A battery pack active balancing management method and system

Abstract

The present invention relates to the technical field of battery management systems, and discloses a method and system for active balancing management of battery packs, wherein one method for active balancing management of battery packs comprises: collecting battery pack parameter data, and dynamically decomposing the system through a time scale decomposition algorithm; constructing a probabilistic state transition model, and performing probabilistic prediction of the future state of the battery pack; constructing three parallel predictive controllers; expanding traditional fuzzy rules into a probabilistic form, and quantifying decision uncertainty through Bayesian fuzzy reasoning; constructing a risk-adjusted objective function based on conditional risk value, and generating balancing control decisions and their confidence intervals; adjusting the execution intensity, and outputting balancing current instructions; and realizing comprehensive management of the battery pack from second-level transient response to monthly-level health evolution through multi-time scale data decomposition and a parallel predictive control architecture, thereby breaking the limitations of traditional single-time scale controllers.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and more particularly, to a method and system for active equalization management of battery packs. Background Art

[0002] Battery packs are widely used in modern applications such as electric vehicles, energy storage systems, and portable devices. In these applications, battery packs are typically composed of multiple single cells connected in series and parallel. Due to factors such as manufacturing process variations, uneven temperature distribution, and varying degrees of aging, parameters such as state of charge, state of health, and voltage levels of each single cell can become inconsistent, resulting in degraded battery pack performance, reduced capacity, and shortened service life.

[0003] To address this issue, battery management systems typically use balancing technology to balance differences between individual cells. Traditional balancing techniques primarily rely on passive balancing based on voltage or charge differences, such as resistive shunt balancing, capacitor shuttling balancing, and transformer coupling balancing. These methods rely heavily 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 research has begun to introduce predictive control into the field of battery balancing, establishing dynamic battery models to predict future conditions and optimize balancing strategies.

[0004] However, current battery balancing management still faces multiple technical challenges: First, traditional predictive control methods usually only consider a single time scale, making it difficult to simultaneously handle the short-term dynamic response and long-term health status evolution of the battery pack balancing process; second, most existing balancing decision models use deterministic methods, which cannot effectively quantify the uncertainty in the prediction and control process. Especially in abnormal situations such as sensor failure or data loss, the decision reliability is significantly reduced; third, most balancing systems lack effective risk assessment and management mechanisms, and are unable to adjust decision-making strategies based on uncertainty, which can easily lead to performance degradation 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, so as to improve the prediction accuracy, decision 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 to solve the technical problems in related technologies such as difficulty in simultaneously processing multi-time scale dynamics, inability to quantify decision uncertainty, and lack of risk-aware decision-making.

[0007] The present invention provides a battery pack active balancing management method, comprising the following steps:

[0008] Collect battery pack parameter data and decompose the system dynamics into three time scale components: short-term dynamics, medium-term dynamics, and long-term dynamics using a time scale decomposition algorithm. The time scale decomposition algorithm includes: applying wavelet transform to perform multi-scale decomposition on the pre-processed time series data, where the short-term dynamic component is at the second to minute level to capture transient responses and rapid fluctuations; the medium-term dynamic component is at the hour to day level to reflect diurnal changes and charge and discharge cycles; and the long-term dynamic component is at the week to month level to reflect aging trends and seasonal changes;

[0009] Based on the time scale component, a probabilistic state transition model is constructed to probabilistically predict the future state of the battery pack and output the state prediction distribution. The probabilistic state transition model is implemented based on the variational autoencoder architecture, including an encoder network and a decoder network. The posterior distribution is estimated through variational inference, the model parameters are estimated using the Bayesian learning method, and the particle filter algorithm is used for multi-step probabilistic prediction;

[0010] Based on the time scale component and the state prediction distribution, three parallel predictive controllers are constructed: short-term controller, medium-term controller and long-term controller;

[0011] Based on the state prediction distribution and the predicted controller output, the traditional fuzzy rules are expanded into a probabilistic form, and the decision uncertainty is quantified through Bayesian fuzzy reasoning, which includes:

[0012] Construct a Bayesian fuzzy rule library, where each Bayesian fuzzy rule is expressed in a probabilistic form;

[0013] Bayesian learning method is used to estimate fuzzy set parameters;

[0014] Probabilistic fuzzy reasoning based on Bayesian fuzzy rules and parameter distribution;

[0015] Calculate the mean, variance, confidence interval, and entropy of the inference output distribution to quantify decision uncertainty;

[0016] Based on the state prediction distribution, the predicted controller output and the Bayesian fuzzy inference results, a risk-adjusted objective function is constructed based on the conditional value at risk to generate the equilibrium control decision and its confidence interval;

[0017] According to the output of the predictive controller, the Bayesian fuzzy reasoning result and the balanced control decision, the execution intensity is adjusted according to the confidence interval, and the balanced current instruction is output. The execution intensity adjustment function for adjusting the execution intensity is:

[0018]

[0019] Where η(t) represents the execution intensity adjustment coefficient at time t, η base is the basic execution strength, exp(·) represents the natural exponential function, k decayTo implement the intensity attenuation coefficient, - u is the control action normalization factor, u lower and u upper They represent the lower and upper bounds of the confidence interval of the control action, respectively.

[0020] Furthermore, collecting battery pack parameter data includes: collecting battery pack parameter data, including voltage, current, temperature and internal resistance, wherein voltage and current signals are sampled at a higher frequency, temperature is sampled at a medium frequency, and internal resistance data is sampled at a low frequency.

[0021] Furthermore, the short-term controller is based on the model predictive control framework, using a linearized state space model with a prediction range of seconds to minutes; the medium-term controller adopts a nonlinear model predictive control architecture and uses a neural network state transition model with a prediction range of 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 of weeks to months.

[0022] Furthermore, the neural network state transfer model adopts an architecture combining a bidirectional long short-term memory network and a residual connection. The model structure includes:

[0023] The input layer receives the mid-term state vector and control actions;

[0024] The feature extraction layer consists of two layers of bidirectional long short-term memory networks;

[0025] The residual connection layer directly connects the input features to the deep features;

[0026] The output layer uses a fully connected network to predict the state at the next moment.

[0027] Furthermore, in the risk-adjusted objective function constructed based on the conditional value at risk, the risk-adjusted objective function includes maximizing the expected return term and minimizing the extreme risk term.

[0028] The present invention provides a battery pack active balancing management system for executing the above-mentioned battery pack active balancing management method, comprising:

[0029] Multi-time scale data processing module, used to collect battery pack parameter data and perform time scale decomposition, decomposing the data into dynamic components of different frequencies;

[0030] Probabilistic state prediction module, used to build a state transition probability model, achieve multi-step prediction of the battery pack's future state and quantify prediction uncertainty;

[0031] A multi-level parallel control module is used to build three predictive controllers for short, medium and long periods of time based on different time scales to collaboratively manage the battery pack balancing process;

[0032] Bayesian fuzzy reasoning module, which is used to probabilistically expand fuzzy rules and quantify decision uncertainty through parameter learning and probabilistic reasoning;

[0033] The risk-aware decision-making module is used to construct risk metrics and risk-adjusted objective functions, and generate equilibrium control strategies with confidence intervals;

[0034] The adaptive integrated execution module is used to fuse the outputs of each controller and adjust the execution intensity according to the uncertainty level to generate the final balanced control signal.

[0035] The beneficial effects of the present invention are:

[0036] Through multi-timescale data decomposition and a parallel predictive control architecture, comprehensive management of battery pack health evolution from seconds-level transient response to months-level health evolution is achieved, breaking the limitations of traditional single-timescale controllers.

[0037] Through the Bayesian fuzzy inference framework, the uncertainty of equilibrium decisions is accurately quantified, and probabilistic decisions with confidence intervals can be output, rather than single-point estimates from traditional methods. In abnormal situations such as data loss or sensor failure, the system can still maintain high equilibrium efficiency and improve robustness.

[0038] The risk-aware decision-making mechanism based on conditional value at risk can accurately measure system risks in extreme situations, reducing the probability of abnormal events while maintaining balanced efficiency and improving system security.

[0039] Through dynamic weight adjustment and execution strength adaptation, it can intelligently balance control objectives at different time scales based on the current state and uncertainty level, achieving true proactive balance management. Compared with fixed weight methods, it improves adaptability under variable working conditions.

[0040] Providing probability distribution and confidence intervals for each equilibrium decision makes the decision-making process more transparent and explainable, which not only facilitates human-machine collaboration and system supervision, but also improves fault diagnosis efficiency.

[0041] It has achieved a technological leap from "passive response" to "active prediction and risk perception" in active balancing management of battery packs, providing a new technical path for high-performance and high-reliability battery management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a battery pack active balancing management method in the present invention;

[0043] Figure 2 is a flow chart of the multi-time-scale data processing steps in the present invention;

[0044] Figure 3is a flow chart of the probability state prediction step in the present invention;

[0045] Figure 4 It is a flowchart of the multi-level parallel predictive control steps in the present invention;

[0046] Figure 5 It is a flow chart of the Bayesian fuzzy reasoning steps in the present invention;

[0047] Figure 6 is a flow chart of the risk perception decision generation steps in the present invention;

[0048] Figure 7 It is a flow chart of the adaptive integration execution steps in the present invention. DETAILED DESCRIPTION

[0049] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0050] At least one embodiment of the present invention discloses a battery pack active balancing management method, such as Figures 1 to 7 As shown, including:

[0051] Step 1: Collect battery pack parameter data and decompose the system dynamics into three time scale components: short-term dynamics, medium-term dynamics, and long-term dynamics through a time scale decomposition algorithm;

[0052] This step processes the battery pack monitoring data through a time scale decomposition algorithm, dynamically decomposing the battery status into components at different time scales, providing a basis for subsequent multi-level parallel predictive control.

[0053] It includes the following sub-steps:

[0054] Step 1.1, data collection;

[0055] 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 10Hz), temperature is sampled at a medium frequency (such as 1Hz), and internal resistance data is sampled at a low frequency (such as 0.1Hz). The collected data set is recorded as:

[0056] D=x (0) ,x (2) ,…,x (n)

[0057] Where D represents the original data set collected, x (1) 、x (2) 、x (n) They represent the data vectors at the 1st, 2nd, and nth sampling moments respectively, and n is the total number of samples.

[0058] Step 1.2, data preprocessing;

[0059] The collected raw data was preprocessed, including outlier detection, missing value filling, and signal smoothing. Outlier detection used the Z-score method to identify data points that deviated from the mean by more than 3 standard deviations; missing values ​​were repaired using a forward filling method combined with linear interpolation; and signal smoothing used a sliding median filter to remove high-frequency noise. The preprocessed dataset is recorded as:

[0060]

[0061] Among them D clean represents the preprocessed dataset, Represent the preprocessed data vectors at the 1st, 2nd, and nth sampling moments respectively, and n is the total number of samples.

[0062] Step 1.3, time scale decomposition;

[0063] Apply wavelet transform to perform multi-scale decomposition on the preprocessed time series data, and decompose the signal x(t) into a linear combination of different frequency components:

[0064] x(t)=x s (t)+x m (t)+x l (t);

[0065] Where x(t) represents the original time series signal, x s (t) represents the short-term dynamic component (seconds to minutes), capturing transient responses and rapid fluctuations; x m (t) represents the medium-term dynamic component (hour to day level), reflecting diurnal changes and charge-discharge cycles; x l (t) represents the long-term dynamic component (weekly to monthly level), reflecting aging trends and seasonal changes.

[0066] The decomposition process uses discrete wavelet transform, and the formula is as follows:

[0067]

[0068] Where W ψ (f)(a scale , b shift ) is the wavelet coefficient, indicating that at scale a scale and pan b shiftThe wavelet transform result under a scale is the scale parameter, which controls the expansion and contraction of the wavelet; b shift is the translation parameter, which controls the position of the wavelet; represents the integral over the entire time domain; ψ * represents the complex conjugate of the wavelet basis function ψ; f(t) is the time series signal to be decomposed.

[0069] By selecting the wavelet coefficients corresponding to different scale parameters, signal components of different time scales are reconstructed.

[0070] Step 1.4, feature extraction;

[0071] Features are extracted from each time scale component to construct a feature vector. The short-term component mainly extracts statistical features such as mean, standard deviation, peak value, and rate of change; the medium-term component extracts features such as periodic characteristics, trend slope, and fluctuation amplitude; and the long-term component mainly extracts degradation features such as capacity decay rate and impedance growth rate. Finally, the feature vector sets of three time scales are obtained:

[0072]

[0073]

[0074] Among them F s represents the short-term feature vector set, Represent the short-term feature vectors of the first, second, and nth sampling moments respectively; F m represents the mid-term feature vector set, Represent the mid-term feature vectors of the first, second, and nth sampling moments respectively; F l represents the long-term feature vector set, They represent the long-term feature vectors of the first, second, and nth sampling moments respectively, and n is the total number of samples.

[0075] These feature vectors serve as input to subsequent prediction and control models.

[0076] Step 2: Based on the time scale component, a probabilistic state transition model is constructed to probabilistically predict the future state of the battery pack and output the state prediction distribution;

[0077] Based on the multi-timescale feature vectors obtained in Step 1, this step constructs a probabilistic state transition model. Using a Bayesian inference framework, this model probabilistically predicts the battery pack's future state, thereby quantifying the uncertainty of the prediction results. The short-term, medium-term, and long-term data components decomposed in Step 1 serve directly as input to the probabilistic prediction model in this step, ensuring that the model fully captures the dynamic characteristics of different timescales.

[0078] It includes the following sub-steps:

[0079] Step 2.1, constructing the probability state transition model;

[0080] Construct a probabilistic state transition model to describe the evolution of the battery pack state from time t to time t+1:

[0081] p(s t+1 |s t ,a t );

[0082] Where p represents the conditional probability distribution function, s t+1 and s t Represent the state vectors at time t+1 and time t (including the voltage, SOC, temperature and other parameters of each single battery), a t Represents the balancing action vector at time t (including the current magnitude and direction of each balancing channel);

[0083] The model is based on the variational autoencoder (VAE) architecture, which includes the encoder network q φ (z|s) and decoder network p θ (s′|z,a), where z is the latent state variable, φ and θ enc is the network parameter;

[0084] The loss function of the variational autoencoder is:

[0085]

[0086] Where L(θ enc ,φ;s,a,s′) is the loss function of the variational autoencoder, φ and θ enc denote the encoder and decoder network parameters respectively, s and s′ denote the state vectors at time t and time t+1 respectively, a denotes the equilibrium action vector at time t, is the reconstruction error, q φ (z|s) represents the encoder network, p θ (s′|z, a) represents the decoder network, z is the potential state variable, KL[q φ (z|s)||p(z)] is the KL divergence regularization term; p(z) represents the prior distribution of the latent variable z; β vae To trade off parameters, we control the structure of the latent space.

[0087] In a specific implementation, the VAE model in this invention adopts the following structure: the encoder network consists of a three-layer fully connected neural network. 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 use the LeakyReLU activation function. The output layer consists of two parts: the mean and the logarithmic variance, which are used to parameterize the distribution of the latent variable Z. The decoder network also uses a three-layer fully connected network. It takes the latent variable Z and the control action a as input, processes them through the middle layers of 64 and 128 neurons, and finally outputs the predicted state at the next moment.

[0088] An example of the application of this model in an electric vehicle power battery pack is as follows: for a battery pack consisting of 96 lithium-ion battery cells, the state of each cell includes four parameters: voltage, SOC, surface temperature, and internal resistance, and the control action includes the current value of 48 balancing channels. Under high-speed driving conditions, the model receives the current state and control actions of all battery cells as input and predicts the battery state distribution after 5 minutes. The prediction results include not only the expected state value of each battery, but also the uncertainty of each parameter. For example, the predicted SOC value of battery #32 is 85% ± 2.1% (95% confidence interval). This enables the system to adjust the balancing strategy based on the prediction uncertainty, ensuring that excessive balancing operations are avoided when the prediction accuracy is insufficient.

[0089] Step 2.2, Bayesian parameter learning;

[0090] The Bayesian learning method is used to estimate the model parameters, using the prior distribution p(θ enc ) and p(φ), and calculate the posterior distribution p(θ enc ,φ|D). Since the posterior distribution is difficult to calculate directly, variational inference is used to approximate the solution:

[0091] q ω (θ enc ,φ)≈p(θ enc ,φ|D);

[0092] where q ω (θ enc ,φ) represents the approximate posterior distribution determined by the parameter ω, p(θ enc ,φ|D) represents the posterior distribution of the parameter given the observed data;

[0093] The optimal parameter ω is solved by minimizing the KL divergence:

[0094] KL(q ω (θ enc ,φ)||p(θ enc ,φ|D))

[0095] where KL(·||·) represents the KL divergence (Kullback-Leibler divergence); ω represents the parameter of the variational distribution, which is used to measure the difference between two probability distributions;

[0096] In the actual implementation, the Monte Carlo Variational Inference (MCVI) algorithm is used to sample multiple sets of parameters to form an integrated model:

[0097]

[0098] in represents M sampled by MCVI algorithm param Group parameters, where i1 represents the index of the parameter group, M param Indicates the total number of sampled parameter groups; and They represent the encoder and decoder network parameters of the i1th group of samples respectively.

[0099] Step 2.3, multi-step probability prediction;

[0100] Based on the learned probabilistic state transition model, a rolling prediction of the battery pack state in the next N steps is performed. Given the current state s t and the planned equilibrium action sequence:

[0101] a t ,a t+1 ,…,a t+N-1 ;

[0102] where a t 、a t+1 、a t+N-1 They represent the equilibrium action vectors at time t, time t+1, and time t+N-1, respectively, and N represents the number of predicted time steps;

[0103] Predict the state distribution at each moment in the future:

[0104] p(s t+1 |s t ,a t );

[0105] p(s t+2 |s t , a t ,a t+1 )=∫p(s t+2 |s t+1 ,a t+1 )p(s t+1 |s t ,a t )ds t+1 ;

[0106]

[0107]

[0108] Where p(s t+1 |s t , a t ) represents the state s at a given time t t and action a t Under the condition, the state s at time t+1 t+1 The conditional probability distribution of p(s t+2 |s t+1 , a t+1 ) represents the state s at time t+1 t+1 and action a t+1 Under the condition, the state s at time t+2 t+2 The conditional probability distribution of p(s t+N |s t+N-1 , a t+N-1 ) represents the state s at a given time t+N-1 t+N-1 and action a t+N-1 Under the condition, the state s at time t+N t+N The conditional probability distribution of p(s t+N-1 |s t , a t ,...,a t+N-2 ) indicates a given initial state s t and action sequence a t ,...,a t+N-2 Under the condition, the state s at time t+N-1 t+N-1 The conditional probability distribution of ; ∫ represents the integral operation;

[0109] Due to the complexity of integral calculation, the particle filter algorithm is used to approximate the solution, sampling K at each time step. par Particles:

[0110]

[0111] in Respectively represent the first, second, and Kth time t+k par particle samples, K par Indicates the number of particles used in each time step in the particle filter algorithm.

[0112] Step 2.4, quantification of prediction uncertainty;

[0113] Quantify the uncertainty of the prediction results, including epistemic uncertainty (uncertainty in model parameters) and aleatoric uncertainty (inherent randomness in the system).

[0114] The predicted state s of the jth battery j，t+k, calculate its 95% confidence interval and prediction variance Used to evaluate the reliability of prediction results and guide subsequent risk perception decisions;

[0115] Confidence intervals and prediction variances are obtained through sampling statistics of multiple sets of parameters and multiple particles.

[0116] Where j represents the index of the single battery, k represents the time step index, and represents the kth time step in the future; s j,t+k represents the predicted state of the jth single cell at time t+k; represents the 95% confidence interval for the predicted state, where For the lower bound, For the upper bound; It represents the variance of the predicted state and is used to quantify the magnitude of the prediction uncertainty.

[0117] Epistemic uncertainty refers to the prediction uncertainty caused by inaccurate estimation of model parameters, and stochastic uncertainty refers to the prediction uncertainty caused by the randomness of the system itself.

[0118] Step 3: Based on the time scale component and the state prediction distribution, three parallel predictive controllers are constructed: short-term controller, medium-term controller, and long-term controller;

[0119] Based on the multi-timescale data processing results from step 1 and the probabilistic state prediction output from step 2, this step constructs three parallel predictive controllers, corresponding to the short-term, medium-term, and long-term equilibrium tasks, to achieve full-time equilibrium control. The prediction uncertainty quantification results generated in step 2 provide important decision-making basis for the controllers in this step, enabling the control strategy to adapt to prediction results with different confidence levels.

[0120] It includes the following sub-steps:

[0121] Step 3.1, short-term controller construction;

[0122] Construct a short-term prediction controller C s The prediction range is from seconds to minutes, and it is mainly responsible for dealing with transient responses 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:

[0123]

[0124] in, represents the short-term state vector at time t+1, is the short-term state vector, For short-term control actions, is the output observation, and are process noise and measurement noise respectively, A s 、B s and C s represent the state transfer matrix, control input matrix and output matrix respectively.

[0125] The objective function of the short-term controller is:

[0126]

[0127] Among them J s represents the objective function of the short-term controller, N s is the short-term prediction time domain length (such as 60 seconds), Represents weighted Q s The L2 norm of Represents weighted R s The L2 norm of is the reference trajectory, Q s and R s They are the weight matrices for outputting tracking error and control cost, respectively, to balance tracking accuracy and control cost; Represents the output at time t+i predicted at time t; represents the reference trajectory at time t+i, Indicates the control action at time t+i-1 calculated at time t.

[0128] The optimal control sequence is obtained by solving the optimization problem:

[0129]

[0130] in Represents the solution to minimize the objective function J s The optimization problem is the control sequence

[0131] And apply the rolling optimization strategy, only executing the current step control action:

[0132]

[0133] in It means that the control action actually executed is equal to the control action obtained by optimization at the current moment.

[0134] Step 3.2, mid-term controller construction;

[0135] Constructing a mid-term prediction controller C m The prediction range is from hours to days, and it is mainly responsible for optimizing daily balancing efficiency and charge and discharge cycle management.

[0136] The mid-term controller adopts a nonlinear model predictive control architecture using a neural network state transition model:

[0137]

[0138] in represents the mid-term state vector at time t+1, f m is the neural network model, θ net are network parameters, is the mid-term state vector, For mid-term control action, is the mid-term process noise.

[0139] The objective function of the mid-term controller is:

[0140]

[0141] Among them J m represents the objective function of the mid-term controller, N m is the length of the medium-term prediction time domain (e.g. 24 hours); Represents weighted Q m The L2 norm of Represents weighted R m The L2 norm of Represents weighted P m The L2 norm of represents the state at time t+i predicted at time t, represents the control action at time t+i-1 calculated at time t; is the reference state, Q m Represents the weight matrix of state tracking error, R m Represents the weight matrix of control cost, P m represents the terminal cost matrix; Represents the t+N predicted at time t m The state at the moment is the predicted value of the state at the end of the prediction time domain.

[0142] Solve nonlinear optimization problems through numerical optimization methods to obtain mid-term control actions

[0143] The neural network state transition model f in the present invention m The architecture of Bi-directional Long Short-Term Memory (Bi-LSTM) combined with residual connection is adopted. The model structure includes: the input layer receives the mid-term state vector and control actions The feature extraction layer consists of two layers of Bi-LSTM, each layer contains 128 hidden units, which can capture both forward and backward temporal dependencies. The residual connection layer directly connects the input features to the deep features, alleviating the gradient vanishing problem and improving training stability. The output layer uses a fully connected network to predict the state at the next moment. The network is trained using a sequence-to-sequence (seq2seq) framework and uses a teacher-forcing strategy to improve long-term prediction accuracy.

[0144] An example of the application of this neural network model in a stationary energy storage power station scenario is as follows: For a cascaded energy storage system consisting of 500 battery cells, in a photovoltaic power generation supporting energy storage application, the system needs to predict the battery pack state changes over the next 24 hours. The model inputs include the current medium-term state characteristics of each battery cell (SOC mean and variance, capacity decay trend, daily cycle depth, etc.) and the planned charge and discharge curve. The model can predict the battery pack state parameters for each period over the next 24 hours. For example, if it predicts that the SOC difference between the battery packs in the system will increase to 8.3% after 12 hours, it is appropriate to schedule a two-hour active balancing process to control the difference within 5%. Based on these predictions, the medium-term controller generates a balancing plan for the next day, scheduling balancing operations during periods of low grid load while avoiding peak photovoltaic power generation periods, thereby ensuring balancing results and maximizing the economic benefits of the energy storage system.

[0145] Step 3.3, long-term controller construction;

[0146] Build a long-term prediction controller C1 with a prediction range of weeks to months, which is mainly responsible for maintaining the consistency of the battery pack's health status and extending the cycle life;

[0147] The long-term controller is based on a health state model and a degradation model, combined with a periodic optimization strategy:

[0148]

[0149] in represents the health status of the jth battery at time t, represents the health status of the jth battery at time t+1, is the attenuation coefficient, f deg is the degradation function, represents the state of charge of the jth battery at time t, represents the discharge depth of the jth battery at time t, represents the charge and discharge current of the jth battery at time t, represents the temperature of the jth battery at time t.

[0150] The objective function of the long-term controller is:

[0151]

[0152] Among them J l represents the objective function of the long-term controller, E[·] represents the expected value, and N l is the length of the long-term prediction time domain (such as 30 days), γ disc is the discount factor, Represents the health status vector of all n batteries at time t+i, SOH t+i is the average health status of the battery pack; ||·||2 represents the L2 norm, i.e., the Euclidean distance; λ soh The weight coefficient representing the weakest battery health state, Indicates the minimum health status of all batteries at time t+i.

[0153] The long-term controller outputs the control strategy π1, which guides the parameter adjustment of the medium-term and short-term controllers.

[0154] Step 3.4, hierarchical control coordination;

[0155] Implement a coordination mechanism among three time-scale controllers. The upper controller provides constraints and target references to the lower controller, and the lower controller feeds back execution results to the upper controller. Specifically:

[0156] Long-term controller C l For the mid-term controller C m Provide health status target SOH target and SOC operating range constraints (SOC min , SOC max );

[0157] Mid-term controller C m is the short-term controller C s Provide reference trajectory and control constraints

[0158] 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.

[0159] Among them (SOC min , SOC max ) represents the SOC operating range constraint, where SOC min is the minimum SOC limit, SOC max 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.

[0160] Step 4: Based on the state prediction distribution and the predicted controller output, the traditional fuzzy rules are expanded into a probabilistic form, and the decision uncertainty is quantified through Bayesian fuzzy reasoning;

[0161] After obtaining the probabilistic state prediction results from step 2 and the multi-level control outputs from step 3, this step expands the traditional fuzzy inference system into a Bayesian fuzzy inference framework. This accurately quantifies the uncertainty of equilibrium decisions and improves the system's decision-making reliability under data uncertainty. The preliminary control decisions generated by each controller in step 3 are further optimized in this step using Bayesian fuzzy inference, enhancing the adaptability and robustness of the decisions.

[0162] It includes the following sub-steps:

[0163] Step 4.1, Bayesian fuzzy rule construction;

[0164] Construct a Bayesian fuzzy rule library, where each Bayesian fuzzy rule is expressed in a probabilistic form:

[0165]

[0166] in represents the i2th Bayesian fuzzy rule, x in is the input variable, They represent the input fuzzy set, output fuzzy set and credibility of the i2th Bayesian fuzzy rule respectively; y out is the output variable; IF and THEN are the condition and conclusion keywords in the fuzzy rules, and is represents the affiliation.

[0167] Based on expert knowledge and historical data of battery balancing management, a rule base is built that contains the following types of rules:

[0168] Voltage difference rule: determines the equalization direction and strength based on the voltage difference between batteries;

[0169] SOC balancing rule: guides the balancing strategy based on the battery SOC distribution characteristics;

[0170] Temperature safety rules: adjust or limit the balancing current according to the temperature status;

[0171] Health status protection rules: adopt differentiated balancing strategies based on differences in battery health status;

[0172] Each type of rule contains multiple sub-rules, and the total number of rules is between 50 and 100, covering all possible battery status combinations.

[0173] Step 4.2, fuzzy set parameter learning;

[0174] The Bayesian learning method is used to estimate the fuzzy set parameters. i and B i Parameterized as parameter vector θ A and θ B , learn the posterior distribution p(θ A ,θ B |D).

[0175] Specifically, for the Gaussian blur membership function:

[0176]

[0177] where μ A (x in θ A ) represents the input x in The membership degree of fuzzy set A, σ mem represents the standard deviation parameter of the Gaussian blur membership function, exp(·) represents the natural exponential function, and m A represents the mean parameter of the Gaussian blur membership function; θ A The parameter vector representing the fuzzy set A is expressed as:

[0178] θ A =(m A ,σ mem );

[0179] Use the Markov Chain Monte Carlo method to get the posterior distribution p(θ A ,θ B |D) Sampling multiple sets of parameters Constitute a parameter set.

[0180] where p(θ A ,θ B |D) represents the posterior distribution of the fuzzy set parameters after given historical data D, represents N sampled by MCMC method fuz Group parameters, where j1 represents the index of the parameter group, N fuz Indicates the total number of parameter groups sampled.

[0181] Step 4.3, probabilistic fuzzy reasoning;

[0182] Based on Bayesian fuzzy rules and parameter distribution, probabilistic fuzzy reasoning is performed. Given an input x in,0 , according to each set of parameters Calculate the fuzzy inference results separately Get the output probability distribution p(y out |x in，0) rather than a single fixed value. The reasoning process uses the Mamdani fuzzy reasoning model, which includes the following steps:

[0183] Fuzzification, calculate the input x in，0 For each input fuzzy set Membership

[0184] Rule triggering, calculation of each rule Trigger strength:

[0185]

[0186] in Represents input x in，0 Fuzzy Set The membership degree, Representation Rules The trigger strength, Representation Rules credibility;

[0187] Fuzzy reasoning, calculate the output fuzzy set of each rule:

[0188]

[0189] in Representation Rules The output fuzzy set after reasoning, min means taking the minimum value operation, Represents the output value y out Fuzzy Set The degree of membership;

[0190] Fuzzy aggregation, merging the output fuzzy sets of all rules:

[0191]

[0192] Where B(y out ) represents the output fuzzy set after aggregation, Indicates taking the maximum value among all rules i2;

[0193] Defuzzification, using the centroid method to calculate the crisp output value:

[0194]

[0195] where y out,0 represents the clear output value after defuzzification, ∫ represents the integral operation, represents the weighted average calculated using the centroid method;

[0196] By repeating the above reasoning process for multiple sets of parameters, the output probability distribution p(y out |xin，0 );

[0197] Step 4.4, quantify the uncertainty of reasoning;

[0198] Based on the results of probabilistic fuzzy reasoning, the uncertainty of decision making is quantitatively analyzed. out |x in，0 ), calculate the following uncertainty indicators:

[0199] The mean is used as the final decision value:

[0200] μ out =E(y out |x in,0 )=∫y out ·p(y out ||x in,0 )d yout ;

[0201] where μ out represents the mean of the output distribution, E(y out |x in，0 ) indicates that given input x in,0 Output y under the condition out The expected value of ∫y out ·p(y out |x in,0 )dy out Indicates the integration of the product of the output value and its probability density;

[0202] Variance quantifies decision uncertainty:

[0203]

[0204] in Indicates the variance of the inference result, Var(y out |x in，0 ) indicates that given input x in,0 Output y under the condition out The variance of (y out -y out ) 2 Represents the square of the difference between the output value and the mean;

[0205] The 95% confidence interval represents the decision range:

[0206] (y out，lower ,y out，upper );

[0207] Where (y out，lower ,y out，upper) Represents a 95% confidence interval for the output, where yout, lower is the lower bound, and y out，upperFor the upper bound;

[0208] Entropy represents the uncertainty of decision information:

[0209] H(y out |xin,0)=-∫p(y out |x in ,0).logp(y out |x in,0 )dyout;

[0210] Among them, H(y out |x in，0 ) indicates that given input x in，0 Output y under the condition out The entropy of, log represents the natural logarithm, ∫ represents the integral operation, p(y out |x in , 0 represents the probability distribution of the output;

[0211] These uncertainty indicators are passed to the subsequent risk perception decision module to guide the formulation of equilibrium control strategies.

[0212] An example of the application of a Bayesian fuzzy inference system in battery pack balancing management for electric buses is as follows: In an electric bus using a 300kWh lithium battery pack, when the voltage difference between the third and seventh battery modules reaches 150mV, a traditional deterministic fuzzy system directly outputs a balancing current of 500mA. However, the Bayesian fuzzy inference system of the present invention takes into account the ±2°C fluctuation in the current battery temperature sensor reading and the uncertainty of the SOC estimation, and generates a probability distribution for the balancing current with a mean of 480mA, a 95% confidence interval of [420mA, 540mA], and a variance of 18mA. 2 , with an entropy value of 3.2. Further analysis revealed that the uncertainty in this decision primarily stemmed from unstable temperature sensor readings. The system automatically adjusted the execution intensity to 0.85, ultimately outputting a balancing current of 408 mA, avoiding the risk of over-balancing due to information uncertainty. The system also flagged the temperature sensor as a possible fault and reminded maintenance personnel to conduct a focused inspection during the next maintenance visit, improving system maintainability.

[0213] Step 5: Based on the state prediction distribution, the prediction controller output and the Bayesian fuzzy inference results, a risk-adjusted objective function is constructed based on the conditional value at risk to generate the equilibrium control decision and its confidence interval;

[0214] Combining the probabilistic prediction results from step 2, the controller output from step 3, and the Bayesian fuzzy inference results from step 4, this step constructs a risk-adjusted objective function based on the Conditional Value at Risk (CVaR) theory. This accurately measures and balances the risks of equilibrium decisions, ensuring system safety in extreme situations. The decision uncertainty indicator output from step 4 provides a key input for the risk assessment in this step, making risk measurement more accurate and reliable.

[0215] It includes the following sub-steps:

[0216] Step 5.1, risk measurement indicator construction;

[0217] A risk metric based on conditional value at risk is formed, which is defined as follows:

[0218]

[0219] in Indicates that at the confidence level α risk Conditional value at risk, X loss is a random loss variable; Indicates the loss variable X loss Greater than or equal to the value at risk Expected loss under the conditions; At the confidence level α risk The risk value under the value is that the probability of loss exceeding this value is no more than 1-α risk ,Right now

[0220] In battery balancing management, the loss function is defined as:

[0221]

[0222] Where L(s,a) represents the total loss function of taking action a in state s, L imbalance 、L safety 、L efficiency They represent the imbalance loss function, security risk loss function, and energy efficiency loss function respectively, and λ loss，1 ,λ loss，2 and λ loss,3 are the loss weight coefficients of imbalance loss function, safety risk loss function, and energy efficiency loss function respectively.

[0223] Step 5.2, risk adjustment objective function construction;

[0224] Form a risk-adjusted objective function based on CVaR risk measurement:

[0225]

[0226] Among them J CVaR represents the risk-adjusted objective function, E(R(s,a)) represents the expected value of the reward function R(s,a), and R(s,a) is the equilibrium reward function (i.e., the negative loss function -L(s,a)); λ risk is the risk aversion coefficient, which controls the trade-off between expected returns and extreme risks; α risk is the confidence level (usually 0.95 or 0.99); Indicates that at the confidence level 1-α risk The conditional risk value under negative rewards, -R(s,a) represents a negative reward function, which is equivalent to the loss function L(s,a).

[0227] The first term of the objective function maximizes the expected return, and the second term minimizes the extreme risk. The weighted balance of the two yields the optimal strategy.

[0228] Step 5.3, scenario sampling and risk assessment;

[0229] Use the Monte Carlo method to assess decision risk and sample possible future scenarios. The specific steps are as follows:

[0230] Generate M from the probabilistic state transition model traj Sample future state trajectories:

[0231]

[0232] Where 1:N represents the trajectory from time step 1 to time step N, which contains N consecutive time steps; Represent the 1st, 2nd and Mth traj trajectory samples, M traj Indicates the total number of sampled trajectories.

[0233] For each sampled trajectory, calculate the reward under control action a

[0234] Sort the reward samples and determine the VaR threshold That is, it satisfies:

[0235]

[0236] in Indicates that at the confidence level α risk The risk value threshold under risk Indicates the confidence level of the risk assessment;

[0237] In actual calculation, the corresponding position after sorting is α risk ·M trajSample reward values ​​for a position; calculate the CVaR risk metric:

[0238]

[0239] in Indicates that at the confidence level 1-α risk Conditional value at risk under negative rewards, Indicates that after sorting, the risk )M traj to M traj Sum of negative reward samples within the range, represents the i3th reward sample after sorting, Represents the normalization coefficient when calculating CVaR.

[0240] Step 5.4, optimize and solve the risk perception decision;

[0241] Based on the risk-adjusted objective function, the optimal equilibrium decision is solved by numerical optimization method:

[0242] a * (t) = argmin a J CVaR ;

[0243] where a * (t) represents the optimal control action at time t, argmin a J CVaR Represents the objective function J CVaR Minimize action a;

[0244] The following optimization strategies are specifically adopted:

[0245] For discrete action spaces, a sampling-based optimization method is used, using the Cross-Entropy Method (CEM);

[0246] For continuous action spaces, use gradient descent or conjugate gradient methods;

[0247] Consider action constraints during optimization;

[0248] Finally, the risk-optimal equilibrium control action a is obtained * (t).

[0249] Step 5.5, decision confidence interval calculation;

[0250] Based on the risk optimization results, calculate the confidence interval of the decision (a lower , a upper ), reflecting the credibility of the decision;

[0251] The confidence interval is obtained by the following steps:

[0252] Use parametric or nonparametric methods to estimate the decision distribution p(a|s t );

[0253] Calculate the decision mean μ a As the final decision value;

[0254] Calculate β based on the distribution dec Confidence interval (usually β dec =0.95);

[0255] The confidence interval information will be passed to the adaptive integration module to adjust the decision execution strength.

[0256] Among them (a lower , a upper ) represents the confidence interval of the decision, a lower is the lower bound, a upper is the upper bound, p(a|s t ) means in state s t The probability distribution of action a under the condition, μ a Represents the mean of the decision distribution, as the final decision value, β dec It represents the confidence level of the confidence interval, usually 0.95, indicating a 95% confidence level.

[0257] Step 6: Based on the output of the predictive controller, the Bayesian fuzzy inference result, and the balanced control decision, the execution intensity is adjusted according to the confidence interval and the balanced current instruction is output;

[0258] After completing the multi-level control calculations in step 3, the Bayesian fuzzy reasoning in step 4, and the risk-aware decision-making in step 5, this step adaptively fuses the outputs of the three time-scale controllers, dynamically adjusting the control weights based on the current system state and uncertainty level to generate the final balanced control instructions. The decision confidence intervals generated in step 5 are directly used to adjust the execution intensity in this step, ensuring that the system adopts a more conservative control strategy when uncertainty is high.

[0259] It includes the following sub-steps:

[0260] Step 6.1, dynamic weight calculation;

[0261] Calculate the dynamic fusion weight γ of the three time scale controller outputs s (t), γ m (t) and γ l (t); Weight calculation is based on the following factors:

[0262] Current system status characteristics: such as voltage imbalance, SOC distribution, temperature change rate, etc.;

[0263] Forecast uncertainty level: the width of the confidence interval of the forecast results at each time scale;

[0264] Historical control effect evaluation: the recent performance and adaptability of each controller;

[0265] The weight calculation adopts the adaptive weighting algorithm:

[0266]

[0267] where γ s (t), γ m (t) and γ l (t) represents the fusion weights of the short-term controller, the medium-term controller, and the long-term controller at time t, respectively, and γ i4 (t) represents the fusion weight of controller i4 at time t, where represents the short-term, medium-term and long-term controllers respectively, exp(·) represents the natural exponential function, and are the weight characteristics of controller i4 and controller j3 respectively, ∑ represents the summation symbol, and is calculated by combining the above three factors:

[0268]

[0269] in represents the weight eigenvalue of controller i4 at time t, f state 、f uncertainty and f history They are state characteristic function, uncertainty evaluation function and historical effect evaluation function, β fus，1 , β fus，2 and β fus，3 are the fusion balance coefficients of the state characteristic function, uncertainty evaluation function and historical effect evaluation function respectively.

[0270] Step 6.2, control output fusion;

[0271] Based on the calculated dynamic weights, the outputs of the three time-scale controllers are fused to generate the final balancing control action:

[0272]

[0273] where u * (t) represents the final fusion control action at time t, and are the optimal control outputs of the short-term, medium-term and long-term controllers respectively.

[0274] The fusion process ensures the continuous and smooth transition of control actions and avoids system oscillations caused by sudden changes.

[0275] Step 6.3, perform intensity adjustment;

[0276] According to the decision confidence interval (u lower ,u upper ) width, adaptively adjust the execution intensity of the balanced control action. When uncertainty is high (the confidence interval is wide), reduce the execution intensity to reduce risk; when uncertainty is low (the confidence interval is narrow), maintain the original execution intensity. The execution intensity adjustment function is:

[0277]

[0278] Where η(t) represents the execution intensity adjustment coefficient at time t, η base is the basic execution strength (usually 1), k decay To implement the intensity attenuation coefficient, - u is the control action normalization factor, exp(·) represents the natural exponential function, u lower and u upper are the lower and upper bounds of the confidence interval of the control action, respectively.

[0279] The final actual execution control action is:

[0280] u actual (t) = η(t)·u * (t);

[0281] where u actual (t) represents the actual execution control action at time t.

[0282] Step 6.4, generating a control signal for the equalization circuit;

[0283] The integrated balancing control action is converted into a specific balancing circuit control signal, including parameters such as balancing direction, balancing current, and balancing time. The control signal is transmitted to the balancing circuit driver module of the battery management system (BMS) through the communication interface to realize the control of the actual balancing hardware.

[0284] The control signal generation takes into account hardware limitations and safety constraints to ensure that the balanced current does not exceed the maximum allowed by the system while avoiding efficiency losses caused by frequent switching. The continuous control signal is converted into a discrete switching control sequence through pulse-width modulation (PWM) or other suitable modulation methods.

[0285] A battery pack active balancing management system, used to execute the above-mentioned battery pack active balancing management method, comprising:

[0286] Multi-time scale data processing module, used to collect battery pack parameter data and perform time scale decomposition, decomposing the data into dynamic components of different frequencies;

[0287] Probabilistic state prediction module, used to build a state transition probability model, achieve multi-step prediction of the battery pack's future state and quantify prediction uncertainty;

[0288] A multi-level parallel control module is used to build three predictive controllers for short, medium and long periods of time based on different time scales to collaboratively manage the battery pack balancing process;

[0289] Bayesian fuzzy reasoning module, which is used to probabilistically expand fuzzy rules and quantify decision uncertainty through parameter learning and probabilistic reasoning;

[0290] The risk-aware decision-making module is used to construct risk metrics and risk-adjusted objective functions, and generate equilibrium control strategies with confidence intervals;

[0291] The adaptive integrated execution module is used to fuse the outputs of each controller and adjust the execution intensity according to the uncertainty level to generate the final balanced control signal.

[0292] Here, the present invention provides an implementation example:

[0293] This implementation was validated in an electric logistics vehicle from a new energy commercial vehicle manufacturer. This vehicle utilizes a CATL lithium titanate battery pack consisting of 16 battery modules, each composed of 12 cells connected in series, for a total of 192 cells. The battery management system is equipped with an active balancing circuit capable of ±2A balancing current regulation and an embedded computing platform to execute the algorithm in this implementation.

[0294] Logistics vehicles primarily operate in urban express delivery, operating 12 hours daily. These vehicles are characterized by frequent starts and stops, large load fluctuations, and significant ambient temperature variations. Traditional balancing management methods face multiple challenges in this scenario: First, frequent acceleration and deceleration make the battery load highly dynamic, making accurate balancing control difficult. Second, 24 / 7 operation leads to uneven battery temperature distribution, exacerbating battery consistency issues. Third, operating conditions vary significantly across routes and seasons, making a universal balancing strategy difficult to apply. Finally, sensor data quality is affected by factors such as vehicle vibration, resulting in uncertainty.

[0295] The system collects complete battery data every second, including cell voltage (3.5V-2.2V), temperature (0-60°C), bus current (-150A to +150A), and estimated SOC (0-100%). The collected data is decomposed into components at three time scales using a wavelet transform. Table 1 shows an example of the voltage components of a battery cell at different time scales during a typical workday.

[0296] Table 1: Multi-timescale voltage components of battery cell #45 during a typical workday

[0297] Time point Original voltage (V) Short-term component (mV) Medium-term component (V) Long-term component (V) 8:00 2.476 +24.8 2.423 2.428 10:30 2.315 -32.6 2.352 2.430 13:15 2.563 +56.2 2.504 2.431 16:40 2.388 -15.5 2.407 2.432 19:20 2.429 +7.3 2.425 2.433

[0298] The system uses a variational autoencoder model to predict the battery pack state and quantify the uncertainty of the prediction. During a delivery mission after a courier vehicle departs from the warehouse, the system predicts the battery module state of charge (SOC). The prediction includes a mean and a 95% confidence interval, reflecting the uncertainty level of the prediction at different points in time.

[0299] Based on the prediction results, the system generates a hierarchical equilibrium control strategy. In a complete operation cycle, the outputs of the three different time scale controllers and the fusion results are shown in Table 2:

[0300] Table 2: Controller output and fusion results at different time scales (balanced current, unit: mA)

[0301]

[0302] The system's Bayesian fuzzy inference results for the case where some batteries in the battery pack have abnormally high temperatures are shown in Table 3:

[0303] Table 3: Bayesian fuzzy inference results under abnormal temperature conditions (balanced current, unit: mA)

[0304]

[0305] This table compares the decision differences between the traditional deterministic fuzzy system and the Bayesian fuzzy system of this embodiment under the same circumstances.

[0306] The system generates risk-aware decisions based on conditional risk value and adaptively adjusts the execution intensity based on the decision confidence. The decision results under different risk conditions are shown in Table 4:

[0307] Table 4: Decision-making and execution results under different risk conditions

[0308]

[0309] After six months of application on the aforementioned electric logistics vehicle, this implementation was compared with the original balancing management system, primarily validating the following two key technical effects:

[0310] Full-time balanced management effect:

[0311] The comparison of the balancing effect of this embodiment and the traditional single time scale balancing method under different working conditions is shown in Table 5:

[0312] Table 5: Comparison of the balancing effect between this embodiment and the traditional method

[0313] Evaluation Metrics Traditional methods This embodiment Improvement rate Voltage balancing convergence time (minutes) 42.5 20.8 51.1％↓ Voltage consistency (mV, standard deviation) 24.6 8.1 67.1％↓ SOC consistency (percentage points, standard deviation) 3.8 1.2 68.4％↓ Energy consumption during balancing (Wh / time balancing) 18.3 15.5 15.3％↓ Balanced operation frequency (times / day) 5.2 2.8 46.2％↓ Cycle life improvement (estimated) - +12.5％ -

[0314] Uncertainty and risk management effectiveness:

[0315] The robustness and security of this implementation and the traditional deterministic equalization method under data uncertainty or sensor failure conditions are compared, as shown in Table 6:

[0316] Table 6: Comparison of uncertainty and risk management effectiveness

[0317]

[0318] The above data demonstrates that this implementation achieves significant technical benefits compared to traditional methods: In terms of full-time balancing management, balancing convergence time is reduced by 51.1%, and state consistency is improved by approximately 67%. In terms of uncertainty and risk management, the system maintains balancing efficiency of 60%-75% even under abnormal conditions, and the incidence of safety incidents is reduced by 76.2%. These results fully demonstrate the technical advantages and innovative value of this invention in the field of active battery pack balancing management.

[0319] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A battery pack active balancing management method, characterized in that: The following steps are involved: Collect battery pack parameter data and decompose the system dynamics into three time scale components: short-term dynamics, medium-term dynamics, and long-term dynamics using a time scale decomposition algorithm. The time scale decomposition algorithm includes: applying wavelet transform to perform multi-scale decomposition on the pre-processed time series data, where the short-term dynamic component is at the second to minute level to capture transient responses and rapid fluctuations; the medium-term dynamic component is at the hour to day level to reflect diurnal changes and charge and discharge cycles; and the long-term dynamic component is at the week to month level to reflect aging trends and seasonal changes; Based on the time scale component, a probabilistic state transition model is constructed to probabilistically predict the future state of the battery pack and output the state prediction distribution. The probabilistic state transition model is implemented based on the variational autoencoder architecture, including an encoder network and a decoder network. The posterior distribution is estimated through variational inference, the model parameters are estimated using the Bayesian learning method, and the particle filter algorithm is used for multi-step probabilistic prediction; Based on the time scale component and the state prediction distribution, three parallel predictive controllers are constructed: short-term controller, medium-term controller and long-term controller; Based on the state prediction distribution and the predicted controller output, the traditional fuzzy rules are expanded into a probabilistic form, and the decision uncertainty is quantified through Bayesian fuzzy reasoning, which includes: Construct a Bayesian fuzzy rule library, where each Bayesian fuzzy rule is expressed in a probabilistic form; Bayesian learning method is used to estimate fuzzy set parameters; Probabilistic fuzzy reasoning based on Bayesian fuzzy rules and parameter distribution; Calculate the mean, variance, confidence interval, and entropy of the inference output distribution to quantify decision uncertainty; Based on the state prediction distribution, the predicted controller output and the Bayesian fuzzy inference results, a risk-adjusted objective function is constructed based on the conditional value at risk to generate the equilibrium control decision and its confidence interval; According to the output of the predictive controller, the Bayesian fuzzy reasoning result and the balanced control decision, the execution intensity is adjusted according to the confidence interval, and the balanced current instruction is output. The execution intensity adjustment function for adjusting the execution intensity is: Where η(t) represents the execution intensity adjustment coefficient at time t, η base is the basic execution strength, exp(·) represents the natural exponential function, k decay To implement the intensity attenuation coefficient, - u is the control action normalization factor, u lower and u upper They represent the lower and upper bounds of the confidence interval of the control action, respectively.

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

3. The battery pack active balancing management method according to claim 1, characterized in that: The short-term controller is based on a model predictive control framework and uses a linearized state space model with a prediction range of seconds to minutes. The medium-term controller adopts a nonlinear model predictive control architecture and uses a neural network state transition model with a prediction range of 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 of weeks to months.

4. The battery pack active balancing management method according to claim 3, characterized in that: The neural network state transfer model adopts an architecture that combines a bidirectional long short-term memory network with a residual connection. The model structure includes: The input layer receives the mid-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.

5. The battery pack active balancing management method according to claim 1, characterized in that: In the risk-adjusted objective function constructed based on the conditional value at risk, the risk-adjusted objective function includes maximizing the expected return term and minimizing the extreme risk term.

6. A battery pack active balancing management system, characterized in that: A method for actively balancing a battery pack according to any one of claims 1 to 5, comprising: Multi-time scale data processing module, used to collect battery pack parameter data and perform time scale decomposition, decomposing the data into dynamic components of different frequencies; Probabilistic state prediction module, used to build a state transition probability model, achieve multi-step prediction of the battery pack's future state and quantify prediction uncertainty; A multi-level parallel control module is used to build three predictive controllers for short, medium and long periods of time based on different time scales to collaboratively manage the battery pack balancing process; Bayesian fuzzy reasoning module, which is used to probabilistically expand fuzzy rules and quantify decision uncertainty through parameter learning and probabilistic reasoning; The risk-aware decision-making module is used to construct risk metrics and risk-adjusted objective functions, and generate equilibrium control strategies with confidence intervals; The adaptive integrated execution module is used to fuse the outputs of each controller and adjust the execution intensity according to the uncertainty level to generate the final balanced control signal.

Images