Intelligent management method of lithium battery energy storage system

Abstract

The invention discloses an intelligent management method of a lithium battery energy storage system. The intelligent management method comprises the following steps: acquiring voltage, current, temperature and acoustic emission signals of each single battery in real time; based on the electricity and temperature data, dynamically correcting target intrinsic parameters of the electrochemical model by using an online parameter identification algorithm, and generating an incremental electrochemical model; inputting real-time voltage and current data into the model, solving internal states such as negative electrode surface concentration and overpotential, and predicting the residual life in multiple scenes based on a parameter degradation trend; solving an optimal control instruction set by adopting a reinforcement learning algorithm by taking maximization of system life, energy efficiency and power grid service quality as targets and taking internal state and life prediction as constraints; after the instruction is executed, self-adaptive calibration of model parameters is triggered by comparing the error between the actual voltage and the predicted voltage and combining acoustic emission signal monitoring. According to the invention, accurate sensing of the internal state of the battery, predictive management of the service life and dynamic optimization of the control strategy are realized.

Description

Technical Field

[0001] This invention relates to the field of lithium battery energy storage technology, and in particular to an intelligent management method for lithium battery energy storage systems. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy sources, electrochemical energy storage, especially lithium-ion battery energy storage systems, plays an increasingly important role in grid peak and frequency regulation and renewable energy grid integration due to their advantages such as flexible configuration and rapid response. However, the inherent characteristics of lithium-ion batteries, such as performance degradation and the risk of thermal runaway, pose serious challenges to the long-term operational safety, economy, and reliability of large-scale energy storage systems. Therefore, achieving accurate state sensing, health prediction, and optimized control of energy storage batteries has become a key technology for ensuring their efficient and stable operation throughout their entire life cycle.

[0003] Currently, lithium battery management technologies are mainly divided into two categories: methods based on equivalent circuit models and data-driven methods. The former simulates the external characteristics of the battery using circuit components such as resistors and capacitors, resulting in low computational complexity and ease of real-time control. However, the models lack physical meaning and struggle to reveal the internal aging mechanisms of the battery. The latter utilizes machine learning algorithms such as neural networks to directly learn the battery's input-output relationships from historical data, demonstrating good fitting capabilities in certain scenarios. Furthermore, existing advanced solutions are beginning to combine simple electrochemical models with state estimation algorithms to estimate the battery's state of charge (SOC) and state of health (SOH).

[0004] In the process of developing this invention, the inventors discovered that although existing technologies have made some progress, they still have significant limitations. First, neither equivalent circuit models nor shallow data-driven models can accurately and in real-time acquire the microscopic state parameters reflecting the internal electrochemical processes of the battery. This results in state assessment remaining at a "black box" or "gray box" level, failing to provide a deep basis for refined management. Second, existing health state predictions are mostly based on statistical extrapolation of historical data, failing to closely integrate with the physicochemical degradation mechanisms within the battery. This leads to insufficient long-term prediction accuracy and an inability to adapt to changing future operating scenarios. Finally, the system's optimization control objectives are often singular (e.g., only pursuing efficiency or delaying degradation), and the control strategies are mostly static or open-loop. There is a lack of a closed-loop optimization framework that can comprehensively consider the battery's internal state, long-term lifespan, and external grid requirements, and can adjust in real-time based on feedback, making it difficult to maximize the overall system utility. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide an intelligent management method for a lithium battery energy storage system to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, in a first aspect, the present invention provides an intelligent management method for a lithium battery energy storage system, the method comprising the following steps:

[0007] S10: Real-time acquisition of voltage, current, temperature, and acoustic emission signals of each battery cell in the lithium battery energy storage system during operation; based on the voltage, current, and temperature data, using an online parameter identification algorithm, dynamically correcting the target intrinsic parameters in a preset basic electrochemical model to generate an incremental electrochemical model that matches the current actual electrochemical state of each battery cell; the target intrinsic parameters include the solid-phase diffusion coefficients of the positive and negative electrodes and the reaction rate constant;

[0008] S20: The real-time collected voltage and current data are used as inputs and substituted into the incremental electrochemical model corresponding to each battery cell to solve the internal state parameters of each battery cell in real time. The internal state parameters include the negative electrode solid phase surface concentration and the positive and negative electrode overpotentials. Based on the current trend of the change of the target intrinsic parameters, the remaining service life of each battery cell under various preset future operating scenarios is predicted by extrapolation algorithm.

[0009] S30: With the optimization objective of maximizing the overall remaining service life, current cycle efficiency, and quality benefit index of grid frequency regulation service of the lithium battery energy storage system, and with the internal state parameters and predicted remaining service life of each battery cell as constraints, the optimal system-level control instruction set is solved in real time using a reinforcement learning algorithm. The system-level control instruction set includes differentiated charge and discharge current instructions for inconsistent battery cells, and the maximum allowable charge and discharge power at the system level.

[0010] S40: The optimal system-level control instruction set is sent to the battery management system for execution; and after execution, the error between the actual voltage response of the battery cell and the predicted voltage of the incremental electrochemical model is obtained. At the same time, the characteristic changes of the acoustic emission signal are monitored. If the error exceeds a preset threshold and / or the acoustic emission signal shows abnormal characteristics, the online parameter identification algorithm is triggered to update and calibrate the target intrinsic parameters of the incremental electrochemical model.

[0011] The above technical solution has the following beneficial effects:

[0012] This technical solution achieves real-time and accurate perception of the internal microstate of the battery (such as negative electrode surface concentration and overpotential) by constructing an incremental electrochemical model. Combined with multi-timescale lifetime prediction and multi-objective optimization, it forms a closed-loop intelligent management system. It elevates battery management from the traditional mode based on external characteristics to a precise control level based on internal electrochemical state. By integrating multi-source information such as acoustic emission, it achieves safety early warning and model self-calibration. Ultimately, while ensuring safety, it significantly improves the overall lifetime, operating efficiency, and quality of grid service of the energy storage system, achieving synergistic optimization of safety, economy, and reliability. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of an intelligent management method for a lithium battery energy storage system according to an embodiment of the present invention;

[0014] Figure 2 This is a detailed flowchart of step S10 in an embodiment of the present invention;

[0015] Figure 3 This is a detailed flowchart of step S20 in an embodiment of the present invention;

[0016] Figure 4 This is a detailed flowchart of step S30 in an embodiment of the present invention;

[0017] Figure 5 This is a detailed flowchart of step S40 in an embodiment of the present invention;

[0018] Figure 6 This is a functional block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Example 1

[0021] Figure 1 This is an overall flowchart of an intelligent management method for a lithium battery energy storage system according to an embodiment of the present invention. Figure 1 As shown, the method includes the following steps:

[0022] S10: Real-time acquisition of voltage, current, temperature, and acoustic emission signals of each battery cell in the lithium battery energy storage system during operation; based on the voltage, current, and temperature data, using an online parameter identification algorithm, dynamically correcting the target intrinsic parameters in a preset basic electrochemical model to generate an incremental electrochemical model that matches the current actual electrochemical state of each battery cell; the target intrinsic parameters include the solid-phase diffusion coefficients of the positive and negative electrodes and the reaction rate constant;

[0023] Specifically, the voltage data can be acquired through an analog-to-digital converter module with a sampling accuracy of no less than ±0.001V and a sampling frequency of no less than 10ms / time; the current data can be acquired through a Hall current sensor with an accuracy of no less than ±0.1A and a sampling frequency of no less than 10ms / time; the temperature data can be acquired through a thermistor with an accuracy of no less than ±0.5℃, with at least two sensors set for each battery cell, and a sampling frequency of no less than 50ms / time; the acoustic emission signal can be acquired through a piezoelectric acoustic emission sensor with a sampling rate of no less than 1MHz, a frequency response range of 100kHz-1MHz, and a sampling frequency of no less than 100ms / time.

[0024] S20: The real-time collected voltage and current data are used as inputs and substituted into the incremental electrochemical model corresponding to each battery cell to solve the internal state parameters of each battery cell in real time. The internal state parameters include the negative electrode solid phase surface concentration and the positive and negative electrode overpotentials. Based on the current trend of the change of the target intrinsic parameters, the remaining service life of each battery cell under various preset future operating scenarios is predicted by extrapolation algorithm.

[0025] S30: With the optimization objective of maximizing the overall remaining service life, current cycle efficiency, and quality benefit index of grid frequency regulation service of the lithium battery energy storage system, and with the internal state parameters and predicted remaining service life of each battery cell as constraints, the optimal system-level control instruction set is solved in real time using a reinforcement learning algorithm. The system-level control instruction set includes differentiated charge and discharge current instructions for inconsistent battery cells, and the maximum allowable charge and discharge power at the system level.

[0026] S40: The optimal system-level control instruction set is sent to the battery management system for execution; and after execution, the error between the actual voltage response of the battery cell and the predicted voltage of the incremental electrochemical model is obtained. At the same time, the characteristic changes of the acoustic emission signal are monitored. If the error exceeds a preset threshold and / or the acoustic emission signal shows abnormal characteristics, the online parameter identification algorithm is triggered to update and calibrate the target intrinsic parameters of the incremental electrochemical model.

[0027] like Figure 2As shown, in some embodiments, step S10 may include:

[0028] S101: Load the preset basic electrochemical model for each battery cell and set the initial values ​​of the target intrinsic parameters;

[0029] In this embodiment, the basic electrochemical model adopts a pseudo-two-dimensional (P2D) electrochemical model based on the Doyle-Fuller-Newman DFN model. This model can describe the electrochemical processes inside a lithium-ion battery, including solid-phase diffusion, liquid-phase transport, and electrode reaction kinetics. When loading this basic model for each battery cell, the initial values ​​of the target intrinsic parameters are set according to the battery cell's factory specifications: for example, for lithium iron phosphate (LFP) batteries, the initial value of the positive electrode solid-phase diffusion coefficient is set to 1.0 × 10⁻⁶. -14 m 2 / s, the negative electrode solid-phase diffusion coefficient is set to 3.0×10 -14 m 2 / s, the initial value of the reaction rate constant is set to 1.0×10⁻¹¹ m. 2.5 / (mol 0.5 These initial values ​​can be stored in the non-volatile memory of the battery management system and retrieved upon system startup. The initial values ​​are set taking into account the battery's chemistry, design capacity, and typical operating temperature range to ensure reasonable accuracy of the model in the initial stage.

[0030] S102: Using a first time interval of seconds or milliseconds, based on the voltage and current data collected in real time at the second or millisecond level, the least squares method or gradient descent method is used as the online parameter identification algorithm to identify and correct the reaction rate constant in the target intrinsic parameter online, so as to obtain the corrected reaction rate constant.

[0031] Specifically, the first time interval is set to 100 milliseconds to accommodate rapid changes in the battery's dynamic response. Real-time voltage and current data are acquired via a high-precision sensor at a sampling frequency of 1 kHz and pre-processed with filtering to eliminate noise. During online parameter identification, recursive least squares (RLS) is used for real-time correction: the error between the predicted output voltage of the basic electrochemical model and the actual measured voltage is used as the objective function, and the reaction rate constant is iteratively updated by minimizing the sum of squared errors. Specifically, the forgetting factor of the RLS algorithm is set to 0.99 to balance the influence of old and new data. For example, in each iteration, the Jacobian matrix is ​​calculated based on the current voltage and current data, and the parameter estimates are updated. Alternatively, stochastic gradient descent can be used, with the learning rate set to 0.001 for stable convergence. Through this step, the reaction rate constant can be tracked in real-time by battery aging or changes in operating conditions; for example, when the battery temperature increases, the reaction rate constant is corrected to 1.2 × 10⁻⁶.-11 m 2.5 / (mol 0.5 ·s) to improve model accuracy.

[0032] S103: Using a second time interval on the order of minutes, based on the average value of the temperature data collected within the first time interval, temperature compensation is performed on the positive and negative electrode solid-phase diffusion coefficients using the Arrhenius formula. Based on the initial values ​​of the compensated positive and negative electrode solid-phase diffusion coefficients, combined with the voltage and current data, the least squares method or extended Kalman filter algorithm is used to correct the positive and negative electrode solid-phase diffusion coefficients in the target intrinsic parameters, thereby obtaining the corrected positive and negative electrode solid-phase diffusion coefficients.

[0033] Specifically, the second time interval is set to 5 minutes because the solid-phase diffusion coefficient changes slowly and does not require frequent updates. First, based on the temperature data collected within the first time interval (100 milliseconds) (10 sampling points per second), the average temperature over 5 minutes is calculated, for example, 25°C. Then, the Arrhenius equation is used to temperature compensate for the solid-phase diffusion coefficients of the positive and negative electrodes: For the positive electrode diffusion coefficient, whose activation energy is 20 kJ / mol, the temperature-compensated diffusion coefficient is calculated according to the following formula:

[0034]

[0035] Among them, D corrected D is the corrected diffusion coefficient. initial The initial diffusion coefficient is Ea = 20 kJ / mol; the activation energy is R; the ideal gas constant is T. avg The average temperature is Tref = 298K, which is the reference temperature (25℃).

[0036] For example, if the average temperature is 30℃, the initial value of the positive electrode diffusion coefficient is 1.0 × 10⁻⁶. -14 m 2 / s becomes 1.2×10 after compensation. -14 m 2 / s. Next, using the compensated value as the initial value, and combining it with voltage and current data collected over the past 5 minutes (e.g., 3000 data points), the Extended Kalman Filter (EKF) algorithm is used for correction: EKF linearizes the electrochemical model, using voltage as the observed variable and the diffusion coefficient as the state variable, and iteratively optimizes through a prediction-update step to minimize the observation error. Specifically, the process noise covariance is set to 1×10⁻⁶. -10 The observation noise covariance is set to 1×10. -5 This step ensures filter stability. It accurately captures slow drifts in the diffusion coefficient, such as diffusion performance degradation due to battery cycling.

[0037] S104: Substitute the modified reaction rate constant and the modified positive and negative electrode solid-phase diffusion coefficients into the basic electrochemical model to generate the incremental electrochemical model at the current moment.

[0038] Specifically, after completing the parameter corrections in S102 and S103, the updated parameter values ​​are substituted into the corresponding equations of the basic P2D electrochemical model, replacing the original parameters. For example, the corrected reaction rate constant is used to adjust the reaction kinetics term in the Butler-Volmer equation, while the corrected solid-phase diffusion coefficient is used to update the diffusion term in Fick's diffusion law. This generates an incremental electrochemical model that matches the actual state of the current battery cell. This model maintains the structure of the basic model, but the parameter values ​​have been dynamically adjusted, thus enabling a more accurate simulation of the battery's internal state. The generated model is immediately used to solve for the internal state parameters in subsequent steps (e.g., S20), achieving closed-loop optimization. The entire model update process runs in real-time on the embedded processor of the battery management system, with controllable computational overhead, ensuring timely system response.

[0039] In some embodiments, in step S103, the temperature compensation of the solid-phase diffusion coefficients of the positive and negative electrodes using the Arrhenius formula is specifically achieved through the following formula:

[0040]

[0041] Among them, D s (T) represents the initial values ​​of the solid-phase diffusion coefficients of the positive and negative electrodes after compensation with temperature data T, and D s 0 is the preset value of the solid-phase diffusion coefficient of the positive and negative electrodes at the reference temperature, E a_D Let R be the diffusion activation energy, T be the ideal gas constant, and T be the average value of the temperature data. The diffusion activation energy is the energy barrier that a particle must overcome to diffuse within a material; its magnitude reflects the sensitivity of the diffusion process to temperature changes. The diffusion activation energy is obtained by measuring the diffusion coefficient under different temperature conditions and fitting the experimental data using the Arrhenius relation.

[0042] Specifically, in step S102, when the reaction rate constant is identified and rapidly corrected online, the voltage data and current data used are steady-state operating data after removing transient data at the moment of charge-discharge switching.

[0043] Specifically, in step S10, the acoustic emission signal does not participate in the correction process of the target intrinsic parameters, but is used solely for subsequent model credibility assessment and safety warning.

[0044] like Figure 3 As shown, in some embodiments, step S20 may include:

[0045] S201: The voltage and current data collected in real time are used as inputs and substituted into the incremental electrochemical model corresponding to each battery cell. By solving the electrochemical differential algebraic equations described by the incremental electrochemical model, the negative electrode solid phase surface concentration and positive and negative electrode overpotentials of each battery cell are calculated in real time.

[0046] In this step, the real-time acquired battery cell terminal voltage U(t) and charging / discharging current I(t) (current is positive for discharging and negative for charging) are used as known inputs and substituted into the incremental electrochemical model generated in step S10 after parameter correction. This model is described by a set of electrochemical differential-algebraic equations. Solving this set of equations to obtain the internal state parameters is an initial-boundary value problem. The specific solution process is as follows: First, the solid-phase diffusion equation (Fick's second law) is discretized along the electrode particle radius (e.g., using the finite difference method), transforming the partial differential equations into a set of ordinary differential equations. Then, this set of ordinary differential equations is combined with the Butler-Folmer equation describing the electrode reaction kinetics and the system voltage balance equation to form a complete system of differential-algebraic equations. In actual solution, numerical methods are used, such as the Euler method, Runge-Kutta method combined with Newton's iteration method, to solve in real time on an embedded processor with a fixed time step (e.g., 100 milliseconds). By solving the problem, the lithium-ion concentration on the surface of the negative electrode active material particles of each battery cell at the current moment can be directly calculated. This value has been normalized for ease of use. At the same time, the positive electrode overpotential η_pos(t) and the negative electrode overpotential η_neg(t) are calculated. These internal state parameters are a direct reflection of the electrochemical processes inside the battery and cannot be directly measured by external sensors, but they are crucial for assessing the battery's health status and potential risks.

[0047] S202: Based on historical and current intrinsic parameter data of the target, establish a degradation trajectory model for the solid-phase diffusion coefficients and reaction rate constants of the positive and negative electrodes of each battery cell as a function of time.

[0048] This step quantifies the degradation patterns of key intrinsic parameters (i.e., target intrinsic parameters) of each battery cell over time and with use. The system continuously records and stores historical data sequences of the positive electrode solid-phase diffusion coefficient D_s_pos, negative electrode solid-phase diffusion coefficient D_s_neg, and reaction rate constant k_eff, periodically identified in step S10 during past operation (e.g., the past few months). Based on these time-series data, an empirical degradation trajectory model is established for each parameter. For example, for the solid-phase diffusion coefficient, which degrades relatively slowly, a linear degradation model is used: D_s(t) = D_s0 - α*t, where D_s0 is the initial value and α is the degradation rate, obtained by fitting historical data using the least squares method. For the reaction rate constant, which decays with increasing cycle number, an exponential decay model is used: k_eff(n) = k_eff0*exp(-β*n), where n is the equivalent cycle number and β is the decay coefficient. The system independently establishes and maintains a degradation trajectory model for each target intrinsic parameter of each battery cell. The model parameters (α, β) are updated periodically by refitting with the latest data, so that the degradation trajectory model can adapt to the actual aging path of the battery.

[0049] S203: For the preset multiple future operating scenarios, the value of the target intrinsic parameter at the current moment is used as the initial condition and input into the corresponding degradation trajectory model. The change of the target intrinsic parameter on the future time axis is deduced through the extrapolation algorithm until any target intrinsic parameter reaches its failure threshold. This moment is used as the predicted end of life, and the time from the current moment to the end of life is calculated as the remaining service life of the battery cell in the corresponding future operating scenario.

[0050] This step is used to predict the remaining useful life (RUL) of individual battery cells under different assumptions. Various preset future operating scenarios may include: maintaining the current average charge / discharge rate (e.g., 0.5C); frequent participation in grid frequency regulation (high rate, shallow charge / discharge cycles); and serving as a backup power source (floating charge most of the time, occasional discharge). For each future operating scenario to be evaluated, the prediction process is as follows: First, the target intrinsic parameter values ​​(D_s_pos(t0), D_s_neg(t0), k_eff(t0)) latest identified in step S10 at the current time (t0) are used as the initial conditions for the degradation trajectory model. Then, based on the stress level (average current, temperature) under the scenario assumptions, the corresponding degradation trajectory model is run to extrapolate the changes in these parameters forward (in the future time direction). The extrapolation is performed using discrete iterative calculations in the time domain. The failure threshold is a pre-set critical value for a parameter based on battery life termination criteria. For example, the battery life is considered terminated when the cathode solid-phase diffusion coefficient decays to 50% of its initial value, or the reaction rate constant decays to 30% of its initial value. During the iterative simulation, the predicted value of each target intrinsic parameter is monitored in real time. When the predicted value of any parameter first reaches or exceeds its corresponding failure threshold, the simulation stops, and the corresponding future time point t_end is recorded. The predicted remaining lifespan (RUL) of the battery cell under this specific future operating scenario is then t_end - t0. By predicting multiple scenarios, the expected lifespan of the battery under different operating modes can be obtained, providing key constraint information for subsequent optimization control.

[0051] In some embodiments, in step S201, the electrochemical differential algebraic equations described by the incremental electrochemical model include: a solid-phase diffusion equation, which describes the relationship between the diffusion process of lithium ions in the electrode spherical particles and the spatiotemporal distribution of solid-phase lithium ion concentration through the solid-phase diffusion coefficient; the Butler-Folmer equation, which characterizes the electrochemical reaction kinetics on the electrode surface by relating the reaction current density to the positive and negative electrode overpotentials through the reaction rate constant; and a voltage balance equation, which establishes the balance relationship between the battery terminal voltage and the open-circuit voltage, the positive and negative electrode overpotentials, and the ohmic internal resistance voltage drop, wherein the open-circuit voltage is a function of the normalized concentration of lithium ions on the negative electrode solid-phase surface.

[0052] The electrochemical differential-algebraic equations described by the incremental electrochemical model include the following formulas:

[0053] Solid-state diffusion equation, which describes the diffusion process of lithium ions within electrode spherical particles:

[0054]

[0055] Among them, c sLet D be the solid-phase lithium-ion concentration, which is a function of time t and the radial position r of the electrode spherical particles. s The solid-phase diffusion coefficient is one of the intrinsic parameters of the target; this is achieved by solving the solid-phase diffusion equation and substituting the boundary condition r = R. s R s Given the radius of the spherical electrode particle, the calculated solid-phase lithium ion concentration c on the particle surface is... s , where is the desired surface concentration of the negative electrode solid phase;

[0056] The Butler-Folmer equation describes the electrochemical reaction kinetics at the electrode surface:

[0057]

[0058] Where j is the reaction current density, i0 is the exchange current density, which is related to the reaction rate constant k in the target intrinsic parameters, η is the positive and negative electrode overpotentials, and α a and α c Where F is the transfer coefficient, R is the Faraday constant, T is the ideal gas constant, and T is the temperature of the cell.

[0059] The voltage balance equation establishes the relationship between the terminal voltage and the internal state of the battery cell:

[0060] V=U(θ)-η-IR Ω ;

[0061] Where V is the battery terminal voltage, U(θ) is the open-circuit voltage, which is a function of the normalized lithium-ion concentration θ on the negative electrode solid surface, where θ is the ratio of the lithium-ion concentration on the negative electrode solid surface to the maximum lithium-ion concentration, I is the battery current, and R... Ω It is the internal resistance of the Ohm.

[0062] The real-time solution is achieved by simultaneously solving the above system of equations using numerical methods.

[0063] Specifically, in step S201, the electrochemical differential algebraic equations described by the incremental electrochemical model are solved numerically using the Runge-Kutta method or the backward difference method; in step S202, the degradation trajectory model is implemented using linear regression, exponential fitting, the Arrhenius model, or the SEI film growth kinetics model; in step S203, the preset multiple future operating scenarios include at least: a scenario operating at rated power, a scenario operating at peak power, and a scenario operating under predetermined temperature fluctuation conditions; in step S203, for each target intrinsic parameter, its failure threshold is defined as the value corresponding to the parameter value decaying to 80% of its factory nominal value.

[0064] In some embodiments, in step S203, the extrapolation algorithm employs a multi-scene probability extrapolation algorithm based on adaptive unscented Kalman filtering, and the execution of the extrapolation algorithm includes the following steps:

[0065] S2031: Based on the historical intrinsic parameter data of the target, the uncertainty range of each model parameter in the degradation trajectory model is determined by the maximum likelihood estimation method, and a probability distribution model of the degradation trajectory model parameters is established.

[0066] S2032: For each future operating scenario, create an independent prediction thread, and perform the following operations in each prediction thread: randomly sample a set of parameter values ​​from the probability distribution model of the parameters of the degradation trajectory model; substitute the randomly sampled parameter values ​​into the degradation trajectory model; and perform deterministic extrapolation using the target intrinsic parameter value at the current moment as the initial condition.

[0067] S2033: In each of the deterministic extrapolation iterations, an unscented Kalman filter algorithm is used to correct the prediction results of the target intrinsic parameters obtained from the degenerate trajectory model, resulting in corrected predicted values ​​of the target intrinsic parameters; wherein, the unscented Kalman filter algorithm uses a process noise covariance matrix to characterize the uncertainty of the degenerate trajectory model, and the process noise covariance matrix is ​​obtained by calculating the variance of the historical data of the target intrinsic parameters within a sliding time window and adaptively adjusting it;

[0068] S2034: When any of the modified intrinsic target parameter prediction values ​​obtained from step S2033 exceeds three times the standard deviation of the probability distribution corresponding to its failure threshold for the first time, the life end is determined to be reached.

[0069] S2035: Repeat steps S2032 to S2034 until the preset minimum number of times are reached to obtain the Monte Carlo statistical distribution of the remaining service life, and take the mathematical expectation of the Monte Carlo statistical distribution as the final predicted remaining service life.

[0070] In an alternative embodiment, in step S203, the extrapolation algorithm is implemented through a hybrid prediction model, which consists of a temporal convolutional network module, an attention mechanism module, and a fully connected regressor in sequence; its lifetime prediction process includes the following steps:

[0071] S2031': Input the historical time series data of the target intrinsic parameter into the temporal convolutional network module. The temporal convolutional network module extracts and outputs multi-scale temporal features that characterize the short-term fluctuations and long-term degradation trends of the parameter through dilated convolutional layers with different dilation coefficients.

[0072] In this step, a historical time-series dataset of the target intrinsic parameters (i.e., the solid-phase diffusion coefficients of the positive and negative electrodes and the reaction rate constant) is first constructed for each battery cell. This dataset contains parameter values ​​collected at fixed time intervals (e.g., one data point is recorded after each parameter identification in step S10, on a minute or hourly basis), forming a time series. The sequence length can be set as needed, for example, containing data points from the past 1000 time steps. Subsequently, this time-series data is input into a Temporal Convolutional Network (TCN) module for feature extraction. The TCN module contains multiple residual blocks, each containing a series of dilated convolutional layers with expansion coefficients that grow exponentially (e.g., 1, 2, 4, 8, ...), thereby rapidly expanding the receptive field as the network depth increases. This allows for the simultaneous capture of local short-term fluctuations (e.g., instantaneous parameter changes due to temperature abrupt changes) and global long-term degradation trends (e.g., slow parameter decay due to battery aging) in the time series. Each dilated convolutional layer is followed by a ReLU activation function and layer normalization to improve training stability. After processing by the TCN module, a high-level feature tensor rich in multi-scale information is output, which effectively encodes the key patterns in the evolution of parameters throughout history.

[0073] S2032': Input the multi-scale temporal features into the attention mechanism module; the attention mechanism module generates a corresponding attention weight vector by calculating the importance score of each time step in the multi-scale temporal features, and multiplies the attention weight vector with the original features to obtain the weighted temporal features, which strengthen the feature representation at the parameter abrupt change inflection point;

[0074] In this step, the multi-scale temporal feature tensor output by the TCN module is input into an attention mechanism module. This module employs a scaled dot product attention mechanism, where the query, key, and value vectors are all obtained by mapping the input feature tensor through different linear transformation layers. The attention mechanism first calculates the similarity score between the feature vector of each time step in the feature sequence and the feature vectors of all time steps (including itself). Then, it normalizes these scores into an attention weight vector using a Softmax function. Each element in this vector represents the importance score of the corresponding time step feature for the current lifetime prediction task. In particular, this mechanism can automatically learn and assign higher weights to time steps in the parameter sequence that experience significant abrupt changes or inflection points (e.g., a sharp drop in the diffusion coefficient due to internal short circuits). Finally, the resulting attention weight vector is multiplied element-wise with the original multi-scale temporal features to achieve feature weighting. This process effectively highlights features related to key events in battery performance degradation and suppresses interference from irrelevant noise, thereby generating a weighted temporal feature representation focused on important degradation events.

[0075] S2033': Input the weighted time-series features into the fully connected regressor; the fully connected regressor is configured to output the parameters of a Gaussian mixture model, thereby defining the probability distribution of the remaining service life of the battery cell; the final predicted value of the remaining service life is the mathematical expectation of the probability distribution, which is obtained by calculating the weighted sum of the weights, mean and variance of the Gaussian distributions of all components in the Gaussian mixture model.

[0076] In this step, the attention-weighted temporal feature tensor (compressed into a fixed-length feature vector by a global average pooling layer) is input into a fully connected regressor. This regressor is not a simple point estimate output, but is designed to output a set of parameters defining a Gaussian Mixture Model (GMM). Specifically, for a preset of K Gaussian distribution components (e.g., K=3), the regressor's output layer has 3K neurons, each corresponding to a weight (π) for one of the K components. k , satisfying Σπ k =1), mean (μ) k (representing possible RUL values) and variance ( (Representing uncertainty). These output parameters are constrained by an appropriate activation function, such as the weights π. k Variance is measured using the Softmax function. The exponential function ensures a positive value. Therefore, the remaining lifespan of a single battery cell is modeled as a probability distribution function described by a Generic Model (GMM), which effectively represents the inherent uncertainty of the prediction results. Finally, the point prediction value of the remaining lifespan is taken as the mathematical expectation (i.e., the mean) of this GMM. The formula for calculating the mathematical expectation E[RUL] is:

[0077] This probabilistic output method not only provides the expected lifespan but also implicitly includes prediction confidence information, offering richer decision-making basis for subsequent risk-aware optimization control. The entire hybrid prediction model needs to be trained offline on a historical battery aging dataset, with the training objective being to minimize the negative log-likelihood loss between the predicted RUL and the actual RUL.

[0078] like Figure 4 As shown, in some embodiments, step S30 may include:

[0079] S301: The overall remaining service life is quantified as the weighted average of the remaining service life of all battery cells in the lithium battery energy storage system; the current cycle energy efficiency is quantified as the ratio of net output energy to total energy consumption in this charge-discharge cycle; and the quality benefit index of the grid frequency regulation service is quantified as a comprehensive score of the accuracy and response speed of the lithium battery energy storage system's response frequency deviation.

[0080] In this step, the three optimization objectives are first precisely quantified. The overall remaining useful life (RULsys) is calculated using a weighted average: RULsys = Σ(wi × RULi), where RULi is the predicted remaining useful life of the i-th battery cell under the current operating scenario, and the weight wi is dynamically allocated based on the current health status of the battery cell and its importance in the system. The current cycle efficiency (ηc) is calculated as: ηc = |Eout| / Ein, where Eout is the total energy released to the grid or load in a complete charge-discharge cycle, and Ein is the total energy absorbed from the grid or power source in that cycle. The grid frequency regulation service quality benefit index (Qg) is calculated using a preset scoring function: Qg = α × A + β × V, where A is calculated based on the root mean square error of the deviation between the actual frequency response and the command frequency, V is calculated based on the delay time of the response command, and α and β are weighting coefficients. These three quantified indicators together constitute the objective function of the multi-objective optimization problem.

[0081] S302: Construct a multi-dimensional state vector using the negative electrode solid phase surface concentration, positive and negative electrode overpotentials, and predicted remaining lifespan of each battery cell, as the state space of the reinforcement learning environment;

[0082] In this step, the design of the state space S needs to comprehensively reflect the current state of the system. For an energy storage system containing N battery cells, the dimension of the state vector S is 4N (each cell contributes 4 state variables: negative electrode solid surface concentration Cs,n, positive electrode overpotential ηp, negative electrode overpotential ηn, and predicted remaining service life RULi). That is, S = [Cs,n,1,ηp,1,ηn,1,RUL1,Cs,n,2,ηp,2,ηn,2,RUL2,...,Cs,n,N,ηp,N,ηn,N,RULN]. These state variables are obtained in real time from the calculation results of step S20. The negative electrode solid surface concentration is directly related to the risk of lithium deposition, the overpotential reflects the degree of electrochemical polarization, and RUL characterizes long-term durability. Before inputting into the agent, the state vector is normalized to scale all variables to the [0,1] interval to accelerate the convergence of the learning process.

[0083] S303: Construct a multi-dimensional action vector using the set of values ​​for the differentiated charging and discharging current command and the set of values ​​for the maximum allowable charging and discharging power at the system level, as the action space of the reinforcement learning agent;

[0084] In this step, the action space A defines the control commands that the agent can execute. It is also a multi-dimensional vector containing N+1 dimensions: the first N dimensions correspond to the differentiated charge / discharge current commands Ii (unit: A) for N individual battery cells, and their values ​​are determined by the maximum allowable charge / discharge current of each individual battery cell; the N+1th dimension corresponds to the maximum allowable charge / discharge power Pmax (unit: kW) at the system level, and its values ​​are determined by the converter capacity and grid dispatch requirements. Therefore, the action vector A = [I1, I2, ..., IN, Pmax]. The agent outputs a specific action vector in each control cycle, which is a set of specific control commands. In this embodiment, a continuous action space is used for more refined optimization.

[0085] S304: Design a reward function, which includes a first reward item positively correlated with the overall remaining useful life, a second reward item positively correlated with the current cycle energy efficiency, and a third reward item positively correlated with the quality benefit index of the grid frequency regulation service;

[0086] In this step, the reward function R is the core of guiding the agent to learn optimization strategies. It is designed as a weighted sum of three reward terms: R = w1R1 + w2R2 + w3R3. The first reward term R1 is positively correlated with the overall remaining lifetime, for example, R1 = k1 * (RULsys - RULmin) / (RULmax - RULmin). This term encourages the agent to take actions that extend the overall lifetime of the system. The second reward term R2 is positively correlated with the current cycle energy efficiency, for example, R2 = k2 * ηc. This term incentivizes the agent to improve energy conversion efficiency. The third reward term R3 is positively correlated with the grid frequency regulation service quality benefit index, for example, R3 = k3 * Qg. This term ensures the system's contribution to grid service. The weighting coefficients w1, w2, and w3 are used to balance the relative importance of the three objectives.

[0087] S305: Based on the current state space, the reinforcement learning algorithm selects an action in the action space by combining exploration and utilization. After execution, it calculates the reward according to the reward function, iteratively updates its decision-making strategy, and finally outputs the optimal system-level control instruction set.

[0088] This step describes the specific execution process of the reinforcement learning algorithm. Taking the proximal policy optimization algorithm as an example, its online learning process is as follows: The agent's decision policy is represented by a neural network πθ. The agent observes the current state st. Based on the current policy πθ, it randomly selects action at with a certain probability, or selects the action at with the highest expected reward with the highest probability. Action at is a set of control instructions, which is issued to the BMS for execution. Under the action at, the environment transitions to a new state st+1 and generates an immediate reward rt. This experience tuple (st, at, rt, st+1) is stored in the experience replay buffer. Subsequently, the algorithm samples a batch of experience data from the buffer, calculates the advantage function and policy gradient, and iteratively updates the policy network parameters θ using stochastic gradient descent, aiming to maximize the expected long-term cumulative discount reward. Through numerous iterations, the policy network gradually learns an optimal policy πθ* that can intelligently balance lifetime, energy efficiency, and service quality. Finally, using this trained optimal policy, based on the real-time state st, it outputs the current optimal system-level control instruction set At = [I1, I2, ..., IN, Pmax], realizing real-time optimization control of the system.

[0089] In some embodiments, in step S301, the comprehensive score Q of the accuracy of the response frequency deviation and the response speed of the lithium battery energy storage system is... f Calculated based on the following formula:

[0090]

[0091] Wherein, RMSE is the root mean square error between the actual output power of the lithium battery energy storage system and the grid dispatch command; T d P is the response delay time required for the lithium battery energy storage system to receive a command and for the actual power to reach 90% of the command requirement; max The system's rated power is used to normalize the RMSE; T max The maximum allowable delay time required by the power grid, used for T d Normalize; w a and w s These are the weighting coefficients for accuracy and response speed, respectively, and w a and w s The sum is 1;

[0092] In step S304, the reward function adopts a weighted summation form, expressed as: R = w1R1 + w2R2 + w3R3, where R1, R2, and R3 correspond to the first reward item, the second reward item, and the third reward item, respectively, and w1, w2, and w3 are the corresponding weight coefficients, and the weight coefficients can be dynamically adjusted according to the operating mode of the lithium battery energy storage system.

[0093] In step S305, the reinforcement learning algorithm is a proximal policy optimization algorithm, a soft actor-commentator algorithm, or a deep deterministic policy gradient algorithm; the update of the decision policy adopts an offline learning method based on experience replay.

[0094] like Figure 5 As shown, in some embodiments, step S40 may specifically include:

[0095] S401: The optimal system-level control instruction set is sent to the battery management system for execution, and the actual voltage response data and acoustic emission signal of each battery cell in the lithium battery energy storage system after executing the system-level control instruction set are collected simultaneously.

[0096] Upon receiving the optimal system-level control command set (including differentiated charge / discharge current commands and the system-level maximum allowable charge / discharge power) generated in step S30, the central controller transmits this command set to the main controller of the battery management system (BMS) in real time via communication protocols such as CAN bus or Ethernet. The BMS drives the power conversion system (PCS) and the equalization circuits of each battery cell to perform corresponding charge / discharge operations according to the command set. Simultaneously, the data acquisition system starts synchronously at a high sampling rate (e.g., 1kHz), measuring the terminal voltage of each battery cell using a high-precision voltage sensor connected in parallel, as the actual voltage response data; and collecting acoustic emission signals generated by the battery under electrochemical changes and mechanical stress using an acoustic emission sensor attached to the battery casing. All data is timestamped to ensure data synchronization.

[0097] S402: Take the historical voltage data and historical current data collected during the execution of the system-level control instruction set as input, substitute them into the incremental electrochemical model, and obtain the predicted voltage sequence for the corresponding time period; calculate the root mean square error between the actual voltage response data and the predicted voltage sequence as the error;

[0098] This step, after a complete cycle of instruction set execution (e.g., a single charge / discharge pulse lasting several minutes), extracts historical data collected during that time period: including historical current data I(t) as model input and historical actual voltage data U_actual(t) for comparison. The historical current data I(t) is substituted into the currently active incremental electrochemical model, and simulation calculations are performed at the same sampling time intervals, outputting the model-predicted voltage sequence U_predicted(t) for the corresponding time period. Subsequently, the root mean square error (RMSE) between the actual voltage sequence and the predicted voltage sequence is calculated using the formula: RMSE = sqrt[mean((U_actual(t) - U_predicted(t))] 2The RMSE value quantitatively characterizes the predictive accuracy of the electrochemical model under current operating conditions.

[0099] S403: Perform time-frequency analysis on the collected acoustic emission signal, extract its energy distribution characteristics in different frequency bands, and compare the energy distribution characteristics with the preset reference energy distribution under normal operating conditions. If the energy of any frequency band exceeds a preset multiple of the reference value, it is determined to be the abnormal feature.

[0100] This step preprocesses the acoustic emission signal acquired in step S401 (denoising and filtering), and then performs time-frequency analysis using Short-Time Fourier Transform (STFT) or wavelet transform to decompose the signal into multiple preset frequency bands (e.g., 0-50kHz, 50-150kHz, 150-300kHz). The signal energy (i.e., the sum of squares of amplitude) within each frequency band is calculated to obtain the energy distribution feature vector [E_band1, E_band2, ..., E_bandN] of the current acoustic emission signal. This feature vector is then compared band-by-band with the baseline energy distribution vector [E_base1, E_base2, ..., E_baseN] obtained through statistical analysis of a large amount of normal operating condition experimental data. If the energy E_bandi of any current frequency band i exceeds a preset multiple (e.g., 3 times) of its corresponding baseline energy E_basei, it is determined that the frequency band has an energy anomaly, and thus the acoustic emission signal exhibits abnormal characteristics. This abnormal characteristic indicates a fault inside the battery, such as lithium plating, separator damage, or a micro-short circuit.

[0101] S404: If the root mean square error calculated in step S402 exceeds the preset threshold, and / or if step S403 determines that the abnormal feature has occurred, then a model calibration trigger signal is generated.

[0102] This step sets a preset RMSE threshold (e.g., 20mV), which is determined based on model accuracy requirements and battery characteristics. The RMSE calculated in step S402 is compared with this threshold. Simultaneously, the judgment result of step S403 is checked. If any of the following conditions are met, or both are met: RMSE > preset threshold; the acoustic emission signal is determined to have abnormal characteristics, then the logic judgment unit of the central controller will immediately generate a high-level model calibration trigger signal. This signal serves as a flag for re-identifying the start parameters.

[0103] S405: In response to the model calibration trigger signal, the online parameter identification algorithm is triggered to update and calibrate the target intrinsic parameters of the incremental electrochemical model based on the actual voltage response data collected in step S401, the historical current data used in step S402, and the temperature data for the corresponding time period obtained from the real-time acquisition in step S10.

[0104] When the model calibration trigger signal is valid, the system immediately interrupts the current normal operation process and calls the online parameter identification algorithm (recursive least squares method or extended Kalman filter) described in step S10. During calibration, data collected within the time window preceding the trigger time (e.g., the time period corresponding to the just executed instruction set) is used as the input dataset, including: the actual voltage response data U_actual(t) as the model output observation, the historical current data I(t) as the model input excitation, and the average temperature data T_avg for the corresponding time period (obtained from the real-time temperature acquisition in step S10). The algorithm aims to minimize the voltage prediction error and re-optimizes and updates the estimated values ​​of the target intrinsic parameters (solid-phase diffusion coefficients of positive and negative electrodes and reaction rate constants). After the update is completed, the incremental electrochemical model is refreshed with the new parameter values ​​to re-match it with the actual state of the battery cell, thereby completing a closed-loop calibration of the model and ensuring the accuracy of subsequent state estimation and control optimization. After the calibration process is completed, the system resumes normal operation.

[0105] In some embodiments, in step S402, when calculating the root mean square error, a sliding time window is used to align the actual voltage response data and the predicted voltage sequence; in step S403, the time-frequency analysis uses wavelet transform or short-time Fourier transform; in step S403, the preset multiple is dynamically adjusted according to the type and life stage of the battery cell, and the range is set to 2 to 5 times; in step S405, the recursive least squares method is used to update and calibrate the target intrinsic parameters.

[0106] In some embodiments, the method may further include the following steps:

[0107] S50: Digital Twin Synchronization and Advanced Simulation Steps: A high-fidelity digital twin is established for the lithium battery energy storage system. This digital twin includes the incremental electrochemical model of each battery cell and its connection topology. Before issuing the optimal system-level control instruction set to the actual battery management system, the instruction set is executed on the digital twin in an accelerated simulation manner to predict the changing trend of key system parameters over a period of time. If the prediction results show that any parameter will exceed the safety boundary, the instruction set is rejected and step S30 is triggered to recalculate the optimization.

[0108] In practical implementation, a high-fidelity digital twin is established for the physical lithium battery energy storage system in a server or high-performance edge computing unit. This twin is constructed by integrating the latest incremental electrochemical models of each battery cell and strictly following the electrical series / parallel topology of the actual system, thereby accurately reproducing the dynamic characteristics of the physical system in virtual space. Before sending the optimal system-level control instruction set generated in step S30 to the actual battery management system for execution, the instruction set is first sent to the digital twin. Based on its internal model, the digital twin executes the corresponding operations of the instruction set at a simulation speed much faster than real-time (e.g., 10 times faster) and predicts the evolution trend of key internal state parameters of each battery cell (e.g., negative electrode solid surface concentration, overpotential, temperature) and the overall system performance in a short time domain (e.g., the next 5 minutes). During the simulation, it monitors in real time whether these predicted parameters will exceed the preset safe operating boundaries (e.g., whether the negative electrode surface concentration of any battery cell will fall below the lithium plating threshold, whether the temperature will exceed the upper limit, and whether the voltage will exceed the extreme value). If the prediction results show that all parameters are within the safety boundary, the original instruction set is authorized to be issued to the physical system for execution; if any parameter is predicted to have a risk of exceeding the boundary, the instruction set is immediately rejected, and a flag signal is fed back to the optimization controller in step S30, triggering it to recalculate the optimization based on the current system state to generate a safer new instruction set, thereby avoiding potential risks in advance before actual action and achieving early safety protection.

[0109] In some embodiments, the method may further include the following steps:

[0110] S60: Abnormal operating condition simulation steps based on adversarial generative networks: using the characteristics of historical abnormal data of the learning system through adversarial generative networks to generate virtual data simulating various extreme operating conditions and failure modes; injecting these virtual data into the training process of the reinforcement learning algorithm to enhance the decision robustness of the algorithm when facing rare abnormal situations.

[0111] In specific implementation, firstly, a dataset containing various abnormal events (such as internal short circuits, lithium plating, and abnormally increased connection impedance) that have occurred in the historical operation of the lithium battery energy storage system is collected and constructed. This dataset includes time series of important parameters recorded when the abnormality occurred, such as voltage, current, temperature, and corresponding acoustic emission signal characteristics. A conditional generative adversarial network (GAN) is trained using this abnormal data. The generator learns the potential distribution characteristics of various abnormal data, while the discriminator is responsible for distinguishing between real abnormal data and the virtual data generated by the generator. After sufficient training, the generator can generate realistic virtual multi-parameter time series data that conforms to specific abnormal pattern characteristics based on the input conditional labels (such as fault type and severity). Subsequently, in the offline training phase of the reinforcement learning algorithm (such as the PPO algorithm) described in step S30, this generated virtual data, covering various rare extreme operating conditions and fault modes, is injected into the simulation environment with a certain probability, constructing training scenarios such as a sudden drop in the voltage of a single battery cell or an abnormal increase in the system temperature gradient, forcing the reinforcement learning agent to attempt decision-making under these challenging abnormal states. By conducting extensive training iterations in this enhanced environment that injects anomalous data, reinforcement learning agents can learn robust strategies for handling various sudden anomalies while ensuring system safety. For example, when a suspected internal short circuit is detected, the charging and discharging power can be proactively reduced to prevent the accident from escalating, thereby improving the robustness of their decision-making and the safety of the system in the face of uncertainty and rare failures in real operation.

[0112] Example 2

[0113] This invention also provides an intelligent management system for a lithium battery energy storage system, comprising:

[0114] The data acquisition module is used to collect voltage, current, temperature and acoustic emission signals of each battery cell in the lithium battery energy storage system in real time during operation.

[0115] The model update module is used to dynamically correct the target intrinsic parameters in a preset basic electrochemical model based on the voltage data, current data, and temperature data using an online parameter identification algorithm, so as to generate an incremental electrochemical model that matches the current actual electrochemical state of each battery cell; the target intrinsic parameters include the solid-phase diffusion coefficients of the positive and negative electrodes and the reaction rate constant;

[0116] The state estimation and lifetime prediction module is used to take the real-time collected voltage and current data as input, substitute them into the incremental electrochemical model corresponding to each battery cell, and solve the internal state parameters of each battery cell in real time. The internal state parameters include the negative electrode solid phase surface concentration and the positive and negative electrode overpotentials. Based on the current trend of the change of the target intrinsic parameters, the module uses an extrapolation algorithm to predict the remaining lifetime of each battery cell under various preset future operating scenarios.

[0117] The optimization control module is used to optimize the overall remaining service life, current cycle efficiency, and quality benefit index of the grid frequency regulation service of the lithium battery energy storage system as optimization objectives. The internal state parameters and predicted remaining service life of each battery cell are used as constraints. The module uses a reinforcement learning algorithm to solve for the optimal system-level control instruction set in real time. The system-level control instruction set includes differentiated charge and discharge current instructions for inconsistent battery cells and the maximum allowable charge and discharge power at the system level.

[0118] The instruction execution and model calibration module is used to issue the optimal system-level control instruction set to the battery management system for execution; and after execution, to obtain the error between the actual voltage response of the battery cell and the predicted voltage of the incremental electrochemical model, while monitoring the characteristic changes of the acoustic emission signal. If the error exceeds a preset threshold and / or the acoustic emission signal shows abnormal characteristics, the online parameter identification algorithm is triggered to update and calibrate the target intrinsic parameters of the incremental electrochemical model.

[0119] Furthermore, the model update module includes:

[0120] The model loading unit is used to load the preset basic electrochemical model for each battery cell and set the initial values ​​of the target intrinsic parameters.

[0121] The first parameter correction unit is used to identify and correct the reaction rate constant in the target intrinsic parameter online at a first time interval of seconds or milliseconds, based on the voltage and current data collected in real time at the second or millisecond level, using the least squares method or gradient descent method, to obtain the corrected reaction rate constant.

[0122] The second parameter correction unit is used to perform temperature compensation on the positive and negative electrode solid-phase diffusion coefficients based on the average value of the temperature data collected in the first time interval at a second time interval on the order of minutes, using the Arrhenius formula. Based on the initial value of the compensated positive and negative electrode solid-phase diffusion coefficients, and combined with the voltage data and current data, the unit uses the least squares method or the extended Kalman filter algorithm to correct the positive and negative electrode solid-phase diffusion coefficients in the target intrinsic parameters, thereby obtaining the corrected positive and negative electrode solid-phase diffusion coefficients.

[0123] The model generation unit is used to substitute the modified reaction rate constant and the modified positive and negative electrode solid-phase diffusion coefficients into the basic electrochemical model to generate the incremental electrochemical model at the current moment.

[0124] Furthermore, the second parameter correction unit performs temperature compensation on the solid-phase diffusion coefficients of the positive and negative electrodes using the Arrhenius formula, specifically through the following formula:

[0125] D(T)=D×exp[-Ea_D / R×(1 / T-1 / T_ref)]

[0126] Where D(T) is the initial value of the solid-phase diffusion coefficient of the positive and negative electrodes after compensation with temperature data T, D is the preset value of the solid-phase diffusion coefficient of the positive and negative electrodes at the reference temperature, Ea_D is the diffusion activation energy, R is the ideal gas constant, and T is the average value of the temperature data.

[0127] Furthermore, the state estimation and lifetime prediction module includes:

[0128] The state estimation unit is used to take the real-time collected voltage and current data as input and substitute them into the incremental electrochemical model corresponding to each battery cell. By solving the electrochemical differential algebraic equations described by the incremental electrochemical model, the negative electrode solid phase surface concentration and positive and negative electrode overpotentials of each battery cell are calculated in real time.

[0129] The degradation trajectory modeling unit is used to establish a degradation trajectory model for each battery cell based on historical and current intrinsic parameter data of the target, showing the changes of its positive and negative electrode solid-phase diffusion coefficients and reaction rate constants over time.

[0130] The lifespan prediction unit is used to take the value of the target intrinsic parameter at the current moment as the initial condition for the preset multiple future operating scenarios, input it into the corresponding degradation trajectory model, and extrapolate the changes of the target intrinsic parameter on the future time axis through the extrapolation algorithm until any target intrinsic parameter reaches its failure threshold. This moment is taken as the predicted lifespan end point, and the duration from the current moment to the lifespan end point is calculated as the remaining service life of the battery cell in the corresponding future operating scenario.

[0131] Furthermore, in the state estimation unit, the electrochemical differential-algebraic equations described by the incremental electrochemical model include:

[0132] The solid-phase diffusion equation describes the relationship between the diffusion process of lithium ions in electrode spherical particles and the spatiotemporal distribution of solid-phase lithium ion concentration through the solid-phase diffusion coefficient.

[0133] The Butler-Folmer equation, which correlates reaction current density with positive and negative electrode overpotentials through the reaction rate constant, characterizes the electrochemical reaction kinetics on the electrode surface.

[0134] The voltage balance equation establishes a balance between the battery terminal voltage and the open-circuit voltage, the positive and negative electrode overpotentials, and the voltage drop across the ohmic internal resistance, wherein the open-circuit voltage is a function of the normalized concentration of lithium ions on the negative electrode solid surface.

[0135] Furthermore, in the lifetime prediction unit, the execution of the extrapolation algorithm includes:

[0136] The probability distribution modeling subunit is used to determine the uncertainty range of each model parameter in the degradation trajectory model based on the historical intrinsic parameter data of the target and to establish the probability distribution model of the degradation trajectory model parameters.

[0137] The Monte Carlo sampling subunit is used to create an independent prediction thread for each future running scenario. It randomly samples a set of parameter values ​​from the probability distribution model of the parameters of the degradation trajectory model, substitutes the randomly sampled parameter values ​​into the degradation trajectory model, and performs deterministic extrapolation with the target intrinsic parameter value at the current moment as the initial condition.

[0138] The prediction correction subunit is used to correct the prediction results of the target intrinsic parameters obtained from the degenerate trajectory model by using an unscented Kalman filter algorithm in each deterministic extrapolation iteration, so as to obtain the corrected predicted values ​​of the target intrinsic parameters.

[0139] The life end determination subunit is used to determine that the life end has been reached when the predicted value of any modified intrinsic parameter of the target exceeds three times the standard deviation of the probability distribution corresponding to its failure threshold for the first time.

[0140] The result statistics subunit is used to repeatedly perform Monte Carlo sampling and prediction correction to reach a preset minimum number of times to obtain the Monte Carlo statistical distribution of the remaining lifetime, and take the mathematical expectation of the Monte Carlo statistical distribution as the final predicted remaining lifetime.

[0141] Furthermore, the optimization control module includes:

[0142] The target quantification unit is used to quantify the overall remaining service life as a weighted average of the remaining service life of all battery cells in the lithium battery energy storage system, quantify the current cycle energy efficiency as the ratio of net output energy to total energy consumption in this charge-discharge cycle, and quantify the quality benefit index of the grid frequency regulation service as a comprehensive score of the accuracy and response speed of the lithium battery energy storage system's response frequency deviation.

[0143] A state space construction unit is used to construct a multi-dimensional state vector based on the negative electrode solid phase surface concentration, positive and negative electrode overpotentials and predicted remaining lifetime of each battery cell, as the state space of the reinforcement learning environment.

[0144] The action space construction unit is used to construct a multi-dimensional action vector using the set of values ​​for the differentiated charging and discharging current command and the set of values ​​for the maximum allowable charging and discharging power at the system level, as the action space of the reinforcement learning agent.

[0145] The reward function design unit is used to design a reward function, which includes a first reward item positively correlated with the overall remaining useful life, a second reward item positively correlated with the current cycle energy efficiency, and a third reward item positively correlated with the quality benefit index of the grid frequency regulation service.

[0146] The optimization unit is used to select actions in the action space based on the current state space by combining exploration and utilization of the reinforcement learning algorithm. After execution, the reward is calculated according to the reward function, and the decision strategy is iteratively updated to finally output the optimal system-level control instruction set.

[0147] Furthermore, in the target quantization unit, the comprehensive score Qf of the accuracy of the response frequency deviation and the response speed of the lithium battery energy storage system is calculated based on the following formula:

[0148] Qf = w a ×(1-RMSE / Pmax)+w×(1-Td / Tmax);

[0149] Wherein, RMSE is the root mean square error between the actual output power of the lithium battery energy storage system and the grid dispatch command; Td is the response delay time required for the lithium battery energy storage system to reach 90% of the command requirement from receiving the command; Pmax is the rated power of the system; Tmax is the maximum allowable delay time required by the grid; w a and w are the weighting coefficients for accuracy and response speed, respectively, and w a The sum of w and is 1.

[0150] Furthermore, the instruction execution and model calibration module includes:

[0151] The instruction execution and data acquisition unit is used to send the optimal system-level control instruction set to the battery management system for execution, and simultaneously acquire the actual voltage response data and acoustic emission signals of each battery cell in the lithium battery energy storage system after executing the system-level control instruction set;

[0152] The error calculation unit is used to take historical voltage data and historical current data collected during the execution of the system-level control instruction set as input, substitute them into the incremental electrochemical model, and obtain the predicted voltage sequence for the corresponding time period; and calculate the root mean square error between the actual voltage response data and the predicted voltage sequence as the error.

[0153] The acoustic emission signal analysis unit is used to perform time-frequency analysis on the acquired acoustic emission signal, extract its energy distribution characteristics in different frequency bands, and compare the energy distribution characteristics with the preset reference energy distribution under normal operating conditions. If the energy of any frequency band exceeds a preset multiple of the reference value, it is determined to be the abnormal feature.

[0154] A trigger signal generation unit is used to generate a model calibration trigger signal if the root mean square error exceeds the preset threshold and / or if the abnormal feature is determined to occur.

[0155] The model calibration unit is used to trigger the online parameter identification algorithm in response to the model calibration trigger signal, and update and calibrate the target intrinsic parameters of the incremental electrochemical model based on the actual voltage response data, the historical current data, and the temperature data for the corresponding time period obtained from the real-time acquisition module.

[0156] Furthermore, when calculating the root mean square error, the error calculation unit uses a sliding time window to align the actual voltage response data and the predicted voltage sequence.

[0157] Example 3

[0158] See Figure 6 This application also provides an electronic device 600, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform an intelligent management method for a lithium battery energy storage system as described in the foregoing method embodiments. This application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform an intelligent management method for a lithium battery energy storage system as described in the foregoing method embodiments. Figure 6As shown, electronic device 60 may include a processing unit (e.g., central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of electronic device 600. Processing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604. The following devices may be connected to I / O interface 605: input devices 606 including, for example, touch screen, touchpad, keys, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 607 including, for example, liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 608 including, for example, magnetic tape, hard disk, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data.

[0159] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. An intelligent management method for a lithium battery energy storage system, characterized in that, Includes the following steps: S10: Real-time acquisition of voltage, current, temperature and acoustic emission signals of each battery cell in the lithium battery energy storage system during operation; Using an online parameter identification algorithm, the target intrinsic parameters in a pre-set basic electrochemical model are dynamically corrected to generate an incremental electrochemical model that matches the current actual electrochemical state of each battery cell. S20: The voltage and current data collected in real time are used as inputs and substituted into the incremental electrochemical model corresponding to each battery cell to solve the internal state parameters of each battery cell in real time. S30: The optimization objective is to maximize the overall remaining service life, current cycle efficiency, and quality benefit indicators of the grid frequency regulation service of the lithium battery energy storage system. S40: Send the optimal system-level control instruction set to the battery management system for execution; After execution, the error between the actual voltage response of the battery cell and the predicted voltage of the incremental electrochemical model is obtained, while the characteristic changes of the acoustic emission signal are monitored.

2. The method as described in claim 1, characterized in that, Step S10 includes: S101: Load the preset basic electrochemical model for each battery cell and set the initial values ​​of the target intrinsic parameters; the target intrinsic parameters include the solid-phase diffusion coefficients of the positive and negative electrodes and the reaction rate constant; S102: At a first time interval of seconds or milliseconds, based on the real-time collected voltage and current data, the reaction rate constant in the target intrinsic parameter is identified and corrected online to obtain the corrected reaction rate constant; S103: Using a second time interval on the order of minutes, based on the average value of the temperature data collected within the first time interval, perform temperature compensation on the positive and negative electrode solid-phase diffusion coefficients, and correct the positive and negative electrode solid-phase diffusion coefficients in the target intrinsic parameters to obtain the corrected positive and negative electrode solid-phase diffusion coefficients. S104: Substitute the modified reaction rate constant and the modified positive and negative electrode solid-phase diffusion coefficients into the basic electrochemical model to generate the incremental electrochemical model at the current moment.

3. The method as described in claim 1, characterized in that, In step S20, the internal state parameters include the negative electrode solid phase surface concentration and the positive and negative electrode overpotentials; and based on the current trend of the change of the target intrinsic parameters, the remaining service life of each battery cell under various preset future operating scenarios is predicted by extrapolation algorithm.

4. The method as described in claim 1, characterized in that, Step S20 includes: S201: The voltage and current data collected in real time are used as inputs and substituted into the incremental electrochemical model corresponding to each battery cell to calculate the negative electrode solid phase surface concentration and positive and negative electrode overpotential of each battery cell in real time. S202: Based on historical and current intrinsic parameter data of the target, establish a degradation trajectory model for the solid-phase diffusion coefficients and reaction rate constants of the positive and negative electrodes of each battery cell as a function of time. S203: For the preset multiple future operating scenarios, the value of the target intrinsic parameter at the current moment is used as the initial condition and input into the corresponding degradation trajectory model. The change of the target intrinsic parameter on the future time axis is deduced through the extrapolation algorithm until any target intrinsic parameter reaches its failure threshold. This moment is taken as the predicted end of life and as the remaining service life of the battery cell in the corresponding future operating scenario.

5. The method as described in claim 4, characterized in that, In step S201, the electrochemical differential algebraic equations described by the incremental electrochemical model include: solid-phase diffusion equation, Butler-Folmer equation, and voltage balance equation.

6. The method as described in claim 4, characterized in that, In step S202, the degradation trajectory model adopts linear regression, exponential fitting, Arrhenius model or SEI film growth kinetic model.

7. The method as described in claim 4, characterized in that, In step S203, the extrapolation algorithm employs a multi-scenario probability extrapolation algorithm based on adaptive unscented Kalman filtering.

8. The method as described in claim 4, characterized in that, In step S203, the extrapolation algorithm is implemented through a hybrid prediction model, which includes a temporal convolutional network module, an attention mechanism module, and a fully connected regressor connected in sequence.

9. The method as described in claim 1, characterized in that, In step S30, the internal state parameters and predicted remaining service life of each battery cell are used as constraints. The optimal system-level control instruction set is solved in real time using a reinforcement learning algorithm. The system-level control instruction set includes differentiated charge and discharge current instructions for inconsistent battery cells and the maximum allowable charge and discharge power at the system level.

10. The method as described in claim 1, characterized in that, Step S40 further includes: if the error exceeds a preset threshold and / or the acoustic emission signal exhibits abnormal characteristics, triggering the online parameter identification algorithm to update and calibrate the target intrinsic parameters of the incremental electrochemical model.

Images