A pseudo open circuit voltage guided long short term memory network method for estimating state of charge of a power lithium battery

Abstract

The present application relates to a kind of pseudo open-circuit voltage guide long short-term memory network power lithium battery SOC estimation method, belong to lithium ion battery technical field, the method includes: by signal decomposition and steady component reconstruction strategy, pseudo open-circuit voltage is extracted online;The pseudo open-circuit voltage extracted is as key mechanism feature, and real-time current, voltage jointly constitute input feature matrix;The input feature matrix is input into LSTM neural network model, and model training is carried out;Based on the model after training, power lithium battery SOC estimation is carried out.The present application does not need additional sensor or complex battery model, only by signal decomposition and feature reconstruction under the mechanism guidance realizes feature enhancement.Experimental results show that the proposed POCVG-LSTM algorithm effectively improves the SOC estimation accuracy, and shows its feasibility and advantage in practical application.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and in particular to a method for estimating the state of charge (SOC) of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network. Background Technology

[0002] As the core energy storage unit of clean energy systems, the accurate estimation of the state of charge (SOC) of lithium-ion batteries is crucial for improving system safety, reliability, and operational economy. However, SOC cannot be directly measured by physical sensors, and the electrochemical reactions and dynamic polarization effects inside the battery are highly nonlinear and complex. Traditional model-based or data-driven estimation methods struggle to simultaneously achieve high accuracy, robustness, and adaptability under dynamic operating conditions. Although neural network-based intelligent estimation algorithms have shown some potential, their performance is still limited by insufficient input feature representation capabilities and the complex nonlinear mapping relationship between features and SOC, making it difficult to balance high-accuracy estimation with strong robustness.

[0003] To address these challenges, exploring new features suitable for neural network estimation is essential. Currently, existing neural network algorithms directly use the detection signal as input, requiring the neural network to learn the SOC estimation function from complex and variable signals. However, the multi-timescale coupled polarization effect within lithium batteries leads to strong nonlinearity in the detection signal. Therefore, exploring a strategy that can effectively deconstruct the detection signal and improve the accuracy of neural network estimation algorithms remains a pressing technical challenge.

[0004] The accuracy of State of Charge (SOC) estimation directly affects the performance of the battery management system. Existing estimation methods based on traditional mathematical models and data-driven approaches each have their own limitations. Methods based on traditional mathematical models include the ampere-hour integration method, the open-circuit voltage method, and filtering algorithms based on equivalent circuit models. The ampere-hour integration method estimates SOC through an integral algorithm, but this open-loop estimation method results in accumulated errors that cannot be corrected through feedback. The open-circuit voltage algorithm solves for SOC by offline fitting and calibration of the correspondence function between open-circuit voltage and SOC. However, obtaining the open-circuit voltage often requires a long period of rest to allow the internal electrochemical and physical processes of the battery to stabilize. This makes the open-circuit voltage algorithm unsuitable for online real-time application. Algorithms based on equivalent circuit models rely heavily on a precise description of the complex dynamic behavior inside the battery. These methods typically require constructing a high-precision model sufficient to characterize the internal electrochemical processes, physical diffusion effects, and dynamic material responses of the battery, and depend on the accurate online or offline identification of model parameters. However, in practical applications, especially under dynamically changing operating conditions and battery aging, constructing models that can accurately simulate multi-timescale physicochemical processes with stably identifiable parameters presents significant challenges, often limiting the robustness and generalization ability of such methods. Data-driven algorithms do not require precise mechanistic explanations and can directly learn the complex nonlinear mapping between input features such as voltage and current and SOC from historical data. However, model performance heavily depends on the quality of the input features. Most current studies directly use raw time-series voltage and current as inputs, and these features have a weak direct physical correlation with SOC.

[0005] Extracting input features strongly correlated with SOC can improve the estimation accuracy of neural networks while ensuring their robustness. Purely data-driven models are limited by data quality and physical consistency. Exploring new input features that can be used in neural network algorithms has gradually become a focus of researchers. Designing features using physical knowledge can enhance the model's learning ability and reduce its blind dependence on data. The construction of new input features can not only improve the estimation accuracy and robustness of neural network algorithms, but also enhance the algorithm's physical interpretation ability. Yue et al. improved the accuracy and robustness of neural network SOC estimation by adding the battery's surface temperature as a new input feature. Han et al. introduced ultrasonic detection signals to improve SOC estimation accuracy. C32 et al. used analytical equivalent circuit model parameters as new input features to increase the dimensionality of input features and provide more learning information for the neural network. Liu et al. used Gaussian filters to smooth the detection signals to construct new features. Wu et al. introduced detection stress signals as new input features. Qiu et al. proposed a composite neural network algorithm to improve estimation accuracy. Bao et al. proposed a deep learning method with dual channels, achieving accurate SOC estimation through optimization of different network model combinations. Tang et al. fused a physical analytical model and a data-driven model to estimate SOC by utilizing the Gaussian distribution fusion principle. Most existing neural network algorithms improve SOC estimation accuracy by adding sensors or constructing complex models. There is a lack of research analyzing the components of existing detection signals that are strongly correlated with SOC, starting from the internal mechanisms of lithium batteries. Summary of the Invention

[0006] This invention addresses the issue of decreased SOC estimation accuracy caused by multi-timescale coupling effects in voltage measurements under dynamic operating conditions. It proposes a steady-state component deconstruction approach, enhancing feature representation by effectively separating the pseudo-open-circuit voltage (POCV) in the coupled signal. Based on this, a PVO-guided long short-term memory (LSTM) network-based SOC estimation method for power lithium batteries is proposed. This method reconstructs the extracted POCV component and the actual measured voltage and current signals into an input feature matrix with clear physical guidance. The strong correlation between POCV and SOC is then used to guide the training and estimation of the LSTM neural network, thereby achieving high-precision and robust SOC estimation under complex dynamic conditions.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for estimating the state of charge (SOC) of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network includes: The pseudo-open circuit voltage is extracted online by using a signal decomposition and steady-state component reconstruction strategy. The extracted pseudo-open-circuit voltage is used as a key mechanism feature, which together with the real-time current and voltage constitutes the input feature matrix. The input feature matrix is ​​input into the LSTM neural network model for model training; Based on the trained model, the SOC of the power lithium battery is estimated.

[0008] Optionally, the pseudo-open-circuit voltage can be extracted online through a signal decomposition and steady-state component reconstruction strategy, including: Construct an equivalent circuit model; wherein, the equivalent circuit model is used to mathematically describe the polarization dynamics process inside the battery; The voltage components are analyzed using the equivalent circuit model; wherein, the voltage components include: ohmic voltage drop, polarization voltage, and open-circuit voltage components; Based on the voltage components, a signal mode decomposition algorithm is introduced to extract and reconstruct the steady-state pseudo-open-circuit voltage characteristics that approximately characterize the electrochemical equilibrium state from high-frequency dynamic variables.

[0009] Optionally, the polarization kinetics process inside the battery includes: as charge carriers reciprocate between the positive and negative electrodes, electrochemical polarization and diffusion polarization occur simultaneously inside the battery; The electrochemical polarization, originating from the hysteresis of the charge transfer process, manifests as a short-time high-frequency response and is modeled in the equivalent circuit as a result of the ohmic polarization internal resistance. Rohm Polarization resistance Rtc With polarization capacitor Ctc The resulting parallel structure; The diffusion polarization, caused by the ion concentration gradient, exhibits a long-term, low-frequency dynamic, and is correspondingly equivalent to the diffusion polarization internal resistance. Rd With diffusion polarization capacitance Cd Parallel networks; In the equivalent circuit model, the open-circuit voltage characterizing the battery's equilibrium state is... UOC It is modeled as a voltage source that varies slowly with SOC.

[0010] Optionally, analyzing the voltage components through the equivalent circuit model includes: ; in, The measured voltage at the k-th sampling time is... Let be the open-circuit voltage at the k-th sampling time. Let be the high-frequency ohmic voltage drop at the k-th sampling time. Let be the high-frequency load polarization voltage drop at the k-th sampling time. Let the low-frequency diffusion voltage drop be the value at the k-th sampling time. k This represents the k-th sampling time.

[0011] Optionally, the ohmic voltage drop is: ; in, Let be the operating current at the k-th sampling time, and k+1 be the (k+1)-th sampling time. This represents the internal resistance of ohmic polarization; The polarization voltage includes: high-frequency charge transfer polarization voltage drop and low-frequency diffusion voltage drop; the expression is as follows: ; ; in, The high-frequency load polarization voltage drop at the (k+1)th sampling time. Let be the high-frequency load polarization voltage drop at the k-th sampling time. Let be the polarization resistance at the k-th sampling time. Let be the polarization capacitance at the k-th sampling time. The time interval between two adjacent sampling times. The low-frequency diffusion voltage drop at the (k+1)th sampling time. Let the low-frequency diffusion voltage drop be the value at the k-th sampling time. Let be the diffusion polarization internal resistance at the k-th sampling time. Let be the diffusion polarization capacitance at the k-th sampling time, and e be the natural exponent; The open-circuit voltage is: ; in, Let be the open-circuit voltage at the k-th sampling time. This represents the steady-state component during the open-circuit voltage variation process. To compensate for the voltage.

[0012] Optionally, the pseudo-open circuit voltage is: ; in, To measure voltage, The highest frequency component, It is a second high frequency component. For the lowest frequency variable, For a monotonic or approximately monotonic signal, This is a pseudo-open circuit voltage.

[0013] Optionally, the LSTM neural network model adopts a single-layer structure, and each LSTM cell includes three multiplication gate units for information control, namely, input gate ik, forget gate fk, and output gate ok.

[0014] Optionally, model training includes: By leveraging the strong correlation between pseudo-open circuit voltage and SOC, the LSTM neural network model is guided to learn the SOC mapping relationship, thereby achieving high-precision regression prediction of SOC.

[0015] The beneficial effects of this invention are as follows: This invention proposes a power lithium battery SOC estimation method based on feature extraction and mechanism-guided estimation, using a pseudo-open-circuit voltage-guided long short-term memory network. This method innovatively extracts the pseudo-open-circuit voltage (POCV), which is strongly correlated with SOC, by deeply analyzing the electrochemical and physical processes within the battery across multiple time scales. Based on this, the POCV is reconstructed from measured voltage and current signals into a feature matrix with clear physical meaning. The strong correlation between POCV and SOC guides an LSTM neural network to establish a more accurate nonlinear mapping relationship, thereby achieving high-precision and robust SOC estimation under complex dynamic conditions. This method requires no additional sensors or complex battery models; feature enhancement is achieved solely through mechanism-guided signal decomposition and feature reconstruction. Experimental results demonstrate that this method effectively improves SOC estimation accuracy, showcasing its feasibility and advantages in practical applications. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the complete framework and process of the POCVG-LSTM estimation algorithm in an embodiment of the present invention; wherein, (a) is the coupling analysis of the external characteristics of the battery through equivalent circuit modeling, (b) is the comparative analysis of the measured voltage and the open circuit voltage, (c) is the acquisition of pseudo open circuit voltage features, (d) is the precise mathematical calculation of the coupled polarization components, (e) is the extracted POCV components, and (f) is the input features of the LSTM neural network. Figure 2 This is a schematic diagram of the SOC prediction results of POCVG-LSTM under three different BBDST conditions according to an embodiment of the present invention; Figure 2 (a) shows the estimated SOC at 15℃; Figure 2 (b) represents the estimation error of SOC at 15°C; Figure 2 (c) shows the estimated SOC at 25℃; Figure 2 (d) represents the estimation error of SOC at 25℃; Figure 2 (e) shows the estimated SOC at 35°C; Figure 2(f) represents the estimation error of SOC at 35℃; Figure 3 This is a schematic diagram illustrating the calculation of the evaluation parameters of maximum error (ME), root mean squared error (RMSE), mean absolute error (MAE), and mean square error (MSE) under the BBDST condition according to an embodiment of the present invention. Figure 3 (a) Calculation of ME under BBDST conditions; Figure 3 (b) Calculated RMSE under BBDST conditions; Figure 3 (c) is the calculated MAE under the BBDST condition; Figure 3 (d) is the calculated MSE under the BBDST condition. Detailed Implementation

[0018] 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.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The pseudo-OCV-guided estimation concept in this embodiment: Carrier transport and diffusion within lithium batteries are immutable electrochemical and physical phenomena, inducing electrochemical polarization and diffusion polarization. Internal polarization effects and complex external conditions during battery charging and discharging lead to dynamic voltage drops and noise interference. This results in a strong nonlinearity and time-varying nature in the relationship between the measured terminal voltage and the state of charge (SOC), severely limiting the accuracy of SOC estimation based on neural networks. To overcome the challenges of SOC estimation caused by dynamic voltage drops and noise, this embodiment proposes constructing a "pseudo-open-circuit voltage" as a solution. Traditional methods directly use the dynamic terminal voltage, while this method innovatively employs signal decomposition technology based on the frequency domain characteristics of the polarization voltage to extract a stable-state component with low interference and high linear correlation with SOC from the dynamic voltage as a pseudo-open-circuit voltage. This aims to fundamentally alleviate the nonlinearity and time-varying nature of the mapping relationship. The pseudo-open-circuit voltage provides the neural network with a more stable and physically interpretable input feature. This not only avoids the static requirement for obtaining open-circuit voltage from the outset, but also significantly improves the overall accuracy of the SOC estimation model and its adaptability to complex real-world operating conditions. The pseudo-open-circuit voltage-guided estimation strategy is as follows: Figure 1 As shown.

[0021] This feature reconstruction strategy, by introducing pseudo-open-circuit voltages with clear physical meaning, provides neural networks with a more direct and purer SOC state representation, effectively guiding the model to distinguish intrinsic information and polarization noise in dynamic voltages, thereby significantly enhancing the model's learning efficiency and generalization performance across operating conditions.

[0022] This embodiment proposes a method for estimating the state of charge (SOC) of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network, including: The pseudo-open circuit voltage is extracted online by using a signal decomposition and steady-state component reconstruction strategy. The extracted pseudo-open-circuit voltage is used as a key mechanism feature, which together with the real-time current and voltage constitutes the input feature matrix. The input feature matrix is ​​fed into the LSTM neural network model for model training; Based on the trained model, the SOC of the power lithium battery is estimated.

[0023] This embodiment addresses the issue of decreased SOC estimation accuracy caused by multi-timescale coupling effects in voltage measurements under dynamic conditions. It proposes a steady-state component deconstruction approach, effectively separating the pseudo-open-circuit voltage (POCV) in the coupled signal to enhance feature representation capabilities. Based on this, a POCVG-LSTM algorithm is proposed. This algorithm reconstructs the extracted POCV components and the actual measured voltage and current signals into an input feature matrix with clear physical guidance. It then utilizes the strong correlation between POCV and SOC to guide the training and estimation of the LSTM neural network, thereby achieving high-precision and robust SOC estimation under complex dynamic conditions. The contributions of this embodiment are as follows: (1) The voltage signal characteristics of battery polarization response under dynamic operating conditions were revealed. The polarization response mechanism of power lithium battery under dynamic current excitation was systematically analyzed, the multi-timescale physicochemical causes of nonlinear changes in terminal voltage were clarified, and its signal composition was analyzed, providing a theoretical basis for subsequent feature deconstruction.

[0024] (2) A method for online extraction of pseudo-open-circuit voltage based on signal decomposition and reconstruction is proposed. Addressing the multi-scale coupling characteristics of terminal voltage, an innovative strategy is proposed to extract pseudo-open-circuit voltage online through signal decomposition and steady-state component reconstruction. This method can effectively separate the steady-state pseudo-open-circuit voltage, which characterizes the intrinsic state of the battery and has strong anti-interference capabilities, and construct a mechanism-guided feature with clear physical meaning.

[0025] (3) A POCVG-LSTM neural network estimation algorithm was proposed. The extracted pseudo-open-circuit voltage was used as a key mechanistic feature, which, together with the real-time current and voltage, constituted the input feature matrix. Utilizing the strong correlation between the pseudo-open-circuit voltage and the state of charge (SOC), the LSTM neural network was guided to learn the complex nonlinear mapping relationship between SOC and external variables during the dynamic process. This method significantly improved the model estimation accuracy and generalization performance.

[0026] Furthermore, through signal decomposition and steady-state component reconstruction strategies, the pseudo-open-circuit voltage is extracted online, including: Construct an equivalent circuit model; whereby the equivalent circuit model is used to mathematically describe the polarization dynamics process inside the battery. The voltage components are analyzed using an equivalent circuit model; these components include: ohmic voltage drop, polarization voltage, and open-circuit voltage. Based on voltage components, a signal mode decomposition algorithm is introduced to extract and reconstruct the steady-state pseudo-open-circuit voltage characteristics that approximate the electrochemical equilibrium state from high-frequency dynamic variables.

[0027] Furthermore, the polarization kinetics process inside the battery includes: as charge carriers reciprocate between the positive and negative electrodes, electrochemical polarization and diffusion polarization occur simultaneously inside the battery. Electrochemical polarization, originating from the hysteresis of the charge transfer process, manifests as a short-time high-frequency response and is modeled in the equivalent circuit as a result of the ohmic polarization internal resistance. Rohm Polarization resistance Rtc With polarization capacitor Ctc The resulting parallel structure; Diffusion polarization, caused by the ion concentration gradient, exhibits long-term, low-frequency dynamics and is correspondingly equivalent to diffusion polarization internal resistance. Rd With diffusion polarization capacitance Cd Parallel networks; In the equivalent circuit model, the open-circuit voltage characterizes the battery's equilibrium state. UOC It is modeled as a voltage source that varies slowly with SOC.

[0028] Specifically, in this embodiment, the pseudo-OCV enhanced feature extraction algorithm is as follows: To address the challenge of decreased SOC estimation accuracy caused by multi-timescale polarization coupling in the terminal voltage signal of power lithium batteries, this embodiment proposes a physics-inspired signal deconstruction and enhancement feature reconstruction method. Its core lies in analyzing the voltage components through an equivalent circuit model and reconstructing the pseudo-open-circuit voltage feature representing the quasi-equilibrium state from the dynamic response terminal voltage using signal decomposition techniques. This method first constructs an equivalent circuit model to mathematically describe the polarization dynamics process inside the battery. Then, it analyzes the composition of the terminal voltage, separating its ohmic voltage drop, polarization voltage, and open-circuit voltage components. Based on this, a signal mode decomposition algorithm is introduced to extract and reconstruct the steady-state pseudo-open-circuit voltage feature that approximately represents the electrochemical equilibrium state from high-frequency dynamic variables. This pseudo-open-circuit voltage feature is insensitive to transient changes in operating conditions, has a strong correlation with SOC, and possesses a clear physical meaning. Figure 1 (d) and Figure 1 As shown in (e), the enhanced pseudo-open-circuit voltage feature can be extracted from the original measurement signal after the above processing. Using this feature, along with the voltage and current measurements, as input to the neural network can reduce the difficulty for the model to directly learn the SOC mapping from the mixed signals and guide it to focus on more generalized state representations, thereby systematically improving the performance of the estimation model in complex dynamic applications.

[0029] Figure 1 (a) shows the coupled analysis of the battery's external characteristics through equivalent circuit modeling. With the reciprocating movement of charge carriers between the positive and negative electrodes, electrochemical polarization and diffusion polarization occur simultaneously inside the battery. Electrochemical polarization originates from the hysteresis of the charge transfer process, manifesting as a short-time high-frequency response, and is modeled in the equivalent circuit as a result of the ohmic polarization internal resistance. Rohm Polarization resistance Rtc With polarization capacitor Ctc The resulting parallel structure. Diffusion polarization is caused by the ion concentration gradient, exhibiting long-term low-frequency dynamics, and is correspondingly equivalent to the diffusion polarization internal resistance. Rd With diffusion polarization capacitance Cd A parallel network. Furthermore, the open-circuit voltage characterizes the battery's equilibrium state. UOC It is modeled as a voltage source that varies slowly with SOC. This equivalent circuit model physically decouples the multi-timescale characteristics of the battery response, laying a clear modeling foundation for subsequent voltage signal deconstruction and pseudo-open-circuit voltage feature extraction. Figure 1 (d) represents the precise mathematical calculation of the coupled polarization component. The measured voltage includes both the electrochemical polarization voltage and the diffusion polarization voltage, as shown in equation (1).

[0030] (1); In equation (1), where, The measured voltage at the k-th sampling time is... Let be the open-circuit voltage at the k-th sampling time. Let be the high-frequency ohmic voltage drop at the k-th sampling time. Let be the high-frequency load polarization voltage drop at the k-th sampling time. Let be the low-frequency diffusion voltage drop at the k-th sampling time, where k is the k-th sampling time. By quantifying the contribution of each polarization component, the multi-timescale dynamic response mixed in the terminal voltage signal can be deconstructed from a physical perspective. This provides a key theoretical basis for subsequently separating and reconstructing the pseudo-open-circuit voltage representing the quasi-equilibrium state from the dynamic voltage. The specific mathematical models of each polarization voltage drop are shown in equations (2), (3), and (4).

[0031] (2); (3); (4); In equation (2), The operating current is determined by the charging and discharging behavior. In equations (3) and (4), The time interval between two adjacent sampling times is determined by the sensor. e is the natural exponent. For ease of analysis, the parameters of the equivalent model are considered constant values ​​over a short period of time. It can be seen from formula (2) that the high-frequency characteristics of the ohmic voltage drop are determined by the operating current, and the complex charging and discharging behavior makes... Uohm Dynamic high-frequency changes. Similarly, UTC and Ud It includes both the current excitation response and the full voltage response of the parallel RC structure. However, since there is a polynomial single correspondence between SOC and OCV, as shown in equation (5).

[0032] (5); (6); In equation (5), Let SOC be a polynomial function. From equation (5), it can be seen that UOC will not undergo abrupt changes. The formula for calculating SOC is shown in equation (6). In equation (6), This is the Coulomb efficiency factor. Q This represents the current capacity of the battery. U OC It can be regarded as a coupling of a steady-state variable and a high-frequency component, as shown in equation (7).

[0033] (7); In equation (7), for Steady-state components during the change process The compensation voltage is caused by the instantaneous fluctuation of the SOC within a short time scale, and is manifested as the high-frequency dynamic disturbance component in the OCV. The voltage Uk is decomposed into voltage components from high frequency to low frequency using the empirical mode decomposition algorithm. The decomposition process of the voltage components is shown in equations (8) to (12).

[0034] (8); (9); (10); (11); (12); In equation (9), For decomposition function The minimum envelope function. For decomposition function The maximum value envelope function. Where the subscript n indicates the nth high-frequency component being decoupled. Let be the mean function of the envelope obtained in the i-th solution process for solving a certain component. Let be the component function obtained in the i-th solution process for solving a certain component. When When the mean is not 0, Treat it as a new decomposition function and execute formulas (9) and (10). When When the mean is 0, The component is considered as a solution with frequency variation, as shown in equation (11). When solving for the next component, the new decomposition function is obtained by equation (12). The above component solution process is repeated until the new decomposition function is a monotonic function. The decomposed voltage is shown in equation (13).

[0035] (13); In the decomposition result of formula (13), the components are arranged in order from high frequency to low frequency. This is the highest frequency component. The frequency is slightly lower than . The lowest frequency variable. Residual component. This is a monotonic or nearly monotonic signal, representing the ultra-low frequency trend term in the voltage signal. This component is the POCV to be extracted.

[0036] To improve SOC estimation accuracy without adding extra hardware, this embodiment proposes a feature reconstruction-based estimation strategy. The core of this strategy lies in extracting feature components highly correlated with SOC from existing measurement information. First, the nonlinear coupling relationship in the battery's external characteristics is analyzed and decoupled using an equivalent modeling method. Second, a mode decomposition algorithm is employed to separate and extract the Point of Contention (POCV), which characterizes the battery's internal steady-state trend, from the voltage signal. Finally, the POCV is reconstructed with existing measurement information to form an enhanced input feature. This input feature leverages the strong correlation between POCV and SOC to guide the neural network to learn the SOC mapping relationship more quickly and accurately, thereby achieving high-precision regression prediction of SOC. Based on the above principles, this embodiment proposes a POCVG-LSTM estimation method, the algorithm structure of which is as follows: Figure 1 As shown.

[0037] Figure 1 The complete framework and process of the proposed POCVG-LSTM estimation algorithm are systematically presented. Specifically, Figure 1 (a) Coupled analysis of battery external characteristics is performed through equivalent circuit modeling, providing... Figure 1 The precise mathematical calculation of the coupled polarization components in sub-(d) provides the model basis. Figure 1 The mathematical analysis of (d) further provides Figure 1 The determination of the POCV component in (e) lays the theoretical basis. Figure 1 The comparative analysis of measured voltage and open-circuit voltage in (b) inspired a key approach to extracting POCV steady-state characteristics from dynamic voltage signals. Ultimately, Figure 1 The POCV component extracted by (e) and Figure 1 The measured voltage obtained from the (c) pseudo-open-circuit voltage characteristic, along with the current, is reconstructed together as... Figure 1 The input features of the LSTM neural network in (f) are used to construct a complete SOC estimation model. The LSTM network structure used in this embodiment is shown in Table 1.

[0038] Table 1 LSTM Network Structure As shown in Table 1, this embodiment employs a single-layer structure in the LSTM neural network and uses fewer training epochs. This is mainly because the introduced POCV features have a strong correlation with SOC, providing a clear mapping relationship for the network, thus achieving an effective improvement in estimation accuracy without a complex network architecture. Each LSTM cell contains three multiplication gate units for information control, i.e., the input gates. ik Forgotten Gate fk Output gate OK Its definition is shown in equation (14).

[0039] (14); (15); In equation (14), For activation function, For the input gate weight parameters, For the input gate bias parameters, Forget gate weight parameters, Forget gate bias parameters, For the output gate weight parameters, These are the output gate bias parameters. The input features when k is... This is the output value at time k-1. The current time-stamped value of the LSTM cell. Compared with the current output value The calculation is shown in equation (15). In equation (15), tanh is the activation function. The input features at time k are... It is reconstructed from POCV and voltage and current, as shown in Equation (16); (16); From the network input feature structure shown in Equation (16), the measured voltage, operating current, and extracted POCV together constitute a multi-dimensional feature input. In this feature reconstruction, POCV, with its explicit linear mapping relationship with SOC, provides a reliable steady-state reference for the network, effectively guiding the model to converge toward the true value. Meanwhile, the measured voltage and current are mainly responsible for characterizing the dynamic behavior of the system under random operating conditions, capturing high-frequency response components to finely depict the transient fluctuation process of SOC. This "steady-state guidance-dynamic compensation" feature fusion mechanism helps to improve the overall accuracy and adaptability of SOC estimation.

[0040] To evaluate the performance of the proposed POCVG-LSTM model in SOC prediction, this embodiment uses four commonly used evaluation metrics: maximum error (ME), root mean squared error (RMSE), mean absolute error (MAE), and mean square error (MSE). The calculation formulas for each metric are shown in Equation (17).

[0041] (17); In equation (17), This represents the predicted SOC value obtained at time k using the estimation method. This represents the reference true value of SOC at the same time, and n represents the total prediction step size (total number of samples).

[0042] SOC estimation results under the Beijing bus dynamic stress test (BBDST) conditions: To verify the accuracy of SOC estimation, this embodiment uses the SOC value calculated by the ampere-hour integration method as the baseline true value. For the BBDST test, the SOC estimation model is trained using data from battery 1 under the BBDST condition. Then, this model is directly applied to battery 2 to evaluate its cross-battery generalization performance. This design is more in line with real-world application scenarios because in real battery management system deployments, it is usually not possible to use historical data from the same battery for both model training and online prediction simultaneously. The SOC results predicted by POCVG-LSTM under three different BBDST conditions are shown below. Figure 2 As shown. Among them, Figure 2 (a) shows the estimated SOC at 15℃; Figure 2 (b) represents the estimation error of SOC at 15℃; Figure 2 (c) represents the estimated SOC at 25℃; Figure 2 (d) represents the estimation error of SOC at 25℃; Figure 2 (e) represents the estimated SOC at 35°C; Figure 2 (f) represents the estimation error of SOC at 35℃.

[0043] exist Figure 2In the diagram, S1 represents the baseline SOC value obtained based on the ampere-hour integration method. S2 is the estimated value from the traditional LSTM model. S3 is the estimated value from the POCVG-LSTM model proposed in this embodiment. It can be seen that both algorithms can effectively track the trend of the baseline value during the dynamic change of SOC. Further observation from the error curves reveals that the estimation error fluctuation of the POCVG-LSTM algorithm is significantly smaller than that of the LSTM algorithm. This is because there is a strong correlation between POCVG and SOC, which can provide more effective state guidance for the POCVG-LSTM model, thereby helping the model converge to the vicinity of the true SOC faster and maintain more stable estimation performance. The evaluation parameters ME, RMSE, MAE, and MSE are calculated under the BBDST condition as follows: Figure 3 As shown. Among them, Figure 3 (a) is the calculated ME under the BBDST condition; Figure 3 (b) represents the calculated RMSE under the BBDST condition; Figure 3 (c) represents the calculated MAE under the BBDST condition; Figure 3 (d) represents the calculated MSE under the BBDST condition.

[0044] from Figure 3 The comparison results clearly show that the POCVG-LSTM algorithm proposed in this embodiment demonstrates significant advantages over the traditional LSTM algorithm in several key performance indicators, validating the effectiveness of the model improvement. Specifically, the POCVG-LSTM algorithm excels in comprehensive control of estimation error and improvement of prediction accuracy, reducing ME by more than 30.96%. More importantly, the POCVG-LSTM algorithm also performs exceptionally well in measuring the average deviation between predicted and true values, reducing RMSE by more than 8.58% and MAE by more than 2.96%. This reflects the model's stronger ability to suppress large-amplitude errors and less overall volatility in prediction results. From the perspective of squared loss, MSE is reduced by more than 16.43%. The systematic improvement in all indicators indicates that the POCV guidance mechanism introduced in this embodiment can effectively provide state reference and direction guidance in the SOC estimation process under BBDST conditions, thereby significantly improving the model's positioning accuracy and estimation stability.

[0045] This embodiment addresses the issue of reduced SOC estimation accuracy in neural networks caused by multi-timescale polarization coupling effects within power lithium batteries. It innovatively proposes a POCVG-LSTM guided estimation algorithm based on the decomposition of the acquired voltage signal. Through in-depth analysis of the coupling polarization phenomenon in the carrier transport process, theoretical support is provided for extracting POCV based on empirical mode decomposition. The study reveals a significant linear correlation between the decoupled steady-state POCV and SOC, providing an effective guiding signal for the neural network model to quickly locate the true SOC value. Based on this theoretical understanding, this embodiment constructs a POCVG-LSTM algorithm, which demonstrates superior estimation performance under various temperature conditions (15℃, 25℃, 35℃) and different operating conditions. Experimental results show that compared with the traditional LSTM model, the proposed algorithm achieves a systematic improvement in multiple evaluation indicators. In summary, the POCVG-LSTM algorithm exhibits significant advantages in estimation accuracy, robustness, and temperature adaptability.

[0046] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for estimating the state of charge (SOC) of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network, characterized in that, include: The pseudo-open circuit voltage is extracted online by using a signal decomposition and steady-state component reconstruction strategy. The extracted pseudo-open-circuit voltage is used as a key mechanism feature, which together with the real-time current and voltage constitutes the input feature matrix. The input feature matrix is ​​input into the LSTM neural network model for model training; Based on the trained model, the SOC of the power lithium battery is estimated.

2. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 1, characterized in that, The pseudo-open circuit voltage is extracted online using a signal decomposition and steady-state component reconstruction strategy, including: Construct an equivalent circuit model; wherein, the equivalent circuit model is used to mathematically describe the polarization dynamics process inside the battery; The voltage components are analyzed using the equivalent circuit model; wherein, the voltage components include: ohmic voltage drop, polarization voltage, and open-circuit voltage components; Based on the voltage components, a signal mode decomposition algorithm is introduced to extract and reconstruct the steady-state pseudo-open-circuit voltage characteristics that approximately characterize the electrochemical equilibrium state from high-frequency dynamic variables.

3. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 2, characterized in that, The polarization dynamics process inside the battery includes: as charge carriers reciprocate between the positive and negative electrodes, electrochemical polarization and diffusion polarization occur simultaneously inside the battery. The electrochemical polarization originates from the hysteresis of the charge transfer process and manifests as a short-time high-frequency response. In the equivalent circuit, it is modeled as a parallel structure consisting of the ohmic polarization internal resistance Rohm, the polarization resistance Rtc, and the polarization capacitance Ctc. The diffusion polarization, caused by the ion concentration gradient, exhibits a long-term low-frequency dynamic and is correspondingly equivalent to a parallel network of diffusion polarization internal resistance Rd and diffusion polarization capacitance Cd. In the equivalent circuit model, the open-circuit voltage UOC, which characterizes the battery's equilibrium state, is modeled as a voltage source that varies slowly with SOC.

4. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 2, characterized in that, Analyzing the voltage components using the equivalent circuit model includes: ； in, U To measure voltage, UOC Open circuit voltage, Uohm For high-frequency ohmic voltage drop, UTC For high-frequency load transfer polarization voltage drop, Ud For low-frequency diffusion voltage drop, k This represents the k-th sampling time.

5. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 4, characterized in that, The ohmic voltage drop is: ； in, The current is the operating current, and k+1 is the (k+1)th sampling time. This represents the internal resistance of ohmic polarization; The polarization voltage includes: high-frequency charge transfer polarization voltage drop and low-frequency diffusion voltage drop; the expression is as follows: ； ； in, The high-frequency load polarization voltage drop at the (k+1)th sampling time. Let be the high-frequency load polarization voltage drop at the k-th sampling time. Let be the polarization resistance at the k-th sampling time. Let be the polarization capacitance at the k-th sampling time. The time interval between two adjacent sampling times. The low-frequency diffusion voltage drop at the (k+1)th sampling time. Let the low-frequency diffusion voltage drop be the value at the k-th sampling time. Let be the diffusion polarization internal resistance at the k-th sampling time. Let be the diffusion polarization capacitance at the k-th sampling time, and e be the natural exponent; The open-circuit voltage is: ； in, Let be the open-circuit voltage at the k-th sampling time. This represents the steady-state component during the open-circuit voltage variation process. To compensate for the voltage.

6. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 1, characterized in that, The pseudo-open circuit voltage is: ； in, To measure voltage, The highest frequency component, It is a sub-high frequency component. For the lowest frequency variable, For a monotonic or approximately monotonic signal, This is a pseudo-open circuit voltage.

7. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 1, characterized in that, The LSTM neural network model adopts a single-layer structure. Each LSTM cell includes three multiplication gate units for information control, namely the input gate ik, the forget gate fk, and the output gate ok.

8. The method for estimating the SOC of a power lithium battery using a pseudo-open-circuit voltage-guided long short-term memory network according to claim 1, characterized in that, Model training includes: By leveraging the strong correlation between pseudo-open circuit voltage and SOC, the LSTM neural network model is guided to learn the SOC mapping relationship, thereby achieving high-precision regression prediction of SOC.

Images