A battery soc / sob joint prediction model training method and system based on dynamic physical constraint weight and subtask loss
Patent Information
- Application Number
- CN202611136215.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-10-13
AI Technical Summary
这些纯数据驱动方法虽然能够从电池循环测试数据中学习复杂时序规律，但其训练过程通常只关注预测值与标签之间的误差最小化，缺乏对电池物理机理的有效约束，容易出现以下问题：SOC/SOH预测值越出合理边界、SOH退化趋势出现异常回升、搁置阶段SOC发生突变、充放电工步方向不一致等
1）分任务损失提高任务可解释性：SOC和SOH物理意义不同，SOC随充放电快速变化，SOH随循环缓慢退化。本发明分别计算SOC和SOH损失，能够更清楚地分析并调节两类任务的学习效果。
Smart Images
Figure CN122904715A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of battery state estimation, battery management system, artificial intelligence model training and physical information neural network technology, and relates to a training method and system for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-specific losses. Background Technology
[0002] Lithium-ion batteries are widely used in electric vehicles, energy storage power stations, and mobile devices. With their advantages such as high energy density, long cycle life, and low self-discharge rate, they have become a core energy storage component in the new energy field. State of Charge (SOC) characterizes the percentage of remaining battery capacity, while State of Health (SOH) characterizes the degree of battery capacity degradation. Both are important parameters for battery management systems to estimate range, allocate energy, provide safety warnings, and assess battery life. Because battery operation is influenced by multiple factors, including current, voltage, temperature, strain, operating conditions, and cycle aging, joint SOC / SOH prediction exhibits significant nonlinearity, time-varying characteristics, and strong operating condition correlation.
[0003] In existing technologies, SOC / SOH estimation methods based on deep learning have been extensively studied. For example, Transformer-based energy storage battery life prediction algorithms utilize state monitoring data such as voltage, current, and temperature to mine battery health information; time-series Transformer-based feature extraction networks estimate SOH by extracting semantic information features from voltage, current, and temperature data. While these purely data-driven methods can learn complex time-series patterns from battery cycle test data, their training process typically focuses only on minimizing the error between the predicted value and the label, lacking effective constraints on the battery's physical mechanisms. This can easily lead to the following problems: SOC / SOH predicted values exceeding reasonable boundaries, abnormal rebound in SOH degradation trends, sudden changes in SOC during the storage phase, and inconsistent charging and discharging steps.
[0004] To address the aforementioned issues, recent research has focused on integrating physical information into neural networks. Physics-Informed Neural Networks (PINNs) combine physical principles with data-driven models, enabling them to effectively adhere to fundamental physical laws and boundary conditions, thus providing stronger interpretability and generalization capabilities. For example, existing research has embedded partial differential equations describing lithium-ion battery degradation into loss functions to constrain model training; other studies have proposed lithium-ion battery health state estimation methods that integrate physical information diffusion-causality Transformers, using a hybrid loss function combining prediction error and physical constraint loss to train the model. Furthermore, for joint estimation of State of Charge (SOC) and State of Health (SOH), existing methods employ multi-timescale estimation strategies, performing SOH estimation at a macroscopic scale and SOC estimation at a microscopic scale.
[0005] However, existing technologies still have the following shortcomings: 1) Imbalance between SOC and SOH task learning: SOC changes rapidly during charging and discharging, belonging to a dynamic process on the order of seconds or minutes; SOH degrades slowly with cyclic aging, belonging to a long-term process on the order of weeks or months. The two have different physical meanings and significantly different scales of change. Existing methods usually supervise SOC and SOH as a whole, using a single mean squared error loss to constrain both types of tasks simultaneously, making it difficult for the model to meet the learning needs of tasks with rapid changes in state of charge and slow degradation of health status. 2) Lack of sequence-level task sub-supervision: Most existing methods only focus on the prediction accuracy of the final time step, ignoring the model's ability to learn the trend of state changes throughout the entire time window. Since battery state changes have strong temporal continuity, supervising only the final output is insufficient to guarantee the temporal consistency and physical rationality of the prediction sequence. 3) Fixed physical constraint weights lead to training instability: Although existing methods introduce physical constraint losses to enhance the physical consistency of model outputs, the physical constraint weights of these methods are usually fixed. If strong physical constraints are applied in the early stages of training, the model may not have learned the basic data patterns, leading to optimization difficulties or convergence to local optima. Conversely, if the weights of physical constraints are insufficient in the later stages of training, it is difficult to effectively enhance physical consistency. Existing technologies lack a mechanism for dynamically adjusting the weights of physical constraints as the training process progresses. 4) Lack of systematic optimization of training strategies: Existing methods lack a systematic training strategy design that coordinates with the dynamic scheduling of physical constraints in aspects such as learning rate scheduling and overfitting prevention, making it difficult to guarantee the stability and generalization performance of model training.
[0006] Therefore, there is a need for an optimization method for the loss function and training strategy of Transformer models that fuse structural physical information. Without changing the dataset or the main model structure, this method should improve the model's training stability, prediction accuracy, and physical consistency through the systematic combination of task-based loss design, dynamic physical weight scheduling, learning rate scheduling, and early stopping mechanism. Summary of the Invention
[0007] In view of this, the purpose of this invention is to provide a training method and system for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-specific losses. First, the battery cycle test data is read, time-series analyzed, process-by-step encoded, labeled, and physically derived features constructed. Second, fixed-length time-series samples of the same cycle are constructed and input into the PIT-v2 model. Third, the final time-step supervision loss is split into SOC data supervision loss and SOH data supervision loss, and the sequence supervision loss is split into SOC sequence supervision loss and SOH sequence supervision loss. Then, the physical constraint loss weights are dynamically scaled according to the current training round. Finally, model training is completed by combining cosine learning rate scheduling and an early stopping mechanism.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A training method for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-specific losses, the method specifically includes the following steps: S1. Obtain battery cycle test data, which includes cycle number, step type, total time, current, voltage, capacity, first strain signal V1, second strain signal V2, first temperature T1 and second temperature T2. S2. Preprocess the battery cycle test data to generate time in seconds, step code, whether to put aside, time interval between adjacent sampling points and time interval within the same cycle, and construct SOC and SOH tags. S3. Construct physical derived features based on the preprocessed data, and combine the original sensor features, process time features, and physical derived features into model input features; S4. Construct a fixed-length time series sample within the same loop, input the time series sample into the Transformer model that fuses structural-level physical information, and obtain the SOC / SOH prediction sequence and the final time step SOC / SOH prediction value. S5. Calculate the final time step supervision loss of SOC and the final time step supervision loss of SOH respectively, and construct the sub-task data supervision loss based on the two. S6. Calculate the SOC sequence supervision loss and the SOH sequence supervision loss respectively, and construct the sub-task sequence supervision loss based on the two. S7. Calculate the dynamic physical constraint weight scaling factor based on the current training round, and use the dynamic physical constraint weight scaling factor to adjust the effective weights of multiple physical constraint loss terms. S8. The total training loss is obtained by weighting and summing the task-specific data supervision loss, the task-specific sequence supervision loss, and the dynamically weighted physical constraint loss. S9. Perform backpropagation based on the total training loss, and train the battery SOC / SOH joint prediction model by combining learning rate scheduling and early stopping mechanism.
[0009] Furthermore, in step S3, the physical derived features include absolute current value, power, average temperature, temperature difference, average strain, strain difference, ampere-hour integral increment, and stress factor.
[0010] Furthermore, in step S4, the structure-level physical information fusion Transformer model includes a data encoder, a physical encoder, a gating fusion module, and a state update output module; in step S5, the final time step supervision loss of SOC is the mean square error between the SOC predicted value and the SOC label, the final time step supervision loss of SOH is the mean square error between the SOH predicted value and the SOH label, and the sub-task data supervision loss is the weighted sum of the final time step supervision loss of SOC and the final time step supervision loss of SOH.
[0011] Further, in step S6, the SOC sequence supervision loss is the mean square error between the SOC predicted sequence and the SOC true sequence, and the SOH sequence supervision loss is the mean square error between the SOH predicted sequence and the SOH true sequence; the sub-task sequence supervision loss is the weighted sum of the SOC sequence supervision loss and the SOH sequence supervision loss.
[0012] Furthermore, in step S7, the physical constraint loss includes SOC / SOH boundary constraint loss, SOH monotonic degradation loss, SOC ampere-hour integral consistency loss, idle state SOC stability loss, step direction consistency loss, temperature-strain degradation loss, and sequence smoothing loss; the dynamic physical constraint weight scaling factor takes a smaller value in the early stage of training, gradually increases in the middle stage of training, and reaches a preset maximum value in the later stage of training; the dynamic physical constraint weight scaling factor includes a warm-up stage, a climbing stage, and a stabilization stage; in the warm-up stage, the physical constraint weight is maintained at the minimum proportion of the basic weight; in the climbing stage, the physical constraint weight gradually increases from the minimum proportion to the maximum proportion; in the stabilization stage, the physical constraint weight is maintained at the maximum proportion.
[0013] Furthermore, in step S9, the learning rate scheduling adopts cosine annealing scheduling, so that the learning rate is gradually reduced from the initial learning rate to the preset minimum learning rate; the early stopping mechanism is used to stop training when the verification loss has not produced an improvement greater than a preset threshold for multiple consecutive training rounds.
[0014] Furthermore, the training history stores the training epochs, learning rate, dynamic physical weight scaling factor, SOC data supervision loss, SOH data supervision loss, SOC sequence supervision loss, SOH sequence supervision loss, physical constraint loss, total training loss, total validation loss, and number of no-boost epochs. The total training loss is expressed as the sum of the data supervision loss, sequence supervision loss, and dynamic physical constraint loss, wherein the dynamic physical constraint loss is obtained by multiplying each physical constraint loss by its dynamic effective weight and then summing the results.
[0015] The present invention also provides a training system for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-based loss. The system includes a data preprocessing module, a physical feature construction module, a time-series sample construction module, a model prediction module, a task-based supervision loss calculation module, a dynamic physical weight scheduling module, a joint loss calculation module, a learning rate scheduling module, an early stopping control module, and a model saving module; the system is configured to execute the method described above.
[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.
[0017] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.
[0018] The beneficial effects of this invention are as follows: 1) Task-specific loss improves task interpretability: SOC and SOH have different physical meanings. SOC changes rapidly with charging and discharging, while SOH degrades slowly with cycling. This invention calculates SOC and SOH losses separately, enabling a clearer analysis and adjustment of the learning effects of the two types of tasks.
[0019] 2) Sequence-level task supervision enhances temporal consistency: This invention not only supervises the final time step prediction, but also supervises the SOC prediction sequence and the SOH prediction sequence separately, enhancing the model's ability to learn the trend of state changes within the entire window.
[0020] 3) Dynamic physical weights improve training stability: physical constraints are weakened in the early stages of training and gradually strengthened in the middle and later stages of training, so that the model learns the data patterns first and then strengthens physical consistency.
[0021] 4) Learning rate scheduling improves convergence stability: By using cosine learning rate scheduling, the learning rate is gradually reduced in the later stages of training, thus reducing loss oscillations.
[0022] 5) Early stopping mechanism reduces the risk of overfitting: Training is stopped when the validation loss does not improve significantly over a long period of time to avoid ineffective training and overfitting.
[0023] 6) Complete training history facilitates engineering analysis and patent implementation: The training history saves the learning rate, dynamic physical weights, task-specific loss and no-boost rounds, which is beneficial for reproducing experiments and optimizing models.
[0024] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the training method based on dynamic physical constraint weights and task-specific loss. Figure 2 This is a schematic diagram of the task-specific loss function, including SOC data loss, SOH data loss, SOC sequence loss, and SOH sequence loss; Figure 3 This is a schematic diagram of dynamic physical constraint weight scheduling, including the warm-up phase, the climb phase, and the stabilization phase. Figure 4 This is a schematic diagram of the overall structure of the training strategy, including task-specific loss, dynamic physical constraints, learning rate scheduling, and early stopping mechanism. Detailed Implementation
[0026] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0027] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures, and should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0028] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0029] like Figure 1 As shown, the battery SOC / SOH joint prediction model training method provided by this invention based on dynamic physical constraint weights and task-specific losses includes the following steps: S1. Obtain battery cycle test data, which includes cycle number, step type, total time, current, voltage, capacity, first strain signal V1, second strain signal V2, first temperature T1 and second temperature T2. S2. Preprocess the battery cycle test data to generate time in seconds, step code, whether to put aside, time interval between adjacent sampling points and time interval within the same cycle, and construct SOC and SOH tags. S3. Construct physical derived features based on the preprocessed data, and combine the original sensor features, process time features, and physical derived features into model input features; S4. Construct a fixed-length time series sample within the same loop, input the time series sample into the Transformer model that fuses structural-level physical information, and obtain the SOC / SOH prediction sequence and the final time step SOC / SOH prediction value. S5. Calculate the final time step supervision loss of SOC and the final time step supervision loss of SOH respectively, and construct the sub-task data supervision loss based on the two. S6. Calculate the SOC sequence supervision loss and the SOH sequence supervision loss respectively, and construct the sub-task sequence supervision loss based on the two. S7. Calculate the dynamic physical constraint weight scaling factor based on the current training round, and use the dynamic physical constraint weight scaling factor to adjust the effective weights of multiple physical constraint loss terms. S8. The total training loss is obtained by weighting and summing the task-specific data supervision loss, the task-specific sequence supervision loss, and the dynamically weighted physical constraint loss. S9. Perform backpropagation based on the total training loss, and train the battery SOC / SOH joint prediction model by combining learning rate scheduling and early stopping mechanism.
[0030] This embodiment uses battery cycle test data as input. The data fields include cycle number, step type, total time, current (A), voltage (V), capacity (Ah), V1 (V), V2 (V), T1 (°C) and T2 (°C).
[0031] Example 1: Data preprocessing and sample construction.
[0032] The battery cycle test data is read, and the total time is parsed into seconds. A step code is generated based on the step type, where charging is 1, discharging is -1, and resting is 0. A resting flag is generated, and delta_t_s and cycle_delta_t_s are calculated. A SOC tag is constructed based on capacity and rated capacity, and a SOH tag is constructed based on maximum capacity and rated capacity within the same cycle. Subsequently, physical derived features such as absolute current, power, average temperature, temperature difference, average strain, strain difference, ampere-hour integral increment, and stress factor are constructed, and these features are normalized together with the original features to form the model input.
[0033] Example 2: Structure-level physical information fusion Transformer prediction.
[0034] Fixed-length time series samples are input into the PIT-v2 model. The model includes a Data Encoder, a Physics Encoder, a Gate Fusion, and a State Update Head. The Data Encoder is used to extract data-driven representations, the Physics Encoder is used to extract physical mechanism representations, the Gate Fusion is used to fuse the two, and the State Update Head is used to output the SOC prediction sequence and the SOH prediction sequence.
[0035] Example 3: SOC / SOH sub-task data supervision loss.
[0036] Let the predicted SOC value at the final time step be SOCpred, the true SOC value be SOCtrue, the predicted SOH value be SOHpred, and the true SOH value be SOHtrue. Then, calculate the SOC data supervision loss and the SOH data supervision loss separately. Add them together according to their weights to obtain the sub-task data supervision loss. In this way, the error changes of the SOC task and the SOH task can be observed separately during the training process. Figure 2This is a schematic diagram illustrating the loss function for each task.
[0037] Example 4: SOC / SOH task sequence supervision loss.
[0038] For each time step in the predicted sequence, calculate the SOC sequence supervision loss and the SOH sequence supervision loss, and construct the sequence supervision loss according to preset weights. This loss is used to constrain the model not only to be accurate at the final time step, but also to maintain reasonable state changes over the complete sequence within the window.
[0039] Example 5: Physical constraint loss.
[0040] Physical constraint losses include SOC / SOH boundary constraints, SOH monotonically degrading constraints, SOC ampere-hour integral consistency constraints, idle-state SOC stability constraints, step-direction consistency constraints, temperature-strain degradation constraints, and sequence smoothing constraints. Boundary constraints are used to limit SOC / SOH to a reasonable range; SOH monotonically degrading constraints are used to suppress abnormal SOH rises; SOC ampere-hour integral consistency constraints are used to ensure that SOC changes are consistent with current and time integrals; idle-state SOC stability constraints are used to ensure that SOC does not change abruptly in the near-zero current stage; step-direction consistency constraints are used to ensure that SOC does not decrease during charging and does not increase during discharging; temperature-strain degradation constraints are used to enhance degradation consistency under high temperature, high current, or high strain conditions; and sequence smoothing constraints are used to suppress local high-frequency jitter in the SOC / SOH prediction curve.
[0041] Example 6: Dynamic physical weight scheduling.
[0042] Set the minimum and maximum scaling factors for physical constraint weights, as well as the number of warm-up and climbing rounds. During the warm-up phase, the physical constraint weights are kept at a low proportion; during the climbing phase, the physical constraint weights gradually increase with each training round; and during the stabilization phase, the physical constraint weights are kept at their maximum proportion. Through this strategy, the model prioritizes learning data patterns in the early stages of training and gradually reinforces physical patterns in the later stages. Figure 3 This is a schematic diagram of dynamic physical constraint weight scheduling.
[0043] Example 7: Learning rate scheduling and early stopping.
[0044] A cosine learning rate scheduler is used to gradually reduce the learning rate from its initial value to the minimum learning rate. During training, the validation loss is monitored. If the validation loss does not improve by more than the minimum improvement threshold for several consecutive rounds, early stopping is triggered and the optimal model is saved.
[0045] Example 8: Training Records and Model Saving.
[0046] Figure 4This is a schematic diagram of the overall training strategy structure. The training history includes the number of training epochs, learning rate, dynamic physical weight scaling factor, SOC data supervision loss, SOH data supervision loss, SOC sequence supervision loss, SOH sequence supervision loss, physical loss, total training loss, total validation loss, and number of epochs without boosting. If the validation loss reaches the current optimum, the optimal model weights, normalized parameters, and training configuration are saved.
[0047] This invention is not limited to the embodiments described above. Without departing from the core idea of this invention, the task weights, physical constraint base weights, dynamic physical weight scheduling curves, learning rate scheduling methods, number of early stopping patience rounds, and sequence lengths can all be adjusted according to different battery types, data sampling frequencies, and engineering application scenarios.
[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A training method for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-specific losses, characterized in that, The method specifically includes the following steps: S1. Obtain battery cycle test data, which includes cycle number, step type, total time, current, voltage, capacity, first strain signal V1, second strain signal V2, first temperature T1 and second temperature T2. S2. Preprocess the battery cycle test data to generate time in seconds, step code, whether to put aside, time interval between adjacent sampling points and time interval within the same cycle, and construct SOC and SOH tags. S3. Construct physical derived features based on the preprocessed data, and combine the original sensor features, step time features, and physical derived features into model input features; S4. Construct a fixed-length time series sample within the same loop, input the time series sample into the Transformer model that fuses structural-level physical information, and obtain the SOC / SOH prediction sequence and the final time step SOC / SOH prediction value. S5. Calculate the final time step supervision loss of SOC and the final time step supervision loss of SOH respectively, and construct the sub-task data supervision loss based on the two. S6. Calculate the SOC sequence supervision loss and the SOH sequence supervision loss respectively, and construct the sub-task sequence supervision loss based on the two. S7. Calculate the dynamic physical constraint weight scaling factor based on the current training round, and use the dynamic physical constraint weight scaling factor to adjust the effective weights of multiple physical constraint loss terms. S8. The total training loss is obtained by weighting and summing the task-specific data supervision loss, the task-specific sequence supervision loss, and the dynamically weighted physical constraint loss. S9. Perform backpropagation based on the total training loss, and train the battery SOC / SOH joint prediction model by combining learning rate scheduling and early stopping mechanism.
2. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 1, characterized in that, In step S3, the physical derived features include absolute current value, power, average temperature, temperature difference, average strain, strain difference, ampere-hour integral increment, and stress factor.
3. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 2, characterized in that, In step S4, the structure-level physical information fusion Transformer model includes a data encoder, a physical encoder, a gating fusion module, and a state update output module; In step S5, the final time step supervision loss of SOC is the mean square error between the predicted SOC value and the SOC label, the final time step supervision loss of SOH is the mean square error between the predicted SOH value and the SOH label, and the sub-task data supervision loss is the weighted sum of the final time step supervision loss of SOC and the final time step supervision loss of SOH.
4. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 3, characterized in that, In step S6, the SOC sequence supervision loss is the mean square error between the SOC predicted sequence and the SOC true sequence, and the SOH sequence supervision loss is the mean square error between the SOH predicted sequence and the SOH true sequence; the sub-task sequence supervision loss is the weighted sum of the SOC sequence supervision loss and the SOH sequence supervision loss.
5. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 4, characterized in that, In step S7, the physical constraint loss includes SOC / SOH boundary constraint loss, SOH monotonic degradation loss, SOC ampere-hour integral consistency loss, idle state SOC stability loss, step direction consistency loss, temperature-strain degradation loss, and sequence smoothing loss; the dynamic physical constraint weight scaling factor takes a small value in the early stage of training, gradually increases in the middle stage of training, and reaches a preset maximum value in the later stage of training; the dynamic physical constraint weight scaling factor includes a warm-up stage, a climbing stage, and a stabilization stage; in the warm-up stage, the physical constraint weight is maintained at the minimum proportion of the basic weight; in the climbing stage, the physical constraint weight gradually increases from the minimum proportion to the maximum proportion; During the stable phase, the physical constraint weights remain at their maximum proportion.
6. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 5, characterized in that, In step S9, the learning rate scheduling adopts cosine annealing scheduling, so that the learning rate is gradually reduced from the initial learning rate to the preset minimum learning rate; the early stopping mechanism is used to stop training when the verification loss has not produced an improvement greater than a preset threshold for multiple consecutive training rounds.
7. The battery SOC / SOH joint prediction model training method based on dynamic physical constraint weights and task-specific losses according to claim 6, characterized in that, The training history stores the training epochs, learning rate, dynamic physical weight scaling factor, SOC data supervision loss, SOH data supervision loss, SOC sequence supervision loss, SOH sequence supervision loss, physical constraint loss, total training loss, total validation loss, and number of no-boost epochs. The total training loss is expressed as the sum of the data supervision loss, sequence supervision loss, and dynamic physical constraint loss, wherein the dynamic physical constraint loss is obtained by multiplying each physical constraint loss by its dynamic effective weight and then summing the results.
8. A training system for a joint prediction model of battery SOC / SOH based on dynamic physical constraint weights and task-specific losses, characterized in that, The system includes a data preprocessing module, a physical feature construction module, a time-series sample construction module, a model prediction module, a task-based supervision loss calculation module, a dynamic physical weight scheduling module, a joint loss calculation module, a learning rate scheduling module, an early stopping control module, and a model saving module; the system is configured to perform the method of any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.
10. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1 to 7.