A distributed battery state estimation and compensation method based on voltage and current sampling

Abstract

The present invention relates to the technical field of battery management systems, and specifically to a distributed battery state estimation and compensation method based on voltage and current sampling, comprising: synchronously collecting voltage and current waveforms, extracting ripple characteristics and slope mutation points to construct an original characteristic matrix, and generating a characteristic tensor with environmental correction through a coupling compensator; then constructing a state parameter space based on adaptive particle injection and an electrochemical-thermodynamic coupling model. A multi-source data fusion engine is used to achieve distributed collaborative estimation of health status index and state of charge confidence interval, and a multi-state dynamic prediction model is used to complete SOC, SOH and temperature trend prediction and residual compensation correction; based on the prediction results and uncertainty boundaries, an energy compensation optimization model is constructed to search for the Pareto optimal compensation strategy. The present invention can be widely used in the intelligent management and control of battery energy storage systems and electric power equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and in particular to a distributed battery state estimation and compensation method based on voltage and current sampling. Background Art

[0002] In new energy systems, especially in energy storage and electric power equipment, distributed battery packs serve as core energy carriers, and their operating status directly impacts the safety, lifespan, and energy efficiency of the entire system. Due to the large number of battery cells, significant differences in operating environments, and inconsistent aging rates, accurately sensing multi-dimensional parameters such as each cell's state of health (SOH), state of charge (SOC), and temperature is crucial to improving system performance. In recent years, with advances in sensor technology and edge computing capabilities, research on battery state estimation based on locally sampled data has continued to advance. At the same time, integrated modeling and optimized control for state prediction and energy regulation have become a key development direction for efficient and intelligent battery management.

[0003] However, existing technologies still face several bottlenecks when it comes to state estimation and compensation control of distributed battery systems: First, most methods rely only on a single feature or static model for state estimation, which makes it difficult to adapt to the dynamic evolution of batteries under time-varying working conditions, resulting in insufficient estimation accuracy; second, state prediction generally lacks uncertainty modeling and cannot provide a credible interval to support compensation decisions; third, optimization strategies mostly adopt static allocation or heuristic methods, and lack a multi-objective optimal scheduling mechanism based on state evolution trends; fourth, the safety verification link of the compensation strategy is weak, which may lead to risks such as voltage exceeding the limit and temperature anomalies, restricting the efficient and coordinated operation of distributed battery systems. Summary of the Invention

[0004] Based on the above objectives, the present invention provides a distributed battery state estimation and compensation method based on voltage and current sampling, and a distributed battery state estimation and compensation method for state evolution prediction, integration of multi-dimensional perception and optimization control, so as to improve the system intelligence level and operation safety.

[0005] A distributed battery state estimation and compensation method based on voltage and current sampling includes the following steps:

[0006] S1: Synchronously obtain the voltage and current waveforms of each battery cell, and generate the original feature matrix of ripple characteristics and slope mutation points through the multi-dimensional parameter extraction module;

[0007] S2: Input the original characteristic matrix into the coupling compensator, calculate the dynamic compensation factor according to the difference in the operating conditions of adjacent battery cells, and generate a characteristic tensor with environmental parameter correction;

[0008] S3: constructing an equivalent model network with a time-varying topological structure based on the characteristic tensor, updating the model parameters through an adaptive particle injection algorithm, and generating a state parameter space of electrochemical-thermodynamic coupling characteristics;

[0009] S4: Using a multi-source data fusion engine to perform distributed collaborative calculations on the state parameter space, and outputting the health status index and state of charge confidence interval of each battery cell.

[0010] S5: Based on the health status index and state of charge confidence interval obtained in S4, a multi-state dynamic prediction model of the battery pack is constructed to predict the state evolution trend of each battery cell in multiple future control cycles. The prediction content includes the state of charge (SOC), state of charge (SOH) and temperature change trends.

[0011] S6: Based on the prediction results and operation goals, an energy compensation optimization model is constructed. Based on the state constraints and energy scheduling requirements, an optimization algorithm is used to generate the optimal compensation strategy for each battery unit.

[0012] Optionally, the S1 includes:

[0013] S11: synchronously acquiring the voltage and current waveforms of each battery cell, wherein the voltage signal and current signal of each battery cell are respectively collected by a differential voltage sensor and a Hall effect current sensor, and the voltage and current waveforms are sampled at equal intervals at a preset sampling frequency to generate a discrete voltage and current waveform sequence;

[0014] S12: performing sliding window filtering on the discretized voltage and current waveform sequence, using a Savitzky-Golay filter to eliminate high-frequency noise, and outputting smoothed voltage and current waveform data;

[0015] S13: Inputting the smoothed voltage and current waveform data into a multi-dimensional parameter extraction module, extracting the ripple amplitude and phase information within a preset frequency band (1kHz-10kHz) through fast Fourier transform, and calculating the ripple peak-to-peak value and harmonic distortion rate within each time window to generate a ripple feature;

[0016] S14: In the multi-dimensional parameter extraction module, a first-order difference calculation is performed on the smoothed voltage and current waveform data, and a mutation point detection algorithm based on a dynamic threshold is used. If the absolute value of the difference exceeds three standard deviations of the current window average, it is marked as a slope mutation point, and the position coordinates and amplitude change of each slope mutation point are recorded;

[0017] S15: Align the ripple feature vector with the position coordinates and amplitude variation of the slope mutation point according to the timestamp, and combine them into an original feature matrix containing time domain-frequency domain joint features.

[0018] Optionally, the S2 includes:

[0019] S21: Inputting the original characteristic matrix into a coupling compensator, establishing an associated topological structure of adjacent battery cells based on the physical connection relationship of the battery cells, and determining a set of adjacent cells for each battery cell;

[0020] S22: For each battery cell, calculate the difference in operating condition between the battery cell and all battery cells in the adjacent cell set.

[0021] S23: Generate a dynamic compensation factor according to the operating condition difference, wherein the dynamic compensation factor is negatively correlated with the operating condition difference.

[0022] Optionally, the S2 further includes:

[0023] S24: performing a tensor outer product operation on the dynamic compensation factor and the original characteristic matrix according to the unit dimension, and superimposing the environmental parameter correction term (including the temperature gradient matrix and the internal resistance offset matrix) to generate a characteristic tensor with environmental parameter correction;

[0024] S25: normalizing the feature tensor with environmental parameter correction, eliminating dimensional differences through Z-score normalization, and outputting the normalized feature tensor.

[0025] Optionally, the S3 includes:

[0026] S31: Based on the standardized feature tensor, construct an initial equivalent model network, wherein the equivalent model network is composed of a variable-order RC network, wherein each battery cell corresponds to a network node, and the connection weights between nodes are initialized to the cosine similarity of the standardized feature tensors of adjacent cells;

[0027] S32: Dynamically adjust the topological connection weights of the equivalent model network according to the real-time status parameters of the battery cell, which include SOC, SOH, and temperature.

[0028] S33: Adopting adaptive particle injection algorithm to update parameters of equivalent model network;

[0029] S34: Perform electrochemical-thermodynamic coupling solution on the equivalent model network after particle injection, and generate a set of electrochemical-thermodynamic coupling parameters for each battery cell by combining the single-particle model equation and the heat conduction equation.

[0030] S35: Fusing the electrochemical-thermodynamic coupling parameter set with the topological weight of the equivalent model network to update the state parameter space.

[0031] Optionally, the S4 includes:

[0032] S41: Input the state parameter space into a multi-source data fusion engine, distribute the electrochemical-thermodynamic coupling parameter set and topological weight to the corresponding distributed computing nodes according to the battery cell number, and generate a local state vector for each node;

[0033] S42: In each distributed computing node, local state estimation is performed based on the local state vector, lithium ion concentration distribution and temperature gradient field are iteratively corrected using an extended Kalman filter, and a local health state index estimate and a state of charge probability density function are output;

[0034] S43: Perform global collaborative calculation on the estimated local health status indexes of all distributed computing nodes through a consistency algorithm to obtain a globally consistent health status index.

[0035] Optionally, the S4 further includes:

[0036] S44: In the multi-source data fusion engine, performing Bayesian fusion on the state-of-charge probability density function, generating a joint probability distribution through Monte Carlo sampling, and calculating a state-of-charge confidence interval at a preset confidence level of 95%;

[0037] S45: Perform anomaly detection on the global consistency health state index and the state of charge confidence interval, eliminate battery cell data exceeding a preset deviation threshold, and output a final health state index and state of charge confidence interval.

[0038] Optionally, the S5 includes:

[0039] S51: Based on the health state index and the state of charge confidence interval, construct a multi-state dynamic prediction model for the battery pack, wherein the input of the multi-state dynamic prediction model is the lithium ion concentration distribution, temperature gradient field and internal resistance change rate at the current moment, and the output is the state evolution trend prediction value for the next N control cycles;

[0040] S52: Based on the health state index and the state of charge confidence interval, parameter training is performed on the multi-state dynamic prediction model to generate a reduced-order single particle model embedded with a temperature coupling term and a residual compensation weight.

[0041] S53: Based on the reduced-order single-particle model and the residual compensation weight, perform a multi-step rolling prediction to generate a corrected prediction sequence for the next N control cycles;

[0042] S54: Based on the modified prediction sequence, uncertainty quantification is performed, and the probability distribution of the state evolution trend is generated through Monte Carlo sampling, and the SOC fluctuation range, SOH decay rate range and temperature change bandwidth under the preset confidence level of 95% are extracted.

[0043] S55: Based on the probability distribution, the state evolution trend prediction value and the uncertainty quantification result are encapsulated into a prediction data packet according to the timestamp.

[0044] Optionally, the S6 includes:

[0045] S61: Based on the state evolution trend prediction value and uncertainty quantification result in the prediction data packet, an energy compensation optimization model is constructed. The objective function of the energy compensation optimization model is to minimize energy loss and maximize battery pack life and temperature stability. The constraints include the state of charge confidence interval, health status index threshold and temperature change bandwidth.

[0046] S62: Generate a state-action space of the multi-objective optimization problem based on the objective function and the constraints, where the dimension of the state space is the number of battery cells × the number of prediction cycles, and the action space includes the charge and discharge power distribution ratio and thermal management strategy of each battery cell.

[0047] S63: Using the NSGA-II algorithm to perform multi-objective optimization on the state-action space, screening the Pareto optimal solution set through non-dominated sorting and congestion calculation, and generating a set of candidate compensation strategies.

[0048] S64: Performing dynamic feasibility verification on the candidate compensation strategy set, specifically including:

[0049] Based on the electrochemical-thermodynamic coupling characteristics, candidate compensation strategies are simulated and executed to detect whether voltage exceeding the limit, temperature exceeding the tolerance, or internal resistance mutation is triggered;

[0050] Eliminate strategies that violate the preset safety boundaries and retain the set of compensation strategies that meet the safety constraints as the final feasible strategy set;

[0051] S65: Based on the operating target weight (energy efficiency priority or life priority), the optimal compensation strategy is selected from the set of feasible compensation strategies, and converted into a PWM duty cycle instruction of the bidirectional DC / DC converter of each battery cell and a cooling fan speed control signal, and output to the battery management system for execution.

[0052] Beneficial effects of the present invention:

[0053] This invention uses a multi-dimensional parameter extraction module to simultaneously collect voltage and current waveforms, extracting ripple characteristics and slope mutation points. This, combined with a coupling compensator, dynamically corrects operating condition variability and constructs a characteristic tensor with environmental parameter corrections, enabling state estimation with time-varying adaptive capabilities. By employing an adaptive particle injection algorithm and an electrochemical-thermodynamic coupling model, the accuracy of estimation of key state parameters such as SOC, SOH, and temperature is significantly improved, providing a reliable foundation for subsequent control strategies.

[0054] This paper proposes a multi-state dynamic prediction model that integrates a reduced-order single-particle model with an LSTM residual compensation mechanism to achieve a rolling correction forecast of battery state evolution trends over multiple future control cycles. Combining Monte Carlo sampling with Bayesian uncertainty quantification, it generates confidence-based predictions for the SOC fluctuation range, SOH decay rate, and temperature bandwidth. This mechanism enhances the credibility of predictions, provides data support for dynamic compensation strategies, and effectively supports early intervention and health management scheduling.

[0055] This paper constructs an energy compensation optimization model that integrates prediction results with confidence bounds. Using an improved NSGA-II algorithm, it searches for a Pareto-optimal compensation strategy in the state-action space. Dynamic feasibility verification is then performed using electrochemical-thermodynamic constraints. Through a target weighting mechanism, priorities can be flexibly shifted between energy efficiency, battery life, and temperature stability. The resulting PWM control instructions and thermal management strategy ensure that the compensation process operates within safe boundaries, balancing safety, battery life, and overall system efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0058] Figure 2 Schematic diagram of the S4 process of an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0060] like Figure 1-2 As shown, a distributed battery state estimation and compensation method based on voltage and current sampling includes the following steps:

[0061] S1: Synchronously obtain the voltage and current waveforms of each battery cell, and generate the original feature matrix of ripple characteristics and slope mutation points through the multi-dimensional parameter extraction module;

[0062] S2: The original characteristic matrix is ​​input into the coupling compensator, and the dynamic compensation factor is calculated according to the difference in the operating conditions of adjacent battery cells to generate the characteristic tensor with environmental parameter correction;

[0063] S3: Construct an equivalent model network with a time-varying topological structure based on the characteristic tensor, update the model parameters through an adaptive particle injection algorithm, and generate the state parameter space of electrochemical-thermodynamic coupling characteristics;

[0064] S4: A multi-source data fusion engine is used to perform distributed collaborative calculations on the state parameter space, and output the health status index and state of charge confidence interval of each battery cell.

[0065] S5: Based on the health status index and state of charge confidence interval obtained in S4, a multi-state dynamic prediction model for the battery pack is constructed to predict the state evolution trend of each battery cell in multiple future control cycles. The prediction content includes SOC (state of charge), SOH (state of health) and temperature change trend;

[0066] S6: Based on the prediction results and operation goals, an energy compensation optimization model is constructed. Based on the state constraints and energy scheduling requirements, an optimization algorithm is used to generate the optimal compensation strategy for each battery unit.

[0067] S1 includes:

[0068] S11: synchronously acquiring the voltage and current waveforms of each battery cell, wherein the voltage signal and current signal of each battery cell are respectively collected by a differential voltage sensor and a Hall effect current sensor, and the voltage and current waveforms are sampled at equal intervals at a preset sampling frequency to generate a discrete voltage and current waveform sequence;

[0069] S12: Perform sliding window filtering on the discretized voltage and current waveform sequence, use Savitzky-Golay filter to eliminate high-frequency noise, and output smoothed voltage and current waveform data;

[0070] S13: Input the smoothed voltage and current waveform data into the multi-dimensional parameter extraction module, extract the ripple amplitude and phase information within the preset frequency band (1kHz-10kHz) through fast Fourier transform, and calculate the ripple peak-to-peak value and harmonic distortion rate in each time window to generate ripple characteristics;

[0071] S14: In the multi-dimensional parameter extraction module, a first-order difference calculation is performed on the smoothed voltage and current waveform data. A mutation point detection algorithm based on a dynamic threshold is used. If the absolute value of the difference exceeds three standard deviations of the current window average, it is marked as a slope mutation point, and the position coordinates and amplitude change of each slope mutation point are recorded.

[0072] S15: Align the ripple feature vector with the position coordinates of the slope mutation point and the amplitude change according to the timestamp, and combine them into an original feature matrix containing the time domain-frequency domain joint features.

[0073] S2 includes:

[0074] S21: Inputting the original characteristic matrix into the coupling compensator, establishing the associated topological structure of adjacent battery cells based on the physical connection relationship of the battery cells, and determining the adjacent cell set of each battery cell;

[0075] S22: For each battery cell, calculate the operating condition difference between it and all battery cells in the adjacent cell set, where the operating condition difference is calculated using the following formula:

[0076] Working condition difference ;

[0077] in, 、 Represent the original feature matrices of the current battery cell and the adjacent cells respectively, Represents the Euclidean norm (L2 norm), which is used to measure the distance between two feature matrices. 、 : Respectively represent the temperature parameters of the current battery cell and the adjacent cells, 、 are the internal resistance parameters of the current battery cell and the adjacent cells, 、 、 are the normalized weight coefficients of the original feature difference, temperature difference, and internal resistance difference, respectively, and are used to adjust the weight of the impact of the differences in each dimension on the working condition difference.

[0078] S23: Generate a dynamic compensation factor based on the working condition difference, where the dynamic compensation factor is negatively correlated with the working condition difference. The specific expression is:

[0079] Dynamic compensation factor ;

[0080] in, is the attenuation coefficient, which is used to adjust the sensitivity of the dynamic compensation factor to changes in working condition differences. is a natural exponential function, indicating an exponential growth relationship.

[0081] S2 also includes:

[0082] S24: Performing a tensor outer product operation on the dynamic compensation factor and the original feature matrix according to the unit dimension, and superimposing the environmental parameter correction term (including the temperature gradient matrix and the internal resistance offset matrix) to generate a feature tensor with environmental parameter correction;

[0083] S25: Normalize the feature tensor with environmental parameter correction, eliminate dimensional differences through Z-score normalization, and output the standardized feature tensor.

[0084] S3 includes:

[0085] S31: Based on the normalized feature tensor, an initial equivalent model network is constructed. The equivalent model network consists of a variable-order RC network, where each battery cell corresponds to a network node, and the connection weights between nodes are initialized to the cosine similarity of the normalized feature tensors of adjacent cells;

[0086] The variable-order RC network is initially configured as a 2nd-order structure (i.e., containing the resistor-capacitor submodel R0-R1-C1) and can be dynamically expanded to a 4th-order model based on battery aging to accommodate complex dynamic processes.

[0087] The cosine similarity calculation is based on the last dimension feature vector of each battery cell in the standardized feature tensor, which is used to measure the feature similarity between different cells and thus initialize the connection weights between nodes;

[0088] S32: Based on the real-time status parameters of the battery cells, which include SOC, SOH, and temperature, dynamically adjust the topological connection weights of the equivalent model network. The specific adjustment rules are as follows:

[0089] ;

[0090] in, represents the connection weight between node i and node j in the equivalent model network, is the SOC (state of charge) difference between the battery cells corresponding to node i and node j, is a natural exponential function, k is a scaling factor, which is used to control the sensitivity of the SOC difference to the connection weight. 、 are the normalized feature tensors corresponding to node i and node j, is the Euclidean norm, It is the maximum norm value of all current node feature tensors and is used for normalization.

[0091] The dynamic update of the topology connection weight is performed with 5 control cycles as one adjustment cycle to ensure the timeliness of the network response to changes in the system operation status.

[0092] S33: Adopting the adaptive particle injection algorithm to update the parameters of the equivalent model network, specifically including:

[0093] Particle generation: Based on the Monte Carlo sampling method, 1000 virtual particles are generated. Each particle includes electrochemical parameters and thermodynamic parameters. The electrochemical parameters include the lithium ion diffusion coefficient and the reaction rate constant, and the thermodynamic parameters include the heat generation rate and the heat dissipation coefficient.

[0094] Particle injection: When the error between the predicted value output by the equivalent model network and the actual sampled data satisfies the voltage prediction error greater than 10mV or the temperature prediction error greater than 1°C, the particle swarm is sorted and screened based on the residual covariance matrix estimated by the Kalman filter algorithm, and the optimal particle is selected for injection. The particle injection node corresponds to the battery cell with significant prediction error;

[0095] S34: Solve the electrochemical-thermodynamic coupling of the equivalent model network after particle injection, and generate a set of electrochemical-thermodynamic coupling parameters for each battery cell by combining the single particle model equation and the heat conduction equation;

[0096] The single particle model equation is used to describe the diffusion behavior of lithium ion concentration in electrode particles. The single particle model equation used is:

[0097] Where c is the lithium ion concentration, D is the diffusion coefficient, j is the electrochemical reaction current density, and F is the Faraday constant;

[0098] The heat conduction equation is used to calculate the temperature distribution inside the battery and its changing trend. The heat conduction equation used is: ;

[0099] in, is the density, is the specific heat capacity, T is the temperature, k is the thermal conductivity, is the heat production rate;

[0100] The set of electrochemical-thermodynamic coupling parameters obtained by simultaneous decoupling includes: local lithium ion concentration, instantaneous temperature rise rate, reaction heat and Joule heat terms.

[0101] S35: Fuse the electrochemical-thermodynamic coupling parameter set with the topological weight of the equivalent model network to update the state parameter space;

[0102] The fusion adopts weighted average method, in which the parameter fusion weight of each pair of battery cells is determined by their topological connection weight Determine that the updated state parameter space includes the lithium ion concentration distribution, temperature gradient field and internal resistance change rate of the battery cell;

[0103] Each battery cell corresponds to a 12-dimensional state vector, and the state parameter space is output to the S4 multi-source data fusion engine.

[0104] S4 includes:

[0105] S41: Input the state parameter space into the multi-source data fusion engine, distribute the electrochemical-thermodynamic coupling parameter set and topological weight to the corresponding distributed computing nodes according to the battery cell number, and generate the local state vector of each node;

[0106] Distributed computing nodes are independent computing units on the edge side. Each battery cell corresponds to an embedded edge device (such as an ARM Cortex-M7). The memory resources of this device are pre-allocated to 1.5 times the state parameter space to ensure the stability of data buffering and filtering calculations.

[0107] The local state vector includes the lithium ion concentration distribution (6 dimensions), the temperature gradient field (3 dimensions), and the internal resistance change rate (3 dimensions), totaling 12-dimensional structured feature vectors;

[0108] S42: In each distributed computing node, local state estimation is performed based on the local state vector, and the lithium ion concentration distribution and temperature gradient field are iteratively corrected using an extended Kalman filter to output a local health status index estimate and a state of charge probability density function.

[0109] The state equation of the extended Kalman filter is established in the single particle model framework to dynamically estimate the concentration diffusion and heat conduction states;

[0110] The observation equation is established based on the voltage and current waveform response model, and the real-time sampling waveform is used as a reference for filtering update;

[0111] The state of charge probability density function is fitted by a Gaussian mixture model, and the number of mixtures is set to 3 to reflect the uncertainty distribution of SOC estimation under complex operating conditions;

[0112] The extended Kalman filter iterative process is as follows:

[0113] Prediction steps:

[0114] ;

[0115] ;

[0116] Update steps:

[0117] ;

[0118] ;

[0119] ;

[0120] in, is the current state estimate (including concentration and temperature), is the state covariance matrix, is the state transition function, is the observation model, 、 is the covariance of process noise and measurement noise, is the Kalman gain, is the current measured value, is a nonlinear observation function;

[0121] The final output includes: local health status index estimate and state of charge probability density function;

[0122] S43: Performing a global collaborative calculation of the estimated local health status index values ​​of all distributed computing nodes using a consensus algorithm, specifically including:

[0123] Data exchange: Data communication is carried out between physically adjacent distributed computing nodes to exchange their respective local health status index estimates and their covariance matrices;

[0124] Weight update: The collaborative weighted value is calculated based on the inverse matrix of the covariance matrix, and the consistent health status index is obtained by fusion. The fusion weight is expressed as follows:

[0125] ;

[0126] in, is the covariance matrix of the output of the i-th node, is the estimated value of the local health status index output by the i-th node, is the fusion weight, is the global consistency health status index.

[0127] The S4 also includes:

[0128] S44: In the multi-source data fusion engine, perform Bayesian fusion on the state-of-charge probability density function, generate a joint probability distribution through Monte Carlo sampling, and calculate the state-of-charge confidence interval at a preset confidence level of 95%;

[0129] Bayesian fusion uses the following expression: ;

[0130] in, is the prior SOC distribution, is the likelihood function, is the posterior distribution.

[0131] Monte Carlo sampling uses the importance resampling method. The number of Monte Carlo sampling particles is set to 5000. Random samples are fused on the SOC probability density functions output by multiple nodes to finally generate a joint probability distribution.

[0132] Extract 95% confidence intervals based on the joint distribution: ;

[0133] S45: Perform anomaly detection on the global consistency health state index and the state of charge confidence interval, eliminate battery cell data that exceeds a preset deviation threshold, and output the final health state index and state of charge confidence interval.

[0134] The anomaly detection threshold includes the following two criteria:

[0135] The health status index deviation is greater than 5%;

[0136] The confidence interval width of the state of charge is greater than 3%;

[0137] Once the abnormal condition is met, the battery cell will be marked as an abnormal cell and the compensation strategy generation mechanism in step S6 will be triggered;

[0138] The final output of all normal units will be input into the multi-state dynamic prediction model of the next stage as a trusted data source to support prediction and control decisions in long-term operation.

[0139] S5 includes:

[0140] S51: Based on the health status index and the state of charge confidence interval, a multi-state dynamic prediction model for the battery pack is constructed. The input of the multi-state dynamic prediction model is the current lithium ion concentration distribution, temperature gradient field, and internal resistance change rate. The output is the predicted value of the state evolution trend in the next N control cycles.

[0141] Input variables: 12 dimensions in total, including lithium ion concentration distribution (6 dimensions), temperature gradient field (3 dimensions), and internal resistance change rate (3 dimensions);

[0142] Output variables: SOC, SOH and temperature for each prediction period;

[0143] The model architecture adopts a physical modeling and data-driven fusion strategy to provide a unified platform for subsequent dynamic prediction and compensation;

[0144] S52: Based on the health status index and state of charge confidence interval, the multi-state dynamic prediction model is trained to generate a reduced-order single-particle model with embedded temperature coupling terms and residual compensation weights. Specifically, the model includes:

[0145] (1) Electrochemical characteristic modeling: Based on the lithium ion concentration distribution and time-domain evolution equation, the lithium ion diffusion kinetic equation is generated through a reduced-order single-particle model, and the temperature coupling term is embedded based on the heat generation rate output by the thermodynamic equation;

[0146] The reduced-order single-particle model retains the cathode particle diffusion process: ;

[0147] At the same time, the change of electrolyte concentration is ignored to reduce the calculation complexity of the model.

[0148] The coupled temperature diffusion term is expressed as: ;

[0149] in, is the reference diffusion coefficient, is the activation energy, R is the universal gas constant, and T is the current temperature;

[0150] (2) Machine learning correction: Based on the residual sequence of the historical state of charge confidence interval and the true value, a long short-term memory network (LSTM) is used to generate residual compensation weights. The steps include:

[0151] (2.1) Construct a training dataset to collect the spatial data of the battery's state parameters during its historical operation, including 12 dimensions: lithium ion concentration distribution, temperature gradient field, and internal resistance change rate. At the same time, record the errors between the predicted SOC, SOH, and temperature values ​​at each moment and the actual values, and generate the corresponding residual sequence to form supervised learning samples.

[0152] (2.2) According to a fixed time window, the state parameter vectors of multiple consecutive periods are organized into a sliding input sequence as the input of the LSTM model. The sliding window length can be set according to the prediction accuracy requirements to capture the impact of state evolution on future residual change trends.

[0153] (2.3) Design a long short-term memory network (LSTM) structure, which consists of an input layer for receiving the state parameter sequence, a 64-unit LSTM hidden layer for extracting time-dependent features, and an output layer for outputting three residual compensation weights, corresponding to the prediction correction factors of SOC, SOH, and temperature, respectively.

[0154] (2.4) Based on the historical residual mean as the training target, the LSTM model is iteratively trained to optimize its parameters so that it can automatically learn the error correction trend according to the state parameter sequence, thereby outputting the optimal residual compensation weight and improving the model prediction stability.

[0155] (2.5) During the actual operation of the model, the latest state parameter sequence is input in each control cycle, and the trained LSTM network is called to output the residual compensation weight at the current moment. The SOC, SOH and temperature prediction results are dynamically corrected based on the residual mean of the cycle.

[0156] S53: Based on the reduced-order single-particle model and the residual compensation weight, a multi-step rolling forecast is performed to generate a revised forecast sequence for the next N control cycles, including:

[0157] (1) Initialize the multi-state dynamic prediction model and take the state parameter space in the current control cycle as the initial input, including lithium ion concentration distribution, temperature gradient field and internal resistance change rate, to generate the SOC, SOH and temperature prediction values ​​of the next cycle.

[0158] (2) For each prediction result output by the model, the historical residual mean within the current control period is combined with the residual compensation weight output by LSTM to perform dynamic correction to form the first step of the corrected prediction value. The compensation process is applicable to all three state parameters;

[0159] (3) Feedback the corrected prediction value to the model input, update the state parameter space, and use it as the model input for the next control cycle to achieve temporal recursion and iteration of the model state;

[0160] (4) Repeat the prediction-correction-feedback rolling mechanism until the complete prediction path for the next N control cycles is completed, and a corrected prediction sequence including SOC, SOH, and temperature is obtained;

[0161] (5) During each round of rolling forecasting, the lithium ion concentration and temperature state variables within the model are automatically updated to ensure that each prediction step is based on the revised state of the previous cycle, thereby improving the consistency and accuracy of the forecast sequence;

[0162] S54: Based on the modified prediction sequence, uncertainty quantification is performed. The probability distribution of the state evolution trend is generated through Monte Carlo sampling. The SOC fluctuation range, SOH decay rate range and temperature change bandwidth under the preset confidence level of 95% are extracted. Specifically, the following are included:

[0163] (1) Based on the modified prediction sequence generated above, a Monte Carlo sampling model is constructed. Noisy disturbances are introduced in each control cycle. A large number of state evolution trajectories are formed through repeated sampling. The total number of sampled particles is set to 2000, and the sampling interval is 1 control cycle to construct the time evolution probability space of SOC, SOH and temperature.

[0164] (2) All particle sampling results are statistically processed, and the kernel density estimation method is used to generate continuous probability distribution functions for SOC, SOH and temperature variables respectively. The upper and lower quantiles at the 95% confidence level are extracted from the distribution, and the corresponding SOC fluctuation range, SOH decay rate range and temperature change bandwidth are output as the result indicators of uncertainty quantification.

[0165] S55: Based on the probability distribution, the state evolution trend prediction value and the uncertainty quantification result are encapsulated into a prediction data packet according to the timestamp.

[0166] S6 includes:

[0167] S61: Based on the state evolution trend prediction value and uncertainty quantification results in the prediction data packet, an energy compensation optimization model is constructed. The objective function of the energy compensation optimization model is to minimize energy loss and maximize battery pack life and temperature stability. The constraints include the state of charge confidence interval, health status index threshold, and temperature change bandwidth.

[0168] (1) Extract the predicted SOC, SOH and temperature values ​​of each battery cell in the next N control cycles and the corresponding 95% confidence intervals in the prediction data packet to form the time series state input set required for optimization modeling.

[0169] (2) Set the objective function of the energy compensation optimization model and clarify the optimization direction, including three sub-goals: one is to minimize the total energy loss, the second is to minimize the SOH decay rate to extend the life of the battery pack, and the third is to minimize the temperature change fluctuation to improve thermal stability.

[0170] (3) Define the optimization constraints, including: the SOC must fall within its predicted confidence interval, the health status index cannot be lower than the set threshold, and the temperature change amplitude must not exceed the set bandwidth boundary, to ensure that the optimization results are executed within the safe operating range.

[0171] S62: Generate a state-action space for the multi-objective optimization problem based on the objective function and the constraints, where the dimension of the state space is the number of battery cells × the number of prediction cycles, and the action space includes the charge and discharge power allocation ratio and thermal management strategy of each battery cell;

[0172] (1) Construct the state space, and form a two-dimensional grid with the number of battery cells and the number of predicted cycles. Each grid node contains a state vector that records the SOC, SOH, temperature, and uncertainty boundary information of the cell at that moment.

[0173] (2) Design the action space. The optional actions of each battery unit in each control cycle include the charge and discharge power distribution ratio (such as 0%, 25%, 50%, 75%, 100%) and the thermal management strategy (such as fan speed level, whether liquid cooling is turned on or not), forming a combination strategy set.

[0174] (3) Discretize the state-action mapping function, use a modeler (such as RC model and thermal coupling model) to perform predictive response simulation on the impact of each set of actions on the state space, and construct a state-action transfer model for optimizer evaluation.

[0175] S63: Use the NSGA-II algorithm to perform multi-objective optimization on the state-action space. Filter the Pareto optimal solution set through non-dominated sorting and congestion calculation to generate a set of candidate compensation strategies, including:

[0176] (1) Initialize the NSGA-II population. The individuals in the population are composed of state-action pairs. The objective function value is obtained through simulation to obtain the corresponding energy loss, SOH decay rate and temperature fluctuation range for subsequent sorting.

[0177] (2) Perform non-dominated sorting, determine the dominance level of each individual according to the objective function value, and evaluate the diversity of the solution set through crowding calculation, give priority to retaining boundary solutions and non-dense area solutions, and form a Pareto frontier candidate set.

[0178] (3) Iteratively update the population based on the elite retention strategy, generate a new generation of compensation strategy individuals through crossover and mutation operations, and continuously optimize the solution set until the set algebraic upper limit or convergence threshold is reached.

[0179] S64: Dynamically verify the feasibility of the candidate compensation strategy set, specifically including:

[0180] Based on the electrochemical-thermodynamic coupling characteristics, candidate compensation strategies are simulated and executed to detect whether voltage exceeding the limit, temperature exceeding the tolerance, or internal resistance mutation is triggered;

[0181] Eliminate strategies that violate the preset safety boundaries and retain the set of compensation strategies that meet the safety constraints as the final feasible strategy set;

[0182] S65: Based on the operating target weight (energy efficiency priority or life priority), the optimal compensation strategy is selected from the set of feasible compensation strategies, and converted into a PWM duty cycle instruction of the bidirectional DC / DC converter of each battery cell and a cooling fan speed control signal, and output to the battery management system for execution.

[0183] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0184] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A distributed battery state estimation and compensation method based on voltage and current sampling, characterized in that: The following steps are involved: S1: Synchronously obtain the voltage and current waveforms of each battery cell, and generate the original feature matrix of ripple characteristics and slope mutation points through the multi-dimensional parameter extraction module; S2: Input the original characteristic matrix into the coupling compensator, calculate the dynamic compensation factor according to the difference in the operating conditions of adjacent battery cells, and generate a characteristic tensor with environmental parameter correction; S3: constructing an equivalent model network with a time-varying topological structure based on the characteristic tensor, updating the model parameters through an adaptive particle injection algorithm, and generating a state parameter space of electrochemical-thermodynamic coupling characteristics; S4: Using a multi-source data fusion engine to perform distributed collaborative calculations on the state parameter space, and outputting a health status index and a state of charge confidence interval for each battery cell; S5: Based on the health status index and state of charge confidence interval obtained in S4, a multi-state dynamic prediction model of the battery pack is constructed to predict the state evolution trend of each battery cell in multiple future control cycles. The prediction content includes the state of charge (SOC), state of charge (SOH) and temperature change trends. S6: Based on the prediction results and operation goals, an energy compensation optimization model is constructed. Based on the state constraints and energy scheduling requirements, an optimization algorithm is used to generate the optimal compensation strategy for each battery unit.

2. A distributed battery state estimation and compensation method based on voltage and current sampling according to claim 1, characterized in that: Said S1 comprises: S11: synchronously acquiring the voltage and current waveforms of each battery cell, wherein the voltage signal and current signal of each battery cell are respectively collected by a differential voltage sensor and a Hall effect current sensor, and the voltage and current waveforms are sampled at equal intervals at a preset sampling frequency to generate a discrete voltage and current waveform sequence; S12: performing sliding window filtering on the discretized voltage and current waveform sequence, using a Savitzky-Golay filter to eliminate high-frequency noise, and outputting smoothed voltage and current waveform data; S13: Inputting the smoothed voltage and current waveform data into a multi-dimensional parameter extraction module, extracting the ripple amplitude and phase information within a preset frequency band through fast Fourier transform, and calculating the ripple peak-to-peak value and harmonic distortion rate within each time window to generate a ripple feature; S14: In the multi-dimensional parameter extraction module, a first-order difference calculation is performed on the smoothed voltage and current waveform data, and a mutation point detection algorithm based on a dynamic threshold is used. If the absolute value of the difference exceeds three standard deviations of the current window average, it is marked as a slope mutation point, and the position coordinates and amplitude change of each slope mutation point are recorded; S15: Aligning the ripple feature with the position coordinates and amplitude variation of the slope mutation point according to the timestamp, and combining them into an original feature matrix containing time domain-frequency domain joint features.

3. The distributed battery state estimation and compensation method based on voltage and current sampling according to claim 2 is characterized in that: The S2 includes: S21: Inputting the original characteristic matrix into a coupling compensator, establishing an associated topological structure of adjacent battery cells based on the physical connection relationship of the battery cells, and determining a set of adjacent cells for each battery cell; S22: For each battery cell, calculating the degree of difference in operating condition between the battery cell and all battery cells in the adjacent cell set; S23: Generate a dynamic compensation factor according to the operating condition difference, wherein the dynamic compensation factor is negatively correlated with the operating condition difference.

4. A distributed battery state estimation and compensation method based on voltage and current sampling according to claim 3, characterized in that: Said S2 further comprises: S24: performing a tensor outer product operation on the dynamic compensation factor and the original feature matrix according to the unit dimension, and superimposing the environmental parameter correction term to generate a feature tensor with environmental parameter correction; S25: normalizing the feature tensor with environmental parameter correction, eliminating dimensional differences through Z-score normalization, and outputting the normalized feature tensor.

5. A distributed battery state estimation and compensation method based on voltage and current sampling according to claim 4, characterized in that: The S3 includes: S31: Based on the standardized feature tensor, construct an initial equivalent model network, wherein the equivalent model network is composed of a variable-order RC network, wherein each battery cell corresponds to a network node, and the connection weights between nodes are initialized to the cosine similarity of the standardized feature tensors of adjacent cells; S32: Dynamically adjust the topological connection weights of the equivalent model network according to the real-time state parameters of the battery cell, which include SOC, SOH, and temperature; S33: Adopting adaptive particle injection algorithm to update parameters of equivalent model network; S34: Solve the electrochemical-thermodynamic coupling of the equivalent model network after particle injection, and generate a set of electrochemical-thermodynamic coupling parameters for each battery cell by combining the single particle model equation and the heat conduction equation; S35: Fusing the electrochemical-thermodynamic coupling parameter set with the topological weight of the equivalent model network to update the state parameter space.

6. A distributed battery state estimation and compensation method based on voltage and current sampling according to claim 5, characterized in that: The S4 includes: S41: Input the state parameter space into a multi-source data fusion engine, distribute the electrochemical-thermodynamic coupling parameter set and topological weight to the corresponding distributed computing nodes according to the battery cell number, and generate a local state vector for each node; S42: In each distributed computing node, local state estimation is performed based on the local state vector, lithium ion concentration distribution and temperature gradient field are iteratively corrected using an extended Kalman filter, and a local health state index estimate and a state of charge probability density function are output; S43: Perform global collaborative calculation on the estimated local health status indexes of all distributed computing nodes through a consistency algorithm to obtain a globally consistent health status index.

7. A distributed battery state estimation and compensation method based on voltage and current sampling according to claim 6, characterized in that: Said S4 further comprises: S44: In the multi-source data fusion engine, performing Bayesian fusion on the state-of-charge probability density function, generating a joint probability distribution through Monte Carlo sampling, and calculating a state-of-charge confidence interval at a preset confidence level of 95%; S45: Perform anomaly detection on the global consistency health state index and the state of charge confidence interval, eliminate battery cell data exceeding a preset deviation threshold, and output a final health state index and state of charge confidence interval.

8. The distributed battery state estimation and compensation method based on voltage and current sampling according to claim 7 is characterized in that: The S5 includes: S51: Based on the health state index and the state of charge confidence interval, construct a multi-state dynamic prediction model for the battery pack, wherein the input of the multi-state dynamic prediction model is the lithium ion concentration distribution, temperature gradient field and internal resistance change rate at the current moment, and the output is the state evolution trend prediction value for the next N control cycles; S52: Based on the health state index and the state of charge confidence interval, perform parameter training on the multi-state dynamic prediction model to generate a reduced-order single-particle model embedded with a temperature coupling term and a residual compensation weight; S53: Based on the reduced-order single-particle model and the residual compensation weight, perform a multi-step rolling prediction to generate a corrected prediction sequence for the next N control cycles; S54: Based on the modified prediction sequence, uncertainty quantification is performed, a probability distribution of the state evolution trend is generated through Monte Carlo sampling, and the SOC fluctuation range, SOH decay rate range and temperature change bandwidth under a preset confidence level of 95% are extracted; S55: Based on the probability distribution, the state evolution trend prediction value and the uncertainty quantification result are encapsulated into a prediction data packet according to the timestamp.

9. The distributed battery state estimation and compensation method based on voltage and current sampling according to claim 8, characterized in that: The S6 includes: S61: Based on the state evolution trend prediction value and uncertainty quantification result in the prediction data packet, construct an energy compensation optimization model, wherein the objective function of the energy compensation optimization model is to minimize energy loss and maximize battery pack life and temperature stability, and the constraints include the state of charge confidence interval, the health state index threshold, and the temperature change bandwidth; S62: Generate a state-action space for the multi-objective optimization problem based on the objective function and the constraints, where the state space dimension is the number of battery cells × the number of prediction cycles, and the action space includes the charge and discharge power allocation ratio and thermal management strategy of each battery cell; S63: Using the NSGA-II algorithm to perform multi-objective optimization on the state-action space, screening the Pareto optimal solution set through non-dominated sorting and congestion calculation, and generating a set of candidate compensation strategies; S64: Performing dynamic feasibility verification on the candidate compensation strategy set, specifically including: Based on the electrochemical-thermodynamic coupling characteristics, candidate compensation strategies are simulated and executed to detect whether voltage exceeding the limit, temperature exceeding the tolerance, or internal resistance mutation is triggered; Eliminate strategies that violate the preset safety boundaries and retain the set of compensation strategies that meet the safety constraints as the final feasible strategy set; S65: Select the optimal compensation strategy from the set of feasible compensation strategies based on the operating target weight, convert it into a PWM duty cycle instruction of the bidirectional DC / DC converter of each battery cell and a cooling fan speed control signal, and output it to the battery management system for execution.

Images