094] In step (1) of the present invention, multi-level data collection and analysis are used to generate a discharge feature dataset that includes discharge depth, discharge frequency, and energy consumption. This fully captures the dynamic behavior of the battery under frequent power outages. Compared with traditional single-parameter analysis, this method improves the comprehensiveness and accuracy of feature extraction, laying a solid foundation for health status assessment.

[0095] (2) Analyze the discharge characteristic data set to obtain the voltage recovery curve after deep discharge, and determine the dynamic recovery characteristics based on the voltage recovery curve;

[0096] The step (2) comprises the following steps:

[0097] (2.1) collecting a voltage sampling point sequence from the end of discharge to the recovery process through the distribution terminal, and using data smoothing to eliminate noise in the voltage sampling point sequence to obtain voltage time series characteristic data;

[0098] Among them, the voltage time series characteristic data is obtained by drawing a continuous voltage recovery curve based on the processed voltage sampling point sequence;

[0099] (2.2) calculating the voltage rise per unit time based on the voltage time series characteristic data, obtaining the remaining battery charge change value through the state of charge calculation unit, and obtaining voltage recovery characteristic data;

[0100] Among them, before obtaining the battery remaining power change value through the charge state calculation unit, it is necessary to calculate the voltage change rate during the recovery process in combination with the voltage sampling timestamp;

[0101] (2.3) collecting internal resistance sampling values ​​for the voltage recovery characteristic data, and performing three-layer decomposition and reconstruction of the internal resistance sampling values ​​using wavelet transform to obtain battery polarization characteristic data;

[0102] (2.4) Performing piecewise linear fitting based on the battery polarization characteristic data to obtain a voltage recovery curve after deep discharge, and determining dynamic recovery characteristics based on the voltage recovery curve.

[0103] Specifically, after piecewise linear fitting, the polarization recovery time constant is calculated in combination with the voltage sampling time window, and a recursive neural network is used to establish a voltage recovery prediction model to obtain dynamic characteristic data of the recovery process; a Gaussian mixture distribution is established based on the dynamic characteristic data of the recovery process, and the expectation maximization method is used to estimate the probability density of the recovery curve to obtain the dynamic recovery characteristics of the voltage recovery curve; a time series correlation matrix is ​​established for the dynamic recovery characteristics, and an adaptive threshold is used to segment the recovery process to obtain voltage recovery characteristic parameters at different stages.

[0104] In step (2), the discharge feature dataset contains multi-dimensional discharge information of distribution terminal batteries under frequent power outage scenarios, providing basic data for further analysis of battery health status. This step focuses on the dynamic characteristics of the voltage recovery process after deep discharge by deeply mining the dataset, generating a voltage recovery curve and extracting dynamic recovery features that reflect battery polarization and recovery capabilities. These features can effectively characterize the performance degradation trend of batteries under complex operating conditions.

[0105] In step (2.1), the battery management unit collects voltage data at a high frequency to capture subtle changes during the voltage recovery process. To eliminate environmental interference and sensor noise, a sliding average filter is used with a window size of seven sampling points. The mean voltage within the window is calculated to smooth the data curve. The filtered voltage time series characteristic data clearly reflects the rapid rise in the early stage of recovery and the slow stabilization trend in the later stage, generating a continuous voltage change sequence that provides reliable input for analysis.

[0106] In step (2.2), based on the smoothed voltage time series characteristic data, the ratio of the voltage difference between adjacent sampling points to the time interval is calculated to obtain the voltage rise rate. Simultaneously, the state of charge monitoring module records the changes in the remaining battery charge in real time. Combining the voltage rise rate with the charge increment generates voltage recovery characteristic data, which comprehensively characterizes the dynamic behavior of the recovery process.

[0107] In step (2.3), internal resistance, a key parameter reflecting the degree of battery polarization, is acquired using the AC impedance method. A 1000 Hz small-signal excitation is applied, and the voltage and current responses are recorded to calculate the internal resistance value. To extract the effective information from the internal resistance signal, a wavelet transform is employed, using the db4 wavelet basis function for a three-layer decomposition, retaining high-energy coefficients for signal reconstruction. The reconstructed internal resistance signal generates polarization characteristic data, which clearly reflects the changes in the battery's internal state during the voltage recovery phase, complementing dynamic characteristic analysis.

[0108] In step (2.4), the polarization characteristic data and voltage recovery characteristic data are integrated, and a piecewise linear fitting method is used to divide the recovery process into two stages: fast recovery and slow recovery. The polarization recovery time constant of each stage is calculated. For example, the time constant of the fast recovery stage is approximately 20 seconds, and that of the slow recovery stage is approximately 60 seconds. To further improve prediction accuracy, a long short-term memory network is used to construct a voltage recovery prediction model with a time window of 300 seconds. The model is trained to capture the temporal dependence of voltage recovery. The model outputs dynamic characteristic data of the recovery process, which can accurately predict the voltage recovery trend at different discharge depths.

[0109] Furthermore, to quantify the diversity of the voltage recovery process, a Gaussian mixture model was constructed based on the dynamic feature data. The recovery curve is assumed to consist of two Gaussian components, corresponding to the fast and slow recovery phases. The expectation-maximization algorithm was used to iteratively optimize the model parameters and estimate the mean and variance of each component. For example, the mean of the fast recovery phase is approximately 15 seconds, while the mean of the slow recovery phase is 60 seconds. The resulting dynamic recovery features can characterize the probability distribution of the voltage recovery curve and reflect the performance differences of the battery in different recovery phases.

[0110] In step (2) of the present invention, dynamic recovery features are generated through multi-level analysis of the voltage recovery process, including voltage rise rate, power increment, internal resistance change, and polarization time constant. These features comprehensively reflect the recovery ability and polarization characteristics of the battery after deep discharge. Compared with traditional methods that rely solely on voltage curve analysis, this method significantly improves the accuracy and predictive ability of feature extraction by combining wavelet transform with long-short-term memory network. The generation of dynamic recovery features provides a key basis for health status assessment and ensures that the assessment model can adapt to complex working conditions under frequent power outages.

[0111] (3) If the voltage recovery time of the dynamic recovery feature exceeds the preset threshold, the corresponding discharge depth and discharge times are extracted from the discharge feature data set to obtain the change trend of the internal resistance parameter;

[0112] The step (3) comprises the following steps:

[0113] (3.1) Obtain the discharge depth value and the cumulative number of discharges in the corresponding time period based on the discharge feature data set to obtain the discharge feature sequence;

[0114] Among them, the voltage recovery time is calculated based on the dynamic recovery characteristic data, and the preset recovery time threshold is compared to determine whether the recovery time exceeds the limit. The discharge depth value and the cumulative number of discharges in the corresponding time period are extracted by querying the discharge characteristic data set to obtain the discharge characteristic sequence.

[0115] (3.2) performing linear temperature compensation on the internal resistance measurement value according to the ambient temperature data at the corresponding moment of the discharge characteristic sequence to obtain compensated internal resistance data;

[0116] Wherein, a sampling window is set according to the discharge characteristic sequence, and an AC impedance sweep method is used to measure the internal resistance value;

[0117] (3.3) establishing a random forest regressor for the compensated internal resistance data, constructing a feature matrix using the discharge depth value and the cumulative number of discharges, and training an internal resistance predictor in combination with the state of charge change to obtain a predicted internal resistance value;

[0118] (3.4) A time series is constructed for the predicted internal resistance value, and the internal resistance variation curve is fitted using the least squares method.

[0119] Specifically, after fitting the internal resistance change curve using the least squares method, the internal resistance growth rate is calculated, and the internal resistance change curve is divided into stages using piecewise linear fitting to obtain the internal resistance parameter change trend. Based on the internal resistance parameter change trend, a time series correlation matrix is ​​constructed, and the sliding average method is used to extract trend features to generate segmented internal resistance change features.

[0120] In step (3), the voltage recovery time, as the core indicator of the dynamic recovery characteristic, directly reflects the battery's self-recovery ability after deep discharge. When the recovery time exceeds the preset threshold, it indicates that the battery performance may have significantly degraded, and further analysis of the change pattern of the discharge characteristics and internal resistance parameters is required. This step extracts key information from the discharge characteristic data set, combines internal resistance measurement and prediction, and generates the change trend of the internal resistance parameter, providing accurate data support for health status assessment.

[0121] In step (3.1), the voltage recovery time is calculated by analyzing dynamic recovery characteristic data, such as the time from discharge termination to voltage stabilization. A preset threshold is determined based on the battery type and application scenario; if the recovery time exceeds this threshold, it indicates that the battery's recovery capability has weakened. At this point, the records for the corresponding time period are queried from the discharge characteristic dataset to extract the discharge depth value and cumulative discharge count. A discharge characteristic sequence containing these parameters is generated, providing targeted data for subsequent internal resistance analysis.

[0122] In step (3.2), based on the discharge characteristic sequence, the sampling window was set to 60 seconds, and the internal resistance was measured using the AC impedance method. The frequency sweep range covered 0.1 Hz to 1000 Hz, and 20 frequency points were selected to ensure measurement accuracy. A temperature compensation coefficient of 0.3% per degree Celsius was used for correction to generate compensated internal resistance data. The compensation process ensured the comparability of internal resistance values ​​at different temperatures and improved the reliability of data analysis.

[0123] In step (3.3), the compensated internal resistance data is used to construct a random forest regression model to predict the internal resistance change. The feature matrix contains discharge depth, voltage recovery time, discharge times and charge state change, where the discharge depth is divided into 6 intervals from 20% to 80%, the recovery time is divided into 8 intervals from 100 seconds to 500 seconds, and the discharge times are divided into 10 intervals from 0 to 1000 times. The charge state change is obtained by monitoring the incremental amount of electricity during the recovery process. When training the model, the number of decision trees is set to 100, and the five-fold cross-validation optimization parameters are used to generate the predicted internal resistance value. The predicted value can reflect the dynamic change trend of the internal resistance under different working conditions.

[0124] In step (3.4), the predicted internal resistance values ​​are organized into a time series in chronological order. The least squares method is used to fit the internal resistance curve to capture its long-term trend. To eliminate measurement noise, a Gaussian kernel function is used for smoothing, with a kernel width of 30 data points. The smoothed internal resistance curve clearly shows the growth pattern of internal resistance with increasing cycle number, providing high-quality data for trend analysis.

[0125] And, by smoothing the internal resistance curve, the internal resistance growth rate is calculated. By using piecewise linear fitting method, the internal resistance change is divided into stable period, growth period and rapid growth period, corresponding to 0 to 200 times, 200 to 400 times and 400 times or more cycles. The internal resistance growth characteristics of each stage reflect the different degradation stages of battery performance, and the internal resistance parameter change trend is generated.

[0126] Moreover, based on the internal resistance parameter change trend, a time series correlation matrix is constructed to analyze the correlation of internal resistance values at different stages. By using the moving average method, the window size is set to 180 data points, and the internal resistance change characteristics of each stage are extracted. For example, the stable period internal resistance fluctuation is less than 1%, the growth period annual growth rate is about 10%, and the rapid growth period annual growth rate is more than 20%. The generated feature data can quantify the dynamic law of internal resistance change and reflect the continuous impact of deep discharge on battery life.

[0127] In step (3) of the present application, by systematic analysis of internal resistance parameters, combined with discharge characteristic sequence and environmental temperature compensation, a high-precision internal resistance change trend is generated. Compared with the traditional method which only relies on single internal resistance measurement, this method comprehensively captures the nonlinear characteristics of internal resistance change with discharge depth and cycle number through random forest regression and piecewise fitting. The generation of internal resistance change trend provides a key basis for evaluating the state of health of storage battery, which can effectively support operation and maintenance decision, and reduce the risk of equipment failure caused by battery degradation.

[0128] (4) Fusing dynamic recovery characteristics and internal resistance parameter change trend, constructing multi-parameter data set, the multi-parameter data set includes discharge depth, voltage recovery time, internal resistance value and terminal voltage fluctuation, and the multi-parameter data set is subjected to standardization processing to obtain a comprehensive feature matrix;

[0129] The step (4) comprises the following steps:

[0130] (4.1) obtaining voltage recovery time data in dynamic recovery characteristics and internal resistance parameter change trend curve, and obtaining initial multi-parameter data according to discharge depth record and terminal voltage fluctuation measurement value;

[0131] Wherein, the initial multi-parameter data is obtained by establishing an original parameter data table through discharge depth record and terminal voltage fluctuation measurement value;

[0132] (4.2) calculating upper and lower quartiles according to the initial multi-parameter data by using the interquartile range method, and correcting abnormal data by median replacement method to obtain corrected multi-parameter data;

[0133] Wherein, the abnormal value determination threshold is set according to 1.5 times of the interquartile range;

[0134] (4.3) performing zero-mean normalization processing on the corrected multi-parameter data, and performing multi-scale decomposition on the terminal voltage fluctuation record using discrete wavelet transform to obtain filtered voltage data;

[0135] The data is normalized by calculating the mean and standard deviation of each parameter, and the maximum and minimum method is used to map the discharge depth to the interval [0, 1] to obtain standardized parameter data. The terminal voltage fluctuation is recorded using the data in the standardized parameter data.

[0136] (4.4) The principal component analysis method is used to extract the characteristic vector based on the filtered voltage data and the standardized parameter data, and the comprehensive characteristic matrix is ​​obtained through orthogonal transformation matrix operation.

[0137] The number of principal components is determined by calculating eigenvalues ​​to obtain reduced-dimensional feature data, an orthogonal transformation matrix is ​​constructed for the reduced-dimensional feature data, and the coordinates of the feature data are transformed by matrix multiplication to obtain a comprehensive feature matrix.

[0138] In step (4), the dynamic recovery characteristics and the change trend of the internal resistance parameters respectively characterize the recovery ability of the battery after discharge and the internal performance degradation law. The integration of these two types of data can fully reflect the health status of the battery under complex working conditions.

[0139] In step (4.1), the data acquisition unit acquires key parameters from the battery management system at the power distribution terminal. Voltage recovery time data is extracted from the dynamic recovery characteristics, reflecting the duration from discharge termination to voltage stabilization. The internal resistance parameter trend curve provides information on the dynamic changes in the battery's internal impedance. The depth of discharge record indicates the percentage of battery charge consumed. Terminal voltage fluctuation measurements capture voltage instability during the recovery process. These parameters are aligned through timestamps to generate an initial multi-parameter dataset, laying the foundation for subsequent processing.

[0140] In step (4.2), the initial multi-parameter data set may contain outliers due to sensor errors or environmental interference. To ensure data quality, the interquartile range method is used to detect outliers. Taking the voltage recovery time as an example, the upper quartile of the data set is calculated to be 450 seconds, the lower quartile is 150 seconds, and the interquartile range is 300 seconds. The outlier threshold is set to 1.5 times the interquartile range, that is, 450 seconds. Data outside this range is considered abnormal. Similarly, the upper quartile of the internal resistance value is 18 milliohms, the lower quartile is 12 milliohms, and the abnormal threshold is 21 milliohms. Abnormal data are corrected by the median replacement method. For example, the outlier value of the voltage recovery time is replaced with the median value of the data set, 300 seconds, and the outlier value of the internal resistance is replaced with 15 milliohms. The corrected multi-parameter data set eliminates abnormal interference and improves the reliability of the data.

[0141] In step (4.3), the corrected multi-parameter dataset is standardized to eliminate differences in the dimensions and scales of the different parameters. Zero-mean normalization converts the data into a standard normal distribution by calculating the mean and standard deviation of each parameter. Furthermore, a maximum-minimum normalization method is used to map the depth of discharge to the range of 0 to 1 to ensure a uniform data distribution. Other parameters, such as voltage recovery time and internal resistance, are also standardized to generate a standardized parameter dataset. This enhances the comparability between multiple parameters and provides a unified scale for feature extraction.

[0142] Furthermore, terminal voltage fluctuation data contains high-frequency noise, which can affect feature extraction accuracy. Using the discrete wavelet transform (DWT) method, a three-layer decomposition using the db4 wavelet basis function is performed to decompose the signal into low-frequency and high-frequency components. The low-frequency coefficients are retained for signal reconstruction, while high-frequency noise is filtered out to generate a smooth voltage fluctuation curve. The reconstructed curve retains the main trends of the voltage fluctuation, such as the periodic changes in voltage fluctuation during the recovery process, while eliminating the influence of transient interference. The filtered voltage fluctuation data, combined with other standardized parameters, provides high-quality input for feature fusion.

[0143] In step (4.4), based on the filtered voltage fluctuation data and the standardized parameter data set, the principal component analysis method is used to perform feature dimensionality reduction and generate a comprehensive feature matrix. Principal component analysis extracts the main variation information in the data by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. For example, by analyzing 1000 sets of standardized data, four eigenvalues ​​were obtained, namely 2.8, 0.8, 0.3 and 0.1, with a cumulative contribution rate of 95%. The first two principal components are selected to construct a two-dimensional feature space, in which the first principal component reflects the comprehensive change of voltage recovery time and internal resistance value, and the second principal component characterizes the correlation between discharge depth and terminal voltage fluctuation. Through orthogonal transformation, the standardized data is projected into the principal component space to generate a comprehensive feature matrix. Each data point in the matrix contains two principal component values, which retains the core information of the original data while reducing the dimension. The converted data shows a clear distribution pattern in the feature space. For example, the low internal resistance data points are concentrated in the first quadrant, and the high internal resistance data points are distributed in the third quadrant, reflecting the state differences of the battery under different working conditions. The generation of a comprehensive feature matrix significantly improves data processing efficiency and provides a reliable basis for subsequent multi-parameter fusion analysis.

[0144] In step (4) of the present invention, through multi-level data cleaning and feature extraction, the generated comprehensive feature matrix can effectively integrate the multi-dimensional information of discharge depth, voltage recovery time, internal resistance value and terminal voltage fluctuation. Compared with the traditional method that relies only on the processing of a single parameter, this method improves the data quality and feature expression ability through interquartile range anomaly detection, wavelet transform denoising and principal component analysis dimensionality reduction. The comprehensive feature matrix not only retains the key characteristics of the original data, but also optimizes the computational complexity through dimensionality reduction, providing strong support for the construction of a high-precision health status assessment model.

[0145] (5) The comprehensive feature matrix is ​​weighted by a multi-parameter fusion algorithm to identify the nonlinear relationship between discharge depth and voltage recovery time, and the feature weights and correlation coefficients obtained by the internal resistance are quantified. The health status feature vector is obtained based on the feature weights and correlation coefficients, and a health status assessment model is constructed.

[0146] The step (5) comprises the following steps:

[0147] (5.1) Weight distribution is performed on the comprehensive feature matrix according to the fusion calculation unit to obtain the initial value of the parameter weight;

[0148] Among them, a radial basis kernel function is established for the discharge depth and voltage recovery time data, and the kernel width parameter is selected to perform nonlinear mapping on the characteristic data. The characteristic correlation matrix is ​​calculated based on the mapped data to obtain the initial value of the parameter weight;

[0149] (5.2) Using the sliding window method to quantify the correlation between the internal resistance and the discharge depth based on the initial value of the parameter weight, the feature weight is optimized by the gradient descent method to obtain the feature weight coefficient;

[0150] Wherein, the Pearson correlation coefficient is calculated according to the initial value of the parameter weight;

[0151] (5.3) A combined feature function is constructed based on the feature weighting coefficients, a support vector regressor is used to fit the nonlinear relationship between discharge depth and recovery time, a three-layer perceptron network is constructed, and the network parameters are trained through a back propagation algorithm to obtain a health status assessment model.

[0152] The radial basis kernel function was selected as the kernel function, and the penalty factor and kernel parameters were determined through a grid search method to construct a three-layer perceptron network. The number of input layer nodes was set to be the same as the feature dimension, the number of hidden layer nodes was twice that of the input layer, and the number of output layer nodes was equal to the number of health status levels. The network parameters were trained using a backpropagation algorithm to obtain a state mapping relationship. A dynamic update mechanism was established for this state mapping relationship, and a sliding time window was used to extract features from newly added data. The model parameters were updated online using a mini-batch stochastic gradient descent method to obtain a health status assessment model. Feature sensitivity was calculated based on the health status assessment model, and the model's response to each input feature was quantified using a perturbation analysis method to obtain a feature importance ranking.

[0153] In step (5), the comprehensive feature matrix integrates multidimensional information such as discharge depth, voltage recovery time, internal resistance, and terminal voltage fluctuation, providing a basis for capturing the complex laws of battery health status. This step reveals the deep correlation between parameters through feature weighting and nonlinear fitting, and constructs a high-precision health status assessment model, aiming to improve the adaptability and reliability of the assessment results.

[0154] In step (5.1), the fusion calculation unit first performs a preliminary analysis of the parameters in the comprehensive feature matrix and uses a Gaussian radial basis kernel function to achieve nonlinear mapping of the feature data. The kernel function width parameter is set to 0.8 to balance the local and global nature of the mapping. Initial weights are generated by calculating the correlation matrix between the parameters. These initial values ​​reflect the potential contribution of each parameter to the health status and provide a starting point for subsequent optimization.

[0155] In step (5.2), the Pearson correlation coefficient is used to evaluate the linear correlation between the internal resistance value and other parameters. For example, the correlation coefficient between the internal resistance value and the depth of discharge is calculated to be 0.72, and the correlation coefficient with the voltage recovery time is 0.68, indicating that the internal resistance has a significant impact on the battery state. In order to capture dynamic changes, the sliding window method is used, the window size is set to 300 seconds, and the correlation trend between parameters is analyzed section by section. Based on the correlation analysis results, the weights are optimized using the gradient descent method, the learning rate is set to 0.01, and the iteration is performed 100 times to adjust the weights to minimize the prediction error. The optimized feature weight coefficients are, for example, 0.42 for depth of discharge, 0.34 for voltage recovery time, 0.24 for internal resistance, and 0.2 for terminal voltage fluctuation, reflecting the relative importance of each parameter.

[0156] In step (5.3), the combined feature function integrates multi-parameter information using weighted coefficients to generate a comprehensive feature representation. A support vector regression model was used to fit the nonlinear relationship between discharge depth and voltage recovery time. A Gaussian radial basis kernel function was selected, and a grid search was used to determine a penalty factor of 1.2 and a kernel parameter of 0.5. During training, the model input 1000 sets of feature data and output a predicted value for voltage recovery time. The resulting nonlinear mapping model accurately describes the complex relationships between parameters, providing a reliable basis for subsequent state assessment.

[0157] Furthermore, based on a nonlinear mapping model, a three-layer perceptron neural network was constructed. The network parameters were trained using a backpropagation algorithm to generate a state mapping relationship. The network structure was designed with four nodes in the input layer, corresponding to the four parameters of the comprehensive feature matrix; eight nodes in the hidden layer, used to capture nonlinear interactions between features; and three nodes in the output layer, representing the three states of health, subhealth, and decline. The training data consisted of 1,000 sets of historical records, with a learning rate set to 0.01 and a batch size of 32. The backpropagation algorithm optimized the network parameters by minimizing cross-entropy loss. After 100 training cycles, the model achieved a classification accuracy of 95% on the validation set. The state mapping relationship maps input features to clear health status levels, providing an efficient tool for real-time monitoring.

[0158] At the same time, a dynamic update mechanism was designed to adapt to the dynamic operating conditions of distribution terminals. A 600-second sliding window was used to extract features of newly added data, such as depth of discharge and internal resistance, every 300 seconds. A small-batch stochastic gradient descent method was used to update model parameters online, processing 16 groups of samples per batch with a learning rate set to 0.005. This update mechanism ensures that the model can quickly adapt to new data. For example, in scenarios with frequent power outages, the model can adjust state predictions based on the latest discharge records, improving robustness to complex operating conditions. The generated health status assessment model can output battery status assessment results in real time.

[0159] Furthermore, feature sensitivity analysis is used to assess the model's dependence on various parameters. By applying a 5% perturbation to the input features, the changes in the model output are observed. For example, perturbing the depth of discharge resulted in an 8% change in the model output, while perturbing the internal resistance resulted in a 6% change. The voltage recovery time and terminal voltage fluctuation were 4% and 3%, respectively. The analysis results indicate that depth of discharge and internal resistance are key factors influencing health status assessment. The generated feature importance ranking is: depth of discharge, internal resistance, voltage recovery time, and terminal voltage fluctuation. This guides the optimization of the monitoring strategy, ensuring a focus on highly sensitive parameters.

[0160] In step (5) of the present invention, the health status assessment model generated through systematic analysis of the multi-parameter fusion algorithm can effectively capture the nonlinear relationship between the discharge depth and the voltage recovery time, and quantify the contribution weight of each parameter. Compared with the traditional single parameter evaluation, this method significantly improves the fitting accuracy and generalization ability of the model through the combination of radial basis kernel function, support vector regression and neural network. The dynamic update mechanism further enhances the adaptability of the model to new working conditions and ensures the reliability of the assessment in frequent power outage scenarios. Feature sensitivity analysis provides operation and maintenance personnel with a clear monitoring focus and reduces the risk of assessment errors caused by improper parameter selection.

[0161] (6) Extract the weight of the internal resistance feature from the health state feature vector, calculate the degree of deviation between the voltage recovery time and the standard recovery time, and combine the degree of deviation and the internal resistance weight to predict the remaining battery life and obtain the life prediction value;

[0162] The step (6) comprises the following steps:

[0163] (6.1) Smoothing the internal resistance data using a sliding average method for the health state feature vector, calculating the internal resistance change rate based on the smoothed data, and obtaining the internal resistance feature weight;

[0164] Wherein, the internal resistance feature data is extracted from the health state feature vector by a feature extractor;

[0165] (6.2) obtaining a time difference using a difference calculation method based on the internal resistance characteristic weight and the measured recovery time, calculating the recovery time deviation using a cumulative deviation function, and obtaining time deviation data;

[0166] The time difference is the time difference between the measured recovery time and the baseline recovery time;

[0167] (6.3) Using the cycle number and internal resistance change rate to construct a time series feature for the time deviation data, and training a capacity prediction model using a long short-term memory network to obtain a capacity decay curve;

[0168] wherein a capacity decay predictor is established for the time deviation data;

[0169] (6.4) Extracting life characteristics using a rolling time window based on the capacity decay curve, establishing a remaining life prediction model using a grey prediction method, and obtaining life prediction data;

[0170] (6.5) Compensating the predicted lifespan using a temperature coefficient, establishing a mapping relationship between temperature and lifespan using an exponential function, and obtaining a predicted lifespan value for a baseline operating condition;

[0171] (6.6) Taking the standard recovery time as the benchmark, the difference between the voltage recovery time and the standard recovery time is normalized to obtain a standardized deviation metric. Combined with the weight of the internal resistance feature, the weighted fusion of the deviation metric and the internal resistance weight is calculated to generate a health degradation score. The health degradation score is used as the input of the life prediction model, and a preliminary life estimate is output. The remaining life is calculated to generate a life prediction value that represents the health status of the battery.

[0172] The step (6.6) comprises the following steps:

[0173] (6.6.1) Obtaining a measured voltage recovery time based on the measured operating parameters, and subtracting a reference recovery time from the measured voltage recovery time to obtain a time difference sequence;

[0174] (6.6.2) Performing maximum and minimum normalization on the time difference sequence, and obtaining a battery degradation score by linearly combining the deviation metric value and the internal resistance characteristic weight;

[0175] The difference sequence is normalized and mapped using a maximum-minimum normalization method to obtain a deviation metric value. The deviation metric value and the internal resistance characteristic weight are linearly combined according to a preset proportional coefficient, and the combined result is converted to the interval [0, 1] using a numerical mapping function to obtain a battery degradation score.

[0176] (6.6.3) Using the battery degradation score to construct a long short-term memory network, and training the long short-term memory network using a sliding time window to obtain a preliminary lifespan estimate;

[0177] The input layer dimension is set to be the same as the feature dimension, the number of hidden layer nodes is twice that of the input layer, and a sliding time window is used for sequence prediction to obtain a preliminary lifespan estimate.

[0178] (6.6.4) Calculate the cumulative deviation between the preliminary life estimate and the actual life, and use an exponential weighting method to compensate for the cumulative deviation to generate a life prediction value that represents the health status of the battery.

[0179] Among them, random fluctuations in the prediction results are eliminated by mean filtering to obtain a corrected life value, the prediction error standard deviation is calculated based on the corrected life value, three times the standard deviation is used to determine the fluctuation range of the prediction results, and the weighted average method is used to synthesize the prediction results to obtain the life prediction value.

[0180] In step (6), the health state feature vector integrates multi-dimensional parameter information. The internal resistance feature weight and voltage recovery time deviation are key indicators for evaluating battery degradation. This step generates a high-precision life prediction value through in-depth analysis of the feature vector, combined with time series prediction and operating condition correction, providing a scientific basis for battery operation and maintenance management at distribution terminals.

[0181] In step (6.1), the internal resistance characteristic data reflects the degree of aging of the battery's internal performance. After extracting the internal resistance value sequence from the health state feature vector, the sliding average method is used to smooth the data with a window size of 180 seconds to eliminate measurement fluctuations. The internal resistance change rate is calculated by calculating the ratio of the difference between the internal resistances of adjacent windows to the number of cycles. Based on the change rate, the internal resistance characteristic weight is generated to reflect the relative importance of internal resistance to life prediction. Smoothing and change rate calculation ensure the stability and representativeness of the internal resistance characteristic.

[0182] In step (6.2), the degree of deviation of the voltage recovery time characterizes the decline of the battery's recovery ability. The benchmark recovery time is determined according to standard operating conditions, for example, 180 seconds at a 50% depth of discharge. If the measured recovery time is 240 seconds, the difference is 60 seconds. Through the cumulative deviation function, the difference sequence of multiple cycles is weighted and accumulated, and the weight decreases with the number of cycles, for example, the weight of the most recent cycle is 0.5, and the weight of the previous cycle is 0.3. The calculation results show that the deviation can reach 0.45 after 1000 cycles, reflecting a significant decrease in recovery ability. The generated time deviation data provides dynamic degradation information for prediction.

[0183] In step (6.3), the time deviation data and the internal resistance change rate together constitute a time series feature sequence, which reflects the degradation law of battery performance with cycles. The long short-term memory network is used to capture the long-term dependence of the sequence. The network sets 10 time steps, and each step corresponds to the feature data of 100 cycles. The input features include the number of cycles, the internal resistance change rate and the deviation, and the output is the capacity prediction value. For example, the initial capacity is 100 ampere hours, which drops to 90 ampere hours after 500 cycles and 85 ampere hours after 1000 cycles. The training process uses 1000 sets of data with a learning rate of 0.01 to generate a capacity decay curve, which clearly shows the nonlinear decrease trend of capacity with the number of cycles.

[0184] In step (6.4), the capacity decay curve provides the basic data for life prediction. Using a rolling time window of 300 cycles, the capacity decay rate and internal resistance variation characteristics within each window are extracted. The gray prediction method uses a small amount of data to generate a remaining life model. The initial life prediction data can reflect the expected life of the battery under the current operating conditions.

[0185] At the same time, the initial lifespan prediction data was corrected for operating conditions, introducing a temperature compensation mechanism. Using an exponential function to map the relationship between temperature and lifespan, a lifespan prediction value under baseline operating conditions was generated. Temperature significantly affects battery lifespan. For example, for every 10°C increase in temperature, the lifespan loss rate increases by 15%. Under 35°C operating conditions, the initial predicted lifespan of 1500 cycles required correction. Using an exponential function, setting a baseline temperature of 25°C, and calculating the temperature coefficient, the lifespan was adjusted to 1275 cycles. The revised prediction value more closely matches actual operating conditions, enhancing the model's applicability.

[0186] In step (6.5), the validation function evaluates the model's accuracy by comparing the predicted values ​​with historical lifespan data. This historical data contains the actual lifespans of multiple battery groups under different operating conditions. For example, the measured lifespan of a battery at 35°C was 1300 cycles, close to the predicted value of 1275. Error analysis shows that the prediction error range is plus or minus 8%. Using the triple standard deviation method, a confidence interval is generated between 1173 and 1377 cycles. This confidence interval provides a reliable reference for operational and maintenance decisions, ensuring the engineering applicability of the prediction results.

[0187] In step (6.6), a battery degradation score is generated based on the time difference sequence and the internal resistance feature weights. A long-short-term memory (LSTM) network is used for lifespan prediction, and the results are optimized through bias compensation. The difference sequence between the measured voltage recovery time and the baseline value of 180 seconds is normalized using the maximum and minimum values ​​and mapped to a range of 0 to 1, generating a deviation metric value, for example, ranging from 0.2 to 0.4. The degradation score is calculated using a linear combination, with a weight of 0.6 for the deviation metric and 0.4 for the internal resistance. For example, a 20% increase in internal resistance results in a degradation score of approximately 0.5. The LSTM network inputs 4-dimensional features, has 8 hidden nodes, and uses a 360-second sliding window to predict lifespan, outputting a preliminary estimate of approximately 1200 cycles. Bias compensation uses an exponential weighting method with a weight coefficient of 0.8, resulting in a revised lifespan of 1150 cycles with a confidence interval of 1000 to 1300 cycles. The final lifespan prediction is a weighted average of the multi-window results, with a weight of 0.5 for the recent prediction, enhancing the dynamic adaptability of the prediction.

[0188] In step (6) of the present invention, multi-level feature extraction and model training are used to generate a lifespan prediction value that integrates multi-dimensional information such as internal resistance change, recovery time deviation, and temperature effects. Compared to traditional methods that rely solely on a single parameter, the combination of long-short-term memory networks and gray prediction improves prediction accuracy. Deviation compensation and confidence interval analysis further optimize the reliability of the results, providing precise guidance for maintenance strategies for distribution terminal batteries and reducing equipment risks caused by misjudgment of lifespan.

[0189] (7) Analyze the correlation between discharge depth and voltage recovery time, adjust the feature weights in the multi-parameter fusion algorithm according to the life prediction value, optimize the correlation analysis, and retrain the health status assessment model through the adjusted algorithm;

[0190] The step (7) comprises the following steps:

[0191] (7.1) The correlation coefficient is calculated using the Pearson algorithm based on the discharge depth data and the voltage recovery time series, and the characteristic correlation curve is obtained by polynomial fitting;

[0192] Among them, the characteristic correlation coefficient is obtained through the characteristic correlation curve;

[0193] (7.2) constructing a weight optimization function for the feature correlation curve, using the gradient descent method to calculate the weight update amount, and obtaining the updated feature weight through the dynamic learning rate;

[0194] (7.3) constructing a parameter mapping relationship using a recursive neural network based on the updated feature weights, and selecting a hidden layer structure through cross-validation to obtain a feature combination function;

[0195] Among them, a multi-parameter fusion algorithm is reconstructed according to the updated feature weights, a three-layer recursive neural network is used to construct a parameter mapping relationship, the number of input layer nodes is set to be the same as the feature dimension, and the hidden layer structure is selected through five-fold cross validation to obtain a feature combination function;

[0196] (7.4) The mean square error is used as the loss function for the feature combination function, and the network parameters are trained through error back propagation to obtain the retrained health status assessment model.

[0197] A state evaluator is constructed based on the feature combination function, using mean squared error as the loss function. Network parameters are trained through error backpropagation to obtain an optimized evaluator. A prediction function is constructed based on the optimized evaluator, and the prediction results are filtered using exponential smoothing. Online calibration is performed using a dynamic threshold method to obtain an optimized state assessment model. A validation mechanism is established for the optimized state assessment model, using a holdout method to partition the training and test sets. The model is evaluated using performance metrics to obtain a retrained health state assessment model.

[0198] In step (7), the correlation between discharge depth and voltage recovery time directly affects the accuracy of battery health assessment, while the life prediction value provides feedback information for model optimization. This step aims to improve the model's adaptability to complex operating conditions and its prediction reliability by quantifying the correlation between parameters, dynamically adjusting feature weights, and retraining the health assessment model.

[0199] In step (7.1), the deviation between the predicted life and the actual life reflects the prediction error of the model. For example, the predicted life is 1500 cycles, the actual life is 1350 times, the deviation is 150 times, and the calculated adjustment coefficient is 0.15. The Pearson correlation coefficient is used to quantify the correlation between the depth of discharge and the voltage recovery time. The result shows that the correlation coefficient is 0.82, indicating that there is a strong positive correlation between the two. To further capture the nonlinear relationship, a third-order polynomial is used to fit the correlation data. The fitting accuracy reaches 0.95, generating a smooth characteristic correlation curve. The curve shows that when the depth of discharge increases from 40% to 60%, the voltage recovery time shows an accelerating growth trend, providing a data basis for weight optimization.

[0200] In step (7.2), the characteristic correlation curve provides basic information for weight optimization. A weight optimization function is constructed with the goal of minimizing the prediction error. The input includes the correlation coefficient and the adjustment coefficient. The weights are optimized using the adaptive gradient descent method, and the initial learning rate is set to 0.05. If the absolute value of the gradient exceeds 0.1, the learning rate is reduced to 0.8 times; if it is less than 0.01, the learning rate is increased to 1.2 times. After 50 iterations, for example, the discharge depth weight is adjusted from 0.4 to 0.45, the voltage recovery time weight is adjusted from 0.35 to 0.3, and the internal resistance weight remains at 0.25. The optimized weights more accurately reflect the impact of each parameter on the health status and improve the expressive power of the model.

[0201] In step (7.3), the optimized feature weights are used to update the multi-parameter fusion algorithm. A three-layer recursive neural network is used. The four nodes in the input layer correspond to the discharge depth, voltage recovery time, internal resistance, and temperature parameters. The number of hidden layer nodes is determined to be 8 through five-fold cross-validation. The output layer has one node representing the health status score. The training data contains 1000 sets of records. The hidden layer structure with the lowest error in the validation set is 8 nodes, with a mean square error of 0.04. The generated feature combination function integrates parameter information by weighting features, which can effectively map multidimensional inputs to health status outputs, providing an efficient model for subsequent evaluation.

[0202] Furthermore, a health status evaluator was constructed based on the feature combination function. Using mean squared error as the loss function, network parameters were optimized through error backpropagation to generate an optimized evaluation model. The evaluator uses the feature combination function as input and is trained using stochastic gradient descent with a batch size of 32. If the validation set error does not decrease for five consecutive cycles, the learning rate is reduced to 0.5. After 1000 cycles of training, the training set error dropped to 0.03, and the validation set error was 0.035. The optimized evaluation model can accurately predict battery health status, for example, accurately classifying batteries in sub-health conditions, with a prediction accuracy of 92%.

[0203] In step (7.4), to improve prediction stability, the output of the evaluation model is processed using exponential smoothing, with a smoothing coefficient of 0.8 to reduce the impact of short-term fluctuations. For example, after the predicted health score fluctuates from 0.7 to 0.72, the smoothed result stabilizes at 0.71. The dynamic threshold calibration mechanism sets the threshold to plus or minus 10% of the predicted value. When the output exceeds the threshold, online calibration is triggered, adjusting the model parameters to adapt to the new operating conditions. The real-time optimized model can quickly respond to data changes in frequent power outage scenarios, ensuring the reliability of the evaluation results.

[0204] Furthermore, the dataset was divided into training and test sets in an 8:2 ratio using a holdout method. Performance metrics included accuracy and mean squared error (MSE). On the test set, the model achieved 92% accuracy and a mean squared error of 0.035. Validation results demonstrated that the model can effectively distinguish between healthy, subhealthy, and degraded states, for example correctly identifying subhealthy batteries at a depth of discharge of 60%. The final evaluation model, tested under multiple operating conditions, demonstrated high generalization capabilities, providing precise support for battery management.

[0205] In step (7) of the present invention, the health status assessment model generated through correlation analysis and weight optimization significantly improves the ability to capture the dynamic relationship between discharge depth and voltage recovery time. Compared with the traditional static weight method, this method achieves dynamic adjustment of feature weights and optimization of model structure through adaptive gradient descent and recursive neural network. Exponential smoothing and dynamic calibration mechanisms further enhance the stability of the model in real-time applications and reduce the risk of misjudgment.

[0206] (8) The retrained health status assessment model is used to analyze the discharge data under frequent power outage scenarios, determine the rotation priority, and generate a battery rotation plan.

[0207] The step (8) comprises the following steps:

[0208] (8.1) Use a sliding time window to obtain time series data on discharge depth and discharge times, calculate the capacity loss rate using the capacity decay curve, and obtain the battery health score;

[0209] The battery health status value is calculated through the optimized health status assessment model, and then the time series data of discharge depth and discharge times are obtained;

[0210] (8.2) constructing a state transition matrix for the battery health score, calculating the probability distribution of the normal state, the over-discharge state, and the deep discharge state through the state transition matrix, and obtaining a rotation index value;

[0211] A rotation index function is constructed based on the battery health score, and a state transition matrix of the discharge condition is established using a hidden Markov chain. Three states, normal, over-discharge, and deep discharge, are set. The rotation urgency is calculated through the state probability distribution to obtain the rotation index value.

[0212] (8.3) Feature classification of load power, discharge depth, and discharge duration is performed based on the rotation index value, and classification boundaries are calculated using a support vector machine to obtain rotation priority data;

[0213] Wherein, a multi-layer decision tree is established according to the rotation index value, and load power, discharge depth and discharge duration are used as classification features;

[0214] (8.4) Genetic encoding is performed on the rotation priority data, and a crossover mutation operation is performed on the rotation interval through the genetic encoding to obtain a battery rotation plan.

[0215] Among them, a rotation timing optimizer is constructed for the rotation priority data, a genetic algorithm is used to encode the rotation interval, the crossover probability and the mutation probability are set, the rotation timing is optimized by the fitness function, and the rotation sorting result is obtained, an execution rule base is constructed according to the rotation sorting result, the matching degree between the remaining capacity and the load power is used as the execution threshold, the rotation trigger condition is determined by threshold comparison, and the rotation execution sequence is obtained, and the rotation plan is dynamically adjusted using online monitoring data to obtain the rotation plan.

[0216] In step (8), under frequent power outages, the battery's discharge behavior directly affects its health status and power supply reliability. The optimized health status assessment model can accurately capture changes in battery performance. This step analyzes discharge data and combines state transition and classification algorithms to generate a scientific rotation plan to ensure balanced use of the battery pack and extend its overall life.

[0217] In step (8.1), the health status assessment model analyzes battery performance based on multi-parameter characteristics. A 300-second sliding time window is used to count the depth of discharge and the number of discharges. For example, a battery pack undergoes 15 discharges within 24 hours, with an average depth of discharge of 45%. Combined with the capacity decay model, the capacity loss rate of each cycle is calculated, for example, 0.02%, reflecting a slight degradation of battery performance. The discharge data and model output are combined to generate a health score. For example, a score of 85 indicates that the battery is in a healthy but transitional state that requires attention. The high or low health score provides an intuitive basis for rotation decisions.

[0218] In step (8.2), the rotation index function takes the health score as input and analyzes the dynamic evolution of the discharge condition through the hidden Markov model. Three states are defined: the normal state is a discharge depth of less than 30%, the over-discharge state is 30% to 60%, and the deep discharge state is more than 60%. The state transition matrix shows that the probability of the normal state turning into the over-discharge state is 0.3, and the probability of turning into the deep discharge state is 0.1. Based on the state probability distribution, the rotation urgency is calculated. For example, a value of 0.65 indicates that the battery needs to be rotated first. The urgency index quantifies the degradation risk of the battery under the current operating conditions and provides a basis for subsequent priority division.

[0219] In step (8.3), the rotation urgency is combined with multi-dimensional features for classification analysis. A classification model is constructed, and the features include load power, discharge depth and discharge duration, with thresholds set as 2000 watts and 1000 watts for high, medium and low load power, 35% and 55% for discharge depth, and 120 minutes as the standard for discharge duration. A support vector machine is used, and the radial basis kernel function is selected. The parameters are optimized through grid search to divide the urgency into three levels: emergency, priority and normal. For example, a battery with an urgency of 0.65 and a load power of 1500 watts is classified as priority. The generated priority ranking data clarifies the rotation order of each battery.

[0220] In step (8.4), the rotation sequence optimization aims to balance battery usage and power supply demand. A genetic algorithm is used with a population size of 100, a crossover probability of 0.8, and a mutation probability of 0.1. Real number encoding is used, and the chromosome length corresponds to the number of battery packs. For example, 10 batteries correspond to a 10-dimensional encoding. The fitness function combines the uniformity of the rotation interval with the load matching degree. After 200 generations of iteration, the optimal rotation sequence is generated. For example, batteries with high priority are assigned shorter rotation intervals, while batteries with low priority have longer service life. The optimized sequence ensures balanced degradation of the battery pack.

[0221] Furthermore, based on the rotation sequence, an execution rule base is constructed. The rotation trigger conditions are set by the matching degree between the remaining capacity and the load demand, and a rotation execution plan is generated. The rule base is based on actual working conditions. For example, a load of 1500 watts for 4 hours requires a remaining capacity of no less than 7.2 kWh. If the capacity of a battery is less than 1.2 times the demand, a rotation is triggered. The execution plan specifies the rotation time point for each battery. For example, high-priority batteries are rotated every 48 hours. The setting of the rule base improves the executability of the rotation operation and ensures uninterrupted power supply.

[0222] Furthermore, a feedback correction mechanism dynamically optimizes the rotation plan using hourly updated monitoring data. For example, if a battery's depth of discharge unexpectedly rises to 65% during operation, causing its health score to drop to 80, the correction mechanism will prioritize it and shorten the rotation interval to 24 hours. Online adjustments ensure the plan adapts to unexpected operating conditions. Monitoring data shows that the adjusted battery pack capacity variance is kept within 5%, significantly improving operational balance. The resulting rotation plan balances health status with actual needs, optimizing operational efficiency.

[0223] In step (8) of the present invention, the rotation plan generated by multi-level discharge data analysis and optimization algorithm effectively balances battery health and power supply reliability. Compared with the traditional fixed-cycle rotation, this method accurately divides the rotation priority through the hidden Markov model and support vector machine, and the genetic algorithm further optimizes the timing arrangement. The online correction mechanism enhances the dynamic adaptability of the plan and ensures stable operation in frequent power outage scenarios. The monitoring results show that the health score dispersion of the battery pack is reduced after rotation, and the overall life is effectively extended, providing strong support for the reliable operation and maintenance of the distribution terminal.

Claims

1. A method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion, characterized in that: The following steps are involved: (1) By collecting the discharge data of the distribution terminal battery under frequent power outage scenarios, the discharge depth and discharge times of deep discharge are obtained to obtain the discharge characteristic data set; (2) Analyze the discharge characteristic data set to obtain the voltage recovery curve after deep discharge, and determine the dynamic recovery characteristics based on the voltage recovery curve; (3) If the voltage recovery time of the dynamic recovery feature exceeds the preset threshold, the corresponding discharge depth and discharge times are extracted from the discharge feature data set to obtain the change trend of the internal resistance parameter; (4) Integrate the dynamic recovery characteristics and the change trend of the internal resistance parameters to construct a multi-parameter data set. The multi-parameter data set includes discharge depth, voltage recovery time, internal resistance value and terminal voltage fluctuation. The multi-parameter data set is standardized to obtain a comprehensive feature matrix; (5) The comprehensive feature matrix is ​​weighted by a multi-parameter fusion algorithm to identify the nonlinear relationship between discharge depth and voltage recovery time, and the feature weights and correlation coefficients obtained by the internal resistance are quantified. The health status feature vector is obtained based on the feature weights and correlation coefficients, and a health status assessment model is constructed. (6) Extract the weight of the internal resistance feature from the health state feature vector, calculate the degree of deviation between the voltage recovery time and the standard recovery time, and combine the degree of deviation and the internal resistance weight to predict the remaining battery life and obtain the life prediction value; (7) Analyze the correlation between discharge depth and voltage recovery time, adjust the feature weights in the multi-parameter fusion algorithm according to the life prediction value, optimize the correlation analysis, and retrain the health status assessment model through the adjusted algorithm; (8) The retrained health status assessment model is used to analyze the discharge data under frequent power outage scenarios, determine the rotation priority, and generate a battery rotation plan.

2. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (1) comprises the following steps: (1.1) collecting voltage sampling point data during the discharge process, and processing the sampling point data using a median filter method to obtain initial discharge characteristic data; (1.2) monitoring the discharge current according to the initial discharge characteristic data, and calculating the remaining capacity percentage by multiplying the current sampling value by the time to obtain the discharge number characteristic data; (1.3) Based on the discharge number characteristic data and the discharge cut-off voltage data, the average load power is calculated using a sliding time window to obtain discharge energy characteristic data; (1.4) classifying the discharge conditions according to the discharge energy characteristic data to obtain a data set of corresponding relationships between discharge depth and discharge times; (1.5) Using the support vector regression method to model the corresponding relationship dataset, outputting a discharge feature dataset.

3. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (2) comprises the following steps: (2.1) collecting a voltage sampling point sequence from the end of discharge to the recovery process through the distribution terminal, and using data smoothing to eliminate noise in the voltage sampling point sequence to obtain voltage time series characteristic data; (2.2) calculating the voltage rise per unit time based on the voltage time series characteristic data, obtaining the remaining battery charge change value through the state of charge calculation unit, and obtaining voltage recovery characteristic data; (2.3) collecting internal resistance sampling values ​​for the voltage recovery characteristic data, and performing three-layer decomposition and reconstruction of the internal resistance sampling values ​​using wavelet transform to obtain battery polarization characteristic data; (2.4) Performing piecewise linear fitting based on the battery polarization characteristic data to obtain a voltage recovery curve after deep discharge, and determining dynamic recovery characteristics based on the voltage recovery curve.

4. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (3) comprises the following steps: (3.1) Obtain the discharge depth value and the cumulative number of discharges in the corresponding time period based on the discharge feature data set to obtain the discharge feature sequence; (3.2) performing linear temperature compensation on the internal resistance measurement value according to the ambient temperature data at the corresponding moment of the discharge characteristic sequence to obtain compensated internal resistance data; (3.3) establishing a random forest regressor for the compensated internal resistance data, constructing a feature matrix using the discharge depth value and the cumulative number of discharges, and training an internal resistance predictor in combination with the state of charge change to obtain a predicted internal resistance value; (3.4) A time series is constructed for the predicted internal resistance value, and the internal resistance variation curve is fitted using the least squares method.

5. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (4) comprises the following steps: (4.1) Obtaining the voltage recovery time data and internal resistance parameter change trend curve in the dynamic recovery characteristics, and obtaining the initial multi-parameter data based on the discharge depth record and terminal voltage fluctuation measurement value; (4.2) calculating the upper and lower quartiles using the interquartile range method based on the initial multi-parameter data, and correcting the abnormal data using the median substitution method to obtain the corrected multi-parameter data; (4.3) performing zero-mean normalization processing on the corrected multi-parameter data, and performing multi-scale decomposition on the terminal voltage fluctuation record using discrete wavelet transform to obtain filtered voltage data; (4.4) The principal component analysis method is used to extract the characteristic vector based on the filtered voltage data and the standardized parameter data, and the comprehensive characteristic matrix is ​​obtained through orthogonal transformation matrix operation.

6. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (5) comprises the following steps: (5.1) Weight distribution is performed on the comprehensive feature matrix according to the fusion calculation unit to obtain the initial value of the parameter weight; (5.2) Using the sliding window method to quantify the correlation between the internal resistance and the discharge depth based on the initial value of the parameter weight, the feature weight is optimized by the gradient descent method to obtain the feature weight coefficient; (5.3) A combined feature function is constructed based on the feature weighting coefficients, a support vector regressor is used to fit the nonlinear relationship between discharge depth and recovery time, a three-layer perceptron network is constructed, and the network parameters are trained through a back propagation algorithm to obtain a health status assessment model.

7. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (6) comprises the following steps: (6.1) Smoothing the internal resistance data using a sliding average method for the health state feature vector, calculating the internal resistance change rate based on the smoothed data, and obtaining the internal resistance feature weight; (6.2) obtaining a time difference using a difference calculation method based on the internal resistance characteristic weight and the measured recovery time, calculating the recovery time deviation using a cumulative deviation function, and obtaining time deviation data; (6.3) Using the cycle number and internal resistance change rate to construct a time series feature for the time deviation data, and training a capacity prediction model using a long short-term memory network to obtain a capacity decay curve; (6.4) Extracting life characteristics using a rolling time window based on the capacity decay curve, establishing a remaining life prediction model using a grey prediction method, and obtaining life prediction data; (6.5) Compensating the predicted lifespan using a temperature coefficient, establishing a mapping relationship between temperature and lifespan using an exponential function, and obtaining a predicted lifespan value for a baseline operating condition; (6.6) Taking the standard recovery time as the benchmark, the difference between the voltage recovery time and the standard recovery time is normalized to obtain a standardized deviation metric. Combined with the weight of the internal resistance feature, the weighted fusion of the deviation metric and the internal resistance weight is calculated to generate a health degradation score. The health degradation score is used as the input of the life prediction model, and a preliminary life estimate is output. The remaining life is calculated to generate a life prediction value that represents the health status of the battery.

8. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 7, characterized in that: The step (6.6) comprises the following steps: (6.6.1) Obtaining a measured voltage recovery time based on the measured operating parameters, and subtracting a reference recovery time from the measured voltage recovery time to obtain a time difference sequence; (6.6.2) Performing maximum and minimum normalization on the time difference sequence, and obtaining a battery degradation score by linearly combining the deviation metric value and the internal resistance characteristic weight; (6.6.3) Using the battery degradation score to construct a long short-term memory network, and training the long short-term memory network using a sliding time window to obtain a preliminary lifespan estimate; (6.6.4) Calculate the cumulative deviation between the preliminary life estimate and the actual life, and use an exponential weighting method to compensate for the cumulative deviation to generate a life prediction value that represents the health status of the battery.

9. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (7) comprises the following steps: (7.1) The correlation coefficient is calculated using the Pearson algorithm based on the discharge depth data and the voltage recovery time series, and the characteristic correlation curve is obtained by polynomial fitting; (7.2) constructing a weight optimization function for the feature correlation curve, using the gradient descent method to calculate the weight update amount, and obtaining the updated feature weight through the dynamic learning rate; (7.3) constructing a parameter mapping relationship using a recursive neural network based on the updated feature weights, and selecting a hidden layer structure through cross-validation to obtain a feature combination function; (7.4) The mean square error is used as the loss function for the feature combination function, and the network parameters are trained by error back propagation to obtain the retrained health status assessment model.

10. The method for evaluating the health status of batteries at distribution terminals based on multi-parameter fusion according to claim 1, characterized in that: The step (8) comprises the following steps: (8.1) Use a sliding time window to obtain time series data on discharge depth and discharge times, calculate the capacity loss rate using the capacity decay curve, and obtain the battery health score; (8.2) constructing a state transition matrix for the battery health score, calculating the probability distribution of the normal state, the over-discharge state, and the deep discharge state through the state transition matrix, and obtaining a rotation index value; (8.3) Feature classification of load power, discharge depth, and discharge duration is performed based on the rotation index value, and classification boundaries are calculated using a support vector machine to obtain rotation priority data; (8.4) Genetic encoding is performed on the rotation priority data, and a crossover mutation operation is performed on the rotation interval through the genetic encoding to obtain a battery rotation plan.

Images