Model optimization method for neural network for state estimation of energy storage battery

Abstract

This invention relates to the field of model optimization technology and discloses a model optimization method for neural networks used in energy storage battery state estimation. The method includes: first, acquiring module observation signals and performing coordinate transformation; then, extracting the principal mode direction of the pressure array and removing temperature interference to obtain assembly pressure parameters free from thermal expansion effects; next, learning the reversible expansion baseline and integrating the pressure residual along the charge flux to obtain the clamping compliance memory; subsequently, converting this memory into a memory-gated time hidden layer of the neural network and combining it with the spatial propagation mechanism of adjacent modules to output state estimates; simultaneously, constructing a multi-field mechanism residual layer based on the state estimates to complete offline optimization without fixed weights; finally, recursively calculating the memory in online operation, updating the adaptation parameters using unsupervised real-time residuals, and outputting the state results and uncertainty indices. This scheme effectively improves the adaptability and accuracy of state estimation under complex assembly conditions.

Description

Technical Field

[0001] This invention relates to the field of model optimization technology, and more specifically, to a model optimization method for neural networks used for state estimation of energy storage batteries. Background Technology

[0002] In energy storage battery systems, cells are typically constrained into a densely stacked form by devices such as end plates, tie rods, cooling plates, and compression pads. During long-term charge and discharge operation, cells not only undergo reversible expansion but also irreversible expansion due to factors such as aging, side reactions, gas generation, and interfacial film growth. Meanwhile, the compression pads in the clamping device also continuously experience elastic and viscoelastic relaxation and cyclic softening.

[0003] Therefore, the stacking pressure experienced at the module endplate is not solely determined by the current state of charge or health, but is influenced by the current electrochemical state, past clamping history, compression pad flexibility, and thermal expansion. Existing technologies using neural networks for energy storage battery state estimation typically use only voltage, current, temperature, or directly acquired pressure amplitudes as input signals, and employ conventional absolute error or root mean square error as optimization targets.

[0004] This conventional approach fails to adequately consider the historical bias of the clamping boundary, easily misinterpreting historical bias caused by loosening of the clamping device or changes in preload as aging changes in the battery cell itself. This deficiency is particularly evident in energy storage scenarios, where energy storage systems operate in a shallow cycle state for extended periods and experience frequent daily charging and discharging. Inconsistent module replacement cycles within the rack mean that slow changes in the clamping boundary continuously alter the distribution of signals received by the neural network. Consequently, the same batch of signal sequences exhibits significant drift in state of charge or health estimation under different preloads, varying degrees of compression pad aging, or different endplate stiffnesses, making it difficult to meet the system's requirement for long-term, high-precision status monitoring. Summary of the Invention

[0005] This invention provides a model optimization method for neural networks used in energy storage battery state estimation, which solves the technical problems mentioned in the background art.

[0006] This invention provides a model optimization method for neural networks used in energy storage battery state estimation, applicable to energy storage battery systems comprising battery cells and clamping devices, including:

[0007] Acquire module observation signals covering multiple consecutive sampling times, and extract normalized ampere-hour charge characteristics and thermal field gradient characteristics at each time point;

[0008] The main force spatial modes of the pressure sensor array are extracted and temperature interference is removed to obtain the assembly pressure characteristic parameters at each time point after removing the influence of thermal expansion.

[0009] Integrating the assembly pressure characteristic parameter and the residual term of the reversible expansion baseline along the cumulative charge and discharge capacity, the clamping flexibility memory amount, which includes the historical cumulative effect, is recursively obtained.

[0010] The clamping flexibility memory is converted into memory gating, injected into the neural network, and the time hidden state is transmitted between time steps. Combined with the spatial propagation mechanism, the estimated parameters of the battery at each time step are output. The estimated parameters include the state of charge, health state, and core temperature.

[0011] Based on the estimated parameters, a mechanism residual term is constructed that covers terminal voltage, charge dynamics, thermal equilibrium, and mechanical boundary characteristics;

[0012] By combining the state supervision error of the estimated parameters with the mechanism residuals of each term, an overall objective function based on trainable weight coefficients and adaptive regularization terms is constructed to complete the joint offline optimization of basic network parameters and adaptive parameters.

[0013] During online operation, the clamping flexibility memory is continuously recursively calculated, and the adaptive parameters are updated unsupervisedly only using the real-time mechanism residual term, outputting the state estimation results and uncertainty index.

[0014] The beneficial effects of this invention are as follows: by converting the historical bias and compliance changes of the endplate pressure into calculable memory parameters and injecting them into a neural network to dynamically adjust the boundary of the mechanism mapping, the invention effectively avoids the model misjudging the assembly history slack of clamping components as cell aging or state drift. This invention enables the model to continuously output highly reliable estimates of state of charge, state of health, and core temperature under complex operating conditions such as long-term operation, preload decay, and irreversible expansion, significantly improving the adaptability and long-term evaluation accuracy of the energy storage battery management system under different assembly conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart of the model optimization method for neural networks used in energy storage battery state estimation according to the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0017] like Figure 1As shown, a model optimization method for neural networks used in energy storage battery state estimation is applied to an energy storage battery system comprising cells and clamping devices, including:

[0018] Acquire module observation signals covering multiple consecutive sampling times, and extract normalized ampere-hour charge characteristics and thermal field gradient characteristics at each time point;

[0019] The main force spatial modes of the pressure sensor array are extracted and temperature interference is removed to obtain the assembly pressure characteristic parameters at each time point after removing the influence of thermal expansion.

[0020] Integrating the assembly pressure characteristic parameter and the residual term of the reversible expansion baseline along the cumulative charge and discharge capacity, the clamping flexibility memory amount, which includes the historical cumulative effect, is recursively obtained.

[0021] The clamping flexibility memory is converted into memory gating, injected into the neural network, and the time hidden state is transmitted between time steps. Combined with the spatial propagation mechanism, the estimated parameters of the battery at each time step are output. The estimated parameters include the state of charge, health state, and core temperature.

[0022] Based on the estimated parameters, a mechanism residual term is constructed that covers terminal voltage, charge dynamics, thermal equilibrium, and mechanical boundary characteristics;

[0023] By combining the state supervision error of the estimated parameters with the mechanism residuals of each term, an overall objective function based on trainable weight coefficients and adaptive regularization terms is constructed to complete the joint offline optimization of basic network parameters and adaptive parameters.

[0024] During online operation, the clamping flexibility memory is continuously recursively calculated, and the adaptive parameters are updated unsupervisedly only using the real-time mechanism residual term, outputting the state estimation results and uncertainty index.

[0025] This implementation is applied to an energy storage battery system that includes cells and clamping devices. Each energy storage module in the system is equipped with a current acquisition unit, a voltage acquisition unit, a surface temperature sensing array, an end plate pressure sensing array, and an ambient temperature acquisition unit. All acquisition units are aligned in time by a synchronous clock, and the synchronization error does not exceed 10% of the sampling period.

[0026] For any module m in the energy storage system and any current time k in the discrete sampling sequence, calculate the system timestamp corresponding to the current time. The system timestamp corresponding to the previous sampling time The difference is used to obtain the adjacent sampling interval. The calculation formula is:

[0027]

[0028] Wherein, module m is the smallest energy storage unit in the energy storage battery system with an independent cell stacking structure and clamping device, and each module has a unique physical identifier; k is the time index of the discrete sampling sequence, and its value is a positive integer; This is the system timestamp corresponding to the k-th sampling time, in seconds; The time interval between two adjacent sampling times is measured in seconds. The value range is determined by the system sampling frequency, with a typical range of 0.1 seconds to 10 seconds.

[0029] Based on adjacent sampling intervals The module current sampled by module m at the k-th sampling time. The nominal capacity of module m Introducing a time dimension conversion constant Calculate the increment of electricity change at the k-th sampling time. The calculation formula is:

[0030]

[0031] in, Let be the charging and discharging current of module m at the kth sampling time, in amperes. The current sign convention is that the value is positive during the discharging process, negative during the charging process, and 0 during the resting process. The nominal ampere-hour capacity of module m, in ampere-hours, is determined by the cell's factory parameters and the series-parallel topology of the cells within the module. It is a time dimension conversion constant used to convert second-level time intervals into hour units. It is dimensionless and has a fixed value of 1 / 3600. It is the normalized increment of the charge change at the k-th sampling time relative to the previous time. It is dimensionless and represents the relative change of the module's charging / discharging capacity relative to its nominal capacity at the current time.

[0032] By setting an initial energy baseline at the start of the sampling sequence and accumulating the energy change increments time-by-time, the normalized ampere-hour energy characteristic of module m at the k-th sampling time is obtained. The calculation formula is:

[0033]

[0034] The initial energy reference is the normalized ampere-hour energy characteristic corresponding to the starting time k=0 of the sampling sequence. , dimensionless, with a value range of [0,1], is obtained by mapping the full charge calibration state or the static open circuit voltage at the start of the module; Let be the normalized ampere-hour capacity characteristic of module m at the k-th sampling time. It is dimensionless and represents the ratio of the cumulative charge and discharge capacity of the module from the sampling start time to the current time to the nominal capacity.

[0035] The average surface temperature of the module is calculated based on the temperature sensor array deployed on the surface of module m at the k-th sampling time. The calculation formula is:

[0036]

[0037] in, The total number of temperature sensing points deployed on the surface of module m, which is a positive integer and is determined by the module size and temperature monitoring requirements. The typical value range is 4 to 16. The temperature reading collected by module m at the k-th sampling time and the s-th temperature sensing point is expressed in Kelvin. Let be the average surface temperature of module m at the k-th sampling time, in Kelvin.

[0038] Construct a spatial physical adjacency weight matrix based on the physical topology of temperature sensing points. Multiply the temperature sensing reading vector at the k-th sampling time with the spatial physical adjacency weight matrix to obtain the thermal field gradient characteristics of module m at the k-th sampling time. The calculation formula is:

[0039]

[0040] in, For dimension The spatial physical adjacency weight matrix is ​​dimensionless. The values ​​of the elements in the matrix are determined by the physical installation position of the temperature sensing points. The matrix elements corresponding to two adjacent sensing points have non-zero positive values, while the matrix elements corresponding to non-adjacent sensing points have 0 values. The sum of the elements in each row of the matrix is ​​normalized to 1. For dimension The column vector, in Kelvin, represents the temperature spatial distribution gradient characteristics of module m at the k-th sampling time.

[0041] For the original input parameters of each feature channel, based on the dispersion statistics of the historical training dataset, scale normalization is performed on the input parameters at the current time step to obtain the robust normalized features of the corresponding feature channel. The calculation formula is:

[0042]

[0043] Where 'a' is the feature channel index, covering all input feature channels including module current, module terminal voltage, module average temperature, thermal field gradient features, assembly pressure feature parameters after removing the influence of thermal expansion, clamping flexibility memory, and clamping flexibility memory time difference value. Let be the original input parameter of the a-th feature channel at the k-th sampling time, and its dimensions are determined by the physical properties of the corresponding feature channel; It is a median calculation operator, calculated based on all valid data of the corresponding feature channel in the historical training dataset D, and its dimensions are consistent with the original input parameters of the corresponding feature channel; Let be the minimum bias constant corresponding to the a-th feature channel, whose dimensions are consistent with the square of the difference between the original input parameters of the corresponding feature channel, and whose value is . Multiply by the squared range of the corresponding feature channel in the historical training dataset D to avoid the abnormal operation of the denominator being 0; Let be the robust normalized feature of the 'a'-th feature channel at the 'k'-th sampling time, which is dimensionless. It is the set of all original input parameters for the a-th feature channel in the historical training dataset D.

[0044] Subtract the median baseline of the corresponding pressure sensing point in the historical training dataset from the pressure reading vector collected by the pressure sensing array at the k-th sampling time on the m-endplate of module m to obtain the centered pressure vector. The calculation formula is:

[0045]

[0046] in, For dimension Pressure sensing array vector, The total number of pressure sensing points deployed on the m-end plate of module is a positive integer, with a typical range of 4 to 12. The elements in the vector are the pressure readings collected by the corresponding pressure sensing points at the k-th sampling time, in Pascals. For dimension The median baseline vector is calculated based on all valid data of the corresponding pressure sensor points in the historical training dataset D, and the unit is Pascal. For dimension The centered pressure vector, in Pascals, is used to eliminate the fixed reference bias of the pressure sensing signal.

[0047] The central pressure vectors at all observation times are stacked row-wise to form a sequence matrix. Singular value decomposition is performed on the sequence matrix, and the right singular vector corresponding to the largest singular value is extracted as the main force space mode. The calculation formula is:

[0048]

[0049]

[0050] Wherein, SVD is the singular value decomposition operator; The sequence matrix is ​​formed by stacking the centered pressure vectors at K consecutive observation times in rows, with dimension . K is the total length of the observation sequence, and its value is a positive integer; For dimension The left singular matrix, For dimension A singular value diagonal matrix, where the elements on the diagonal are arranged in descending order. For dimension The right singular matrix; For dimension A unit column vector, with the first element being 1 and the rest being 0, is used to extract the right singular vector corresponding to the maximum singular value; For dimension The main force spatial mode vector is dimensionless and represents the dominant spatial characteristics of the pressure distribution on the module end plate.

[0051] Multiplying the transpose of the main stress spatial mode with the central pressure vector at the current moment yields the main mode pressure scalar of module m at the k-th sampling moment. The calculation formula is:

[0052]

[0053] in, The main force-bearing spatial mode The transpose row vector with dimension . ; Let be the dominant mode pressure scalar of module m at the k-th sampling time, in Pascals, representing the amplitude corresponding to the dominant spatial mode of force on the end plate.

[0054] Introducing temperature and time reference bases, the bias reference term, dimensionless average temperature, and dimensionless average temperature change rate are sequentially concatenated to form the temperature influence interpretation vector. The calculation formula is:

[0055]

[0056] in, This is a temperature reference standard, with the unit being Kelvin, and a fixed value of 298.15 Kelvin. It serves as a time reference standard, with the unit being seconds, and a fixed value of 3600 seconds. For dimension A column vector, with all elements being dimensionless, is used to characterize the set of linear influence factors of temperature on pressure signals.

[0057] Based on historical training datasets, the temperature stripping coefficient that minimizes the sum of squared errors between the principal modal pressure scalar and the fitted value is obtained through least squares parameter optimization. The calculation formula is:

[0058]

[0059] in, For dimension The temperature stripping coefficient column vector is used, with all elements being dimensionless; arg min is the minimization operator, and parameter optimization is achieved through linear least squares fitting; the summation operation is performed based on all valid sampling times within the historical training dataset D, b T It is the transpose of the temperature stripping coefficient vector b, with a dimension of 1×3 and dimensionless.

[0060] Subtracting the product of the temperature stripping coefficient and the temperature effect interpretation vector from the principal modal pressure scalar yields the assembly pressure characteristic parameter of module m at the k-th sampling time, after removing the thermal expansion effect. The calculation formula is:

[0061]

[0062] in, The assembly pressure characteristic parameter of module m at the k-th sampling time after removing the influence of thermal expansion is expressed in Pascals. It represents the effective signal of cell stacking assembly pressure after eliminating temperature-related thermal expansion interference.

[0063] Based on the increment of charge change at the k-th sampling time, the charging and discharging flux direction parameter is extracted. The calculation formula is:

[0064]

[0065] Wherein, sgn is the sign function, which outputs 1 when the input value is greater than 0, -1 when the input value is less than 0, and 0 when the input value is equal to 0; This is a dimensionless flux direction parameter used to characterize the direction of charge flow during charging and discharging at the current moment. It takes a value of 1 during the discharging process, a value of -1 during the charging process, and a value of 0 during the resting process.

[0066] The normalized ampere-hour charge characteristics, dimensionless average temperature, and flux direction parameters are combined to form the input parameters of the reversible expansion baseline mapping model. The calculation formula is:

[0067]

[0068] in, For dimension The input parameter column vector, with all elements being dimensionless, serves as the input variable for the reversible dilatational baseline mapping model.

[0069] Based on historical training datasets, a cost function containing fitting error and smoothness constraints is constructed. The minimum value of the cost function is obtained by solving the gradient descent optimization algorithm, thus yielding a reversible dilated baseline mapping model. Optimal parameters The cost function is calculated using the following formula:

[0070]

[0071] in, For the reversible dilated baseline mapping model, a three-layer fully connected neural network structure is adopted, with 3 neurons in the input layer, 16 neurons in the hidden layer, and the activation function is the hyperbolic tangent function. The output layer has 1 neuron and the activation function is a linear function. This is the set of trainable parameters for the mapping model, including the weight matrix and bias vector of the fully connected layers; The pressure reference is in Pascals, and its value is the maximum absolute value of the principal modal pressure scalar within the historical training dataset D. The fitting error weighting factor is dimensionless, initially set to 0, and updated synchronously during the optimization process. is the parameter smoothness constraint factor, which is dimensionless, initially set to 0, and updated synchronously during the optimization process; For input parameters The second-order partial derivative operator; It is the Frobenius norm; Input parameters The empirical probability distribution measure on the historical training dataset D is obtained through Gaussian kernel density estimation; the integral term is implemented using the Monte Carlo numerical integration method, with 1000 sampling points randomly generated based on the input parameter distribution within the historical training dataset D; the optimization algorithm employs the Adam optimizer, with an initial learning rate of... The maximum number of iterations is 10,000, and the convergence threshold is... Training is terminated prematurely when the change in the cost function over 50 consecutive iterations is less than the convergence threshold.

[0072] By inputting the baseline input parameters at the current time into the trained reversible dilated baseline mapping model, the reversible dilated baseline of module m at the k-th sampling time is obtained. The calculation formula is:

[0073]

[0074] in, Let be the reversible expansion baseline of module m at the kth sampling time, in Pascals, representing the pressure reference value corresponding to the expansion of the cell caused by the reversible electrochemical reaction under the current state of charge, temperature, and charge / discharge direction.

[0075] Subtracting the reversible expansion baseline at the current moment from the assembly pressure characteristic parameter after removing the influence of thermal expansion yields the pressure residual term of module m at the k-th sampling moment. The calculation formula is:

[0076]

[0077] in, The pressure residual term of module m at the kth sampling time is expressed in Pascals. It represents the pressure deviation signal caused by slow-varying effects such as changes in the flexibility of the clamping structure and irreversible expansion of the battery cell after removing the influence of reversible expansion.

[0078] The initial value of the integral is set to 0. The pressure residual accumulated along the capacity at the previous time step is added to the product of the pressure residual term and the change in electrical quantity at the current time step. The pressure residual accumulated along the capacity at the k-th sampling time of module m is then obtained recursively. The calculation formula is:

[0079]

[0080] Wherein, the initial value of the integral is the value corresponding to the starting time k=0 of the sampling sequence. The unit is Pascal; Let be the cumulative integral value of the pressure residual along the capacity of module m at the k-th sampling time, in Pascals, representing the cumulative effect of pressure residual with charge and discharge capacity.

[0081] The initial throughput value is set to 0. The cumulative power throughput of module m at the k-th sampling time is obtained by adding the absolute value of the power change increment at the current time to the cumulative power throughput of the previous time step. The calculation formula is:

[0082]

[0083] Wherein, the initial throughput value is the value corresponding to the starting time k=0 of the sampling sequence. Dimensionless; Let be the cumulative charge throughput of module m at the k-th sampling time. It is dimensionless and represents the total charge / discharge capacity cycle from the sampling start time to the current time.

[0084] Dividing the current pressure residual along the capacity accumulation integral value by the sum of the current power throughput accumulation and the minimum bias constant, we obtain the clamping flexibility memory of module m at the k-th sampling time. The calculation formula is:

[0085]

[0086] in, The smallest bias constant is dimensionless and takes the value of This is used to avoid arithmetic errors where the denominator is 0; The clamping flexibility memory of module m at the kth sampling time is expressed in Pascals. It represents the historical cumulative flexibility change characteristics of the module clamping structure and the cell stacking system, including comprehensive memory information of slow-changing effects such as preload decay, compression pad aging, and irreversible cell expansion.

[0087] Calculate the time difference value based on the clamping flexibility memory between the current time and the previous time. The calculation formula is:

[0088]

[0089] in, The time difference value of the clamping flexibility memory, in Pascals, represents the instantaneous rate of change of the clamping flexibility memory.

[0090] Robustly normalized features from all feature channels, along with the dimensionless clamping flexibility memory and its time difference robustly normalized features, are concatenated to form the model input sequence vector. The calculation formula is:

[0091]

[0092] in, This represents the robust normalization characteristic of the module current. This represents the robust normalization characteristic of the module terminal voltage. This represents the robust normalization characteristic of the module's average temperature. This is a robust normalized feature of the thermal field gradient. Robust normalization characteristics of assembly pressure characteristic parameters to remove the effects of thermal expansion. To provide robust normalization features for holding flexibility memory, Robust normalization features for holding the time difference of flexibility memory; all features are dimensionless parameters; For dimension The model input sequence vector is dimensionless and serves as the input to the time series encoder.

[0093] A time series encoder is constructed using gated recurrent units. The current time step model input sequence vector and the temporal hidden state from the previous time step are input into the encoder to obtain the current time step temporal hidden state. The calculation formula is:

[0094]

[0095] in, For gated loop unit encoders, These are the trainable basic network parameters for the encoder; the encoder adopts a single-layer unidirectional structure, the hidden layer dimension H is fixed at 64, the activation function is the hyperbolic tangent function, and the activation functions for the reset gate and update gate are the sigmoid function; The temporal hidden state at the (k-1)th sampling time has a dimension of Dimensionless; the initial hidden state is the state corresponding to the starting time k=0 of the sampling sequence. A vector with all zero values ​​and dimension 1. Dimensionless; The temporal hidden state at the k-th sampling time has a dimension of , dimensionless, characterizes the coding features of the time-series observation signal of the module.

[0096] The first correlation weight matrix is ​​used to map the combination of dimensionless clamping flexibility memory, time difference, and corresponding dimensionless pressure features. After superimposing the bias vector and processing with an activation function, a memory gating is generated. The calculation formula is:

[0097]

[0098] in, This is the first correlation weight matrix, with dimension . Dimensionless; Let be the bias vector, with dimension . Dimensionless; It is a sigmoid activation function with an output value range of (0,1) and is dimensionless. For dimension The memory gating vector is dimensionless and is used to control the injection strength of the clamping flexibility memory into the temporal hidden state.

[0099] The memory-gated state is then multiplied element-wise with the feature branch vector processed by the second correlation weight matrix, and this multiplication is added to the current time-stack hidden state to form the memory-injected hidden state. The calculation formula is:

[0100]

[0101] in, This is the second correlation weight matrix, with dimension . Dimensionless; This is the Hadamard product operator, which is the element-wise multiplication of corresponding elements of two vectors. For dimension The memory-injected hidden state is dimensionless and represents the temporal coding features that integrate the clamping flexibility historical memory information.

[0102] Based on the physical adjacency relationship between adjacent modules and the cosine similarity of the memory-injected hidden state, the spatial propagation weight between module m and its neighboring module n at the k-th sampling time is calculated. The calculation formula is:

[0103]

[0104] in, This is an identifier for the physical adjacency relationship between module m and module n. The value is 1 when the two modules are physically adjacent in the energy storage rack, and 0 otherwise. Identifier of physical adjacency with module m The set consisting of all adjacent modules; This is a cosine similarity calculation operator with an output value range of [-1, 1], which is dimensionless. This is a temperature coefficient, dimensionless, with a fixed value of 0.1, used to adjust the smoothness of the weight distribution; Let be the spatial propagation weight between module m and its neighboring module n at the k-th sampling time. This weight is dimensionless, and the sum of the weights corresponding to all neighboring modules is normalized to 1. For edge modules without neighboring modules, Since it is an empty set, the sum of the spatial propagation weights is 0, and there is no need to perform feature aggregation of adjacent modules.

[0105] The memory-injected hidden states of adjacent modules are weighted and aggregated based on spatial propagation weights. After being superimposed with the memory-injected hidden state of the current module, they are processed by feature fusion activation to generate the aggregated global hidden state. The calculation formula is:

[0106]

[0107] in, The spatial feature mapping weight matrix has dimensions of . Dimensionless; It is a linear rectifier activation function with leakage, and the slope of the negative half axis is fixed at 0.1, which is dimensionless; For dimension The aggregated global hidden state is dimensionless and represents the global encoding features that integrate temporal memory information and spatial association information of adjacent modules.

[0108] A fully connected layer is used to construct the state decoder, which aggregates the global hidden state input to the decoder and outputs the battery estimation parameters of module m at the k-th sampling time through separate channels. including state of charge Health status Core temperature The calculation formula is:

[0109]

[0110]

[0111]

[0112]

[0113] in, , , Here is the weight matrix of the decoder output layer, with dimensions 1. Dimensionless; , , Let be the bias vector of the decoder output layer, with dimensions 1. ,in , Dimensionless The unit is Kelvin; It is a sigmoid activation function with an output value range of (0,1), which is dimensionless and is used to constrain the range of values ​​of charged state and healthy state to conform to physical meaning; Let be the estimated state of charge of module m at the k-th sampling time. It is dimensionless and takes values ​​in the range [0,1]. Let be the health status estimate of module m at the k-th sampling time. It is dimensionless and its value ranges from [0,1]. Let be the estimated core temperature of the battery cell for module m at the k-th sampling time, in Kelvin; For dimension The estimated parameter column vector.

[0114] Based on the calibrated open-circuit voltage mapping function, equivalent internal resistance mapping function, and polarization dynamic model, the estimated terminal voltage of module m at the k-th sampling time is calculated. And recursively calculate the polarization memory vector for the next time step. The calculation formula is:

[0115]

[0116]

[0117] in, The open-circuit voltage mapping function is constructed based on the correspondence between open-circuit voltage, state of charge, and temperature obtained from cell calibration tests. It is implemented using a two-dimensional lookup table method combined with linear interpolation. The input is the state of charge and average temperature, and the output is the cell open-circuit voltage under the corresponding operating conditions, in volts. The battery equivalent internal resistance mapping function is obtained based on the hybrid pulse power characteristic test calibration. It is implemented by a two-dimensional lookup table method combined with linear interpolation. The input is the healthy state and average temperature, and the output is the battery DC internal resistance under the corresponding operating condition, in ohms. This is the polarization voltage drop mapping function, which calculates the polarization voltage drop based on the polarization memory vector, with the unit being volts, and is implemented using a linear mapping. For dimension The polarization memory vector, dimensionless, corresponds to the polarization capacitor voltage state in the second-order polarization equivalent circuit model, with initial values... It is a vector of all zeros; , The polarization transfer parameter matrix has dimensions of 1. and Dimensionless, constructed based on calibrated polarization time constant and impedance parameters, modulated by state of charge and average temperature, and implemented using lookup table method combined with linear interpolation; This is the estimated terminal voltage of module m at the k-th sampling time, in volts.

[0118] The actual terminal voltage collected by module m at the kth sampling time The terminal voltage mechanism residual term is obtained by subtracting the terminal voltage prediction from the terminal voltage prediction. The calculation formula is:

[0119]

[0120] in, Let be the actual terminal voltage collected by module m at the kth sampling time, in volts; The terminal voltage mechanism residual term, in volts, represents the deviation between the estimated and measured terminal voltage values ​​and is used to constrain the model output to conform to the battery electrical mechanism characteristics.

[0121] Based on the current state of charge estimate and the next time step's single-step prediction of the state of charge, combined with the temperature-modulated Coulomb efficiency, the residual term of the charge dynamics mechanism is calculated. The calculation formula is:

[0122]

[0123] in, The state of charge at the next time step is a dimensionless, dimensionless prediction obtained from the state estimate at the k-th sampling time. Calculated using the same mapping model; The coulomb efficiency mapping function is obtained based on cell calibration tests and is implemented using a one-dimensional lookup table method combined with linear interpolation. The input is the average temperature, and the output is the coulomb efficiency under the corresponding operating conditions. It is dimensionless and has a value range of (0.99, 1). This is a dimensionless residual term representing the charge dynamics mechanism. It characterizes the deviation between the charge state evolution and the ampere-hour integral mechanism and is used to constrain the model output to conform to the battery charge dynamics characteristics.

[0124] Based on battery thermodynamic parameters, core temperature estimates, and single-step predictions, and combined with heat generation and dissipation mechanisms, the residual term of the thermal equilibrium mechanism is calculated. The calculation formula is:

[0125]

[0126] in, This is a single-step prediction of the core temperature at the next time step, obtained based on the state estimate at the k-th sampling time, in Kelvin. Calculated using the same mapping model; The equivalent heat capacity of the battery module is expressed in joules per Kelvin, calculated based on the mass and specific heat capacity of the cells and structural components within the module. The entropy thermal coefficient of the battery is expressed in volts per Kelvin. It is obtained by calibration based on the temperature characteristics of open-circuit voltage and is implemented by a one-dimensional lookup table method combined with linear interpolation. The input is the state of charge. The equivalent heat dissipation coefficient of the battery module to the environment, measured in watts per Kelvin, is obtained from calibration tests based on the module's thermal management structure. The ambient temperature collected at the k-th sampling time is expressed in Kelvin. This is the residual term of the thermal equilibrium mechanism, in watts, which characterizes the deviation between the core temperature evolution and the thermal equilibrium mechanism, and is used to constrain the model output to conform to the thermodynamic characteristics of the battery.

[0127] Based on the clamping flexibility memory modulation clamping boundary stiffness mapping function, structural overall shape variable mapping function, and assembly memory bias mapping function, the estimated endplate force of module m at the k-th sampling time is calculated. The calculation formula is:

[0128]

[0129] in, The clamping boundary pressure stiffness mapping function, in Pascals per meter, is obtained from pressure-displacement calibration tests under different preloads and compression pad aging degrees. It is implemented using a one-dimensional lookup table method combined with linear interpolation. The input is the clamping flexibility memory value, and the output is the clamping boundary pressure stiffness under the corresponding working condition, which represents the pressure change corresponding to the unit deformation. The function is a mapping function for the overall shape of the battery cell stack structure, with units in meters. It is obtained based on the calibration test of the battery cell expansion characteristics and is implemented by a two-dimensional lookup table method combined with linear interpolation. The input is the state of charge and the state of health, and the output is the overall shape of the battery cell stack under the corresponding operating conditions, including reversible and irreversible expansion effects. The assembly memory bias mapping function, in Pascals, is obtained based on the calibration test of the clamping flexibility memory. It is implemented using a one-dimensional lookup table method combined with linear interpolation. The input is the clamping flexibility memory and the output is the assembly pressure bias under the corresponding working condition. This is the estimated force on the end plate of module m at the kth sampling time, in Pascals.

[0130] The assembly pressure characteristic parameter, after removing the influence of thermal expansion, is subtracted from the estimated stress on the end plate to obtain the mechanical boundary mechanism residual term. The calculation formula is:

[0131]

[0132] in, The mechanical boundary mechanism residual term, in Pascals, characterizes the deviation between the measured assembly pressure and the predicted mechanism value, and is used to constrain the model output to conform to the mechanical boundary characteristics of the battery stack structure.

[0133] Hyperbolic cosine smoothed robust error function is used to evaluate various error losses. The formula for calculating the error function is as follows:

[0134]

[0135] in, is a smooth and robust error function, with an output value of a non-negative dimensionless scalar, which can reduce the impact of outliers on the optimization process; 'a' is the input dimensionless error value.

[0136] Based on the calibration labels in the historical training dataset, a smooth and robust error function is used to calculate the total state-supervised loss. The calculation formula is:

[0137]

[0138] Wherein, Y is a set of state parameter labels, which includes three types of labels: state of charge, health state, and core temperature; The calibration label value of the r-th state parameter at the k-th sampling time is obtained through offline calibration testing; The reference value for the r-th state parameter is used for dimensionless processing. The reference value for the charged state and healthy state is 1, and the reference value for the core temperature is 1. The summation operation is performed based on all valid sampling times and all modules within the historical training dataset D. The total loss of state monitoring is a dimensionless scalar.

[0139] After dimensionless processing of the residual terms of each mechanism, the corresponding loss is calculated using a smoothing robust error function. The calculation formula is as follows:

[0140]

[0141]

[0142]

[0143]

[0144] in, The voltage reference is in volts, and the value is the nominal voltage of the battery cell. The summation operation is performed based on all valid sampling times and all modules in the historical training dataset D. , , , These are the mechanistic residual losses corresponding to terminal voltage, charge dynamics, thermal equilibrium, and mechanical boundary, respectively, and all loss terms are dimensionless scalars.

[0145] By introducing trainable weight coefficients corresponding to each error, the loss value of each error is correlated with its corresponding weight coefficient, and the resulting sum is calculated to obtain the basic optimization objective. The calculation formula is:

[0146]

[0147] in, It is the total set of trainable parameters consisting of the basic network parameters and the adaptation parameters, including all trainable parameters of the time sequence encoder, state decoder, memory gating mapping weights, and mechanical boundary mapping weights; This is the set of trainable weight coefficients corresponding to each error. Let be the trainable weight coefficient corresponding to the r-th error term. It is dimensionless and initially takes the value of 0. Based on the optimization objective, a dimensionless scalar, L is used to optimize the weights of the state supervision loss and the residual losses of each mechanism by adaptively weighting the state supervision loss. r S is the loss value corresponding to the r-th error. r is the trainable logarithmic weight parameter corresponding to the r-th error term, dimensionless, and initially set to 0. It is the adaptive weighting coefficient corresponding to the r-th error, dimensionless, and its value range is... It is used to automatically balance the optimization contribution of different loss terms.

[0148] Extract the set of structural weight parameters from the neural network, and construct an adaptive regularization term based on the feature norm of the weight parameters and the distribution variance of the clamping flexibility memory. The calculation formula is:

[0149]

[0150] Where L is the set of structural weight parameters consisting of the encoder, decoder, gating mapping weights, and spatial feature mapping weights in the neural network; The regularization baseline coefficient is dimensionless and has a fixed value. ; The variance calculation operator is calculated based on the dimensionless clamping flexibility memory in the historical training dataset D, and is dimensionless. The smallest bias constant is dimensionless and takes the value of ; Let be the square of the Frobenius norm of the weight matrix W, which is dimensionless; The adaptive regularization term is a dimensionless scalar used to constrain the complexity of network weights and avoid model overfitting. The regularization strength is adaptively adjusted according to the distribution dispersion of the clamping flexibility memory.

[0151] The basic optimization objective is superimposed with the adaptive regularization term to form the overall objective function. The calculation formula is:

[0152]

[0153] in, The overall objective function is a dimensionless scalar that serves as the solution objective for offline optimization.

[0154] The L-BFGS algorithm based on second-order curvature information is used to solve the overall objective function. A one-dimensional line search is combined to determine the dynamic descent step size, completing the joint optimization and update of the basic network parameters and the adaptation parameters. The parameter update calculation formula is as follows:

[0155]

[0156] Where t is the index of the number of optimization iterations, and its value is a positive integer; For the overall objective function with respect to the parameter set The first-order gradient vector; For the overall objective function with respect to the parameter set The second-order Hessian matrix; For the Moore-Penrose generalized inverse operator of a matrix; Let be the dynamic descent step size for the t-th iteration, dimensionless, determined by a one-dimensional Armijo line search to ensure stable convergence during the iteration process; the historical vector storage size of the L-BFGS algorithm is set to 10, the maximum number of iterations is set to 5000, and the convergence threshold is set to... The optimization is terminated early when the change in the overall objective function over 20 consecutive iterations is less than the convergence threshold. During the optimization process, the parameters of the reversible dilatational baseline mapping model are kept frozen, and only the parameter set included in the overall objective function is adjusted. With weight coefficient set Perform the update.

[0157] The historical training dataset covers the entire lifecycle of energy storage batteries, including charge / discharge rates from 0.1C to 3C, ambient temperatures from -20°C to 60°C, state of charge (SOC) from 0 to 1, state of health (SOC) from 0.7 to 1, preload from 500 Pascals to 5000 Pascals, and compression pad aging from 0 to 1000 cycles. Calibration labels within the dataset were obtained through offline calibration testing, SOC labels were obtained by integrating full charge / discharge capacity, SOC labels were obtained through capacity decay testing, and core temperature labels were collected by temperature sensors embedded within the cells. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter optimization, the validation set for hyperparameter tuning and early stop detection, and the test set for verifying model generalization performance.

[0158] During the online operation phase, the current observation sequence is continuously fed into the network that has completed offline optimization. The various mechanistic residual terms for the current period are calculated, and after dimensionless processing, they are merged to construct an unsupervised online residual objective function. The calculation formula is:

[0159]

[0160] in, The set of adaptation parameters at the k-th sampling time includes trainable parameters of memory gating mapping weights and mechanical boundary mapping weights, which are dimensionless; It is an unsupervised online residual objective function, a dimensionless scalar, constructed only based on the currently observable mechanistic residual terms, requiring no calibration labels, and enabling online unsupervised optimization.

[0161] Freeze the basic network parameters in the backbone network, and perform optimization updates on the adaptation parameters using only the unsupervised online residual objective function. Combine this with a one-dimensional line search to determine the dynamic step size. The parameter update calculation formula is as follows:

[0162]

[0163]

[0164] in, The dynamic step size for online updates at the k-th sampling time is dimensionless and determined through a one-dimensional Armijo line search. The basic network parameters are the weights of the offline optimized time-series encoder and state decoder, which remain fixed during online operation to avoid shifts in the network's main features. The triggering conditions for online updates are: absolute value of module charging / discharging current greater than 0.05C, state of charge between 0.1 and 0.9, and average temperature between 0 and 45 degrees Celsius. Parameter updates are executed when all triggering conditions are met. Prohibited operating conditions for online updates include: static condition, fault condition, module replacement condition, and preload adjustment condition. All parameter updates are frozen when these conditions are present. The adaptive parameter settings are within a legal value range, and amplitude limiting is performed during updates. If the value exceeds the range, it automatically reverts to the previous set of legal parameters. The maximum number of iterations within a single sampling step is set to 5, and the convergence threshold is set to... The iteration is terminated early when the convergence condition is met.

[0165] By integrating the state decoder that optimizes the basic parameters offline and updates the adaptive parameters online, the output module m provides the real-time battery state estimation result at the k-th sampling time. The calculation formula is:

[0166]

[0167] in, To integrate offline optimized basic parameters Adaptation parameters with online updates State decoder; The final output is a real-time battery state estimation result, which includes the final estimates of state of charge, state of health, and core temperature.

[0168] Based on the current clamping flexibility memory and the statistical characteristics of the historical training dataset, a continuous uncertainty index for assessing the risk of deviation from the physical structure boundary is calculated. The calculation formula is:

[0169]

[0170] in, The mean calculation operator is calculated based on the clamping flexibility memory in the historical training dataset D, and the unit is Pascal; This is an external risk scale variable, dimensionless, and can be adjusted according to the system risk management needs. The default value is 0. The minimum bias constant, in Pascals squared, has a value of [value missing]. Multiply by the squared range of the flexibility memory within the historical training dataset D to avoid computational anomalies where the denominator is 0; It is a continuous uncertainty index, dimensionless, with a value greater than or equal to 1. The larger the value, the greater the deviation of the current module's clamping structure boundary state from the historical normal operating conditions, and the greater the uncertainty of the corresponding state estimation result. It can be used for health warning and risk management of battery systems.

[0171] For the initial stage of the sampling sequence, when the cumulative power throughput... When the value is less than 0.01, the clamping flexibility memory value remains at its initial value of 0, and online parameter updates are paused. Normal recursion and updates will resume once the cumulative charge throughput exceeds the threshold. For stationary operation, when the absolute value of the charging and discharging current is less than 0.01C, the incremental change in charge, the residual pressure integral value, and the cumulative charge throughput remain unchanged, and online parameter updates are paused. For numerical stability processing, a sliding window is set along the cumulative capacity integral value and the cumulative charge throughput, with a window length of 10,000 sampling steps. Historical data exceeding the window will no longer participate in the cumulative calculation to avoid numerical overflow and drift during long-term operation.

[0172] The calibration procedure for the open-circuit voltage mapping function is as follows: Place the cell under test in a constant-temperature chamber with ambient temperatures set to 0°C, 25°C, 45°C, and 60°C. At each temperature point, fully charge the cell and then discharge it at a constant current rate of 0.05C. After discharging 5% of the nominal capacity, allow it to stand for 2 hours. Record the terminal voltage after standing as the open-circuit voltage at the corresponding state of charge and temperature. Construct a two-dimensional lookup table for all test point data, and use linear interpolation to achieve continuous mapping between adjacent data points. The calibration procedure for the equivalent internal resistance and polarization dynamic parameters is as follows: Using a hybrid pulse power characteristic test method, apply a 10-second 1C discharge pulse to the cell at each calibration temperature and state of charge point. Record the voltage response during the pulse process. Obtain the DC internal resistance, polarization time constant, and polarization impedance parameters through least-squares fitting, and construct a lookup table for the corresponding mapping function.

[0173] The calibration procedure for the clamping boundary pressure stiffness mapping function is as follows: place the battery cell stack and clamping device in a universal pressure testing machine, apply a compressive load at a rate of 0.1 mm / min, record the end plate pressure under different compression displacements, and simultaneously simulate compression pad replacement tests with different aging degrees to obtain the pressure-displacement correspondence under different flexibility states, and construct a lookup table for the mapping function.

[0174] The calibration procedure for the coulomb efficiency mapping function is as follows: At different temperature points, perform full charge and discharge cycle tests on the battery cell, record the ratio of discharge capacity to charge capacity as the coulomb efficiency at the corresponding temperature, and construct a one-dimensional lookup table.

[0175] The calibration procedure for the entropy-heat coefficient is as follows: at different states of charge, the battery cell is placed in a constant temperature environment at different temperatures, and the change in open-circuit voltage at different temperatures is recorded. The rate of change of open-circuit voltage with temperature is obtained by linear fitting, which is the entropy-heat coefficient of the corresponding state of charge. A one-dimensional lookup table is then constructed.

[0176] The weight matrices of all fully connected layers are initialized using the He normal distribution method, with the following initialization formula:

[0177]

[0178] in, This corresponds to the number of input neurons in a fully connected layer. It is a normal distribution operator; this initialization method can effectively alleviate the gradient vanishing and gradient explosion problems in the training process of deep networks. The physical meaning of the weight matrix is ​​the linear mapping coefficient from input features to output features, which is dimensionless.

[0179] The bias vectors of all fully connected layers are uniformly initialized to all zero vectors. The physical meaning of the bias vector is the reference offset of the corresponding output feature, and the dimension is consistent with the dimension of the corresponding output feature.

[0180] The initial value of the hidden state of the gated cyclic unit encoder is uniformly set to an all-zero vector. This initial value corresponds to the reference hidden state when there is no timing information input and is dimensionless. The initial value of the polarization memory vector is uniformly set to an all-zero vector, which corresponds to the reference value when the cell is in a non-polarized state and is dimensionless.

[0181] Trainable weight coefficients The initial weights are uniformly set to 0, corresponding to the initial weight balance of each loss, which is dimensionless.

[0182] The rule for determining anomalies in mechanistic residuals is as follows: For each mechanistic residual at each sampling time, calculate the ratio of its value to the mean of the absolute values ​​of the residuals at the 100 historical valid sampling times. If the ratio is greater than 3, the residual is determined to be an outlier.

[0183] The rules for handling abnormal residuals are as follows: when a single residual is determined to be abnormal, the loss contribution of the corresponding residual is suspended at that sampling moment, and only the remaining normal residuals are used to construct the online objective function; when two or more residuals are determined to be abnormal, the online parameter update is completely suspended at that sampling moment, while the valid parameter values ​​at the current moment are retained, and the online update is restarted after the residuals at the next three consecutive sampling moments have all returned to normal.

[0184] The sliding window filtering rule for the residual terms is as follows: all mechanistic residual terms used for online updates are processed by sliding window mid-range filtering over 5 sampling times. The filtered residual values ​​are used to construct the online objective function to further suppress the interference of random noise on the online update.

[0185] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A model optimization method for neural networks used in energy storage battery state estimation, applied to an energy storage battery system including cells and clamping devices, characterized in that, include: Acquire module observation signals covering multiple consecutive sampling times, and extract normalized ampere-hour charge characteristics and thermal field gradient characteristics at each time point; The main force spatial modes of the pressure sensor array are extracted and temperature interference is removed to obtain the assembly pressure characteristic parameters at each time point after removing the influence of thermal expansion. Integrating the assembly pressure characteristic parameter and the residual term of the reversible expansion baseline along the cumulative charge and discharge capacity, the clamping flexibility memory amount, which includes the historical cumulative effect, is recursively obtained. The clamping flexibility memory is converted into memory gating, injected into the neural network, and the time hidden state is transmitted between time steps. Combined with the spatial propagation mechanism, the estimated parameters of the battery at each time step are output. The estimated parameters include the state of charge, health state, and core temperature. Based on the estimated parameters, a mechanism residual term is constructed that covers terminal voltage, charge dynamics, thermal equilibrium, and mechanical boundary characteristics; By combining the state supervision error of the estimated parameters with the mechanism residuals of each term, an overall objective function based on trainable weight coefficients and adaptive regularization terms is constructed to complete the joint offline optimization of basic network parameters and adaptive parameters. During online operation, the clamping flexibility memory is continuously recursively calculated, and the adaptive parameters are updated unsupervisedly only using the real-time mechanism residual term, outputting the state estimation results and uncertainty index.

2. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, Acquire module observation signals covering multiple consecutive sampling times, and extract normalized ampere-hour charge characteristics and thermal field gradient characteristics at each time point, including: Calculate the adjacent sampling interval between the current time and the previous time; Based on the adjacent sampling interval, the module current in the module observation signal, and the nominal capacity, calculate the increment of power change at each time point; An initial energy baseline is set, and the energy change increment is accumulated step by step to obtain the normalized ampere-hour energy characteristics at each time step; Calculate the average temperature of the readings at each temperature sensor point on the module surface in the module observation signal; The temperature vector, composed of the readings of each temperature sensing point, is multiplied by the spatial physical adjacency weight matrix constructed based on the physical topology of the sensors to obtain the thermal field gradient characteristics at each time point. By utilizing the dispersion parameters of each feature channel on historical training data, the current input is scaled and normalized to obtain the robust normalized features of each parameter.

3. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, The main force spatial modes of the pressure sensor array are extracted and temperature interference is removed to obtain the assembly pressure characteristic parameters at each time point after removing the influence of thermal expansion, including: Subtract the historical reference value from the pressure sensor array vector containing all endplate pressure sensor readings to obtain the centered pressure vector. Principal component feature extraction is performed on the sequence matrix formed by the central pressure vector at all observation times, and the first principal component feature vector is extracted as the main force space mode. Multiply the main force spatial mode by the central pressure vector at the current moment to obtain the main mode pressure scalar; The temperature effect interpretation vector is formed by sequentially concatenating the bias reference input, the extracted average temperature, and the rate of change of the average temperature over time. The temperature stripping coefficient is obtained by parameter optimization, which minimizes the error between the product of the main modal pressure scalar, the temperature stripping coefficient, and the temperature effect interpretation vector. Subtracting the product of the temperature stripping coefficient and the temperature effect interpretation vector from the principal modal pressure scalar yields the assembly pressure characteristic parameter that removes the effect of thermal expansion.

4. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 2, characterized in that, Integrating the assembly pressure characteristic parameter and the residual term of the reversible expansion baseline along the cumulative charge-discharge capacity, the clamping flexibility memory amount, including the historical cumulative effect, is recursively derived, including: Extract the positive and negative signs of the charging and discharging current at each moment as flux direction parameters; Using historical datasets, a reversible dilated baseline mapping model is trained based on a cost function constructed from fitting error weighting factors and parameter smoothness constraint factors. The normalized ampere-hour energy characteristics, average temperature, and flux direction parameters are input into the reversible expansion baseline mapping model to obtain the reversible expansion baseline at the current moment. Subtract the reversible expansion baseline at the current moment from the assembly pressure characteristic parameter to obtain the pressure residual term; The initial value of the integral is set to zero. The pressure residual cumulative integral value along the capacity at the previous moment is added to the product of the pressure residual term and the current change in the electrical quantity, and the pressure residual cumulative integral value along the capacity at the current moment is obtained recursively. The initial throughput value is set to zero. The cumulative power throughput of the previous moment is added to the absolute value of the current power change increment, and the cumulative power throughput of the current moment is recursively calculated. The clamping flexibility memory value is obtained by dividing the current pressure residual along the capacity cumulative integral value by the sum of the current power throughput cumulative value and the minimum bias constant.

5. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, The clamping flexibility memory is converted into memory gating, injected into the neural network, and the temporal hidden state is propagated between time steps. Combined with the spatial propagation mechanism, the estimated parameters of the battery at each time step are output, including: The normalized parameters, together with the clamping flexibility memory value and its time difference value at the current moment, are concatenated to form the model input sequence vector; The model input sequence vector is input together with the preset initial hidden state or the time hidden state of the previous time step into the time series encoder to obtain the time hidden state of the current time step. The combination of the clamping flexibility memory amount and the corresponding pressure feature is mapped using the first correlation weight matrix, and the memory gating is generated after activation processing. The memory gate is multiplied element-wise with the feature branch vector processed by the second correlation weight matrix and then superimposed onto the time-hidden state at the current moment to form a memory-injected hidden state. The spatial propagation weight is calculated based on the physical adjacency relationship between adjacent modules. The memory injection hidden states of adjacent modules are weighted and aggregated based on the spatial propagation weight, and an aggregated global hidden state is generated after feature fusion. The aggregated global hidden state is input into the state decoder of the neural network, and the output is the estimated parameters including the state of charge, the health state, and the core temperature.

6. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, Based on the estimated parameters, a mechanistic residual term is constructed that covers terminal voltage, charge dynamics, thermal equilibrium, and mechanical boundary characteristics, including: Using the open-circuit voltage estimate mapped by the estimated parameters, subtract the battery equivalent internal resistance voltage drop and polarization voltage drop estimates to obtain the terminal voltage estimate, and then subtract it from the actual terminal voltage in the module observation signal to obtain the terminal voltage mechanism residual term. The difference between the current and previous states of charge is subtracted from the ampere-hour integral theoretical increment modulated by the estimated parameters to obtain the residual term of the charge dynamics mechanism. The product of the core temperature change rate and the equivalent heat capacity is calculated using the battery thermodynamic parameters. The heat generated by the internal resistance Joule heat, the relatively reversible heat generated including the battery entropy heat coefficient, and the equivalent heat dissipation to the environment are then subtracted from this product to obtain the residual term of the thermal equilibrium mechanism. Multiply the clamping boundary stiffness parameter modulated by the clamping flexibility memory by the overall structural deformation including reversible and irreversible expansion effects to obtain the estimated stress on the end plate. The mechanical boundary mechanism residual term is obtained by subtracting the assembly pressure characteristic parameter after removing the influence of thermal expansion from the estimated stress value of the end plate.

7. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, By combining the state supervision error of the estimated parameters with the mechanistic residuals of each term, an overall objective function based on trainable weight coefficients and adaptive regularization terms is constructed to complete the joint offline optimization of the basic network parameters and adaptive parameters, including: The state supervision error with calibration labels and each independent mechanism residual term are evaluated using a smooth robust error function. By introducing the trainable weight coefficients corresponding to each error, the loss value of each error is correlated with its corresponding trainable weight coefficient, and the correlation decay calculation is performed and summed to obtain the basic optimization objective. Extract the predetermined structural weight parameters from the neural network, and construct the adaptive regularization term by combining their feature norms with the distribution variance of the clamping flexibility memory on the training dataset; The overall objective function is formed by superimposing the basic optimization objective with the adaptive regularization term. The overall objective function is solved using a parameter optimization algorithm based on second-order curvature information, and the basic network parameters and the adaptation parameters are optimized and updated by combining the dynamic descent step size.

8. The model optimization method for neural networks used for state estimation of energy storage batteries according to claim 1, characterized in that, During online operation, the clamping compliance memory is continuously recursively calculated. Unsupervised updates are performed on the adaptation parameters only using the real-time mechanistic residual term. The output state estimation results and uncertainty indices include: During the online operation phase, the current observation sequence is continuously fed into the network that has completed offline optimization, and the mechanism residual terms of each term in the current period are calculated and merged to construct an unsupervised online residual objective function; The basic network parameters in the backbone network are frozen, and optimization updates are performed only using the unsupervised online residual objective function relative to the adaptation parameters, where the adaptation parameters represent the set of mapping weights used to control network memory gating and endplate stiffness. The state decoder, which incorporates updated adaptive parameters, outputs real-time battery state estimates. Calculate the dispersion parameter of the current clamping flexibility memory value from the historical mean, and output the dispersion parameter as the continuous uncertainty index for assessing the risk of deviation from the current physical structure boundary after scaling and mapping.

Images