Active equalization management method and system for battery pack

Abstract

The invention relates to the technical field of battery management systems, and discloses an active equalization management method and system for a battery pack, and the method comprises the steps: collecting the parameter data of the battery pack, and carrying out the dynamic decomposition of the system through a time scale decomposition algorithm; constructing a probabilistic state transition model, and carrying out probabilistic prediction on the future state of the battery pack; three parallel predictive controllers are constructed; a traditional fuzzy rule is expanded into a probability form, and decision uncertainty is quantified through Bayesian fuzzy reasoning; constructing a risk adjustment type objective function based on the conditional value-at-risk, and generating a balance control decision and a confidence interval thereof; adjusting the execution intensity, and outputting an equalization current instruction; through multi-time-scale data decomposition and parallel prediction control architecture, comprehensive management of the battery pack from second-level transient response to month-level healthy evolution is realized, and the limitation of a traditional single-time-scale controller is broken.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and more specifically, to a battery pack active balancing management method and system. Background Art

[0002] Battery packs are widely used in modern applications such as electric vehicles, energy storage systems, and portable devices. In these applications, a battery pack is usually composed of multiple single cells connected in series and parallel. Due to factors such as manufacturing process differences, uneven temperature distribution, and different aging degrees, parameters such as the state of charge, health state, and voltage level of each single cell will show inconsistencies, resulting in a decline in battery pack performance, a reduction in capacity, and a shortening of service life.

[0003] To solve this problem, battery management systems usually adopt balancing techniques to balance the differences between single cells. Traditional balancing techniques mainly perform passive balancing based on voltage or power differences, such as resistive shunt balancing, capacitor shuttling balancing, and transformer coupling balancing methods. These methods mainly rely on currently collected battery parameters for decision-making and lack the ability to predict future operating conditions. With the development of model predictive control technology, some studies have begun to introduce predictive control into the battery balancing field, predicting future states by establishing a battery dynamic model and optimizing the balancing strategy.

[0004] However, the current battery balancing management still faces multiple technical challenges: First, traditional predictive control methods usually only consider a single time scale and are difficult to simultaneously handle the short-term dynamic response and long-term health state evolution during the battery pack balancing process; Second, most existing balancing decision models adopt deterministic methods and cannot effectively quantify the uncertainties in the prediction and control processes. Especially in abnormal situations such as sensor failures or data missing, the decision-making reliability is significantly reduced; Third, most balancing systems lack an effective risk assessment and management mechanism and cannot adjust the decision-making strategy according to uncertainties, easily leading to a decline in performance or even safety accidents due to over-balancing or under-balancing.

[0005] Therefore, there is an urgent need to provide a battery pack active balancing management method that can perform multi-time scale prediction, uncertainty quantification, and risk-aware decision-making to improve the prediction accuracy, decision-making reliability, and environmental adaptability of the balancing system. Summary of the Invention

[0006] The present invention provides a battery pack active balancing management method and system, which solve the technical problems in the related art that it is difficult to simultaneously handle multi-time scale dynamics, cannot quantify decision-making uncertainties, and lack risk-aware decision-making.

[0007] The present invention provides a battery pack active balancing management method, including the following steps: Collect parameter data of the battery pack, and decompose the system dynamics into three time-scale components, i.e., short-term dynamics, medium-term dynamics, and long-term dynamics, through a time-scale decomposition algorithm; Based on the time-scale components, construct a probabilistic state transition model to probabilistically predict the future state of the battery pack and output a state prediction distribution; Based on the time-scale components and the state prediction distribution, construct three parallel predictive controllers, namely a short-term controller, a medium-term controller, and a long-term controller; According to the state prediction distribution and the output of the predictive controllers, extend traditional fuzzy rules to a probabilistic form and quantify decision-making uncertainty through Bayesian fuzzy inference; According to the state prediction distribution, the output of the predictive controllers, and the results of Bayesian fuzzy inference, construct a risk-adjusted objective function based on conditional value at risk to generate an equilibrium control decision and its confidence interval; According to the output of the predictive controllers, the results of Bayesian fuzzy inference, and the equilibrium control decision, and adjust the execution intensity according to the confidence interval to output an equilibrium current command.

[0008] Further, the collection of the battery pack parameter data includes: collecting parameter data of the battery pack, including voltage, current, temperature, and internal resistance, where the voltage and current signals are sampled at a relatively high frequency, the temperature is sampled at a medium frequency, and the internal resistance data is sampled at a low frequency.

[0009] Further, the time-scale decomposition algorithm includes: applying wavelet transform to perform multi-scale decomposition on the preprocessed time-series data, where the short-term dynamic component is from seconds to minutes, capturing transient responses and rapid fluctuations; the medium-term dynamic component is from hours to days, reflecting daily variations and charge-discharge cycles; the long-term dynamic component is from weeks to months, reflecting aging trends and seasonal variations.

[0010] Further, the probabilistic state transition model is implemented based on a variational autoencoder architecture, including an encoder network and a decoder network, estimating the posterior distribution through variational inference, estimating the model parameters using Bayesian learning methods, and performing multi-step probabilistic prediction using a particle filter algorithm.

[0011] Further, the short-term controller is based on a model predictive control framework, using a linearized state-space model, with a prediction range from seconds to minutes; the medium-term controller adopts a non-linear model predictive control architecture, using a neural network state transition model, with a prediction range from hours to days; the long-term controller is based on a health state model and a degradation model, combined with a periodic optimization strategy, with a prediction range from weeks to months.

[0012] Further, the neural network state transition model adopts an architecture combining a bidirectional long short-term memory network and a residual connection, and the model structure includes: The input layer receives the medium-term state vector and control actions; The feature extraction layer consists of two layers of bidirectional long short-term memory networks; The residual connection layer directly connects the input features to the deep features; The output layer uses a fully connected network to predict the state at the next moment.

[0013] Furthermore, the Bayesian fuzzy inference includes: Construct a Bayesian fuzzy rule base, and each Bayesian fuzzy rule is expressed in a probability form; Use the Bayesian learning method to estimate the parameters of the fuzzy set; Conduct probability fuzzy inference based on the Bayesian fuzzy rules and parameter distribution; Calculate the mean, variance, confidence interval and entropy of the inference output distribution to quantify the decision uncertainty.

[0014] Furthermore, in the risk-adjusted objective function constructed based on conditional value at risk, the risk-adjusted objective function includes maximizing the expected return term and minimizing the extreme risk term.

[0015] Furthermore, the execution intensity adjustment function for adjusting the execution intensity is: ; where represents the execution intensity adjustment coefficient at time , is the basic execution intensity, represents the natural exponential function, is the execution intensity decay coefficient, is the control action normalization factor, and represent the lower and upper bounds of the confidence interval of the control action respectively.

[0016] The present invention provides a battery pack active equalization management system for implementing the above-mentioned battery pack active equalization management method, including: A multi-time scale data processing module for collecting battery pack parameter data and performing time scale decomposition to decompose the data into dynamic components of different frequencies; A probability state prediction module for constructing a state transition probability model to achieve multi-step prediction of the future state of the battery pack and quantify the prediction uncertainty; A multi-level parallel control module for constructing short, medium and long-term prediction controllers based on different time scales to collaboratively manage the equalization process of the battery pack; A Bayesian fuzzy inference module for probabilistically expanding the fuzzy rules and quantifying the decision uncertainty through parameter learning and probability inference; A risk perception decision-making module, which is used to construct risk measurement indicators and a risk-adjusted objective function, and generate an equilibrium control strategy with a confidence interval; An adaptive integration execution module, which is used to fuse the outputs of each controller and adjust the execution intensity according to the uncertainty level, and generate a final equilibrium control signal.

[0017] The beneficial effects of the present invention are as follows: Through multi-time-scale data decomposition and a parallel predictive control architecture, comprehensive management of the battery pack from second-level transient response to monthly-level health evolution is achieved, breaking the limitations of traditional single-time-scale controllers; Through a Bayesian fuzzy inference framework, precise quantification of the uncertainty of the equalization decision is realized, and a probabilistic decision with a confidence interval can be output, rather than the single-point estimation of traditional methods. In abnormal situations such as data loss or sensor failure, the system can still maintain a high equalization efficiency, and the robustness is improved; Based on the conditional value-at-risk risk perception decision-making mechanism, the system risk in extreme situations can be accurately measured, the probability of abnormal events is reduced while maintaining the equalization efficiency, and the system safety is improved; Through the dynamic weight adjustment and execution intensity adaptive mechanism, the control objectives of different time scales can be intelligently balanced according to the current state and uncertainty level, realizing true active equalization management. Compared with the fixed weight method, the adaptability is improved under variable working conditions; Provide a probability distribution and a confidence interval for each equalization decision, making the decision-making process more transparent and interpretable. This not only facilitates human-machine collaboration and system supervision, but also improves the fault diagnosis efficiency; Realize the technical leap of the active equalization management of the battery pack from "passive response" to "active prediction and risk perception", providing a new technical path for high-performance and high-reliability battery management systems. Description of the Drawings

[0018] Figure 1 is a flowchart of a method for active equalization management of a battery pack in the present invention; Figure 2 is a flowchart of the multi-time-scale data processing steps in the present invention; Figure 3 is a flowchart of the probability state prediction steps in the present invention; Figure 4 is a flowchart of the multi-level parallel predictive control steps in the present invention; Figure 5 is a flowchart of the Bayesian fuzzy inference steps in the present invention; Figure 6 is a flowchart of the risk perception decision generation steps in the present invention; Figure 7It is a flowchart of the adaptive integration execution steps in the present invention. Detailed implementation manners

[0019] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the scope of protection of the content of this specification, changes can be made to the functions and arrangements of the elements discussed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0020] In at least one embodiment of the present invention, a method for active equalization management of a battery pack is disclosed. As Figures 1 to 7 shown, it includes: Step 1: Collect battery pack parameter data, and decompose the system dynamics into three time-scale components, namely short-term dynamics, medium-term dynamics, and long-term dynamics, through a time-scale decomposition algorithm. In this step, the battery pack monitoring data is processed by a time-scale decomposition algorithm to decompose the battery state dynamics into components of different time scales, providing a basis for subsequent multi-level parallel predictive control.

[0021] Specifically, it includes the following sub-steps: Step 1.1: Data acquisition; Collect parameter data such as voltage, current, temperature, and internal resistance of the battery pack. The sampling frequency is configured according to different signal characteristics: voltage and current signals are sampled at a higher frequency (such as 10 Hz), temperature is sampled at a medium frequency (such as 1 Hz), and internal resistance data is sampled at a low frequency (such as 0.1 Hz). The collected data set is denoted as: ; where represents the collected original data set, , , respectively represent the data vectors at the 1st, 2nd, th sampling moments, and is the total number of samplings.

[0022] Step 1.2: Data preprocessing; Preprocess the collected original data, including outlier detection, missing value filling, and signal smoothing. Outlier detection identifies data points that deviate from the mean by more than 3 standard deviations through the Z-score method; missing values are repaired using the forward filling method combined with linear interpolation; signal smoothing uses a sliding median filter to remove high-frequency noise. The preprocessed data set is denoted as: ; where Denote the preprocessed dataset, , , respectively denote the preprocessed data vectors at the 1st, 2nd, th sampling moments, and

[0023] Step 1.3, Time-scale decomposition; Apply wavelet transform to perform multi-scale decomposition on the preprocessed time series data, and decompose the signal into a linear combination of different frequency components: ; where denotes the original time series signal, denotes the short-term dynamic component (seconds to minutes), capturing transient responses and rapid fluctuations; denotes the medium-term dynamic component (hours to days), reflecting daily variations and charge-discharge cycles; denotes the long-term dynamic component (weeks to months), reflecting aging trends and seasonal variations.

[0024] The decomposition process uses discrete wavelet transform, and the formula is as follows: ; where is the wavelet coefficient, representing the result of wavelet transform at scale and translation ; is the scale parameter, controlling the stretching of the wavelet; is the translation parameter, controlling the position of the wavelet; represents the integration over the entire time domain; represents the complex conjugate of the wavelet basis function ; is the time series signal to be decomposed.

[0025] By selecting the wavelet coefficients corresponding to different scale parameters, signal components of different time scales are reconstructed.

[0026] Step 1.4, Feature extraction; Extract features from each time-scale component to construct a feature vector. Statistical features such as mean, standard deviation, peak value, and change rate are mainly extracted from the short-term component; features such as periodic characteristics, trend slope, and fluctuation amplitude are extracted from the medium-term component; degradation features such as capacity attenuation rate and impedance growth rate are mainly extracted from the long-term component. Finally, three feature vector sets of different time scales are obtained: ; ; ; wherein represents the short - term feature vector set, , , respectively represent the short - term feature vectors at the 1st, 2nd, and th sampling times; represents the medium - term feature vector set, , , respectively represent the medium - term feature vectors at the 1st, 2nd, and th sampling times; represents the long - term feature vector set, , , respectively represent the long - term feature vectors at the 1st, 2nd, and th sampling times, is the total number of samplings.

[0027] These feature vectors are used as the input for subsequent prediction and control models.

[0028] Step 2: Based on the time - scale components, construct a probabilistic state - transition model to probabilistically predict the future state of the battery pack and output the state prediction distribution; Based on the multi - time - scale feature vectors obtained in Step 1, this step constructs a probabilistic state - transition model and uses the Bayesian inference framework to probabilistically predict the future state of the battery pack, thereby quantifying the uncertainty of the prediction results. The short - term, medium - term, and long - term data components decomposed in Step 1 are directly used as the input for the probabilistic prediction model in this step to ensure that the model can fully capture the dynamic characteristics of different time scales.

[0029] Specifically, it includes the following sub - steps: Step 2.1: Construction of the probabilistic state - transition model; Construct a probabilistic state - transition model to describe the evolution process of the battery - pack state from time to time: ; wherein represents the conditional probability distribution function, and respectively represent time and time state vectors (including parameters such as the voltage, SOC, and temperature of each single - cell battery), represents time balancing action vector (including the magnitude and direction of the current in each balancing channel); The model is implemented based on the Variational Autoencoder (VAE) architecture and includes an encoder network and a decoder network , where is the latent state variable, and are network parameters; The loss function of the variational autoencoder is: ; where is the loss function of the variational autoencoder, and represent the encoder and decoder network parameters respectively, and represent the state vectors at time and represents the equilibrium action vector at time is the reconstruction error, represents the encoder network, represents the decoder network, is the latent state variable, is the KL divergence regularization term; represents the prior distribution of the latent variable ; is the trade-off parameter that controls the structure of the latent space.

[0030] In a specific implementation, the VAE model in the present invention adopts the following structure: The encoder network consists of three fully-connected neural networks. The number of nodes in the input layer is the same as the dimension of the battery state vector. The middle layers have 128 and 64 neurons respectively, and the LeakyReLU activation function is used. The output layer has two parts, the mean and the log variance, which are used to parameterize the distribution of the latent variable . The decoder network also uses three fully-connected networks, takes the latent variable and the control action as inputs, processes them through the middle layers with 64 and 128 neurons, and finally outputs the predicted state at the next moment.

[0031] An application example of this model in an electric vehicle's power battery pack is as follows: For a battery pack composed of 96 lithium-ion battery cells, where the state of each single cell includes four parameters: voltage, SOC, surface temperature, and internal resistance, and the control actions include the current values of 48 equalization channels. Under high-speed driving conditions, the model receives the states and control actions of all current battery cells as inputs and predicts the battery state distribution 5 minutes later. The prediction results not only include the expected state values of each battery but also the uncertainties of each parameter. For example, the predicted SOC value of battery #32 is 85% ± 2.1% (95% confidence interval), enabling the system to adjust the equalization strategy based on the prediction uncertainties and ensuring the avoidance of excessive equalization operations when the prediction accuracy is insufficient.

[0032] Step 2.2, Bayesian parameter learning; Use the Bayesian learning method to estimate the model parameters, using the prior distribution and , and calculate the posterior distribution of the parameters through the observed data . Since the posterior distribution is difficult to calculate directly, use variational inference for approximate solution: ; where represents the approximate posterior distribution determined by the parameters , represents the posterior distribution of the parameters given the observed data; Solve for the optimal parameters by minimizing the KL divergence: ; where represents the KL divergence (Kullback-Leibler divergence); represents the parameters of the variational distribution, which is used to measure the difference between two probability distributions; In actual implementation, use the Monte Carlo variational inference (MCVI) algorithm to sample multiple groups of parameters to jointly form an ensemble model: ; where represents the groups of parameters sampled by the MCVI algorithm, where represents the index of the parameter group, represents the total number of sampled parameter groups; and respectively represent the encoder and decoder network parameters of the th group of samples.

[0033] Step 2.3, multi-step probability prediction; Based on the learned probabilistic state transition model, perform a rolling prediction of the battery pack state for the next steps. Given the current state and the planned sequence of balancing actions: ; where , , respectively represent the balancing action vectors at time, time, time, represents the predicted number of time steps; Predict the state distribution at each future time: ; ; ; where represents the conditional probability distribution of the state at time and action ; represents the conditional probability distribution of the state at time given the state at time and action ; represents the conditional probability distribution of the state at time given the state at time and action ; represents the conditional probability distribution of the state at time given the initial state and the action sequence ; represents the integral operation; Since the integral calculation is complex, the particle filter algorithm is used for approximate solution, and particles are sampled for each time step: ; where , , , respectively represent the 1st, 2nd, and at A particle sample, indicating the number of particles used at each time step in the particle filter algorithm.

[0034] Step 2.4, Prediction uncertainty quantification; Quantify the uncertainty of the prediction results, including epistemic uncertainty (model parameter uncertainty) and stochastic uncertainty (inherent randomness of the system).

[0035] For the predicted state of the th single cell, calculate its 95% confidence interval and prediction variance to evaluate the reliability of the prediction results and guide subsequent risk perception decisions; The confidence interval and prediction variance are obtained through sampling statistics of multiple groups of parameters and multiple particles.

[0036] Wherein represents the index of the single cell, represents the time step index, indicating the th future time step; represents the th predicted state of the single cell at time; represents the 95% confidence interval of this predicted state, where is the lower bound, is the upper bound; represents the variance of the predicted state, used to quantify the magnitude of the prediction uncertainty.

[0037] Epistemic uncertainty refers to the prediction uncertainty caused by inaccurate model parameter estimation, and stochastic uncertainty refers to the prediction uncertainty caused by the randomness of the system itself.

[0038] Step 3, Based on the time scale component and the state prediction distribution, construct three parallel prediction controllers: a short-term controller, a medium-term controller, and a long-term controller; Based on the multi-time scale data processing results in Step 1 and the probabilistic state prediction output in Step 2, this step constructs three parallel prediction controllers, corresponding to short-term, medium-term, and long-term equilibrium tasks respectively, to achieve full-time domain equilibrium control. The prediction uncertainty quantification results generated in Step 2 provide an important decision-making basis for the controllers in this step, enabling the control strategy to adapt to prediction results at different confidence levels.

[0039] Specifically, it includes the following sub-steps: Step 3.1, Short-term controller construction; Construct a short-term prediction controller , the prediction range is from seconds to minutes, mainly responsible for dealing with the transient response and rapid fluctuations of the battery pack. The short-term controller is based on the Model Predictive Control (MPC) framework and uses a linearized state-space model: ; ; where denotes the short-term state vector at time is the short-term state vector, is the short-term control action, is the output observation, and are the process noise and measurement noise respectively, , and denote the state transition matrix, control input matrix and output matrix respectively.

[0040] The objective function of the short-term controller is: ; where denotes the objective function of the short-term controller, is the short-term prediction horizon length (e.g., 60 seconds), denotes the L2 norm with weight , denotes the L2 norm with weight , is the reference trajectory, and are the weight matrices of the output tracking error and control cost respectively, balancing the tracking accuracy and control cost; denotes the output predicted at time for time ; denotes the reference trajectory at time denotes the control action at time calculated at time ;

[0041] The optimal control sequence is obtained by solving the optimization problem: ; where denotes solving the optimization problem of minimizing the objective function , the optimization variable is the control sequence , and the receding horizon optimization strategy is applied, only executing the current step control action: ; where Indicates that the actually executed control action is equal to the control action at the current moment obtained by optimization.

[0042] Step 3.2, construction of the medium-term controller; Construct a medium-term predictive controller , with a prediction range from hours to days, mainly responsible for optimizing daily balancing efficiency and charge-discharge cycle management.

[0043] The medium-term controller adopts a non-linear model predictive control architecture and uses a neural network state transition model: ; where represents the medium-term state vector at time is the neural network model, is the network parameter, is the medium-term state vector, is the medium-term control action, is the medium-term process noise.

[0044] The objective function of the medium-term controller is: ; where represents the objective function of the medium-term controller, is the medium-term prediction horizon length (e.g., 24 hours); represents the L2 norm with weight , represents the L2 norm with weight , represents the L2 norm with weight ; represents the state predicted at time at time , represents the control action calculated at time at time ; is the reference state, represents the weight matrix of the state tracking error, represents the weight matrix of the control cost, represents the terminal cost matrix; represents the state predicted at time at time , that is, the state prediction value at the end of the prediction horizon.

[0045] Solve the non-linear optimization problem through numerical optimization methods to obtain the medium-term control action .

[0046] Neural Network State Transition Model in the Present Invention Adopts an architecture that combines a bidirectional long short-term memory network (Bi-LSTM) with residual connections. The model structure includes: an input layer that receives the medium-term state vector and control actions ; the feature extraction layer consists of two layers of Bi-LSTM, each layer containing 128 hidden units, which can capture both forward and backward temporal dependencies simultaneously; the residual connection layer directly connects the input features to the deep features, alleviating the vanishing gradient problem and improving training stability; the output layer uses a fully connected network to predict the state at the next moment . This network is trained through a sequence-to-sequence (seq2seq) framework, and the teacher forcing strategy is used to improve the long-term prediction accuracy.

[0047] An application example of this neural network model in the scenario of a fixed energy storage power station is as follows: For a cascade utilization energy storage system composed of 500 battery units, in the application of energy storage supporting photovoltaic power generation, the system needs to predict the state changes of the battery pack in the next 24 hours. The model inputs include the medium-term state characteristics (SOC mean and variance, capacity attenuation trend, daily cycle depth, etc.) of each current battery unit and the planned charge and discharge curve. The model can predict the state parameters of the battery pack at each time period within the next 24 hours. For example, after predicting 12 hours in the future, the SOC difference of each battery pack in the system will expand to 8.3%. At this time, it is advisable to arrange a 2-hour active equalization process to control the difference within 5%. Based on these prediction results, the medium-term controller generates the equalization plan for the next day, arranges the equalization operation during the low grid load period, and at the same time avoids the peak photovoltaic power generation period, which not only ensures the equalization effect but also maximizes the economic benefits of the energy storage system.

[0048] Step 3.3, construction of the long-term controller; Construct a long-term prediction controller , with a prediction range from weekly to monthly level, mainly responsible for maintaining the consistency of the battery pack health state and extending the cycle life; The long-term controller is based on the health state model and the degradation model, combined with a periodic optimization strategy: ; Wherein represents the health state of the th battery at the moment, represents the health state of the th battery at the moment, is the attenuation coefficient, is a degradation function, indicating the th battery's state of charge at time, indicating the th battery's depth of discharge at time, indicating the th battery's charge and discharge current at time, indicating the th battery's temperature at time.

[0049] The objective function of the long-term controller is: ; where represents the objective function of the long-term controller, represents the expected value, is the long-term prediction horizon length (e.g., 30 days), is the discount factor, indicates at time the health state vector of all batteries, is the average health state of the battery pack; represents the L2 norm, i.e., the Euclidean distance; represents the weight coefficient of the weakest battery's health state, indicating the minimum value of the health states among all batteries at

[0050] The long-term controller outputs a control strategy to guide the parameter adjustment of the medium-term and short-term controllers.

[0051] Step 3.4, hierarchical control coordination; Implement the coordination mechanism of the three time-scale controllers. The upper-layer controller provides constraint conditions and target references for the lower layer, and the lower-layer controller feeds back the execution results to the upper layer. Specifically: The long-term controller provides the health state target and the SOC operating range constraint for the medium-term controller ; The medium-term controller provides the reference trajectory and the control constraint for the short-term controller .

[0052] Through this hierarchical coordination mechanism, the unification of control objectives at different time scales is achieved, ensuring that short-term responses do not compromise long-term health goals.

[0053] Among them represents the SOC operating range constraint, where is the minimum SOC limit, is the maximum SOC limit; represents the constraint range of short-term control actions, where is the minimum control action limit, is the maximum control action limit.

[0054] Step 4: According to the state prediction distribution and the output of the predictive controller, expand the traditional fuzzy rules into a probabilistic form, and quantify the decision-making uncertainty through Bayesian fuzzy inference; After obtaining the probabilistic state prediction results in Step 2 and the multi-level control output in Step 3, this step expands the traditional fuzzy inference system into a Bayesian fuzzy inference framework, realizes the accurate quantification of the balanced decision-making uncertainty, and improves the decision-making reliability of the system under data uncertainty conditions. The preliminary control decisions generated by each controller in Step 3 are further optimized through Bayesian fuzzy inference in this step, enhancing the adaptability and robustness of the decision-making.

[0055] Specifically, it includes the following sub-steps: Step 4.1: Construction of Bayesian fuzzy rules; Construct a Bayesian fuzzy rule base, and each Bayesian fuzzy rule is expressed in a probabilistic form: ; Among them represents the th Bayesian fuzzy rule, is the input variable, , , respectively represent the input fuzzy set, output fuzzy set, and credibility of the th Bayesian fuzzy rule; is the output variable; and are the condition and conclusion keywords in the fuzzy rule, represents the membership relationship.

[0056] According to the expert knowledge and historical data of battery equalization management, construct a rule base containing the following types of rules: Voltage difference rule: Determine the equalization direction and intensity based on the voltage difference between batteries; SOC equalization rule: Guide the equalization strategy based on the SOC distribution characteristics of the batteries; Temperature safety rule: Adjust or limit the equalization current according to the temperature state Health state protection rules: adopt a differential equalization strategy based on the difference in battery health state; Each type of rule contains multiple sub - rules, and the total number of rules is between 50 and 100, covering various possible combinations of battery states.

[0057] Step 4.2, fuzzy set parameter learning; Use the Bayesian learning method to estimate the fuzzy set parameters, parameterize the fuzzy set and into parameter vectors and , and learn the posterior distribution of the parameters through historical data . .

[0058] Specifically, for the Gaussian fuzzy membership function: ; where represents the membership degree of the input to the fuzzy set , represents the standard deviation parameter of the Gaussian fuzzy membership function, represents the natural exponential function, represents the mean parameter of the Gaussian fuzzy membership function; represents the parameter vector of the fuzzy set , expressed as: ; Use the Markov chain Monte Carlo method to sample multiple groups of parameters from the posterior distribution , forming a parameter set.

[0059] where represents the posterior distribution of the fuzzy set parameters given the historical data , represents the groups of parameters sampled by the MCMC method, where represents the index of the parameter group, represents the total number of sampled parameter groups.

[0060] Step 4.3, probabilistic fuzzy inference; Based on the Bayesian fuzzy rules and parameter distribution, perform probabilistic fuzzy inference. Given the input , calculate the fuzzy inference result respectively according to each group of parameters , obtaining the probability distribution of the output instead of a single definite value. The inference process uses the Mamdani fuzzy inference model, including the following steps: Fuzzification, calculate the input For each input fuzzy set Membership degree ; Rule triggering, calculate the triggering strength of each rule : ; Where Represents the membership degree of the input To the fuzzy set , Represents the triggering strength of the rule , Represents the credibility of the rule ; Fuzzy inference, calculate the output fuzzy set of each rule: ; Where Represents the output fuzzy set after the rule Inference, Represents the minimum value operation, Represents the output value To the fuzzy set Membership degree; Fuzzy aggregation, merge the output fuzzy sets of all rules: ; Where Represents the aggregated output fuzzy set, Represents taking the maximum value over all rules ; Defuzzification, calculate the crisp output value using the centroid method: ; Where Represents the crisp output value after defuzzification, Represents the integral operation, Represents the weighted average calculated using the centroid method; By repeating the above inference process for multiple groups of parameters, the output probability distribution is obtained ; Step 4.4, inference uncertainty quantification; Based on the probability fuzzy inference result, quantify the uncertainty of the decision. For the inference output distribution , calculate the following uncertainty indicators: The mean value is used as the final decision value: ; Where Represents the mean value of the output distribution, represents the expected value of the output under the given input , which represents the integration of the product of the output value and its probability density; The variance quantifies the decision uncertainty: ; where represents the variance of the inference result, represents the variance of the output under the given input , which represents the square of the difference between the output value and the mean; The 95% confidence interval represents the decision range: ; where represents the 95% confidence interval of the output, where is the lower bound, is the upper bound; The entropy represents the decision information uncertainty: ; where represents the entropy of the output under the given input , represents the natural logarithm, represents the integration operation, represents the probability distribution of the output; These uncertainty metrics are passed to the subsequent risk-aware decision-making module to guide the formulation of the balancing control strategy.

[0061] An example of the application of the Bayesian fuzzy inference system in the balancing management of the electric bus battery pack is as follows: In an electric bus using a 300 kWh lithium battery pack, when the voltage difference between the 3rd and 7th battery modules is detected to reach 150 mV, the traditional deterministic fuzzy system will directly output a balancing current of 500 mA. However, the Bayesian fuzzy inference system of the present invention takes into account the ±2 °C fluctuation in the current battery temperature sensor readings and the uncertainty in the SOC estimation, and generates a probability distribution of the balancing current, with a mean of 480 mA, a 95% confidence interval of [420 mA, 540 mA], and a variance of 18 mA 2, the entropy value is 3.2. Further analysis by the system found that the uncertainty of this decision mainly stems from the unstable readings of the temperature sensor. Therefore, the execution intensity was automatically adjusted to 0.85, and the final output balanced current was 408 mA, avoiding the risk of excessive balancing caused by information uncertainty. At the same time, the system marked that the temperature sensor may be faulty and reminded the maintenance personnel to focus on inspection during the next maintenance, improving the maintainability of the system.

[0062] Step 5: Based on the state prediction distribution, the output of the prediction controller, and the Bayesian fuzzy inference result, construct a risk-adjusted objective function based on conditional value at risk, and generate an equilibrium control decision and its confidence interval; Combining the probability prediction result in Step 2, the output of the controller in Step 3, and the Bayesian fuzzy inference result in Step 4, in this step, a risk-adjusted objective function is constructed based on the theory of Conditional Value at Risk (CVaR) to achieve precise measurement and trade-off of the equilibrium decision risk and ensure the safety of the system in extreme situations. The decision uncertainty index output in Step 4 provides a key input for the risk assessment in this step, making the risk measurement more accurate and reliable.

[0063] Specifically, it includes the following sub-steps: Step 5.1: Construction of risk measurement index; Form a risk measurement index based on conditional value at risk, defined as follows: ; where represents the conditional value at risk at the confidence level , is the random loss variable; represents the expected loss under the condition that the loss variable is greater than or equal to the value at risk ; is the value at risk at the confidence level , indicating that the probability of the loss exceeding this value is not greater than , that is ; In battery equalization management, the loss function is defined as: ; where represents the total loss function of taking action in the state , , , respectively represent the unbalance loss function, the safety risk loss function, and the energy efficiency loss function, , and They are the loss weight coefficients of the imbalance loss function, the safety risk loss function, and the energy efficiency loss function, respectively.

[0064] Step 5.2, construction of the risk-adjusted objective function; A risk-adjusted objective function based on CVaR risk measurement is formed: ; where represents the risk-adjusted objective function, represents the expected value of the reward function of, is the balanced reward function (i.e., the negative loss function ); is the risk aversion coefficient, controlling the trade-off between expected return and extreme risk; is the confidence level (usually taken as 0.95 or 0.99); represents at the confidence level the conditional value at risk of the negative reward, represents the negative reward function, equivalent to the loss function .

[0065] The first term of the objective function maximizes the expected return, and the second term minimizes the extreme risk. The two are weighted and balanced to obtain the optimal strategy.

[0066] Step 5.3, scenario sampling and risk assessment; The Monte Carlo method is used to evaluate the decision risk and sample the possible future scenarios. The specific steps are as follows: Generate future state trajectory samples from the probabilistic state transition model: ; where represents the trajectory from time step 1 to time step N, containing N consecutive time steps; , , respectively represent the 1st, 2nd, and th trajectory samples, represents the total number of sampled trajectories.

[0067] For each sampled trajectory, calculate the reward under the control action ; Sort the reward samples to determine the VaR threshold , that is, satisfy: ; where represents at the confidence level The value-at-risk threshold below represents the probability measure indicating the confidence level of the risk assessment; In actual calculations, corresponding to the sorted position at the reward sample value; calculate the CVaR risk measure: ; where represents the conditional value-at-risk of negative rewards at the confidence level below, represents the summation of negative reward samples sorted within the range from to , represents the th sorted reward sample, represents the normalization coefficient when calculating CVaR.

[0068] Step 5.4, optimize and solve the risk-aware decision; Based on the risk-adjusted objective function, solve for the optimal equilibrium decision through numerical optimization methods: ; where represents the optimal control action at time , represents the action that minimizes the objective function ; ; Specifically, adopt the following optimization strategies: For the discrete action space, use the sampling-based optimization method, the Cross-Entropy Method (CEM); For the continuous action space, use gradient descent or conjugate gradient method; Consider action constraints during the optimization process; Finally, obtain the risk-optimal equilibrium control action .

[0069] Step 5.5, calculate the decision confidence interval; Based on the risk optimization result, calculate the confidence interval of the decision, reflecting the credibility of the decision; The confidence interval is obtained through the following steps: Estimate the decision distribution using parametric or non-parametric methods ; Calculate the decision mean as the final decision value; Based on the distribution, calculate the confidence interval (usually ); The confidence interval information will be passed to the adaptive integration module for adjusting the decision execution intensity.

[0070] Among them represents the confidence interval of the decision, is the lower bound, is the upper bound, represents the probability distribution of the action under the condition of the state , represents the mean value of the decision distribution and serves as the final decision value, represents the confidence level of the confidence interval, usually taking 0.95, indicating a 95% confidence level.

[0071] Step 6: Based on the output of the predictive controller, the results of Bayesian fuzzy inference, and the balanced control decision, and adjust the execution intensity according to the confidence interval, and output the balanced current command; After completing the multi-level control calculation in Step 3, the Bayesian fuzzy inference in Step 4, and the risk perception decision in Step 5, this step realizes the adaptive fusion of the outputs of the three time-scale controllers, dynamically adjusts the control weights according to the current system state and uncertainty level, and generates the final balanced control command. The decision confidence interval generated in Step 5 is directly used for adjusting the execution intensity in this step to ensure that the system can adopt a more conservative control strategy when the uncertainty is high.

[0072] Specifically, it includes the following sub-steps: Step 6.1: Dynamic weight calculation; Calculate the dynamic fusion weights , and of the outputs of the three time-scale controllers; The weight calculation is based on the following factors: Current system state characteristics: such as voltage imbalance degree, SOC distribution, temperature change rate, etc.; Prediction uncertainty level: the confidence interval width of the prediction results of each time scale; Historical control effect evaluation: the recent performance and adaptability of each controller; The weight calculation adopts an adaptive weighting algorithm: ; Among them , and respectively represent the fusion weights of the short-term controller, the medium-term controller, and the long-term controller at time , represents the fusion weight of the controller at time , where they represent the short-term, medium-term, and long-term controllers respectively, denotes the natural exponential function, and are the weight characteristics of controller and controller respectively, denotes the summation symbol, and is calculated by combining the above three factors: ; where denotes the weight characteristic value of controller at time respectively, , and are the state characteristic function, the uncertainty evaluation function, and the historical effect evaluation function respectively, , and are the fusion balance coefficients of the state characteristic function, the uncertainty evaluation function, and the historical effect evaluation function respectively.

[0073] Step 6.2, control output fusion; Based on the calculated dynamic weights, fuse the outputs of the three time-scale controllers to generate the final balanced control action: ; where denotes the final fusion control action at time respectively, , and are the optimal control outputs of the short-term, medium-term, and long-term controllers respectively.

[0074] The fusion process ensures a continuous and smooth transition of the control action, avoiding system oscillations caused by mutations.

[0075] Step 6.3, execution intensity adjustment; According to the width of the decision confidence interval , adaptively adjust the execution intensity of the balanced control action. When the uncertainty is high (the confidence interval is wide), reduce the execution intensity to reduce risks; when the uncertainty is low (the confidence interval is narrow), maintain the original execution intensity. The execution intensity adjustment function is: ; where denotes the execution intensity adjustment coefficient at time , is the basic execution intensity (usually 1), is the execution intensity attenuation coefficient, is the control action normalization factor, denotes the natural exponential function, and are respectively the lower bound and the upper bound of the confidence interval of the control action.

[0076] The final actual executed control action is: ; where represents the actual executed control action at time .

[0077] Step 6.4, generating the balancing circuit control signal; Convert the fused balancing control action into specific balancing circuit control signals, including parameters such as the balancing direction, the magnitude of the balancing current, and the balancing time. The control signal is transmitted to the balancing circuit drive module of the Battery Management System (BMS) through the communication interface to realize the control of the actual balancing hardware.

[0078] The generation of the control signal takes into account hardware limitations and safety constraints to ensure that the balancing current does not exceed the maximum value allowed by the system, and at the same time avoids efficiency losses caused by frequent switching. Through Pulse-Width Modulation (PWM) or other suitable modulation methods, the continuous control signal is converted into a discrete switching control sequence.

[0079] A battery pack active balancing management system for implementing the above-mentioned battery pack active balancing management method, including: A multi-time scale data processing module for collecting battery pack parameter data and performing time scale decomposition to decompose the data into dynamic components of different frequencies; A probability state prediction module for constructing a state transition probability model to realize multi-step prediction of the future state of the battery pack and quantify the prediction uncertainty; A multi-level parallel control module for constructing short, medium, and long-term prediction controllers based on different time scales to collaboratively manage the battery pack balancing process; A Bayesian fuzzy inference module for probabilistically expanding fuzzy rules and quantifying decision-making uncertainty through parameter learning and probabilistic inference; A risk-aware decision-making module for constructing risk measurement indicators and a risk-adjusted objective function to generate a balancing control strategy with a confidence interval; An adaptive integration execution module for fusing the outputs of each controller and adjusting the execution intensity according to the uncertainty level to generate the final balancing control signal.

[0080] Here, the present invention provides an implementation example: This implementation mode has been verified in an electric logistics vehicle of a new energy commercial vehicle manufacturer. This vehicle model uses CATL's lithium titanate battery pack, which contains 16 battery modules. Each module is composed of 12 single cells connected in series, for a total of 192 single cells. The battery management system is equipped with an active equalization circuit that can adjust the equalization current by ±2A, and is also equipped with an embedded computing platform to execute the algorithm of this implementation mode.

[0081] The main operation scenario of the logistics vehicle is urban express delivery, working 12 hours a day, with characteristics such as frequent starts and stops, large load fluctuations, and obvious differences in ambient temperature. Traditional equalization management methods face multiple challenges in such scenarios: First, the frequent acceleration and deceleration make the battery load highly dynamic, making it difficult to perform accurate equalization control; Second, the all-weather operation causes uneven battery temperature distribution, exacerbating the battery consistency problem; Third, the working conditions vary greatly on different routes and in different seasons, and it is difficult for the equalization strategy to be uniformly applicable; Finally, the quality of sensor data is affected by factors such as vehicle vibration, resulting in uncertainty.

[0082] The system collects complete battery data every 1 second, including the voltage of each single cell (3.5V - 2.2V), temperature (0 - 60°C), bus current (-150A to +150A), and estimated SOC (0 - 100%), etc. The collected data is decomposed into components of three time scales through wavelet transform. An example of the voltage components of a certain battery single cell at different time scales on a typical working day is shown in Table 1: Table 1: Multi-time scale voltage components of battery single cell #45 on a typical working day

[0083] The system predicts the state of the battery pack based on the variational autoencoder model and quantifies the prediction uncertainty at the same time. During a delivery task after the express vehicle departs from the warehouse, the system predicts the SOC of the battery module. The prediction includes the mean value and the 95% confidence interval, reflecting the uncertainty level of the prediction at different time points.

[0084] The system generates an equalization control strategy hierarchically according to the prediction results. The outputs of the three different time scale controllers and the fused results during a complete operation cycle are shown in Table 2: Table 2: Outputs of different time scale controllers and fused results (equalization current, unit: mA)

[0085] The Bayesian fuzzy inference results of the system for some batteries in the battery pack with abnormal temperature rise are shown in Table 3: Table 3: Bayesian fuzzy inference results in case of temperature anomaly (equalization current, unit: mA)

[0086] This table compares the decision-making differences between the traditional deterministic fuzzy system and the Bayesian fuzzy system of this embodiment under the same circumstances.

[0087] The system generates risk-aware decisions based on conditional value at risk and adaptively adjusts the execution intensity according to the decision confidence. The decision-making results under different risk conditions are shown in Table 4: Table 4: Decisions and Execution Results under Different Risk Conditions

[0088] After applying this embodiment to the above electric logistics vehicle for 6 months and comparing it with the original balanced management system, the following two key technical effects are mainly verified: Full-time-domain balanced management effect: The comparison of the balanced effects between this embodiment and the traditional single-time-scale balancing method under different working conditions is shown in Table 5: Table 5: Comparison of Balanced Effects between this Embodiment and Traditional Methods

[0089] Uncertainty and risk management effect: The comparison of the robustness and safety between this embodiment and the traditional deterministic balancing method under the condition of data uncertainty or sensor failure is shown in Table 6: Table 6: Comparison of Uncertainty and Risk Management Effects

[0090] From the above data, it can be seen that this embodiment has achieved significant technical effects compared with the traditional method: in terms of full-time-domain balanced management, the balanced convergence time is reduced by 51.1%, and the state consistency is increased by about 67%; in terms of uncertainty and risk management, the system can still maintain an equilibrium efficiency of 60%-75% under abnormal conditions, and the incidence of safety events is reduced by 76.2%. These results fully verify the technical advantages and innovative value of the present invention in the field of active balanced management of battery packs.

[0091] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An active equalization management method for a battery pack, characterized in that, It includes the following steps: Collect the parameter data of the battery pack, and decompose the system dynamics into three time-scale components: short-term dynamics, medium-term dynamics, and long-term dynamics through the time-scale decomposition algorithm; Based on the time-scale components, construct a probabilistic state transition model to probabilistically predict the future state of the battery pack and output the state prediction distribution; Based on the time-scale components and the state prediction distribution, construct three parallel predictive controllers: a short-term controller, a medium-term controller, and a long-term controller; According to the state prediction distribution and the output of the predictive controller, extend the traditional fuzzy rules to a probabilistic form, and quantify the decision-making uncertainty through Bayesian fuzzy inference; According to the state prediction distribution, the output of the predictive controller, and the results of Bayesian fuzzy inference, construct a risk-adjusted objective function based on conditional value at risk, and generate an equilibrium control decision and its confidence interval; According to the output of the predictive controller, the results of Bayesian fuzzy inference, and the equilibrium control decision, and adjust the execution intensity according to the confidence interval, and output the equilibrium current command.

2. The active balancing management method for the battery pack according to claim 1, wherein The collection of the parameter data of the battery pack includes: collecting the parameter data of the battery pack, including voltage, current, temperature, and internal resistance, where the voltage and current signals are sampled at a high frequency, the temperature is sampled at a medium frequency, and the internal resistance data is sampled at a low frequency.

3. The active equalization management method for the battery pack according to claim 1, wherein The time-scale decomposition algorithm includes: applying wavelet transform to perform multi-scale decomposition on the preprocessed time-series data, where the short-term dynamic component is from seconds to minutes, capturing transient responses and rapid fluctuations; the medium-term dynamic component is from hours to days, reflecting daily variations and charge-discharge cycles; the long-term dynamic component is from weeks to months, reflecting aging trends and seasonal variations.

4. The active balancing management method for the battery pack according to claim 1, wherein The probabilistic state transition model is implemented based on the variational autoencoder architecture, including an encoder network and a decoder network, estimating the posterior distribution through variational inference, estimating the model parameters using the Bayesian learning method, and performing multi-step probabilistic prediction using the particle filter algorithm.

5. The battery pack active balancing management method according to claim 1, wherein The short-term controller is based on the model predictive control framework, uses a linearized state-space model, and the prediction range is from seconds to minutes; the medium-term controller adopts a non-linear model predictive control architecture, uses a neural network state transition model, and the prediction range is from hours to days; The long-term controller is based on the health state model and the degradation model, combined with a periodic optimization strategy, and the prediction range is from weeks to months.

6. The battery pack active equalization management method according to claim 5, wherein The neural network state transition model adopts an architecture combining a bidirectional long short-term memory network and a residual connection. The model structure includes: The input layer receives the medium-term state vector and the control action; The feature extraction layer consists of two layers of bidirectional long short-term memory networks; The residual connection layer directly connects the input features to the deep features; The output layer uses a fully connected network to predict the next state.

7. The active balancing management method for a battery pack according to claim 1, wherein The Bayesian fuzzy inference includes: Construct a Bayesian fuzzy rule base, and each Bayesian fuzzy rule is expressed in a probabilistic form; Use the Bayesian learning method to estimate the fuzzy set parameters; Perform probabilistic fuzzy inference based on the Bayesian fuzzy rules and the parameter distribution; Calculate the mean, variance, confidence interval, and entropy of the inference output distribution to quantify the decision-making uncertainty.

8. The active equalization management method for a battery pack according to claim 1, characterized in that In the construction of the risk-adjusted objective function based on conditional value at risk, the risk-adjusted objective function includes maximizing the expected return term and minimizing the extreme risk term.

9. The active equalization management method for a battery pack according to claim 1, wherein The execution intensity adjustment function for adjusting the execution intensity is as follows: ； wherein represents the execution intensity adjustment coefficient at the moment , is the basic execution intensity represents the natural exponential function is the execution intensity attenuation coefficient is the control action standardization factor and respectively represent the lower bound and the upper bound of the confidence interval of the control action 10. A battery pack active balancing management system, characterized in that, To execute an active equalization management method for a battery pack according to any one of claims 1-9, comprising: A multi-time scale data processing module, configured to collect battery pack parameter data and perform time scale decomposition to decompose the data into dynamic components of different frequencies; A probability state prediction module, configured to construct a state transition probability model to achieve multi-step prediction of the future state of the battery pack and quantify prediction uncertainty; A multi-level parallel control module, configured to construct short-term, medium-term, and long-term prediction controllers based on different time scales to collaboratively manage the equalization process of the battery pack; A Bayesian fuzzy inference module, configured to perform probability extension on fuzzy rules and quantify decision uncertainty through parameter learning and probability inference; A risk-aware decision-making module, configured to construct a risk metric index and a risk-adjusted objective function to generate an equalization control strategy with a confidence interval; An adaptive integration execution module, configured to fuse the outputs of each controller and adjust the execution intensity according to the uncertainty level to generate a final equalization control signal.

Images