WIPI wireless power dynamic balancing method for grid connection of energy storage system

Abstract

The invention relates to the technical field of control of power electronics and energy storage systems, in particular to a WIPI wireless power dynamic balancing method for grid connection of an energy storage system, which comprises the following steps: constructing a fractional order digital twinborn model, and predicting the charge state and temperature rise of a module; based on the prediction result, a demand matrix and a loss matrix are obtained by using a three-layer GraphSAGE and a twenty-quantum-bit variable component sub-optimizer respectively, and the demand matrix and the loss matrix are fused into a power coupling target according to a preset weight; performing target thresholding to generate a 32 * 32 magnetic-acoustic super-structure pixel map, finishing magnetic field shaping at a microsecond level, and driving a GaN inverter to output a matched phase and frequency at the same time; power flow and temperature data are collected through a high-precision analog-to-digital converter, a power transport matrix is iteratively corrected by adopting an entropy regularization Sinkhorn algorithm, and pixels are updated in real time; and finally, writing the feedback Hash into the alliance chain storage to complete millisecond-level credible traceability. According to the invention, rapid, accurate and traceable balance of wireless energy of the energy storage module is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power electronics and energy storage system control, and particularly to a WIPI wireless power dynamic equalization method for energy storage system grid connection. Background Art

[0002] During the grid-connected operation of an energy storage power station, the state of charge imbalance is likely to occur among battery modules due to differences in aging rate and thermal environment, resulting in capacity waste and local overheating. Wireless inductive power interchange (WIPI) can continuously redistribute the module power flow without interrupting the DC bus, which is the key to improving the available capacity and thermal safety. Existing dynamic equalization mostly adopts the "switch slicing" method (Chinese invention patent, publication number: CN114899923B). For example, monomers are temporarily bypassed or series-connected to an equalization resistor, and then cycled in sequence according to the voltage error. This idea relies on mechanical / solid-state switches, and the action frequency is limited by the device life. Moreover, the cut-out process destroys the continuity of the bus; the algorithm is only based on the instantaneous voltage, ignoring magnetic and thermal coupling, and the adjustment accuracy is not high; when the system voltage level increases, the number of switches and the failure risk increase exponentially, and the maintenance cost rises significantly. Summary of the Invention

[0003] Aiming at the many problems existing in the above-mentioned prior art, the present invention provides a WIPI wireless power dynamic equalization method for energy storage system grid connection. The present invention is based on a fractional-order digital twin model to predict the future state of charge and temperature rise trends of the battery. The graph neural network outputs a demand-oriented matrix, and the quantum optimizer outputs a loss-optimal matrix. The two are fused according to a fixed weight to generate a power coupling target; the target regulates a 32×32 magneto-acoustic metamaterial array through threshold mapping and synchronously drives a GaN inverter to achieve uninterrupted energy redistribution. The real-time power and temperature are sampled at high resolution and then enter the Sinkhorn loop to quickly obtain the optimal transmission gradient and update the pixels, forming a closed loop. The scheme significantly reduces the state of charge difference among modules, suppresses bus inrush current, and reduces the inverter harmonics.

[0004] A WIPI wireless power dynamic equalization method for energy storage system grid connection, the method comprising:

[0005] Establish a digital twin model, and integrate system parameters representing coil geometry, electrochemical impedance of battery modules, and coil coupling into a unified model for describing the electromagnetic topology and electrochemical state of the energy storage system;

[0006] Collect and synchronize electrical data and magnetic flux data, and perform fractional-order dynamics prediction based on the digital twin model to obtain state of charge data, temperature constraint data, and generate control optimization initialization data;

[0007] Generate power coupling control target data based on the predicted results and initialization data, and generate magneto-acoustic metamaterial tuning data and execute control instruction data accordingly;

[0008] Drive the inverter according to the execute control instruction data, tune the magneto-acoustic metamaterial device based on the magneto-acoustic metamaterial tuning data to achieve wireless power coupling, measure the power flow data and temperature data in real time, calculate the coupling correction gradient using the optimal transmission algorithm to update the magneto-acoustic metamaterial device, and form feedback data from the power flow data, temperature data and execute control instruction data to return to the power coupling control target data generation step, so as to realize the wireless power dynamic balance between battery modules within the closed loop.

[0009] Preferably, the system parameters consist of the following data: coil center coordinates, coil normal vector, coil effective area, battery module zero-frequency internal resistance value, battery module polarization resistance value, and coupling coefficients between each pair of coils.

[0010] Preferably, the digital twin model is established through the following steps:

[0011] In the electromagnetic part, a finite element mesh with an element size not greater than 0.5 mm is used to calculate the coil coupling impedance matrix; in the electrochemistry part, a fractional-order equivalent circuit containing a constant phase element is used to fit the battery module impedance curve; with the goal of minimizing the mean square error between the model output and the measured impedance spectrum, a sparse identification dynamics algorithm with a penalty coefficient is used to determine the fractional order and circuit parameters.

[0012] Preferably, the synchronization of electrical data and magnetic flux data is completed using a time-sensitive optical fiber network, and the synchronization jitter is not greater than 50 nanoseconds.

[0013] Preferably, the order of the fractional-order dynamics prediction takes a value between 0.80 and 0.90, and this order is determined by minimizing the mean square error of the state of charge within a 10-second prediction window.

[0014] Preferably, the power coupling control target data is linearly fused by the coupling weight matrix output by the graph neural network and the optimal coupling matrix output by the variational quantum optimizer with weight coefficients of 0.60 and 0.40, where the graph neural network adopts a 3-layer GraphSAGE structure with 256 hidden nodes in each layer, and the variational quantum optimizer uses 20 qubits and a parameterized quantum circuit with a depth of 6 layers.

[0015] Preferably, the magneto-acoustic metamaterial tuning data is a pixel matrix of 32 rows and 32 columns, where when the element value in the coupling matrix target data is greater than 0.35, the corresponding pixel is set to the on state, and the rest of the pixels are set to the off state, and the pixel switching completion time is not greater than 60 microseconds.

[0016] Preferably, the power flow data is collected at a rate of 1,000,000 samples per second through an 18-bit analog-to-digital converter, and the temperature data is collected at a rate of 200,000 samples per second through a 16-bit analog-to-digital converter.

[0017] Preferably, the optimal transmission algorithm uses the entropy-regularized Sinkhorn algorithm, with an entropy regularization coefficient of 0.05 and 30 iterations. The iteration stops when the difference in the two-norm of the power transport matrices obtained from two consecutive iterations is less than.

[0018] Preferably, after generating the feedback data, the SHA-256 hash algorithm is used to calculate the hash value and write it into the blockchain for storage. The blockchain node completes the transaction confirmation including the hash value and the Coordinated Universal Time (UTC) timestamp within 200 milliseconds.

[0019] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0020] The present invention shapes the magnetic flux through an array of magneto-acoustic metamaterial devices, realizing instantaneous power routing without physical switching and eliminating high-frequency switching stress.

[0021] The present invention realizes the forward prediction and global optimal allocation of charge-temperature-rise-loss through fractional-order digital twin + GraphSAGE-VQO fusion optimization.

[0022] The present invention realizes the adaptive correction of the power flow within a 100Hz closed loop through the fast iteration of entropy-regularized Sinkhorn.

[0023] The present invention realizes millisecond-level trusted data traceability through blockchain hash evidence storage, facilitating operation and maintenance and compliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a schematic flow diagram of the method of the present invention;

[0025] Figure 2 is a schematic diagram of the composition of the digital twin model in the present invention;

[0026] Figure 3 is a schematic diagram of the control optimization fusion architecture in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, obviously, one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present disclosure.

[0028] The terms used herein are for describing specific embodiments only and are not intended to limit the present disclosure. The terms "including", "comprising" and the like used herein indicate the presence of the described features, steps, operations and / or components, but do not preclude the presence or addition of one or more other features, steps, operations or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those of ordinary skill in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0030] As Figure 1 shown, a WIPI wireless power dynamic equalization method for grid connection of an energy storage system, the method comprising:

[0031] As Figure 2 shown, establish a digital twin model, and integrate system parameters characterizing coil geometry, battery module electrochemical impedance and coil coupling into a unified model for describing the electromagnetic topology and electrochemical state of the energy storage system;

[0032] The coil - coil coupling system includes:

[0033] Coils: Planar induction coils mounted outside each battery module, which can be PCB spiral lines, Litz wires or three - dimensional shaped windings, serving as wireless power transfer ports and not located inside the battery cells.

[0034] Coil geometric parameters: Central coordinates, normal vectors, effective areas, number of turns, compensation capacitance values, used for calculating self - inductance and radiation distribution.

[0035] Coil coupling system: Composed of all coil mutual inductances M ij , coupling coefficients and their compensation networks, which can be abstracted as a multi - port magnetic - coupled transmission network, determining the instantaneous energy flow direction between modules and being the core of the electromagnetic topology.

[0036] Electromagnetic topology - electrochemical state unified digital twin model, equivalent impedance matrix of the external induction network:

[0037]

[0038] L represents the coil self - inductance matrix (H), M represents the coil mutual - inductance matrix (H), C represents the compensation capacitance matrix (F), C -1 is its inverse matrix, R Cu represents the coil copper - loss resistance vector (Ω), and ω represents the angular frequency (rad·s−1).

[0039] Fractional-order equivalent electrochemical impedance of the battery module:

[0040]

[0041] R0 represents the internal resistance at zero frequency (Ω); R t represents the polarization resistance (Ω); τ represents the diffusion time constant (s); j represents the imaginary unit, satisfying j 2 = -1, which is used to represent the phase information of the AC impedance; α represents the fractional-order (0 < α ≤ 1) to jointly obtain the unified fractional-order state equation:

[0042]

[0043] represents the fractional derivative of the Caputo form and order β,

[0044] x(t) represents the state vector, including: coil current i ij , the state of charge SOC of the module i , polarization potential η i , module temperature T i , coupling coefficient κ ij .

[0045] A and B represent the system matrices obtained by online sparse identification. A is the system matrix and B is the output matrix; u(t) represents the inverter drive voltage amplitude-phase vector.

[0046] A - System matrix, including four blocks:

[0047]

[0048] A mm describes the coupling between coil currents (derived from L, M, R Cu ); A ee describes the diffusion-polarization dynamics between electrochemical states (derived from R0, R t , τ, α); The cross blocks A me , A em couple the thermal-electric feedback between current-temperature-rise-polarization.

[0049] If the system contains N battery modules and N coupling coils, then:

[0050]

[0051] i(t) - Coil current vector of length N SOC(t) - State of charge vector of length N. η(t) - Polarization potential vector of length N. T(t) - Module temperature vector of length N. Mutual inductance κ ijNot as a separate state, but reflected in A mm and A me in the coefficients; when the magneto-acoustic pixel matrix changes the mutual inductance, these two sub-blocks are refreshed online synchronously.

[0052] This model simultaneously characterizes the electromagnetic topology (through Z em ) and the electrochemical state (through Z ec and the fractional-order terms), realizing an integrated digital twin.

[0053] Deducing the electrochemical state from electromagnetic and chemical impedance includes:

[0054] Online swept-frequency injection, where the inverter superimposes a small-amplitude swept-frequency signal outside the operating frequency and collects the total impedance Z tot .

[0055] Magnetic coupling deconvolution, using the known Z em numerical value to eliminate the coupling term and obtain

[0056] Sparse identification dynamics, using SINDy-f to minimize the mean square error to obtain R0, R t , τ, α, and update the electrochemical sub-model in real time.

[0057] Fractional-order observer estimation, writing the new parameters into A, B, and inversely calculating SOC i , η i , T i .

[0058] This process relies on externally measurable voltage-current-magnetic flux to reconstruct the internal electrochemical state, without the need to arrange sensors inside the battery cell, ensuring safety and maintainability.

[0059] Coil geometric vector, each coil C i is described by the following triple: center coordinate normal vector equivalent cross-sectional area S i > 0. Measurement method: blue light three-dimensional scanner, point cloud density 1000 points / mm -2 , geometric reconstruction error ≤ 0.05mm.

[0060] Electrochemical impedance parameters, the impedance spectrum of each battery module is collected in the frequency band of 0.1Hz - 5kHz, and fitted using a fractional-order equivalent circuit containing a constant phase element (CPE):

[0061]

[0062] Among them, R 0,i represents the zero-frequency internal resistance; Rt,i represents the polarization resistance; τ i represents the time constant; α i represents the fractional order. Fitting method: Genetic algorithm initialization + Least squares iteration; Mean square error ≤ 4%.

[0063] Coil coupling coefficient, the mutual inductance matrix M is calculated by the finite element method (element side length 0.5mm), and obtained after calibration of the measured mutual inductance:

[0064]

[0065] where L i and L j are the self-inductance values respectively. The relative error between the simulated and measured mutual inductances ≤ 3%.

[0066] Unified state space structure, select the Caputo form, fractional order derivative starting from 0, and establish the state equation:

[0067]

[0068] y(t) = Cx(t)

[0069] where, represents the Caputo fractional order derivative operator starting from 0 and with order β. β represents the fractional order, the measured range is 0.80 - 0.90, and the comprehensive performance is the best when β = 0.85 after comparison. x(t) represents the state vector, including the state of charge SOC of each battery module i , temperature T i and coil coupling coefficient κ ij . u(t) represents the inverter drive voltage amplitude-phase vector, which is composed of the inverter output voltage amplitude V k and phase φ k . y(t) represents the output vector, which is composed of the DC bus current I bus , bus power P bus and maximum temperature difference ΔT max . A represents the system matrix. B represents the output matrix, which is calculated from the coil topology and inverter interface parameters. C represents the output matrix, which is a set of row vectors for selecting the corresponding state components. A, B, and C are updated online every 1s by the sparse identification dynamics algorithm (penalty coefficient 1×10 -4 ).

[0070] Model calibration and verifiability, impedance spectrum fitting error:

[0071]

[0072] The mean square error of the state of charge in a ten-second rolling window is 0.9%, which is 42% lower than that of the integer order model.

[0073] Example 1 was deployed in an energy storage station with a capacity of 1 MWh and a bus voltage of 1200 V. After importing 46 pairs of coil data, A dimension 94×94 was generated. After continuous operation for 24 h, the standard deviation of SOC decreased from 3.5% to 1.4%.

[0074] The effect of adjusting the fractional order β based on Example 1 in Example 2 is shown in Table 1:

[0075] Table 1

[0076] β Predicted mean square error Inverter peak current (A) 0.80 1.5％ 132 0.85 0.9％ 118 0.90 1.1％ 125

[0077] In Example 3, based on Example 1, the number of hidden nodes of GraphSAGE was reduced from 256 to 64, the prediction error increased to 1.3%, and the video memory occupancy decreased by 62%.

[0078] Through the unified digital twin model of the present invention, the impedance spectrum fitting error is controlled within 5%, and the state of charge prediction error ≤ 1%. Compared with the integer order model, the prediction error is reduced by 42%, and the peak current of the inverter is reduced by 11%. The online update period is 23 ms, meeting the requirement of a 1 s control period, verifying the real-time availability of the model.

[0079] Preferably, the system parameters consist of the following data: coil center coordinates, coil normal vector, coil effective area, zero-frequency internal resistance value of the battery module, polarization resistance value of the battery module, and coupling coefficient between each pair of coils.

[0080] The present invention clearly defines the system parameters as: coil center coordinate p i , coil normal vector n i , coil effective area S i , zero-frequency internal resistance value R of the battery module 0,i , polarization resistance value R of the battery module t,i , and coil-coil coupling coefficient k ij . These parameters jointly drive the digital twin kernel of wireless power dynamic balancing. The specific acquisition, fusion, and verification processes are as follows.

[0081] For electromagnetic geometric parameters, coordinate and normal vector acquisition, a blue light three-dimensional scanner (accuracy 0.02 mm) was used to collect point clouds of all coils in the busbar; the coil center p i and normal vector n i were calculated using Poisson-Surface reconstruction. Critical value test: When the point cloud density dropped to 200 points / mm -2 , the reconstruction error increased to 0.18 mm, resulting in a 6% increase in the downstream mutual inductance calculation error, verifying the necessity of high-density scanning.

[0082] Calculation of the effective area, obtained by intercepting the scanned grid and projecting the cross-sectional normal vector:

[0083]

[0084] where r is the cross-sectional boundary vector. The area error is better than 0.6%. The closed boundary curve of the i-th grid cell of the cross-section, which is composed of the connection lines of sampling points, lies in the cross-sectional plane, and the curve direction is in the counterclockwise convention; dr represents the differential displacement vector along, which is in the same direction as the curve tangent vector and has a length approaching zero;

[0085] Electrochemical impedance parameters, broadband EIS sampling: 0.1 Hz–5 kHz, amplitude 5 mA, four-point probe to eliminate wire impedance. The fast fitting process includes:

[0086] 1. Generate candidate [R 0,i , R t,i , τ i , α i using the genetic algorithm.

[0087] 2. Converge after 20 rounds of least squares iteration, mean square error:

[0088]

[0089] controlled within 0.04.

[0090] 3. Online update period: Perform a fast EIS (6 s sweep frequency) correction every 30 min to solve the temperature-life drift.

[0091] Coupling coefficient κ ij , finite element mutual inductance matrix:

[0092]

[0093] Φ ij is the link magnetic flux on C j when the current L j flows through C i . I j represents the excitation current (A) in the coil C j , which is the reference current amplitude used for calculating the mutual inductance.

[0094] Simulation-experiment closed loop, using 46 pairs of coils as the reference, the average deviation between the measured mutual inductance and the simulation value is 3%. If the κ ij error exceeds 5%, the system triggers adaptive magnetic field shaping at the control layer to ensure the coupling accuracy.

[0095] Unify the mapping to the fractional-order state space, core relationship:

[0096]

[0097] y(t) = Cx(t))

[0098] represents the Caputo fractional derivative of order β starting from 0;

[0099] x(t) represents the state vector, including: coil current i ij , module state of charge SOC i , polarization potential η i , module temperature T i , coupling coefficient κ ij ;

[0100] u(t) includes the inverter output amplitude V k and phase φ k ;

[0101] y(t) includes bus current I bus , power P bus , maximum temperature difference ΔT max ;

[0102] A, B, C are updated every 1 s on the GPU using the sparse identification dynamics algorithm (penalty coefficient 1×10 -4 ).

[0103] Quantitative verification of technical effects, deployed at a 1 MWh site for 24 h: the standard deviation of SOC decreased from 3.5% to 1.4%. Compared with the linear integer-order model of wireless power equalization, the prediction error decreased by 42% and the peak current of the inverter decreased by 11%. The online update takes 23 ms, meeting the 1 s control cycle and leaving a 40% computational margin.

[0104] Example 1 (basic example), all system parameters are obtained according to the above process, generating a 94×94 matrix in dimension A, and achieving the goals of charge balance and thermal balance.

[0105] Example 2 (extended example: change in the number of hidden nodes), based on Example 1, the number of hidden nodes of GraphSAGE is reduced from 256 to 64. The prediction error increases to 1.3%, and the video memory occupancy decreases by 62%, indicating that it can be flexibly trimmed according to the computing power of edge devices.

[0106] Preferably, the digital twin model is established through the following steps:

[0107] In the electromagnetic part, a finite element mesh with an element size not greater than 0.5 mm is used to calculate the coil coupling impedance matrix; in the electrochemistry part, a fractional-order equivalent circuit containing a constant phase element is used to fit the impedance curve of the battery module; with the goal of minimizing the mean square error between the model output and the measured impedance spectrum, a sparse identification dynamics algorithm with a penalty coefficient is used to determine the fractional order and circuit parameters.

[0108] For the coil coupling impedance matrix, a three-dimensional finite element model is established for all coils on the bus side and the module side. The side length of the mesh element is taken as 0.5 mm, and it takes 28 s to solve the mutual inductance matrix M once on the Nvidia A40 GPU. The critical value test shows that when the side length of the element increases to 1 mm, the mutual inductance error rises from 3% to 7%, and the inverter phase compensation amount increases by 4°, proving that 0.5 mm is the feasible upper limit.

[0109] For the coil resistance R i and the mutual inductance M ij Under the known conditions, the coupling impedance matrix is written as:

[0110]

[0111] L represents the self-inductance diagonal matrix, and C is the compensation capacitance diagonal matrix. The calculation results are input as known terms for subsequent sparse identification.

[0112] For the fractional-order equivalent circuit, the impedance spectrum sampling frequency band is 0.1 Hz–5 kHz, and the amplitude is 5 mA. The original curve suppresses the measurement noise through Savitzky–Golay filtering (window 11 points, polynomial order 3). Select the R0||R t -CPE|| topology, impedance expression:

[0113]

[0114] Among them, R0 is the internal resistance at zero frequency, R t is the polarization resistance, τ is the time constant, and α is the fractional-order.

[0115] For the sparse identification dynamics algorithm (SINDy-f), the objective function:

[0116]

[0117] is the parameter vector to be identified, and λ = 1×10 -4 is the sparse penalty coefficient.

[0118] Initialize 50 candidate Θ using the differential evolution algorithm. Iterate 40 rounds of L-BFGS on the GPU, with 8 frequency point data in each round of gradient batch; the average convergence time is 3.1 s. Constraint settings: 0 < α < 1, R t > 0, τ > 0.

[0119] Model coupling and online refresh. After the identification is completed, Z em and Z ecThe combined mapping is projected onto the fractional - order state matrix A. This matrix is updated online with temperature drift and aging, and the refresh period is 1 s. The time taken for a single update on an RTXA5000 GPU is 23 ms, meeting the 1 - s scheduling window requirement of the control layer. The implementation effects are shown in Table 2 as follows:

[0120] Table 2

[0121] Index The present invention Without fractional-order control <![CDATA[Impedance spectrum fitting error ε Z > ≤4％ ≈9％ Mean square error of state of charge in ten seconds 0.9％ 1.6％ Inverter peak current 118A 132A

[0122] Verification of technical effects: The measured data is from a 1 - MWh prototype, which proves that after the cooperation of fractional - order modeling and sparse identification, the prediction accuracy is improved by 44%, the device stress is reduced by 11%, which is directly translated into narrower bus - voltage ripple and lower switching losses.

[0123] Basic implementation example: Running the above three steps under 46 pairs of coils and a 0.5 - mm grid, we get α = 0.85. The standard deviation of SOC drops from 3.5% to 1.4%.

[0124] Extended implementation example (grid size 1 mm): When the grid side length is widened to 1 mm, ε Z rises to 6.8%, the SOC error increases to 1.2%, and the peak current increases by 7%. This shows that a fine grid is the key to ensuring the equalization accuracy.

[0125] Through the hierarchical modeling process of finite - element - CPE - SINDy - f, the present invention realizes an accurate digital mapping of the electromagnetic - electrochemical coupling behavior in the WIPI wireless power dynamic equalization application, which can be stably refreshed at the second - level clock in the field, supporting the subsequent real - time closed - loop operation of fractional - order prediction and controller.

[0126] Collect and synchronize electrical data and magnetic - flux data, perform fractional - order dynamics prediction based on the digital - twin model to obtain state - of - charge data, temperature - constraint data, and generate control - optimization initialization data;

[0127] The composition of the sampling signal includes:

[0128] 1. Electrical data:

[0129] Module - terminal voltage V i and current I i ; 18 - bit ADC, 1 MS / s -1 .

[0130] Bus - voltage V bus and bus - current I bus ; also 1 MS / s -1 ; reflecting the sub - station - level load disturbance. Sweep - frequency injection signal δV(t); a small - amplitude sine sequence superimposed on the PWM sideband, used for online extraction of the electrochemical impedance spectrum.

[0131] These electrical quantities are timestamp-aligned and the instantaneous power P is calculated. i = V i I i and a frequency-domain impedance curve is formed for model updating of R0, R t , τ, α.

[0132] 2. Flux data:

[0133] The three-axis magnetic induction intensity B i = [B x,i , B y,i , B z,i ; The tunneling magnetoresistance array samples at 10 kS / s -1 The data is used for online inversion of the mutual inductance M of the coil ij , correction of the coupling submatrix M, and ensuring the electromagnetic topology is adaptive to temperature drift and mechanical vibration.

[0134] The expression of the fractional-order kinetic prediction model is:

[0135]

[0136] represents the Caputo fractional-order derivative with the order β = 0.85.

[0137] x(t) represents the state vector, including: the instantaneous current i of the coil ij ; the state of charge SOC of the module i ; the polarization potential η i ; the temperature T of the module i .

[0138] A represents the system matrix, which is composed of the coil geometry L, the mutual inductance M, the copper loss R Cu and the electrochemical impedance parameters R0, R t , τ, α combination. B represents the input matrix, corresponding to the inverter drive voltage vector u(t). u(t) represents the amplitude-phase of the output of each GaN inverter.

[0139] The derivation process of the state of charge and temperature constraints includes:

[0140] The inverter injects δV(t) in the broadband band and samples V i , I i to obtain the total impedance Z tot .

[0141] Subtract the known Z tot from Z em , and the residual is obtained by SINDy-f to get R0, R t , τ, α.

[0142] Write the new parameters into A, and use the fractional-order Kalman observer to fuse the instantaneous Coulomb integral ∫I i dt, and output the corrected SOC i .

[0143] Current power P i Combined with the thermal network model (C th , R th ), the temperature rise can be predicted, and the temperature constraint vector can be obtained

[0144] The control optimization initialization includes: SOC with a prediction window length of 10 s i and are spliced into the initialization vector The second-order approximation H0 is obtained by the above-mentioned automatic Jacobian differentiation, providing a warm start for the GNN-VQO hybrid optimizer, and enabling the power coupling target to converge within a 0.1 s cycle.

[0145] On the one hand, the present invention clarifies the specific sources of "electrical data" and "magnetic flux data" and their uses in the model; on the other hand, it elaborates on how the fractional-order digital twin inversely calculates the SOC i and temperature rise online by means of these data, thereby supporting the closed-loop control of wireless power dynamic balance.

[0146] To achieve the second-level wireless power dynamic balance of the WIPI energy storage system during grid-connected operation, the present invention sets a "synchronous observation - fractional-order prediction" layer at the control front end. This layer estimates the charge evolution curve and the upper bound of the temperature rise of each module 10 s in advance through the accurately time-synchronized electrical - magnetic flux data stream and the digital twin model constructed in the previous section, and compresses the results into the initialization quantity that can be directly used by the optimizer.

[0147] Synchronous observation principle: The electrical variables (bus voltage, current, and 46-way module terminal voltage) and the magnetic flux variables (46-way three-axis magnetic induction intensity) are uniformly stamped with 64-bit PTP timestamps by the time-sensitive fiber optic network (TSN) at the hardware layer. The 40 ns jitter ensures the alignment of the two types of signals within the same time step, providing in-phase input for subsequent coupled prediction.

[0148] Fractional-order dynamics principle: The model uses the Caputo fractional-order framework with an order β = 0.85 to describe the multi-domain coupling of "coil coupling - electrochemical impedance - heat diffusion". β is obtained through a 0.80 - 0.90 grid scan to obtain the minimum mean square error, and at the same time, the peak inverter current can be suppressed within 120 A.

[0149] Temperature constraint principle: The thermal behavior of the module is simplified to a first-order RC network, and the real-time I 2 R loss and switching loss drive the temperature node; if the predicted temperature in the 10 s window will exceed 55 °C, the temperature weight w T= 4 Write control vector to enhance heat dissipation priority.

[0150] Initialization quantity principle, the end SOC output by the prediction layer and the temperature upper limit are concatenated into a 92-dimensional vector, and accompanied by the Hessian obtained by second-order gradient approximation; the two are used as the "warm start" of the graph neural-quantum hybrid optimizer, significantly shortening the convergence steps.

[0151] In practical applications, for sampling and time synchronization, the DC bus quantity uses an 18-bit ADC at 1MS / s -1 , and the flux array uses 10kS / s -1 . All data is re-sampled to a 0.1s step size after TSN time synchronization; experiments show that if the step size is relaxed to 0.2s, the charge prediction error increases by 0.4%, so 0.1s is fixed as a compromise value.

[0152] Signal purification, the third-level Daubechies-4 wavelet retains the 2nd and 3rd-level low-frequency subbands, and then performs independent component analysis to strip the 20kHz switching harmonics. The noise power spectral density drops from 6×10 -9 to 8×10 -11 V 2 / Hz.

[0153] Fractional-order numerical advancement:

[0154] x k+1 = x k + h 0.85 (Ax k + Bu k )

[0155] h = 0.1s

[0156] where x k is the state vector (46 SOCs, 46 temperatures, 1035 coupling coefficients), u k is the inverter amplitude-phase vector, and A and B are re-calculated by the digital twin every second. The RTXA5000 GPU has a batch size of 128 in parallel, and the prediction for 10 steps at a time takes 4.2ms.

[0157] Temperature update, for each module, the temperature rise is calculated based on the heat capacity C th = 4.1kJ K -1 , thermal resistance Rth = 0.23KW -1 . If the predicted upper limit exceeds 55°C, then write w T .

[0158] Initialization packaging, the 46 SOCs and 46 temperature upper limits at the 10th step moment form χ0, and at the same time calculate the second-order gradient approximation H0; the two together with w TThe co-fed controller optimizer reduces the number of iterations from 52 to 32. The implementation effects are shown in Table 3 as follows:

[0159] Table 3

[0160] Index The present invention Integer-order Kalman-SOC 10s rolling prediction MSE 0.9％ 1.6％ Temperature warning lead 0.8s 0.1s Inverter peak current 118A 132A

[0161] Preferably, the synchronization of electrical data and magnetic flux data is completed using a time-sensitive optical fiber network, and the synchronization jitter is no greater than 50 nanoseconds.

[0162] In the WIPI wireless power dynamic equalization closed loop, electrical quantities (voltage, current) and magnetic fluxes (tunnel magnetoresistance three-axis magnetic induction) must be written into the predictor at the "same physical instant". If the time alignment error exceeds 100 nanoseconds, the fractional-order difference kernel will amplify the phase shift error into a charged prediction drift. Therefore, the present invention uses a time-sensitive optical fiber network (TSN-FN) to complete cross-domain synchronization and suppresses the end-side jitter within 50 ns.

[0163] TSN-FN is based on the IEEE802.1AS-rev precision clock: at the site, a master clock node stabilized by an OCXO distributes Sync / Follow-Up frames to 12 boundary switching nodes to form a hierarchical clock tree. Both the electrical measurement board (ADCFPGA) and the magnetic flux measurement board (TMRFPGA) are embedded with a hardware timestamp unit HTS-64, which writes the 64-bit PTP timestamp into the sampling FIFO immediately upon receiving the Sync frame, achieving "sampling-time-binding in one go". Since HTS-64 injects timestamps at the PHY-MAC boundary without passing through the operating system stack, the single-hop jitter is limited within 8 ns; the time base error of a five-hop link is theoretically The component information is shown in Table 4.

[0164] Table 4

[0165]

[0166] Time step resampling - all channels first enter the circular buffer and are then packed and sent by FPGA DMA at a fixed step of 0.1 s. If the time difference between two consecutive timestamps exceeds 50 ns, the jitter satellite window is triggered, and the predictor pauses and waits for the next frame alignment.

[0167] Absolute jitter test, a dual-channel oscilloscope compares the electrical-flux trigger signals, and the 24-hour statistical results show that the RMS jitter is 27 ns and the peak-to-peak is 48 ns, meeting the upper limit of 50 ns.

[0168] Regarding the impact on prediction accuracy, the synchronization jitter is artificially widened to 200 ns: the mean square error of the 10-s charged prediction increases from 0.9% to 1.3%, and the peak current of the inverter increases by 6 A; after restoring to 50 ns, the indexes fall back, proving the necessity of strictly controlling the jitter.

[0169] Regarding the impact on grid connection stability, a 5kW step discharge test was conducted at a 1MWh site: the precise synchronization version triggered the temperature warning 0.8s in advance, and the peak value of the bus inrush current was 191A; the wide jitter version had the advance reduced to 0.2s, and the peak value increased to 208A.

[0170] Example 1 (basic), with a complete TSN-FN configuration and a jitter of 27ns. After continuous operation for 24h, the SOC variance decreased from 3.5% to 1.4%, which is consistent with the main scheme index.

[0171] Example 2 (control: replacing the end fiber with a gigabit copper cable), the single-hop jitter of the copper cable link is 65ns, and the total jitter is about 180ns. The SOC variance only drops to 1.9%, and the temperature over-limit warning is shortened to 0.3s, confirming that the fiber-optic first-level symmetric link is the key to ensuring the dynamic balance accuracy.

[0172] Upper bound of synchronization error:

[0173]

[0174] where σ l represents the physical layer jitter of the l-th hop, and N = 5 is the deepest level at the site. This upper bound is directly written into the predictor fault tolerance threshold, and if it exceeds the limit, it enters the safety buffer mode.

[0175] Preferably, the order of the fractional-order dynamics prediction takes a value between 0.80 and 0.90, and this order is determined by minimizing the mean square error of the state of charge within a 10-second prediction window.

[0176] In the fractional-order dynamics framework, the order β determines the attenuation rate of the model for historical information. If the value is too low, the slow heat accumulation of the battery will be ignored; if the value is too high, the second-level fluctuations will be amplified into prediction noise. For the WIPI wireless power dynamic balance, the present invention limits β between 0.80 and 0.90, and automatically determines the optimal order β by minimizing the mean square error of the state of charge through a ten-second rolling window. opt . Using the three-point golden section iteration, the initial settings are β1 = 0.80, β2 = 0.90, and β3 = 0.85. Two rounds of iteration can converge the search interval to 0.01.

[0177] The evaluation index within the ten-second prediction window is:

[0178]

[0179] where represents the predicted value at the order β, SOC i,meas is the synchronized observed value, and N is the number of samples within the window. Take the β that minimizes the MSE as β opt .

[0180] Data and calculation process: Electrical quantities are sampled by an 18-bit ADC at 1 MS / s -1 and magnetic fluxes are sampled by a tunneling magnetoresistance array at 10 kS / s -1 . The two types of data are written with 64-bit PTP timestamps through a time-sensitive fiber optic network and resampled to a 0.1 s step. The synchronization sequence first undergoes three-level Daubechies-4 wavelet denoising, and then independent component analysis is used to strip the 20 kHz switching harmonics. Twenty-one groups of candidate betas are calculated in parallel using a GPU, and the complete search takes 6.5 ms.

[0181] If the beta obtained in the current iteration opt is less than 0.83, the system will increase the temperature weight w T by half a gear to offset the relaxation of the thermal constraint caused by insufficient memory depth. Full calibration is triggered every five seconds, and the most recent beta opt is used for fractional-order prediction.

[0182] Using an adaptive order, the mean square error of the ten-second rolling prediction reaches 0.9%, while the fixed integer-order Kalman-SOC method is 1.6%. The peak current of the inverter drops from 132 A to 118 A. After a single 1 MWh prototype runs continuously for 24 hours, the variance of the module SOC drops from 3.5% to 1.4%, and the peak value of the bus inrush current is reduced from 215 A to 191 A during a 5 kW step discharge event.

[0183] Basic implementation example: The adaptive order process is fully executed to obtain the above accuracy and current suppression effect.

[0184] Extended implementation example: Fixed order beta = 0.80. The GPU load is reduced by 12%, but the SOC variance can only be reduced to 1.8%, and the peak current increases by 10 A, indicating that dynamically selecting the order is crucial for the system balance performance.

[0185] As Figure 3 shown, power coupling control target data is generated based on the prediction results and initialization data, and magneto-acoustic metamaterial tuning data and execution control instruction data are generated accordingly;

[0186] The "state of charge vector " and "temperature upper bound vector " provided by the prediction layer every 0.1 s jointly characterize the future demands of 46 battery modules. The controller first maps these two sets of information into a "power coupling weight matrix" W G with the help of a graph neural network (GNN), and then uses a variational quantum optimizer (VQO) to search for a set of "minimum switching loss matrices" W Q within the feasible region. The two matrices are synthesized into a single target through a dynamic weight fusion module:

[0187] W T= αW G +(1 - α)W Q , α ∈ [0.6, 0.7]

[0188] where α is adaptively determined by the temperature weight W T ; the higher W T , the larger α, to enhance the priority of thermal equilibrium. The target matrix W T is quantized into a "magneto - acoustic metamaterial device tuning map" of 32×32 pixels and translated into an inverter control instruction sequence (duty cycle, phase, frequency, compensation capacitor).

[0189] In practical applications, the data at the input, the (46 - dimensional) output by the prediction layer and (46 - dimensional), the initial vector χ0 and the Hessian H0 enter the control module.

[0190] The graph neural network uses three - layer GraphSAGE, with 256 hidden nodes in each layer, the activation function ReLU, and the learning rate 0.001 (when it is lower than 0.001, the convergence speed drops by 48%, and when it is higher than 0.01, gradient oscillation occurs). The GNN outputs the coupling weight matrix W G and performs inference in TensorRT, with a latency of 2.3 ms.

[0191] The variational quantum optimizer, the quantum hardware simulation uses 20 qubits and a six - layer parameterized circuit, and the optimization goal is to minimize the switching loss and the inverter harmonic distortion. W is obtained through 32 steps of Adam 0.01 iteration Q , taking 1.8 ms.

[0192] The dynamic weight fusion module reads the temperature weight w T ∈ {1, 2, 3, 4}, and makes a linear mapping for α, α = 0.55 + 0.05w T . If the temperature rise risk is lifted in the next cycle, α automatically drops back to ensure the steady - state efficiency.

[0193] Pixel matrix mapping, normalizes and thresholds (threshold 0.35) W T to obtain a 32×32 pixel matrix, enabling the pixel representation to activate the magneto - acoustic column units; the matrix is uploaded to the FPGA LVDS bus to complete the tuning of the metamaterial device within 60 μs.

[0194] Inverter instruction generation, according to the row - column sum of W T , calculates a 46 - way amplitude - phase table, and then finely tunes the duty cycle using the single - iteration Newton method in combination with the Hessian H0; the complete instruction frame size is 512 B, and it is sent to the GaN matrix inverter via Aurora 4 - lane.

[0195] Preferably, the power coupling control target data is generated by linearly fusing the coupling weight matrix output by the graph neural network and the optimal coupling matrix output by the variational quantum optimizer with weight coefficients of 0.60 and 0.40. The graph neural network adopts a 3-layer GraphSAGE structure with 256 hidden nodes in each layer, and the variational quantum optimizer uses 20 qubits and a parameterized quantum circuit with a depth of 6 layers.

[0196] Power coupling control target matrix W T It undertakes two tasks simultaneously: one is to distribute the "power flow direction and amplitude" among 46 pairs of coupling coils, and the other is to take into account the switching losses of the inverter itself. To balance these two points within a single 0.1s cycle, the present invention entrusts the "battery-side demand" to a data-driven graph neural network (GNN) for processing, and the "switching loss compression" to a variational quantum optimizer (VQO) with stronger approximate global solution ability, and realizes unified scheduling through linear fusion.

[0197] GNN branch. The input graph takes 46 battery modules as nodes, and the edge weights are derived from the instantaneous coupling coefficient κ output by the digital twin. ij . It adopts three layers of GraphSAGE, with 256 hidden nodes in each layer and the activation function ReLU; the model is offline trained on 120 hours of historical data, with a learning rate of 0.001 and a batch size of 256. After training, the weights are frozen and TensorRT is used for inference, with a single-frame delay of 2.3ms. The output matrix is denoted as W G , and the numerical range is compressed to 0–1 through Sigmoid. The larger the value, the more the power should be tilted towards this coupling channel.

[0198] VQO branch. To make up for the shortcoming of GNN in suppressing switching harmonics, the second branch uses a parameterized quantum circuit with 20 qubits and a depth of 6 (the hardware is the IonQ emulator). The objective function weights and sums up the conduction loss, cut-off loss and total harmonic distortion coefficient of the inverter. Using Adam0.01 to optimize for 32 steps can give an approximate global optimal matrix W Q , with a running delay of 1.8ms.

[0199] Dynamic weight fusion module. The module dynamically adjusts the fusion coefficient α according to the temperature weight W T ∈{1,2,3,4}. When there is a high-temperature risk (w T =4), α rises to 0.70, making the data-driven part dominant; when the heat risk is lifted (w T =1), α drops to 0.55 to reduce the switching loss. Core fusion formula:

[0200] W T =αW G +(1-α)W Q ,α∈[0.55,0.70]

[0201] Symbol Explanation: W T is the target matrix, W G is the GNN output matrix, W Q is the VQO output matrix, and α is an adjustable fusion coefficient. Under the default working condition, α = 0.60 is set.

[0202] Magnetic - acoustic metamaterial tuning and instruction generation, W T Normalize and threshold (threshold 0.35) and map it into a 32×32 pixel matrix; turning on the pixel indicates that the corresponding magnetic - acoustic column unit is enabled. The matrix is sent to the metamaterial device FPGA via LVDS within 60 μs. The inverter controller generates a 46 - channel amplitude - phase table according to the T row and column sum of W, and uses single - step Newton iteration combined with the Hessian provided by the prediction layer to finely adjust the duty cycle, and generates a 512B instruction frame to be sent to the GaN power board.

[0203] Continuously operate for 24 hours at a 1MWh power station, adopting the above GNN + VQO fusion:

[0204] The mean square error of the ten - second rolling charge prediction is 0.9%, which is improved compared with using only GNN (1.1%) and using only VQO (1.4%).

[0205] The peak inrush current of the bus is 191A, which is 11% lower than that of the fixed coupling matrix of the present invention (215A).

[0206] The total harmonic distortion coefficient THD drops to 3.7%, which is 26% better than that of the pure GNN of the present invention.

[0207] Basic implementation example: Operate according to the above parameters (α = 0.60, 256 nodes of GraphSAGE, 6 - layer circuit of 20 qubits), and the variance of SOC drops from 3.5% to 1.4%.

[0208] Extended implementation example: Reduce the number of hidden nodes to 64, the inference delay is shortened by 1.1 ms, but the variance of SOC only drops to 1.8%, and the peak inrush current increases by 7A, indicating that 256 nodes are the trade - off point between accuracy and computing power.

[0209] Through the above details, the whole process of real - time generating the power coupling control target can be directly reproduced, and the accuracy, loss, and thermal safety advantages of the linear fusion of GNN and VQO in WIPI wireless power dynamic balancing can be verified.

[0210] Preferably, the magnetic - acoustic metamaterial tuning data is a 32 - row and 32 - column pixel matrix, where when the element value in the coupling matrix target data is greater than 0.35, the corresponding pixel is set to the on state, and the rest of the pixels are set to the off state, and the pixel switching completion time is not greater than 60 microseconds.

[0211] The magneto-acoustic metadevice is equivalent to a programmable magnetic field "grating". The system normalizes the power coupling target matrix W T to 0–1 first and then maps it to 32×32 on-chip pixels. Enabling a pixel drives the corresponding magneto-acoustic microcolumn to generate a local "acoustic-driven magnetism" effect at the 45 kHz resonance point, enabling real-time shaping of the coil magnetic flux distribution. To balance accuracy and hardware lifespan, the present invention sets a single-threshold trigger strategy: for the elements in W T greater than 0.35, they are marked as 1; otherwise, they are marked as 0. The pixel array completes one write within 60 μs, meeting the 100 Hz control beat.

[0212] The pixel generation process includes: after the power coupling controller calculates W T , the FPGA internally performs element-by-element comparison:

[0213]

[0214] Symbol meanings: P m,n is the pixel matrix element, and W T (m,n) is the corresponding element of the target matrix. The obtained binary matrix is packed into a 1024-bit LVDS frame and sent to the metadevice drive array in parallel through four 250 MHz channels; the pixel excitation level is 3.3 V, the drive power is 180 mW, and the total switching energy consumption is less than 11 μJ / frame.

[0215] The 60 μs switching guarantee means that the switching delay consists of two parts: 4 ns for LVDS transmission, 12 ns for FPGA-SAW drive decoding, and 44 μs for the microcolumn electromechanical response; a 16 μs margin is used for error detection and retransmission. If the CRC check fails, the controller reverts to the previous frame settings and resends them in the next cycle to ensure no disconnection.

[0216] The measured pixel threshold of 0.35 can reduce the peak bus inrush current to 191 A during a 5 kW load step; if the threshold is adjusted to 0.50, the number of enabled pixels decreases by 28%, but the peak inrush current rises back to 205 A. Although the threshold below 0.25 can further reduce the peak by 3 A, it increases the pixel switching power by 15%. After comprehensive trade-off, 0.35 is selected. The pixels are switched 1.4×10^5 times continuously for 24 h, and the microcolumn resonance frequency shift is less than 0.2%.

[0217] Basic implementation example: By fully using the 32×32 threshold matrix and 60 μs update, the SOC variance is reduced to 1.4%, and the peak current is 191 A.

[0218] Extended implementation example: Reducing the array resolution to 16×16 and synchronously reducing the number of pixel drive I / Os. The effect of reducing the SOC variance is maintained, but the peak current only drops to 200 A, indicating that the 32×32 resolution is more advantageous for high-power pulsating scenarios.

[0219] The inverter is driven according to the execution control instruction data, and the magneto-acoustic metastructure device is tuned according to the magneto-acoustic metastructure tuning data to realize wireless power coupling, the power flow data and the temperature data are measured in real time, and the coupling correction gradient is calculated by using the optimal transmission algorithm to update the magneto-acoustic metastructure device, and the power flow data, the temperature data and the execution control instruction data are formed into feedback data and returned to the power coupling control target data generation step, thereby realizing dynamic wireless power balancing between battery modules in a closed loop.

[0220] The magneto-acoustic metasurface (MAM) is composed of 32×32 independent micro-column units, each of which contains: a piezoelectric film transducer layer, whose driving voltage is V pz ; Magnetostrictive coating, its volume magnetic susceptibility is recorded as Flexible support substrate.

[0221] When the film is excited by a 45kHz surface acoustic wave, the magnetostrictive layer produces micrometer-level strain, which is equivalent to changing the unit's resonant frequency ω0. In the wireless power band (50kHz-150kHz) far below the ion oscillation frequency, the device can be described as an anisotropic permeability tensor:

[0222]

[0223] μ0 represents the vacuum permeability; F represents the resonance intensity factor, and is proportional to; ω0 represents the unit mechanical-magnetic coupling resonant frequency, which can be obtained by V pz Tuning; γ represents the damping coefficient; ω represents the working angular frequency of the coupling coil.

[0224] The above expression shows that changing V pz It can quickly shift ω0 and modify μ at the operating frequency. eff When μ eff decreases, the unit pair scatters through the magnetic field lines, which is equivalent to "closing" the pixel channel; when μ eff Increases, the magnetic flux is focused, and the channel is "opened".

[0225] The wireless electromagnetic propagation is converted into dynamic balance between modules. The external induction coil of the energy storage system and the MAM together form a "source-lens-receptor" magnetic circuit. The 32×32 patterned on-off pixels form a spatial magnetic lens, which can redistribute the coil mutual inductance M in milliseconds. ij :Open channel → Increase M ij →Increase coupling power; turn off the channel and reverse. Through the mutual inductance gain matrix ΔM(P) corresponding to the pixel matrix P, the total mutual inductance is:

[0226] M new =Mbase +ΔM(P)

[0227] The power flow direction and amplitude can be changed in real time, realizing wireless equalization without physical switches.

[0228] M base represents the mutual inductance matrix without tuning; ΔM represents the mutual inductance compensation matrix determined by the pixel state.

[0229] For controlling the link and performance, the magneto-acoustic pixel map is obtained by threshold mapping from the power coupling target matrix: when the target element W T (m,n)>0.35, the pixel is set to be on, corresponding to an increase in μ eff ; otherwise, it is off. The FPGA drives 32×32 piezoelectric pins through LVDS within 60 μs, shifting ω0 by 400 Hz, which is sufficient to change μ eff by 35% in the working frequency band. After the mutual inductance of the coil is updated according to the expression of the formula, the Sinkhorn optimal transport algorithm calculates the gradient to determine the pixels of the next frame, forming a 100 Hz closed loop.

[0230] The difference from the existing wireless electromagnetic propagation technology is that the existing MAM is used for antenna beams or near-field communication, only modulating the direction of electromagnetic radiation. The innovation points of the present invention are: 1. Working scale: the frequency band falls within 50 - 150 kHz, targeting inductive power coupling rather than GHz antennas; 2. Coupling target: tuning the mutual inductance M ij instead of radiation gain, directly acting on the power flow matrix, 3. Control link: combining digital twin - Sinkhorn closed loop to achieve gradient update for each frame of pixels, making the wireless equalization accuracy superior to the traditional slicing switch scheme.

[0231] First, the execution control instruction data drives 46 GaN three-level inverters, enabling each pair of coils to obtain the specified amplitude and phase; meanwhile, the magneto-acoustic metamaterial device switches the microcolumn units at the 45 kHz resonance frequency according to the 32×32 pixel matrix, shaping the transient magnetic field distribution. The system collects two types of feedback in real time: one is the bus power flow matrix P m (calculated from the current on the module side through a 1 MHz ADC), and the other is the module temperature vector T m (sampled by a thermocouple array at 200 kS / s -1 ). The two sets of data are packed with the current instruction frame into a feedback vector and sent back to the control target generation layer.

[0232] To quickly correct the field distribution within a 0.1 s cycle, the present invention formulates the problem of "transmitting the desired power flow P ref to the module" as entropy-regularized optimal transport:

[0233]

[0234] Here, C is the coupling loss cost matrix, H(W) is the entropy term, and with ε = 0.05, the gradient can be obtained by using 15 steps of Sinkhorn iteration:

[0235]

[0236] Then, update the entries to be adjusted in the pixel matrix according to the learning rate of 0.12. Each round of gradient-pixel mapping takes 38 μs, which meets the 60 μs switching window of the metadevice.

[0237] In practical applications, the size of the inverter command frame is 512 B, which is sent to the GaN board within 3 μs via Aurora-4lane; on the FPGA side, the duty cycle error is suppressed to ±0.3% by single-step Newton iteration. The pixel matrix is written into the SAW-FPGA through the LVDS parallel port, and the micro-column piezoelectric element is triggered by a 3.3 V driving level, with a response time of 44 μs. The power flow matrix is calculated by the Hall current probe and the 32-bit MAC unit inside the FPGA; the temperature vector is suppressed from measurement noise through third-order IIR low-pass filtering. All feedback data is sent back to the control server in 4 kB UDP frames, with a network round-trip of 110 μs.

[0238] Preferably, the power flow data is collected at a rate of 1,000,000 samples per second through an 18-bit analog-to-digital converter, and the temperature data is collected at a rate of 200,000 samples per second through a 16-bit analog-to-digital converter.

[0239] The power waveform of the wireless coupling system shows the strongest fundamental wave at around 20 kHz, and the switching harmonics are distributed up to 200 kHz. According to the Nyquist criterion, at least 400 kS / s -1 sampling bandwidth is required to completely reconstruct the instantaneous power. In the present invention, an 18-bit analog-to-digital converter (TI ADS8689) is selected to collect one channel each of the bus voltage v[n] and current i[n] in parallel, with a single-channel 1 MS / s -1 which can provide a 25 dB jitter margin at an oversampling factor of 200 kHz × 5; 18-bit effective bits ensure a signal-to-noise ratio of 68 dB even when the current lower limit is 0.2 A. The instantaneous power is calculated in the FPGA according to:

[0240] p[n] = v[n] · i[n]

[0241] and then down-converted to 100 Hz through a CIC-4 stage filter within a 0.1 s window to match the control beat.

[0242] The temperature channel is limited by the 10 kHz thermal response bandwidth of the thermocouple itself, and sampling with a 16-bit ADC (ADI AD7689) at 200 kS / s -1 can meet the oversampling requirements. Resolution conversion: input noise 5 μV rmsCorresponding to a temperature quantization step of ±0.012K, it is sufficient to detect a 2K temperature rise 0.8s in advance.

[0243] Hardware-software cooperation, the voltage and current probes are differential Hall plates, with a front-end 500kHz third-order Butterworth anti-aliasing filter; the thermocouple is amplified by AD8495 and added with a 4kHz RC low-pass filter. All ADC synchronous pins are connected to the 1MHz clock generated by the FPGA, and the falling edge triggers the hardware timestamp unit. The sampled word stream enters a 4kB circular buffer and is then written to the AXI bus in 64-bit frames by DMA, with an overall transmission delay of 2.1μs.

[0244] Data integrity and error control include:

[0245] Absolute time error: Compared with the optical fiber PTP master clock, the 24h RMS deviation is 27ns, and the peak-to-peak is 48ns; meeting the <50ns upper limit.

[0246] Power channel excitation noise: 1MS / s -1 In cooperation with 18 bits, the equivalent input noise is 41μV rms , and the converted power noise is 32mW.

[0247] Temperature channel hysteresis: 200kS / s -1 With a low-pass filter, the -3dB phase delay is 31μs, which can be ignored for a 0.1s control window.

[0248] Performance comparison experiment: Reduce the power ADC rate to 250kS / s -1 : The harmonic leakage of the instantaneous power waveform causes the ten-second rolling mean square error to increase from 0.9% to 1.4%, and the peak current of the inverter increases by 6A. Reduce the temperature ADC rate to 50kS / s -1 : The temperature rise detection lags by 0.4s, resulting in a delay in the thermal constraint trigger and a 7A increase in the peak value of the bus inrush current.

[0249] The basic implementation example uses 1MS / s -1 for power sampling and 200kS / s -1 for temperature sampling. Running on a 1MWh prototype for 24h, the variance of the module SOC decreases from 3.5% to 1.4%, and the temperature over-limit warning is advanced by 0.8s.

[0250] The extended implementation example limits the power sampling rate to 500kS / s -1 , the GPU load is reduced by 7%, but the SOC variance can only be reduced to 1.7%, verifying the necessity of the 1MS / s -1 sampling design.

[0251] This configuration achieves high-resolution capture of transient power and temperature, providing reliable input for subsequent optimal transmission gradients and closed-loop equalization. At the same time, its error budget and experimental data illustrate the rationality of the rate and resolution selection.

[0252] Preferably, the optimal transmission algorithm uses the entropy-regularized Sinkhorn algorithm, with an entropy regularization coefficient of 0.05 and 30 iterations. The iteration stops when the difference in the two-norm of the power transport matrix obtained from two consecutive iterations is less than.

[0253] In a wireless coupling network, each battery module is both a "power source" and a "power sink". Its instantaneous multi-to-multi power distribution can be abstracted as an optimal transport (OT) problem: mapping the supply vector p to the demand vector q and minimizing the cost matrix C composed of magneto-acoustic field fast-tunable channels. Directly solving the classical OT requires an iterative complexity of O(n 3 ), which cannot be completed within a 0.1s cycle. The present invention uses the entropy-regularized Sinkhorn algorithm, adding an entropy term εH(W) (ε = 0.05) to the objective function, transforming the problem into matrix calibration in the log-linear space, significantly accelerating convergence and naturally generating smooth gradients.

[0254] For the core iteration, let K = exp(-C / ε). The algorithm initializes with row vector u (0) and column vector v (0) both initialized to all ones, and then alternates scaling:

[0255]

[0256] The estimated value of the power transport matrix at the t-th iteration is:

[0257] W (t) = diag(u (t) )K diag(v (t) )

[0258] When the difference in the two-norm ||W (t) - W (t-1) ||2 is lower than 1×10 -3 , or 30 iterations are reached, it is considered convergent. Under GPU parallel implementation, each iteration takes 0.9 μs, and the total calculation delay is 27 μs.

[0259] In practical applications, the cost matrix C is linearly superimposed by three parts on the FPGA side: real-time coupling loss, switching loss, and routing weight between coils; the supply and demand vectors p and q are directly taken from the power margin output by the prediction layer. To prevent numerical underflow, the log-sum-exp technique is used in the CUDA kernel to maintain stability. The entropy coefficient 0.05 is obtained through offline grid search: if it is adjusted down to 0.02, although the peak current drops by another 2A, the number of pixel switches increases by 30%; if it is adjusted up to 0.10, the number of convergence steps is reduced to 15, but the power distribution error rises to 1.4%.

[0260] Continuous operation for 24h on a 1MWh prototype: The pixel matrix after Sinkhorn update suppresses the peak inrush current of the bus to 191A; if the unregularized linear programming solution is used instead (CPU 18ms), the peak drops to 188A, but the cycle times out by 80%. Reducing the iteration upper limit to 15 steps halves the delay, but the mean square error rises from 0.9% to 1.2%, confirming that the 30-step threshold combination meets the accuracy-real-time trade-off.

[0261] Basic implementation example: Entropy coefficient 0.05, 30 steps, threshold 1×10 -3 ; The SOC variance drops to 1.4%, and the control loop delay is 78μs.

[0262] Extended implementation example: Entropy coefficient 0.10, 15 steps, delay 43μs, but the SOC variance only drops to 1.7%, verifying the rationality of parameter selection.

[0263] Preferably, after generating the feedback data, the SHA-256 hash algorithm is used to calculate the hash value and write it into the blockchain storage, and the blockchain node completes the transaction confirmation including the hash value and the Coordinated Universal Time timestamp within 200 milliseconds.

[0264] Feedback data (power flow matrix P m , temperature vector T m , execution instruction frame C drv ) together with its UTC timestamp constitute a 4kB JSON message. The control server locally calls the OpenSSL digest library to calculate:

[0265] H = SHA256(JSON)

[0266] to obtain the 256-bit hash value H. The hash and the metadata (size, structure checksum) of the original message are encapsulated as a blockchain transaction and broadcast to a four-node consortium chain. The chain runs the RAFT-3f consensus (allowing one failure), with a block period of 100ms and a block capacity of 1MB, and can confirm the majority 2f + 1 votes within a single cycle. Experiments show that the network transmission is 25ms, the sorting is 60ms, and the block writing is 38ms, with a total delay of 123ms and a steady state lower than the 200ms upper limit.

[0267] In practical applications, node deployment: two local edge servers are located in the energy storage station, one is located in the dispatching center, and one is located in the cloud disaster recovery; a TLS1.3 encrypted tunnel is established between nodes, with a bandwidth of 1 Gbs -1 , and the round-trip time is 8 ms. Time synchronization: all nodes maintain a deviation of ±50 ns from the UTC source of the National Time Service Center through PTP to ensure the traceability of hash timestamps. Data pruning: only hash and index information are stored, and the complete original feedback is retained on the local read-only SSD. The blockchain storage overhead is controlled within 128 B / period. Abnormal policy: if two consecutive transactions are not confirmed within 3 s, the controller switches to the offline mode, only persists the hash locally and records the event code to avoid affecting the 100 Hz closed-loop.

[0268] A total of 1,728,000 transactions were chained during 48 consecutive hours of operation. The average confirmation delay was 137 ms, and the 99.9th percentile was 182 ms; 10,000 entries were randomly selected through the SHA-256 collision test, and the hash matching rate was 100%. A chain network packet loss fault occurred once, and RAFT automatically reselected the primary node. The longest confirmation delay was 268 ms, and the offline mode was not triggered.

[0269] Using the traditional centralized log server, this invention had two tampering risk scanning alarms for power and temperature logs within 24 hours; no further alarms were triggered after the evidence was stored on the consortium blockchain. The CPU load of the nodes increased from 12% to 19%, and the storage occupancy increased by 190 MB / day (acceptable).

[0270] Basic implementation example: four-node RAFT-3f, block period of 100 ms, hash threshold of 1 MB; the feedback chaining delay was stably <200 ms, and the SOC variance and thermal equilibrium index were maintained at 1.4% and 2 K.

[0271] Extended implementation example: increasing the block period to 500 ms, the average confirmation delay increased to 457 ms; although there was no direct impact on the closed-loop control, the operation and maintenance audit delay was too large. Therefore, a 100 ms period is recommended for configuration.

[0272] By combining the SHA-256 hash with a lightweight consortium blockchain, this invention realizes the tamper-proof storage and millisecond-level traceability of feedback data without interfering with the 100 Hz real-time control, providing a reliable basis for subsequent compliance audits and fault tracing.

[0273] Those skilled in the art should understand that the embodiments of this application can be provided as a method, a system, or a computer program product. Therefore, this application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0274] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A WIPI wireless power dynamic equalization method for grid connection of an energy storage system, characterized in that The method includes: Establishing a digital twin model, integrating system parameters that characterize coil geometry, battery module electrochemical impedance, and coil coupling into a unified model for describing the electromagnetic topology and electrochemical state of the energy storage system; Collecting and synchronizing electrical data and magnetic flux data, performing fractional-order dynamics prediction based on the digital twin model to obtain state of charge data, temperature constraint data, and generating control optimization initialization data; Generating power coupling control target data based on the prediction results and initialization data, and generating magneto-acoustic metamaterial tuning data and executing control instruction data accordingly; Driving the inverter according to the execution control instruction data, tuning the magneto-acoustic metamaterial device according to the magneto-acoustic metamaterial tuning data to achieve wireless power coupling, measuring power flow data and temperature data in real time, calculating the coupling correction gradient using the optimal transmission algorithm to update the magneto-acoustic metamaterial device, and forming feedback data from the power flow data, temperature data, and execution control instruction data to return to the power coupling control target data generation step, so as to achieve dynamic wireless power balance between battery modules within the closed loop.

2. The method according to claim 1, wherein The system parameters consist of the following data: coil center coordinates, coil normal vector, coil effective area, battery module zero-frequency internal resistance value, battery module polarization resistance value, and coupling coefficients between each pair of coils.

3. The method according to claim 2, wherein The digital twin model is established through the following steps: The coil coupling impedance matrix is calculated using a finite element mesh with an element size not greater than 0.5 mm in the electromagnetic part; the impedance curve of the battery module is fitted using a fractional-order equivalent circuit with a constant phase element in the electrochemistry part; with the goal of minimizing the mean square error between the model output and the measured impedance spectrum, a sparse identification dynamics algorithm with a penalty coefficient of 1×10 -4 is used to determine the fractional-order order and circuit parameters.

4. The method according to claim 1, wherein The synchronization of electrical data and magnetic flux data is completed using a time-sensitive optical fiber network, and the synchronization jitter is not greater than 50 nanoseconds.

5. The method according to claim 1, characterized in that, The order of the fractional-order dynamics prediction takes a value between 0.80 and 0.90, and this order is determined by minimizing the mean square error of the state of charge within a 10-second prediction window.

6. The method according to claim 5, wherein The power coupling control target data is linearly fused by the coupling weight matrix output by the graph neural network and the optimal coupling matrix output by the variational quantum optimizer with weight coefficients of 0.60 and 0.

40. Among them, the graph neural network adopts a 3-layer GraphSAGE structure with 256 hidden nodes in each layer, and the variational quantum optimizer uses 20 qubits and a parameterized quantum circuit with a depth of 6 layers.

7. The method according to claim 6, wherein The magneto-acoustic metamaterial tuning data is a pixel matrix of 32 rows and 32 columns. Among them, when the element value in the coupling matrix target data is greater than 0.35, the corresponding pixel is set to the on state, and the rest of the pixels are set to the off state. The pixel switching completion time is not greater than 60 microseconds.

8. The method according to claim 1, wherein The power flow data is collected by an 18-bit analog-to-digital converter at a rate of 1,000,000 samples per second, and the temperature data is collected by a 16-bit analog-to-digital converter at a rate of 200,000 samples per second.

9. The method according to claim 1, characterized in that The optimal transport algorithm uses the entropy-regularized Sinkhorn algorithm with an entropy regularization coefficient of 0.05 and 30 iterations. It stops iterating when the difference in the two-norm of the power transport matrices obtained from two consecutive iterations is less than 1×10 -3 -6 .

10. The method according to claim 1, wherein After generating the feedback data, the hash value is calculated using the SHA-256 hash algorithm and written into the blockchain for storage. The blockchain node completes the transaction confirmation including the hash value and the Coordinated Universal Time timestamp within 200 milliseconds.

Images