A lithium ion battery cross-condition migration state of charge estimation method based on temperature adaptive physical parameter adaptation network

Abstract

The application discloses a lithium ion battery cross-condition migration state of charge estimation method based on temperature adaptive physical parameter adaptive network. The method integrates a physical model into a loss function to enable a neural network to learn physical knowledge, and establishes models with different parameters for batteries at different temperatures. A double-window sliding sampling strategy is adopted, a Transformer encoder is used to capture long-range global dependencies, and a time convolution network is used to extract recent fine dynamics; a temperature adaptive physical parameter adaptive network based on a monotonic neural network is introduced to learnably model the temperature variation law of battery capacity, internal resistance and coulomb efficiency; a physical information loss function with temperature adaptive weights and a preheating mechanism is designed to dynamically weaken the physical constraint loss in the low temperature interval, so as to avoid the interference of model deviation on estimation. The method can accurately estimate the battery SOC in the low temperature condition.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle battery management technology, specifically to a method for estimating the state of charge of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network. Background Technology

[0002] As the core energy storage component of new energy vehicle power systems, the accurate and real-time estimation of the state of charge (SOC) of lithium-ion batteries is crucial for the safe and reliable operation of the battery management system (BMS). SOC reflects the ratio of the battery's current stored energy to its rated capacity, and is a core parameter for assessing driving range, preventing overcharging and over-discharging, and ensuring driving safety. Since SOC cannot be directly measured and must be estimated using indirect parameters (current, voltage, temperature, etc.), this problem remains a challenge in engineering.

[0003] The current mainstream SOC estimation methods can be divided into three categories: (1) Model-based methods, such as the ampere-hour integral method and the Kalman filter series (EKF, UKF, AEKF), etc. The ampere-hour integral method has a prominent cumulative error problem. Kalman-type methods rely on accurate electrochemical equivalent circuit model (ECM) parameters, but battery parameters change nonlinearly with temperature and aging degree, making it difficult to identify model parameters; (2) Data-driven deep learning methods, such as LSTM and Transformer, have strong nonlinear fitting capabilities, but usually have a strong dependence on the distribution of training data, and the generalization performance decreases in cross-operating conditions; (3) Physical Information Neural Network (PINN) method, which embeds physical constraints into the neural network training process, which can improve interpretability and generalization ability to a certain extent. However, existing PINN methods mostly apply physical equations in the form of hard constraints. Under low temperature conditions, the ECM model itself has a large deviation, and hard constraints will pull the prediction results off. The performance drops sharply under extreme low temperature conditions such as -10℃.

[0004] Furthermore, the actual driving conditions of new energy vehicles are diverse, with significant differences in the dynamic range of current under each condition. When models trained on conventional source domain conditions are directly transferred to target conditions, the error often increases dramatically due to the shift in feature distribution. How to achieve efficient adaptation with a small amount of target domain data while ensuring estimation accuracy in low-temperature scenarios is a core problem that urgently needs to be solved in the field of lithium-ion battery SOC estimation.

[0005] To address this, this invention proposes a SOC estimation method that integrates temperature-adaptive physical constraints, dual-scale temporal feature extraction, and efficient parameter transfer learning. This method significantly improves estimation accuracy under low-temperature conditions while maintaining cross-condition generalization. Summary of the Invention

[0006] This invention addresses the aforementioned technical problems by proposing a method for estimating the state of charge (SOC) of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network. This method integrates a data-driven deep learning model with electrochemical mechanism constraints, employs a dual-window sliding sampling strategy to extract multi-scale temporal features, and introduces a temperature-adaptive physical parameter adaptation network based on a monotonic neural network to learnably model the temperature-dependent variations in battery capacity, internal resistance, and coulombic efficiency. Furthermore, a physical information loss function incorporating temperature-adaptive weights and a preheating mechanism is designed to dynamically reduce the physical constraint loss in the low-temperature range. This method can accurately estimate the SOC of batteries under low-temperature conditions and maintain excellent estimation accuracy in cross-condition migration scenarios, effectively solving the technical challenges of low low-temperature estimation accuracy and weak cross-condition generalization ability in existing technologies.

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies. This invention provides a method for migrating the state of charge (SOC) of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network. The estimation method has higher SOC estimation accuracy under extreme low temperature conditions and good cross-domain generalization ability across various driving conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for migrating the state of charge (SOC) of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network. The estimation method includes the following steps: Step 1: Preprocess the time-series data of lithium-ion batteries collected under different temperature conditions and driving conditions to extract three-dimensional feature sequences of current, voltage, and temperature. Then, normalize the features using a global standardization method based on full-condition data. For the normalized feature sequences, a dual-window sliding sampling strategy is used to construct long-window and short-window samples.

[0009] Step 2: Input the long window samples into the Transformer temporal backbone encoder to extract global temporal features, and input the short window samples into the temporal convolutional network (TCN) to extract recent local dynamic features. Combine the two features to output the initial prediction value of SOC.

[0010] Step 3: Establish a temperature-adaptive electrochemical equivalent circuit model through a learnable physical parameter adaptation network, and perform monotonic network modeling of the temperature variation of battery capacity Q(T), internal resistance R(T), and coulombic efficiency η(T).

[0011] Step 4: The model is trained end-to-end using a physical information loss function, which includes data supervision loss, coulomb count consistency constraint, voltage consistency constraint and physical parameter regularization term, and introduces temperature adaptive weights and a preheating mechanism.

[0012] Step 5: Based on the model trained under source domain conditions, freeze the backbone encoder parameters and only perform target domain adaptive fine-tuning on the physical parameter adaptation network to achieve efficient generalization of SOC estimation across operating conditions and temperatures.

[0013] The global standardization method described in step 1 specifically involves: collecting data on the target battery under all operating conditions ( Using raw data across the entire temperature range, the global mean of the three-dimensional characteristics of current, voltage, and temperature is calculated together. and standard deviation With standardized formula Features are normalized, and the normalization parameters are persistently saved to ensure that a completely consistent normalization space is used during the training and inference phases. In the dual-window sliding sampling strategy, the long window length... Each time step, short window length Each time step is a subsequence at the end of the long window, and the two windows correspond to the same prediction time, thus achieving multi-scale time series modeling in a lightweight manner.

[0014] The learnable physical parameter adaptation network described in step 3 is implemented using a monotonic neural network, and the network weights are constrained to be positive to ensure... and With temperature monotonically increasing, The physical priors decrease monotonically with temperature; the monotonicity constraint effectively prevents parameter extrapolation failure under extreme low-temperature conditions.

[0015] In step 4, a temperature-adaptive weighting function is introduced into the physical information loss function. Under low-temperature conditions, the weight of physical constraint loss is automatically reduced, allowing data-driven loss to dominate parameter updates in the low-temperature range, thus avoiding... Model systematic bias The prediction generates negative interference; at the same time, a linear preheating mechanism is introduced to gradually introduce physical constraints in the early stage of training to ensure the model's convergence stability.

[0016] The parameter fine-tuning strategy described in step 5 only fine-tunes the physical parameters of the adaptation network, and the amount of target domain labeled data used is the total amount of target domain data. Within, and adopt Layered sampling strategy to ensure full Interval coverage.

[0017] Furthermore, the data preprocessing method described in step 1 is specifically as follows: Raw data of the target battery across the entire temperature range under DST, FUDS, and US06 operating conditions are collected. The global mean and standard deviation of the three-dimensional characteristic current, voltage, and temperature are calculated. All features are normalized using a standardized formula to ensure that the source and target domains share a unified feature distribution space. Compared to methods that only use source domain training data to fit normalized parameters, this method effectively avoids the target domain data from exceeding the normalization range under high dynamic conditions, improving the robustness of cross-condition estimation. A long window length is set. Each time step, short window length The nth time step; for the nth time step in the time series; A prediction point, with a long window for timing. All three-dimensional features within, the short window takes the last part of the long window. Each time step, i.e., location Subsequence at the target; The truth of every moment The two windows share the same feature data. The shorter window is a subset of the longer window and does not require additional data reading, thus achieving efficient multi-scale time series modeling.

[0018] Furthermore, the monotonic neural network described in step 3 specifically refers to: Normalized battery temperature For input, use The activation function processes the weights, ensuring they remain positive to achieve monotonicity of the output with respect to the input; where... Network and The network output increases monotonically with temperature. The network output decreases monotonically with temperature, which is consistent with the physical law that lithium-ion batteries experience capacity decay, increased internal resistance, and reduced efficiency at low temperatures. The monotonicity constraint can effectively prevent model extrapolation failure under extreme low and high temperature conditions where data is scarce.

[0019] Furthermore, the specific expression for the physical information loss function described in step 5 is as follows: ; in For prediction With reality The mean squared error loss; To address the consistency loss of the Coulomb count, adjacent time steps are constrained based on the Coulomb count equation. Consistency of rate of change; For voltage consistency loss, prediction is based on equivalent circuit model constraints. The error between the reconstructed terminal voltage and the measured terminal voltage is calculated using... Loss is used to enhance robustness to outliers; For physical parameter regularization terms, constraints , , Within reasonable physical limits; This is a temperature-adaptive weighting function that applies smaller weights to physical constraint terms under low-temperature conditions, thus avoiding the negative impact of physical model biases on performance under these conditions. It is estimated that negative interference will occur; , This is an adjustable loss weighting coefficient.

[0020] Furthermore, the temperature adaptive weighting function Defined as: ; in This represents the average battery temperature of the current batch of samples. This is the lower limit of the weighting at low temperatures. To adjust the slope, This is the neutral temperature reference point; under extremely low temperatures... Approaching This significantly reduces the physical constraint loss, allowing data to drive the update of the loss-dominant parameters; at room temperature Approaching The physical constraints are functioning normally.

[0021] Furthermore, the model training described in step 5 also includes a warm-up mechanism: During the initial warm-up phase of training, the weighting coefficients... from linear growth to The effective weights of the physical constraint terms are After the preheating phase ends Hengwei The preheating mechanism ensures that the model prioritizes data supervision loss during the parameter initialization phase. Once the prediction converges to a reasonable region, physical constraints are gradually introduced to avoid unstable gradients caused by physical constraint loss in the early stages of training.

[0022] Furthermore, the efficient parameter fine-tuning strategy described in step 5 is specifically as follows: Freeze all parameters of the Transformer encoder and temporal convolutional network, and only unfreeze and update the physical parameters in the adaptation network. , , The parameters of the monotonic network and the parameters of the equivalent circuit residual compensation network; the number of fine-tunable parameters accounts for approximately [a certain percentage] of the total number of model parameters. Within; fine-tuning data adopted Stratified sampling, uniformly covering the target domain dataset of The interval, with a small number of labeled samples, achieves efficient cross-condition adaptation.

[0023] The present invention also provides an estimation system for the state of charge migration of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network, comprising: a data preprocessing module, a time-series feature extraction module, a physical parameter adaptation module, a physical information loss training module, and a domain adaptive fine-tuning module.

[0024] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores executable instructions, and the processor executes the instructions to implement the above-described state of charge estimation method.

[0025] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described state of charge estimation method.

[0026] Compared with existing technologies, this invention provides a method for estimating the state of charge of lithium-ion batteries across operating conditions based on a temperature-adaptive physical parameter adaptation network, which has the following advantages: 1. This invention addresses the problem of large deviations in equivalent circuit models under low-temperature conditions by innovatively introducing a temperature-adaptive loss weight mechanism. In extreme low-temperature scenarios of -10°C, the intervention weight of physical constraint loss is automatically reduced, allowing the model to primarily rely on data-driven learning of complex nonlinear characteristics under low-temperature conditions. According to experimental data, the SOC estimation RMSE of the method in this invention under extreme low-temperature conditions of -10°C is 8.96%, which is about 37% lower than the 14.25% of the traditional end-to-end method, and more than 30% lower than the SOC estimation RMSE of the hard-constraint physical information neural network method under the -10°C condition.

[0027] 2. Based on a parameter-efficient fine-tuning strategy, this invention achieves efficient domain adaptation from DST training scenarios to FUDS and US06 scenarios with a very small amount of target domain data, significantly reducing data annotation costs in practical applications. According to experimental data, the RMSE of this invention in the DST→FUDS transfer scenario is 3.15%, which is about 11% lower than the 3.54% of direct transfer; and the RMSE in the DST→US06 transfer scenario is 3.77%, which is about 9% lower than the 4.14% of direct transfer.

[0028] 3. This invention models Q(T), R(T), and η(T) using a monotonic neural network. The learned parameters satisfy the physical laws of lithium-ion batteries, namely, Q(T) and η(T) increase monotonically with temperature, and R(T) decreases monotonically with temperature. The model has good interpretability and will not produce physically unreasonable extrapolation results in the temperature boundary region where data is scarce.

[0029] 4. The RMSE of the method of the present invention is 4.27% under mild low temperature conditions of 0°C, which is about 54% lower than that of the traditional end-to-end method of 9.30%; the RMSE is 2.35% under normal temperature conditions of 25°C, which is about 64% lower than that of the traditional end-to-end method of 6.45%, and it maintains excellent estimation accuracy under all temperature conditions. Attached Figure Description

[0030] Figure 1 A flowchart of the overall method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the dual-window sliding sampling strategy provided in an embodiment of the present invention; Figure 3 This is a diagram of the temperature adaptive physical parameter adaptation network structure provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the cross-condition adaptive fine-tuning strategy provided in an embodiment of the present invention; Figure 5 This is a comparison chart of SOC estimation results under various temperature conditions provided in the embodiments of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1 to 5 This invention provides a method and system for estimating the state of charge (SOC) of lithium-ion batteries based on temperature-adaptive physical constraints and efficient parameter transfer learning. The method is applied to real-time SOC estimation of lithium-ion batteries in actual driving scenarios of new energy vehicles. It includes five core steps: data preprocessing, two-way time-series feature extraction, temperature-adaptive physical parameter adaptation, physical information loss function training, and cross-operating condition adaptive fine-tuning. Specifically, it includes the following steps: This invention relates to the application of lithium-ion batteries in actual driving scenarios of new energy vehicles. Real-time estimation is used as the application background. Batteries in publicly available datasets , , , , Under multiple temperature conditions , , Cyclic test data under three operating conditions are used to verify the proposed method. The experimental battery number is... The feature list includes current. ,Voltage and temperature .

[0033] Step S1: Refer to Appendix Figure 1 The overall process of the method of the present invention includes the following steps: First, the raw battery data is globally standardized and sampled using a data preprocessing module; then, global and local time-series features are extracted using a Transformer encoder and a TCN network, respectively, through two time-series feature extraction modules; next, a temperature-adaptive equivalent circuit model is established through a physical parameter adaptation module; then, end-to-end training is performed using a physical information loss function that includes temperature-adaptive weights through a physical loss training module; finally, cross-domain generalization is achieved through a cross-operating condition adaptive fine-tuning module.

[0034] Specifically, in the data preprocessing module, the specific processing steps are as follows: reading from the dataset... , , , Under multiple temperature conditions , , Three working conditions Raw data, extract current ,Voltage ,temperature Three columns as a feature The list is used as the estimation target.

[0035] All operating conditions of the target battery The data are combined to fit normalized parameters and calculate the global mean vector of the three-dimensional features of current, voltage, and temperature. and standard deviation vector Using standardized formulas Normalize all samples. Persist the normalization parameters in scaler.pkl format for loading during inference.

[0036] This step employs a combined standardization approach covering all operating conditions, rather than solely using the source domain. The reason for the data is: The dynamic range of current under operating conditions is significantly greater than Operating conditions, if only The normalization norm of data fitting will lead to High-current samples under operating conditions exceed the normalization range, causing a significant shift in feature distribution and reducing the model's ability to generalize across operating conditions.

[0037] See appendix Figure 2For the normalized time series, a dual-window sliding sampling strategy is used to construct the model input samples. Let the length of the long window be... Short window length Short window start offset .

[0038] For the sequence in the th Based on the predicted points, construct the following sample: Long window : Take time The three-dimensional feature sequence after internal normalization; short window :Pick Middle position The subsequence, i.e., the nearest Features of each time step; Target value : for time place Truth value.

[0039] The physical significance of the dual-window design is that when the sampling interval is 1 second, the long window covers approximately... A time series of minutes, making Capable of capturing the global dynamic patterns of a battery throughout a complete charge and discharge cycle; short window coverage of the nearest... ,make Capable of finely depicting rapid acceleration and braking. The effect of the rate of change. Both windows are from the same continuous series, requiring no additional data. It efficiently enables multi-scale modeling, allowing the model to accurately depict the impact of rapid acceleration and braking on the rate of change of SOC.

[0040] Step S2: Two-way temporal feature extraction and Initial forecast.

[0041] The model backbone consists of two parallel encoders: Encoder: Accept As input For batch size. First, use linear projection to... 3D feature mapping to The latent space is added, and a learnable positional encoding is incorporated ( ); then after Layers are composed of multi-head self-attention ( ) and feedforward network constitute Encoder layer; extracting the hidden states at the final time step. As a global feature representation.

[0042] Temporal Convolutional Networks: Acceptance As input. Employing dilated causal convolution (... )of Structure, expansion rate The exponential increase ensures that the receptive field covers the entire area. Each time step; extract features from the last time step of the output sequence. .

[0043] Feature fusion: and After concatenation, the output is mapped through a fully connected layer. Initial predicted value .

[0044] The mathematical expression for the self-attention mechanism is:

[0045] in These are query, key, and value matrices, respectively. As the key dimension. Multi-head attention through Multiple parallel attention heads capture the temporal dependencies in different subspaces, and the calculation formula is as follows:

[0046] in .

[0047] Step S3: Temperature-adaptive physical parameter adaptation network.

[0048] See Figure 3 The figure illustrates the overall structure of the temperature-adaptive physical parameter adaptation network and its connection to the equivalent circuit model. This network takes battery temperature information as input and generates key parameters such as battery capacity, internal resistance, and coulombic efficiency through a temperature-related physical parameter mapping module. These parameters are then further incorporated into the equivalent circuit model and the physical information loss function to constrain the SOC prediction process. Through this structure, the model can adaptively adjust physical parameters according to different temperature conditions, enabling subsequent SOC estimation to simultaneously possess the nonlinear expressiveness of a data-driven model and the interpretability of a battery physics model.

[0049] The physical parameter adaptation network consists of three monotonic neural networks and an equivalent circuit residual compensation network, which realizes temperature adaptive modeling of battery physical characteristics.

[0050] Monotonic Neural Networks: Using Normalized Temperature For input, by using The function processes the original weights to force the weights of each layer of the network to be positive, thereby achieving monotonicity of the output with respect to the input.

[0051] Battery capacity function Modeled using a monotonically increasing network, it reflects the physical law of the decrease in usable battery capacity at low temperatures.

[0052] Internal resistance function Modeled using a monotonically decreasing network, it reflects the increasing internal resistance at low temperatures.

[0053] Coulomb efficiency function Modeled using a monotonically increasing network, it reflects the law of decreasing charging and discharging efficiency at low temperatures.

[0054] This conforms to the physical law that the usable capacity of lithium-ion batteries decreases at low temperatures but recovers as the temperature rises; the output of the R(T) network decreases monotonically with temperature, which conforms to the physical law that the internal resistance of lithium-ion batteries increases at low temperatures but decreases as the temperature rises. The monotonicity constraint can effectively prevent model extrapolation failure under extreme low and high temperature conditions where data is scarce.

[0055] All three monotonic networks use input Dimensions, Hidden Layers Dimensions, Output A small-scale structure with a total number of parameters accounting for approximately [percentage missing] of the total model parameters. .

[0056] Equivalent circuit model : Use first order The equivalent circuit, whose terminal voltage model is represented as:

[0057] in for The voltages of the parallel circuits satisfy a first-order recurrence relation:

[0058] Its parameters are defined as follows:

[0059] Model: Pre-trained The network will With temperature as input, the corresponding open-circuit voltage is output. The network's parameters are frozen during the inference phase.

[0060] Voltage residual compensation network: Introducing a lightweight residual network right Reconstructing the voltage for compensation and correcting system errors, the formula is updated as follows:

[0061] Step S4: Design and training of the physical information loss function.

[0062] The physical information loss function designed in this invention consists of a weighted combination of four terms, and its overall form is as follows: (7) Data supervision loss :for Predicted value Compared with the true value Mean square error between:

[0063] Set to 1.0.

[0064] Coulomb count consistency constraint Random sampling within a short window Anchor points Calculate and predict SOC in Change within a step And the theoretical change calculated based on the Coulomb counting equation:

[0065] in and The learnable physical parameters output by the temperature compensation module. The sampling interval is denoted as .

[0066] To prevent gradient explosion caused by extreme current integral values ​​(such as large current pulses under US06 conditions), this invention normalizes the relative error of the variation deviation before applying Huber loss, as shown in the following formula:

[0067] in This indicates that a gradient truncation operation is applied to the denominator to prevent the normalization process itself from participating in backpropagation and causing the denominator to be maliciously optimized to zero. To prevent division-by-zero errors in the numerical stability term, the inflection point parameter of the Huber loss is set to... The loss function degenerates into squared loss when the error is small and into linear loss when the error is large, thereby further suppressing abnormal gradients.

[0068] The consistency constraint applies only to Applying a short-term window effectively blocks the long-term cumulative effect of Coulomb integration error, and the corresponding loss weight is set as follows: .

[0069] Voltage consistency constraints :right The reconstructed terminal voltage was compared with the measured terminal voltage using... Loss to enhance robustness:

[0070] set up .

[0071] Physical parameter regularization :constraint Implemented in the form of a penalty function, setting... .

[0072] Temperature adaptive weighting :

[0073] in , , .exist hour, The physical constraint weights are automatically reduced. ;exist hour, The physical constraints are close to normal strength.

[0074] Preheating mechanism At the beginning of training Within steps, from linear growth to .

[0075] Step S5: Adaptive fine-tuning across operating conditions.

[0076] See Figure 4 The figure illustrates the overall process of the cross-operating condition adaptive fine-tuning strategy of this invention. The model is first trained on source domain operating condition data to learn the general time-series features and temperature-related physical parameter expressions required for battery SOC estimation. Then, when transferring to the target operating condition, an efficient parameter fine-tuning method is used, updating only the physical parameters closely related to the target operating condition, while keeping the core time-series feature extraction network parameters unchanged. Through this process, the model can complete cross-operating condition adaptation with a small amount of labeled target domain data, reducing the amount of data and computational cost required to retrain the entire network.

[0077] The specific modules of the cross-condition adaptive fine-tuning strategy framework are explained as follows: Parameter Freeze: Freeze Encoder All parameters of the feature fusion layer are unfrozen and updated only in the physical parameter adaptation network. , , The number of fine-tunable parameters in the monotonic network and voltage residual compensation network accounts for less than 5% of the total number of parameters in the model, that is, about 150,000 learnable weights need to be fine-tuned, while the approximately 3 million parameters of the backbone network remain frozen.

[0078] Target domain data sampling: using Layered sampling strategy, The interval is divided into equal parts Uniform sampling is performed across all intervals, ensuring that the fine-tuning samples cover the entire range. Range. The size of the labeled samples used for fine-tuning is controlled within the total data of the target domain. Within.

[0079] Fine-tuning training: Fine-tuning the unfrozen parameters with a small learning rate, turning off the warm-up mechanism, and directly using the full physical information loss function for training.

[0080] Migration logic description: The backbone encoder is already in... The model has acquired a general ability to extract time-series features under various operating conditions. The main difference between the FUDS and US06 operating conditions and the DST operating condition lies in the different dynamic characteristics of the current. This difference is captured by the physical parameter adaptation network. Since the physical parameters satisfy the monotonicity constraint and have a clear physical meaning, the model only needs a very small amount of data to complete cross-domain adaptation.

[0081] Experimental verification and effect analysis The method of this invention was experimentally verified using DST, FUDS, and US06 data of lithium-ion lithium iron phosphate batteries at temperatures of -10°C, 0°C, 25°C, 40°C, and 50°C. Evaluation metrics included root mean square error (RMSE), mean absolute error (MAE), and maximum error (MaxErr), all expressed as %SOC.

[0082] Low-temperature performance and cross-cell generalization are shown in Table 1 below:

[0083] See attached document Figure 5 As shown in Table 1, under extremely harsh conditions of -10°C inter-cell migration, the SOC estimation RMSE of the method of this invention is approximately 8.96%. In contrast, the traditional end-to-end method overfits to local noise and suffers catastrophic forgetting under small sample conditions, with an RMSE as high as 14.25%. Referring to Table 1, this invention reduces the estimation error in extreme low-temperature scenarios by approximately 37% by using physical information loss. Under ambient temperature of 25°C and mild low temperature of 0°C, the method of this invention remains stable, with RMSEs controlled at 2.35% and 4.27%, respectively, significantly better than the 6.45% and 9.30% of direct inter-cell migration without fine-tuning, perfectly meeting the engineering robustness accuracy requirements in uncontrolled environments.

[0084] Cross-condition generalization validation: To verify the model's ability to handle severe current fluctuations, in the DST→FUDS transfer scenario, the baseline RMSE of direct transfer was approximately 3.54%, and fine-tuning was performed using only 5% of the target domain labeled data. Referring to Table 2, the overall RMSE of the FUDS method of this invention was reduced to 3.15%. For the DST→US06 scenario, where acceleration and deceleration are more frequent and severe, the direct transfer RMSE was 4.14%. After fine-tuning, this invention successfully compressed the error to 3.77%, as shown in Table 2 below.

[0085] The above experimental results verify that the loss function containing physical information combined with the parameter efficient transfer mechanism proposed in this invention can not only overcome the low-temperature failure problem of traditional methods in cross-cell transfer with very little target domain data, but also demonstrate excellent robustness across strong dynamic conditions.

[0086] Based on the above embodiments, the present invention also provides a method for migrating the state of charge (SOC) of a lithium-ion battery across operating conditions based on a temperature-adaptive physical parameter adaptation network. Estimation systems include: Data preprocessing module: used to perform global standardization and normalization on battery time series data under multiple temperatures and operating conditions, and to construct a dual-window sliding sample. Temporal feature extraction module: includes a Transformer encoder for processing long window samples and a TCN for processing short window samples, used to extract multi-scale temporal features and output initial SOC prediction values; Physical parameter adaptation module: includes a monotonic neural network and an ECM residual compensation network for temperature modeling of battery capacity Q(T), internal resistance R(T) and coulombic efficiency η(T); Physical information loss function training module: The system is trained end-to-end using a weighted combination loss of data supervision loss, coulomb count consistency constraint, voltage consistency constraint and physical parameter regularization term, and low temperature robustness is improved by temperature adaptive weights and preheating mechanism; Domain Adaptive Fine-Tuning Module: Under the premise of freezing the trunk encoder parameters, only the physical parameter adaptation module is fine-tuned in the target domain to achieve cross-operating condition migration generalization.

[0087] The present invention also provides an electronic device, including a processor and a memory, wherein the memory is used to store at least one executable instruction, and the processor is used to execute the executable instruction to implement the above-described state of charge estimation method.

[0088] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described state of charge estimation method.

[0089] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for estimating the state of charge (SOC) of a lithium-ion battery across operating conditions based on a temperature-adaptive physical parameter adaptation network, characterized in that, Includes the following steps: Step 1: Preprocess the time series data of lithium-ion batteries collected under different temperature conditions and different driving conditions, extract the three-dimensional feature sequences of current, voltage and temperature, normalize the features using a global standardization method based on full-condition data, and construct long window samples and short window samples using a dual-window sliding sampling strategy. The long window is used to capture the long-range global dependence of the battery, and the short window is used to focus on the recent fine dynamics. Step 2: Input the long window samples into the Transformer temporal backbone encoder to extract global temporal feature representations; Short-window samples are input into a temporal convolutional network to extract recent local dynamic features; the two feature outputs are then fused. Initial predicted value; Step 3: Establish a temperature-adaptive equivalent circuit model using a learnable physical parameter adaptation network, wherein the physical parameter adaptation network includes information about battery capacity. Internal resistance Coulomb efficiency A monotonic neural network that models the relationship between temperature and temperature variations; Step 4: Train the model using the physical information loss function, which is a weighted combination of four parts: data supervision loss, coulomb count consistency constraint, voltage consistency constraint, and physical parameter regularization term. Step 5: Based on the model trained under source domain conditions, freeze the backbone encoder weights using an efficient parameter fine-tuning strategy, and perform target domain adaptive fine-tuning on the physical parameter adaptation network to achieve adaptation to different operating conditions and temperatures. Estimated cross-domain generalization.

2. The method according to claim 1, characterized in that, The data preprocessing method described in step 1 is as follows: Collect raw data of the target battery under DST, FUDS and US06 conditions across the entire temperature range, calculate the global mean and standard deviation of three-dimensional characteristic current, voltage and temperature, and normalize all features using a standardized formula to ensure that the source domain and target domain share a unified feature distribution space. Compared to methods that only use source domain training data to fit normalized parameters, the proposed method effectively avoids the target domain data from exceeding the normalization range under high dynamic conditions, improving the robustness of cross-condition estimation and allowing for the setting of a long window length. Each time step, short window length The nth time step; for the nth time step in the time series; One prediction point, long window for time measurement All three-dimensional features within, short window takes the last part of long window. Each time step, i.e., location Subsequence at the target; The truth of every moment The two windows share the same feature data. The shorter window is a subset of the longer window, so no additional data needs to be read, thus achieving efficient multi-scale time series modeling.

3. The method according to claim 1, characterized in that, The monotonic neural network mentioned in step 3 is specifically: Normalized battery temperature For input, use The activation function processes the weights, ensuring they remain positive to achieve monotonicity of the output with respect to the input; where... Network and Network output increases monotonically with temperature. The network output decreases monotonically with temperature, which is consistent with the physical law that lithium-ion batteries experience capacity decay, increased internal resistance, and reduced efficiency at low temperatures; the monotonicity constraint prevents model extrapolation failure under extreme low and high temperature conditions where data is scarce.

4. The method according to claim 1, characterized in that, The specific expression for the physical information loss function mentioned in step 5 is as follows: ； in For prediction With reality The mean squared error loss; To address the consistency loss of the Coulomb count, adjacent time steps are constrained based on the Coulomb count equation. Consistency of rate of change; For voltage consistency loss, prediction is based on equivalent circuit model constraints. The error between the reconstructed terminal voltage and the measured terminal voltage is calculated using... Loss is used to enhance robustness to outliers; For physical parameter regularization terms, constraints , , Within reasonable physical limits; This is a temperature-adaptive weighting function that applies smaller weights to physical constraint terms under low-temperature conditions, thus avoiding the negative impact of physical model biases on performance under these conditions. It is estimated that negative interference will occur; , This is an adjustable loss weighting coefficient.

5. The method according to claim 4, characterized in that, The temperature adaptive weighting function Defined as: ； in This represents the average battery temperature of the current batch of samples. This is the lower limit of the weighting at low temperatures. To adjust the slope, This is the neutral temperature reference point; under extremely low temperatures... Approaching This significantly reduces the physical constraint loss, allowing data to drive the update of the loss-dominant parameters; at room temperature Approaching The physical constraints are functioning normally.

6. The method according to claim 1, characterized in that, The model training described in step 5 also includes a warm-up mechanism: During the initial warm-up phase of training, the weighting coefficients... from linear growth to The effective weights of the physical constraint terms are After the preheating phase ends Hengwei The preheating mechanism ensures that the model prioritizes data supervision loss during the parameter initialization phase. Once the prediction converges to a reasonable region, physical constraints are gradually introduced to avoid unstable gradients caused by physical constraint loss in the early stages of training.

7. The method according to claim 1, characterized in that, The efficient parameter fine-tuning strategy described in step 5 is as follows: Freeze all parameters of the Transformer encoder and temporal convolutional network, and only unfreeze and update the physical parameters in the adaptation network. , , The parameters of the monotonic network and the parameters of the equivalent circuit residual compensation network; the number of fine-tunable parameters accounts for approximately [a certain percentage] of the total number of model parameters. Within; fine-tuning data adopted Stratified sampling, uniformly covering the target domain dataset of The interval, with a small number of labeled samples, achieves efficient cross-condition adaptation.

8. A system for implementing the method of any one of claims 1-7, characterized in that, include: The data preprocessing module is used to perform global standardization and normalization on battery time series data under multiple temperatures and operating conditions, and to construct a dual-window sliding sample. The temporal feature extraction module includes a Transformer encoder for processing long window samples and a temporal convolutional network for processing short window samples, which is used to extract multi-scale temporal features and generate initial SOC prediction values. Physical parameter adaptation module, including battery capacity Internal resistance Coulomb efficiency A monotonic neural network for temperature modeling is used to achieve adaptive characterization of battery physical properties as temperature changes. The physical loss training module is used to train the system end-to-end with a weighted combination loss of data supervision loss, coulomb count consistency constraint, voltage consistency constraint and physical parameter regularization term, and improves low temperature robustness through temperature adaptive weights and preheating mechanism. The domain adaptive fine-tuning module is used to fine-tune the target domain only for the physical parameter adaptation module while freezing the trunk encoder parameters, so as to achieve cross-operating condition migration generalization.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; the memory is used to store at least one executable instruction, and the processor is used to execute the executable instruction to implement the method for estimating the state of charge of a lithium-ion battery across operating conditions based on a temperature-adaptive physical parameter adaptation network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for estimating the state of charge of a lithium-ion battery across operating conditions based on a temperature-adaptive physical parameter adaptation network as described in any one of claims 1 to 7.

Images