Rul estimation method for energy storage battery based on gda-mhda-timesnet model
Patent Information
- Application Number
- CN202611003894.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-07
- Publication Date
- 2026-10-09
AI Technical Summary
具体而言，多数数据驱动方法依赖单一健康指标（如SOH）或直接使用原始充放电时序数据作为输入，未能充分挖掘多维度、多尺度时序特征（如CCCT、CVCT、DVCT等）中所蕴含的退化趋势信息，导致特征表达与寿命衰退过程之间关联性不足，限制了模型对复杂退化模式的建模能力
本发明的有益效果是：为解决现有储能电池RUL估计方法存在的数据样本稀缺、特征单一以及模型泛化能力弱的问题，提出了一种基于GDA-MHDA-TimesNet模型的储能电池RUL估计方法，通过用采用高斯数据增强与多头差异化注意力机制相结合的双策略，从特征维度和时序维度协同增强原始数据，有效提升了样本多样性与模型鲁棒性。同时，模型以能够捕捉多尺度时序依赖的TimesNet为核心架构，输入包括恒流充电时间、恒压充电时间、放电电压变化时间及电池健康状态在内的多维时序特征，以学习电池容量的非线性衰退规律。本发明能够有效提升RUL估计的准确性。实验结果验证了其在复杂工况下的优越性和稳定性，具有较高的估计精度，为电池管理系统的预测性维护提供了有效的技术方案。
Smart Images
Figure CN122883028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery management technology, and specifically to a method for estimating the RUL of energy storage batteries based on the GDA-MHDA-TimesNet model. Background Technology
[0002] In energy storage power stations, the remaining useful life (RUL) of a battery is a key indicator that characterizes the future usable life of a battery, predicts the trend of capacity decay and the dynamics of performance degradation, and is also an important basis for realizing predictive maintenance and optimizing operation and maintenance strategies for energy storage systems.
[0003] Current mainstream methods for predicting the remaining useful life (RUL) of energy storage batteries still have significant shortcomings in feature construction and model building. Specifically, most data-driven methods rely on a single health indicator (such as SOH) or directly use raw charge-discharge time-series data as input, failing to fully explore the degradation trend information contained in multi-dimensional and multi-scale time-series features (such as CCCT, CVCT, DVCT, etc.). This results in insufficient correlation between feature representation and the life-degradation process, limiting the model's ability to model complex degradation patterns. In addition, existing methods mostly use traditional recurrent neural networks or basic attention mechanisms for time-series modeling, which have limited ability to capture the nonlinear and multi-stage degradation characteristics exhibited during battery aging and are insufficient in modeling long-term dependencies. At the same time, the difficulty in obtaining full-life-cycle data, the limited sample size, and the homogeneous data distribution further exacerbate the risk of model overfitting, reducing its generalization performance and robustness under real-world multi-condition working conditions.
[0004] Therefore, there is an urgent need to develop a RUL prediction method that can effectively integrate multi-dimensional temporal degradation features, enhance small sample modeling capabilities, and accurately model nonlinear degradation processes, so as to improve the accuracy, stability, and engineering applicability of energy storage battery lifetime prediction. Summary of the Invention
[0005] To address the problems existing in the prior art, this invention provides a method for estimating the RUL of energy storage batteries based on the GDA-MHDA-TimesNet model, which solves the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for estimating the RUL of energy storage batteries based on the GDA-MHDA-TimesNet model, characterized by comprising the following steps: S1. Obtain full life cycle charge and discharge data of the energy storage battery under different test conditions; S2. Extract four health characteristics related to battery aging from the test data; S3. Gaussian data augmentation and multi-head differential attention mechanism are used to augment the data from the feature dimension and time series dimension; S4. The enhanced four-dimensional temporal features are transformed into fixed-length input sequences using the sliding window method, and the enhanced dataset is divided according to different ratios of training set and test set. S5. Construct a RUL estimation model with TimesNet as the core architecture, take the enhanced time series as the model input, obtain the final RUL estimation model through iterative training, and output the RUL prediction value. S6. Use the test set on the pre-trained RUL estimation model to obtain the estimated RUL value.
[0007] Preferably, in step S1, under different ambient temperature conditions (24°C and 0°C), the operating condition data of the energy storage battery under constant current and constant voltage charging and constant current discharging strategies are collected to obtain model training samples; the collected data fully covers the current, voltage and temperature information during battery operation.
[0008] Preferably, in step S2, the health characteristics include the following four types: constant current charging time (CCCT), constant voltage charging time (CVCT), discharge voltage change time (DVCT), and battery health state (SOH); the extraction formula for the health characteristics is as follows:
[0009]
[0010]
[0011]
[0012] In the formula, This indicates the time taken to charge to the cutoff voltage during the constant current charging phase. This indicates the constant current charging stage, specifically the charging start time. This indicates the time from the constant voltage charging stage to the cutoff current. This indicates the constant voltage charging stage, specifically the charging start time. This indicates the moment when the voltage first drops to the set termination voltage during the constant current discharge phase. This indicates the moment when the voltage first drops to the set starting voltage during the constant current discharge phase; Preferably, in step S3, the Gaussian data augmentation and multi-head differential attention mechanism specifically include the following steps: S31. Gaussian Data Augmentation (GDA): Based on the local statistical properties of the original feature data, Gaussian noise with a mean of 0 and adaptive standard deviation is added to each feature value. The noise intensity is dynamically adapted according to the importance of the feature and its correlation with the battery's state of harmonics (SOH). The core mathematical expression of this strategy is as follows:
[0013]
[0014]
[0015] In the formula: where, These are the original eigenvalues. These are the eigenvalues after GDA enhancement; Standard Gaussian noise; For dynamically adjusted noise standard deviation, This is the noise intensity adjustment coefficient, with a value ranging from 0.1 to 0.3; The standard deviation of the original feature data. The values are the feature weights, where the core features range from 0.1 to 0.3 and the auxiliary features range from 0.5 to 0.8. , These are the upper and lower limits of the physical constraints on the features, used to prune invalid samples; S32. Multi-Head Differential Attention Mechanism (MHDA): MHDA enhancement focuses on the temporal dimension and mainly includes four core operations: time warp, scaling, window slicing, and noise jittering. Time warp simulates the non-uniform degradation characteristics of SOH during battery aging by nonlinearly stretching or compressing time series. Scaling adaptively scales time series feature values to simulate feature changes under different charge / discharge rates and ambient temperatures. Window slicing locally truncates and interpolates long time series feature sequences, enhancing the model's ability to capture local degradation features. Noise jittering adds random noise related to feature amplitude to simulate sensor measurement errors in actual operation. The core mathematical expression of this strategy is as follows:
[0016]
[0017]
[0018] In the formula: where, This is the original time-series index. For the distorted time-series index, The length of the time series sequence; The value represents the distortion amplitude, ranging from 0.05 to 0.1. This is a nonlinear coefficient, with a value ranging from 0.001 to 0.005; The temporal characteristics after time distortion These are time-series features after scaling. This is the scaling factor, with a value ranging from 0.05 to 0.15. It is a uniformly distributed random variable.
[0019] Preferably, in step S4, the original multidimensional temporal feature sequence is divided using a sliding window method to generate a sample sequence suitable for input to the deep learning model. The sliding window length is set to 8 and the prediction step size to 1, meaning that the four-dimensional features of eight consecutive charge-discharge cycles are used as input to predict the RUL corresponding to the ninth cycle. Simultaneously, the dataset is divided according to a strict temporal order, with the enhanced data split according to different ratios between the training and test sets. Specific splitting schemes include four scenarios: a training set to test set ratio of 2:8, 4:6, 6:4, and 8:2.
[0020] Preferably, the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model is based on the TimesNet model and specifically includes three parts: an embedding layer, a Times Block module, and a regression head. The first part, the embedding layer, maps the original four-dimensional input features to a higher-dimensional feature space through a fully connected layer, improving the model's expressive power. The feature mapping expression of the embedding layer is as follows:
[0021] In the formula, Input temporal features into the model; Output for the embedding layer; , Weights and biases for the embedding layer; The second part, the Times Block module, contains three parallel one-dimensional convolutional branches, employing kernels of 3, 5, and 7 different sizes to capture local dependencies at short, medium, and long time scales. The input features are fused through average pooling of the branch outputs, and then normalized using residual connections and LayerNorm, which enhances information transfer efficiency and alleviates the gradient vanishing problem during training. Multiple Times Blocks are cascaded and stacked to progressively extract deeper temporal patterns. The specific expression for the Times Block multi-scale convolutional fusion is as follows:
[0022]
[0023]
[0024]
[0025] In the formula, One-dimensional convolution outputs with different convolution kernels; For average pooling output; This is the result of multi-scale feature fusion; The dimension adaptation results output by the embedding layer are used for residual connections; The third part, the regression head, flattens the high-dimensional features output by the Times Block and performs nonlinear mapping through two fully connected layers, ultimately outputting the RUL prediction result. The specific prediction expression of the regression head is as follows:
[0026]
[0027]
[0028] In the formula, This is the final output of the Times Block; , , , For the parameters of the fully connected layer in the regression head, This is the final RUL prediction value.
[0029] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: an energy storage battery RUL estimation device based on the GDA-MHDA-TimesNet model for estimating the RUL of an energy storage battery, the device comprising the following modules: Test data acquisition module: Acquires battery test data under different test conditions; Data feature extraction module: Extracts four health features related to battery aging from the test data; Data Augmentation Module: The extracted health features are augmented using a dual-strategy data augmentation method that combines GDA and MHDA. While strictly preserving the core physical meaning of the data, the distribution space of the samples is effectively expanded, and the diversity and representativeness of time-series health features are significantly improved, thereby constructing a richer and more generalizable augmented sample library. Data partitioning module: The sliding window method is used to convert multidimensional time series data into a fixed-length input sequence suitable for the input model, and the augmented dataset is partitioned according to different ratios of training set and test set in strict chronological order; Pre-training module: The processed data is input into the pre-built energy storage battery RUL estimation model based on GDA-MHDA-TimesNet model for iterative training to obtain the pre-trained RUL estimation model; RUL estimation module: Uses the test set to train a pre-trained RUL estimation model to obtain the estimated RUL value.
[0030] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: an electronic device includes a processor and a memory for storing one or more programs; when the one or more programs are executed by the processor, the processor executes the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model.
[0031] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model. The beneficial effects of this invention are as follows: To address the problems of scarce data samples, limited features, and weak model generalization ability in existing energy storage battery RUL estimation methods, a new energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model is proposed. This method employs a dual strategy combining Gaussian data augmentation and multi-head differential attention mechanisms to synergistically enhance the original data from both feature and time-series dimensions, effectively improving sample diversity and model robustness. Simultaneously, the model uses TimesNet, capable of capturing multi-scale time-series dependencies, as its core architecture. Inputs include multi-dimensional time-series features such as constant current charging time, constant voltage charging time, discharge voltage change time, and battery health status to learn the nonlinear degradation law of battery capacity. This invention effectively improves the accuracy of RUL estimation. Experimental results verify its superiority and stability under complex operating conditions, demonstrating high estimation accuracy and providing an effective technical solution for predictive maintenance of battery management systems. Attached Figure Description
[0032] Figure 1 This is a schematic diagram of the process steps of the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model in an embodiment of the present invention. Figure 2 This is a schematic diagram of the energy storage battery RUL estimation method framework based on the GDA-MHDA-TimesNet model in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the GDA-MHDA-TimesNet model; Figure 4This is a comparison chart of different features before and after data augmentation. Figure 5a A comparison chart of RUL estimation results for battery B0005 using different methods at an ambient temperature of 24°C; Figure 5b A comparison chart of RUL estimation results for battery B0006 using different methods at an ambient temperature of 24°C; Figure 5c This is a comparison chart showing the RUL estimation results of battery B0007 using different methods at an ambient temperature of 24°C. Figure 5d A comparison chart of RUL estimation results for battery B0018 using different methods at an ambient temperature of 24°C; Figure 6a A comparison chart of RUL estimation results for CS2_35 battery using different methods at an ambient temperature of 1°C; Figure 6b A comparison chart of RUL estimation results for CS2_36 battery using different methods at an ambient temperature of 1°C; Figure 6c A comparison chart of RUL estimation results for CS2_37 battery using different methods at an ambient temperature of 1°C; Figure 6d A comparison chart of RUL estimation results for CS2_38 battery using different methods at an ambient temperature of 1°C; Figure 7 This is a schematic diagram of the energy storage battery RUL estimation device module based on the GDA-MHDA-TimesNet model in an embodiment of the present invention. Figure 8 This is a schematic diagram of the electronic device structure in an embodiment of the present invention; In the diagram, 110 is the test data acquisition module; 120 is the data feature extraction module; 130 is the data preprocessing module; 140 is the data feature fusion module; 150 is the pre-training module; 160 is the SOH estimation module; 210 is the processor; and 220 is the memory. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] Example 1 This invention provides a technical solution: a method for estimating the RUL of energy storage batteries based on the GDA-MHDA-TimesNet model, such as... Figure 1and Figure 2 As shown, the specific steps include the following: S1. Obtain battery test data under different test conditions; Operating data of the energy storage battery under constant current and constant voltage charging and constant current discharging strategies were collected under different ambient temperature conditions (24°C and 0°C) to obtain model training samples. The collected data comprehensively covers the current, voltage, and temperature information during battery operation.
[0035] The test data includes current, voltage, and battery casing temperature collected during battery operation, all acquired using high-precision sensors. Meanwhile, the Recovery Under Length (RUL) prediction target during model training is defined as the remaining number of cycles corresponding to the battery capacity decaying to the failure threshold. The core associated state of health (SOH) is the basis for RUL prediction, and SOH is calculated as follows:
[0036] In the formula: This indicates the battery's health status. This is the current actual capacity of the battery; This is the initial rated capacity of the battery.
[0037] By comprehensively analyzing constant current / constant voltage charging time, discharge voltage change time, and multi-dimensional time series data such as voltage, current, and temperature, the battery capacity degradation trajectory can be accurately tracked, thereby enabling reliable prediction of the remaining battery life (RUL).
[0038] It is understood that the RUL estimation can be integrated into the battery management system (BMS), energy storage power station monitoring platform, vehicle control unit, or portable devices with edge computing capabilities (such as industrial tablet PCs and smart monitoring terminals) to provide key support for predictive maintenance and life management of batteries.
[0039] S2. Extract four health characteristics related to battery aging from the test data; Health characteristics such as constant current charging time (CCCT), constant voltage charging time (CVCT), constant current discharge voltage change time (DVCT), and state of health (SOH) are extracted from battery test data. The extraction formulas for some of these health characteristics are as follows:
[0040]
[0041]
[0042]
[0043] In the formula, This indicates the time taken to charge to the cutoff voltage during the constant current charging phase. This indicates the constant current charging stage, specifically the charging start time. This indicates the time from the constant voltage charging stage to the cutoff current. This indicates the constant voltage charging stage, specifically the charging start time. This indicates the moment when the voltage first drops to the set termination voltage during the constant current discharge phase. This indicates the moment when the voltage first drops to the set starting voltage during the constant current discharge phase. These characteristics can effectively characterize the performance evolution of a battery during charging and discharging, especially its dynamic behavior closely related to aging mechanisms. By extracting and analyzing these time-series health characteristics, battery degradation modes can be accurately identified, and their health status can be assessed, thus providing crucial input for predicting remaining lifespan. The evolution trends of these characteristics have significant engineering guiding value for predicting battery lifespan end-points, warning of potential performance degradation, and optimizing the operation and maintenance strategies of energy storage systems.
[0044] S3. The extracted health features are augmented using a dual-strategy approach based on GDA and MHDA, which enriches the data diversity of the samples.
[0045] During the data acquisition process, obtaining full life cycle aging data of energy storage batteries faces challenges such as long cycle, high cost, and scarcity of effective samples. Furthermore, due to factors such as sensor noise and transmission delay, the obtained time-series health feature data often suffers from a single distribution and contains outliers, which severely limits the generalization ability and prediction accuracy of data-driven models.
[0046] To address the aforementioned issues, this invention utilizes Gaussian data augmentation to process the extracted health features from the feature dimension. Based on the local statistical characteristics of the original feature data, Gaussian noise with a mean of 0 and adaptive standard deviation is added to each feature value. The noise intensity is dynamically adapted according to the importance of the feature and its correlation with the battery's state of health (SOH). The calculation formula is as follows:
[0047]
[0048]
[0049] In the formula: where, These are the original eigenvalues. These are the eigenvalues after GDA enhancement; Standard Gaussian noise; For dynamically adjusted noise standard deviation, This is the noise intensity adjustment coefficient, with a value ranging from 0.1 to 0.3; The standard deviation of the original feature data. The values are the feature weights, where the core features range from 0.1 to 0.3 and the auxiliary features range from 0.5 to 0.8. , These are the upper and lower limits of the physical constraints for the features, used to prune invalid samples.
[0050] Simultaneously, a multi-head differentiated attention mechanism is employed to enhance the temporal dimension, primarily comprising four core operations: time warp, scaling, window slicing, and noise jitter. Time warp introduces nonlinear time perturbations to simulate the non-uniformity of battery aging rates, enabling the model to adapt to different degradation rhythms. Scaling adaptively adjusts feature values to correspond to changes in feature amplitude at different charge / discharge rates or ambient temperatures, enhancing the model's robustness to fluctuations in operating conditions. Window slicing locally truncates and reconstructs long sequences, forcing the model to focus more on feature patterns in local degradation stages, improving the ability to identify key degradation segments. Noise jitter adds random noise related to feature amplitudes, simulating sensor measurement errors in actual operation. These four operations work together on the original temporal features, significantly expanding the distribution space of training samples while strictly preserving the core physical meaning and overall trend of the data, effectively constructing a richer and more representative augmentation sample library. The core mathematical expression of this strategy is as follows:
[0051]
[0052]
[0053] In the formula: where, This is the original time-series index. For the distorted time-series index, The length of the time series sequence; The value represents the distortion amplitude, ranging from 0.05 to 0.1. This is a nonlinear coefficient, with a value ranging from 0.001 to 0.005; The temporal characteristics after time distortion These are time-series features after scaling. This is the scaling factor, with a value ranging from 0.05 to 0.15. It is a uniformly distributed random variable.
[0054] Table 1 shows the comparison of standard deviations of different features under different enhancement strategies.
[0055] Figure 4 The image shows a comparison of data augmentation before and after different features. S4. The enhanced four-dimensional temporal features are transformed into fixed-length input sequences using the sliding window method, and the enhanced dataset is divided according to different ratios of training set and test set.
[0056] To eliminate the impact of dimensional differences between different features on model training, this invention employs a sliding window method to transform the original multidimensional temporal features (including CCCT, DVCT, CVCT, and SOH) to construct an input sequence suitable for deep learning models. Considering both the actual sample size and the dynamic characteristics of the battery aging process, the sliding window length is set to 8 and the prediction step size to 1. This means that the four-dimensional features from eight consecutive charge-discharge cycles are used as model input to predict the RUL of the next cycle. This setting effectively captures the nonlinear decay trend of SOH within a short window while preserving the dependencies between temporal health features, ensuring model training efficiency while providing sufficient contextual information for the model.
[0057] To further enhance the model's generalization ability and the rationality of the evaluation, this invention adopts a strategy of strictly dividing the dataset according to time sequence. The enhanced dataset is divided into training and test sets, with four different ratios: 2:8, 4:6, 6:4, and 8:2. The division process strictly adheres to the chronological order, ensuring that all loop cycles in the test set are later than those in the training set. This effectively avoids the risk of information leakage and guarantees the reliability and practical applicability of the model training and evaluation process.
[0058] S5. Construct a RUL estimation model with TimesNet as the core architecture, take the enhanced time series as the model input, obtain the final RUL estimation model through iterative training, and output the RUL prediction value.
[0059] The energy storage battery RUL estimation method based on GDA-MHDA-TimesNet proposed in this invention uses the time-series deep neural network TimesNet as its core architecture. The model structure mainly consists of three parts: an embedding layer, a Times Block module, and a regression head. Its specific structure is as follows: Embedding Layer: The model first performs feature upscaling and nonlinear mapping on the input four-dimensional temporal health features (CCCT, CVCT, DVCT, SOH) to enhance feature expressive power. Specifically, the input feature sequence constructed by the sliding window is mapped to a high-dimensional space through a fully connected layer:
[0060] In the formula, Input temporal features into the model; Output for the embedding layer; , The embedding layer weights and biases are defined. This embedding process not only extends low-dimensional features to the hidden layer dimension, but also strengthens the interaction between features through nonlinear transformations, resulting in a high-dimensional temporal embedding. This lays the foundation for subsequent multi-scale modeling; The Times Block module is a core component of TimesNet, designed to extract degradation-related patterns from multi-scale temporal dependencies. The module contains three parallel convolutional branches, each employing a different sized one-dimensional convolutional kernel for the temporal embeddings. Perform local feature extraction:
[0061] In the formula, each The degradation trends at short, medium, and long scales are captured by using a convolution kernel size k. To reduce feature dimensionality and enhance stability, the outputs of each branch are further downsampled by pooling.
[0062] Next, the multi-scale features are fused along the channel dimension:
[0063] Finally, residual connections and layer normalization are used to preserve initial feature information and stabilize the training process.
[0064] in, The original embedded features are linearly projected to adjust their dimensions to accommodate residual computation. This design enables the model to simultaneously capture the dynamics of battery degradation in both local variations and overall trends, effectively learning nonlinear temporal patterns during the degradation process.
[0065] Regression head: In obtaining temporal feature representation Next, it needs to be mapped to RUL predictions. First, the temporal dimension is flattened to transform it into a feature vector suitable for fully connected layers:
[0066] Subsequently, nonlinear regression is implemented using a two-layer feedforward network. Dropout regularization is added to the first layer to prevent overfitting.
[0067] Finally, the output layer maps the latent features to scalar RUL predictions:
[0068] In the formula, , , , For the parameters of the fully connected layer in the regression head, This is the final RUL prediction value.
[0069] To prevent overfitting, the model introduces Dropout regularization between fully connected layers and uses the GELU activation function instead of the traditional ReLU, effectively mitigating the overfitting risk of deep networks when dealing with the nonlinear features of battery degradation. During the model training phase, the Adam optimizer is selected with an initial learning rate of 0.001 and a weight decay coefficient of 0.01. The objective function is constructed using the MSE (mean squared error) loss function, and gradient pruning is employed to suppress gradient explosion. Furthermore, an early stopping mechanism is implemented on the validation set, automatically terminating training when the MSE metric stops decreasing for 50 consecutive epochs. This significantly reduces redundant computation while ensuring convergence stability, ensuring the model has optimal generalization ability and resource utilization efficiency.
[0070] The key hyperparameters of the proposed GDA-MHDA-TimesNet model were determined after multiple rounds of grid search and validation set tuning, and the specific settings are shown in Table 2.
[0071]
[0072] S6. Use the test set on the pre-trained RUL estimation model to obtain the estimated RUL values. The pre-training and testing process of this model is as follows: First, energy storage battery samples were obtained (using the publicly available CALCE (lithium battery, ambient temperature 1℃) and NASA (lithium battery, ambient temperature 24℃) datasets). These datasets collected data on current, voltage, and temperature of lithium batteries tested at the corresponding ambient temperatures. After performing steps S1 to S5 on the datasets, the datasets were divided into training, validation, and test sets in chronological order: the first 60% of the data in each dataset was used as the training set, the middle 10% as the validation set (used to verify model effectiveness), and the remaining data as the test set (used to evaluate model performance).
[0073] Subsequently, the loss function for training the RUL estimation model is set to mean squared error (MSE), which is calculated as follows:
[0074] In the formula, This is the estimated RUL value. This is the actual value of RUL.
[0075] During the model training phase, a dynamically decaying learning rate scheduling strategy is adopted, specifically: the initial learning rate is set to 0.001, and thereafter, every 50 training epochs, the learning rate is smoothly reduced by a decay factor of 0.9. Furthermore, the Adam optimizer is used to iteratively update the model parameters with an initial learning rate of 0.001.
[0076] To prevent overfitting during training, an early stopping mechanism is introduced, which continuously monitors changes in the loss function on the validation set. A patience value of 50 is set; if the validation set loss does not decrease significantly within 50 consecutive training epochs, training is automatically terminated, and the optimal model parameters from that period are retained.
[0077] After the model is trained, it can be applied to online or offline estimation of the range-limited liability (RUL) of energy storage batteries. To intuitively verify the superiority of the proposed method, we present comparison curves of the RUL prediction results of our method with those of various baseline models on two publicly available datasets. Figures 5a to 5d The actual degradation curves of four batteries (B0005, B0006, B0007, and B0018) in the NASA dataset (ambient temperature 24°C) are compared with the prediction results of each model. Figures 6a to 6d The results for four batteries (CS2_35, CS2_36, CS2_37, and CS2_38) in the CALCE dataset (ambient temperature 1℃) are shown. The figures demonstrate that after fusing the GDA and MHDA dual data augmentation strategies, the model exhibits better generalization performance under limited data sample conditions, especially in the early stages of capacity degradation (sparse data, indistinct degradation characteristics). The predicted trajectory matches the true value better, significantly suppressing early overfitting and improving the accuracy and reliability of the prediction.
[0078] To quantitatively evaluate the predictive performance of the proposed GDA-MHDA-TimesNet model, this study selected Transformer, Informer, and the original TimesNet model without data augmentation as baselines on two publicly available battery datasets: NASA (24℃) and CALCE (1℃). Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used as evaluation metrics. Specific results are shown in Tables 3 and 4. It can be seen that in all eight battery samples tested, the model of this invention achieved optimal or near-optimal results on most evaluation metrics, especially in the two key error metrics of RMSE and MAE, showing a significant advantage over existing mainstream time-series models. This verifies the effectiveness and advancement of the proposed fusion strategy in battery RUL prediction tasks.
[0079]
[0080]
[0081] This invention presents a method for estimating the relative degradation level (RUL) of energy storage batteries based on the GDA-MHDA-TimesNet model. This method extracts multiple time-series health features and introduces a dual-strategy data augmentation approach combining Gaussian data augmentation and multi-head differential attention mechanisms to construct a high-quality sample library suitable for battery degradation modeling. Based on this, using TimesNet as the core architecture, it effectively integrates local and global degradation patterns at different time scales during battery degradation. Experiments show that this method exhibits excellent estimation performance under various temperature conditions, including 24℃ and 0℃, and significantly outperforms existing mainstream methods in evaluation metrics such as MAE, RMSE, and MAPE. It effectively addresses the problems of insufficient feature information utilization, weak noise resistance, and insufficient modeling ability for complex degradation patterns inherent in single-feature or traditional decomposition methods, thus improving the accuracy and robustness of RUL estimation.
[0082] Based on the same inventive concept as the above-described method embodiments, this application also provides a storage battery RUL estimation device based on the GDA-MHDA-TimesNet model. This device can achieve the functions provided by the above-described method embodiments, such as... Figure 7 As shown, the device includes the following modules: Test data acquisition module 110: Acquires battery test data under different test conditions; Data feature extraction module 120: Extracts four health features related to battery aging from the test data; Data Augmentation Module 130: The extracted health features are augmented using a dual-strategy data augmentation method that combines GDA and MHDA. While strictly preserving the core physical meaning of the data, the distribution space of the samples is effectively expanded, and the diversity and representativeness of time-series health features are significantly improved, thereby constructing a richer and more generalizable augmented sample library. Data partitioning module 140: Uses the sliding window method to convert multidimensional time series data into a fixed-length input sequence suitable for the input model, and strictly partitions the augmented dataset according to different ratios of training and test sets in chronological order; Pre-training module 150: Inputs the processed data into the pre-built energy storage battery RUL estimation model based on the GDA-MHDA-TimesNet model for iterative training to obtain the pre-trained RUL estimation model; RUL estimation module 160: Uses the test set to obtain the estimated RUL value from the pre-trained RUL estimation model. Validates the effectiveness of the model.
[0083] Based on the same inventive concept as the above-described method embodiments, this application also provides an electronic device, such as... Figure 8 As shown, the device includes: a processor 210; and a memory 220 for storing one or more programs; When the one or more programs are executed by the processor 210, the processor performs the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model.
[0084] The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model specifically includes the following: Obtain battery test data under different test conditions; Four health characteristics related to battery aging were extracted from the test data; Gaussian data augmentation and multi-head differential attention mechanisms are used to augment the data from both feature and temporal dimensions. The enhanced four-dimensional temporal features are transformed into fixed-length input sequences using the sliding window method, and the enhanced dataset is divided according to different ratios of training and test sets. A RUL estimation model based on TimesNet is constructed. The enhanced time series is used as the model input, and the final RUL estimation model is obtained through iterative training. The RUL prediction value is then output. The test set is used to obtain the estimated RUL value from the pre-trained RUL estimation model; Based on the same inventive concept as the above-described method embodiments, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by processor 210, implements the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model.
[0085] The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model specifically includes the following: Obtain battery test data under different test conditions; Four health characteristics related to battery aging were extracted from the test data; Gaussian data augmentation and multi-head differential attention mechanisms are used to augment the data from both feature and temporal dimensions. The enhanced four-dimensional temporal features are transformed into fixed-length input sequences using the sliding window method, and the enhanced dataset is divided according to different ratios of training and test sets. A RUL estimation model based on TimesNet is constructed. The enhanced time series is used as the model input, and the final RUL estimation model is obtained through iterative training. The RUL prediction value is then output. The test set is used to obtain the estimated RUL value from the pre-trained RUL estimation model; In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0086] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0087] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0088] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.
[0089] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0090] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."
[0091] The use of "first" and "second" in the embodiments is merely to distinguish similar objects and does not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.
[0092] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for estimating the relative uptime (RUL) of an energy storage battery, characterized in that, A dual strategy of Gaussian Data Augmentation (GDA) and Multi-Head Differentiated Attention (MHDA) is employed to augment time-series data of battery health features, and estimation is performed based on the TimesNet model. The steps include: S1. Obtain full life cycle charge and discharge data of the energy storage battery under different test conditions; S2. Extract multi-dimensional time-series features of CCCT, CVCT, DVCT and SOH that reflect the aging pattern of the battery from the test data; S3. Gaussian data augmentation and multi-head differential attention mechanism are used to augment the data from the feature dimension and time series dimension; S4. The enhanced four-dimensional temporal features are transformed into fixed-length input sequences using the sliding window method, and the enhanced dataset is divided according to different ratios of training set and test set. S5. Construct a RUL estimation model with TimesNet as the core architecture, take the enhanced time series as the model input, obtain the final RUL estimation model through iterative training, and output the RUL prediction value. S6. Use the test set on the pre-trained RUL estimation model to obtain the estimated RUL value.
2. The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model according to claim 1, characterized in that: In step S1, under different ambient temperature conditions (24°C and 0°C), the operating condition data of the energy storage battery under constant current and constant voltage charging and constant current discharging strategies are collected to obtain model training samples. The collected data fully covers the current, voltage and temperature information during battery operation.
3. The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model according to claim 1, characterized in that: In step S2, the health characteristics include the following four types: constant current charging time (CCCT), constant voltage charging time (CVCT), discharge voltage change time (DVCT), and battery state of health (SOH); the extraction formula for the health characteristics is as follows: In the formula, This indicates the time taken to charge to the cutoff voltage during the constant current charging phase. This indicates the constant current charging stage, specifically the charging start time. This indicates the time from the constant voltage charging stage to the cutoff current. This indicates the constant voltage charging stage, specifically the charging start time. This indicates the moment when the voltage first drops to the set termination voltage during the constant current discharge phase. This indicates the moment when the voltage first drops to the set starting voltage during the constant current discharge phase.
4. The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model according to claim 1, characterized in that: In step S3, the Gaussian data augmentation and multi-head differential attention mechanism specifically include the following steps: S31. Gaussian Data Augmentation (GDA): Based on the local statistical properties of the original feature data, Gaussian noise with a mean of 0 and adaptive standard deviation is added to each feature value. The noise intensity is dynamically adapted according to the importance of the feature and its correlation with the battery's state of harmonics (SOH). The core mathematical expression of this strategy is as follows: In the formula: where, These are the original eigenvalues. These are the eigenvalues after GDA enhancement; Standard Gaussian noise; For dynamically adjusted noise standard deviation, This is the noise intensity adjustment coefficient, with a value ranging from 0.1 to 0.3; The standard deviation of the original feature data. The values are the feature weights, where the core features range from 0.1 to 0.3 and the auxiliary features range from 0.5 to 0.
8. , These are the upper and lower limits of the physical constraints on the features, used to prune invalid samples; S32. Multi-Head Differential Attention Mechanism (MHDA): MHDA enhancement focuses on the temporal dimension and mainly includes four core operations: time warp, scaling, window slicing, and noise jittering. Time warp simulates the non-uniform degradation characteristics of SOH during battery aging by nonlinearly stretching or compressing time series. Scaling adaptively scales time series feature values to simulate feature changes under different charge / discharge rates and ambient temperatures. Window slicing locally truncates and interpolates long time series feature sequences, enhancing the model's ability to capture local degradation features. Noise jittering adds random noise related to feature amplitude to simulate sensor measurement errors in actual operation. The core mathematical expression of this strategy is as follows: In the formula: where, This is the original time-series index. For the distorted time-series index, The length of the time series sequence; The value represents the distortion amplitude, ranging from 0.05 to 0.
1. This is a nonlinear coefficient, with a value ranging from 0.001 to 0.005; The temporal characteristics after time distortion These are time-series features after scaling. This is the scaling factor, with a value ranging from 0.05 to 0.
15. It is a uniformly distributed random variable.
5. The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model according to claim 1, characterized in that: In step S4, the original multidimensional temporal feature sequence is divided using a sliding window method to generate a sample sequence suitable for input to the deep learning model. The sliding window length is set to 8 and the prediction step size is 1, that is, the four-dimensional features of 8 consecutive charge-discharge cycles are used as input to predict the RUL corresponding to the 9th cycle. At the same time, the dataset is divided according to a strict time order, and the enhanced data is divided according to different ratios of training set and test set. The specific division schemes include four cases where the ratio of training set to test set is 2:8, 4:6, 6:4 and 8:
2.
6. The method for estimating the RUL of an energy storage battery based on the GDA-MHDA-TimesNet model according to claim 1, characterized in that: The energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model has the TimesNet model as its core and specifically includes three parts: an embedding layer, a Times Block module, and a regression head. The first part, the embedding layer, maps the original four-dimensional input features to a higher-dimensional feature space through a fully connected layer, improving the model's expressive power. The feature mapping expression of the embedding layer is as follows: In the formula, Input temporal features into the model; Output for the embedding layer; , Weights and biases for the embedding layer; The second part, the Times Block module, contains three parallel one-dimensional convolutional branches, which use three different sizes of convolutional kernels (3, 5, and 7) to capture local dependencies at short, medium, and long time scales. The input features are fused through average pooling of the outputs of each branch, and then normalized using residual connections and LayerNorm. This enhances information transmission efficiency and alleviates the gradient vanishing problem during training. Multiple Time Blocks are cascaded and stacked to gradually extract deep temporal patterns. The specific expression for the Time Block multi-scale convolution fusion is as follows: In the formula, One-dimensional convolution outputs with different convolution kernels; For average pooling output; This is the result of multi-scale feature fusion; The dimension adaptation results output by the embedding layer are used for residual connections; The third part, the regression head, flattens the high-dimensional features output by the Times Block and performs nonlinear mapping through two fully connected layers, ultimately outputting the RUL prediction result. The specific prediction expression of the regression head is as follows: In the formula, This is the final output of the Times Block; , , , For the parameters of the fully connected layer in the regression head, This is the final RUL prediction value.
7. A storage battery RUL estimation device based on the GDA-MHDA-TimesNet model according to any one of claims 1-6, characterized in that: The device includes the following modules: Test data acquisition module (110): Acquires battery test data under different test conditions; Data feature extraction module (120): Extracts four health features related to battery aging from the test data; Data augmentation module (130): The extracted health features are augmented using a dual-strategy data augmentation method that combines GDA and MHDA. Under the premise of strictly preserving the core physical meaning of the data, the distribution space of the samples is effectively expanded, and the diversity and representativeness of the time-series health features are significantly improved, thereby constructing a richer and more generalizable augmented sample library. Data partitioning module (140): The sliding window method is used to convert multidimensional time series data into a fixed-length input sequence suitable for the input model, and the enhanced dataset is partitioned according to different ratios of training set and test set in strict time order; Pre-training module (150): Input the processed data into the pre-built energy storage battery RUL estimation model based on GDA-MHDA-TimesNet model for iterative training to obtain the pre-trained RUL estimation model; RUL estimation module (160): Uses the test set to the pre-trained RUL estimation model to obtain the estimated RUL value.
8. An electronic device, characterized in that: The electronic device includes a processor (210) and a memory (220) for storing one or more programs; When the one or more programs are executed by the processor (210), the processor performs the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that: It stores a computer program that, when executed by a processor (210), implements the energy storage battery RUL estimation method based on the GDA-MHDA-TimesNet model as described in any one of claims 1-6.