All-vanadium redox flow battery multi-physics field multi-scale agent modeling method

Abstract

The invention belongs to the technical field of energy storage of all-vanadium redox flow batteries, particularly relates to a multi-physics-field multi-scale agent modeling method of an all-vanadium redox flow battery, and aims to solve the problem that multi-physics-field coupling in the battery cannot be efficiently processed. The method comprises the following steps: constructing a structured agent model, and forming a topological structure for integrally describing a chemical field, a fluid mechanical field and a thermodynamic field of the all-vanadium redox flow battery by connecting a plurality of linear dynamic operator blocks and a plurality of static nonlinear blocks in a differentiatable manner. And in the structured proxy model, establishing an inter-field coupling relationship among the chemical field, the hydrodynamic field and the thermodynamic field by taking the output variable of the first physical field as the input variable of the second physical field. And different dynamic response characteristics are configured for simulating linear dynamic operator blocks of different dynamic speed physical processes. And training the structured agent model by using the historical data of the operation of the all-vanadium redox flow battery, and reconstructing the internal state of the all-vanadium redox flow battery in real time by using the trained model.

Description

Technical Field

[0001] This invention relates to the field of vanadium redox flow battery energy storage technology, and in particular to a multi-physics, multi-scale proxy modeling method for vanadium redox flow batteries. Background Technology

[0002] Vanadium redox flow batteries, as an electrochemical energy storage technology with large-scale energy storage potential, have their operating state determined by the complex coupling of multiple physical fields, including electrochemistry, fluid dynamics, and thermodynamics. The battery interior simultaneously exhibits dynamic characteristics across multiple timescales, ranging from rapid electrochemical reactions on the millisecond scale to slow thermal diffusion processes on the hourly scale. The electrolyte flow rate directly affects ion transport and reaction rates at the electrode surface, while Joule heating and reaction heat generated by the electrochemical process drive the evolution of the temperature field. Temperature, in turn, inversely regulates reaction kinetics and electrolyte properties, forming a complex, highly nonlinear, closed-loop interactive system. Therefore, accurately reconstructing the battery's internal state is crucial for achieving optimized system control and lifetime prediction.

[0003] Currently, modeling methods for vanadium redox flow batteries can be mainly divided into two categories: partial differential equation models based on physical mechanisms and purely data-driven black-box models. While high-fidelity mechanistic models can precisely describe the physical processes of multi-field coupling, their computational cost is high and solution time is long, making it difficult to meet the millisecond-level real-time monitoring and control requirements of battery management systems. Conversely, traditional equivalent circuit models or data-driven models such as neural networks, although computationally efficient, generally lack physical constraints, are prone to accumulating errors in long-term predictions, have poor extrapolation properties, and their model structures cannot naturally express the bidirectional coupling mechanism and wide-timescale dynamics between multiple physical fields.

[0004] Furthermore, some block-structured models, such as Wiener and Hammerstein, have limited interpretability and are difficult to flexibly represent coupled systems with multiple inputs and outputs. Meanwhile, emerging methods such as physically-informed neural networks still face challenges such as high training overhead, uncertain convergence, and difficulties in engineering deployment. In summary, existing technologies still struggle to achieve a good balance between physical consistency of models, multi-field coupling expressiveness, unified description across multiple time scales, and computational efficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-physics, multi-scale proxy modeling method for vanadium redox flow batteries, aiming to solve the problem that existing technologies cannot efficiently handle the coupling of multiple physics fields and dynamics at multiple time scales within the battery.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a multi-physics, multi-scale surrogate modeling method for vanadium redox flow batteries, comprising: S1: constructing a structured surrogate model, which interconnects multiple linear dynamic operator blocks and multiple static nonlinear blocks in a differentiable manner to form a topology for integrating and describing the chemical, hydrodynamic, and thermodynamic fields of the vanadium redox flow battery; S2: within the structured surrogate model, establishing coupling connections between the linear dynamic operators and the static nonlinear blocks to correlate the outputs / inputs of the chemical, hydrodynamic, and thermodynamic fields, thereby explicitly realizing multi-physics coupling within the structured surrogate model; S3: configuring different dynamic response characteristics for the linear dynamic operator blocks to simulate physical processes at different dynamic velocities, in order to handle the multi-timescale dynamic problems caused by multi-physics coupling; S4: training the structured surrogate model using historical data from the operation of the vanadium redox flow battery, and using the trained model to reconstruct the internal state of the vanadium redox flow battery in real time.

[0007] In step S1, the input-output relationship of the linear dynamic operator block is defined by a rational transfer function, which is used to perform an infinite impulse response filtering operation to capture the dynamic process of the physical field; the static nonlinear block includes a multi-layer feedforward neural network, which is used to perform a memoryless nonlinear transformation on the input vector to describe the nonlinear characteristics of the physical field.

[0008] In step S1, the structured proxy model includes: a first linear dynamic operator block and a first static nonlinear block for describing the chemical field; a second linear dynamic operator block for describing the fluid dynamic field; and a second static nonlinear block and a third linear dynamic operator block for describing the thermodynamic field.

[0009] The inter-field coupling relationship in step S2 includes: the electrolyte flow rate output by the second linear dynamic operator block as the input of the first linear dynamic operator block; the battery terminal voltage output by the first static nonlinear block as the input of the second static nonlinear block; and the battery temperature output by the third linear dynamic operator block as the input of the first static nonlinear block.

[0010] In step S3, the order of the numerator and denominator polynomials of the linear dynamic operator block simulating a fast-changing physical process is lower than that of the linear dynamic operator block simulating a slow-changing physical process.

[0011] The training process in step S4 includes: obtaining input sequences and corresponding output sequences from historical data to construct a training dataset, wherein the input sequences include battery current and pump speed commands, and the output sequences include measured battery terminal voltage and battery temperature; inputting the input sequences into a structured proxy model to obtain a prediction sequence; and optimizing the parameters of all linear dynamic operator blocks and static nonlinear blocks in the structured proxy model through a backpropagation algorithm to minimize the error between the prediction sequence and the output sequence.

[0012] The internal states reconstructed in real time in step S4 include: estimated values ​​of measurable states and reconstructed values ​​of unmeasurable internal states; wherein, the estimated values ​​of measurable states include terminal voltage and temperature; and the reconstructed values ​​of unmeasurable internal states include electrode surface ion concentration, overpotential, heat generation rate and electrolyte flow rate.

[0013] The input to the first linear dynamic operator block is the flow rate, and the output is the ion concentration on the electrode surface; the input to the first static nonlinear block includes the ion concentration on the electrode surface, temperature, and current, and the output is the battery terminal voltage.

[0014] The input to the second linear dynamic operator block is the pump speed command, and the output is the electrolyte flow rate.

[0015] The second static nonlinear block takes battery current and terminal voltage as inputs and outputs the total heat generation rate; the third linear dynamic operator block takes the total heat generation rate as input and outputs the battery temperature.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The multi-physics, multi-scale proxy modeling method for vanadium redox flow batteries provided in this application adopts a structured architecture of differentially interconnected linear dynamic operator blocks and static nonlinear blocks. Each physical field is configured with a dedicated functional block (the chemical field corresponds to the first linear dynamic operator block and the first static nonlinear block, the fluid field corresponds to the second linear dynamic operator block, and the thermal field corresponds to the second static nonlinear block and the third linear dynamic operator block). The function of each block directly corresponds to the physical process (e.g., the linear dynamic operator block captures dynamic processes such as ion diffusion and fluid flow, and the static nonlinear block describes nonlinear characteristics such as electrochemical reaction kinetics and heat generation rate), ensuring that the physical meaning of the model is clear. At the same time, through explicit inter-field coupling pathways (the electrolyte flow rate of the fluid field is input into the chemical field, the current / voltage of the chemical field is input into the thermal field, and the temperature of the thermal field is fed back to the chemical field), the closed-loop interaction of fluid-chemistry-heat within the battery is accurately reproduced. This avoids the physical blindness of pure black box models and overcomes the problem of poor coupling flexibility of traditional block structure models, achieving a deep integration of physical interpretability and data-driven adaptability.

[0017] 2. The method provided in this application uses differentiated dynamic characteristic configuration: low-order operators are used to achieve fast response in fast-changing processes, high-order operators are used to capture long-term memory in slow-changing processes, and intermediate-order transitions are used in medium-dynamic processes. Without splitting the model, the dynamics of the entire time scale can be adaptively captured through the same surrogate model. It can track the ion concentration fluctuations of the electrochemical reaction in milliseconds in real time, and accurately simulate the temperature change trend of thermal diffusion in hours. It solves the compatibility problem of traditional segmented modeling, realizes a unified dynamic description of the entire operating condition time range, and provides unified modeling support for battery control from instantaneous response, such as current change adjustment, to long-term operation monitoring, such as long-term charge and discharge temperature control.

[0018] 3. The method provided in this application offers dual safeguards through physical constraints and end-to-end training. On one hand, the structured architecture and explicit coupling pathway provide physical prior constraints for the model, preventing data noise from causing the model to deviate from physical laws. On the other hand, based on the historical operating data of the vanadium redox flow battery (input sequence includes battery current and pump speed commands, output sequence includes battery terminal voltage and temperature), all parameters (linear dynamic operator block coefficients, static nonlinear block neural network weights / biases) are optimized through the backpropagation algorithm, forcing the model to fit the battery's true dynamic response and minimizing the error between predicted and measured values. This design enables the model to maintain high accuracy under known operating conditions and maintain stability based on physical constraints under unknown operating conditions, avoiding the accumulation of errors in long-term predictions and meeting the needs of long-term stable operation monitoring of large-scale energy storage systems. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multiphysics multiscale proxy modeling method for vanadium redox flow batteries provided in an embodiment of this application; Figure 2 This is a multi-field coupling field relationship diagram provided in the embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] For example, refer to Figure 1 This application provides a multi-physics, multi-scale proxy modeling method for vanadium redox flow batteries, including: S1: Construct a structured proxy model. The structured proxy model forms a topology for integrating the chemical, hydrodynamic, and thermodynamic fields of an all-vanadium redox flow battery by interconnecting multiple linear dynamic operator blocks and multiple static nonlinear blocks in a differentiable manner.

[0022] For example, in step S1, the input-output relationship of the linear dynamic operator block G-block is defined by a rational transfer function, used to perform an infinite impulse response filtering operation to capture the dynamic processes of the physical field. Its core mathematical expression is: in, It is a time delay operator. and It is a polynomial that defines the dynamics of the system: A rational transfer function is equivalent in the time domain to a linear difference equation with constant coefficients, that is: The input-output relationship of the entire system can be represented as: G-block uses the above recursive equation to process the input sequence. The infinite impulse response filter, whose internal state depends on the current value of the input, the past value of the input, and the past value of the output.

[0023] The static nonlinear block F-block comprises a multi-layer feedforward neural network used to perform a memoryless nonlinear transformation on the input vector to describe the nonlinear properties of the physical field. The input-output mapping relationship of the multi-layer feedforward neural network is expressed as: in, For the input vector, For the output vector, and For neural network parameters, The activation function is, but is not limited to, the hyperbolic tangent function, the ReLU function, or the Sigmoid function.

[0024] The structured proxy model interconnects multiple G-blocks and multiple F-blocks in a differentiable manner to form a topology that includes parallel, cascaded, or feedback connections, and integrates the chemical, hydrodynamic, and thermodynamic fields of the all-vanadium redox flow battery.

[0025] For example, in step S1, the structured proxy model includes: a first linear dynamic operator block and a first static nonlinear block for describing a chemical field; a second linear dynamic operator block for describing a fluid dynamic field; and a second static nonlinear block and a third linear dynamic operator block for describing a thermodynamic field.

[0026] As one possible implementation, the input of the first linear dynamic operator block is the flow rate, and the output is the ion concentration on the electrode surface; the input of the first static nonlinear block includes the ion concentration on the electrode surface, temperature, and current, and the output is the battery terminal voltage.

[0027] As one possible implementation, the input to the second linear dynamic operator block is the pump speed command, and the output is the electrolyte flow rate.

[0028] As one possible implementation, the second static nonlinear block takes battery current and terminal voltage as inputs and outputs the total heat generation rate; the third linear dynamic operator block takes the total heat generation rate as input and outputs the battery temperature.

[0029] More specifically, the chemical field of the all-vanadium redox flow battery describes the changes in vanadium ion concentration and the electrochemical reaction kinetics. The ion convection-diffusion equation in the electrolyte is as follows: in, It is the diffusion coefficient. It's the flow rate. Regarding ion concentration, the dynamic characteristics of the above partial differential equations in the structured surrogate model are represented by a dedicated G-block, namely the first linear dynamic operator block, denoted as... Approximate simulations are then performed. Input Corresponding to the flow rate, Output Trends in ion concentration at specific locations .

[0030] The electrochemical reaction kinetics at the electrode surface are described by the Butler-Volmer equation: in, For current density, For exchange current density, and The mass transfer coefficient is . This is the universal gas constant. It is Faraday Changshu. For temperature, This is an overpotential.

[0031] The complete terminal voltage equation can be expressed as: in, For current, Given the battery's internal resistance, the theoretical potential of the electrodes in equilibrium is determined by the Nernst equation: In the structured surrogate model, this static nonlinear relationship is represented by a dedicated F-block, also known as the first static nonlinear block, denoted as... Approximate simulations are then performed. input vector Including electrode surface ion concentration ,temperature and current , Output Corresponding to battery terminal voltage .

[0032] The hydrodynamic field of an all-vanadium redox flow battery describes the flow dynamics of the electrolyte in the battery circuit. The electrolyte flow is described by a simplified form of the Navier-Stokes equations: in, The electrolyte density, The dynamic viscosity of the electrolyte Volume force term, This refers to fluid pressure, which is related to the pump speed command. In the structured surrogate model, the aforementioned fluid dynamics are represented by a dedicated G-block, also known as the second linear dynamic operator block, denoted as... Approximate simulations are then performed. Input Corresponding to the pump speed command, Output Corresponding to the flow rate of the electrolyte .

[0033] Furthermore, the thermodynamic field of the all-vanadium redox flow battery describes the heat generation and temperature distribution changes of the battery system. The heat generation rate of the system mainly consists of ohmic heat and reaction heat, and can be approximately expressed as: In the structured surrogate model, the aforementioned nonlinear relationship is represented by a dedicated F-block, also known as the second static nonlinear block, denoted as... Approximate simulations are then performed. input vector At least including battery current Terminal voltage , Output Corresponding to the total heat generation rate .

[0034] The temperature change of the battery stack is described by the energy conservation equation: in, For thermal conductivity, Let be the average density of the fuel cell material. In the structured surrogate model, the aforementioned thermal diffusion dynamics are represented by a dedicated G-block, also known as the third linear dynamic operator block, denoted as . Approximate simulations are then performed. Input Corresponding to the rate of heat generation , Output Corresponding to the battery temperature .

[0035] S2: Within the structured proxy model, a coupling connection is established between the linear dynamic operator and the static nonlinear block to correlate the outputs / inputs of the chemical field, fluid dynamics field, and thermodynamic field, so as to explicitly realize multiphysics coupling within the structured proxy model.

[0036] For example, the coupling relationship in step S2 includes: the electrolyte flow rate output by the second linear dynamic operator block as the input of the first linear dynamic operator block; the battery terminal voltage output by the first static nonlinear block as the input of the second static nonlinear block; and the battery temperature output by the third linear dynamic operator block as the input of the first static nonlinear block.

[0037] More specifically, the interaction between the three physical fields is achieved by establishing the following specific data flow connections in the structured proxy model: The The output, i.e., the electrolyte flow rate As an input signal, it is input to The input vector is used to simulate the effect of flow rate on ion transport rate and reactant supply at the electrode surface, thereby coupling fluid dynamics with the electrochemical process.

[0038] Depend on The battery current obtained from the relevant model calculation and battery terminal voltage It is fed into the system as an input signal. The input vector is used to simulate the effect of energy loss from electrochemical processes as the primary heat source on the thermal field.

[0039] The The output, i.e., battery temperature. It is fed back as an input signal to The input vector. This connection simulates the effect of temperature on the electrochemical reaction rate constant. Electrolyte conductivity and ion diffusion coefficient This has a significant impact, thus achieving feedback coupling between the thermodynamic field and the chemical field.

[0040] S3: To simulate physical processes with different dynamic velocities, the linear dynamic operator block is configured with different dynamic response characteristics to handle multi-timescale dynamic problems caused by multi-physics coupling.

[0041] For example, in step S3, the order of the numerator and denominator polynomials of the linear dynamic operator block simulating a fast-changing physical process is lower than that of the linear dynamic operator block simulating a slow-changing physical process. For example, simulating charge transfer dynamics or rapid concentration fluctuations... The order of its numerator and denominator polynomials and Configured to a relatively low value, enabling it to respond quickly to input changes and capture high-frequency dynamics; used for simulating pressure and velocity dynamics in fluid systems. The order of its numerator and denominator and Configured to an intermediate value between chemical reaction and thermal diffusion; used for simulating slowly varying processes, such as thermal diffusion or overall temperature isostatics. The order of its numerator and denominator polynomials and It is configured to a relatively high value; and the above order satisfies the following relationship: .

[0042] S4: Use historical data from the operation of the vanadium redox flow battery to train the structured surrogate model, and use the trained model to reconstruct the internal state of the vanadium redox flow battery in real time.

[0043] For example, the training process in step S4 includes: obtaining input sequences and corresponding output sequences from historical data to construct a training dataset, wherein the input sequences include battery current and pump speed commands, and the output sequences include measured battery terminal voltage and battery temperature; inputting the input sequences into a structured proxy model to obtain a prediction sequence; and optimizing the parameters of all linear dynamic operator blocks and static nonlinear blocks in the structured proxy model through a backpropagation algorithm to minimize the error between the prediction sequence and the output sequence.

[0044] More specifically, obtain lengths from historical data. The sequence, to construct the dataset Among them, the structured proxy model input , Battery current, For pump speed commands, the model outputs... , To measure the battery terminal voltage, The measured battery temperature. Input sequence The input is fed into the structured proxy model, forward computation is performed, and the predicted sequence output by the model is obtained. Then, the mean squared error loss between the predicted and actual measured values ​​is calculated: The aforementioned loss function forces the structured proxy model to learn the battery's true voltage and temperature dynamic response. The loss is calculated using the backpropagation algorithm. For all trainable parameters The gradient. Parameters Including all G-blocks ( coefficient and all F-blocks ( The neural network weights and biases.

[0045] For example, the internal state reconstructed in real time in step S4 includes: the estimated value of the measurable state and the reconstructed value of the unmeasurable internal state; wherein, the estimated value of the measurable state includes the terminal voltage and temperature; and the reconstructed value of the unmeasurable internal state includes the electrode surface ion concentration, overpotential, heat generation rate and electrolyte flow rate.

[0046] After the structured proxy model is trained, at each sampling time... The input vector collected in real time Input is a pre-deployed fixed-parameter model. The fixed-parameter model performs a single-step forward computation, outputting estimates of the measurable states and reconstructed values ​​of the unmeasurable internal states. The estimates of the measurable states include the terminal voltages. and temperature This is used for comparison with sensor data. Reconstructed values ​​of unmeasurable internal states include... Output electrode surface ion concentration and overpotential , Output heat generation rate ,as well as The flow rate of the output electrolyte These variables together constitute the battery's real-time internal state vector. .

[0047] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0048] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-physics, multi-scale proxy modeling method for all-vanadium redox flow batteries, characterized in that, include: S1: Construct a structured proxy model, which interconnects multiple linear dynamic operator blocks and multiple static nonlinear blocks in a differentiable manner to form a topology for integrating the chemical, hydrodynamic, and thermodynamic fields of a vanadium redox flow battery; S2: Within the structured proxy model, establish coupling connections between the linear dynamic operators and the static nonlinear blocks to correlate the outputs / inputs of the chemical, hydrodynamic, and thermodynamic fields, thereby explicitly realizing multi-physics coupling within the structured proxy model; S3: To simulate physical processes at different dynamic velocities, configure different dynamic response characteristics for the linear dynamic operator blocks to handle the multi-timescale dynamic problems caused by the multi-physics coupling; S4: The structured proxy model is trained using historical data from the operation of the vanadium redox flow battery, and the trained model is used to reconstruct the internal state of the vanadium redox flow battery in real time.

2. The multiphysics, multiscale proxy modeling method for all-vanadium redox flow batteries according to claim 1, characterized in that, In step S1, the input-output relationship of the linear dynamic operator block is defined by a rational transfer function, which is used to perform an infinite impulse response filtering operation to capture the dynamic process of the physical field; the static nonlinear block includes a multi-layer feedforward neural network, which is used to perform a memoryless nonlinear transformation on the input vector to describe the nonlinear characteristics of the physical field.

3. The multi-physics, multi-scale proxy modeling method for all-vanadium redox flow batteries according to claim 2, characterized in that, In step S1, the structured proxy model includes: a first linear dynamic operator block and a first static nonlinear block for describing the chemical field; a second linear dynamic operator block for describing the fluid dynamic field; and a second static nonlinear block and a third linear dynamic operator block for describing the thermodynamic field.

4. The multiphysics, multiscale proxy modeling method for all-vanadium redox flow batteries according to claim 3, characterized in that, The coupling relationship in step S2 includes: the electrolyte flow rate output by the second linear dynamic operator block as the input of the first linear dynamic operator block; the battery terminal voltage output by the first static nonlinear block as the input of the second static nonlinear block; and the battery temperature output by the third linear dynamic operator block as the input of the first static nonlinear block.

5. The multiphysics, multiscale proxy modeling method for all-vanadium redox flow batteries according to claim 3, characterized in that, In step S3, the order of the numerator and denominator polynomials of the linear dynamic operator block simulating a fast-changing physical process is lower than that of the linear dynamic operator block simulating a slow-changing physical process.

6. The multiphysics, multiscale proxy modeling method for an all-vanadium redox flow battery according to claim 1, characterized in that, The training process in step S4 includes: obtaining input sequences and corresponding output sequences from historical data to construct a training dataset, wherein the input sequences include battery current and pump speed commands, and the output sequences include measured battery terminal voltage and battery temperature; inputting the input sequences into the structured proxy model to obtain a prediction sequence; and optimizing the parameters of all linear dynamic operator blocks and static nonlinear blocks in the structured proxy model through a backpropagation algorithm to minimize the error between the prediction sequence and the output sequence.

7. The multiphysics, multiscale proxy modeling method for an all-vanadium redox flow battery according to claim 1, characterized in that, The internal states reconstructed in real time in step S4 include: estimated values ​​of measurable states and reconstructed values ​​of unmeasurable internal states; wherein, the estimated values ​​of measurable states include terminal voltage and temperature; and the reconstructed values ​​of unmeasurable internal states include electrode surface ion concentration, overpotential, heat generation rate and electrolyte flow rate.

8. The multiphysics, multiscale proxy modeling method for an all-vanadium redox flow battery according to claim 3, characterized in that, The input to the first linear dynamic operator block is the flow rate, and the output is the ion concentration on the electrode surface; the input to the first static nonlinear block includes the ion concentration on the electrode surface, temperature, and current, and the output is the battery terminal voltage.

9. The multi-physics, multi-scale proxy modeling method for all-vanadium redox flow batteries according to claim 3, characterized in that, The input to the second linear dynamic operator block is the pump speed command, and the output is the electrolyte flow rate.

10. The multiphysics, multiscale proxy modeling method for an all-vanadium redox flow battery according to claim 3, characterized in that, The second static nonlinear block takes battery current and terminal voltage as input and outputs total heat generation rate as output; the third linear dynamic operator block takes the total heat generation rate as input and outputs battery temperature as output.

Images