Anti-interference multi-constraint fusion lithium battery soc estimation method

Abstract

The application discloses an anti-interference type multi-constraint fusion lithium battery SOC estimation method and belongs to the technical field of battery state of charge estimation. A physical information multi-task learning model is constructed to identify SOC. The model adopts a multi-task LSTM framework. Hidden state features output by a backbone network of the model are respectively sent to two branches. Branch one is a full connection layer of a main task of SOC prediction, and outputs an SOC prediction value. Branch two is a full connection layer of an auxiliary task of voltage reconstruction. The branch two predicts and reconstructs voltage according to the hidden state features, the SOC prediction value and the original input current. The model adopts a joint loss function. Based on the same variance uncertainty principle, dynamic weights of each loss term in the total loss are adaptively adjusted. The loss terms include an SOC loss, a voltage reconstruction loss and two-phase physical constraint losses, namely, an SOC direction constraint loss and a rate constraint loss. The application can realize high-precision, high-physical-consistency and strong-robustness SOC estimation.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a battery state of charge (SOC) estimation technique, specifically an anti-interference multi-constraint fusion lithium battery SOC estimation method. Background Technology

[0002] Lithium-ion batteries, with their advantages of high energy density and long cycle life, have been widely used in key areas such as new energy vehicles, large-scale energy storage systems, and battery reuse. State of Charge (SOC), as a core parameter characterizing the remaining usable capacity of a battery, is crucial for accurately estimating the range prediction accuracy of new energy vehicles and the energy dispatch efficiency of energy storage systems. It is also a key prerequisite for avoiding overcharging and over-discharging, ensuring safety, and extending battery lifespan.

[0003] Currently, existing SOC estimation methods are mainly divided into three categories: traditional methods, model-based methods, and data-driven methods, but all have significant drawbacks. Among traditional methods, the ampere-hour integration method relies on high-precision current sensors and accurate initial SOC values. Measurement errors and capacity decay lead to cumulative deviations, with errors increasing significantly after long-term cycling. The open-circuit voltage method requires the battery to be left undisturbed for a long time to reach equilibrium, which cannot meet the needs of real-time applications. Furthermore, the hysteresis effect at high temperatures and the drift in the OCV-SOC relationship (the curve corresponding to the open-circuit voltage and the battery's state of charge) caused by battery aging further affect the estimation accuracy. Model-based methods, represented by Kalman filter algorithms, have a closed-loop correction mechanism, but they have high computational complexity, are sensitive to the accuracy of the battery model, and are prone to filter divergence under dynamic operating conditions, making them difficult to adapt to complex and ever-changing real-world application scenarios.

[0004] With the development of artificial intelligence technology, data-driven methods, especially deep learning models, have been widely used in the field of SOC estimation due to their powerful feature extraction and nonlinear fitting capabilities. Models such as CNN-LSTM, BiLSTM, and Transformer have shown certain estimation potential. However, existing data-driven methods still have several key shortcomings: First, they lack consideration of the intrinsic physical laws of batteries, treating SOC estimation as a simple sequential regression task and ignoring the electrochemical relationship between SOC and voltage and current. This leads to decoupling between state and observation, and when the predicted SOC value does not match its corresponding actual voltage, it cannot be effectively penalized, and the estimation result may deviate from the true battery state. Second, they are prone to physically inconsistent predictions. In noisy data or unseen operating conditions, logical errors that violate basic physical common sense, such as SOC decrease during continuous charging, may occur, and traditional loss functions are difficult to effectively penalize such errors. Third, they are overly sensitive to dynamic operating conditions. The SOC estimation curve is prone to unreasonable high-frequency jitter or instantaneous jumps, resulting in poor numerical stability and failing to meet the continuous and smooth characteristics of battery SOC changes. Fourth, they have insufficient generalization ability. In wide temperature ranges of 0℃-45℃ and battery aging scenarios, due to the lack of sufficient integration of physical constraints, the model performance is prone to significant degradation, making it difficult to meet the robustness requirements in practical applications. These shortcomings make it difficult for existing methods to balance estimation accuracy, physical consistency, and adaptability to complex operating conditions, thus limiting their reliable application in battery management systems (BMS). Summary of the Invention

[0005] To address at least one of the aforementioned problems, this invention provides an anti-interference multi-constraint fusion lithium battery SOC estimation method. This invention constructs a Physical Information Multi-Task Learning Framework (PIML-Frame), which uses a Long Short-Term Memory Network (LSTM) as the core temporal feature extraction module. Through multi-task collaborative training, physical constraint embedding, and dynamic weight adjustment, it achieves high-precision, high-physical-consistency, and robust SOC estimation.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An anti-interference, multi-constraint fusion lithium battery SOC estimation method includes the following steps:

[0008] S1. Collect current-voltage time series data during the operation of lithium-ion batteries, construct current-voltage time series samples for model training, and label the current-voltage time series samples as real SOC sequence and real voltage sequence.

[0009] S2. Construct a multi-task learning model for physical information. The model adopts a multi-task LSTM architecture. Its backbone network is used to read data (past current-voltage data) and extract useful information as hidden state features. The output hidden state features are fed into two branches: Branch 1 is a fully connected layer for the main task of SOC prediction, which outputs the predicted SOC value; Branch 2 is a fully connected layer for the auxiliary task of voltage reconstruction. The input of this branch is the multi-source fusion feature formed by concatenating the hidden state features, the predicted SOC value, and the original input current in the feature dimension. The output is a reconstructed voltage sequence aligned with the input sequence. The multi-task learning model for physical information adopts a joint loss function. The loss terms include SOC direction constraint loss, rate constraint loss, SOC loss, and voltage reconstruction loss. Based on the principle of homoscedastic uncertainty, the dynamic weights of each loss term in the total loss are adaptively adjusted. The SOC direction constraint loss is used to penalize cases where the direction of SOC change is inconsistent with the current direction, while the rate constraint loss is based on the rate of change of SOC.

[0010] S3. Use the current and voltage time-series samples from step S1 to train the physical information multi-task learning model constructed in step S2.

[0011] S4. Collect the timing data of the lithium battery and use the trained physical information to predict the SOC value using a multi-task learning model.

[0012] As a specific embodiment of the present invention, the formula for calculating the SOC orientation constraint loss is as follows:

[0013]

[0014]

[0015] In the formula, This represents the SOC directional constraint loss; Indicates the length of the sequence; This indicates taking the maximum value; Represents the first in the sequence One time step; Indicates the first The current at any given moment is positive during the charging state. The current is negative during discharge. ; Indicates the first The SOC difference at time points; and Indicates the first Time and the SOC value at time t; To represent the threshold; It represents the absolute value.

[0016] Setting a threshold in the SOC directional constraint loss is primarily because, during actual battery operation, when the absolute value of the current is very small (close to zero), the battery is in a near-static state. At this point, the change in SOC is extremely weak, and it may even fluctuate slightly due to factors such as measurement noise. If directional constraints are strictly applied under these conditions, these noises might be mistakenly interpreted as violations of physical laws, leading to unnecessary penalties and affecting the model's training results. To avoid this, a masking mechanism is introduced into the directional constraint loss calculation. A current threshold is set; when the absolute value of the current is less than this threshold, the directional constraint loss is ignored, meaning it is not included in the total loss calculation. This effectively reduces the interference of noise on the constraints when the current is close to zero, allowing the directional constraints to function more effectively.

[0017] After applying SOC directional constraint loss, during charging, when the current... When this happens, it means that external electrical energy is being input into the battery, and the battery's SOC should increase, that is... According to the above calculation formula, the directional constraint loss is 0 at this time, indicating that the model prediction conforms to physical laws and no penalty is imposed; while when the current However, SOC decreased ( When this occurs, it violates the physical law that the State of Charge (SOC) should increase during charging. According to the above calculation formula, the directional constraint loss is positive at this time, thus imposing a penalty. The same applies to the discharge situation.

[0018] The formula for calculating the rate constraint loss is as follows:

[0019]

[0020]

[0021] In the formula, Indicates rate constraint loss, This represents the second difference of SOC; and Indicates the first Time and The SOC difference at time points; This represents the threshold.

[0022] This invention achieves the rate constraint objective by constraining the second-order difference of the SOC. When the absolute value of the second-order difference of the SOC is within a threshold... When the SOC is within a certain range, the loss is 0; when it exceeds this threshold, the loss is the value of the excess portion, and the greater the excess, the greater the loss. From the perspective of the battery's physical characteristics, the change in SOC is a relatively smooth process, and its rate of change does not undergo drastic abrupt changes. This is determined by the electrochemical reaction rate inside the battery. When the SOC changes linearly, , The value is negative. The value of is 0, and the rate constraint loss is 0, indicating that the predicted SOC change conforms to the smooth physical characteristics. When the SOC undergoes a sudden change, such as a step-like change, its rate of change will change significantly instantaneously, leading to second-order difference. The absolute value increases. When Exceeding the threshold hour, The value is positive. When the value is positive, the rate constraint loss increases, the model will be penalized, and thus be guided to learn a smooth SOC change trend to avoid abrupt changes that do not conform to physical laws.

[0023] As a specific embodiment of the present invention, the formula for calculating the joint loss function is as follows:

[0024]

[0025] In the formula, Indicates the total loss; Indicates task The corresponding losses, among which, Indicates SOC loss. Indicates voltage reconstruction loss; Is related to the task The square of the learnable noise parameter corresponding to the loss The uncertainty of the model's prediction for this task was quantified.

[0026] The joint loss function employs dynamic weights based on the principle of homoscedastic uncertainty, which can intelligently balance the complex relationship between data fitting and physical constraints, significantly improving the training stability and final performance of the model under varying conditions.

[0027] Compared with the prior art, the present invention has at least the following beneficial effects:

[0028] This invention effectively solves the core problem of state-observation decoupling in traditional SOC estimation. Existing data-driven methods often treat SOC estimation as a simple sequence regression task, neglecting the electrochemical correlation between SOC and voltage / current, leading to potential numerical biases or physical inconsistencies in the prediction results. This invention, by adding a voltage reconstruction auxiliary task involving multi-source information fusion, forces the model to learn the intrinsic correlation between historical dynamics, the current state, and external stimuli. This ensures that SOC estimation not only pursues numerical accuracy but also closely aligns with the electrochemical operating laws of the battery, significantly improving the physical interpretability and rationality of the estimation results.

[0029] This invention solves the problems of physically inconsistent predictions and numerical instability. Purely data-driven models are prone to exhibiting physically incompatible behaviors under noisy or complex operating conditions, such as SOC drops during charging and drastic jumps in the SOC curve, which are difficult to constrain effectively using traditional loss functions. This invention, by embedding explicit SOC direction and rate constraints into the loss function, ensures that the direction of SOC change is consistent with the current direction, while also guaranteeing the continuous and smooth characteristics of the SOC curve. This avoids unreasonable logical errors and numerical abrupt changes, significantly improving the reliability and stability of the model under dynamic operating conditions.

[0030] This invention significantly enhances the model's generalization ability in complex environments. Existing methods are insufficiently adaptable to factors such as temperature changes and battery aging, and their accuracy decreases significantly under extreme conditions such as low and high temperatures. The voltage reconstruction task and physical constraints of this invention work synergistically to dynamically capture the impact of temperature on battery electrochemical characteristics. By using physical constraints to offset prediction biases caused by environmental changes, it can adapt to complex scenarios such as wide temperature ranges and aging without additional parameter adjustments, greatly expanding the application scope of SOC estimation methods.

[0031] This invention balances estimation performance with engineering practicality. Compared to complex architecture models, this invention uses LSTM as the core module, controlling computational complexity while ensuring high accuracy. It eliminates the need for high-precision sensors or complex model parameter identification, reducing hardware deployment costs. Furthermore, a dynamic weighting strategy based on homoscedasticity uncertainty achieves adaptive balancing of multi-task losses, avoiding the tedious manual parameter tuning required in traditional multi-task learning. This improves model training efficiency and stability, enabling direct integration into existing battery management systems (BMS), providing an efficient and reliable SOC estimation solution for new energy vehicles, energy storage power stations, and other fields. Attached Figure Description

[0032] Figure 1 This is a diagram illustrating the overall architecture of the Physical Information Multi-Task Learning Framework (PIML-Frame).

[0033] Figure 2The following are performance comparison charts of each model under normal operating conditions at 25℃, where (a) is the MAE comparison chart and (b) is the RMSE comparison chart.

[0034] Figure 3 Comparison of the SOC estimation curves of each model with the reference values;

[0035] Figure 4 The error distribution of the SOC estimates for each model;

[0036] Figure 5 The graphs show the performance of each model in the generalization experiment, where (a) is the MAE comparison graph and (b) is the RMSE comparison graph.

[0037] Figure 6 The figures show the performance of each comparative model in the ablation experiment, where (a) is the MAE comparison figure, (b) is the RMSE comparison figure, and (c) is the MaxAE comparison figure. Detailed Implementation

[0038] To more clearly illustrate the present invention, specific embodiments are described below. Those skilled in the art should understand that the following description is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0039] This embodiment takes the 18650 lithium-ion battery as the research object and uses the publicly available Dynamic Stress Test (DST) dataset to verify the SOC estimation performance of the proposed Physical Information Multi-Task Learning Model (PIML-Frame). The interference-resistant multi-constraint fusion lithium battery SOC estimation method in this embodiment includes the following steps:

[0040] S1. Collect current-voltage time series data during the operation of lithium-ion batteries, construct current-voltage time series samples for model training, and the labels of the current-voltage time series samples are the real SOC sequence and the real voltage sequence.

[0041] This step first collects current-voltage time-series data during the operation (charge and discharge process) of lithium-ion batteries. The data covers different temperatures, different initial SOCs, and different charge and discharge conditions. Specifically, the dataset covers three typical temperature conditions: 0℃, 25℃, and 45℃. At each temperature, it includes charge and discharge cycle data with initial SOCs of 50% and 80%, completely recording the real-time current and voltage time-series information and the corresponding true SOC values ​​during battery operation. Before the experiment, the data is preprocessed: outliers and missing values ​​are removed to ensure data integrity. Then, current and voltage are extracted as input features for the model. The true SOC value is used as the label for the main SOC prediction task, and the true voltage value is used as the label for the voltage reconstruction auxiliary task. Subsequently, the sliding window method is used to convert the raw data into fixed-length time-series samples, preserving the temporal correlation between data. Finally, the input features and labels are standardized to eliminate the influence of differences in physical dimensions on model training.

[0042] S2. Construct a multi-task learning model for physical information. This model employs a multi-task LSTM architecture, such as... Figure 1 As shown, its backbone network is used to read data (past current-voltage data) and extract useful information as hidden state features. The output hidden state features are fed into two branches: Branch 1 is a fully connected layer for the main task of SOC prediction, which outputs the SOC prediction value; Branch 2 is a fully connected layer for the auxiliary task of voltage reconstruction. The input of this branch is the multi-source fusion feature formed by concatenating the hidden state features, the SOC prediction value, and the original input current in the feature dimension. The output is the reconstructed voltage sequence aligned with the input sequence in time. The physical information multi-task learning model adopts a joint loss function. The loss terms include SOC direction constraint loss, rate constraint loss, SOC loss, and voltage reconstruction loss. Based on the principle of homoscedastic uncertainty, the dynamic weight of each loss term in the total loss is adaptively adjusted. Among them, the SOC direction constraint loss is used to penalize the case where the change direction of SOC is inconsistent with the current direction, and the rate constraint loss is based on the rate of change of SOC.

[0043] In this embodiment, a SOC prediction model based on a Long Short-Term Memory (LSTM) network is used. The LSTM network is a three-layer unidirectional stacked structure. The input is time-series features with a dimension of 2, corresponding to current and voltage features respectively, and the input sequence length is 100. The hidden state dimension of each layer of the LSTM network is 64. The input data is organized in the form of batch, time step, and feature dimension, and the dropout parameter between layers is set to 0.2. The LSTM network models the input sequence step by step and outputs a 64-dimensional hidden state vector at each time step; its initial hidden state and initial cell state are zero-initialized. During model training, the batch size is set to 64, the learning rate is set to 0.001, and the number of training epochs is set to 100.

[0044] Based on the hidden states output by the LSTM network, a SOC prediction branch and a voltage reconstruction branch are constructed. The SOC prediction branch consists of two fully connected layers: first, the 64-dimensional hidden state is input into the first linear transformation layer, mapped to a 32-dimensional feature vector. After processing by the ReLU activation function, it is input into the second linear transformation layer, outputting a 1-dimensional SOC prediction value. The voltage reconstruction branch concatenates the 64-dimensional LSTM hidden state, the 1-dimensional SOC prediction value, and the 1-dimensional current feature at the same time step along the feature dimension, forming a 66-dimensional fused feature vector. This 66-dimensional fused feature vector is mapped to a 32-dimensional feature vector by the first linear transformation layer, activated by ReLU, and then mapped to a 1-dimensional output by the second linear transformation layer, thus obtaining the voltage reconstruction value for the corresponding time step. This dual-branch structure achieves voltage reconstruction simultaneously with SOC prediction, enhancing the model's ability to represent the dynamic characteristics of the battery and improving the accuracy and stability of SOC estimation.

[0045] In this embodiment, both SOC loss and voltage reconstruction loss are calculated using the mean square error between the true and estimated values.

[0046] SOC loss is used to measure the difference between the SOC sequence predicted by the model and the true SOC sequence (label), and its calculation formula is as follows:

[0047]

[0048] in, Indicates SOC loss. Indicates the length of the sequence; Represents the first in the sequence One time step; Indicates the first The real-time SOC value; The model predicts the first... SOC value at time t;

[0049] The voltage reconstruction loss is used to measure the difference between the voltage sequence derived from the model's auxiliary branches and the actual acquired voltage sequence. Its calculation formula is as follows:

[0050]

[0051] In the formula, Indicates voltage reconstruction loss, Indicates the first Real-time, accurate voltage value acquisition; The model predicts the first... The voltage value at that moment;

[0052] Both SOC directional and rate constraints are newly introduced explicit physical constraints. The core objective of the directional constraint is to ensure that the direction of SOC change aligns with the direction of current, which is determined by the fundamental physical laws of battery charging and discharging. The main purpose of the rate constraint is to limit the drastic nature of SOC changes and avoid abrupt changes that contradict the battery's physical characteristics.

[0053] The formula for calculating the SOC directional constraint loss is as follows:

[0054]

[0055]

[0056] In the formula, This represents the SOC directional constraint loss; This indicates taking the maximum value; Represents the first in the sequence One time step; Indicates the first The current at any given moment is positive during the charging state and negative during the discharging state. Indicates the first The SOC difference at time points; and Indicates the first Time and the SOC value at time t; To represent the threshold; It represents the absolute value.

[0057] The formula for calculating the rate constraint loss is as follows:

[0058]

[0059]

[0060] In the formula, Indicates rate constraint loss, This represents the second difference of SOC; and Indicates the first Time and The SOC difference at time points; This represents the threshold.

[0061] The formula for calculating the joint loss function is as follows:

[0062]

[0063] In the formula, Indicates the total loss; Indicates task There are four types of losses. Is related to the task The square of the learnable noise parameter corresponding to the loss The uncertainty of the model's prediction for this task was quantified.

[0064] The joint loss function employs dynamic weights based on the principle of homoscedastic uncertainty, which can intelligently balance the complex relationship between data fitting and physical constraints, significantly improving the training stability and final performance of the model under varying conditions.

[0065] S3. Use the current and voltage time series samples from step S1 to train the physical information multi-task learning model constructed in step S2. During training, the calculation result of the joint loss function is used as the basis for gradient update. The backpropagation algorithm is used to iteratively optimize the parameters of the LSTM network, the two fully connected layers, and the uncertainty parameters of each loss term until the model converges.

[0066] S4. Collect the timing data of the lithium battery and use the trained physical information to predict the SOC value using a multi-task learning model.

[0067] To verify the effectiveness of this invention, in this embodiment, the current and voltage time-series samples are divided into a training set and a test set. The training set uses complete charge-discharge cycle data with an initial SOC of 80% at 25°C. The test set includes standard test data with an initial SOC of 50% at 25°C, as well as generalized test data at 0°C and 45°C. Four mainstream deep learning models—CNN-LSTM, BiLSTM, TCN, and Transformer—are selected as benchmarks. All models use the same input dimension, number of training iterations, and optimizer parameters to ensure fairness. The experiment uses mean absolute error (MAE), root mean square error (RMSE), and maximum absolute error (MaxAE) as performance evaluation metrics to measure the performance of the model trained on the training set on the test set. The results are as follows: Figure 2-6 As shown.

[0068] like Figure 2-4As shown, in the performance comparison test under normal operating conditions of 25℃, the PIML-Frame proposed in this invention exhibits excellent estimation accuracy, with a MAE as low as 0.0020 and an RMSE of 0.0030, significantly outperforming all comparison models. Specifically, compared to the second-best TCN model (MAE=0.0034, RMSE=0.0046), PIML-Frame reduces MAE by 41.2% and RMSE by 34.8%; compared to the CNN-LSTM model (MAE=0.0058), MAE is reduced by 65.5%; and compared to the Transformer model (MAE=0.0100), MAE is reduced by 80.0%, fully demonstrating the technical advantages of multi-task collaborative training and physical constraint embedding.

[0069] Please refer to Figure 5 Generalization performance test results show that PIML-Frame maintains stable high accuracy across a wide temperature range. At 0℃, the battery polarization effect is significantly enhanced, and the nonlinear characteristics of the SOC-voltage curve intensify. However, PIML-Frame's MAE is still only 0.0092, and RMSE is 0.0109, representing a 36.5% reduction in error compared to the BiLSTM model (MAE=0.0144) and a 23.3% reduction in error compared to the Transformer model (MAE=0.0119). At 45℃, accelerated battery side reactions lead to unstable SOC decay rates. PIML-Frame's MAE is 0.0096, and RMSE is 0.0171, while the MAE of the TCN model and Transformer model soars to 0.0170 and 0.0166, respectively, highlighting PIML-Frame's strong adaptability to extreme temperature conditions.

[0070] To verify the function of each core module, this embodiment designed an ablation experiment, constructing three ablation models: one removing the voltage reconfiguration task (WoRecon), one removing the SOC direction constraint (WoTrend), and one removing the SOC rate constraint (WoRate), as well as a baseline model retaining only the basic LSTM structure. Please refer to [reference needed]. Figure 6Test results show that after removing the voltage reconstruction module, the model's MAE increases to 0.0027, and the loss of physical supervision of the voltage signal leads to increased estimation bias. After removing the direction constraint, the model's MAE increases to 0.0023, and a logical error occurs where the SOC change direction contradicts the current direction under charge-discharge switching conditions. After removing the rate constraint, the model's MAE climbs to 0.0033, and the SOC estimation curve exhibits unreasonable jitter, resulting in decreased noise immunity. In contrast, the baseline LSTM model has an MAE of 0.0036 and frequently exhibits physically inconsistent predictions under complex operating conditions. This fully demonstrates that the voltage reconstruction task, direction constraint, and rate constraint modules of this invention are key supports for improving SOC estimation accuracy and ensuring prediction rationality.

[0071] In summary, this embodiment fully verifies through conventional performance comparison, generalization test and ablation test that PIML-Frame's estimation accuracy under normal operating conditions is significantly better than that of existing mainstream models. It can still maintain stable performance in complex environments with a wide temperature range. Moreover, through the synergistic effect of each core module, it effectively solves the problems of state and observation decoupling, insufficient physical consistency and poor generalization ability of traditional models, and can provide reliable technical support for lithium-ion battery SOC estimation.

[0072] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included within the scope of protection of the present invention.

Claims

1. A robust multi-constraint fusion method for estimating the state of charge (SOC) of lithium-ion batteries, characterized in that: Includes the following steps: S1. Collect current-voltage time series data during the operation of lithium-ion batteries, construct current-voltage time series samples for model training, and the labels of the current-voltage time series samples are the real SOC sequence and the real voltage sequence. S2. Construct a physical information multi-task learning model. The physical information multi-task learning model adopts a multi-task LSTM architecture. The hidden state features output by the backbone network are fed into two branches: Branch 1 is a fully connected layer for the main task of SOC prediction, outputting the predicted SOC value; Branch 2 is a fully connected layer for the auxiliary task of voltage reconstruction. The input of Branch 2 is the multi-source fusion feature formed by concatenating the hidden state features, the predicted SOC value, and the original input current in the feature dimension. The output is a reconstructed voltage sequence aligned with the input sequence. The physical information multi-task learning model adopts a joint loss function. The loss terms include SOC direction constraint loss, rate constraint loss, SOC loss, and voltage reconstruction loss. Based on the principle of homoscedastic uncertainty, the dynamic weights of each loss term in the total loss are adaptively adjusted. The SOC direction constraint loss is used to penalize cases where the direction of SOC change is inconsistent with the current direction, and the rate constraint loss is based on the rate of change of SOC. S3. Train the physical information multi-task learning model using the current and voltage time-series samples; S4. Collect time-series operating data of lithium batteries and use the trained physical information to predict using a multi-task learning model. SOC value; The formula for calculating the SOC directional constraint loss is as follows: In the formula, This represents the SOC directional constraint loss; Indicates the length of the sequence; This indicates taking the maximum value; Represents the first in the sequence One time step; Indicates the first The current at any given moment is positive during the charging state and negative during the discharging state. Indicates the first The SOC difference at time points; and Indicates the first Time and the SOC value at time t; To represent the threshold; It represents the absolute value.

2. The anti-interference multi-constraint fusion lithium battery SOC estimation method according to claim 1, characterized in that, The formula for calculating the rate constraint loss is as follows: In the formula, This represents the rate constraint loss; Indicates the length of the sequence; Represents the first in the sequence One time step; This represents the second difference of SOC; and Indicates the first Time and The SOC difference at time points; This represents the threshold.

3. The anti-interference multi-constraint fusion lithium battery SOC estimation method according to claim 1, characterized in that, The formula for calculating the joint loss function is as follows: In the formula, Indicates the total loss; Indicates task The corresponding losses, among which, Indicates SOC loss. Indicates voltage reconstruction loss; Is related to the task The learnable noise parameter corresponding to the loss.

Images