Lithium ion battery energy storage capacity dynamic prediction method

Abstract

The invention discloses a lithium ion battery energy storage capacity dynamic prediction method, which comprises the following steps: acquiring cyclic aging data of different lithium ion batteries and carrying out data preprocessing to obtain a voltage-capacity curve; performing numerical derivation on the preprocessed voltage-capacity curve to obtain a dV / dQ curve; determining recurrence plot parameters according to the dV / dQ sequence and generating a corresponding battery characteristic recurrence plot; forming a sample pair by the battery characteristic recurrence plot and the corresponding capacity value to train a battery capacity correlation prediction model; and generating a battery characteristic recurrence plot of the lithium ion battery to be predicted in real time, inputting the battery characteristic recurrence plot into the trained battery capacity correlation prediction model to obtain a corresponding capacity prediction value, if an error between the capacity prediction value and the actual capacity value exceeds a threshold value, retraining the battery capacity correlation prediction model, and if the error does not exceed the threshold value, retraining the battery capacity correlation prediction model. And if so, outputting the capacity prediction value as a capacity prediction result. According to the invention, a more reliable and intelligent solution is provided for battery health management of the energy storage system.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and more specifically to a method for dynamically predicting the energy storage capacity of lithium-ion batteries. Background Technology

[0002] With the acceleration of the global energy transition, the penetration rate of renewable energy sources such as wind power and photovoltaics in the power system continues to increase. However, the intermittent and volatile characteristics of renewable energy pose serious challenges to the stability and supply-demand balance of the power system. Electrochemical energy storage systems, as a core technology for smoothing power fluctuations and realizing the spatial and temporal transfer of energy, have become a key infrastructure supporting the construction of new power systems. Among various energy storage technologies, lithium-ion batteries for energy storage, with their advantages of high energy density, excellent charge and discharge performance, and long cycle life, are widely used in large-scale energy storage power stations, new energy vehicles, distributed energy systems, and other scenarios, becoming the mainstream technology choice in the current energy storage field.

[0003] Battery capacity is a core indicator for measuring the State of Health (SOH) of lithium-ion batteries, directly determining the energy management, operational safety, and service life of energy storage systems. Accurately predicting the capacity degradation trend of lithium-ion batteries not only provides a basis for optimizing charging and discharging strategies and formulating operation and maintenance plans for energy storage systems, but also effectively prevents sudden safety accidents such as thermal runaway and sudden capacity drops caused by battery performance degradation. This is of great significance for ensuring the safe and efficient operation of energy storage systems throughout their entire lifecycle. Therefore, lithium-ion battery capacity prediction technology has become one of the core directions in the research and development of battery management systems (BMS) for energy storage systems, and is also a research hotspot in both academia and industry.

[0004] Currently, lithium-ion battery capacity prediction methods are mainly divided into three categories: mechanistic modeling methods, data-driven methods, and hybrid modeling methods. Although each method has its own advantages and a certain theoretical research foundation, they all have significant limitations and are difficult to meet the requirements of energy storage systems in complex environments in actual engineering for prediction accuracy, robustness, and engineering applicability.

[0005] Mechanistic modeling methods are based on the internal physicochemical processes of batteries, describing the capacity decay mechanism by constructing equivalent circuit models and electrochemical models. Some studies improve the accuracy of parameter extraction by refining the equivalent circuit model, or extract characteristic parameters based on the Randle equivalent circuit model combined with electrochemical impedance spectroscopy (EIS) to achieve state of equilibrium (SOH) identification. The advantage of this type of method is its strong theoretical interpretability and ability to reflect the essential laws of battery decay. However, it suffers from drawbacks such as the large scale of lithium-ion batteries used in energy storage, complex operating environments including temperature fluctuations, changes in charge and discharge rates, and frequent switching of operating conditions. This leads to cumbersome parameter identification processes and high model complexity. At the same time, simplified models cannot characterize the complex decay process involving the coupling of multiple physical fields (electric, thermal, and chemical), and cannot adapt to the dynamic changes of actual operating conditions, making them prone to prediction bias in engineering applications.

[0006] Data-driven methods characterize the relationship between battery operating data such as voltage, current, temperature, and cycle count and battery capacity degradation, using machine learning or deep learning algorithms to build predictive models without requiring in-depth understanding of the internal mechanisms. Common methods include Support Vector Machines (SVM), Random Forests (RF), Extreme Learning Machines (ELM), and traditional neural networks. Current research extracts health factors based on partial charge-discharge data and combines principal component analysis and improved Gaussian process regression to achieve SOH estimation. Alternatively, random charge-discharge segments and sparse Gaussian process regression are used to adapt to dynamic operating conditions. These methods have two main problems: First, they require extremely high completeness of health factors, necessitating manual selection of feature parameters that effectively characterize degradation. However, features extracted from one-dimensional time series data are insufficient to fully characterize the degradation behavior of the battery's multi-physics coupling, easily losing key degradation information. Second, under dynamic operating conditions with temperature changes, they rely on open-circuit voltage model filtering. The generalization ability under extreme conditions such as low temperature and high rate is not fully verified, and there is a contradiction between the computational complexity and real-time performance of the model, making it difficult to meet the online prediction requirements of large-scale energy storage systems.

[0007] Hybrid modeling combines the advantages of mechanistic and data-driven models, improving prediction performance through model fusion. However, it still fails to address the core issue of insufficient one-dimensional feature representation. Current mainstream research relies heavily on one-dimensional time series data to mine battery degradation-related features. However, lithium-ion battery capacity degradation involves multiple coupled mechanisms within the battery, such as active material loss, electrolyte decomposition, SEI film growth, and lithium dendrite precipitation. One-dimensional features cannot fully capture these complex dynamic characteristics, resulting in insufficient accuracy in modeling the degradation trend. This is especially true when capacity degradation falls below 80% in the later stages of long-term cycling, where prediction errors increase significantly. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a dynamic prediction method for the energy storage capacity of lithium-ion batteries, which addresses the above-mentioned problems in the prior art. By converting the battery differential voltage IC (dV / dQ) curve, which sensitively reflects the internal electrochemical changes of the battery, into a recursive graph, a dedicated convolutional neural network is used to extract multi-level battery features and realize capacity regression prediction. This method has advantages in prediction accuracy, robustness, and engineering applicability, providing a new approach for the health management of lithium-ion batteries.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for dynamically predicting the energy storage capacity of lithium-ion batteries includes the following steps: Cyclic aging data of different lithium-ion batteries were obtained and preprocessed to obtain the preprocessed voltage-capacity curves; The dV / dQ curve is obtained by numerically differentiating the preprocessed voltage-capacity curve. The recursion graph parameters are determined based on the numerical sequence of the dV / dQ curves, and the corresponding battery characteristic recursion graph is generated to obtain the battery characteristic recursion graph dataset. The battery characteristic recursion graphs in the battery characteristic recursion graph dataset and their corresponding capacity values ​​are used to form a sample to train a battery capacity correlation prediction model. The system generates a recursive graph of the battery characteristics of the lithium-ion battery to be predicted in real time and inputs it into the trained battery capacity correlation prediction model to obtain the corresponding capacity prediction value. The error between the capacity prediction value and the actual capacity value is compared. If the error exceeds the threshold, the recursive graph of the battery characteristics of the lithium-ion battery to be predicted is added to the battery characteristic recursive graph dataset. The battery characteristic recursive graph in the battery characteristic recursive graph dataset and the corresponding capacity value are used to form a sample pair to retrain the battery capacity correlation prediction model. If the error does not exceed the threshold, the capacity prediction value is output as the capacity prediction result.

[0010] Furthermore, the cyclic aging data includes voltage, current, and capacity for each cycle. During data preprocessing, the data from the constant voltage charging stage is first removed and the data from the constant current charging stage is retained. Then, the original voltage-capacity data is processed by moving average filtering to eliminate measurement noise. Finally, the voltage-capacity data is unified to the same measurement step size through linear interpolation.

[0011] Furthermore, when determining the recurrence graph parameters based on the numerical sequence of the dV / dQ curves, the following steps are included: The embedding dimension m is selected based on the length N of the dV / dQ sequence. When N is less than the first threshold, m is taken from the first interval. When N is greater than or equal to the first threshold, m is taken from the second interval. The autocorrelation function of the dV / dQ sequence is calculated, and the time interval at which the autocorrelation function first drops to 1 / e is taken as the delayed sampling time interval τ. Calculate the state vector distance matrix D of the dV / dQ sequence, and take the 5%~10% quantile of D as the distance threshold ε.

[0012] Furthermore, the first threshold is 1000, the first interval is 2~3, and the second interval is 4~5.

[0013] Furthermore, the battery capacity correlation prediction model adopts a convolutional neural network model including an input layer, a feature learning layer group, a spatial information aggregation layer, a global feature aggregation layer, a fully connected layer, and an output layer.

[0014] Furthermore, the feature learning layer group uses three convolutional layers to progressively extract battery aging features at different levels of abstraction. The first layer unit captures the distribution of local recursive points and micro-texture structures in the recursive graph, corresponding to a transient peak in the battery's dV / dQ curve. The second layer unit integrates the bottom-level features to form medium-scale features related to the curve smoothing trend caused by the growth of the SEI film and the intensification of electrode polarization in the mid-term of the battery. The third layer unit performs a higher level of abstraction to extract features that are strongly correlated with the overall capacity decay trajectory and health status of the battery.

[0015] Furthermore, the parameters of the convolutional neural network model are set as follows: 3×3 convolutional kernel size to balance feature capture accuracy and computational efficiency; 3 convolutional layers, with the first layer outputting 32 feature maps, the second layer outputting 64 feature maps, and the third layer outputting 128 feature maps; pooling window size: 2×2; 1 fully connected layer, with the number of neurons determined to be 64 through cross-validation to avoid overfitting due to too many parameters.

[0016] Furthermore, the loss function formula for the convolutional neural network model is as follows:

[0017] in, It is the mean square error function between the predicted and actual battery capacity values. It is a smoothed L1 piecewise loss function between the predicted and actual battery capacity values. These are the weighting coefficients.

[0018] Furthermore, after training the battery capacity correlation prediction model by combining the battery characteristic recursion graphs and corresponding capacity values ​​in the battery characteristic recursion graph dataset, the process also includes: performing basic validation of the battery capacity correlation prediction model; if the error exceeds the standard, adjusting the recursion graph parameters or convolutional neural network model parameters according to the actual engineering situation, and retraining the battery capacity correlation prediction model to ensure engineering applicability.

[0019] Compared with the prior art, the advantages of the present invention are as follows: The technical solution of this invention uses innovative technologies such as one-dimensional feature visualization, two-dimensional feature extraction, and full-process regression prediction as its logic. It transforms the operating data of lithium-ion batteries for energy storage into feature images, and uses innovative battery feature extraction and capacity prediction models to predict the degradation of lithium-ion battery health status. Compared with traditional methods, this invention addresses existing technical difficulties with the following innovations: It innovatively converts the battery dV / dQ curve into an RP image, achieving two-dimensional visualization of a one-dimensional time series, fully preserving the dynamic characteristics of lithium-ion battery degradation trends, and avoiding the limitations of manual feature selection; the customized battery capacity correlation prediction model can automatically extract multi-level features from the RP image, from image feature recognition to prediction of battery health status degradation, improving prediction efficiency without manual intervention, and providing a more reliable and intelligent technical solution for battery health management in energy storage systems. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall method framework of the present invention.

[0021] Figure 2 This is a schematic diagram of the battery's dV / dQ recursive characteristic degradation process.

[0022] Figure 3 This is the architecture of the battery capacity characteristic correlation prediction model of the present invention.

[0023] Figure 4 This is a detailed flowchart of the capacity prediction method according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0025] Recurrence plots (RPs), as a time series visualization tool, can convert one-dimensional time series into two-dimensional images, effectively revealing the periodicity, chaos, and non-stationarity of the series, without requiring manual feature selection. Convolutional Neural Networks (CNNs) possess powerful two-dimensional image feature extraction capabilities, automatically mining multi-level abstract features in images and enabling end-to-end training. Based on this, this embodiment proposes a dynamic prediction method for lithium-ion battery energy storage capacity. This method combines battery feature recurrence plots and a battery capacity correlation prediction model based on convolutional neural networks to predict the capacity of lithium-ion batteries used in energy storage. By converting the battery differential voltage IC (dV / dQ) curve, which sensitively reflects the internal electrochemical changes of the battery, into a recurrence plot, a dedicated convolutional neural network is used to extract multi-level battery features and achieve capacity regression prediction. This method achieves "multi-dimensional feature fusion - complex dynamic capture - efficient prediction throughout the entire process." The advantages of this method in prediction accuracy, robustness, and engineering applicability provide a new approach for lithium-ion battery health management. The core innovation of this method lies in converting a one-dimensional dV / dQ time series into a two-dimensional RP image. Utilizing the powerful image extraction capabilities of CNNs, it represents the battery capacity degradation characteristics, avoiding the limitations of manual feature selection. Furthermore, high-precision model construction and multi-scenario validation ensure the method's effectiveness and robustness. The core content includes: First, a battery data preprocessing and feature visualization method, which extracts the differential voltage (dV / dQ) curve during the constant current charge-discharge cycle of the lithium-ion battery. Based on the embedding dimension m and delay time τ, a state vector is constructed. A recursive graph is generated by calculating the Euclidean distance matrix and the recursive matrix, converting the one-dimensional dV / dQ time series into a two-dimensional feature image, thus visually representing the battery degradation process. Second, a battery capacity correlation prediction model is constructed, using three layers of convolution to progressively extract battery feature information from the recursive graph, combined with global average pooling to compress spatial information. Finally, a fully connected layer achieves accurate battery capacity regression, enabling the prediction of lithium-ion battery capacity changes to determine battery health status, thereby achieving accurate diagnosis of various types of faults in the energy storage system and ensuring its safe operation.

[0026] like Figure 1 As shown, this method consists of five core stages: data acquisition and preprocessing, dV / dQ curve extraction, recursive graph (RP) generation, battery capacity correlation prediction model training, and capacity prediction and validation. The specific logic is as follows: (1) Data acquisition: Acquire charging and discharging data such as voltage, current, capacity, and temperature during the cyclic aging process of lithium-ion batteries, and integrate the dataset with publicly available lithium battery aging datasets and battery charge and discharge cycle test data. (2) Data preprocessing: The original charge and discharge data are filtered to remove noise, interpolated to unify the sampling step size, and the voltage-capacity curves of the charge and discharge stages are extracted. The differential voltage dV / dQ curves that can reflect the changes in the internal electrochemical reaction of the lithium-ion battery are calculated. (3) RP generation: The one-dimensional dV / dQ time series is converted into a two-dimensional recursive graph. By constructing the state vector and calculating the distance matrix and recursive matrix, the capacity decay characteristics of lithium-ion batteries can be visualized. (4) Battery capacity correlation prediction model construction: Using the generated RP image as input, a customized battery capacity correlation prediction model is constructed. By dividing the training set / test set into different data ratios, the model is trained and the mapping relationship between the RP image and the capacity is learned to obtain a convolutional neural network suitable for the RP image of the battery characteristics to be tested. (5) Model training and prediction: The battery capacity correlation prediction model is trained using RP images, and the performance of the model is evaluated using MAE, RMSE and MAPE indicators. Through parameter sensitivity analysis and method model optimization, the accuracy and robustness of the method are ensured, and the battery capacity prediction results are finally output.

[0027] The key points are explained below.

[0028] 1. Battery aging feature extraction Extracting features reflecting the degree of battery aging is a key step in battery capacity prediction. This embodiment selects the differential voltage dV / dQ curve as a carrier of battery health status to represent the capacity degradation process. Since the dV / dQ curve is derived from the voltage-capacity curve, it can sensitively reflect a series of internal electrochemical reactions during the charging and discharging process of lithium-ion batteries, such as the potential plateau of positive electrode delithiation and negative electrode lithium insertion, and the polarization voltage changes caused by SEI film growth. It is an effective indicator for characterizing lithium-ion battery capacity decay and is superior to traditional one-dimensional features such as voltage and current. This invention uses a publicly available lithium battery aging dataset as the research object and can select batteries of different rates to extract the battery aging feature dV / dQ curve through the following steps: 1.1 Data preprocessing, including: Data filtering: Extract cycle aging data for each battery from the dataset, including voltage (V) and capacity (mAh) data during the constant current charge and discharge phase, and remove the constant voltage charging phase which is not sensitive to degradation. Noise Removal: The raw voltage-capacity data is processed using a moving average filter to remove measurement noise caused by voltage fluctuations due to errors. Interpolation standardization: Since the sampling step size for charging and discharging may differ for different cycle numbers, linear interpolation is used to standardize the voltage-capacity data to the same capacity interval, ensuring the consistency of dV / dQ calculation.

[0029] 1.2 dV / dQ curve calculation: Numerical differentiation is performed on the preprocessed voltage-capacity curve (VQ curve) to obtain the dV / dQ curve. Let the capacity sequence be... The corresponding voltage sequence is Then the formula for calculating dV / dQ is: (1) in, (Exclude the beginning and end points to ensure the accuracy of derivative calculation).

[0030] 1.3 Feature Analysis: Based on the dV / dQ curve obtained by the method, the key characteristics corresponding to battery capacity degradation are: Capacity decay pattern: The capacity of all batteries decreases with increasing cycle number. Capacity decay is faster in the early stages of the active material adaptation phase, then gradually levels off and reaches a stable decay stage. A cross-sectional comparison of battery packs shows that batteries with higher rate operation have a slightly faster decay rate than those with lower rate operation, consistent with the pattern of accelerated aging at higher rates. Changes in the dV / dQ curve: such as Figure 2 As shown, with increasing cycle number, the peak height of the dV / dQ curve decreases, the peak position shifts towards higher capacity, and the curve smoothness increases, enhancing the polarization effect. The dV / dQ curve of a lithium-ion battery exhibits a significant peak in the early stages of cycling, which disappears as cycling progresses, reflecting a decrease in electrochemical reaction intensity due to the loss of active material. These characteristics, after RP conversion, form a visualized battery capacity decay pattern, providing rich input features for the CNN model and ensuring that the model can capture the subtle dynamics of battery capacity decay to reflect the battery's health status.

[0031] 2. Method for generating dV / dQ characteristic recursive graph (RP) of lithium-ion batteries Recursive graphs (RPs) are visualization tools that convert one-dimensional time series into two-dimensional images. They effectively reveal the periodicity, non-stationarity, and unpredictability of time series, and are particularly suitable for feature extraction of short time series. Based on the nonlinear dynamic evolution characteristics, multi-timescale response features, and electrochemical impedance spectroscopy migration patterns exhibited by the differential voltage curve of lithium-ion batteries during capacity decay, this embodiment proposes a dedicated recursive graph generation method for lithium-ion battery capacity prediction. This method achieves accurate visualization of the dynamic characteristics of capacity decay in the battery's dV / dQ curve by constructing a multi-dimensional phase space reconstruction framework, an adaptive parameter optimization mechanism, and a decay-sensitive threshold selection strategy. A strong correlation is established between the distribution pattern of recursive points in the image and the battery's health state, thus providing highly discriminative visual features for subsequent capacity prediction.

[0032] 2.1 Basic Definitions and Principles: Let the one-dimensional dV / dQ time series characterizing the battery aging state be: (N is the sequence length, i.e., the number of sampling points in the charging and discharging phase). The core of RP generation is to visualize the recursive characteristics of the sequence by comparing the proximity of state vectors in the reconstructed phase space. If the distance between two state vectors in the phase space is less than a threshold ε, it is determined to be recursive. In a two-dimensional image, points are marked as colored dots if they are not colored, and white dots otherwise. The distribution of recursive points visually reflects the changing patterns of the battery's dV / dQ sequence, thus relating it to the capacity decay of the lithium-ion battery under test. The overall distribution pattern of the recursive points directly reflects the stability and evolution of the battery's aging trajectory.

[0033] 2.2 Adaptive determination of recursive graph parameters for adapting to battery sequence characteristics: The selection of parameters (embedding dimension m, delay time τ, distance threshold ε) in standard RP is universal. This invention, to accurately characterize the unique dynamics of the battery dV / dQ curve, adapts the method for determining these parameters to battery aging analysis. Adaptive selection of embedding dimension m: The embedding dimension m determines the dimension of the reconstructed state vector, affecting the depth of characterization of sequence dynamics. To match the changing complexity of the battery dV / dQ curve morphology at different aging stages, this invention dynamically selects m based on the sequence length N: when N < 1000 (usually corresponding to stages with less data or simpler curve changes), a value is taken from the range of 2 to 3 to avoid overparameterization; when N ≥ 1000, a value is taken from the range of 4 to 5 to fully describe the rich dynamics inherent in the complex curve. This selection strategy aims to match the reconstructed phase space dimension with the actual information capacity of the battery data.

[0034] Optimization of the delay time τ: The delay time τ affects the independence between the components of the state vector. To extract the minimum time interval that can effectively separate the continuous electrochemical states in the battery dV / dQ sequence, this invention employs the autocorrelation function method: calculating the autocorrelation function of the sequence and taking the delay corresponding to its first decrease to the initial value 1 / e as τ. The τ determined by this method can effectively capture the characteristic time scale of related processes within the battery.

[0035] Battery data-driven determination of the distance threshold ε: The distance threshold ε determines the strictness of the "recursion" judgment and directly affects the density of the RP image. To eliminate the influence of differences in the overall amplitude of the curves between different batteries or different cycling stages, and to ensure that the generated RP has a stable and interpretable texture structure, this invention proposes a method for determining ε based on data distribution statistics: First, the Euclidean distance matrix D between all state vectors is calculated, and then the 5% to 10% quantile of all elements in D is taken as ε. This strategy utilizes the distribution characteristics of the battery data itself to set the threshold, so that the generated recursive graph can highlight the state similarity patterns related to aging.

[0036] 2.3 Adaptive RP generation mechanism based on decay characteristics: Based on the adaptively determined parameters mentioned above, the generation of battery characteristics RP follows the core formulas and steps below: 1) Constructing the state vector: To embed a one-dimensional time series into a high-dimensional phase space, it is necessary to define the embedding phase space dimension m and the delayed sampling time interval τ, and construct the state vector. : (2) in, K represents the number of state vectors, i.e., the size of the RP image is K×K. Embedding dimension m: Based on the short sequence characteristics of existing public datasets, m=3 (empirical range is 2-5). This needs to be adjusted according to the sequence length; a smaller value is used for shorter sequences to avoid too few state vectors; a larger value can be used for longer sequences to improve feature representation accuracy. Delay time τ: Calculated using the autocorrelation function method, i.e., finding the first time the autocorrelation function decreases to its maximum value. The time interval is set to τ=1s in this method to ensure that the state vector can reflect the dynamic changes of the sequence and to avoid information redundancy or loss.

[0037] 2) Calculate the distance matrix: The distance matrix D is used to measure the similarity between all state vector pairs in the phase space. It employs the Euclidean distance, which is computationally simple and effectively characterizes vector differences. State vector and Euclidean distance: (3) in, State vector The k-th component, i,j=1,2,...,K. The distance matrix D has a size of K×K, and the smaller the element value, the more similar the two state vectors are.

[0038] 3) Generating the recursive matrix: The recursive matrix R is the core of RP. By setting a distance threshold ε, the distance matrix D is transformed into a binary matrix. (4) in: The unit step Heaviside function is defined as follows: ; ε is the distance threshold used to determine the criterion for recursion, and it is determined using the recursion density method. A specific quantile of all elements in the distance matrix D is selected as ε to ensure the recursion points in the RP image. The density should be moderate, avoiding being too sparse or too dense, which would affect feature extraction; : Represents the state vector and Similarity within the ε-neighborhood corresponds to recursive points in the RP image; : Indicates that the two state vectors are not similar, corresponding to a non-recursive point in the RP image.

[0039] Finally, the recurrence matrix R is visualized as a K×K two-dimensional RP image. From the generated recurrence graph, it can be intuitively observed that when the battery has few initial cycle times, the yellow recurrence region corresponds to the maximum capacity; as the number of cycle times increases, the yellow region gradually shrinks, and the capacity decreases, with the region's position shifting from the lower right corner to the upper left corner, reflecting the dynamic changes in the dV / dQ curve. This provides a clear visual basis for subsequent CNN extraction of battery degradation features. Through the above-mentioned adaptive parameter selection and construction process that closely integrates the statistical characteristics of battery data with the needs of aging analysis, this invention achieves a proprietary transformation from a one-dimensional electrochemical feature sequence of the battery to a two-dimensional recurrence feature map rich in state information, laying the foundation for subsequent deep feature extraction.

[0040] 3. Battery capacity characteristic correlation prediction model CNN is a deep learning model specifically designed for processing grid-structured image data. Its advantage lies in extracting local image features through convolution and reducing model complexity through pooling. In this embodiment, a battery capacity characteristic correlation prediction model is formed through structural customization for visualizing battery aging features. Based on its characteristics, multi-level features of lithium-ion battery degradation can be extracted from the recurrence graph (RP) image. This model not only possesses the inherent local perception, weight sharing, and hierarchical feature extraction capabilities of CNN, but more importantly, its network structure, inter-layer connections, and loss function are specifically optimized for the degradation dynamics inherent in the lithium-ion battery recurrence graph (RP), achieving high-precision end-to-end regression prediction from battery feature images to capacity values.

[0041] 3.1 Battery characteristic adaptation design of the model architecture: This model employs a hierarchical, progressive feature extraction mechanism, and its architecture is as follows: Figure 3 As shown, the model comprises an input layer, a group of feature learning layers, a spatial information aggregation layer, a global feature aggregation layer, a fully connected layer, and an output layer. Each layer is designed to consider the typical change patterns of battery RP images during the aging process, ensuring that the model can extract visual features strongly correlated with capacity decay from the images. The functions of each layer are as follows: 1) Input Layer: Receives two-dimensional image data generated by the recursive graph (RP), with an image size of K×K. The input layer only transmits data and does not perform calculations, ensuring that the original features of the RP image representing the dynamics of battery degradation are fully input into the model.

[0042] Where K is the number of state vectors, determined by the original time series length, the embedding dimension m, and the delay time τ. 2) Feature Learning Layers (Convolutional Layers): This group of layers is the core of the model, specifically designed to automatically and layer-by-layer capture local and global spatial patterns associated with aging states in the battery recurrence relation (RP) image. Its core function is to capture local features in the battery capacity characteristic RP image, collecting changes in the shape, size, and location of decay regions to map the capacity decay pattern of lithium-ion batteries.

[0043] The feature learning layer group uses three convolutional layers to progressively extract battery aging features at different levels of abstraction. The design of each layer maps to the physical process of battery degradation. The first layer unit (outputs 32 feature maps): its learning objective is to identify the basic micro-elements in the recursive graph, capture the local recursive point distribution and micro-texture structure in the RP image, and reflect a certain transient peak in the battery dV / dQ curve.

[0044] The second-layer unit (outputting 64 feature maps) fuses the features from the lower-level layer to form mid-scale features, such as recursive regions of specific shapes, directional textures, and boundary clarity. These features are related to the curve smoothing trend caused by mid-term SEI film growth and increased electrode polarization.

[0045] The third-level unit (outputting 128 feature maps): performs a higher level of abstraction. The extracted features can characterize the overall structural evolution trend of the recursion graph, reflecting the systematic contraction of the recursion region along the diagonal direction and the overall decrease in recursion density within a specific quadrant. These high-level features are considered to have a strong correlation with the overall capacity decay trajectory and state of health (SOH) of the battery.

[0046] Convolutional layers generate feature maps by performing a sliding window calculation between the convolutional kernel and the input image, as shown in the following formula: Let the convolution kernel be The input feature map is The output of the convolution operation is ,but: (5) Where b is the bias term, and (i,j) are the pixel coordinates of the feature map. This invention employs three convolutional layers to progressively improve the level of feature abstraction. The first convolutional layer captures primary features such as edges and textures of the RP image; the second convolutional layer fuses primary features to form intermediate features such as attenuation region boundaries; the third convolutional layer further abstracts to form high-level features directly related to capacity attenuation, namely the color gamut reduction rate.

[0047] 3) Spatial Information Aggregation Layer (Pooling Layer): Interspersed between feature learning layers, this layer reduces the dimensionality of the feature maps, decreases the number of model parameters, avoids overfitting, and enhances the model's stability against small image displacements. This invention supports both max pooling and average pooling; specifically, average pooling is used, as shown in the following formula: (6) Where V is the output value of the pooling layer. Within the pooled window (size) The input feature values, This represents the size of the feature map in the last convolutional layer. Pooling layers and convolutional layers are alternated; with each pooling operation, the feature map size shrinks, and the pooling window size is halved, improving the abstractness of battery capacity features. The purpose of this layer is to progressively compress the spatial size of the data while preserving the key structural information of the battery RP, enhancing the model's robustness to small displacements or deformations in the input image, effectively controlling the overall number of model parameters, and preventing overfitting to a limited battery aging dataset.

[0048] 4) Global Feature Convergence Layer (Global Average Pooling Layer): Located after the last pooling layer, this layer averages all pixel values ​​from each feature map to obtain a one-dimensional feature vector with dimensions equal to the number of feature maps. The purpose of this layer is to compress spatial information, avoiding the surge in model complexity caused by fully connected layers directly processing high-dimensional feature maps, while simultaneously enhancing the model's generalization ability and reducing the risk of overfitting. This layer outputs a one-dimensional vector with dimensions equal to the final number of feature maps. This vector fuses all spatial information related to battery aging in the RP image, providing a compact high-dimensional feature representation for subsequent regression prediction.

[0049] 5) Global Feature Mapping Layer (Fully Connected Layer): This layer maps the one-dimensional feature vector output from the global average pooling layer to the predicted capacity value. The neurons in the fully connected layer are connected to all neurons in the previous layer. Through weighted summation and activation functions, a non-linear transformation is achieved, ultimately outputting the predicted lithium-ion battery capacity through the output layer of a single neuron. This layer is responsible for non-linearly fusing and interacting with all the scattered visual features related to battery capacity extracted from the front end, ultimately mapping them to a continuous scalar output value, i.e., the predicted battery capacity.

[0050] 6): Output layer: A linear activation function is used to output continuous capacity values, thereby achieving capacity regression prediction.

[0051] Based on the feature complexity and prediction requirements of lithium-ion battery RP images, the CNN model parameters of this invention are set as follows: a 3×3 convolutional kernel size to balance feature capture accuracy and computational efficiency; three convolutional layers, with the first layer outputting 32 feature maps, the second layer outputting 64 feature maps, and the third layer outputting 128 feature maps; a pooling window size of 2×2; and a single fully connected layer with 64 neurons determined through cross-validation to avoid overfitting due to excessive parameters.

[0052] 3.2 Loss function design for battery capacity regression: The loss function is the core of guiding model training and optimizing prediction accuracy. Considering the characteristics of regression tasks in battery capacity prediction and the requirements for prediction bias stability in practical engineering, this invention designs a composite weighted loss function to simultaneously optimize the absolute accuracy of battery capacity predictions and the consistency of error distribution.

[0053] The loss function is a weighted sum of mean squared error (MSE) loss and smoothing L1 loss, and its specific form is as follows: The formula for the total loss function is as follows: (7) The mean squared error function calculates the average of the squares of the differences between predicted and true values, emphasizing a strong penalty for large errors and prompting the model to focus on prediction accuracy during periods of rapid capacity reduction. The formula is as follows: (8) in, n This represents the number of samples contained in a training batch. Each sample is a data pair consisting of a "battery characteristic recursion graph - corresponding actual capacity value". For the first i The actual battery capacity value of each sample, that is, the actual capacity value obtained through measurement or calibration, is the target of model learning. For the battery capacity correlation prediction model for the first i The output capacity prediction value is calculated by forward propagation of the recursion graph of each sample.

[0054] The smoothed L1 piecewise loss function is insensitive to small errors but linearly penalizes large errors. That is, when... When the absolute error is less than 1, its behavior is similar to MSE (providing smooth gradients); when the absolute error is greater than or equal to 1, its behavior is similar to L1 loss (less sensitive to outliers or large errors than MSE). This helps improve the training stability of the model during the phase of gradual capacity decay, avoiding convergence difficulties due to sudden gradient changes. The formula is as follows: (9) The weighting coefficients are determined by adjusting the validation set (usually set to 0.6~0.7) to balance the contributions of the two types of losses, so that the model achieves the optimal balance between overall accuracy and error smoothness.

[0055] The design of this composite loss function fully considers the nonlinear characteristics of capacity decay in the lithium-ion battery under test. In the early and middle stages of decay, capacity changes are relatively gradual, and the SmoothL1 loss helps stabilize training. In the later stages of decay, capacity may decrease rapidly, and the MSE loss enhances the model's fitting ability for this stage. Compared with traditional single loss functions, this design is better suited to the accuracy requirements of predicting battery degradation throughout its entire lifespan.

[0056] Based on the above key content, such as Figure 4 As shown, the dynamic prediction method for lithium-ion battery energy storage capacity in this embodiment includes the following steps: S1) Data Acquisition: Obtain cycle aging data for different lithium-ion batteries; Specifically, the cycle aging data is obtained by independently collecting battery data based on existing publicly available lithium-ion battery aging datasets or battery cycle testing platforms, specifying the voltage, current, and capacity for each cycle. Experimental parameters such as ambient temperature, charge / discharge rate, number of cycles, and sampling step size are set during this process. The data includes voltage (V), current (A), capacity (mAh), and number of cycles for each cycle, with a focus on retaining data from the constant current charging phase.

[0057] S2) Data preprocessing: Perform data preprocessing to obtain the preprocessed voltage-capacity curve; Specifically, during data preprocessing, the collected data is first filtered to remove data from the constant voltage charging stage and retain data from the constant current charging stage. Then, the original voltage-capacity data is processed using a moving average filter to eliminate measurement noise. Finally, the voltage-capacity data is unified to the same measurement step size through linear interpolation to ensure the consistency of subsequent dV / dQ calculations.

[0058] S3) Battery dV / dQ curve extraction: Numerical differentiation is performed on the preprocessed voltage-capacity curve to obtain the dV / dQ curve; Specifically, using formula (1), the dV / dQ curve is obtained by differentiating the preprocessed voltage-capacity curve using the central difference method. The obtained curve is then verified to check for any abnormal peaks in the dV / dQ curve. If any are found, the curve needs to be re-filtered to ensure that it reflects the electrochemical changes of the battery.

[0059] S4) Recursive Graph (RP) Generation: Determine the recursive graph parameters based on the numerical sequence of the dV / dQ curves and generate the corresponding battery characteristic recursive graph to obtain the battery characteristic recursive graph dataset; Specifically, when determining the parameters of the recurrence graph based on the numerical sequence of the dV / dQ curve, the following steps are taken: The embedding dimension m is selected based on the length N of the dV / dQ sequence. When N is less than the first threshold, m is taken from the first interval. When N is greater than or equal to the first threshold, m is taken from the second interval. The first threshold is 1000, the first interval is 2~3, and the second interval is 4~5. Calculate the autocorrelation function for the dV / dQ sequence, and take the time interval when the autocorrelation function first drops to 1 / e as the delayed sampling time interval τ, which is usually τ=1~5; After determining the embedding dimension m and the late sampling time interval τ, according to Determine the number of state vectors, and then construct the state vectors of the dV / dQ sequence according to formula (2) to reconstruct the original one-dimensional time series. Then calculate the Euclidean distance between these state vectors according to formula (3) to form the state vector distance matrix D of the dV / dQ sequence. Take the 5%~10% quantile of D as the distance threshold ε.

[0060] When generating the corresponding battery characteristic recurrence graph, the distance matrix D is binarized by setting a threshold ε to generate a recurrence matrix, where an element value of 1 indicates that the corresponding recurrence points are similar to each other, and 0 indicates that they are dissimilar. Finally, the recurrence matrix is ​​visualized as a two-dimensional image, usually distinguishing recurrence points from non-recurrence points with different colors, thereby intuitively displaying the dynamic characteristics of the time series.

[0061] S5) Battery capacity correlation prediction model construction and training: The battery characteristic recursion graph in the battery characteristic recursion graph dataset and the corresponding capacity value are used to form a sample to train the battery capacity correlation prediction model. Specifically, battery characteristic RP images and corresponding capacity values ​​for each cycle number are paired as sample sets, and the training and test sets are divided proportionally to ensure the randomness of the sample data partitioning. A corresponding CNN model is constructed, including an input layer that receives RP images, convolutional layers, pooling layers, global average pooling layers, fully connected layers, and an output layer, finally outputting the final battery capacity prediction result. The model is trained using an early stopping mechanism; training stops if the validation set loss does not decrease for 10 consecutive cycles to prevent overfitting.

[0062] S6) Model Validation and Optimization: Perform basic validation of the battery capacity correlation prediction model. If the error exceeds the standard, adjust the recursive graph parameters or convolutional neural network model parameters according to the actual engineering situation, and retrain the battery capacity correlation prediction model to ensure engineering applicability. In the specific basic validation, the MAE, RMSE, and MAPE of the training set and the test set are calculated to verify and ensure the accuracy of the model and the superiority of RP-CNN. If the error exceeds the standard, the RP parameters (m, τ, ε) or the parameters of the battery capacity correlation prediction model are adjusted according to the actual engineering situation, and the model is retrained to ensure engineering applicability.

[0063] S7) Capacity Prediction: Generate the recursive graph of the battery characteristics of the lithium-ion battery to be predicted in real time and input it into the trained battery capacity association prediction model to obtain the corresponding capacity prediction value. Compare the error between the capacity prediction value and the actual capacity value. If the error exceeds the threshold, add the recursive graph of the battery characteristics of the lithium-ion battery to be predicted to the battery characteristic recursive graph dataset, and use the battery characteristic recursive graph in the battery characteristic recursive graph dataset and the corresponding capacity value to form a sample pair to retrain the battery capacity association prediction model. If the error does not exceed the threshold, output the capacity prediction value as the capacity prediction result.

[0064] Specifically, during the operation of the energy storage system, real-time charging voltage-capacity data of the lithium-ion batteries is collected, and steps S2-S4 are repeated to generate real-time RP images. The real-time RP images are input into a trained battery capacity correlation prediction model, which outputs the predicted capacity value for the current cycle number. The predicted capacity value is transmitted to the energy storage system monitoring platform, and the prediction error is calculated and compared with the actual capacity. If the error exceeds a threshold, the model is retrained.

[0065] In summary, this invention proposes a dynamic capacity prediction method for lithium-ion batteries. By fusing and interacting with a battery feature recursive graph and a battery capacity correlation prediction model, the estimation accuracy of lithium-ion battery capacity is improved. Through a process of "data preprocessing - feature visualization - battery capacity correlation prediction model building - capacity prediction," a mapping relationship between battery operating parameters and health status is established, achieving high-precision prediction of lithium-ion battery capacity. This method significantly improves the accuracy of capacity prediction and has good engineering applicability, providing a reliable technical means for online monitoring and lifespan assessment of battery health status in energy storage systems. Specific technical effects are as follows: High prediction accuracy: The method's error metrics are extremely low on publicly available battery aging datasets. When the training data ratio increases to 70%, both the battery MAE and MAPE are low, and the predicted curves highly coincide with the actual capacity curves. This accurately depicts the degradation trend of batteries from their new state to the later stages of aging, solving the problem of insufficient prediction accuracy for long-term degradation by traditional data-driven methods.

[0066] Strong robustness: Robustness is verified through method comparison and parameter sensitivity analysis. Compared with RP-SVM and 1D-CNN models, the proposed model shows significantly lower MAE, RMSE, and MAPE, smaller and more concentrated errors with no outliers, and strong adaptability to batteries with different rate of change and decay characteristics. When the key RP parameters (m, τ, ε) are within a reasonable range, the prediction error fluctuation is <15%, exhibiting low parameter sensitivity and avoiding prediction failure due to small parameter changes.

[0067] Excellent engineering applicability: The method performs outstandingly in real-world scenarios. The predicted maximum values ​​of MAE, RMSE, and MAPE for the battery consistently remain at low levels. The model training time is short, meeting the online prediction time requirements and satisfying the real-time requirements of the energy storage system's BMS with its second-level sampling cycle. Furthermore, the method requires no complex mechanistic analysis or only conventional voltage and current sensor detection equipment, allowing for direct integration into existing energy storage systems and reducing engineering application costs.

[0068] Significant Value in Health Management: The method can output real-time battery capacity predictions, providing crucial support for lithium-ion battery health management. By predicting capacity degradation trends, battery replacement plans can be developed in advance based on battery capacity ratio standards, avoiding insufficient energy storage system capacity due to battery aging. Simultaneously, combined with prediction error analysis, abnormally degraded batteries can be identified, preventing safety accidents such as thermal runaway, extending the lifespan of energy storage systems, reducing operation and maintenance costs, and providing technical assurance for the safe and efficient operation of energy storage systems in new power systems.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0070] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamically predicting the energy storage capacity of a lithium-ion battery, characterized in that, Includes the following steps: Cyclic aging data of different lithium-ion batteries were obtained and preprocessed to obtain the preprocessed voltage-capacity curves; The dV / dQ curve is obtained by numerically differentiating the preprocessed voltage-capacity curve. The recursion graph parameters are determined based on the numerical sequence of the dV / dQ curves, and the corresponding battery characteristic recursion graph is generated to obtain the battery characteristic recursion graph dataset. The battery characteristic recursion graphs in the battery characteristic recursion graph dataset and their corresponding capacity values ​​are used to form a sample to train a battery capacity correlation prediction model. The system generates a recursive graph of the battery characteristics of the lithium-ion battery to be predicted in real time and inputs it into the trained battery capacity correlation prediction model to obtain the corresponding capacity prediction value. The error between the capacity prediction value and the actual capacity value is compared. If the error exceeds the threshold, the recursive graph of the battery characteristics of the lithium-ion battery to be predicted is added to the battery characteristic recursive graph dataset. The battery characteristic recursive graph in the battery characteristic recursive graph dataset and the corresponding capacity value are used to form a sample pair to retrain the battery capacity correlation prediction model. If the error does not exceed the threshold, the capacity prediction value is output as the capacity prediction result.

2. The method for dynamically predicting the energy storage capacity of lithium-ion batteries according to claim 1, characterized in that, The cyclic aging data includes voltage, current and capacity for each cycle. During data preprocessing, the data from the constant voltage charging stage is first removed and the data from the constant current charging stage is retained. Then, the original voltage-capacity data is processed by moving average filtering to eliminate measurement noise. Finally, the voltage-capacity data is unified to the same measurement step size by linear interpolation.

3. The method for dynamically predicting the energy storage capacity of lithium-ion batteries according to claim 1, characterized in that, When determining the parameters of the recursive graph based on the numerical sequence of the dV / dQ curve, the following are included: The embedding dimension m is selected based on the length N of the dV / dQ sequence. When N is less than the first threshold, m is taken from the first interval. When N is greater than or equal to the first threshold, m is taken from the second interval. The autocorrelation function of the dV / dQ sequence is calculated, and the time interval at which the autocorrelation function first drops to 1 / e is taken as the delayed sampling time interval τ. Calculate the state vector distance matrix D of the dV / dQ sequence, and take the 5%~10% quantile of D as the distance threshold ε.

4. The method for dynamically predicting the energy storage capacity of lithium-ion batteries according to claim 3, characterized in that, The first threshold is 1000, the first interval is 2~3, and the second interval is 4~5.

5. The method for dynamically predicting the energy storage capacity of a lithium-ion battery according to claim 1, characterized in that, The battery capacity correlation prediction model adopts a convolutional neural network model including an input layer, a feature learning layer group, a spatial information aggregation layer, a global feature aggregation layer, a fully connected layer, and an output layer.

6. The method for dynamically predicting the energy storage capacity of a lithium-ion battery according to claim 5, characterized in that, The feature learning layer group uses three convolutional layers to progressively extract battery aging features at different levels of abstraction. The first layer unit captures the distribution of local recursive points and micro-texture structures in the recursion graph, which corresponds to a transient peak in the battery dV / dQ curve. The second layer unit integrates the bottom-level features to form medium-scale features related to the curve smoothing trend caused by the growth of SEI film and the intensification of electrode polarization in the middle stage of the battery. The third-layer unit performs a higher level of abstraction, extracting features that are strongly correlated with the overall capacity decay trajectory and health status of the battery.

7. The method for dynamically predicting the energy storage capacity of a lithium-ion battery according to claim 6, characterized in that, The parameters of the convolutional neural network model are set as follows: 3×3 convolutional kernel size to balance feature capture accuracy and computational efficiency; the first layer unit outputs 32 feature maps, the second layer unit outputs 64 feature maps, and the third layer unit outputs 128 feature maps; pooling window size: 2×2; 1 fully connected layer, and the number of neurons is determined to be 64 through cross-validation to avoid overfitting due to too many parameters.

8. The method for dynamically predicting the energy storage capacity of a lithium-ion battery according to claim 5, characterized in that, The loss function formula for the convolutional neural network model is as follows: in, It is the mean square error function between the predicted and actual battery capacity values. It is a smoothed L1 piecewise loss function between the predicted and actual battery capacity values. These are the weighting coefficients.

9. The method for dynamically predicting the energy storage capacity of a lithium-ion battery according to claim 1, characterized in that, After training the battery capacity correlation prediction model by combining the battery characteristic recursion graphs and corresponding capacity values ​​in the battery characteristic recursion graph dataset, the following steps are also included: performing basic validation of the battery capacity correlation prediction model. If the error exceeds the standard, the recursion graph parameters or convolutional neural network model parameters are adjusted according to the actual engineering situation, and the battery capacity correlation prediction model is retrained to ensure engineering applicability.

Images