Lithium battery SOC rapid estimation method under low-temperature working condition

Abstract

The invention belongs to the field of lithium battery SOC estimation, and particularly relates to a lithium battery SOC rapid estimation method under a low-temperature working condition, and the method comprises the following steps: S1, building a sample data set, S2, screening key features based on a ReliefF algorithm, S3, carrying out SOC estimation based on a machine learning regression model, and S4, carrying out model training and evaluation, high-precision and rapid estimation of the SOC of the lithium battery is realized in a low-temperature environment, and the problems of low precision and complex calculation of a traditional method in a severe cold working condition are effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery SOC estimation, specifically involving a rapid method for estimating the SOC of lithium batteries under low-temperature conditions. Background Technology

[0002] Lithium-ion batteries, with their high energy density and long cycle life, have become the core power source for electric vehicles. State of charge (SOC), as a key indicator reflecting the remaining battery capacity, directly determines the driving safety, range reliability, and battery life of electric vehicles, while effectively alleviating range anxiety and preventing battery abuse such as overcharging and over-discharging. Currently, SOC estimation methods mainly include model-based methods, data-driven methods, and hybrid methods. Although existing technologies have made some progress, significant limitations remain: some model-based methods rely on complex electrochemical mechanism modeling, requiring stringent parameter matching; traditional data-driven methods often rely on massive amounts of monitoring data and do not specifically screen for redundant features in the original voltage and current signals, resulting in high model computational complexity, poor real-time performance, and difficulty in adapting to the dynamic operating scenarios of electric vehicles.

[0003] In low-temperature environments (such as 0°C), the electrochemical characteristics of lithium-ion batteries undergo significant changes, with reduced ion diffusion rates and increased internal resistance, further complicating SOC estimation. Existing methods are prone to accuracy degradation in this scenario. Therefore, developing an SOC estimation scheme that balances low-temperature adaptability, estimation accuracy, and computational efficiency has become a key requirement for promoting the large-scale application of electric vehicles in complex climatic conditions. Based on this background, this patent employs the ReliefF algorithm to filter high-dimensional features of battery voltage and current signals, eliminating redundant information and focusing on key features strongly correlated with SOC. It then combines various machine learning methods for comparative optimization, ultimately achieving a balance between high accuracy and high efficiency. The significance of this research lies in addressing the pain points of traditional methods, such as strong data dependence, computational complexity, and insufficient low-temperature adaptability, providing more engineering-approved SOC estimation technology support for battery management systems. Summary of the Invention

[0004] The technical problem this invention aims to solve is to provide a rapid method for estimating the state of charge (SOC) of lithium batteries under low-temperature conditions. By combining ReliefF feature selection with a machine learning model, high-precision and rapid estimation of lithium battery SOC is achieved in low-temperature environments, effectively solving the problems of low accuracy and computational complexity of traditional methods under severe cold conditions.

[0005] This invention provides a method for rapid estimation of the state of charge (SOC) of a lithium battery under low-temperature operating conditions, comprising the following steps: S1. Sample Dataset Construction: Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. S2. Key feature selection based on the ReliefF algorithm: From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. S3. SOC estimation based on machine learning regression model: Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. S4. Model Training and Evaluation: The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

[0006] Furthermore, the feature importance evaluation using the ReliefF algorithm in S2 includes: S21. Assign initial weights to all features in the feature set; S22. Randomly select samples from the sample dataset. R And search in its own class of samples and in samples of different classes respectively. k The nearest neighbor; S23, Based on the sample R The weights of each feature are iteratively updated by considering the distance differences between its similar and dissimilar neighbors on each feature and the prior probability of the category. S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset.

[0007] Furthermore, the feature importance evaluation using the ReliefF algorithm in S2 includes: S21. For a feature set containing n features, set the initial weight of each feature to 0. S22. From the sample dataset containing m samples, randomly select a sample R, which has a feature vector ( f 1, f 2, , f n Search among similar samples in sample R. k nearest neighbors H j And find the k nearest neighbors in different classes of samples. M j ; S23, According to the sample R Based on the feature differences with its nearest neighbors, update the weights of each feature; for each feature... f i Its weight update model is:

[0008] in, P ( C Let be the prior probability of category C, i.e., the ratio of the number of samples in category C to the number of samples in category D. R The ratio of the total number of samples in the same category. f i For the sample R, the first i One characteristic, h ji For the first j Nearest neighbors of the same type H j The i One characteristic, m ji For the first j Different class nearest neighbors M j The i One characteristic, k This represents the number of nearest neighbors of the same class / different classes. z The number of samplings is the discrete feature diff( f i , R , H j ) and diff( f i , R , M j The formula for calculating ) is:

[0009] in, H j For the first j The nearest neighbor of the same kind, M j For the firstj The nearest neighbors of different classes.

[0010] S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset.

[0011] Furthermore, in S1: The multi-dimensional statistical features extracted from the voltage and current time-series signals are 18-dimensional, totaling 36-dimensional initial features; the 18-dimensional statistical features include: maximum value, minimum value, mean, median, upper quartile, lower quartile, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, margin factor, mean of absolute value, energy, and coefficient of variation; The sliding window method has a window size of 500 data points and a step size of 100 data points.

[0012] Furthermore, in step three, the machine learning regression model used includes at least the LightGBM model.

[0013] Furthermore, the low-temperature operating conditions involved in S1 and S4 are 0°C environments; The construction of the sample dataset and the training and evaluation of the model are based on battery test data under 0°C conditions.

[0014] Furthermore, the model training and evaluation in S4 are based on a specific low-temperature dataset, which is collected from various driving cycle tests in a 0°C environment and divided into a training set and an independent test set according to the driving cycle type. The training set includes data from US06, HWFET, UDDS, LA92, Neural Network, and Cycle 4 driving cycles, while the test set includes data from Cycle 1, Cycle 2, and Cycle 3 driving cycles.

[0015] Furthermore, the sampling time step of the dataset under the 0°C environment is 0.1 seconds; and after each driving cycle test, the battery is charged at a constant current rate of 1C to 4.2V, and then charged at a constant voltage rate of 50mA until the current is lower than the cutoff current.

[0016] This invention also provides a method and apparatus for rapid estimation of the state of charge (SOC) of lithium batteries under low-temperature operating conditions, the apparatus comprising: The sample dataset construction module is used for S1: Sample dataset construction. Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. The key feature filtering module is used for S2: key feature filtering based on the ReliefF algorithm. From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. The SOC estimation module is used for S3: SOC estimation based on a machine learning regression model. Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. The model training and evaluation module is used for S4 model training and evaluation. The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

[0017] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above method.

[0018] The beneficial effects of this invention are that the rapid SOC estimation method for lithium batteries under low-temperature conditions provided by this invention achieves rapid and high-precision estimation of lithium battery SOC under low-temperature conditions through the effective fusion of ReliefF feature selection and machine learning models. This significantly reduces the dependence of traditional methods on complex mechanism models or massive amounts of data, and greatly improves computational efficiency while ensuring estimation accuracy. It provides a practical and feasible technical solution for real-time and reliable state estimation of battery management systems in low-temperature environments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method in this invention; Figure 2 This is a graph of the training set SOC data in this invention; Figure 3This is a graph of the SOC data for the test set in this invention; Figure 4 This is a graph showing the SOC estimation results of the linear regression model in this invention; Figure 5 This is a graph showing the SOC estimation results of the Lasso regression model in this invention; Figure 6 This is a graph showing the SOC estimation results of the support vector regression model in this invention; Figure 7 This is a graph showing the SOC estimation results of the random forest model in this invention; Figure 8 This is a graph showing the SOC estimation results of the lightGBM model in this invention; Figure 9 This is a diagram showing the SOC estimation result of the BP neural network model in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0022] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0023] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0025] like Figures 1-9 As shown, this invention provides a method for rapid estimation of the state of charge (SOC) of a lithium battery under low-temperature operating conditions, comprising the following steps: S1. Sample Dataset Construction: Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. In this step, the original voltage and current time-series signals are transformed into structured, labeled sample data. The sliding window effectively captures the time-series correlation between electrical characteristics and SOC under dynamic operating conditions, providing high-quality, spatiotemporally relevant input for subsequent machine learning modeling.

[0026] S2. Key feature selection based on the ReliefF algorithm: From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. In this step, the core features most relevant to the SOC are automatically identified and selected from the high-dimensional feature set, significantly eliminating redundant information and noise, and reducing data dimensionality. This step not only improves the efficiency of subsequent model training and inference but also enhances the model's ability to focus on key signals, laying a feature foundation for high-precision estimation.

[0027] S3. SOC estimation based on machine learning regression model: Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. In this step, a machine learning model with nonlinear fitting capabilities is constructed using the selected core features. This model can accurately characterize the complex mapping relationship between SOC and voltage characteristics at low temperatures. By comparing various models, the estimator with the best accuracy and efficiency (such as the lightGBM model) can be selected to achieve fast and stable SOC output.

[0028] S4. Model Training and Evaluation: The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

[0029] In this step, training and testing sets strictly divided according to driving cycle type are used for model training and validation to ensure that the model has good generalization ability and adaptability to operating conditions. By comprehensively evaluating accuracy (MSE, R²) and efficiency (training time), it can be quantitatively confirmed whether the selected model meets the practical requirements of being both fast and accurate in low-temperature scenarios.

[0030] The present invention provides a rapid method for estimating the state of charge (SOC) of lithium batteries under low-temperature conditions. By effectively integrating ReliefF feature selection with a machine learning model, it achieves rapid and high-precision estimation of lithium battery SOC under low-temperature conditions. This significantly reduces the dependence of traditional methods on complex mechanistic models or massive amounts of data, and greatly improves computational efficiency while ensuring estimation accuracy. It provides a practical technical solution for real-time and reliable state estimation of battery management systems in low-temperature environments.

[0031] In one embodiment, the feature importance evaluation using the ReliefF algorithm in S2 includes: S21. Assign initial weights to all features in the feature set. In this step, a unified evaluation starting point is established for all features to be evaluated, ensuring that the calculation of feature importance weights starts from a fair and consistent benchmark, providing a clear initial state for subsequent iterative updates.

[0032] S22. Randomly select samples from the sample dataset. R And search in its own class of samples and in samples of different classes respectively. kThe nearest neighbor is calculated; in this step, by randomly sampling and combining nearest neighbor searches of the same and different classes, the ability of features to distinguish categories (SOC intervals) in the local space of the samples can be efficiently simulated. This step ensures that the evaluation process fully explores the data distribution, making weight updates based on representative local comparisons.

[0033] S23, Based on the sample R By analyzing the distance differences between the sample and its neighbors of the same and different classes on various features, and combining this with prior class probabilities, the weights of each feature are iteratively updated. In this step, by repeatedly calculating the differences between the sample and its neighbors on various features and adjusting the weights based on prior class probabilities, the sensitivity and contribution of each feature to changes in SOC can be gradually and stably quantified. This iterative mechanism makes the evaluation results of feature weights statistically robust, effectively capturing the stable and universal correlation between features and SOC.

[0034] S24. After reaching the preset number of iterations, the features are sorted according to their final weights, and features with weights higher than the set threshold are selected to form the core feature subset. In this step, the features are sorted and threshold-filtered based on their final weights after iteration convergence, which can objectively and automatically identify the set of core features most relevant to the SOC. This step directly outputs the high-quality feature subset after dimensionality reduction, providing a refined and efficient input for subsequent machine learning models, fundamentally reducing model complexity and improving computational speed.

[0035] The ReliefF algorithm described in this embodiment achieves automated and objective screening of core features from high-dimensional battery characteristics through a complete and iterative weight calculation and update mechanism. It not only significantly improves the reliability and efficiency of feature screening, but more importantly, it reveals a stronger intrinsic correlation between static voltage distribution characteristics and State of Charge (SOC) under low-temperature conditions through a data-driven approach, achieving a precise and fast SOC estimation model.

[0036] In one embodiment, the feature importance evaluation using the ReliefF algorithm in S2 includes: S21. For a feature set containing n features, set the initial weight of each feature to 0. S22. From the sample dataset containing m samples, randomly select one sample. R It has a feature vector ( f 1, f 2, , f n ); in the sample R Find the k nearest neighbors in the same type of sample H jAnd search in different classes of samples k nearest neighbors M j ; S23, According to the sample R Based on the feature differences with its nearest neighbors, update the weights of each feature; for each feature... f i Its weight update model is:

[0037] in, P ( C Let be the prior probability of category C, i.e., the ratio of the number of samples in category C to the number of samples in category D. R The ratio of the total number of samples in the same category. f i For the sample R, the first i One characteristic, h ji For the first j Nearest neighbors of the same type H j The i One characteristic, m ji For the first j Different class nearest neighbors M j The i One characteristic, k This represents the number of nearest neighbors of the same class / different classes. z The number of samples. Discrete feature diff( f i , R , H j ) and diff( f i , R , M j The formula for calculating ) is:

[0038] in, H j For the first j The nearest neighbor of the same kind, M j For the first j The nearest neighbors of different classes.

[0039] In this step, the sample is calculated. R The normalized difference (diff) between features and those of nearest neighbors is used, and the differences between similar and dissimilar nearest neighbors are contributed to the weights in opposite directions, thus accurately quantifying the discriminative power of each feature for the SOC category. This is achieved by introducing prior class probabilities.P ( C The mechanism weights contributions from neighbors of different classes, automatically balancing potential biases caused by imbalances in sample sizes across different SOC intervals, making the weight evaluation results more globally representative and robust. By iterating this process multiple times, the weights of each feature steadily and gradually converge to a reliable value, thus reliably capturing the complex and nonlinear correlation strength between features and SOC.

[0040] S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset.

[0041] In one embodiment, in S1: The multi-dimensional statistical features extracted from the voltage and current time-series signals are 18-dimensional, totaling 36-dimensional initial features; the 18-dimensional statistical features include: maximum value, minimum value, mean, median, upper quartile, lower quartile, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, margin factor, mean of absolute value, energy, and coefficient of variation; Eighteen statistical features were extracted from both voltage and current signals to comprehensively and multi-dimensionally characterize the battery's electrical behavior during operation. The selected 18 features cover the signal's central tendency (e.g., mean, median), dispersion (e.g., variance, standard deviation), distribution shape (e.g., skewness, kurtosis), energy intensity (e.g., root mean square, energy), and waveform characteristics (e.g., peak factor, impulse factor). This comprehensive feature engineering strategy ensures that the information extracted from the original signal reflects both static statistical characteristics and captures dynamic change patterns, providing a rich and complementary descriptive foundation for subsequent analysis and effectively avoiding the omission of key information due to singular feature representation.

[0042] The sliding window method has a window size of 500 data points and a step size of 100 data points.

[0043] A window length of 500 points is sufficient to cover typical acceleration, deceleration, and cruising phases in most driving cycles, ensuring that statistically significant and representative features can be extracted within each window. A sliding window with a step size of 100 points allows for substantial overlap between adjacent windows. This effectively maintains the continuity of temporal context information, avoids dynamic information breaks caused by excessive window jumps, and ensures that the constructed sample dataset can smoothly and coherently represent the evolution of SOC over time.

[0044] In one embodiment, the machine learning regression model used in step three includes at least the LightGBM model. Choosing the LightGBM model as the SOC estimator allows for near-theoretical-limit estimation accuracy with minimal computation time under complex low-temperature conditions, thus providing a superior technical solution for real-time state estimation of vehicle battery management systems that combines advancement, practicality, and reliability.

[0045] In one embodiment, the low-temperature operating condition involved in S1 and S4 is a 0°C environment; The construction of the sample dataset and the training and evaluation of the model are based on battery test data under 0°C conditions.

[0046] This embodiment strictly limits the entire method process to the typical and harsh low-temperature condition of 0℃, and uses dedicated test data under this environment for all modeling and verification. This ensures that the developed SOC estimation technology directly addresses and solves the estimation problem caused by the increase in internal resistance and sudden change in characteristics of batteries in cold environments. Thus, it provides a fully verified, highly reliable, high-precision, and fast SOC estimation solution that can be directly applied to cold regions or winter scenarios.

[0047] In one embodiment, model training and evaluation in S4 are based on a specific low-temperature dataset. This dataset is collected from various driving cycle tests at 0°C and divided into a training set and an independent test set according to the driving cycle type. The training set includes data from US06, HWFET, UDDS, LA92, Neural Network, and Cycle 4 driving cycles, while the test set includes data from Cycle 1, Cycle 2, and Cycle 3 driving cycles. This embodiment, by employing a dataset strategy that strictly divides the dataset according to driving cycle type, ensures that the model is exposed to and learns the battery dynamic characteristics under various representative driving conditions (such as US06 and UDDS) during the training phase, while using completely independent driving cycles (Cycles 1-3) for testing during the evaluation phase. This effectively verifies the model's generalization ability and robustness to unknown conditions. This division method not only simulates the actual situation of batteries facing diverse driving scenarios in real-world applications but also eliminates evaluation bias caused by data leakage, providing a crucial guarantee for the reliability of the SOC estimation model in practical applications at 0°C.

[0048] In one embodiment, the data set in the 0°C environment has a sampling time step of 0.1 seconds; and after each driving cycle test, the battery is charged at a constant current rate of 1C to 4.2V, and then charged at a constant voltage rate of 50mA until the current is lower than the cutoff current.

[0049] This embodiment effectively captures the rapid dynamic changes in battery voltage and current signals by setting a high sampling time step of 0.1 seconds, providing a high-resolution data foundation for feature extraction and ensuring that the subsequently constructed statistical features can accurately characterize the transient response of SOC. At the same time, the standardized recovery process of strictly adopting 1C constant current charging to 4.2V and completing constant voltage charging with a 50mA cutoff current ensures the consistency of battery state after each test, controlling the repeatability and comparability of experimental conditions from the source. This makes the constructed dataset highly reliable and standardized, providing solid and high-quality data support for the model to achieve stable and reliable SOC estimation in a 0℃ low-temperature environment.

[0050] This invention provides a specific embodiment to further illustrate the detailed implementation of the feature importance evaluation based on the ReliefF algorithm. This embodiment corresponds to the technical solution described in the claims, as follows: To accurately capture the correlation between lithium battery voltage, current signals and SOC, a high-dimensional feature set is first constructed through feature extraction and sliding window processing. Then, the ReliefF algorithm is used to remove redundant information and select core features to support subsequent SOC estimation.

[0051] S1. Sample Dataset Construction: Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. A sliding window method with a set window size and step size is used to process the time-series signals. The statistical features of voltage and current are calculated independently for each window, generating a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. Specifically: For voltage and current time series, 18 statistical features (36 features in total) are extracted for each, comprehensively covering the static distribution and dynamic change characteristics of the signal. The calculation formulas and definitions of each feature are as follows:

[0052]

[0053] To construct spatiotemporal correlation features, a sliding window method was used: the window size was set to 500 and the step size to 100, based on the dataset sampling interval. p (Unit: seconds), each window corresponds to 500 pA continuous signal segment of one second. For each window, the above 18-dimensional features are calculated independently for the voltage and current signals, generating a 36-dimensional feature vector; the SOC value corresponding to the last timestamp of the window is used as the target label to form a sample dataset composed of feature vector and SOC label, that is, a 36-dimensional feature vector-SOC value input-output pair, which effectively captures the spatiotemporal correlation between electrical characteristics and SOC.

[0054] S2. Key feature selection based on the ReliefF algorithm: S2. Key feature selection based on the ReliefF algorithm: From the sample dataset constructed in S1, all high-dimensional feature vectors are extracted to form a feature set. The ReliefF algorithm is used to evaluate the feature importance of this feature set. Based on the calculated feature weights, the features are sorted, and those with weights higher than a set threshold are selected to form the core feature subset used for SOC estimation. The details are as follows: Based on the 36-dimensional feature vector-SOC value input-output pair, the correlation between features and SOC is quantified using the ReliefF algorithm to select core features. The steps are as follows: S21, Feature weight initialization: Set the initial weights of all features to 0; S22, Randomly select samples from the sample dataset. R And search in its own class of samples and in samples of different classes respectively. k nearest neighbors: from the dataset (including m One sample n Randomly select samples from (features) R (Eigenvector) f 1, f 2, , f n )),search k Nearest neighbors of the same type H j (Eigenvector) h j1 , h j2 , , h jn ))and k Different class nearest neighbors M j (Eigenvector) m j1 , m j2 , , m jn )); S23, based on the distance differences between the sample R and its nearest neighbors of the same class and those of different classes on each feature, and combined with the prior probability of the class, iteratively update the weights of each feature: combined with the prior probability of the class. P ( C (Category C sample size and non-Category C sample size) R The weights are updated based on the ratio of the total number of samples in each category. The model is as follows: (1) in, P ( C Let be the prior probability of category C, i.e., the ratio of the number of samples in category C to the number of samples in category D. R The ratio of the total number of samples in the same category. f i For the sample R The i One characteristic, h ji For the first j Nearest neighbors of the same type H j The i One characteristic, m ji For the first j Different class nearest neighbors M j The i There are 3 features, k is the number of nearest neighbors of the same / different classes, and z is the number of samplings. Discrete feature diff( f i , R , H j ) and diff( f i , R , M j The formula for calculating ) is: (2) in, H j For the first j The nearest neighbor of the same kind, M j For the first j The nearest neighbors of different classes.

[0055] S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset: Repeat steps S21-S24 until the preset number of iterations, sorting them according to their final weights. W ( f i ) Sorting features, retaining features with weights higher than the threshold.

[0056] S3. SOC estimation based on machine learning regression model: Using the core feature subset obtained in S2 as input, at least one machine learning regression model is employed to estimate the SOC; the machine learning regression model includes any one or a combination of linear regression, support vector regression, Lasso regression, random forest, BP neural network, or LightGBM; specifically as follows: To achieve accurate and efficient estimation of lithium battery state of charge (SOC), this invention employs six mainstream regression-based machine learning methods. Based on the 8-dimensional voltage core features selected by the ReliefF algorithm, an SOC estimation model is constructed and its performance is compared. Finally, the optimal solution that balances accuracy and efficiency is selected. The core principles, mathematical framework, and adaptability of each method are as follows: 1. Linear Regression Model The linear regression model establishes a direct linear mapping relationship between the selected features and the SOC (Sort of Components). Assuming the SOC is a weighted linear combination of the input features, the expression is: (3) In the formula, θ 0 represents the bias term. θ 1- θ 8 represents the weights of each feature. x 1- x 8 represents the core voltage features after screening. This method solves for the optimal weights by minimizing the sum of the squared errors between the predicted and actual SOC values, i.e., (4) Where m is the number of samples. Its advantages lie in its simple model structure, strong interpretability, and fast training speed, making it a benchmark model for SOC estimation. However, it is difficult to capture the nonlinear relationship between features and SOC.

[0057] 2. Support Vector Regression (SVR) Model Support Vector Regression (SVR) models map input features to a high-dimensional space using kernel functions, constructing an optimal hyperplane to achieve SOC prediction. The core principle is to maximize the margin between the sample and the hyperplane while minimizing the prediction error. A soft margin is introduced to tolerate small biases, and its regression function is defined as: (5) Where ω is the weight vector, φ(x) is the nonlinear mapping function of the features, and b is the bias term. The optimization objective is to minimize the regularization loss function, balance model complexity and prediction accuracy, avoid direct computation in high-dimensional space through kernel functions, effectively handle the nonlinear relationship between features and SOC, and improve estimation robustness.

[0058] 3. Lasso Regression Model The Lasso regression model is a linear regression method with L1 regularization. It adds an L1 norm penalty term to the feature coefficients based on the least squares loss function, and the optimization objective is: (6) Where λ is the regularization parameter, controlling the intensity of the penalty. θ j Feature coefficients. This method can compress redundant or low-correlation feature coefficients to zero, achieve secondary feature selection, simplify the model structure and avoid overfitting, and is suitable for SOC estimation in high-dimensional feature scenarios, but its fitting ability for nonlinear relationships is limited.

[0059] 4. Random Forest Model Random forest models are ensemble learning methods that use bootstrap sampling to generate multiple differential training subsets. Each subset independently constructs a decision tree, and each decision tree node randomly selects a subset of features when splitting. The final SOC prediction result is the average of the predictions from all decision trees. (7) Where T is the number of decision trees. f t ( x The method can effectively capture the complex nonlinear relationship between features and SOC, has strong anti-overfitting ability, and can output feature importance scores to further verify the effectiveness of feature selection, but the training process is relatively time-consuming.

[0060] 5. Backpropagation (BP) neural network The BP neural network model is a multi-layer feedforward neural network consisting of an input layer, hidden layers, and an output layer. The input layer receives 8-dimensional voltage features, and the output layer outputs the predicted SOC value. The network generates prediction results through forward propagation and calculates the prediction error through backpropagation. Gradient descent is used to adjust the weights and biases of each layer to minimize error loss. This network can fit complex nonlinear mapping relationships and has high estimation accuracy, but it requires fine-tuning of hyperparameters (such as the number of hidden layers, the number of neurons, and the learning rate), and training is time-consuming.

[0061] 6. LightGBM model The lightGBM model is a gradient boosting framework based on decision trees, employing a leaf-growing strategy and histogram feature discretization to improve training efficiency. The model corrects the prediction errors of preceding ensemble models by successively constructing decision trees and minimizes the loss function using gradient descent. This method can efficiently process large-scale feature data, quickly capture the interaction relationships between features, and balance estimation accuracy with training speed, making it suitable for resource-constrained or fast-response lithium battery SOC estimation scenarios.

[0062] S4, Model Training and Evaluation: The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

[0063] This invention provides a specific embodiment to verify the practical effect of the aforementioned fast lithium battery SOC estimation method based on the ReliefF algorithm and machine learning fusion model under 0℃ low-temperature conditions. This verification is based on the publicly available Panasonic 18650PF lithium battery 0℃ test dataset. Through the aforementioned feature extraction, filtering, modeling, and evaluation process, the performance of various machine learning models is compared. The specific implementation and results are as follows: Based on the Panasonic 18650PF lithium-ion battery dataset, which was obtained by testing brand-new 2.9Ah Panasonic 18650PF battery cells in a constant temperature chamber using a professional battery testing instrument, the dataset covers multiple test types under five ambient temperatures. After each test, the cells were charged to 4.2V at a 1C rate with a 50mA cutoff current set to ensure battery state consistency, providing high-quality data support for SOC estimation and verification under low-temperature conditions.

[0064] Considering the significant impact of low temperature environment on the electrochemical characteristics of lithium battery, nine driving cycle data at 0℃ were selected as experimental objects, including Cycle 1 to Cycle 4, US06, HWFET, UDDS, LA92 and Neural Network cycles. The data sampling time step was 0.1 seconds, which can provide high-resolution voltage, current and temperature data to meet the training and evaluation requirements of SOC estimation model.

[0065] To avoid overfitting, the raw data at 0°C was divided into training and testing sets based on the driving cycle type. The training set integrates data from six driving cycles: US06, Neural Network, HWFET, Cycle 4, LA92, and UDDS, used for model parameter training. The testing set contains data from three driving cycles: Cycle 1, Cycle 2, and Cycle 3, used to verify the model's generalization performance, achieving a clear separation between training and validation data. Specifically... Figure 1 and Figure 2 As shown.

[0066] The feature ranking results show that voltage-related features rank highly, while current-related features generally have low correlation, with only one current feature (Icv, coefficient of variation) making it into the top 10. The top 8 are all voltage features, namely Vmax (0.722), Vp75 (0.657), Venergy (0.630), Vrms (0.601), Vmean (0.596), Vmedian (0.574), Vp25 (0.482), and Vmin (0.353). The score of Icv, which ranks 9th, is only 0.073, far lower than that of Vmin, which ranks 8th, highlighting that current features have a limited contribution to SOC estimation, and that the correlation between voltage signals and SOC is more significant.

[0067] Table 1 Feature Analysis and Screening Based on ReliefF

[0068] To eliminate redundant information and reduce the computational complexity of the model, the top 8 voltage features by weight were selected as inputs for the subsequent SOC estimation model. This selection result verifies that the static distribution characteristics of battery voltage are directly related to SOC changes, accurately capturing the core influencing factors of SOC during discharge and load changes, thus laying the foundation for improving the model's estimation accuracy and efficiency.

[0069] Based on the 8-dimensional voltage core features selected by the ReliefF algorithm, the aforementioned 6 machine learning algorithms were used to conduct lithium battery SOC estimation experiments at 0℃. The model performance was evaluated from three dimensions: mean square error (MSE), coefficient of determination (R²), and training time. The results showed that there were significant differences in accuracy and efficiency among the different algorithms.

[0070] The BP neural network model has the highest estimation accuracy, with an MSE as low as 0.0014 and an R² of 0.9799. It can accurately capture the complex nonlinear relationship between features and SOC, but it requires high computational costs and a training time of up to 15.108 seconds, which limits its practicality in resource-constrained or fast-response scenarios.

[0071] The lightGBM model performs exceptionally well, achieving a high accuracy close to that of the BP neural network model with an MSE of 0.0023 and an R² of 0.9681. Moreover, its training time is only 0.348 seconds, which is more than 45 times faster than the BP neural network model, thus achieving the optimal balance between accuracy and efficiency.

[0072] The random forest model achieves similar accuracy to the lightGBM model, with an MSE of 0.0024 and an R² of 0.9672, but its training time is 7.634 seconds, making it significantly less efficient than the lightGBM model. The support vector regression model exhibits moderate accuracy, with an MSE of 0.0046 and an R² of 0.9364, and a training time of approximately 1 second, placing it in the middle range between accuracy and efficiency.

[0073] Linear regression and Lasso cross-validation regression were the fastest linear models to train, with times of 0.030 seconds and 0.071 seconds respectively. However, their estimation accuracy was significantly insufficient, with MSEs of 0.0211 and 0.0188, and R² values ​​of only 0.7082 and 0.7391. This result indicates a complex nonlinear relationship between the selected voltage characteristics and SOC, which linear models struggle to adequately fit, while ensemble learning and neural network methods are better suited for this estimation scenario.

[0074] In summary, the lightGBM model demonstrates the optimal balance between accuracy and efficiency in estimating the SOC of lithium batteries under 0℃ low-temperature conditions, providing an ideal solution for the engineering application of practical battery management systems.

[0075] Table 2 Comparison of SOC estimation results

[0076] This invention also provides a method and apparatus for rapid estimation of the state of charge (SOC) of lithium batteries under low-temperature operating conditions, the apparatus comprising: The sample dataset construction module is used for S1: Sample dataset construction. Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. The key feature filtering module is used for S2: key feature filtering based on the ReliefF algorithm. From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. The SOC estimation module is used for S3: SOC estimation based on a machine learning regression model. Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. The model training and evaluation module is used for S4 model training and evaluation. The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

[0077] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0078] The above description is merely an embodiment and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations, modifications, or alterations to the technical solutions of the present invention without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, should fall within the protection scope of the present invention.

Claims

1. A method for rapid estimation of the state of charge (SOC) of a lithium battery under low-temperature operating conditions, characterized in that, Includes the following steps: S1. Sample Dataset Construction: Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. S2. Key feature selection based on the ReliefF algorithm: From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. S3. SOC estimation based on machine learning regression model: Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. S4. Model Training and Evaluation: The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

2. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in claim 1, characterized in that, The feature importance evaluation using the ReliefF algorithm in S2 includes: S21. Assign initial weights to all features in the feature set; S22. Randomly select samples from the sample dataset. R And search in its own class of samples and in samples of different classes respectively. k The nearest neighbor; S23, Based on the sample R The weights of each feature are iteratively updated by considering the distance differences between its similar and dissimilar neighbors on each feature and the prior probability of the category. S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset.

3. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in claim 2, characterized in that, The feature importance evaluation using the ReliefF algorithm in S2 includes: S21. For a feature set containing n features, set the initial weight of each feature to 0. S22. From the sample dataset containing m samples, randomly select one sample. R It has a feature vector ( f 1, f 2, , f n ); in the sample R Searching for similar samples k nearest neighbors H j And search in different classes of samples k nearest neighbors M j ; S23, According to the sample R Based on the feature differences with its nearest neighbors, update the weights of each feature; for each feature... f i Its weight update model is: in, P ( C Let be the prior probability of category C, i.e., the ratio of the number of samples in category C to the number of samples in category D. R The ratio of the total number of samples in the same category. f i For the sample R The i One characteristic, h ji For the first j Nearest neighbors of the same type H j The i One characteristic, m ji For the first j Different class nearest neighbors M j The i One characteristic, k This represents the number of nearest neighbors of the same class / different classes. z The number of samplings is the discrete feature diff( f i , R , H j ) and diff( f i , R , M j The formula for calculating ) is: in, H j For the first j The nearest neighbor of the same kind, M j For the first j The nearest neighbors of different classes; S24. After reaching the preset number of iterations, sort the features according to their final weights, and select the features with weights higher than the set threshold to form the core feature subset.

4. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in any one of claims 1-3, characterized in that, In S1: The multi-dimensional statistical features extracted from the voltage and current time-series signals are 18-dimensional in total, forming a total of 36-dimensional initial features; The 18-dimensional statistical features include: maximum value, minimum value, mean, median, upper quartile, lower quartile, variance, standard deviation, kurtosis, skewness, root mean square, waveform factor, peak factor, impulse factor, margin factor, mean of absolute value, energy, and coefficient of variation. The sliding window method has a window size of 500 data points and a step size of 100 data points.

5. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in any one of claims 1-3, characterized in that, In step three, the machine learning regression model used includes at least the LightGBM model.

6. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in any one of claims 1-3, characterized in that, The low-temperature operating conditions involved in S1 and S4 refer to a 0°C environment; The construction of the sample dataset and the training and evaluation of the model are based on battery test data under 0°C conditions.

7. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in claim 6, characterized in that, The model training and evaluation in S4 are based on a specific low-temperature dataset, which is collected from various driving cycle tests in a 0°C environment and divided into a training set and an independent test set according to the driving cycle type. The training set includes data from US06, HWFET, UDDS, LA92, Neural Network, and Cycle 4 driving cycles, while the test set includes data from Cycle 1, Cycle 2, and Cycle 3 driving cycles.

8. The method for rapid estimation of lithium battery SOC under low-temperature conditions as described in claim 7, characterized in that, The dataset in the 0℃ environment has a sampling time step of 0.1 seconds; and after each driving cycle test, the battery is charged at a constant current rate of 1C to 4.2V, and then charged at a constant voltage rate of 50mA until the current is lower than the cutoff current.

9. A method and apparatus for rapid estimation of the state of charge (SOC) of a lithium battery under low-temperature operating conditions, characterized in that, The device includes: The sample dataset construction module is used for S1: Sample dataset construction. Multi-dimensional statistical features are extracted from the voltage and current time-series signals of the battery. The time-series signals are processed using a sliding window method with a set window size and step size. The statistical features of voltage and current are calculated independently for each window to generate a high-dimensional feature vector. The SOC value corresponding to the last timestamp of the window is used as the target label of the feature vector, thereby constructing a sample dataset composed of feature vectors and SOC labels. The key feature filtering module is used for S2: key feature filtering based on the ReliefF algorithm. From the sample dataset constructed by S1, all high-dimensional feature vectors are extracted to form a feature set; the ReliefF algorithm is used to evaluate the feature importance of the feature set, and the features are sorted according to the calculated feature weights. Features with weights higher than a set threshold are selected to form the core feature subset for SOC estimation. The SOC estimation module is used for S3: SOC estimation based on a machine learning regression model. Using the subset of core features obtained in S2 as input, SOC estimation is performed using at least one machine learning regression model; the machine learning regression model includes any one or a combination of linear regression model, support vector regression model, Lasso regression model, random forest model, BP neural network model or LightGBM model. The model training and evaluation module is used for S4 model training and evaluation. The S3 model was trained on a low-temperature operating condition dataset, and the model accuracy was evaluated using mean squared error and coefficient of determination. The model performance was comprehensively evaluated by combining the training time.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images