Energy-saving optimization method for electric energy storage device

Abstract

The invention relates to the technical field of power supply devices, in particular to an energy-saving optimization method for an electric energy storage device, and the method comprises the steps: fusing a lithium ion battery solid-liquid phase change kinetic equation, a flow battery fluid mechanics equation and a flywheel rotor rigid-flexible coupling equation through a neural differential equation; constructing a dynamic meta-knowledge base containing cross-time scale nonlinear interaction features; performing dynamic parameter compensation on the polarization voltage hysteresis effect and the electrolyte flow resistance coupling effect by using a physical information neural network to generate a decoupled residual feature vector; according to the method, multi-physics field coupling modeling is achieved through the dynamic meta-knowledge base, and the energy efficiency optimization effect of the energy storage device is improved by combining the physical information neural network and the quantum optimization algorithm.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply devices, and particularly to an energy-saving optimization method for an electric energy storage device. Background Art

[0002] Electric energy storage devices store and release electrical energy through electrochemical, physical, or electromagnetic mechanisms. Common types include lithium-ion batteries, flow batteries, supercapacitors, and flywheel energy storage systems. Lithium-ion batteries rely on the redox reaction of lithium ions intercalating and deintercalating between the positive and negative electrodes, and have high energy density and cycle stability; flow batteries use the circulation of electrolyte between external storage tanks and stacks for energy storage, and are suitable for large-scale long-term energy storage scenarios; supercapacitors achieve rapid charge separation based on the electric double layer or pseudocapacitance effect, and have millisecond-level response and millions of cycle characteristics; flywheel energy storage converts electrical energy into mechanical energy through a high-speed rotor and uses magnetic levitation technology to reduce friction losses, and is suitable for short-term high-frequency power compensation. Various energy storage devices achieve power interaction with the power grid through power electronic converters, supporting functions such as frequency regulation, renewable energy consumption, and microgrid grid connection and disconnection switching.

[0003] The technical pain points affecting the energy-saving control accuracy of electric energy storage devices stem from the synergistic effect of multi-dimensional non-linear dynamic characteristics and system-level coupling interference. In the field of electrochemical energy storage, the dynamic hysteresis effect between the internal polarization voltage and the state of charge of lithium-ion batteries leads to parameter mismatch of the equivalent circuit model, especially causing prediction deviation of the charge and discharge cut-off voltage during frequent switching of operating conditions; flow batteries are affected by the dynamic coupling of the ion migration rate of the electrolyte and the pressure drop in the stack flow channels, resulting in lag in the compensation of pump power consumption by the energy management strategy. In power-type energy storage systems, the competitive relationship between the electric double layer charge relaxation effect and the pseudocapacitance reaction kinetics of supercapacitors weakens the linear control accuracy of the terminal voltage in the constant power mode; in flywheel energy storage, due to the interaction between the gyroscopic effect of the high-speed rotor and the non-linear friction force of the bearing, it is difficult to suppress torque fluctuations during the mechanical energy feedback stage. When the power electronic converter realizes multi-modal switching, the cross-modulation interference between the DC bus voltage ripple and the AC grid connection harmonic component exacerbates the attenuation of the phase margin of the power closed-loop control. The above dynamic processes and external environmental disturbances form multi-dimensional time-varying constraints, significantly reducing the robustness of the energy efficiency optimization of the energy storage system on a wide time scale (seconds to hours). Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an energy-saving optimization method for an electric energy storage device, which solves the problems of model mismatch and control deviation of the energy storage device caused by multi-physical field coupling and non-linear dynamic characteristics.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides an energy-saving optimization method for an electric energy storage device, including: Build a multi - physical - field coupling model of the energy storage device. By using neural differential equations to fuse the solid - liquid phase change kinetic equation of lithium - ion batteries, the fluid mechanics equation of flow batteries, and the rigid - flexible coupling equation of flywheel rotors, generate a dynamic meta - knowledge base containing cross - time - scale non - linear interaction characteristics; Based on the physical constraint conditions of the dynamic meta - knowledge base, use a physics - informed neural network to perform dynamic parameter compensation processing on the polarization voltage hysteresis effect and the electrolyte flow resistance coupling effect of the energy storage device, and output a residual feature vector containing the decoupled state - of - charge characteristics; Input the residual feature vector into a multi - agent reinforcement learning framework to generate the hierarchical control parameters of the supercapacitor terminal voltage and the magnetic bearing damping compensation amount of the flywheel energy storage, and form a cooperative control strategy; According to the power distribution threshold of the cooperative control strategy, use a quantum - inspired optimization algorithm to transform the DC bus ripple suppression and harmonic distortion rate optimization objectives into the Hamiltonian constraint conditions of the quantum annealing model, and generate the converter modulation command; In the federated reinforcement learning framework, distill the cloud - side global optimization strategy to the edge controller through an imitation - learning generative adversarial network, and at the same time inject the real - time operating condition data fed back by the edge controller into the multi - physical - field coupling model for parameter recalibration, forming a closed - loop energy - efficiency enhancement loop that crosses physical models, data - driven, and quantum optimization.

[0006] Furthermore, for the energy - saving optimization method of the electric energy storage device described in the present invention, the construction of the multi - physical - field coupling model of the energy storage device includes: Use neural differential equations to fuse the solid - liquid phase change reaction kinetic equation of lithium - ion batteries, the fluid mechanics equation of flow batteries, and the rigid - flexible coupling equation of flywheel rotors; Extract the spatio - temporal correlation features of multi - source heterogeneous data of the energy storage device through a Transformer architecture with spatio - temporal attention mechanisms, and generate a meta - knowledge base containing cross - scale non - linear interaction relationships, which is used to provide physical constraint conditions for the physics - informed neural network.

[0007] Furthermore, for the energy - saving optimization method of the electric energy storage device described in the present invention, the dynamic parameter compensation includes: On the lithium - ion battery side of the energy storage device, embed the Butler - Volmer electrochemical equation into a long - short - term memory network to decouple the cross - sensitivity of the state - of - charge and the health state in real time; On the flow - battery side, build a graph neural network to predict the flow - channel pressure - drop distribution, combine federated learning to aggregate the operation data of multi - node pump units, dynamically optimize the feed - forward compensation coefficient matrix of the electrolyte flow rate, and use the compensation residual as the state - observation input of the multi - agent reinforcement learning.

[0008] Furthermore, in the energy-saving optimization method of the electric energy storage device of the present invention, the residual feature vector is input into a multi-agent reinforcement learning framework to generate hierarchical control parameters for the supercapacitor terminal voltage and the magnetic bearing damping compensation amount of the flywheel energy storage, and the collaborative control strategy formed includes: A hierarchical attention mechanism is adopted for the control of the supercapacitor terminal voltage, and the charge relaxation time constant is matched through a sliding mode parameter dynamic adjustment strategy; A rotor deformation observer based on neural radiance fields is deployed on the flywheel energy storage side, and the magnetic bearing damping compensation amount is generated in combination with the proximal policy optimization algorithm, so that the quantum-inspired optimization algorithm receives the torque fluctuation suppression threshold as a modulation target constraint.

[0009] Furthermore, in the energy-saving optimization method of the electric energy storage device of the present invention, the multi-objective optimization includes: Suppressing the DC bus ripple, optimizing the harmonic distortion rate, and minimizing the switching loss are transformed into the Hamiltonian of the quantum annealing model; The harmonic spectrum distribution characteristics are predicted through a deep residual network, and the distribution characteristics are fed back to the multi-physical field coupling model to update the boundary conditions of the neural differential equation.

[0010] Furthermore, in the energy-saving optimization method of the electric energy storage device of the present invention, the formation of a closed-loop energy efficiency enhancement loop that crosses physical models, data-driven, and quantum optimization includes: Based on the output of the multi-agent reinforcement learning framework, a multi-agent proximal policy optimization algorithm is deployed in the cloud, and the multi-node pump operation data aggregated by the federated learning is fused through a differential privacy protection mechanism; On the edge side, an imitation learning generative adversarial network is used to transform the cloud policy into a lightweight control instruction, and the instruction execution result is reversely injected into the generated dynamic meta-knowledge base to trigger the recalibration of the boundary condition parameters of the neural differential equation.

[0011] Furthermore, in the energy-saving optimization method of the electric energy storage device of the present invention, the federated reinforcement learning framework includes: Based on the physical constraints of the dynamic meta-knowledge base, a hierarchical reward mechanism is designed to map the dynamic efficiency of the energy storage system to the depth of the potential energy well, and a long-term energy efficiency optimization goal is constructed; A dynamic energy efficiency entropy index is generated through a multi-entropy maximization controller, and the index is used to constrain the charge relaxation threshold in the supercapacitor power distribution interval of the generated collaborative control strategy.

[0012] Furthermore, in the energy-saving optimization method of the electric energy storage device of the present invention, it further includes: An adversarial curriculum generator is embedded in the federated learning framework to synthesize training scenarios based on the critical operating condition data stream; The federated learning framework performs differential privacy processing on the operation data of multi-node pumping machines to generate differentially private protected data; The distribution difference degree between the synthetic scenario and the differentially private protected data is evaluated through a graph attention network, and the training course difficulty coefficient of the multi-agent reinforcement learning framework is dynamically adjusted.

[0013] Furthermore, the energy-saving optimization method for the electric energy storage device according to the present invention further includes: Construct a non-steady Pareto front search engine, and dynamically allocate multi-objective optimization weights according to the dynamic energy efficiency entropy index and harmonic spectrum distribution characteristics; Input the weight sequence into a quantum-inspired optimization algorithm, and update the Hamiltonian constraint conditions of the quantum annealing model to achieve real-time alignment of the converter modulation strategy and the dynamic energy efficiency optimization target.

[0014] Furthermore, for the energy-saving optimization method for the electric energy storage device according to the present invention, the closed-loop energy efficiency enhancement loop includes: Inject the synthesized critical operating condition data stream reversely into the dynamic meta-knowledge base to trigger the reparameterization of the fluid dynamics equation of the flow battery in the neural differential equation; The frequency response interval of the flywheel energy storage magnetic bearing damping compensation amount in the generated cooperative control strategy is constrained through a hierarchical reward mechanism and the dynamic energy efficiency entropy index; Synchronously feedback the updated dynamic meta-knowledge base parameters to the physics-informed neural network and the quantum optimization algorithm to form a closed-loop enhancement loop for cross-model parameter compensation, control strategy generation, and modulation optimization.

[0015] Advantages of the present invention; The advantages of the present invention are reflected in that accurate modeling of the multi-physical field coupling model is realized through the dynamic meta-knowledge base, and real-time dynamic compensation of the polarization voltage hysteresis and electrolyte flow resistance effect is carried out by using the physics-informed neural network, effectively solving the cross-scale modeling mismatch problem between the electrochemical energy storage and the mechanical energy storage system; the cooperative mechanism of the multi-agent reinforcement learning framework and the quantum-inspired optimization converts the power distribution parameters and harmonic suppression targets into quantum annealing constraint conditions, significantly improving the robustness of the converter modulation strategy against non-linear dynamic interference; the closed-loop energy efficiency enhancement loop constructed by the federated reinforcement learning framework realizes the dynamic adaptation of the physical mechanism model and the data-driven compensation through real-time feedback data of the edge computing node to drive the re-calibration of the neural differential equation parameters, and finally achieves the full-life cycle energy efficiency optimization of the energy storage system within the time range from seconds to hours, improving the dynamic response accuracy and energy conversion efficiency compared with the traditional method. Description of the drawings

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on the accompanying drawings.

[0017] Figure 1 It is a flowchart of an energy-saving optimization method for an energy storage device provided by an embodiment of the present invention. Specific embodiments

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The technical solutions provided by the embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. To better understand the objectives of the present invention, the present invention is further described in detail below.

[0019] Please refer to Figure 1 , the present invention provides an energy-saving optimization method for an energy storage device, including: Step S101, constructing a multi-physical field coupling model of the energy storage device, fusing the solid-liquid phase change kinetics equation of the lithium-ion battery, the fluid mechanics equation of the flow battery, and the rigid-flexible coupling equation of the flywheel rotor through neural differential equations, generating a dynamic meta-knowledge base containing cross-time-scale nonlinear interaction characteristics. The dynamic meta-knowledge base is a distributed database generated by fusing the multi-physical field coupling model through neural differential equations, storing cross-time-scale characteristics such as the solid-liquid phase change rate of the lithium-ion battery, the electrolyte flow velocity distribution of the flow battery, and the stress gradient of the flywheel rotor, and is used to provide spatio-temporal correlation constraint conditions for the physics-informed neural network; The spatio-temporal attention mechanism includes a temporal attention layer and a spatial attention layer. The temporal attention layer analyzes the correlation between second-level charge-discharge data and hourly-level aging data through a gated recurrent unit, and the spatial attention layer captures the difference in the monomer voltage distribution within the battery stack through a convolutional kernel and fuses multi-modal features through a multi-head mechanism.

[0020] Step S102, based on the physical constraint conditions of the dynamic meta-knowledge base, using the physics-informed neural network to perform dynamic parameter compensation processing on the polarization voltage hysteresis effect and the electrolyte flow resistance coupling effect of the energy storage device, and outputting a residual feature vector containing the decoupled state-of-charge characteristics; The Butler–Volmer electrochemical equation is a differential equation describing the kinetics of electrode reactions, and its expression is:

[0021] i: Current density, unit: A / m²; : Exchange current density, unit: A / m²; α: Transfer coefficient (dimensionless); F: Faraday constant, with a value of 96485 C / mol; η: Overpotential, unit: V; R: Gas constant, with a value of 8.314 J / (mol·K); T: Temperature, unit: K.

[0022] In step S103, input the residual feature vector into the multi-agent reinforcement learning framework to generate the hierarchical control parameters of the supercapacitor terminal voltage and the flywheel energy storage magnetic bearing damping compensation amount, and form a cooperative control strategy; In step S104, according to the power distribution threshold of the cooperative control strategy, use the quantum-inspired optimization algorithm to transform the DC bus ripple suppression and harmonic distortion rate optimization objectives into the Hamiltonian constraint conditions of the quantum annealing model, and generate the converter modulation command; the quantum-inspired optimization algorithm specifically includes the following steps: Map the DC bus ripple suppression objective to the local magnetic field strength of the qubit, transform the harmonic distortion rate optimization objective into the antiferromagnetic coupling relationship between adjacent qubits, and use the quantum annealing algorithm to solve the ground state of the Hamiltonian to generate the converter modulation command.

[0023] In step S105, in the federated reinforcement learning framework, distill the cloud global optimization strategy to the edge controller through the imitation learning generative adversarial network, and at the same time inject the real-time working condition data fed back by the edge controller into the multi-physical field coupling model for parameter recalibration, forming a closed-loop energy efficiency enhancement loop that crosses physical models, data-driven, and quantum optimization.

[0024] When constructing the multi-physical field coupling model of the energy storage device, use the neural differential equation to couple and solve the solid-liquid phase change process kinetic equation of the lithium-ion battery and the Navier-Stokes fluid control equation of the flow battery. At the same time, introduce the rigid-flexible coupling vibration equation of the flywheel rotor, and use the variational autoencoder to extract the features of the temperature field, flow velocity field, and stress field data at multiple time scales. This process generates a dynamic meta-knowledge base containing the second-level charge and discharge response and the hourly cycle decay correlation features, and the meta-knowledge base is stored in the distributed graph database for providing cross-domain constraint conditions for subsequent physical parameter compensation.

[0025] Based on the physical constraint conditions of the dynamic meta-knowledge base, a physical information neural network is designed on the lithium-ion battery side. The Butler-Volmer electrode reaction equation is embedded into the Long Short-Term Memory (LSTM) network in a hard constraint manner to construct a state of charge estimation model with electrochemical mechanism. On the flow battery side, a graph convolutional network is used to model the electrolyte flow channel topology, and combined with the federated learning framework to aggregate the data of multi-pumping station pressure sensors to generate an electrolyte flow velocity compensation matrix. The outputs of the two compensation mechanisms are fused through a feature splicing layer into a residual vector containing the decoupled state of charge and flow resistance characteristics, and its dimension matches the input layer structure of the multi-agent reinforcement learning framework.

[0026] After the residual feature vector is input into the multi-agent reinforcement learning framework, the supercapacitor control branch uses a hierarchical attention mechanism to analyze the charge relaxation process, and dynamically adjusts the sliding mode control surface parameters through the double-delay deep deterministic policy gradient algorithm; the flywheel energy storage branch deploys a rotor deformation observer based on the neural radiance field, and uses the proximal policy optimization algorithm to generate the active damping compensation amount of the magnetic bearing. The outputs of the two branches are fused through a policy fusion layer to generate a power distribution command and a damping control signal, forming a collaborative control strategy that meets the multi-time scale response requirements.

[0027] The power distribution threshold of the collaborative control strategy is mapped to the Hamiltonian of the Ising model through the quantum approximate optimization algorithm, and the goals of suppressing the DC bus ripple and optimizing the harmonic distortion rate are encoded as the ground state search problem of the spin glass system. During the quantum annealing process, an adaptive tuning algorithm is used to dynamically adjust the transverse magnetic field strength, and combined with the harmonic frequency spectrum distribution characteristics predicted by the deep residual network, a carrier phase shift modulation instruction set that meets the multi-objective Pareto optimality is generated.

[0028] In the federated reinforcement learning framework, the multi-agent proximal policy optimization algorithm deployed in the cloud fuses the real-time working condition data uploaded by the edge nodes through the differential privacy mechanism to generate a global optimization strategy; the edge side uses a generative adversarial network architecture for policy distillation to simplify the complex policy into a lightweight control instruction. The working condition data collected after the edge controller executes the instruction is injected into the multi-physical field coupling model through the backpropagation channel, triggering the online calibration of the boundary conditions in the neural differential equation, forming a closed-loop enhancement loop covering physical mechanism modeling, data-driven compensation and quantum optimization modulation. This loop realizes the continuous optimization of the energy efficiency during the whole life cycle of the energy storage system through the parameter iteration update mechanism of the dynamic meta-knowledge base.

[0029] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the construction of the multi-physical field coupling model of the energy storage device includes: Using neural differential equations to fuse the solid-liquid phase change reaction kinetics equation of lithium-ion batteries, the fluid mechanics equation of flow batteries, and the rigid-flexible coupling equation of flywheel rotors; Extract the spatio-temporal correlation features of multi-source heterogeneous data of the energy storage device through the Transformer architecture with spatio-temporal attention mechanism, and generate a meta-knowledge base containing cross-scale non-linear interaction relationships, which is used to provide physical constraint conditions for the physics-informed neural network.

[0030] When constructing the multi-physics field coupling model of the energy storage device, first express the solid-liquid phase change reaction kinetic equation of the lithium-ion battery in the form of partial differential equations, discretize it using the continuous-time neural network architecture in the neural differential equation, and solve the phase change interface evolution process through the adaptive step-size algorithm. At the same time, use the spectral method for spatial discretization of the fluid mechanics equation of the flow battery, and combine the Reynolds-averaged Navier-Stokes model to capture the turbulence effect. The rigid-flexible coupling equation of the flywheel rotor is constructed through the Hamilton variational principle, and the symplectic algorithm is used to maintain the energy conservation characteristics of the numerical calculation. The above three types of equations are coupled through a multi-task learning framework with a shared hidden layer, and the attention mechanism is used to dynamically allocate the weight coefficients of different physical fields at the coupling boundary to form a unified differential operator that can represent cross-time-scale interaction effects.

[0031] During the solution process of the neural differential equation, a variational autoencoder is used to compress the features of the multi-physics field output, and the key modal features of the solid-liquid phase change rate, electrolyte flow rate, and rotor stress distribution are extracted. The feature decoupling ability is enhanced through adversarial generation training, so that the latent variables output by the encoder respectively correspond to the independent action components of chemical kinetics, fluid mechanics, and mechanical dynamics. The finally generated coupling model parameters are stored in the distributed parameter server to form a dynamic meta-knowledge base containing multi-physics field coupling relationships.

[0032] Based on the above coupling model, process the operation data of the energy storage device through the Transformer architecture. Construct a spatio-temporal attention mechanism, in which the time attention layer analyzes the correlation pattern between second-level charge and discharge data and hour-level cycle aging data, and the spatial attention layer captures the monomer distribution difference in the battery stack and the pressure drop gradient of the flow channels of the flow battery. The multi-head attention mechanism is used to process the temperature sensor data, voltage and current time series data, and vibration spectrum data in parallel, and a tensor representation with spatio-temporal correlation is generated through cross-modal feature crossing. The tensor is input into the feed-forward neural network for feature recombination to generate a meta-knowledge graph containing cross-scale non-linear interaction relationships.

[0033] The meta-knowledge graph is lightweight processed using knowledge distillation technology, and node embedding vectors are constructed through a graph neural network, where each node corresponds to specific physical constraint conditions. The edge weights in the graph are dynamically matched with the parameters of the residual connection layer of the physical information neural network, so that the flow channel optimization constraints of the flow battery are automatically mapped to the adjacency matrix update rule of the graph convolutional network, and the phase change constraints of the lithium-ion battery are associated with the forgetting gate control mechanism of the long short-term memory network. Through the online incremental learning mechanism, newly discovered physical constraint relationships during the operation of the energy storage device are continuously supplemented to the meta-knowledge base in real time, forming a continuously evolving cross-domain constraint supply system.

[0034] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the dynamic parameter compensation includes: On the lithium-ion battery side of the energy storage device, the Butler–Volmer electrochemical equation is embedded into the long short-term memory network to decouple the cross-sensitivity of the state of charge and the state of health in real time; On the flow battery side, a graph neural network is constructed to predict the flow channel pressure drop distribution. Combining federated learning to aggregate the operation data of multi-node pump units, the feedforward compensation coefficient matrix of the electrolyte flow rate is dynamically optimized, and the compensation residual is used as the state observation input of the multi-agent reinforcement learning. The specific federated learning framework includes: edge nodes train the graph neural network model through local data sets to generate model gradients; the central server aggregates the gradients using a dynamic weighted average algorithm to update the global model parameters; Gaussian noise is added to the gradients through a differential privacy mechanism.

[0035] In the parameter compensation process of the lithium-ion battery, the activation overpotential term and the concentration overpotential term of the Butler–Volmer electrochemical equation are respectively mapped to the input gate and the forgetting gate of the long short-term memory network, and the time-series modeling of the electrochemistry reaction kinetic characteristics is realized through the gating mechanism. A residual connection structure is used to separate the phase change characteristics of the electrode material and the electrolyte diffusion characteristics, and a two-channel feature extraction network is constructed, where the main channel learns the dynamic change mode of the state of charge, and the auxiliary channel captures the degradation trajectory of the state of health. The outputs of the two channels are processed by orthogonalization constraints, and the cross-sensitivity between state variables is eliminated through covariance matrix decomposition to generate the decoupled state of charge estimation value and the state of health degradation rate parameter.

[0036] For the prediction of the pressure drop in the flow channels of a flow battery, a flow channel topology model is constructed based on a graph neural network. The electrolyte flow channel is decomposed into a set of nodes and edges. The node attributes include the flow velocity, pressure, and temperature sensor data, and the edge weights represent the flow channel cross-sectional change rate and the bending angle. A message passing mechanism is used to aggregate the features of adjacent nodes, and a heat map of the flow channel pressure drop distribution is generated through multiple layers of graph convolution operations. Combining with the federated learning framework, a global model is deployed on the central server to receive the pump speed, valve opening, and pressure difference data uploaded by each edge node. The dynamic weighted average algorithm is used to update the weight parameters of the graph neural network, and a feedforward compensation coefficient matrix adapted to the working conditions of multiple nodes is generated.

[0037] The feedforward compensation coefficient matrix is dynamically optimized through an online incremental learning mechanism. The compensation residual is defined as the integral of the time series error between the actual flow velocity and the predicted value. The sliding window mechanism is used to extract the frequency domain features of the residual signal. The residual feature vector and the decoupling parameters of the lithium-ion battery are processed for spatio-temporal alignment to construct a multi-dimensional state observation tensor, whose dimension matches the input layer structure of the multi-agent reinforcement learning framework. After being normalized, this tensor is input into the policy network to provide real-time state feedback for the power distribution of the supercapacitor and the flywheel damping control.

[0038] In the multi-modal fusion process of the compensation residual, a feature attention mechanism is designed to dynamically allocate the weight coefficients of the parameters of the lithium-ion battery and the flow battery. The differentiable decision tree is used to analyze the correlation between the residual features and the system energy efficiency, generate a feature importance score, and dynamically adjust the feature combination method of the state observation vector in the multi-agent reinforcement learning framework according to the score. The fused feature vector extracts the long-term and short-term dependencies through a temporal convolutional network, providing a state representation basis for generating the cooperative control strategy across time scales.

[0039] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the residual feature vector is input into the multi-agent reinforcement learning framework to generate the hierarchical control parameters of the supercapacitor terminal voltage and the magnetic bearing damping compensation amount of the flywheel energy storage, and the cooperative control strategy formed includes: The hierarchical attention mechanism is adopted for the supercapacitor terminal voltage control, and the sliding mode parameter dynamic adjustment strategy is used to match the charge relaxation time constant; A rotor deformation observer based on the neural radiance field is deployed on the flywheel energy storage side. Combining with the proximal policy optimization algorithm, the magnetic bearing damping compensation amount is generated, so that the quantum-inspired optimization algorithm receives the torque fluctuation suppression threshold as the modulation target constraint.

[0040] In the terminal voltage control of supercapacitors, a hierarchical attention mechanism is constructed to handle spatial and temporal dimensional features. The spatial attention layer analyzes the voltage distribution differences of individual cells within the battery stack and generates a region-sensitive weight matrix through convolutional kernels. The temporal attention layer extracts the relaxation time features of the charge-discharge curve and uses gated recurrent units to capture millisecond-level transient responses and second-level steady-state conversion modes. After the outputs of the two attention layers are fused, a dynamic sliding mode surface parameter adjustment instruction is generated. This instruction dynamically matches the time-varying characteristics of the electric double-layer charge relaxation process of the supercapacitor through an adaptive law, and dynamically aligns the sliding mode control bandwidth with the charge migration rate.

[0041] Based on the dynamic sliding mode surface parameters, a sliding mode observer is designed to estimate the remaining capacity deviation during the charge relaxation process. The sliding mode reaching law is derived through Lyapunov stability analysis, and the output of the observer is coupled with the modulation frequency parameters of the quantum-inspired optimization algorithm to establish a quantitative relationship between the terminal voltage control accuracy and the converter switching loss. The observer residual signal is fed back to the hierarchical attention mechanism to form a parameter self-correction loop.

[0042] On the flywheel energy storage side, a rotor deformation observer based on neural radiance fields is deployed. The deformation data of the rotor surface is collected through a multi-view laser displacement sensor to construct a three-dimensional point cloud spatial distribution model. The implicit neural representation method is used to encode the point cloud data into a continuous radiance field function, and the rotor dynamic deformation surface is reconstructed through voxel rendering technology. The deformation gradient tensor is input into the value function network of the proximal policy optimization algorithm to calculate the torque fluctuation components in each degree-of-freedom direction.

[0043] The proximal policy optimization algorithm adopts a trust region constraint mechanism to generate the multi-phase winding current compensation amount of the magnetic bearing according to the torque fluctuation spectrum distribution characteristics. During the compensation amount generation process, the smoothness constraint of the compensation instruction is maintained through the policy gradient projection algorithm to avoid introducing high-frequency oscillation components into the second harmonic. The generated damping compensation amount threshold is used as a boundary condition to input the quantum annealing optimization process, which restricts the search space of the converter carrier phase-shifting angle, so that the harmonic suppression frequency band of the modulation strategy forms a dynamic avoidance zone with the flywheel rotor resonance frequency band.

[0044] During the execution of the coordinated control strategy, the power distribution coefficient output by the supercapacitor terminal voltage control and the damping compensation amount threshold of the flywheel are coupled through a feature cross-attention mechanism. The cross-attention layer analyzes the coordination relationship between the two in the millisecond-level transient response and second-level steady-state maintenance stages, and generates a joint optimization target weight matrix. This matrix is input into the Hamiltonian construction module of the quantum-inspired optimization algorithm to drive the multi-objective Pareto front search process to converge to the optimal solution of system-level energy efficiency.

[0045] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the multi-objective optimization includes: Convert the DC bus ripple suppression, harmonic distortion rate optimization, and switching loss minimization into the Hamiltonian of a quantum annealing model; Predict the harmonic spectrum distribution characteristics through a deep residual network, and feedback the distribution characteristics to the multi-physics field coupling model to update the boundary conditions of the neural differential equation.

[0046] During the construction of the quantum annealing model, the DC bus ripple suppression target is encoded as the local field strength parameter of the spin system, and its strength is inversely proportional to the effective value of the ripple voltage; the harmonic distortion rate optimization target is mapped to the interaction coefficient between adjacent spins, and the corresponding relationship between the harmonic order and the coupling strength is constructed through the topological sorting algorithm; the switching loss minimization target is transformed into the system ground state energy constraint condition, and the Lagrange multiplier method is used to balance the loss weight and the ripple suppression requirement. The three types of optimization targets are normalized to form a composite Hamiltonian, which is input into the quantum annealing solver for ground state search.

[0047] The deep residual network is constructed using a multi-scale convolutional kernel structure. The input layer receives the output current waveform data of the converter, and the time-frequency domain joint features are extracted through the stacking of residual blocks. The skip connection structure fuses the shallow detail features and the deep semantic features, and the output layer generates a harmonic spectrum distribution vector, which includes the amplitude and phase information of each harmonic and the inter-harmonic clustering features. During the network training process, the adversarial sample enhancement technology is used to improve the robustness to the grid background harmonic interference.

[0048] The harmonic spectrum distribution vector extracts the key frequency band energy ratio parameter through the feature selection module, and generates a boundary condition update vector after being weighted by the attention mechanism. This vector is input into the neural differential equation solver of the multi-physics field coupling model to adjust the viscosity coefficient boundary value in the hydrodynamic equation of the flow battery and correct the interface energy parameter of the lithium-ion battery phase change equation. The updated boundary conditions trigger the recalculation of the coupling model to generate a new set of dynamic meta-knowledge base parameters.

[0049] The optimal spin configuration output by the quantum annealing solution is decoded into a carrier phase shift angle sequence, and combined with the harmonic sensitive frequency band distribution predicted by the deep residual network, the dead time and switching frequency of the pulse width modulation wave are dynamically adjusted. The current waveform data collected during the execution of the modulation strategy forms a closed-loop feedback, and the convolution kernel weights of the residual network are updated through the online learning mechanism, so that the harmonic prediction model continuously adapts to the characteristic drift caused by the aging of the converter.

[0050] The interaction between the multi-objective optimization result and the cooperative control strategy is realized through the strategy fusion layer, which spatially and temporally aligns the modulation instructions generated by the quantum annealing and the power distribution parameters of the multi-agent reinforcement learning. The dynamic time warping algorithm is used to compensate for the control delay during the alignment process, generate a joint optimization instruction set to drive the operation of the energy storage device, and at the same time inject the execution effect data into the federated learning framework for global policy update.

[0051] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the formation of a closed-loop energy efficiency enhancement loop across physical models, data-driven, and quantum optimization includes: Based on the output of the multi-agent reinforcement learning framework, deploy the multi-agent proximal policy optimization algorithm in the cloud, and fuse the multi-node pump operation data aggregated by the federated learning through the differential privacy protection mechanism; On the edge side, use the imitation learning generative adversarial network to convert the cloud policy into lightweight control instructions, and inject the instruction execution result back into the generated dynamic meta-knowledge base to trigger the re-calibration of the boundary condition parameters of the neural differential equation.

[0052] When deploying the multi-agent proximal policy optimization algorithm in the cloud, construct a hierarchical policy network structure, where the global policy network receives the multi-node pump operation data aggregated by the federated learning framework, including pressure, flow rate, and temperature time series information. Use the differential privacy protection mechanism to add noise to the data, and balance the data utility and privacy protection intensity by dynamically adjusting the privacy budget parameter. The output layer of the policy network generates a global optimization policy containing power distribution coefficients and frequency adjustment instructions, which is encrypted and transmitted to the edge node through the secure multi-party computing protocol.

[0053] The generative adversarial network deployed on the edge side includes a policy generator and a discriminator module. After receiving the encrypted cloud policy, the generator extracts key features using knowledge distillation technology and reconstructs them into a control instruction set suitable for edge computing resources through a lightweight convolutional neural network. The discriminator module compares the difference in the action value distribution between the reconstructed instruction and the original policy, and generates an adversarial loss function to guide the optimization of the generator. The generated lightweight instructions include the charge and discharge thresholds of the supercapacitor and the flywheel speed control parameters, which are converted into PWM modulation signals by the edge controller to drive the power module.

[0054] The instruction execution result data is collected by the edge computing node, including the bus voltage fluctuation spectrum, the energy storage unit efficiency parameter, and the device temperature gradient. After feature extraction, the collected data is used to construct a feedback vector, which is transmitted to the storage layer of the dynamic meta-knowledge base through the secure tunnel protocol. The feedback vector triggers the version control mechanism of the meta-knowledge base, and filters out the historical parameter set with the highest correlation with the current working condition for comparison and analysis.

[0055] During the re-calibration of the boundary conditions of the neural differential equation, the Bayesian optimization algorithm is used to calculate the deviation between the feedback vector and the current boundary conditions, and the viscosity coefficient tensor of the Navier-Stokes equation of the flow battery and the flywheel rotor stress relaxation parameters are updated by the gradient descent method. The calibrated differential equation parameters are injected into the online solver of the multi-physics field coupling model to generate updated cross-time scale interaction features, forming a real-time state mirror of the digital twin.

[0056] During the closed-loop optimization process, the updated parameters of the dynamic meta-knowledge base are synchronized to the value function estimator of the cloud policy network to adjust the balance coefficient between policy exploration and exploitation. At the same time, the calibrated boundary conditions are injected into the Hamiltonian construction module of the quantum optimization algorithm as physical constraints to constrain the search space of the converter modulation strategy, forming a cross-layer optimization architecture for cloud-edge collaboration. This architecture improves the two-way mapping accuracy between the digital twin and the physical system through the continuous evolution of the dynamic meta-knowledge base.

[0057] Specifically, for the energy-saving optimization method of the electrical energy storage device described in the present invention, the federated reinforcement learning framework includes: Based on the physical constraints of the dynamic meta-knowledge base, a hierarchical reward mechanism is designed to map the dynamic efficiency of the energy storage system to the depth of the potential energy well and construct a long-term energy efficiency optimization goal. A dynamic energy efficiency entropy index is generated by a multi-entropy maximization controller, and this index is used to constrain the charge relaxation threshold of the supercapacitor power distribution interval in the generated cooperative control strategy.

[0058] When constructing the hierarchical reward mechanism, a potential energy well depth calculation module is designed based on the phase change kinetics parameters of lithium-ion batteries and the flow resistance characteristics of flow batteries in the dynamic meta-knowledge base. The instantaneous efficiency index of the energy storage system is mapped to the gradient change rate of the potential energy function surface, where the local extreme point of the potential energy well corresponds to the optimal working interval. The energy conservation equation of the long-term energy efficiency optimization goal is generated by the Hamiltonian Monte Carlo sampling method, where the kinetic energy term represents the power regulation rate and the potential energy term reflects the cumulative loss of the state of charge deviating from the optimal interval. The potential energy well parameters are dynamically updated through an online learning mechanism and synchronized with the global model parameters in the federated learning framework.

[0059] During the construction of the multi-entropy maximization controller, a differentiable logic programming is used to define the dynamic energy efficiency entropy index. The input layer receives the double-layer voltage distribution data of the supercapacitor and the stress tensor of the flywheel rotor, and extracts the joint time-frequency domain features of the charge relaxation process through a feature cross network. A causal graph model is constructed to analyze the intervention effect of the power distribution action and the charge relaxation time, and a causal intervention reward term is generated to eliminate the spurious correlation between the environmental state and the action selection. The entropy value calculation module combines the Shannon entropy and the Rayleigh entropy to construct a hybrid metric index to constrain the charge migration path of the power distribution strategy in the transient response stage.

[0060] The dynamic energy efficiency entropy index acts on the cooperative control strategy through a policy distillation mechanism, and an entropy constraint module is embedded in the supercapacitor terminal voltage control loop. This module monitors the charge relaxation time constant of the power distribution interval in real time, and when detecting the trend of double-layer oversaturation, it triggers a sliding mode parameter adaptive adjustment mechanism. During the adjustment process, the modulation frequency parameter output by the quantum optimization algorithm is combined to dynamically reconstruct the boundary conditions of the sliding mode control surface, so that the charge relaxation process is always within the safe threshold interval of the entropy value constraint.

[0061] The collaboration between the hierarchical reward mechanism and the multi-entropy controller is achieved through the federated learning architecture. The cloud policy network receives the potential well parameters and entropy value distribution data from multiple nodes, and uses the attention mechanism to generate the weight coefficient matrix. After receiving the global weight coefficients, the edge controller converts the entropy constraint conditions into the feasible domain boundary of the local control instructions through the policy projection algorithm. The characteristic data of the charge relaxation process collected during the instruction execution process is uploaded to the dynamic meta-knowledge base through the differential privacy mechanism, triggering the online calibration of the curvature parameters of the potential energy function surface.

[0062] During the calibration of the potential energy function, the generative adversarial network is used to synthesize the charge migration path data under critical working conditions, and the contrast learning mechanism is used to enhance the representation ability of the potential well for abnormal working conditions. The calibrated potential energy parameters are fed back to the multi-entropy controller to adjust the intervention effect strength coefficient in the causal graph model, forming a dynamic collaborative optimization loop of the reward mechanism and the control strategy. This loop continuously iterates to improve the adaptability of the supercapacitor power distribution strategy to the charge relaxation process, achieving the energy efficiency optimal control of the energy storage system's transient response.

[0063] Specifically, the energy-saving optimization method of the electric energy storage device described in the present invention further includes: Embedding an adversarial curriculum generator within the federated learning framework to synthesize training scenarios based on the critical working condition data stream; Performing differential privacy processing on the multi-node pump operation data by the federated learning framework to generate differentially private protected data; Evaluating the distribution difference degree between the synthesized scenarios and the differentially private protected data through the graph attention network, and dynamically adjusting the training curriculum difficulty coefficient of the multi-agent reinforcement learning framework.

[0064] When constructing the adversarial curriculum generator within the federated learning framework, the Wasserstein generative adversarial network architecture is used to design the scenario synthesis module. The generator receives the feature vectors from the critical working condition data stream, including the time series of the grid frequency mutation event and the spectral characteristics of the load step response, and generates the synthetic scenario data with high information entropy through the deep convolutional network. The discriminator adopts the graph attention network structure, the node attributes include the voltage volatility, temperature gradient, and vibration spectral parameters of the real working conditions, and the edge weights represent the Markov transition probability between the working condition characteristics. The Wasserstein distance between the synthetic scenarios and the differentially private protected data is calculated through the multi-head attention mechanism.

[0065] The graph attention network sets up a hierarchical attention mechanism. The first - layer attention heads analyze the statistical distribution differences of node attributes, and the second - layer attention heads capture the topological structure differences of edge weights. The difference degree calculation module performs weighted fusion on the outputs of each attention head to generate a scalar index representing the realism of the synthetic scene. This index is input into the course difficulty regulator, and through a fuzzy logic controller, it dynamically adjusts the mixing ratio of the synthetic scene and real data in the training course, where the difficulty coefficient is negatively correlated with the exploration rate parameter of the agent's policy network.

[0066] When the dynamically adjusted training course data is input into the multi - agent reinforcement learning framework, a course progressive training strategy is adopted. In the initial stage, strongly perturbed scene data is injected with a high - difficulty coefficient to activate the exploration mechanism of the policy network; as the training process progresses, the proportion of the synthetic scene is gradually reduced and the weight of real data is increased. The value function estimator of the policy network receives the difficulty coefficient as prior knowledge and adjusts the uncertainty range of value estimation through a Bayesian neural network to balance the optimization directions of exploration and exploitation.

[0067] The cooperation between the adversarial course generator and the federated learning framework is achieved through a dual - feedback mechanism. On the one hand, the performance data of the multi - agent policy network in the synthetic scene is fed back to the reward signal generation module of the generator, driving the generator to optimize the criticality features of the scene; on the other hand, the distribution difference degree data evaluated by the graph attention network is injected into the model update rule of the federated aggregation server to dynamically adjust the model contribution weights of each edge node. This mechanism enables the evolution direction of the synthetic scene to be dynamically aligned with the real - world condition distribution, while promoting the robustness improvement of the global policy.

[0068] The dynamic adjustment process of the course difficulty coefficient forms a closed - loop linkage with the parameter calibration of the digital twin. When it is detected that the real - time working condition data fed back by the edge controller exceeds the current course coverage range, a step - type adjustment instruction for the course difficulty coefficient is triggered. This instruction is synchronously sent to the re - initialization module of the cloud - based policy network, enhancing the adaptability of the policy network to the new working condition mode through a partial parameter reset mechanism, and at the same time updating the boundary condition constraint set of the neural differential equation to maintain the mirror accuracy between the digital twin and the physical system.

[0069] Specifically, the energy - saving optimization method of the electric energy storage device described in the present invention further includes: Construct an unsteady Pareto - front search engine, and dynamically allocate multi - objective optimization weights according to the dynamic energy - efficiency entropy index and harmonic spectrum distribution characteristics; Input the weight sequence into a quantum - inspired optimization algorithm, and update the Hamiltonian constraint conditions of the quantum annealing model to achieve the real - time alignment of the converter modulation strategy and the dynamic energy - efficiency optimization goal.

[0070] When constructing a non-steady-state Pareto front search engine, an improved non-dominated sorting genetic algorithm framework is adopted. The input layer receives the time series data of the dynamic energy efficiency entropy index and the frequency domain feature vector of the harmonic spectrum distribution. The long short-term memory network is used to predict the evolution trend of the multi-objective weight vector, and the prediction result is encoded as the crossover probability parameter of the gene sequence. An evolutionary strategy guided by the hypervolume index is set, and the Monte Carlo tree search algorithm is used to select a mutation operator in the strategy space that can simultaneously expand the hypervolume of the objective function and reduce the conflict degree, generating the dynamic topological structure of the non-dominated solution set.

[0071] During the dynamic weight allocation process, a feature attention mechanism is constructed to analyze the correlation strength between the energy efficiency entropy index and the harmonic features, and the contribution degree scores of each objective function are calculated through a differentiable decision tree. The scoring results are input into the fuzzy logic controller to generate an adaptive adjustment coefficient matrix of the weight vector, where the weight value in the harmonic sensitive frequency band forms a non-linear mapping relationship with the transverse magnetic field strength of the quantum annealing model. The weight sequence is synchronized with the converter switching period through the spatio-temporal alignment module to generate an optimized instruction stream with timestamp matching.

[0072] When the weight sequence is input into the quantum-inspired optimization algorithm, the spin encoding technology is used to map the weight value to the coupling strength parameter of the quantum bit. The DC bus ripple suppression target is encoded as the local magnetic field bias, the harmonic distortion rate optimization target is transformed into the antiferromagnetic coupling relationship between adjacent quantum bits, and the switching loss constraint is embodied as the system ground state energy threshold. The updated Hamiltonian constraint conditions are solved by the heat bath algorithm of the quantum annealer, and an optimized sequence of carrier phase shift angles that satisfies multi-objective trade-off is output.

[0073] The real-time alignment process of the converter modulation strategy adopts a sliding window mechanism to couple the angle sequence output by the quantum optimization and the power distribution instruction generated by the multi-agent reinforcement learning framework in space and time. The transmission delay of the control instruction is compensated by the dynamic time warping algorithm to generate a phase-synchronized pulse width modulation waveform. The current harmonic data collected during the execution of the modulation strategy is fed back to the objective function update module of the Pareto front search engine after fast Fourier transform to form a closed-loop optimization link.

[0074] In the closed-loop feedback mechanism, the real-time change data of the harmonic spectrum characteristics triggers the topological structure reorganization of the Pareto front, and the priority rules of non-dominated sorting are dynamically adjusted through the gradient boosting tree algorithm. The reorganized front solution set is distributed to each edge node through the federated learning framework to update the Hamiltonian constraint parameters of the local quantum optimization model. At the same time, the optimization deviation is injected into the dynamic meta-knowledge base to trigger the recalibration operation of the neural differential equation, forming a collaborative enhancement effect across optimization levels.

[0075] Specifically, for the energy-saving optimization method of the electric energy storage device described in the present invention, the closed-loop energy efficiency enhancement loop includes: Inject the synthesized critical operating condition data stream reversely into the dynamic meta-knowledge base, triggering the reparameterization of the hydrodynamic equation of the flow battery in the neural differential equation; Constrain the frequency response interval of the flywheel energy storage magnetic bearing damping compensation amount in the generated cooperative control strategy through the dynamic energy efficiency entropy index of the hierarchical reward mechanism; Synchronously feedback the updated dynamic meta-knowledge base parameters to the cyber-physical neural network and the quantum optimization algorithm, forming a closed-loop enhanced loop for cross-model parameter compensation, control strategy generation, and modulation optimization.

[0076] During the reverse injection of the critical operating condition data stream, use the generative adversarial network to enhance the feature diversity of the synthesized data, and encode the time-frequency domain features of scenarios such as sudden drops in grid frequency and load mutations into high-dimensional tensors. The data stream is transmitted to the incremental learning module of the dynamic meta-knowledge base through a distributed message queue, triggering the reparameterization operation of the Navier-Stokes equation of the flow battery. The reparameterization uses the Bayesian optimization algorithm to calculate the correction amounts of the fluid viscosity coefficient and the turbulence model, and updates the vorticity dissipation term parameters in the differential equation through an online learning mechanism to generate a hydrodynamic feature matrix adapted to extreme operating conditions.

[0077] The calculation module of the dynamic energy efficiency entropy index in the hierarchical reward mechanism receives the flywheel rotor vibration spectrum data, and uses wavelet packet decomposition to extract the energy distribution characteristics of the resonance frequency band. The entropy index is decomposed into a fundamental frequency constraint term and a harmonic suppression term through a dynamic weight allocation algorithm, where the fundamental frequency constraint term is mapped to the upper limit of the frequency response of the magnetic bearing current compensation amount, and the harmonic suppression term corresponds to the phase delay threshold of the damping compensation amount. Construct a feature attention mechanism to analyze the correlation pattern between the compensation amount threshold and the rotor stress tensor, and generate the dynamic boundary conditions of the frequency response interval.

[0078] When the frequency response boundary conditions are input into the cooperative control strategy generation module, a hybrid architecture of sliding mode control and adaptive fuzzy logic is adopted. The sliding mode surface parameters are dynamically adjusted according to the real-time frequency response error, and the fuzzy rule base receives the modulation frequency parameters output by the quantum optimization algorithm to generate a smooth transition instruction for the compensation amount. The flywheel speed fluctuation data collected during the execution of the compensation instruction is used to extract the frequency domain features through fast Fourier transform and fed back to the hierarchical reward mechanism to update the entropy index calculation weight.

[0079] The updated dynamic meta-knowledge base parameters are synchronized to the cyber-physical neural network and the quantum optimization algorithm through a version control protocol. The cyber-physical neural network adopts a transfer learning strategy to align the cross-domain features of the reparameterized hydrodynamic feature matrix and the lithium-ion battery phase change model, and updates the electrochemistry-fluid coupling compensation parameters. After receiving the synchronized parameters, the quantum optimization algorithm reconstructs the spin coupling coefficient in the Hamiltonian constraint condition and adjusts the ground state search path through an annealing temperature scheduling algorithm.

[0080] During the operation of the closed-loop enhancement loop, the edge computing node collects the converter modulation waveform and energy storage unit efficiency data in real time, and generates a feedback vector after feature extraction. The feedback vector is uploaded to the cloud optimization engine through the aggregation channel of the federated learning framework, triggering the version iteration of the dynamic meta-knowledge base and the update of the policy network parameters. The updated control strategy is sent to the edge controller after lightweight compression, forming a full-link optimization system for cross-physical modeling, collaborative control, and quantum modulation, and realizing the continuous improvement of the energy efficiency ratio and device life of the energy storage system.

[0081] The specific implementation manner of the present invention relates to the multi-level optimal control of an electric energy storage device. Aiming at the energy efficiency improvement requirements of lithium-ion batteries, flow batteries, supercapacitors, and flywheel energy storage systems under complex working conditions, dynamic optimization is realized through the deep integration of physical models and data-driven technologies. In the multi-physical field coupling modeling stage of the energy storage device, the neural differential equation is used to implicitly discretize and solve the solid-liquid phase change kinetic equation of the lithium-ion battery, and the adaptive step size algorithm is combined to capture the evolution characteristics of the interface of the electrode active material; the Navier-Stokes fluid equation of the flow battery is spatially discretized by the spectral element method, and the Reynolds stress model is used to predict the turbulent vortex dissipation effect; the rigid-flexible coupling equation of the flywheel rotor is constructed based on the Hamilton variational principle, and the symplectic algorithm is used to maintain the mechanical energy conservation characteristics. The three types of equations are coupled through a multi-task learning framework with a shared hidden layer, and the attention mechanism is used to dynamically allocate the coupling boundary weights, generating a dynamic meta-knowledge base containing the second-level transient response and the hourly cycle decay correlation characteristics.

[0082] During the dynamic parameter compensation process, the activation overpotential term of the Butler–Volmer equation is embedded in the input gate unit of the long short-term memory network on the lithium-ion battery side. The electrode phase change characteristics and the electrolyte diffusion effect are separated through the residual connection structure, and the cross-sensitivity between the state of charge and the state of health is eliminated by covariance matrix decomposition; on the flow battery side, a channel topology model is constructed based on the graph neural network, the node attributes integrate the pressure sensor data and the infrared temperature field distribution, and the operation data of multiple pump stations are aggregated through the federated learning framework to generate an electrolyte flow rate compensation matrix. The compensation residuals extract the frequency domain characteristics through the sliding window mechanism, and a multi-dimensional state observation tensor is constructed and input into the multi-agent reinforcement learning framework.

[0083] During the collaborative control strategy generation stage, the hierarchical attention mechanism is used to analyze the electric double layer relaxation process for the supercapacitor terminal voltage control. The spatial attention layer identifies the voltage distribution differences in the battery stack through the convolution kernel, and the temporal attention layer uses the gated recurrent unit to capture the transient response mode of charge migration; on the flywheel energy storage side, the rotor deformation surface is reconstructed based on the neural radiance field, and the multi-phase winding current compensation amount of the magnetic bearing is generated through the proximal policy optimization algorithm. The compensation amount threshold is used as the torque fluctuation constraint of the quantum optimization algorithm. The policy fusion layer aligns the power distribution parameters and the damping compensation amount in space and time to generate a joint optimization instruction set.

[0084] In the quantum-inspired optimization stage, the DC bus ripple suppression target is encoded as the local field strength parameter of the spin system, and the harmonic distortion rate optimization is mapped to the antiferromagnetic coupling relationship between qubits. The time-frequency joint features of the current waveform are extracted through a deep residual network to predict the distribution of harmonic-sensitive frequency bands. Based on the updated Hamiltonian constraint conditions, the quantum annealer outputs a carrier phase-shift angle sequence, and combines with the dynamic time warping algorithm to compensate for the control delay to generate a phase-synchronized pulse-width modulation waveform.

[0085] In the federated reinforcement learning framework, the cloud policy network fuses the pump pressure data of multiple nodes through the differential privacy mechanism to generate a global optimization strategy; on the edge side, a generative adversarial network is used for policy distillation to simplify the complex control logic into lightweight PWM instructions. The instruction execution data triggers the recalibration of the boundary conditions of the neural differential equation through the backpropagation channel to update the viscosity coefficient and interfacial energy parameters of the dynamic meta-knowledge base. The closed-loop optimization loop synchronizes the cloud and edge parameters through the version control protocol to realize the dynamic mirroring of the digital twin and the physical system, and finally achieves the improvement of the energy efficiency of the energy storage system throughout its life cycle.

[0086] The technical solution of the present invention solves the problems of model mismatch and control deviation of the energy storage device through three technical means: constructing a dynamic coupling model across time scales, establishing a data-driven parameter compensation mechanism, and realizing closed-loop optimization control. First, the neural differential equation is used to fuse the phase change dynamics of lithium-ion batteries, the fluid mechanics of flow batteries, and the mechanical mechanics of flywheels. The spatio-temporal correlation features of multi-source heterogeneous data are extracted through the Transformer architecture to generate a dynamic meta-knowledge base containing cross-scale non-linear interaction features. This model realizes the dynamic weight allocation of the coupling boundary of different physical fields through a multi-task learning framework with shared hidden layers, overcoming the modeling deviation of traditional single physical field models in scenarios of second-level transient response and hour-level cyclic decay.

[0087] In the dynamic parameter compensation stage, the Butler–Volmer electrochemical equation is embedded in the gating mechanism of the long short-term memory network. The phase change of the electrode and the diffusion characteristics of the electrolyte are separated through orthogonalization constraints to eliminate the cross-sensitivity between the state of charge and the state of health. Combining with a graph neural network to construct a flow channel topology model of the flow battery, federated learning aggregates the operation data of multiple node pumps to generate a feedforward compensation coefficient matrix. The compensation residuals are spatio-temporally aligned to form a multi-dimensional state observation vector, providing accurate input features for multi-agent reinforcement learning and solving the problem of the lag response of traditional control strategies to non-linear dynamic characteristics.

[0088] At the level of closed-loop optimization control, the quantum-inspired optimization algorithm maps the power distribution threshold of the cooperative control strategy to the Hamiltonian constraint of the quantum annealing model, and dynamically updates the model boundary conditions through the harmonic spectrum features extracted by the deep residual network. The federated reinforcement learning framework uses the differential privacy mechanism to fuse multi-node real-time data, and realizes the lightweight migration of the cloud strategy to the edge controller through the imitation learning generative adversarial network. The execution results are injected back into the dynamic meta-knowledge base to trigger the recalibration of the parameters of the neural differential equation. This closed-loop architecture continuously corrects the dynamic mismatch of the multi-physical field coupling model through the bidirectional feedback mechanism of the physical model and data-driven, and realizes the adaptive optimization of the control strategy for nonlinear characteristics.

Claims

1. An energy-saving optimization method for an electric energy storage device, characterized in that, Including: Construct a multi - physical - field coupling model of the energy storage device. By fusing the solid - liquid phase change kinetics equation of lithium - ion batteries, the fluid mechanics equation of flow batteries, and the rigid - flexible coupling equation of flywheel rotors through neural differential equations, generate a dynamic meta - knowledge base containing cross - time - scale non - linear interaction characteristics; Based on the physical constraint conditions of the dynamic meta - knowledge base, use a physics - informed neural network to perform dynamic parameter compensation processing on the polarization voltage hysteresis effect and the electrolyte flow resistance coupling effect of the energy storage device, and output a residual feature vector containing the decoupled state - of - charge characteristics; Input the residual feature vector into a multi - agent reinforcement learning framework to generate the hierarchical control parameters of the supercapacitor terminal voltage and the magnetic bearing damping compensation amount of the flywheel energy storage, forming a collaborative control strategy; According to the power distribution threshold of the collaborative control strategy, use a quantum - inspired optimization algorithm to transform the DC bus ripple suppression and harmonic distortion rate optimization objectives into the Hamiltonian constraint conditions of the quantum annealing model, and generate the converter modulation instruction; In the federated reinforcement learning framework, distill the cloud - side global optimization strategy to the edge controller through an imitation - learning generative adversarial network, and at the same time inject the real - time operating condition data fed back by the edge controller into the multi - physical - field coupling model for parameter recalibration, forming a closed - loop energy - efficiency enhancement loop that crosses physical models, data - driven, and quantum optimization.

2. The energy-saving optimization method for the electrical energy storage device according to claim 1, wherein The construction of the multi - physical - field coupling model of the energy storage device includes: Use neural differential equations to fuse the solid - liquid phase change reaction kinetics equation of lithium - ion batteries, the fluid mechanics equation of flow batteries, and the rigid - flexible coupling equation of flywheel rotors; Extract the spatio - temporal correlation characteristics of multi - source heterogeneous data of the energy storage device through a Transformer architecture with spatio - temporal attention mechanisms, and generate a meta - knowledge base containing cross - scale non - linear interaction relationships, which is used to provide physical constraint conditions for the physics - informed neural network.

3. The energy-saving optimization method for the electrical energy storage device according to claim 1, characterized in that The dynamic parameter compensation includes: On the lithium - ion battery side of the energy storage device, embed the Butler–Volmer electrochemical equation into a long - short - term memory network to decouple the cross - sensitivity of the state - of - charge and the state - of - health in real time; On the flow - battery side, construct a graph neural network to predict the flow - channel pressure drop distribution, combine federated learning to aggregate the operation data of multi - node pump machines, dynamically optimize the feed - forward compensation coefficient matrix of the electrolyte flow rate, and use the compensation residual as the state observation input of the multi - agent reinforcement learning.

4. The energy-saving optimization method of the electrical energy storage device according to claim 1, characterized in that, Inputting the residual feature vector into the multi - agent reinforcement learning framework to generate the hierarchical control parameters of the supercapacitor terminal voltage and the magnetic bearing damping compensation amount of the flywheel energy storage, forming a collaborative control strategy includes: Adopt a hierarchical attention mechanism for the supercapacitor terminal voltage control, and match the charge relaxation time constant through a sliding - mode parameter dynamic adjustment strategy; Deploy a rotor deformation observer based on neural radiance fields on the flywheel energy - storage side, and combine the proximal policy optimization algorithm to generate the magnetic bearing damping compensation amount, so that the quantum - inspired optimization algorithm receives the torque fluctuation suppression threshold as the modulation target constraint.

5. The energy-saving optimization method of the electrical energy storage device according to claim 1, characterized in that The multi - objective optimization according to the power distribution threshold of the collaborative control strategy includes: Transform the DC bus ripple suppression, harmonic distortion rate optimization, and switching loss minimization into the Hamiltonian of the quantum annealing model; Predict the harmonic spectrum distribution characteristics through a deep residual network, and feedback the distribution characteristics to the multi-physics field coupling model to update the boundary conditions of the neural differential equation.

6. The energy-saving optimization method for the electrical energy storage device according to claim 1, characterized in that The formation of a closed-loop energy efficiency enhancement loop integrating cross-physical models, data-driven methods, and quantum optimization includes: Based on the output of the multi-agent reinforcement learning framework, deploy the multi-agent proximal policy optimization algorithm in the cloud, and fuse the operation data of multiple-node pump units aggregated by the federated learning through a differential privacy protection mechanism; On the edge side, use an imitation learning generative adversarial network to convert the cloud policy into lightweight control instructions, and inject the instruction execution results back into the generated dynamic meta-knowledge base to trigger the recalibration of the boundary condition parameters of the neural differential equation.

7. The energy-saving optimization method for the electrical energy storage device according to claim 1, characterized in that The federated reinforcement learning framework includes: Based on the physical constraints of the dynamic meta-knowledge base, design a hierarchical reward mechanism, map the dynamic efficiency of the energy storage system to the depth of the potential energy well, and construct a long-term energy efficiency optimization goal; Generate a dynamic energy efficiency entropy index through a multi-entropy maximization controller, and the index is used to constrain the charge relaxation threshold of the supercapacitor power distribution interval in the generated cooperative control strategy.

8. The energy-saving optimization method for the electrical energy storage device according to claim 3, characterized in that It also includes: Embed an adversarial curriculum generator within the federated learning framework to synthesize training scenarios based on the critical operating condition data stream; The federated learning framework performs differential privacy processing on the operation data of multiple-node pump units to generate differentially private protected data; Evaluate the distribution difference degree between the synthesized scenario and the differentially private protected data through a graph attention network, and dynamically adjust the training curriculum difficulty coefficient of the multi-agent reinforcement learning framework.

9. The energy-saving optimization method for the electrical energy storage device according to claim 5, characterized in that It also includes: Construct an unsteady Pareto front search engine, and dynamically allocate multi-objective optimization weights according to the dynamic energy efficiency entropy index and the harmonic spectrum distribution characteristics; Input the weight sequence into a quantum-inspired optimization algorithm to update the Hamiltonian constraint conditions of the quantum annealing model to achieve real-time alignment of the converter modulation strategy and the dynamic energy efficiency optimization goal.

10. The energy-saving optimization method for the electrical energy storage device according to claim 1, characterized in that, The closed-loop energy efficiency enhancement loop includes: Inject the synthesized critical operating condition data stream back into the dynamic meta-knowledge base to trigger the reparameterization of the hydrodynamic equation of the flow battery in the neural differential equation; Constrain the frequency response interval of the flywheel energy storage magnetic bearing damping compensation amount in the generated cooperative control strategy through the hierarchical reward mechanism dynamic energy efficiency entropy index; Synchronously feedback the updated dynamic meta-knowledge base parameters to the physical information neural network and the quantum optimization algorithm to form a closed-loop enhancement loop for cross-model parameter compensation, control strategy generation, and modulation optimization.

Images