Lithium battery state estimation method and system based on improved patch learning and TCN-BMLSTM
Through the improved patch learning method combined with TCN-BMLSTM and patch model, the accuracy and adaptability of lithium battery state estimation in the prior art are solved, and more efficient and accurate lithium battery state monitoring and prediction are achieved.
Patent Information
- Application Number
- CN202411927509.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-25
- Publication Date
- 2025-05-27
AI Technical Summary
The prior art is difficult to accurately and robustly estimate the status of lithium batteries, especially under complex structures and multiple data-driven methods, which are difficult to capture long-term dependencies and subtle changes, and are computationally costly.
The improved patch learning method is adopted to combine the TCN-BMLSTM global model and the TCN-BiGRU/SVR patch model to provide overall trend prediction through the global model, and to correct local errors using the patch model, dynamically update the patch location to improve model adaptability.
It improves the accuracy and robustness of lithium battery state estimation, enhances the model's adaptability and flexibility to new data, reduces calculation costs, and improves the predictive ability of lithium battery aging and environmental changes.
Smart Images
Figure CN120044398A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of battery optimization, and more specifically, relates to a lithium battery state estimation method and system based on improved patch learning and TCN-BMLSTM. Background Art
[0002] With the revolutionary changes driven by new lithium battery technologies, the structure of lithium batteries has become more complex, and it has become more difficult to establish an accurate, robust, and highly generalized model. Existing electrochemical models or equivalent circuit models may need to adapt to new lithium battery chemical compositions and structures, which usually requires a large amount of experimental data and calibration work. In addition, fusion model and data-driven methods often require detailed lithium battery physical or chemical models, which can make the entire system modeling and simulation process complex. At the same time, fusion methods usually involve complex mathematical calculations and a large number of parameter estimations, which can lead to relatively high computational costs. Although models can help understand the internal mechanisms of lithium batteries, their accuracy largely depends on the understanding of lithium battery physical processes and parameter settings. If the model settings are not accurate enough or not applicable to different types of lithium batteries, it may affect the accuracy of predictions.
[0003] Integrating multiple data-driven methods can effectively utilize the unique advantages of each algorithm, thereby significantly improving the accuracy and robustness of the model. When estimating the state of a lithium battery, the choice of which model depends on the characteristics of the data, the complexity of the problem, and the available computing resources. Traditional data-driven fusion models are difficult to capture long-term dependencies in time series data and have low utilization efficiency of input data features, making it difficult to more accurately capture and understand the subtle changes in the state of lithium batteries, and have poor generalization ability and self-adaptability under drastic environmental changes.
[0004] The battery management system (Battery Management System, BMS) is the core of battery safety and efficient operation. Among them, accurate estimation of the state of charge (State Of Charge, SOC) and state of health (State Of Health, SOH) of the battery is crucial for ensuring reliability. Data-driven methods do not require in-depth knowledge of the physical or chemical models of the battery and can adapt to various types and brands of batteries, showing considerable potential in dealing with SOC and SOH estimations with large dataset scales and complex nonlinear problems. Summary of the Invention
[0005] In view of the above defects or improvement requirements of the prior art, the present invention provides a lithium battery state estimation method and system based on improved patch learning and TCN-BMLSTM. By combining a global model (such as TCN-BMLSTM) and a patch model (such as TCN-BiGRU and SVR), this solution can more accurately estimate the state of the lithium battery, including SOC (State of Charge) and SOH (State of Health). The global model provides a prediction of the overall trend, while the patch model corrects the local errors under specific regions or conditions. This integrated learning method not only improves the estimation accuracy, but also enhances the adaptability and flexibility of the model to new data through a dynamic patch update mechanism.
[0006] To achieve the above object, according to one aspect of the present invention, a lithium battery state estimation method based on improved patch learning and TCN-BMLSTM includes the following steps:
[0007] Step 1, collect battery data of the battery management system, extract features that affect the battery health state from the battery data, and establish a data set. Divide the data set into a training set and a test set according to a ratio.
[0008] Step 2, construct a TCN-BMLSTM global model, a TCN-BiGRU patch model, and an SVR patch model.
[0009] Step 3, use the data set to train the global model, the TCN-BiGRU patch model, and the SVR patch model, and search for and determine the patch position according to the model error.
[0010] Step 4, use the test set and the selected patch model to test the global model, and update the global model position or the patch position.
[0011] Step 5, use the updated global model to output the estimated values of the battery's SOC and SOH.
[0012] As a further preference, Step 1 specifically includes: extracting health features according to the IC curve of the battery voltage and current and the overall change trend of the charging voltage curve with battery aging, collecting battery data, using the principal component analysis method to identify important influence parameters of the battery SOH condition, and establishing a data set based on this parameter. Divide the data set into a training set and a test set according to a ratio.
[0013] As a further preference, in Step 2, the TCN-BMLSTM global model includes:
[0014] Bidirectional Mogrifier-LSTM, used to process time series data and capture forward and backward time dependencies;
[0015] The TCN layer increases the receptive field through dilated convolution to handle long-term dependencies;
[0016] Among them, the TCN layer is combined with the bidirectional Mogrifier-LSTM to form a hybrid model.
[0017] As a further preference, in each long short-term memory network unit of the bidirectional Mogrifier-LSTM, the input of this unit and the output of the previous unit are interactively processed. Among them, the number of interactions r determines the frequency of interaction between h n-1 and x n . When i is the iteration index of the number of interactions, i = 1, 2... n, when i is odd, update x n ; when i is even, update h n-1 , and the update process is as follows.
[0018] When i is odd,
[0019] When i is even,
[0020] Among them, is the output of the previous node, is the input of the current node, σ g is the Sigmoid activation function, Q i and R i are weight matrices respectively.
[0021] As a further preference, the bidirectional Mogrifier-LSTM includes two unidirectional deformed long short-term memory network layers with opposite directions. The two unidirectional deformed long short-term memory network layers are symmetrically arranged in structure, and the two unidirectional deformed long short-term memory network layers receive the same input, have opposite information transmission directions, and have different weight and bias parameters.
[0022] As a further preference, the information transfer process in the reverse deformed long short-term memory network layer is as follows:
[0023]
[0024] In the formula, is the state and output of the forward deformed long short-term memory network layer, h t ′ is the state and output of the reverse deformed long short-term memory network layer, z t ′ is the output of the update gate representing at time step t, h t ′ +1 is the predicted value of the hidden state at time step t + 1;
[0025] Finally, the output H of the bidirectional deformable long short-term memory network layer t concatenates h t and h t ′, or reduces the dimension of the hidden state to:
[0026]
[0027] where H t is the output of the bidirectional deformable long short-term memory network layer.
[0028] As a further preference, step three includes the following steps:
[0029] (31) Train the global model using the dataset and calculate the global model error
[0030] (32) Train the TCN-BiGRU patch model using the dataset and calculate the TCN-BiGRU patch model error
[0031] (33) Train the SVR patch model using the dataset and calculate the SVR patch model error
[0032]
[0033] (34) Determine the specific location where the patch needs to be applied by comparing the global model error with the absolute value of the error of the patch model in each region and .
[0034] As a further preference, step (34) includes the following steps:
[0035] (341) Set an update threshold For each region, if is less than and and are both less than then it is considered that the patch needs to be applied to this region to improve the prediction accuracy, and enter step (342); otherwise, retain the use of the global model for estimation;
[0036] (342) If in this region is less than then select the TCN-BiGRU patch model as the patch; otherwise, select the SVR patch model as the patch.
[0037] 9. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 7, characterized in that step four includes the following steps:
[0038] (41) After determining all the areas where patches need to be applied, the corresponding areas will be estimated using the selected patches, while other areas will continue to use the global model;
[0039] (42) Use the test set to estimate the model processed in step (41), and calculate the absolute value of the error between the estimated results of the global model and the patches on the test set and the actual values and
[0040] (43) Set a threshold dynamically determine whether to retain the global model or update it to a patch,
[0041] For the area of the global model: If the performance of the global model is simultaneously worse than that of the patch and its error is higher than the threshold then update it to the patch with the minimum error, otherwise keep using the global model,
[0042] For the patch area: If during the test and are not both greater than then this position continues to remain as a patch area, and use the patch with a smaller absolute value of the corresponding error, otherwise use the global model.
[0043] According to another aspect of the invention, there is also provided a lithium battery state estimation system based on improved patch learning and TCN - BMLSTM, including:
[0044] The first main control module is used to collect battery data of the battery management system, extract features that affect the battery health state from the battery data, and establish a data set, and divide the data set into a training set and a test set according to a ratio;
[0045] The second main control module is used to construct a TCN - BMLSTM global model, a TCN - BiGRU patch model, and an SVR patch model;
[0046] The third main control module is used to train the global model, the TCN - BiGRU patch model, and the SVR patch model using the data set, and search for and determine the patch positions according to the model errors;
[0047] The fourth main control module is used to test the global model using the test set and the selected patch model, and update the global model position or the patch position;
[0048] The fifth main control module is used to output the SOC and SOH estimation values of the battery using the updated global model.
[0049] Generally speaking, compared with the prior art, the above technical solution conceived by the present invention mainly has the following technical advantages:
[0050] 1. The present invention combines the advantages of a temporal convolutional network in capturing long-term dependencies and the powerful ability of a bidirectional deformable long short-term memory network in learning complex time series correlations, enhancing the adaptability of the model to lithium batteries under abnormal working conditions and different environmental conditions. The improved patch learning framework precisely reduces local errors by introducing patches specifically for regions with large model errors. In this framework, a novel patch position search strategy is proposed, which can effectively locate the regions that are most critical for improving model performance. In addition, the patch update mechanism is optimized to ensure that the patches can adapt dynamically as new data is added, thereby enhancing the robustness of the model when facing changes. Specifically, the global model of the present invention effectively captures the long-term dependencies in time series data, can consider both the forward and backward dependencies of time series data simultaneously; enhances the utilization efficiency of input data features, and can more accurately capture and understand the subtle changes in the state of lithium batteries; has strong generalization ability and self-adaptability, and is suitable for predictions under lithium battery aging, environmental temperature changes, and different working conditions; can identify and correct local errors at a fine-grained level, and weaken the greater negative impact generated in practical applications.
[0051] 2. The present invention effectively captures the long-term dependencies and complex dynamics in time series data by introducing various data-driven techniques and deep learning models, such as bidirectional Mogrifier-LSTM and TCN layers, during the training process. This structural design enables the model to better understand the battery behavior patterns, especially in the case of rapid changes in battery state. In addition, data dimensionality reduction is performed through correlation detection and PCA (principal component analysis), reducing information redundancy and improving the generalization ability of the model and the prediction accuracy for unknown data.
[0052] 3. The present invention provides an efficient battery state monitoring and prediction tool for the battery management system (BMS) by intelligently integrating the advantages of different models. This not only helps to monitor the health state of the battery in real time, but also can predict the aging process of the battery, thereby providing decision support for battery maintenance and management. By accurately analyzing the working performance of lithium batteries, this solution can extend the battery life, reduce the risk of battery failure, and improve the economy and safety of battery use. In addition, the dynamic adaptive characteristic of this solution enables the BMS to automatically adjust as the battery ages and the usage conditions change, maintaining its optimal performance state. Description of the Drawings
[0053] Figure 1 It is a schematic diagram of a lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to an embodiment of the present invention;
[0054] Figure 2 It is a schematic diagram of the overall change trend of the IC curve with battery aging involved in the embodiment of the present invention;
[0055] Figure 3 It is a schematic diagram of the overall change trend of the charging voltage curve with battery aging involved in the embodiment of the present invention;
[0056] Figure 4 It is the TCN-BMLSTM network structure diagram involved in the embodiment of the present invention;
[0057] Figure 5 It is a schematic diagram of improved patch learning for lithium battery state estimation involved in the embodiment of the present invention;
[0058] Figure 6 It is a flow chart of improved patch learning for lithium battery state estimation involved in the embodiment of the present invention. Detailed implementation manners
[0059] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0060] Embodiment 1
[0061] A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM provided by the embodiment of the present invention combines the advantages of a temporal convolutional network for capturing long-term dependencies and the powerful ability of a bidirectional deformable long short-term memory network to learn complex time series correlations, enhancing the adaptability of the model to lithium batteries under abnormal working conditions and different environmental conditions. The improved patch learning framework precisely reduces local errors by introducing patches specifically for regions with large model errors. In this framework, a novel patch position search strategy is proposed, which can effectively locate the regions that are most critical for improving model performance. In addition, the patch update mechanism is optimized to ensure that the patches can dynamically adapt as new data is added, thereby enhancing the robustness of the model when facing changes. Specifically as follows:
[0062] Step 1: Collect battery data from the battery management system, extract features that affect the battery health state from the battery data, establish a data set, and divide the data set into a training set and a test set according to a ratio.
[0063] In this step, health features are extracted according to the IC curves of battery voltage and current and the overall change trend of the charging voltage curve with battery aging. Battery data is collected, and the principal component analysis method is used to identify the important influencing parameters of the battery SOH status. Based on these parameters, a data set is established, and the data set is divided into a training set and a test set according to a certain proportion.
[0064] Step 2: Construct a TCN-BMLSTM global model, a TCN-BiGRU patch model, and an SVR patch model.
[0065] In this step, the TCN-BMLSTM global model includes: a bidirectional Mogrifier-LSTM, which is used to process time series data and capture the forward and backward time dependencies; a TCN layer, which increases the receptive field through dilated convolution to process long-term dependencies; among them, the TCN layer is combined with the bidirectional Mogrifier-LSTM to form a hybrid model.
[0066] In each long short-term memory network unit of the bidirectional Mogrifier-LSTM, the input of this unit and the output of the previous unit are interactively processed. Among them, the number of interactions r determines the interaction frequency between h n-1 and x n . When i is the iteration index of the number of interactions, i = 1, 2... n. When i is odd, x n is updated; when i is even, h n-1 is updated. The update process is as follows.
[0067] When i is odd,
[0068] When i is even,
[0069] Among them, is the output of the previous node, is the input of the current node, σ g is the Sigmoid activation function, Q i and R i are weight matrices respectively.
[0070] The bidirectional Mogrifier-LSTM includes two unidirectional deformed long short-term memory network layers with opposite directions. The two unidirectional deformed long short-term memory network layers are symmetrically arranged in structure, and the two unidirectional deformed long short-term memory network layers receive the same input, have opposite information transmission directions, and have different weight and bias parameters.
[0071] The information transfer process in the reverse deformed long short-term memory network layer is as follows:
[0072]
[0073] In the formula, is the state and output of the forward deformation long short-term memory network layer, h t ′ is the state and output of the reverse deformation long short-term memory network layer, z t ′ is the output of the update gate at time step t, h t ′ +1 is the predicted value of the hidden state at time step +1;
[0074] Finally, the output H of the bidirectional deformation long short-term memory network layer t concatenates h t and h t ′, or reduces the dimension of the hidden state to:
[0075]
[0076] In the formula, H t is the output of the bidirectional deformation long short-term memory network layer.
[0077] Step 3: Train the global model, the TCN-BiGRU patch model, and the SVR patch model using the said dataset, and search for and determine the patch positions according to the model errors.
[0078] In this step, train the global model using the said dataset and calculate the global model error Train the SVR patch model using the said dataset and calculate the SVR patch model error By comparing the global model error with the absolute values of the errors of the patch models in each region and to determine the specific positions where patches need to be applied.
[0079] More specifically, set an update threshold For each region, if is less than and and are both less than then it is considered that a patch needs to be applied to this region to improve the prediction accuracy. If in this region is less than then select the TCN-BiGRU patch model as the patch; otherwise, select the SVR patch model as the patch. Otherwise, retain the use of the global model for estimation.
[0080] Step 4: Test the global model using the test set and the selected patch model, and update the global model position or the patch position.
[0081] After determining all the areas where the patch needs to be applied, the corresponding areas will be estimated using the selected patch, while other areas will continue to use the global model; the test set is used to estimate the model, and the absolute value of the error between the estimated results of the global model and the patch on the test set and the actual value is calculated. and Set a threshold Dynamically determine whether to retain the global model or update it to the patch. For the areas of the global model: if the performance of the global model is simultaneously worse than that of the patch and its error is higher than the threshold Then update it to the patch with the smallest error, otherwise keep using the global model. For the patch areas: if during the test and do not both exceed at the same time, then this position continues to remain as a patch area, and the patch with a smaller absolute value of the corresponding error is adopted, otherwise the global model is used.
[0082] Step Five: Use the updated global model to output the estimated values of the battery's SOC and SOH.
[0083] Embodiment 2
[0084] In this embodiment, by combining multiple data-driven technologies, a lithium battery state estimation method with deep learning as the core is developed, aiming to accurately analyze the working performance and aging process of lithium batteries. The main schematic diagram is as shown in the appendix Figure 1 as shown.
[0085] The battery box provides a stable temperature environment. The battery tester is used to apply specific charge and discharge cycles to the battery and accurately measure parameters such as current, voltage, and capacity during the battery's response process. The battery tester is a key device for obtaining battery performance data in the experiment. It can execute various test programs, such as cycle life tests, capacity tests, etc., providing direct experimental data for battery performance evaluation. The computer (host computer), as the control center of the experimental system, is responsible for instructing the battery tester to perform tests according to the preset program and collecting the data generated during the test. The computer is also responsible for preliminary processing and storage of the data, providing support for subsequent data analysis. The test software is a bridge connecting the computer and the battery tester. Users can set test parameters, start test programs, monitor test status, and analyze test results through the software interface. The test software usually has a friendly user interface and powerful data processing functions, and can further analyze, visually display, and generate reports on the collected data, greatly improving the efficiency and accuracy of the experiment.
[0086] Through systematic analysis of the aging test data, it is possible to extract battery health characteristics from complex data, which is a key step in evaluating battery performance and predicting its future state. The health characteristics focus on the behavior of the battery during a relatively stable charging phase, which provides valuable information about the battery's SOH. By analyzing the charging IC curve and voltage change curve in detail, key parameters characterizing the battery's SOH can be extracted.
[0087] The present invention uses the overall change trend of the lithium battery IC curve with aging. From Figure 2 it can be observed that as the battery ages, the two peaks on the IC curve gradually shift to the left, and at the same time, the height of the peaks decreases, showing a trend of becoming flatter. Most IC curves have two peaks. The intensity and position of the peaks can reveal the capacity decline due to the loss of lithium ions or active materials. From the perspective of the external characteristics of the battery, the IC peak can detect the ability of the battery to absorb energy at a specific potential. Extract the heights of the two more obvious IC peaks as health characteristics H1 and H2.
[0088] The curves of the charging voltage varying with time for different SOHs are shown in the appendix Figure 3 as shown. It can be seen that as the battery undergoes more charge-discharge cycles, during the voltage growth stage, the charging voltage at the same moment gradually increases. At the same time, the time required to reach a specific voltage value also gradually shortens. This trend also reflects the change in the electrochemical performance of the battery during repeated use. Extract the time required for the voltage to reach 2.5V and the voltage value after 20 minutes of CC charging as health characteristics H3 and H4.
[0089] To deeply understand the relationship between these parameters and the battery state, correlation detection and PCA (Principal Component Analysis) are used. Correlation detection helps researchers identify which parameters are strongly correlated with the battery SOH condition, while PCA realizes data dimensionality reduction by retaining the main feature information of the original data, effectively reducing information redundancy and training complexity. These data analysis techniques are not only crucial for understanding the aging mechanism of lithium batteries but also provide substantial help for developing more efficient battery management strategies and extending battery life.
[0090] After preparing the data, regarding the selection of hyperparameters for the neural network, due to the high dimensionality of the hyperparameter space, the process of finding the optimal hyperparameter configuration is both computationally intensive and complex. The non-convex hyperparameter space requires the optimization algorithm to have good exploration ability to avoid premature convergence to local optimal solutions, and at the same time, effective strategies are needed to explore the regions where potential better solutions are located. In this embodiment, a population optimization algorithm IDBO is adopted, aiming to enhance the exploration and exploitation capabilities during the hyperparameter search process.
[0091] The processed data is input into the TCN-BMLSTM network. The introduction of TCN is based on its excellent long-sequence data processing ability. Through the stacking of one-dimensional convolutional layers, TCN effectively captures the long-term dependencies in time-series data, which is crucial for understanding the behavior patterns of the battery. Subsequently, the fusion of bidirectional LSTM enables the model to consider both the forward and backward dependencies of time-series data simultaneously. This bidirectional structure provides a more comprehensive perspective on the temporal information for the model, enhancing the understanding and prediction capabilities of the battery state change trends. By capturing the subtle changes in time-series data, the bidirectional network helps improve the accuracy of state estimation, especially in the case of rapid battery state changes. Subsequently, the addition of the deformation module further optimizes the information processing ability of the model. The deformation module enhances the information interaction between modules by dynamically adjusting the input and hidden layer states of the network. This mechanism not only improves the model's ability to handle complex relationships but also enhances the utilization efficiency of the input data features, enabling TCN-BMLSTM to more accurately capture and understand the subtle changes in the battery state.
[0092] Finally, through an improved patch learning, an ensemble learning framework, the global model TCN-BMLSTM is integrated with other patches to reduce the local error in battery estimation. The improved patch learning framework adopts an innovative patch position search algorithm, effectively locking the key data regions to be optimized and accurately positioning the patches. At the same time, a dynamic patch update mechanism is adopted, enabling the model to adjust the patches based on the newly added data, further enhancing the adaptability and flexibility of the model.
[0093] That is, in this embodiment, the ensemble learning framework includes data acquisition and analysis, population hyperparameter optimization, deep learning network model, ensemble learning framework, etc. Regarding the input parameter selection problem of the SOH prediction model, this patent introduces how to extract key health features from the aging data, including the relevant parameters of the charging capacity increment curve and the voltage change curve. And the correlation detection and principal component analysis are carried out on the feature data.
[0094] Embodiment 3
[0095] Based on any one of the above embodiments or a combination of multiple embodiments, the fusion of the temporal convolutional and bidirectional deformable long short-term memory network structure is a brand-new combined structure, and the structure is as shown in the appendix Figure 4As shown. In this model, the input sequence is first preprocessed by a temporal convolutional layer, which effectively captures the long-term temporal dependencies in the sequence data using its unique dilated convolution and causal convolution mechanisms. This design enables the model to look into past information while maintaining computational efficiency, laying the foundation for in-depth analysis. Subsequently, the sequence data is fed into a bidirectional deformable long short-term memory network layer, which is the core of the model. The deformation mechanism significantly enhances the processing ability of the long short-term memory network units. By periodically alternatingly modifying the input and hidden states, the deformable long short-term memory network can understand and analyze the features extracted by the temporal convolutional layer at a deeper level. This mechanism not only enhances the model's ability to capture information but also improves the dynamics of data processing.
[0096] In the attached Figure 4 In each long short-term memory network unit, the input of this unit and the output of the previous unit are interactively processed. After such interactive updates, the two sets of data are fed into a standard long short-term memory network unit for further learning and updating. This mechanism significantly enhances the network's learning ability to capture the sequential relationships between data, thereby improving the model's processing efficiency and prediction accuracy for time series data. The number of interactions r determines the frequency of interaction between h n-1 and x n ; when i = 1, 2... (the iteration index of the number of interactions) and i is odd, x n is updated; when i is even, h n-1 is updated. The update process is as follows.
[0097] When i is odd,
[0098] When i is even,
[0099] where represents the output of the previous node, represents the input of the current node. σ g represents the Sigmoid activation function, and Q i and R i represent the associated weight matrices respectively.
[0100] In addition, the bidirectional structure further deepens the model's understanding of the context information of the sequence data, enabling it to consider both past and future data points simultaneously, thereby achieving more comprehensive and accurate sequence prediction. The integrated model not only improves the ability to identify the internal laws of time series data but also enables the model to understand and predict sequences at multiple scales, thus significantly enhancing the accuracy and reliability of the sequence prediction task.
[0101] The hidden layer of the bidirectional deformable long short-term memory network consists of two unidirectional deformable long short-term memory network layers with opposite directions. These two unidirectional deformable long short-term memory network layers are symmetric in structure. Although their information transmission directions are opposite, they receive the same input and have different weight and bias parameters. The states and outputs of each unit are respectively represented as and h t ′. The information transfer process in the reverse deformable long short-term memory network layer is as follows:
[0102]
[0103] Finally, the output H t of the bidirectional deformable long short-term memory network layer t and h t ′ can be concatenated, or the dimension of the hidden state can be reduced to:
[0104]
[0105] This design allows the bidirectional deformable long short-term memory network to learn both the forward and reverse information of the data simultaneously. Thus, when processing sequence data, it can capture context information more comprehensively and improve the model's ability to understand time series data.
[0106] Example 4
[0107] Based on any of the above embodiments or a combination of multiple embodiments, in the data-driven lithium battery state estimation method, the current research focus tends to minimize the root mean square error of the estimation, but often ignores individual local errors. The patch learning framework combines the overall prediction of the global model and the local correction of the patch to more accurately improve the estimation accuracy. Nevertheless, the accuracy of determining the patch position remains a challenge in the traditional patch learning framework. This study optimizes from the perspective of patch position search and update. Att Figure 5 shows the working schematic diagram of this method. Its key steps cover global model and patch training, patch position search and determination, and patch position update and test set verification.
[0108] The flowchart of patch learning for lithium battery state estimation is as shown in Att Figure 6 .
[0109] First, a global model TCN-BMLSTM based on the bidirectional deformable long short-term memory network is trained using the training set. This model aims to capture and learn the overall trends and patterns in the dataset. After training, the trained global model is tested to generate a series of estimated values. For each estimated value, it is defined as an independent region (the region range can be expanded when the data volume is large), and the absolute value of the error between the estimated value of the global model and the actual value within this region is calculated
[0110] Next, two types of patches, TCN-BiGRU and SVR, are established respectively. These two patches are designed to complement the possible deficiencies of the global model in specific local area predictions. The absolute value of the error between the estimated value and the actual value of the TCN-BiGRU and SVR models in this area is calculated respectively and are used to evaluate the accuracy of the patches in each area. By comparing the absolute values of the errors of the global model and the patches in each area and the specific locations where the patches need to be applied can be determined. Set an update threshold For each area, if is less than and and are both less than then it is considered that the patch needs to be applied in this area to improve the prediction accuracy. If in this area is less than then TCN-BiGRU is preferentially selected as the patch; otherwise, the SVR model is selected. If the above conditions are not met, the global model is retained for estimation.
[0111] After determining all the areas where the patches need to be applied, the corresponding areas will be estimated using the selected patches, while other areas continue to use the global model. To evaluate the effect of the initial patch model, it is used to estimate another test set, and the absolute values of the errors between the estimated results of the global model and the patches on the test set and the actual values are calculated and This step aims to quantify the generalization ability of the model on unseen data.
[0112] According to the test results, dynamically decide whether to retain the global model or update it to the patch through the set threshold For the areas of the global model: If the performance of the global model is simultaneously worse than that of the patch and its error is higher than the threshold then it is updated to the patch with the minimum error, otherwise the global model is retained. For the patch areas: If in the test and are not both greater than then this position continues to be a patch area, and the patch with a smaller absolute value of the corresponding error is adopted, otherwise the global model is used.
[0113] Through the above process, a dynamic adaptive patch model is established, which combines the overall prediction ability of the global model and the optimized performance of the TCN-BiGRU and SVR patches in specific regions, thus achieving high-precision estimation on the entire dataset. This integrated model demonstrates a new strategy for optimizing prediction results by intelligently fusing different models in complex datasets.
[0114] In this embodiment, aiming at the problem that the global optimization learning framework is not fine enough in dealing with local errors, this patent proposes an improved patch learning framework. By optimizing the search for patch positions and update strategies, the model can effectively locate the regions that are most critical for improving its performance and dynamically adapt as new data is added, thereby enhancing its robustness in the face of changes.
[0115] Embodiment 5
[0116] Based on any combination of the above embodiments, this embodiment provides a lithium battery state estimation system based on improved patch learning and TCN-BMLSTM for implementing the above method, including:
[0117] The first main control module is used to collect battery data from the battery management system, extract features that affect the battery health state from the battery data, establish a dataset, and divide the dataset into a training set and a test set according to a ratio;
[0118] The second main control module is used to construct a TCN-BMLSTM global model, a TCN-BiGRU patch model, and an SVR patch model;
[0119] The third main control module is used to train the global model, the TCN-BiGRU patch model, and the SVR patch model using the dataset, and search for and determine patch positions according to the model error;
[0120] The fourth main control module is used to test the global model using the test set and the selected patch model, and update the global model position or the patch position;
[0121] The fifth main control module is used to output the estimated values of the battery's SOC and SOH using the updated global model.
[0122] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included within the protection scope of the present invention.
Claims
1. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM, characterized in that: The following steps are involved: Step 1: Collect battery data from the battery management system, extract features that affect the battery health status from the battery data, and establish a data set, which is divided into a training set and a test set in proportion; Step 2: Build the TCN-BMLSTM global model, TCN-BiGRU patch model, and SVR patch model; Step 3, using the data set to train a global model, a TCN-BiGRU patch model, and an SVR patch model, and searching and determining patch positions based on model errors; Step 4: Use the test set and the selected patch model to test the global model, and update the global model position or the patch position; Step 5: Use the updated global model to output the estimated SOC and SOH of the battery.
2. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 1, characterized in that: Step one specifically includes: extracting health features based on the IC curve of battery voltage and current and the overall change trend of charging voltage curve with battery aging, collecting battery data, using principal component analysis method to identify important influencing parameters of battery SOH status, and establishing a data set based on the parameters, and dividing the data set into training set and test set in proportion.
3. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 1, characterized in that: In step 2, the TCN-BMLSTM global model includes: Bidirectional Mogrifier-LSTM, which is used to process time series data and capture both forward and reverse temporal dependencies; The TCN layer increases the receptive field by dilating convolution and handles long-term dependencies; Among them, the TCN layer is combined with the bidirectional Mogrifier-LSTM to form a hybrid model.
4. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 3, characterized in that: Each long short-term memory network unit of the bidirectional Mogrifier-LSTM will interact with the input of the unit and the output of the previous unit, where the number of interactions r determines h n-1 and x n The frequency of interaction between them, i is the iteration index of the number of interactions, i = 1, 2...n, when i is an odd number, update x n ; When i is an even number, update h n-1 , the update process is as follows, When i is an odd number, When i is an even number, in, is the output of the previous node, is the input of the current node, σ g is the Sigmoid activation function, Q i and R i are weight moments respectively.
5. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 4, characterized in that: The bidirectional Mogrifier-LSTM includes two unidirectional deformable long short-term memory network layers in opposite directions. The two unidirectional deformable long short-term memory network layers are structurally symmetrical, and the two unidirectional deformable long short-term memory network layers receive the same input, have opposite information transmission directions, and have different weights and bias parameters.
6. A lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 5, characterized in that: The information transmission process in the reverse deformation long short-term memory network layer is as follows: In the formula, is the state and output of the positive deformation long short-term memory network layer, h t ′ is the state and output of the reverse deformation long short-term memory network layer, z t ′ represents the output of the update gate at time step t, h t ' +1 is the predicted value of the hidden state at time step +1; Finally, the output H of the bidirectional deformation long short-term memory network layer t h t and h t ′ is concatenated, or the dimension of the hidden state is reduced to: In the formula, H t is the output of the bidirectional deformable long short-term memory network layer.
7. The lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 1, characterized in that: Step three includes the following steps: (31) Use the data set to train the global model and calculate the global model error (32) Use the data set to train the TCN-BiGRU patch model and calculate the TCN-BiGRU patch model error (33) Using the data set to train the SVR patch model, and calculating the SVR patch model error (34) By comparing the global model error The absolute value of the error with the patch model in each area and to determine the specific location where the patch needs to be applied.
8. The lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 7, characterized in that: Step (34) comprises the following steps: (341) Set an update threshold For each region, if Less than and and All less than It is considered that the region needs to be patched to improve the prediction accuracy, and the process goes to step (342). Otherwise, the global model is retained for estimation. (342) If in this area Less than If the model is not good, the TCN-BiGRU patch model is selected as the patch; otherwise, the SVR patch model is selected as the patch.
9. The lithium battery state estimation method based on improved patch learning and TCN-BMLSTM according to claim 7, characterized in that: Step 4 includes the following steps: (41) After all the regions to which patches need to be applied are determined, the corresponding regions will be estimated using the selected patches, while other regions continue to use the global model; (42) Use the test set to estimate the model processed by step (41), and calculate the absolute value of the error between the estimated results of the global model and the patch on the test set and the actual value and (43) Setting the threshold Dynamically decide whether to keep the global model or update it as a patch, For a region of the global model: if the performance of the global model is worse than the patch at the same time, and its error is above a threshold Then update to the patch with the smallest error, otherwise keep using the global model. For patch areas: If in test and Not greater than , the patch area will continue to be maintained at that position, and the patch with a smaller absolute value of the corresponding error will be used, otherwise the global model will be used.
10. A lithium battery state estimation system based on improved patch learning and TCN-BMLSTM, characterized in that: include: The first main control module is used to collect battery data of the battery management system, extract features that affect the battery health status from the battery data, and establish a data set, and divide the data set into a training set and a test set in proportion; The second main control module is used to build the TCN-BMLSTM global model, the TCN-BiGRU patch model, and the SVR patch model; A third main control module is used to train a global model, a TCN-BiGRU patch model and an SVR patch model using the data set, and search and determine patch positions according to model errors; A fourth main control module is used to perform a global model test using the test set and the selected patch model, and update the global model position or the patch position; The fifth main control module is used to output the SOC and SOH estimation values of the battery using the updated global model.