Battery state prediction method based on lithium manganate characteristic modeling
By using a battery state prediction method based on lithium manganese oxide characteristic modeling and employing improved SOCNet and TabNet models, the shortcomings of existing battery state prediction methods in terms of accuracy, adaptability, and efficiency are addressed. This method achieves high-precision, real-time updates and optimization strategies for battery state, thereby extending battery life.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHAOXIAN QIANGNENG POWER SUPPLY CO LTD
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-12
AI Technical Summary
Existing battery state prediction methods are insufficient in terms of accuracy, adaptability, and efficiency. They cannot effectively handle complex features and real-time updates, making it difficult to meet the real-time prediction requirements of battery management systems.
A battery state prediction method based on lithium manganese oxide characteristic modeling is adopted. The improved SOCNet and TabNet models are used, combined with weighted fusion, dynamic update mechanism and multi-objective optimization algorithm, to collect and process battery data in real time and dynamically adjust charging and discharging strategy and temperature control strategy.
It improves the accuracy and adaptability of battery state prediction, ensures real-time updates and accuracy of prediction results, extends battery life, and enhances the efficiency and performance of the battery management system.
Smart Images
Figure CN122017578A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery management system technology, and in particular to a battery state prediction method based on lithium manganese oxide characteristic modeling. Background Technology
[0002] With the increasing popularity of electric vehicles and energy storage devices, real-time prediction and optimized control of battery state by battery management systems (BMS) has become a research hotspot. Existing battery state prediction methods mainly rely on traditional statistical models or simple machine learning algorithms, such as support vector machines (SVM) or decision trees. These methods often face the following problems:
[0003] Low accuracy in battery state prediction: Traditional methods often fail to capture the complex nonlinear relationship between battery health and capacity decay, resulting in significant errors in predictions of remaining battery life and health. Lack of dynamic adaptability: Existing methods often fail to update predictions in real time when dealing with dynamic changes in battery state (such as fluctuations during charging and discharging, and temperature changes), making it difficult to effectively address the adaptability of batteries to different operating environments. Inefficient handling of complex features: Many methods rely on manually selected features, ignoring the complex relationships of high-dimensional features in battery state data. This prevents the model from fully utilizing the potential information in the battery data, affecting prediction accuracy and efficiency. Slow model training: Some existing methods suffer from complex training processes and high computational costs when processing large-scale battery data, making it difficult to meet real-time prediction requirements.
[0004] Therefore, how to provide a battery state prediction method based on lithium manganese oxide characteristic modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a battery state prediction method based on lithium manganese oxide characteristic modeling. This invention fully utilizes battery electrochemical characteristic modeling, an improved SOCNet network structure, a dynamic update mechanism, and an improved TabNet model. It calculates a comprehensive health status assessment value through a weighted fusion method and optimizes the battery's charge / discharge strategy. This method can accurately predict the battery's remaining capacity, internal resistance, health status, and capacity decay, possessing advantages such as high accuracy, high adaptability, and real-time update capability.
[0006] A battery state prediction method based on lithium manganese oxide characteristic modeling according to an embodiment of the present invention includes the following steps:
[0007] Real-time data collection of battery voltage, current, temperature, and charge / discharge cycle count is performed and preprocessed to obtain preprocessed data;
[0008] Based on the electrochemical characteristics of lithium manganese oxide batteries, an improved SOCNet model was established, and the preprocessed data was input into the corresponding sub-model to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery.
[0009] Based on the estimated remaining power, estimated internal resistance, health status assessment, and estimated capacity decay, a weighted fusion model is used to calculate the comprehensive health status assessment of the battery.
[0010] Based on the comprehensive health status assessment value, the charging and discharging strategy of the battery is optimized by dynamically adjusting the charging and discharging rate, temperature control strategy and multi-objective optimization algorithm.
[0011] The improved SOCNet model is dynamically updated after each charge and discharge cycle using an incremental neural network, and the updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are obtained in real time.
[0012] Based on updated estimates of remaining charge, internal resistance, health status, and capacity decay, the future state of the battery is predicted using an improved TabNet model.
[0013] Optionally, the real-time acquisition of battery voltage, current, temperature, and charge / discharge cycle count data, followed by preprocessing to obtain preprocessed data, specifically includes:
[0014] Battery data is collected in real time by sensors, including battery voltage, current, temperature and charge / discharge cycle count data. The sensors include a voltage sensor, a current sensor, a temperature sensor and a cycle counter. The voltage sensor collects the battery terminal voltage, the current sensor collects the battery charge / discharge current, the temperature sensor collects the battery operating temperature, and the cycle counter records the number of charge / discharge cycles of the battery.
[0015] The battery data is time-series aligned, which includes aligning voltage, current, temperature and charge / discharge cycle count data according to timestamps, and interpolating missing values. The interpolation method includes linear interpolation or spline interpolation.
[0016] The battery data is resampled and normalized. The resampling is performed on the collected data periodically according to the battery's working state, either uniformly or non-uniformly. The normalization method includes converting the battery data into values between 0 and 1.
[0017] Data augmentation techniques are employed to expand battery data. These methods include random perturbation of the battery's current and voltage characteristics, time window shifting, and time-series data pruning.
[0018] Optionally, based on the electrochemical characteristics of lithium manganese oxide batteries, an improved SOCNet model is established, and preprocessed data is input into the corresponding sub-model to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery. Specifically, this includes:
[0019] An improved SOCNet model is constructed. The voltage, current, temperature and charge-discharge cycle number of the battery are input. After standardization, the data are input to the input layer. A fully connected layer is used for preliminary feature extraction. Each feature is then standardized and input to the model.
[0020] An electrochemical characteristic subnetwork is constructed. The voltage-current relationship, capacity decay, and dynamic changes of internal resistance of the battery are calculated through a physical model. The charge and discharge data of the battery are processed using a fully connected layer. The nonlinear characteristics of the battery's internal resistance and capacity decay are fitted using the ReLU activation function. The capacity decay of the battery under time-series data is predicted using an LSTM layer. The remaining capacity and health status of the battery are calculated.
[0021] A multi-task learning module is introduced, and a shared layer is used to extract common features of the battery. The shared layer uses a multilayer perceptron for feature extraction, and different tasks are assigned to independent sub-networks for processing. Remaining power prediction, health status assessment, and capacity decay prediction are processed by independent regression sub-networks. An output layer with a linear activation function is used to predict the remaining power, health status, and capacity decay values.
[0022] An adaptive learning rate adjustment mechanism is introduced, using the Adam optimization algorithm as the optimizer, and the learning rate is dynamically adjusted according to changes in battery state during training. The learning rate is automatically adjusted based on the learning rate scheduler.
[0023] A self-attention mechanism is introduced, which uses scaled dot product attention to calculate the similarity between battery state features and dynamically weight them. The battery's voltage, current, and internal resistance features are used as queries, keys, and values. The attention layer is used to weight the features and output them. The contribution of each feature to the prediction is adaptively adjusted through the self-attention mechanism.
[0024] An incremental learning module is added, using an online learning algorithm. Each time the battery state data is updated, the model fine-tunes the parameters based on the new data. By using the mini-batch gradient descent method, the model updates the weights after each charge-discharge cycle.
[0025] A regularization term is added to the loss function. The regularization term uses L2 regularization to constrain the parameters of the improved SOCNet model. The regularization coefficient is selected by cross-validation. The weighting coefficients in the loss function are adjusted according to different battery operating environments.
[0026] Optionally, the step of calculating the battery's comprehensive health status assessment value using a weighted fusion model based on the remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate specifically includes:
[0027] The remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are fed into the weighted fusion model as input data. The input data undergoes feature extraction and linear transformation through a fully connected layer to output intermediate features.
[0028] The weighted fusion model assigns weights to each input data point, uses the Swish activation function to perform a non-linear transformation on each input data point, the Swish activation function performs weighted processing on the input data, and calculates the weighted value of each feature through the Swish activation function. The weighted value is dynamically adjusted based on the battery's historical data and current operating status.
[0029] The weighted estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay are summed to obtain the battery's comprehensive health status assessment.
[0030] The comprehensive health status assessment values are normalized to convert them into a uniform numerical range.
[0031] The normalized comprehensive health status assessment value was optimized using regression analysis, and the model parameters were optimized using the mean squared error loss function.
[0032] Optionally, the optimization of the battery's charging and discharging strategy based on a comprehensive health status assessment value, through dynamic adjustment of charging and discharging rates, temperature control strategies, and multi-objective optimization algorithms, specifically includes:
[0033] Based on the comprehensive health status assessment value, the charging and discharging rate and temperature control strategy of the battery are calculated. When the comprehensive health status assessment value is higher than the preset threshold, the charging rate is set to the maximum value. When the comprehensive health status assessment value is lower than the preset threshold, the charging rate is reduced to a smaller value to reduce damage to the battery. When the estimated internal resistance of the battery is greater than the preset value, the charging current is reduced. When the estimated capacity decay exceeds the preset threshold, the temperature control strategy reduces the charging rate and starts the cooling system.
[0034] The charge / discharge rate and temperature control strategy are optimized by using a particle swarm optimization algorithm. By setting multiple objective functions, including maximizing charging efficiency, minimizing temperature rise, and extending battery life, each particle is iteratively calculated during the optimization process to obtain the optimal charge / discharge rate, charging current, charging time, and temperature control parameters. Based on the comprehensive health status assessment value and internal resistance estimation value, each parameter is dynamically adjusted.
[0035] The charging current, discharging current, charging power, and temperature control system operating modes are adjusted in real time based on the optimized charging and discharging rates, charging current, charging time, and temperature control parameters.
[0036] Optionally, the improved SOCNet model is dynamically updated after each charge and discharge cycle using an incremental neural network to obtain updated estimates of remaining capacity, internal resistance, health status, and capacity decay in real time. Specifically, this includes:
[0037] The improved SOCNet model is dynamically updated by using an incremental neural network. After each charge and discharge cycle, the model is corrected and updated based on real-time collected battery status data to obtain updated estimates of remaining charge, internal resistance, health status, and capacity decay.
[0038] After each charge and discharge cycle, the battery voltage, current, temperature, charge and discharge duration, charging current, and discharge current are collected as incremental data and input into the improved SOCNet model. The model dynamically adjusts the internal weights of the improved SOCNet model by processing the incremental data.
[0039] Calculate the error between the output of the current improved SOCNet model and the actual measurement data, and use the backpropagation algorithm to update the weights in the improved SOCNet model through error backpropagation;
[0040] The network weights of the improved SOCNet model are optimized using the stochastic gradient descent method. The loss is calculated for each incremental data, and the parameters of the improved SOCNet model are adjusted according to the gradient of the loss.
[0041] Through incremental learning, the improved SOCNet model corrects the remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate in real time, and provides feedback on the updated estimate at the end of each charge-discharge cycle.
[0042] Optionally, the prediction of the battery's future state based on updated remaining capacity estimates, internal resistance estimates, health status assessments, and capacity decay estimates, using an improved TabNet model, specifically includes:
[0043] The updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are input into the improved TabNet model.
[0044] The input data is standardized by using a robust scaler method to remove outliers and unify the scale of the input data.
[0045] An adaptive weighted feature selection module is added to the feature selection layer. By calculating the gradient contribution value of each input feature, different weights are dynamically assigned to each feature.
[0046] A multi-head self-attention mechanism extension module is introduced in the self-attention mechanism layer. By computing the relationship between different feature subsets in parallel, each attention head focuses on different aspects of the battery state and assigns different weights. Each head independently performs feature interaction during training, and finally the outputs of multiple heads are merged.
[0047] A differentiable decision tree module is introduced at the decision layer. Nodes and splitting rules are learned through a neural network. The information gain algorithm is used to select the most relevant features at each node for splitting. Each leaf node outputs a predicted value of the future state of the battery. The output of the leaf node is processed by a weighted average method of the neural network layer. An ensemble learning approach is adopted to combine multiple decision trees. The outputs of each tree are fused by a weighted average method. The depth of the tree is dynamically adjusted according to the complexity of the battery state. The complexity is evaluated by calculating the information gain or the Gini index, and the depth is optimized by cross-validation.
[0048] By introducing skip connections, certain intermediate layers are skipped during the input data transmission process, ensuring the stable propagation of information during training.
[0049] Hierarchical residual connections are added between network layers, and the output of each layer is the calculation result of the current layer plus the input information of the previous layer;
[0050] The model was trained using the Adam optimizer and gradient descent, and the network weights were updated by calculating the mean squared error between the output of the improved TabNet model and the actual battery state.
[0051] The beneficial effects of this invention are:
[0052] This invention improves prediction accuracy by employing a multi-layer feature selection module and an adaptive weighting mechanism to dynamically select features most relevant to changes in battery health status. This is particularly effective when handling complex battery state data (such as temperature fluctuations and changes in charge / discharge rates), effectively capturing nonlinear changes and complex relationships within the battery. Secondly, by utilizing an improved TabNet model and a differentiable decision tree module, this invention not only enhances the model's adaptability, enabling it to adjust prediction strategies in real-time according to different battery operating environments and health states, but also improves the model's robustness and computational efficiency, meeting the real-time, accurate prediction requirements of battery management systems. Finally, combined with the dynamic update mechanism of incremental neural networks, this invention can correct the battery state estimate in real-time after each charge / discharge cycle, ensuring the continuous accuracy of the prediction results and providing precise charge / discharge and temperature control strategy support for battery management systems. Overall, this invention provides a highly efficient, accurate, and dynamically adaptive battery state prediction method for battery management systems, effectively extending battery life and improving battery performance. Attached Figure Description
[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0054] Figure 1 This is a flowchart of a battery state prediction method based on lithium manganese oxide characteristic modeling proposed in this invention;
[0055] Figure 2 This is a schematic diagram of the improved SOCNet model in the battery state prediction method based on lithium manganese oxide characteristic modeling proposed in this invention.
[0056] Figure 3 This is a schematic diagram of the improved TabNet model in the battery state prediction method based on lithium manganese oxide characteristic modeling proposed in this invention. Detailed Implementation
[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0058] refer to Figure 1-3 A battery state prediction method based on lithium manganese oxide characteristic modeling includes the following steps:
[0059] Real-time data collection of battery voltage, current, temperature, and charge / discharge cycle count is performed and preprocessed to obtain preprocessed data;
[0060] Based on the electrochemical characteristics of lithium manganese oxide batteries, an improved SOCNet model was established, and the preprocessed data was input into the corresponding sub-model to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery.
[0061] Based on the estimated remaining power, estimated internal resistance, health status assessment, and estimated capacity decay, a weighted fusion model is used to calculate the comprehensive health status assessment of the battery.
[0062] Based on the comprehensive health status assessment value, the charging and discharging strategy of the battery is optimized by dynamically adjusting the charging and discharging rate, temperature control strategy and multi-objective optimization algorithm.
[0063] The improved SOCNet model is dynamically updated after each charge and discharge cycle using an incremental neural network, and the updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are obtained in real time.
[0064] Based on updated estimates of remaining charge, internal resistance, health status, and capacity decay, the future state of the battery is predicted using an improved TabNet model.
[0065] In this embodiment, the real-time acquisition of battery voltage, current, temperature, and charge / discharge cycle count data, followed by preprocessing to obtain preprocessed data, specifically includes:
[0066] Battery data is collected in real time by sensors, including battery voltage, current, temperature and charge / discharge cycle count data. The sensors include a voltage sensor, a current sensor, a temperature sensor and a cycle counter. The voltage sensor collects the battery terminal voltage, the current sensor collects the battery charge / discharge current, the temperature sensor collects the battery operating temperature, and the cycle counter records the number of charge / discharge cycles of the battery.
[0067] The battery data is time-series aligned, which includes aligning voltage, current, temperature and charge / discharge cycle count data according to timestamps to ensure that the data collected by different sensors are synchronized in time, and interpolating missing values. The interpolation methods include linear interpolation or spline interpolation to ensure data integrity and continuity.
[0068] Battery data is resampled and normalized. The resampling is performed on the collected data periodically according to the battery working state, either uniformly or non-uniformly, to ensure that there are enough sample points in each time period to improve the model prediction accuracy. The normalization method includes converting the battery data into values between 0 and 1 to eliminate scale differences between data and ensure that the data is processed under the same standard.
[0069] Data augmentation techniques are employed to expand battery data, thereby improving the robustness of model training and prediction accuracy. Data augmentation methods include random perturbation of battery current and voltage characteristics, time window shifting, and time series data pruning, which simulate data changes under different working environments, enhance data diversity, and reduce overfitting.
[0070] In this embodiment, the step of establishing an improved SOCNet model based on the electrochemical characteristics of lithium manganese oxide batteries, and inputting preprocessed data into the corresponding sub-model to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery specifically includes:
[0071] An improved SOCNet model was constructed, with the battery's voltage, current, temperature, and charge-discharge cycle count data as input. After standardization, the data was input to the input layer. A fully connected layer was used for preliminary feature extraction. Each feature was standardized before being input into the model to ensure numerical consistency and eliminate the influence of different data units.
[0072] An electrochemical characteristic subnetwork is constructed. The voltage-current relationship, capacity decay, and dynamic changes of internal resistance of the battery are calculated through a physical model. The charge and discharge data of the battery are processed using a fully connected layer. The nonlinear characteristics of the battery's internal resistance and capacity decay are fitted using the ReLU activation function. The capacity decay of the battery under time-series data is predicted using an LSTM layer. The remaining capacity and health status of the battery are calculated.
[0073] A multi-task learning module is introduced, and a shared layer is used to extract common features of the battery. The shared layer uses a multilayer perceptron for feature extraction, and different tasks are assigned to independent sub-networks for processing. Remaining power prediction, health status assessment, and capacity decay prediction are processed by independent regression sub-networks. An output layer with a linear activation function is used to predict the remaining power, health status, and capacity decay values.
[0074] An adaptive learning rate adjustment mechanism is introduced, using the Adam optimization algorithm as the optimizer, and the learning rate is dynamically adjusted according to changes in battery state during training. The learning rate is automatically adjusted based on the learning rate scheduler to improve the convergence speed of the model and reduce instability during training.
[0075] A self-attention mechanism is introduced, which uses scaled dot product attention to calculate the similarity between battery state features and dynamically weight them. The battery's voltage, current, and internal resistance features are used as queries, keys, and values. The attention layer is used to weight the features and output them. The contribution of each feature to the prediction is adaptively adjusted through the self-attention mechanism.
[0076] An incremental learning module is added, using an online learning algorithm. Each time the battery state data is updated, the model fine-tunes the parameters based on the new data. By using the mini-batch gradient descent method, the model updates the weights after each charge-discharge cycle, ensuring that the model adapts to the state changes of the battery during long-term use.
[0077] A regularization term is added to the loss function. The regularization term uses L2 regularization to constrain the parameters of the improved SOCNet model, reducing the risk of overfitting. The regularization coefficient is selected by cross-validation to ensure that the model can make stable predictions under various operating conditions. The weighting coefficients in the loss function are adjusted according to different battery operating environments to avoid overfitting under specific conditions.
[0078] In this embodiment, the step of calculating the battery's comprehensive health status assessment value using a weighted fusion model based on the estimated remaining power, estimated internal resistance, health status assessment value, and estimated capacity decay value specifically includes:
[0079] The remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are fed into the weighted fusion model as input data. The input data undergoes feature extraction and linear transformation through a fully connected layer to output intermediate features.
[0080] The weighted fusion model assigns weights to each input data point, uses the Swish activation function to perform a non-linear transformation on each input data point, the Swish activation function performs weighted processing on the input data, and calculates the weighted value of each feature through the Swish activation function. The weighted value is dynamically adjusted based on the battery's historical data and current operating status.
[0081] The weighted remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are summed to obtain the battery's comprehensive health status assessment value. The weighted summation operation ensures that the contribution of each feature to the assessment value is consistent with its weight allocation.
[0082] The comprehensive health status assessment value is normalized to convert it to a uniform numerical range, thereby avoiding the impact of scale differences in different feature data on the model prediction results.
[0083] The normalized comprehensive health status assessment value is optimized using regression analysis, and the model parameters are optimized using the mean squared error loss function, thereby improving the accuracy of the assessment value and the stability of the model.
[0084] In this embodiment, the optimization of the battery's charging and discharging strategy based on a comprehensive health status assessment value, through dynamic adjustment of charging and discharging rates, temperature control strategies, and multi-objective optimization algorithms, specifically includes:
[0085] Based on the comprehensive health status assessment value, the charging and discharging rate and temperature control strategy of the battery are calculated. When the comprehensive health status assessment value is higher than the preset threshold, the charging rate is set to the maximum value. When the comprehensive health status assessment value is lower than the preset threshold, the charging rate is reduced to a smaller value to reduce damage to the battery. When the estimated internal resistance of the battery is greater than the preset value, the charging current is reduced to avoid overheating caused by the increase in internal resistance. When the estimated capacity decay exceeds the preset threshold, the temperature control strategy reduces the charging rate and starts the cooling system to maintain the battery temperature.
[0086] The charge / discharge rate and temperature control strategy are optimized by using a particle swarm optimization algorithm. By setting multiple objective functions, including maximizing charging efficiency, minimizing temperature rise, and extending battery life, each particle is iteratively calculated during the optimization process to obtain the optimal charge / discharge rate, charging current, charging time, and temperature control parameters. Based on the comprehensive health status assessment value and internal resistance estimation value, each parameter is dynamically adjusted to ensure that the battery efficiency, temperature control, and battery life are optimally balanced during the charge / discharge process.
[0087] The battery management system adjusts the charging current, discharging current, charging power, and temperature control mode in real time based on the optimized charge and discharge rate, charging current, charging time, and temperature control parameters. It monitors the battery's voltage, current, and temperature changes in real time and adjusts the battery's charge and discharge rate and temperature control strategy according to the optimization strategy to ensure that the battery temperature is always maintained within a safe range during the charging and discharging process, while maintaining the highest working efficiency.
[0088] In this embodiment, the step of dynamically updating the improved SOCNet model after each charge and discharge cycle using an incremental neural network to obtain updated estimates of remaining capacity, internal resistance, health status, and capacity decay in real time specifically includes:
[0089] The improved SOCNet model is dynamically updated by using an incremental neural network. After each charge and discharge cycle, the model is corrected and updated based on real-time collected battery status data to obtain updated estimates of remaining charge, internal resistance, health status, and capacity decay.
[0090] After each charge and discharge cycle, the battery voltage, current, temperature, charge and discharge duration, charging current, and discharge current are collected as incremental data and input into the improved SOCNet model. The model dynamically adjusts the internal weights of the improved SOCNet model by processing the incremental data.
[0091] The error between the output of the current improved SOCNet model and the actual measurement data is calculated. The weights in the improved SOCNet model are updated by backpropagating the error to gradually reduce the error and improve the accuracy of battery state estimation.
[0092] The network weights of the improved SOCNet model are optimized using stochastic gradient descent. The loss is calculated for each incremental data and the parameters of the improved SOCNet model are adjusted according to the gradient of the loss, thereby optimizing the battery state estimation and avoiding the computational overhead of retraining the entire network.
[0093] Through incremental learning, the improved SOCNet model corrects the remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate in real time, and provides feedback on the updated estimate at the end of each charge-discharge cycle.
[0094] In this embodiment, the prediction of the battery's future state based on updated remaining capacity estimates, internal resistance estimates, health status assessments, and capacity decay estimates, using an improved TabNet model, specifically includes:
[0095] The updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are input into the improved TabNet model.
[0096] The input data is standardized by using a robust scaler method to remove outliers and unify the scale of the input data, so that the input data can be processed at a uniform scale and avoid the impact of differences in feature units.
[0097] An adaptive weighted feature selection module is added to the feature selection layer. By calculating the gradient contribution value of each input feature, different weights are dynamically assigned to each feature to ensure that the most informative features are selected and processed first in each layer of the network. This method can automatically identify the features that are most helpful for battery state prediction, reduce the negative impact of redundant features on model performance, and improve the adaptability of the improved TabNet model to different battery operating conditions.
[0098] A multi-head self-attention mechanism extension module is introduced in the self-attention mechanism layer. By computing the relationship between different feature subsets in parallel, each attention head focuses on different aspects of the battery state and is assigned different weights. Each head independently performs feature interaction during training. Finally, the outputs of multiple heads are merged to ensure that the improved TabNet model can handle complex temporal features and nonlinear relationships in battery state data.
[0099] A differentiable decision tree module is introduced at the decision layer. Nodes and splitting rules are learned through a neural network, enabling end-to-end optimization of the tree structure during training. The information gain algorithm is used to select the most relevant features for splitting at each node. At each split, the differentiability of each splitting decision is calculated using a neural network, allowing the decision tree to be optimized through backpropagation and gradient descent during training. Each leaf node outputs a predicted future state of the battery. The output of the leaf nodes is processed by a weighted average method of the neural network layers, making the output of each leaf node smoother and more stable, avoiding overfitting caused by a single decision path. An ensemble learning approach is adopted to combine multiple decision trees, and the outputs of each tree are fused through a weighted average method. The ensemble learning process improves the prediction accuracy and model stability. The depth of the tree is dynamically adjusted according to the complexity of the battery state. The complexity is evaluated by calculating information gain or Gini index, and the depth is optimized through cross-validation to ensure that the model is neither overfitting nor oversimplifying under different battery states, thereby improving prediction accuracy.
[0100] By introducing skip connections, certain intermediate layers are skipped during the input data transmission process, ensuring the stable propagation of information during training. Skip connections effectively solve the gradient vanishing problem in deep networks, enhancing the training efficiency and predictive ability of the network.
[0101] By adding hierarchical residual connections between network layers, the output of each layer is the calculation result of the current layer plus the input information of the previous layer. This structure better preserves key feature information and prevents the loss of important information in the process of passing through layers. Especially in the prediction of long-term battery degradation, residual connections can effectively improve the prediction accuracy of battery health status. By introducing residual connections, it can be ensured that the network responds more accurately to the long-term degradation of the battery and the nonlinear changes during charging and discharging in the battery status prediction task, and improve the robustness of the model in the training process.
[0102] The model is trained using the Adam optimizer and gradient descent method. The network weights are updated by calculating the mean square error between the output value of the improved TabNet model and the actual battery state, gradually reducing the prediction error and ensuring that the model can accurately predict the remaining battery life, battery health, internal resistance change trend and capacity decay.
[0103] Example 1:
[0104] To verify the feasibility of this invention in practice, it was applied to battery state prediction and optimization control in an electric vehicle battery management system. This system collects real-time data on battery voltage, current, temperature, and charge / discharge cycle count, and combines this data with the improved SOCNet model proposed in this invention to predict the battery's remaining capacity, internal resistance, health status, and capacity decay, thereby achieving intelligent battery management and optimization.
[0105] In this embodiment, in order to ensure the long-term stable operation of the battery and extend its service life, the battery management system needs to monitor the battery status in real time and adjust the charging and discharging strategy, temperature control strategy, etc. according to the battery's health status.
[0106] The battery management system (BMS) of an electric vehicle includes multiple sensors to monitor key data such as battery voltage, current, temperature, and charge-discharge cycle count in real time. This data is transmitted wirelessly to the onboard computing unit and further input into a prediction module based on an improved SOCNet model. In this module, the battery's electrochemical characteristics (such as open-circuit voltage, voltage-current relationship, and charge-discharge curve characteristics) are first modeled to generate estimates of the battery's remaining capacity, internal resistance, state of health, and capacity decay.
[0107] Subsequently, the system uses the TabNet model to predict the future state of the battery based on real-time feedback battery status data. The TabNet model dynamically adjusts the weights of features based on historical input data (such as battery charge / discharge history and battery temperature), thereby more accurately predicting the remaining battery life, changes in battery health, internal resistance trends, and capacity decay rate. Based on these predictions, the battery management system optimizes charging and discharging strategies in real time, and adjusts temperature control strategies according to battery health to prevent the battery from operating in high or low temperature environments, thus ensuring maximum battery performance.
[0108] To verify the effectiveness of this invention, we conducted actual tests, collected data from a battery management system of an electric vehicle under different operating conditions, and input this data into the battery state prediction model based on this invention. Table 1 shows the battery data collected over a period of time, including battery voltage, current, temperature, number of charge-discharge cycles, and predicted values for remaining capacity, internal resistance, health status, and capacity decay.
[0109] Table 1 Battery State Prediction and Optimized Control Data Table
[0110] Charge and discharge cycle Voltage (V) Current (A) Temperature (°C) Number of charge-discharge cycles Estimated remaining battery capacity (Ah) Estimated internal resistance (Ω) Health status assessment value Capacity decay estimate 1 3.7 5 25 100 28.5 0.05 95% 2% 2 3.6 4.5 28 120 27 0.06 92% 4% 3 3.5 4.8 30 150 25.5 0.08 88% 6% 4 3.4 5.2 32 180 24 0.1 85% 8% 5 3.3 5.5 35 200 22.5 0.12 80% 10%
[0111] Analysis of the data in Table 1 shows that as the number of charge-discharge cycles increases, the battery voltage gradually decreases, the internal resistance continuously increases, the health status assessment value gradually decreases, and the capacity decay estimate gradually increases. This indicates that as the battery's usage time increases, its performance undergoes significant degradation. The battery status prediction method of this invention can accurately predict the future health status of the battery and adjust battery management strategies in a timely manner based on the prediction results, avoiding excessive battery degradation and extending battery life.
[0112] As shown in Table 1, the battery state prediction method of this invention enables the battery management system to acquire battery health status, remaining capacity, internal resistance, and capacity decay indicators in a timely and accurate manner, and dynamically adjust charging, discharging, and temperature control strategies based on these prediction results. For example, when the battery temperature rises, the system can predict the decline in battery health status in advance and then adopt appropriate temperature control strategies to avoid damage caused by overheating. Simultaneously, changes in the estimated internal resistance of the battery can also be captured in a timely manner, allowing the system to adjust the charging rate based on this information, reducing the negative impact of overcharging on the battery.
[0113] In summary, this invention, by employing improved SOCNet and TabNet models, accurately predicts battery state, which not only improves the prediction accuracy of the battery management system but also optimizes the battery's charging, discharging, and temperature control strategies in real time, effectively extending battery life and improving the operating efficiency and reliability of electric vehicles and other equipment.
[0114] This embodiment verifies the feasibility and superiority of the battery state prediction method based on the improved SOCNet and TabNet models. By accurately predicting the battery's health state and performance degradation trend in real time, the battery management system can effectively optimize charging and discharging strategies and temperature control strategies, thereby improving the overall battery performance, extending battery life, reducing unnecessary energy loss, and further promoting the intelligent and efficient development of electric vehicles and energy storage devices.
[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A battery state prediction method based on lithium manganese oxide characteristic modeling, characterized in that, Includes the following steps: Real-time data collection of battery voltage, current, temperature, and charge / discharge cycle count is performed and preprocessed to obtain preprocessed data; Based on the electrochemical characteristics of lithium manganese oxide batteries, an improved SOCNet model was established, and the preprocessed data was input into the corresponding sub-model to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery. Based on the estimated remaining power, estimated internal resistance, health status assessment, and estimated capacity decay, a weighted fusion model is used to calculate the comprehensive health status assessment of the battery. Based on the comprehensive health status assessment value, the charging and discharging strategy of the battery is optimized by dynamically adjusting the charging and discharging rate, temperature control strategy and multi-objective optimization algorithm. The improved SOCNet model is dynamically updated after each charge and discharge cycle using an incremental neural network, and the updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are obtained in real time. Based on updated estimates of remaining charge, internal resistance, health status, and capacity decay, the future state of the battery is predicted using an improved TabNet model.
2. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, The system collects real-time data on battery voltage, current, temperature, and charge / discharge cycle count, and preprocesses this data to obtain preprocessed data, specifically including: Battery data is collected in real time by sensors, including battery voltage, current, temperature and charge / discharge cycle count data. The sensors include a voltage sensor, a current sensor, a temperature sensor and a cycle counter. The voltage sensor collects the battery terminal voltage, the current sensor collects the battery charge / discharge current, the temperature sensor collects the battery operating temperature, and the cycle counter records the number of charge / discharge cycles of the battery. The battery data is time-series aligned, which includes aligning voltage, current, temperature and charge / discharge cycle count data according to timestamps, and interpolating missing values. The interpolation method includes linear interpolation or spline interpolation. The battery data is resampled and normalized. The resampling is performed on the collected data periodically according to the battery's working state, either uniformly or non-uniformly. The normalization method includes converting the battery data into values between 0 and 1. Data augmentation techniques are employed to expand battery data. These methods include random perturbation of the battery's current and voltage characteristics, time window shifting, and time-series data pruning.
3. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, Based on the electrochemical characteristics of lithium manganese oxide batteries, an improved SOCNet model is established, and preprocessed data is input into the corresponding sub-models to calculate the estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay of the battery. Specifically, this includes: An improved SOCNet model is constructed. The voltage, current, temperature and charge-discharge cycle number of the battery are input. After standardization, the data are input to the input layer. A fully connected layer is used for preliminary feature extraction. Each feature is then standardized and input to the model. An electrochemical characteristic subnetwork is constructed. The voltage-current relationship, capacity decay, and dynamic changes of internal resistance of the battery are calculated through a physical model. The charge and discharge data of the battery are processed using a fully connected layer. The nonlinear characteristics of the battery's internal resistance and capacity decay are fitted using the ReLU activation function. The capacity decay of the battery under time-series data is predicted using an LSTM layer. The remaining capacity and health status of the battery are calculated. A multi-task learning module is introduced, and a shared layer is used to extract common features of the battery. The shared layer uses a multilayer perceptron for feature extraction, and different tasks are assigned to independent sub-networks for processing. Remaining power prediction, health status assessment, and capacity decay prediction are processed by independent regression sub-networks. An output layer with a linear activation function is used to predict the remaining power, health status, and capacity decay values. An adaptive learning rate adjustment mechanism is introduced, using the Adam optimization algorithm as the optimizer, and the learning rate is dynamically adjusted according to changes in battery state during training. The learning rate is automatically adjusted based on the learning rate scheduler. A self-attention mechanism is introduced, which uses scaled dot product attention to calculate the similarity between battery state features and dynamically weight them. The battery's voltage, current, and internal resistance features are used as queries, keys, and values. The attention layer is used to weight the features and output them. The contribution of each feature to the prediction is adaptively adjusted through the self-attention mechanism. An incremental learning module is added, using an online learning algorithm. Each time the battery state data is updated, the model fine-tunes the parameters based on the new data. By using the mini-batch gradient descent method, the model updates the weights after each charge-discharge cycle. A regularization term is added to the loss function. The regularization term uses L2 regularization to constrain the parameters of the improved SOCNet model. The regularization coefficient is selected by cross-validation. The weighting coefficients in the loss function are adjusted according to different battery operating environments.
4. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, The comprehensive health status assessment value of the battery is calculated using a weighted fusion model based on the estimated remaining power, estimated internal resistance, health status assessment value, and estimated capacity decay value. Specifically, this includes: The remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are fed into the weighted fusion model as input data. The input data undergoes feature extraction and linear transformation through a fully connected layer to output intermediate features. The weighted fusion model assigns weights to each input data point, uses the Swish activation function to perform a non-linear transformation on each input data point, the Swish activation function performs weighted processing on the input data, and calculates the weighted value of each feature through the Swish activation function. The weighted value is dynamically adjusted based on the battery's historical data and current operating status. The weighted estimated remaining capacity, estimated internal resistance, health status assessment, and estimated capacity decay are summed to obtain the battery's comprehensive health status assessment. The comprehensive health status assessment values are normalized to convert them into a uniform numerical range. The normalized comprehensive health status assessment value was optimized using regression analysis, and the model parameters were optimized using the mean squared error loss function.
5. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, The optimization of the battery's charging and discharging strategy based on a comprehensive health status assessment value, through dynamic adjustment of charging and discharging rates, temperature control strategies, and multi-objective optimization algorithms, specifically includes: Based on the comprehensive health status assessment value, the charging and discharging rate and temperature control strategy of the battery are calculated. When the comprehensive health status assessment value is higher than the preset threshold, the charging rate is set to the maximum value. When the comprehensive health status assessment value is lower than the preset threshold, the charging rate is reduced to a smaller value to reduce damage to the battery. When the estimated internal resistance of the battery is greater than the preset value, the charging current is reduced. When the estimated capacity decay exceeds the preset threshold, the temperature control strategy reduces the charging rate and starts the cooling system. The charge / discharge rate and temperature control strategy are optimized by using a particle swarm optimization algorithm. By setting multiple objective functions, including maximizing charging efficiency, minimizing temperature rise, and extending battery life, each particle is iteratively calculated during the optimization process to obtain the optimal charge / discharge rate, charging current, charging time, and temperature control parameters. Based on the comprehensive health status assessment value and internal resistance estimation value, each parameter is dynamically adjusted. The charging current, discharging current, charging power, and temperature control system operating modes are adjusted in real time based on the optimized charging and discharging rates, charging current, charging time, and temperature control parameters.
6. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, The improved SOCNet model is dynamically updated after each charge and discharge cycle using an incremental neural network, and the updated remaining capacity estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are corrected in real time. Specifically, this includes: The improved SOCNet model is dynamically updated by using an incremental neural network. After each charge and discharge cycle, the model is corrected and updated based on real-time collected battery status data to obtain updated estimates of remaining charge, internal resistance, health status, and capacity decay. After each charge and discharge cycle, the battery voltage, current, temperature, charge and discharge duration, charging current, and discharge current are collected as incremental data and input into the improved SOCNet model. The model dynamically adjusts the internal weights of the improved SOCNet model by processing the incremental data. Calculate the error between the output of the current improved SOCNet model and the actual measurement data, and use the backpropagation algorithm to update the weights in the improved SOCNet model through error backpropagation; The network weights of the improved SOCNet model are optimized using the stochastic gradient descent method. The loss is calculated for each incremental data, and the parameters of the improved SOCNet model are adjusted according to the gradient of the loss. Through incremental learning, the improved SOCNet model corrects the remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate in real time, and provides feedback on the updated estimate at the end of each charge-discharge cycle.
7. The battery state prediction method based on lithium manganese oxide characteristic modeling according to claim 1, characterized in that, The prediction of the battery's future state, based on updated remaining capacity estimates, internal resistance estimates, health status assessments, and capacity decay estimates, is achieved using an improved TabNet model. Specifically, this includes: The updated remaining power estimate, internal resistance estimate, health status assessment value, and capacity decay estimate are input into the improved TabNet model. The input data is standardized by using a robust scaler method to remove outliers and unify the scale of the input data. An adaptive weighted feature selection module is added to the feature selection layer. By calculating the gradient contribution value of each input feature, different weights are dynamically assigned to each feature. A multi-head self-attention mechanism extension module is introduced in the self-attention mechanism layer. By computing the relationship between different feature subsets in parallel, each attention head focuses on different aspects of the battery state and assigns different weights. Each head independently performs feature interaction during training, and finally the outputs of multiple heads are merged. A differentiable decision tree module is introduced at the decision layer. Nodes and splitting rules are learned through a neural network. The information gain algorithm is used to select the most relevant features at each node for splitting. Each leaf node outputs a predicted value of the future state of the battery. The output of the leaf node is processed by a weighted average method of the neural network layer. An ensemble learning approach is adopted to combine multiple decision trees. The outputs of each tree are fused by a weighted average method. The depth of the tree is dynamically adjusted according to the complexity of the battery state. The complexity is evaluated by calculating the information gain or the Gini index, and the depth is optimized by cross-validation. By introducing skip connections, certain intermediate layers are skipped during the input data transmission process, ensuring the stable propagation of information during training. Hierarchical residual connections are added between network layers, and the output of each layer is the calculation result of the current layer plus the input information of the previous layer; The model was trained using the Adam optimizer and gradient descent, and the network weights were updated by calculating the mean squared error between the output of the improved TabNet model and the actual battery state.