A wide-temperature-range physical-guided battery state-of-charge estimation method and system

Abstract

The application provides a wide-temperature-range physical guiding battery state-of-charge estimation method and system, and relates to the field of electric variable measurement.The method comprises the following steps: constructing multiple sets of sample data covering a target temperature range, wherein each set of sample data comprises a multi-dimensional feature tensor, and the label is state-of-charge; constructing a dual-channel physical-data decoupling network comprising at least a backbone channel, a compensation channel and a voltage reconstruction layer, wherein the backbone channel is used to generate a reference state-of-charge, the compensation channel is used to output a nonlinear voltage compensation residual error, and the voltage reconstruction layer is used to generate an estimated terminal voltage based on the reference state-of-charge and the nonlinear voltage compensation residual error; training the dual-channel physical-data decoupling network through the multiple sets of sample data; obtaining the multi-dimensional feature tensor of a battery to be evaluated; and estimating the state-of-charge of the battery to be evaluated through the trained dual-channel physical-data decoupling network, thereby improving the accuracy of battery state-of-charge estimation.

Description

Technical Field

[0001] This invention relates to the field of electrical variable measurement, and in particular to a method and system for estimating the state of charge of a battery with wide temperature range physical guidance. Background Technology

[0002] In recent years, with the rapid development of new energy vehicles and large-scale energy storage industries, lithium-ion batteries, as their core power source, have seen performance optimization and accurate state of charge estimation become key areas of technological breakthroughs in the industry. Among the many key technologies of BMS, accurate estimation of SOC (State of Charge) is the foundation for achieving efficient and safe battery operation, directly affecting the system's energy dispatch efficiency and reliability.

[0003] Currently, while mechanistic models such as the ampere-hour integral method and extended Kalman filtering (EKF) have shown some applicability in state estimation of conventional nonlinear systems, the inherent strong nonlinearity between the battery open-circuit voltage (OCV) and state of charge (SOC) still presents numerous challenges for state estimation. This is especially true in real-world operating conditions, where parameters such as battery charge transfer impedance undergo drastic dynamic changes with ambient temperature, cycle number, and current alternation, further increasing the difficulty of high-precision estimation across the entire temperature range. Regarding SOC estimation, existing technologies are primarily limited by complex and variable operating conditions and uncertainties such as environmental noise, making it difficult to obtain high-precision results under extreme conditions. Particularly when batteries are subjected to rapid discharge in extremely cold environments (minus 20 degrees Celsius), high-frequency transient current surges, or sensor zero-point drift, the pre-defined state transition matrices of existing traditional mechanistic models are prone to failure, leading to a significant decrease in estimation accuracy and stability, and severe tracking lag or divergence. Furthermore, existing pure data-driven methods often mix various features during estimation, failing to fully consider the decoupling relationship between the baseline state evolution and the nonlinear polarization residual, thus limiting the system's generalization ability when facing unknown disturbances.

[0004] Therefore, there is a need to provide a wide-temperature-range physically guided method and system for estimating the state of charge of batteries, in order to improve the accuracy of battery state of charge estimation. Summary of the Invention

[0005] This invention provides a wide-temperature-range physically guided battery state-of-charge estimation method, comprising: constructing multiple sets of sample data, wherein the multiple sets of sample data cover a target temperature range, each set of sample data includes a multidimensional feature tensor, wherein the multidimensional feature tensor includes at least terminal voltage, charge / discharge current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and state of charge after defect injection, and the label of the sample data is a reference state of charge obtained based on the ampere-hour integration method; constructing a dual-channel physical-data decoupling network, wherein the dual-channel physical-data decoupling network includes at least a main channel, a compensation channel, and a voltage reconstruction layer, wherein the main channel... The dry channel is used to extract the temporal hidden state of the multidimensional feature tensor through a long short-term memory network to generate a reference state of charge. The compensation channel is used to output a nonlinear voltage compensation residual based on the temporal hidden state of the multidimensional feature tensor through a multilayer perceptron. The voltage reconstruction layer is used to generate an estimated terminal voltage based on the reference state of charge and the nonlinear voltage compensation residual. The dual-channel physical-data decoupling network is trained using multiple sets of sample data to obtain the multidimensional feature tensor of the battery to be evaluated. The state of charge of the battery to be evaluated is estimated based on the multidimensional feature tensor of the battery to be evaluated through the trained dual-channel physical-data decoupling network.

[0006] Furthermore, the state of charge after defect injection is determined, including: a reference state of charge generated based on the ampere-hour integration method; and the state of charge after defect injection is determined based on the drift coefficient of the analog sensor gain drift, the initial deviation parameter, and the reference state of charge. , in, The state of charge after the defect is injected; For reference state of charge; This is the drift coefficient used to simulate sensor gain drift; This is the initial deviation parameter.

[0007] Furthermore, the compensation channel outputs a nonlinear voltage to compensate for the residual based on the temporal hidden state of a multilayer perceptron using a multidimensional feature tensor, including: , in, It is the hyperbolic tangent activation function; The physical boundary adjustment factor is set. To compensate for the output of the channel Nonlinear voltage compensation residual at time step; The weight matrix of the multilayer perceptron in the compensation channel; Extracted from the long short-term memory network in the backbone channel. Hidden states in time sequence This is the offset vector for the compensation channel.

[0008] Furthermore, the voltage reconstruction layer generates an estimated terminal voltage based on a reference state of charge and nonlinear voltage compensation residuals, including: Calculate the nonlinear temperature drift coefficient based on the reference state of charge: , in, for The nonlinear temperature drift coefficient at any given time; The basic reference temperature coefficient; This is the sinusoidal modulation gain factor; Main channel output The reference state of charge at time t; Based on the nonlinear temperature drift coefficient, the terminal voltage is explicitly physically reconstructed to generate an estimated terminal voltage. , in, for Estimated terminal voltage at time; Reference temperature The reference open-circuit voltage below; for The ambient temperature at that moment; for The operating current at any given time, for The dynamic internal resistance at any given time; To compensate for the nonlinear voltage residual of the compensation channel output.

[0009] Furthermore, the loss function used to train the dual-channel physical-data decoupled network includes at least the data-driven fitting loss, the terminal voltage physical constraint loss, and the state boundary constraint loss.

[0010] Furthermore, the data-driven fitting loss is: , in, For data-driven fitting loss, The total length of the sample sequences in the training batch; Predicted by the main channel The reference state of charge at time t; for The state of charge at any given moment; The physical constraint loss of the terminal voltage is: , in, For the physical constraint loss of the terminal voltage, Reconstructed for voltage reconfiguration layer Estimated terminal voltage at time; for The actual measured terminal voltage value at that moment.

[0011] Furthermore, the state boundary constraint loss is: , in, For state boundary constraint loss, This is a linear rectification activation function, used when the main channel output... Reference state of charge at time 1 A nonlinear penalty is applied when the value is less than 0 or greater than 1.

[0012] Furthermore, the loss function is: , in, This represents a combined loss across multiple tasks globally. For data fitting loss; The loss is due to the physical constraint of the terminal voltage. The loss is the state boundary constraint loss; , and Task noise parameters are automatically learned by the dual-channel physical-data decoupling network to characterize homoscedastic uncertainty and dynamically adjust the optimization ratio; This is the physical regularization loss of the open-circuit voltage, used to penalize the nonlinear voltage compensation residual amplitude of the compensation channel output.

[0013] Further, determine the terminal voltage, including: Construct a second-order RC equivalent circuit model, whose state-space equations are expressed as follows: , , in, and These are the time constants of the mid-frequency charge transfer polarization circuit and the low-frequency diffusion polarization circuit, respectively. and These are the internal resistances of mid-frequency charge transfer polarization and low-frequency diffusion polarization, respectively. and They are respectively The voltage of the intermediate frequency polarization capacitor and the voltage of the low frequency polarization capacitor at any given time; and They are respectively The corresponding intermediate frequency polarization capacitor voltage and low frequency polarization capacitor voltage at each time point; The discrete sampling time interval of the system; for The charging and discharging current at any given moment; for The battery terminal voltage at any given time; for The state of charge at any given moment; for The reference open-circuit voltage corresponding to the state of charge at any given moment; Let be the ohmic internal resistance of the battery.

[0014] This invention provides a wide-temperature-range physically guided battery state of charge estimation system, comprising: a data acquisition module for constructing multiple sets of sample data, wherein the multiple sets of sample data cover a target temperature range, each set of sample data includes a multidimensional feature tensor, wherein the multidimensional feature tensor includes at least terminal voltage, charge / discharge current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and state of charge after defect injection, and the sample data is labeled with state of charge; and a model construction module for constructing a dual-channel physical-data decoupling network, wherein the dual-channel physical-data decoupling network includes at least a backbone channel, a compensation channel, and a voltage reconstruction layer, wherein the backbone channel is used to... A Long Short-Term Memory (LSTM) network extracts the temporal hidden state of a multidimensional feature tensor to generate a baseline state of charge (STC). The compensation channel is used to output a nonlinear voltage compensation residual based on the temporal hidden state of the multidimensional feature tensor via a multilayer perceptron. The voltage reconstruction layer is used to generate an estimated terminal voltage based on the baseline STC and the nonlinear voltage compensation residual. A model training module is used to train a dual-channel physical-data decoupling network using multiple sets of sample data. A charge estimation module is used to obtain the multidimensional feature tensor of the battery to be evaluated. The trained dual-channel physical-data decoupling network estimates the STC of the battery to be evaluated based on the multidimensional feature tensor.

[0015] Compared with existing technologies, the wide-temperature-range physically guided battery state-of-charge estimation method and system provided by this invention has at least the following beneficial effects: 1. By introducing a sinusoidal modulation temperature drift coefficient and physical boundary truncation mechanism based on the Nernst equation into the compensation channel, the physical correction of temperature drift and the physical range constraint of the results are completed, which greatly improves the stability of SOC estimation under extreme temperature range. In the extreme verification scenario, the root mean square error of the estimation result can be controlled within 4.91%, and reliable estimation accuracy can still be maintained in extremely cold extrapolation scenarios such as -20℃.

[0016] 2. It significantly reduces the dependence of the dual-channel physical-data decoupling network on ideal sensor data. By artificially injecting defective prior features into the ampere-hour integral sequence to complete model training, the model learns the distribution pattern of sensor drift-type anomaly features in advance, avoiding the problem of overfitting to ideal data. It effectively improves the robustness of the dual-channel physical-data decoupling network in dealing with unknown sensor quantization drift under real and complex working conditions, and adapts to the actual conditions of sensor performance fluctuations in engineering scenarios such as vehicle-mounted applications.

[0017] 3. Relying on a dual-channel decoupled architecture and an adaptive optimization mechanism for multi-task loss with homoscedastic uncertainty, the decoupled optimization of benchmark SOC estimation and voltage compensation is achieved. This balances physical mechanism constraints with data-driven fitting capabilities, maintaining high estimation accuracy under various high-frequency dynamic real-world operating conditions. It balances estimation accuracy, robustness to extreme conditions, and engineering practicality, demonstrating significant value for practical applications. Furthermore, embedding physical mechanisms into the data-driven model framework preserves the fitting capability of pure data-driven models for complex operating conditions while ensuring model interpretability through physical constraints. This addresses the dual problems of high parameter identification difficulty in pure physical models and poor generalization and interpretability in pure data-driven models. Attached Figure Description

[0018] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is a flowchart illustrating a wide-temperature-range physically guided battery state-of-charge estimation method according to some embodiments of this specification; Figure 2 This is a schematic diagram of the structure of a dual-channel physical-data decoupling network according to some embodiments of this specification; Figure 3 This is a comparison chart of test results for Cycle 1 running conditions at 10°C, as shown in some embodiments of this specification. Figure 4 This is an error comparison chart of Cycle 1 operating conditions at 10°C, as shown in some embodiments of this specification. Figure 5 This is a comparison chart of test results under Cycle 2 operating conditions at 10°C, as shown in some embodiments of this specification; Figure 6 This is an error comparison chart of Cycle 2 operating conditions at 10°C, as shown in some embodiments of this specification. Figure 7 This is a comparison chart of test results for Cycle 1 running conditions at -20°C, as shown in some embodiments of this specification; Figure 8 This is an error comparison chart of Cycle 1 operating conditions at -20°C as shown in some embodiments of this specification; Figure 9 This is a comparison chart of test results for Cycle 2 running conditions at -20°C, as shown in some embodiments of this specification; Figure 10This is a comparison chart of errors in Cycle 2 operating conditions at -20°C, as shown in some embodiments of this specification. Figure 11 This is a comparison chart of core evaluation indicators under different temperature conditions, based on some embodiments of this specification; Figure 12 This is a comparison chart of test results under the US06 operating condition at 25°C, as shown in some embodiments of this specification; Figure 13 This is an error comparison chart of the US06 operating conditions at 25°C, as shown in some embodiments of this specification. Figure 14 This is a comparison chart of test results for UDDS operating conditions at 25°C, as shown in some embodiments of this specification. Figure 15 This is an error comparison chart of UDDS operating conditions at 25°C as shown in some embodiments of this specification; Figure 16 This is a comparison chart of test results for the LA92 operating condition at 25°C, based on some embodiments shown in this specification. Figure 17 This is an error comparison chart of the LA92 operating conditions at 25°C, as shown in some embodiments of this specification; Figure 18 This is a comparison chart of test results for HWFET operating conditions at 25°C, based on some embodiments shown in this specification; Figure 19 This is an error comparison chart of HWFET operating conditions at 25°C as shown in some embodiments of this specification; Figure 20 This is a comparison chart of core evaluation indicators under different dynamic test conditions, based on some embodiments of this specification; Figure 21 This is a schematic diagram of a wide-temperature-range physically guided battery state-of-charge estimation system according to some embodiments of this specification. Detailed Implementation

[0019] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0020] Figure 1This is a flowchart illustrating a wide-temperature-range physically guided battery state-of-charge estimation method according to some embodiments of this specification, such as... Figure 1 As shown, a method for estimating the state of charge of a battery with a wide temperature range physically guided charge may include the following steps.

[0021] S110. Construct multiple sets of sample data.

[0022] The dataset consists of multiple sets of sample data covering the target temperature range. Each set of sample data includes a multidimensional feature tensor, which includes at least the terminal voltage. Charging and discharging current Ambient temperature Intermediate frequency charge transfer polarization internal resistance Low-frequency diffusion polarization internal resistance and the state of charge after injection of defects The label for the sample data is the state of charge. The labels for the sample data can be obtained from offline datasets, which provides a more accurate SOC compared to the ampere-hour integration method.

[0023] Specifically, high-precision sensors can be used to collect real-time raw state data of the vehicle or energy storage system during operation over a wide temperature range (i.e., the target temperature range, for example, -20℃ to 25℃), such as terminal voltage, charging and discharging current, and ambient temperature. To overcome high-frequency polarization noise, the raw state data is resampled and dynamically filtered and smoothed.

[0024] To input physical mechanism features that better represent the underlying polarization resistance into the dual-channel physical-data decoupling network, a second-order RC equivalent circuit model is constructed, whose state-space equation is expressed as: , , in, and These are the time constants of the mid-frequency charge transfer polarization circuit and the low-frequency diffusion polarization circuit, respectively. and These are the internal resistances of mid-frequency charge transfer polarization and low-frequency diffusion polarization, respectively. , They are respectively , The corresponding polarization capacitor, , , , , The initial value can be set based on experience or calibration, and the specific value can be obtained online by adaptive extended Kalman filtering based on the terminal voltage information; and They are respectively The voltage of the intermediate frequency polarization capacitor and the voltage of the low frequency polarization capacitor at any given time; and They are respectively The corresponding intermediate frequency polarization capacitor voltage and low frequency polarization capacitor voltage at each time point; The discrete sampling time interval of the system; for The charging and discharging current at any given moment; for The battery terminal voltage at any given time; for The state of charge at any given moment; for The reference open-circuit voltage corresponding to the state of charge at any given time. The specific value is calculated through a pre-constructed OCV-SOC calibration function. Specifically, data points near the open-circuit state are first selected from battery experimental data, such as stable segments where the absolute current value is less than a set threshold and the current change rate is small, and the corresponding SOC and terminal voltage data are extracted. Then, a polynomial fitting is used to obtain the OCV-SOC relationship curve, preferably a sixth-order polynomial fitting is used to obtain the OCV-SOC calibration function. For State of charge at time t Substituting it into the OCV-SOC calibration function, we obtain... The specific value; Let Ω be the internal resistance of the battery. An Adaptive Extended Kalman Filter (AEKF) algorithm is introduced at each time sampling point. The system dynamically performs state prediction, covariance prediction, innovation calculation, Kalman gain update, and state covariance update, thereby enabling online identification of key dynamic physical parameters in the equivalent circuit model and dynamically outputting the impedance characteristic matrix at the current moment, specifically including the intermediate frequency charge transfer polarization internal resistance. With low-frequency diffusion polarization internal resistance Subsequent interaction with ohmic internal resistance Together they constitute dynamic internal resistance It is used to generate estimated terminal voltages and construct physical constraint losses for terminal voltages in the voltage reconstruction layer.

[0025] In the second-order RC equivalent circuit model These are state variables in the AEKF state vector. At each sampling time, they are first predicted by the current integral state equation, and then corrected by combining the measured terminal voltage information for online identification. , , , Physical parameters such as voltage, current, temperature, polarization resistance identified by AEKF, and state of charge after defect injection are input to the subsequent dual-channel physical-data decoupling network for SOC estimation. The voltage reconstruction layer uses the baseline SOC output from the backbone channel of the dual-channel physical-data decoupling network to generate the estimated terminal voltage. Then compare it with the actual measured voltage at the sensor terminal. Construct physical constraint loss.

[0026] In some embodiments, to completely break the reliance of traditional pure data-driven models on shortcut learning of a single high-cleanliness time series feature, an error feature injection mechanism is introduced to determine the charged state after injecting defects, including: Ideal state of charge generated based on ampere-hour integration method; Based on the drift coefficient of the simulated sensor gain drift, the initial deviation parameter, and the ideal state of charge, the state of charge after defect injection is determined: , in, The state of charge after the defect is injected; This is a reference state of charge generated based on the sensor's measured current value using the ampere-hour integration method; This is the drift coefficient used to simulate sensor gain drift; This is the initial deviation parameter. Preferably, to simulate the initial sensor reading deviation during the cold start process of the battery management system, the initial deviation parameter is... The setting is fixed at -5%; simultaneously, to simulate the long-cycle hardware gain drift of the sensor as the operating cycle extends, the sensor gain drift coefficient is adjusted. Set to 2%.

[0027] Terminal voltage Charging and discharging current Ambient temperature Intermediate frequency charge transfer polarization internal resistance Low-frequency diffusion polarization internal resistance and the state of charge after the injection of defects. They are collectively encapsulated and constructed into a structured six-dimensional temporal feature parameter tensor. Its mathematical expression is: , In the formula, for The six-dimensional temporal feature parameter tensor constructed at each moment; , as well as They are respectively The terminal voltage, charging / discharging current, and ambient temperature at any given time; and They are respectively The mid-frequency charge transfer polarization internal resistance and the low-frequency diffusion polarization internal resistance at time t; for The state of charge after injecting the defect at any given moment; This is the symbol for the matrix transpose operation.

[0028] Six-dimensional temporal feature parameter tensor Subsequently, the global boundary conditions are input into the dual-channel physical-data decoupling network as forward inference. By deliberately injecting prior electrical features with obvious defects, the subsequent dual-channel physical-data decoupling network is forced to spontaneously activate the cross-validation logic of the opposite-end voltage conservation mechanism and the impedance physical polarization boundary when performing state tracking. This avoids the risk of unbounded divergence of the dual-channel physical-data decoupling network when encountering unknown noise from outside the domain at the underlying logic level.

[0029] S120. Construct a dual-channel physical-data decoupling network.

[0030] The dual-channel physical-data decoupling network includes at least a backbone channel, a compensation channel, and a voltage reconstruction layer. The backbone channel is used to dynamically maintain and update the temporal hidden state of the multidimensional feature tensor through a long short-term memory network, based on the synergistic effect of the forget gate, input gate, and output gate. This allows for the deep extraction of historical electrochemical evolution information of the battery under alternating dynamic current impacts, including its temporal hidden states. After mapping through a fully connected layer, the output is a reference state of charge that follows the macroscopic principle of charge conservation. The compensation channel is used to capture and quantitatively compensate for the residual nonlinear error generated by the deep learning black box mapping through the multilayer perceptron. Based on the temporal hidden state of the multidimensional feature tensor, it outputs the nonlinear voltage compensation residual. The voltage reconstruction layer is used to generate the estimated terminal voltage based on the reference state of charge and the nonlinear voltage compensation residual.

[0031] In some embodiments, to restrict the fitting degrees of freedom of the black-box network within a reasonable electrochemical thermodynamic polarization threshold, an asymptotic saturation activation function and a physical boundary truncation mechanism are innovatively set at the output of the compensation channel. The compensation channel outputs a nonlinear voltage to compensate for the residual based on the temporal hidden state of the multi-dimensional feature tensor through a multilayer perceptron, including: , in, It is the hyperbolic tangent activation function; The physical boundary adjustment factor is set. To compensate for the output of the channel Nonlinear voltage compensation residual at time step; The weight matrix of the multilayer perceptron in the compensation channel; Extracted from the long short-term memory network in the backbone channel. Hidden states in time sequence This is the bias vector for the compensation channel. Preferably, The voltage is 0.1V, and the Tanh activation function is used at the end of the compensation channel, with an output range of [value missing]. Therefore when At that time, nonlinear voltage compensation residual Restricted to That is, within ±100mV. This cutoff mechanism constrains the compensation network's correction limit for the open-circuit voltage (OCV) curve to a reasonable polarization error range of 100mV, effectively avoiding the neural network from outputting distorted voltage values ​​that significantly violate the mechanism under unknown operating conditions.

[0032] In some embodiments, to endow the state-aware architecture with mechanistic extrapolation capabilities over a wide temperature range, a physical bounding compensation theorem based on the Nernst equation is embedded in the voltage reconstruction layer. The voltage reconstruction layer generates an estimated terminal voltage based on a reference state of charge and nonlinear voltage compensation residuals, including: To address the nonlinear evolution of the reaction entropy change with charge state during the dynamic lithium insertion / extraction process of micron-scale electrode layered materials, a nonlinear temperature drift coefficient modulated by the current estimated state of charge was constructed. The nonlinear temperature drift coefficient was calculated based on a reference state of charge using the following formula: , in, For the recipient Nonlinear temperature drift coefficient of time-referenced state of charge modulation; To be at the standard reference temperature The underlying reference temperature coefficient; This is the sinusoidal modulation gain factor; Main channel output The reference state of charge at time t. This sinusoidal modulation curve can approximately characterize the true thermodynamic properties of layered micromaterials in the middle charge range (medium charge) and the two end charge ranges (full charge and full discharge boundaries).

[0033] In some embodiments, based on the above physical mechanism, the dual-channel physical-data decoupling network is currently... Explicit physical reconstruction of the terminal voltage is performed at all times, and its estimated terminal voltage is... The mathematical reconstruction equation is: , in, for Estimated terminal voltage at time; Reference temperature The reference open-circuit voltage below; for The ambient temperature at that moment; for The operating current at any given time, for The dynamic internal resistance at time t is calculated by summing the ohmic internal resistance and polarization internal resistance obtained from the aforementioned online identification, i.e. ; To compensate for the nonlinear voltage residuals of the compensation channel output, this reconstruction equation seamlessly algebraically fuses the residuals output by the black-box neural network with the electrochemical thermodynamic evolution formula.

[0034] S130. Train the dual-channel physical-data decoupling network using multiple sets of sample data.

[0035] In some embodiments, to coordinate the balance between the pure data fitting accuracy of deep learning and the consistency constraints of thermodynamic physical mechanisms during parameter training backpropagation, the loss function used to train the dual-channel physical-data decoupled network includes at least a data-driven fitting loss, a terminal voltage physical constraint loss, and a state boundary constraint loss. The data-driven fitting loss constrains the approximation of the backbone channel output to the true label; the terminal voltage physical constraint loss constrains the consistency between the reconstructed voltage and the sensor-measured voltage, constructing a spatial energy conservation bound; and the state boundary constraint loss penalizes output over-boundation through a modified linear unit function, ensuring the physical rationality of the state parameters.

[0036] The data-driven fitting loss is: , in, For data-driven fitting loss, The total length of the sample sequences in the training batch; Predicted by the main channel The state of charge at any given moment; for The state of charge at any given moment, i.e., the tag; The physical constraint loss of the terminal voltage is: , in, For the physical constraint loss of the terminal voltage, Reconstructed for voltage reconfiguration layer Estimated terminal voltage at time; for The actual terminal voltage value measured by the sensor at any time, and the physical constraint loss are compared. and The difference indicates that the constrained network prediction results are consistent with the terminal voltage observations.

[0037] The state boundary constraint loss is: , in, For state boundary constraint loss, This is a linear rectification activation function, used when the main channel output... Reference state of charge at time 1 A nonlinear penalty is applied when the value is less than 0 or greater than 1.

[0038] To completely avoid the optimization imbalance caused by manually adjusting various loss weights during backpropagation parameter optimization, this embodiment introduces a three-dimensional adaptive weight adjuster based on homoscedasticity uncertainty into the global training architecture. The loss function is: , in, This represents a combined loss across multiple tasks globally. For data fitting loss; The loss is due to the physical constraint of the terminal voltage. The loss is the state boundary constraint loss; , and The task noise parameters are automatically learned by the dual-channel physical-data decoupling network to characterize homoscedastic uncertainty and dynamically adjust the optimization ratio. , and The variances represent the uncertainty of the data fitting task and the physical interpretable constraint task, respectively, as the network learns and updates its weights spontaneously and online during forward propagation training. The open-circuit voltage physical regularization loss is used to penalize the nonlinear voltage compensation residual amplitude of the compensation channel output, ensuring that the output residual compensation amount is anchored within a thermodynamically / electrochemically reasonable threshold range.

[0039] , In a preferred embodiment, to ensure that the output nonlinear voltage compensation residual is stably anchored within a reasonable electrochemical threshold range, the penalty weight parameter of the regularization loss term is... The preferred setting is 0.1.

[0040] By utilizing a trained dual-channel physical-data decoupling network, and inputting real-time vehicle sensor test condition sequences in practical engineering applications, a stable, high-precision, and divergent real-time output of a robust state-of-charge estimate of the target battery can be achieved.

[0041] S140. Obtain the multidimensional feature tensor of the battery to be evaluated.

[0042] Specifically, the multidimensional feature tensor includes at least the terminal voltage, charging and discharging current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and the state of charge after defect injection.

[0043] S150. Using a trained dual-channel physical-data decoupling network, the state of charge of the battery to be evaluated is estimated based on the multidimensional feature tensor of the battery to be evaluated.

[0044] Specifically, the trained dual-channel physical-data decoupling network, including at least the backbone channel, estimates the state of charge of the battery to be evaluated based on the multidimensional feature tensor of the battery to be evaluated.

[0045] The following section, based on experiments, explains the beneficial effects of a wide-temperature-range physical-guided battery state-of-charge estimation method.

[0046] The test subject in this experiment was a Panasonic NCR18650PF cylindrical battery cell (rated capacity of 2.9Ah). The dataset covers a wide temperature gradient from 25℃ to -20℃, as well as various standard electric vehicle driving conditions including US06, UDDS, LA92, and HWFET.

[0047] Adaptive Kalman Filter (AEKF) and Unscented Kalman Filter (UKF), representing traditional physical mechanism models, and Temporal Convolutional Network (TCN) and Transformer architecture, representing pure data-driven models, were selected as comparison baselines. Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Maximum Absolute Error (MAXe) were uniformly used as core evaluation metrics.

[0048] Combination Figures 3 to 11 The experimental data shown are used to evaluate the cross-temperature domain performance of the sinusoidal modulation temperature compensation mechanism and physical boundary truncation mechanism in this method under extreme thermodynamic boundaries. Significant differences were observed in the error statistics of each model under the 10℃ test set used for interpolation verification and the -20℃ test set used for extrapolation verification.

[0049] Experimental data show that under extremely cold conditions of -20℃ with insufficient feature data coverage, taking the cycle2 test sequence as an example, due to intensified nonlinear polarization, the RMSE of the traditional AEKF model soars to 41.95%, while the RMSEs of the purely data-driven TCN and Transformer models also diverge to 40.70% and 30.87%, respectively, exhibiting severe unbounded divergence. In contrast, the dual-channel physical-data decoupling network proposed in this method effectively suppresses the divergence trend by relying on the physical bounds of the Nernst equation and the constraints of the adaptive loss function, strictly controlling the RMSE to 3.78% and the MAE to 3.10% under the same extremely cold conditions. This result strongly confirms the high reliability and physical mechanism interpretability of the proposed dual-channel physical-data decoupling network in extremely cold extrapolation scenarios.

[0050] Combination Figures 12 to 20 The error evolution curves and statistical indicators shown indicate that in actual vehicle operation, batteries inevitably encounter severe and random high-frequency transient shocks. To comprehensively verify the robustness of the proposed dual-channel physical-data decoupling network in tracking unknown dynamics under highly random current excitation, cross-condition sequence generalization tests were conducted at room temperature. The proposed dual-channel physical-data decoupling network performed forward inference on complete standard continuous driving conditions such as US06, UDDS, LA92, and HWFET, which were not directly involved in training.

[0051] Test results show that, facing dense rapid acceleration and regenerative braking step currents, the proposed dual-channel physics-data decoupling network effectively overcomes the tracking hysteresis of the mechanistic observer and the overfitting bottleneck of the black-box model. Under the drastically alternating US06 condition, the proposed dual-channel physics-data decoupling network achieves an RMSE of only 2.19% and a MAE of 1.98%; under the LA92 condition, the RMSE further decreases to 2.04%. In all test conditions, the core error metrics of the proposed dual-channel physics-data decoupling network comprehensively and significantly outperform the traditional physical model and the deep learning baseline model without physical constraints, demonstrating excellent resistance to out-of-domain distribution migration and high dynamic tracking accuracy.

[0052] Figure 21 This is a schematic diagram of a wide-temperature-range physically guided battery state-of-charge estimation system according to some embodiments of this specification, such as... Figure 21 As shown, a wide-temperature-range physical-guided battery state-of-charge estimation system includes a data acquisition module, a model building module, a model training module, and a charge estimation module.

[0053] The data acquisition module is used to construct multiple sets of sample data, which cover the target temperature range. Each set of sample data includes a multidimensional feature tensor, which includes at least the terminal voltage, charging and discharging current, ambient temperature, mid-frequency charge transfer polarization internal resistance, low-frequency diffusion polarization internal resistance, and state of charge after defect injection. The label of the sample data is the state of charge, and the state of charge after defect injection is used as one of the input features. The model building module is used to construct a dual-channel physical-data decoupling network. The dual-channel physical-data decoupling network includes at least a backbone channel, a compensation channel, and a voltage reconstruction layer. The backbone channel is used to extract the temporal hidden state of the multidimensional feature tensor through a long short-term memory network to generate a reference state of charge. The compensation channel is used to output a nonlinear voltage compensation residual based on the temporal hidden state of the multidimensional feature tensor through a multilayer perceptron. The voltage reconstruction layer is used to generate an estimated terminal voltage based on the reference state of charge and the nonlinear voltage compensation residual. The model training module is used to train a dual-channel physical-data decoupling network using multiple sets of sample data. The charge estimation module is used to obtain the multidimensional feature tensor of the battery to be evaluated. The multidimensional feature tensor includes at least the terminal voltage, charge and discharge current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and state of charge after defect injection. Based on the multidimensional feature tensor of the battery to be evaluated, the state of charge of the battery is estimated by a trained dual-channel physical-data decoupling network.

[0054] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and are considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A method for estimating the state of charge of a battery with wide temperature range physical guidance, characterized in that, include: Multiple sets of sample data are constructed, covering the target temperature range. Each set of sample data includes a multidimensional feature tensor, which includes at least terminal voltage, charging and discharging current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and state of charge after defect injection. The label of the sample data is the reference state of charge obtained based on the ampere-hour integration method. A dual-channel physical-data decoupling network is constructed, wherein the dual-channel physical-data decoupling network includes at least a backbone channel, a compensation channel, and a voltage reconstruction layer. The backbone channel is used to extract the temporal hidden state of the multidimensional feature tensor through a long short-term memory network to generate a reference state of charge. The compensation channel is used to output a nonlinear voltage compensation residual based on the temporal hidden state of the multidimensional feature tensor through a multilayer perceptron. The voltage reconstruction layer is used to generate an estimated terminal voltage based on the reference state of charge and the nonlinear voltage compensation residual. The dual-channel physical-data decoupling network was trained using multiple sets of sample data. Obtain the multidimensional feature tensor of the battery to be evaluated; The state of charge of the battery to be evaluated is estimated based on the multidimensional feature tensor of the battery after training by a dual-channel physical-data decoupling network. Determining the charge state after injecting the defect includes: Reference state of charge generated based on ampere-hour integration method; Based on the drift coefficient of the simulated sensor gain drift, the initial deviation parameter, and the reference state of charge, the state of charge after defect injection is determined: ， in, The state of charge after the defect is injected; For reference state of charge; This is the drift coefficient used to simulate sensor gain drift; These are the initial deviation parameters; The compensation channel outputs a nonlinear voltage to compensate for the residual based on the temporal hidden state of a multilayer perceptron using a multidimensional feature tensor, including: ， in, It is the hyperbolic tangent activation function; The set physical boundary adjustment factor, To compensate for the output of the channel Nonlinear voltage compensation residual at time step; The weight matrix of the multilayer perceptron in the compensation channel; Extracted from the long short-term memory network in the backbone channel. Hidden states in time sequence The offset vector for the compensation channel; The voltage reconstruction layer generates an estimated terminal voltage based on a reference state of charge and nonlinear voltage compensation residuals, including: Calculate the nonlinear temperature drift coefficient based on the reference state of charge: ， in, Let be the nonlinear temperature drift coefficient at time k; The basic reference temperature coefficient; This is the sinusoidal modulation gain factor; The reference state of charge at time k is the output of the main channel; Based on the nonlinear temperature drift coefficient, the terminal voltage is explicitly physically reconstructed to generate an estimated terminal voltage. ， in, for Estimated terminal voltage at time; Reference temperature The reference open-circuit voltage below; for The ambient temperature at that moment; for The operating current at any given time, for The dynamic internal resistance at any given time is determined by the intermediate frequency charge transfer polarization internal resistance. Low-frequency diffusion polarization internal resistance and Ohmic resistance constitute; To compensate for the residual nonlinear voltage output of the compensation channel; The loss function used to train the dual-channel physical-data decoupled network includes at least the data-driven fitting loss, the terminal voltage physical constraint loss, and the state boundary constraint loss.

2. The method for estimating the state of charge of a battery with a wide temperature range physically guided charging, as described in claim 1, is characterized in that... The data-driven fitting loss is: ， in, For data-driven fitting loss, The total length of the sample sequences in the training batch; Predicted by the main channel The reference state of charge at time t; for The state of charge at any given moment; The physical constraint loss of the terminal voltage is: ， in, The physical constraint loss is the terminal voltage loss. Reconstructed for voltage reconfiguration layer Estimated terminal voltage at time; for The actual measured terminal voltage value at that moment.

3. The method for estimating the state of charge of a battery with a wide temperature range physically guided charging, as described in claim 2, is characterized in that... The state boundary constraint loss is: ， in, For state boundary constraint loss, This is a linear rectification activation function used to determine the reference state of charge at time k when the main channel outputs. A nonlinear penalty is applied when the value is less than 0 or greater than 1.

4. The method for estimating the state of charge of a wide-temperature-range physically guided battery according to claim 3, characterized in that, The loss function is: ， in, This represents a combined loss across multiple tasks globally. For data fitting loss; The loss is due to the physical constraint of the terminal voltage. The loss is the state boundary constraint loss; , and Task noise parameters are automatically learned by the dual-channel physical-data decoupling network to characterize homoscedastic uncertainty and dynamically adjust the optimization ratio; This is the open-circuit voltage physical regularization loss, used to penalize the nonlinear voltage compensation residual amplitude of the compensation channel output.

5. The method for estimating the state of charge of a battery with wide temperature range physical guidance according to claim 1, characterized in that, Determine the terminal voltage, including: Construct a second-order RC equivalent circuit model, whose state-space equations are expressed as follows: ， in, and These are the time constants of the mid-frequency charge transfer polarization circuit and the low-frequency diffusion polarization circuit, respectively. and These are the internal resistances of mid-frequency charge transfer polarization and low-frequency diffusion polarization, respectively. and They are respectively The voltage of the intermediate frequency polarization capacitor and the voltage of the low frequency polarization capacitor at any given time; and They are respectively The corresponding intermediate frequency polarization capacitor voltage and low frequency polarization capacitor voltage at each moment; The discrete sampling time interval of the system; for The charging and discharging current at any given moment; for The battery terminal voltage at any given time; for The state of charge at any given moment; for The reference open-circuit voltage corresponding to the state of charge at any given moment; Let be the ohmic internal resistance of the battery.

6. A wide-temperature-range physically guided battery state-of-charge estimation system, characterized in that, The method for estimating the state of charge of a wide-temperature-range physically guided battery as described in claim 1 includes: The data acquisition module is used to construct multiple sets of sample data, which cover the target temperature range. Each set of sample data includes a multidimensional feature tensor, which includes at least terminal voltage, charging and discharging current, ambient temperature, mid-frequency charge transfer polarization resistance, low-frequency diffusion polarization resistance, and state of charge after defect injection. The sample data is labeled with the state of charge. The model building module is used to construct a dual-channel physical-data decoupling network. The dual-channel physical-data decoupling network includes at least a backbone channel, a compensation channel, and a voltage reconstruction layer. The backbone channel is used to extract the temporal hidden state of the multidimensional feature tensor through a long short-term memory network to generate a reference state of charge. The compensation channel is used to output a nonlinear voltage compensation residual based on the temporal hidden state of the multidimensional feature tensor through a multilayer perceptron. The voltage reconstruction layer is used to generate an estimated terminal voltage based on the reference state of charge and the nonlinear voltage compensation residual. The model training module is used to train a dual-channel physical-data decoupling network using multiple sets of sample data. The charge estimation module is used to obtain the multidimensional feature tensor of the battery to be evaluated; through a trained dual-channel physical-data decoupling network, the state of charge of the battery to be evaluated is estimated based on the multidimensional feature tensor of the battery to be evaluated.

Images