SOC prediction method for all-vanadium redox flow battery

Abstract

The invention relates to the technical field of battery management systems, in particular to an SOC prediction method for an all-vanadium redox flow battery, and aims to realize high-precision SOC prediction under complex dynamic working conditions. Comprising the following steps: acquiring current and terminal voltage of the all-vanadium redox flow battery, and normalizing to obtain normalized current and voltage sequences; and inputting the normalized current sequence into the physical information long-short-term memory network, and updating the state of the internal memory unit based on the physical law of the battery equivalent circuit model. Based on the state of the memory unit, the contribution weights of the first polarization voltage state, the second polarization voltage state and the charge state to the terminal voltage prediction are adjusted through a multi-head attention mechanism by utilizing the error between the real-time terminal voltage and the terminal voltage prediction value of the memory unit, and a fused feature representation is generated and input to a full connection layer; and obtaining a predicted value of the state of charge through nonlinear mapping, and judging whether to trigger updating of circuit parameters in the physical information long-short-term memory network according to an error between the real-time terminal voltage and the predicted value of the terminal voltage of the memory unit.

Description

Technical Field

[0001] This invention relates to the field of battery management system technology, and in particular to a method for predicting the state of charge (SOC) of an all-vanadium redox flow battery. Background Technology

[0002] Vanadium redox flow batteries, as a large-scale energy storage technology, are playing an increasingly important role in smart grids and renewable energy grid integration due to their advantages such as independently designable power and capacity, long cycle life, and high safety. State of charge (SOC) is a key internal state characterizing the remaining capacity of a battery. Its accurate estimation is crucial for preventing overcharging and over-discharging, optimizing energy management, and extending battery life, and is one of the core functions of a battery management system.

[0003] However, accurately estimating the state of charge (SOC) of vanadium redox flow batteries under actual dynamic operating conditions faces severe challenges. First, methods based on purely physical models, such as those relying on second-order equivalent circuit models, heavily depend on the accuracy of the model parameters. However, parameters like the battery's polarization resistance and time constant dynamically drift with charge / discharge rates, temperature, and health status, leading to model inaccuracies. Second, purely data-driven methods, such as those based on long short-term memory networks, while capable of capturing dynamic characteristics, suffer from insufficient generalization ability under extreme conditions not covered by training data (such as sudden current changes or deep charge / discharge cycles). Furthermore, their "black box" nature results in a lack of physical interpretability in the predictions, making reliability difficult to guarantee. In addition, existing hybrid methods often only superficially combine physical and data models, failing to achieve deep coupling between physical mechanisms and data patterns within the model. They also generally lack an efficient, adaptive online parameter update mechanism, making it impossible to balance accuracy with computational efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the state of charge (SOC) of an all-vanadium redox flow battery, aiming to achieve high-precision SOC prediction under complex dynamic operating conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for predicting the State of Charge (SOC) of a vanadium redox flow battery, comprising: S1: collecting real-time current and real-time terminal voltage during the operation of the vanadium redox flow battery, and normalizing the real-time current and real-time terminal voltage to obtain a normalized current sequence and a normalized voltage sequence; S2: inputting the normalized current sequence into a physical information long short-term memory (PISM) network, which updates the state of its internal memory cells based on the physical laws of the battery's equivalent circuit model, including a first polarization voltage state, a second polarization voltage state, and a state of charge; S3: based on the memory cell states, utilizing the error between the real-time terminal voltage and the predicted value of the memory cell terminal voltage, dynamically adjusting the contribution weights of the first polarization voltage state, the second polarization voltage state, and the state of charge to the terminal voltage prediction through a multi-head attention mechanism, generating a fused feature representation and inputting it into a fully connected layer, and obtaining the predicted value of the state of charge through nonlinear mapping; S4: determining whether to trigger an update of the circuit parameters in the PISM network based on the error between the real-time terminal voltage and the predicted value of the memory cell terminal voltage in step S3.

[0006] Step S1 also includes establishing a nonlinear mapping relationship between open-circuit voltage and state of charge; the nonlinear mapping relationship is established through a Gaussian process regression model, which is trained using open-circuit voltage data obtained by static measurement at different state of charge calibration points.

[0007] In step S2, the gating parameters of the physical information long short-term memory network are determined by the circuit parameters of the battery equivalent circuit model; among them, the forget gate parameter is determined by the time constant of the RC loop in the equivalent circuit model; the input gate parameter is determined by the polarization resistance, time constant, coulombic efficiency, capacity and sampling interval of the battery in the equivalent circuit model.

[0008] In step S2, the output gate of the physical information long short-term memory network is configured to: combine the updated memory cell state with the input current, multiply the memory cell state with the parameter matrix in the equivalent circuit model to obtain a linear voltage component of the sum of the two polarization voltage values; and multiply the input current with the parameters in the equivalent circuit model. Multiply the values ​​to obtain the ohmic voltage component; simultaneously, substitute the state variables in the memory cell state into the Gaussian process regression model to obtain the nonlinear open-circuit voltage component; finally, superimpose the linear voltage component, the ohmic voltage component, and the nonlinear open-circuit voltage component of the sum of the polarization voltage values ​​to obtain the predicted terminal voltage value.

[0009] Step S3 specifically includes: S31, Query-Key-Value Calculation Sub-step: Multiply the memory cell state and the query vector composed of the terminal voltage prediction error by the corresponding learnable weight matrix to obtain the query matrix, key matrix and value matrix; S32, Attention Fusion Sub-step: Calculate the scaled dot product of the query matrix and the key matrix, obtain the attention weight after normalization, and use the attention weight to perform weighted summation on the value matrix to obtain the fused feature representation; S33, Charge State Prediction Output Sub-step: Input the fused feature representation into the fully connected layer and output the final charge state prediction value.

[0010] In sub-step S31, the query vector consists of the error between the real-time terminal voltage and the terminal voltage prediction value in step S4.

[0011] Step S4 specifically includes: S41, Triggering condition judgment sub-step: Calculate the weighted average of the error between the real-time terminal voltage and the predicted terminal voltage within a preset time window; if the weighted average exceeds a preset threshold, trigger parameter update; S42, Parameter update sub-step: When the triggering condition is met, based on the residual of the polarization voltage mechanism equation and the terminal voltage prediction error, jointly fine-tune the circuit parameters in the physical information long short-term memory network.

[0012] In sub-step S41, the weighting coefficients used in the weighted average calculation increase as time approaches the current moment.

[0013] In sub-step S42, the residual of the mechanistic equation is determined by comparing the difference between the theoretical derivative of the polarization voltage and the model prediction derivative calculated based on the memory cell state.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The SOC prediction method for vanadium redox flow batteries provided in this application embeds the physical laws of the battery's equivalent circuit model into a long short-term memory network, ensuring that the memory cell states strictly follow the dynamic characteristics of an RC circuit. This fundamentally constrains the model output to conform to the battery's working mechanism, reducing prediction deviations under extreme conditions. A nonlinear mapping relationship between open-circuit voltage and state of charge is established through a Gaussian process regression model, accurately fitting their intrinsic correlation. Combined with the linear voltage component output by the physical information long short-term memory network, a linear-nonlinear fusion prediction of the terminal voltage is achieved, effectively correcting the terminal voltage distortion problem caused by polarization effects and improving the basic accuracy of SOC prediction.

[0015] 2. The method provided in this application introduces a multi-head attention mechanism, which dynamically adjusts the contribution ratios of the first polarization voltage, the second polarization voltage, and the state of charge to the prediction result based on the terminal voltage prediction error. This allows the model to adaptively focus on key features under different operating conditions. For example, it focuses on the influence of polarization voltage during high-rate charging and discharging, and strengthens the dominant role of the state of charge under stable operating conditions, further improving the targeting and accuracy of the prediction. The attention mechanism automatically mines the correlation strength of features at different time steps through query-key-value calculation and scaled dot product normalization, avoiding the drawback of traditional models treating all input features equally, making full use of effective information and suppressing noise interference.

[0016] 3. The method provided in this application employs an event-triggered mechanism based on weighted average voltage error. This mechanism determines whether to update parameters by evaluating the error within a preset time window. This avoids the waste of computational resources caused by fixed-period updates and allows for timely response to dynamic changes in battery parameters. When triggering parameter updates, the residual of the polarization voltage mechanism equation and the terminal voltage prediction error are combined for fine-tuning. This ensures that the parameter updates conform to the battery's physical characteristics and optimizes parameter accuracy through backpropagation of data errors, enabling the model to maintain stable predictive performance over the long term. Attached Figure Description

[0017] Figure 1 This is a flowchart of a vanadium redox flow battery SOC prediction method provided in an embodiment of this application. Detailed Implementation

[0018] 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.

[0019] This application provides a method for predicting the state of charge (SOC) of an all-vanadium redox flow battery, exemplarily, such as... Figure 1 As shown. The method includes: S1: Collect the real-time current I, real-time terminal voltage U, open-circuit voltage OCV, and state of charge SOC during the operation of the vanadium redox flow battery, and filter and normalize the collected data. Specifically, normalize the real-time current and real-time terminal voltage to obtain normalized current and voltage sequences.

[0020] For example, real-time current I (negative for discharging, positive for charging), real-time terminal voltage U, and open-circuit voltage OCV are collected during the operation of the vanadium redox flow battery using current and voltage sensors, and the reference true value of the state of charge (SOC) is obtained through experimental measurement methods.

[0021] For example, the normalization formula is as follows: in, Let K be the charging and discharging current at time k. For the minimum current value, This represents the maximum current value. Let be the terminal voltage at time k. Minimum terminal voltage This represents the maximum terminal voltage. Let be the open-circuit voltage at time k. Minimum open-circuit voltage This represents the maximum open-circuit voltage.

[0022] For example, step S1 further includes establishing a nonlinear mapping relationship between open-circuit voltage and state of charge. This nonlinear mapping relationship is established using a Gaussian process regression (GPR) model, which is trained using open-circuit voltage data obtained from static measurements at different state-of-charge calibration points. The output of the Gaussian process regression model is used as the nonlinear component of the output layer of the physical information long short-term memory network, and is superimposed with the linear component of the network's output to achieve terminal voltage prediction.

[0023] Specifically, 11 pre-processed SOC calibration points (intervals of 10%: 0%, 10%, ..., 100%) were used, corresponding to OCV measurements taken under static conditions: in It is the mean function; It is the covariance function. It is a mean function. It is the covariance function. The covariance function can be chosen as a radial basis function.

[0024] Find the optimal value using maximum likelihood estimation. , This allows the OCV predicted by GPR to closely approximate the measured value.

[0025] in (Covariance amplitude) controls the fluctuation of OCV forecast values; (Length scale) determines the sensitivity between SOC and OCV.

[0026] As one possible implementation, this application embodiment also provides a vanadium redox flow battery SOC prediction system for performing a vanadium redox flow battery SOC prediction method. Specifically, the vanadium redox flow battery SOC prediction system includes a nonlinear mapping module for performing the content in step S1.

[0027] Normalized voltage sequences unify the core electrical characteristics of battery operation into a uniform numerical range. This consistency prevents interference with the gating parameter calculation logic of the subsequent Long Short-Term Memory (LSTM) network due to excessively large or small absolute voltage values. It also ensures that the multi-head attention mechanism does not experience weight imbalances due to differences in feature scale when processing associated features, thus improving the overall stability of the model. During the offline training phase, the normalized voltage sequence can serve as a standard label, compared with the terminal voltage predictions output by the PI-LSTM, to optimize the model's initial parameters, such as the covariance magnitude and length scale of the Gaussian process regression model. Standardized voltage data allows the model to more efficiently learn the correlation between voltage and state of charge / polarization voltage, rather than simply fitting absolute voltage values, thereby improving the model's generalization ability under different operating conditions.

[0028] S2: Input the normalized current sequence into the physical information long short-term memory network. The physical information long short-term memory network updates the state of the internal memory cells based on the physical laws of the battery equivalent circuit model. The state of the memory cells includes the first polarization voltage state, the second polarization voltage state, and the state of charge.

[0029] For example, in step S2, the gating parameters of the physical information long short-term memory network are determined by the circuit parameters of the battery equivalent circuit model; wherein, the forget gate parameter is determined by the time constant of the RC loop in the equivalent circuit model; and the input gate parameter is determined by the polarization resistance, time constant, coulombic efficiency, capacity and sampling interval of the battery in the equivalent circuit model.

[0030] As one possible implementation, the vanadium redox flow battery SOC prediction system also includes a physical information long short-term memory module, connected to the nonlinear mapping module, which receives preprocessed data. The physical information long short-term memory module designs gating parameters based on physical constraints, including the calculation of parameters for the forget gate, input gate, and output gate, and updates the battery state using the designed gating parameters.

[0031] The Physical Information Long Short-Term Memory (PSLM) network introduces a second-order equivalent circuit model (ECM), which includes resistance. , , and capacitors , Its expression is as follows: in, Let K be the input current at time k. This is the output voltage. The system state is... , and Resistors and The voltage values ​​at both ends. Let , , , .definition = , = It is a time constant; The sampling interval; S is the coulombic efficiency value; S is the rated capacity of the battery. Let D be the parameter matrix. k Let be the ohmic resistance at time k.

[0032] For example, in step S2, the output gate of the physical information long short-term memory network is configured to: combine the updated memory cell state with the input current, multiply the memory cell state with the parameter matrix in the equivalent circuit model to obtain a linear voltage component of the sum of the two polarization voltage values; and multiply the input current with the parameters in the equivalent circuit model. Multiply the values ​​to obtain the ohmic voltage component; simultaneously, substitute the state variables in the memory cell state into the Gaussian process regression model to obtain the nonlinear open-circuit voltage component; finally, superimpose the linear voltage component, the ohmic voltage component, and the nonlinear open-circuit voltage component of the sum of the polarization voltage values ​​to obtain the predicted terminal voltage value.

[0033] As one possible implementation, the Physical Information Long Short-Term Memory (PI-LSTM) module includes a gating parameter design module and a state update module. The gating parameter design module is primarily used to embed the prior physical laws of the flow battery into the LSTM, resulting in an interpretable PI-LSTM network. The parameters of the LSTM gates are redesigned as forgetting gate variables. Input gate variables and output gate variables , Furthermore, a pre-trained Gaussian process regression (GPR) model is also connected to the output gate to learn the output y. k nonlinear functions .

[0034] Forgotten Gate From the circuit time constant Decide and control the previous memory unit. The proportion retained in the input gate, in a physical sense, describes the attenuation of the capacitor voltage. By polarization resistor Time constant, capacitance Coulomb efficiency The input current is determined by the capacity and sampling interval. This is transformed into a new state increment (the effect of charging and discharging on voltage and SOC). Output gate Combination Give the linear combination of polarization voltage components. Combined with input current Given the ohmic voltage drop of the battery, and then using the Gaussian process regression model. The obtained nonlinear open-circuit voltage components are superimposed to predict the output voltage.

[0035] The PI-LSTM network obtained by the state update module is as follows: in , It is a memory unit.

[0036] S3: Based on the memory cell state, the error between the real-time terminal voltage and the predicted value of the memory cell terminal voltage is used to dynamically adjust the contribution weights of the first polarization voltage state, the second polarization voltage state, and the charge state to the terminal voltage prediction through a multi-head attention mechanism. The fused feature representation is generated and input into the fully connected layer. The predicted value of the charge state is obtained through nonlinear mapping.

[0037] As one possible implementation, the vanadium redox flow battery SOC prediction system also includes a multi-head attention module connected to a physical information long short-term memory module, specifically connected to a state update module. The hidden state output by the state update module serves as the input to the multi-head attention mechanism. Based on this mechanism, the multi-head attention module dynamically adjusts the contribution ratio of polarization voltage (first polarization voltage state U1, second polarization voltage state U2) to the SOC prediction, and outputs the SOC prediction result, i.e., the predicted state of charge, through a fusion layer. Furthermore, for each of the H attention heads, the concatenated hidden state is compared with the corresponding weight matrix (W... q W k and W vThe query, key, and value matrices are multiplied to calculate three matrices: Q, K, and V. Then, the attention score is obtained by scaling the dot product of the Q and K matrices, normalized using the square root of the key vector dimension (dk), and highlighted using a softmax activation function. These attention scores are then weighted onto the value matrix to obtain the outputs (out) of each head. i The contribution weights of each memory unit (U1, U2, SOC) to the terminal voltage are dynamically adjusted based on the terminal voltage error. The outputs of all attention heads are then aggregated in the fusion layer, and the final feature representation is generated by calculating the weighted sum of the attention outputs. For example, step S3 specifically includes: S31, Query-Key-Value Calculation Sub-step: Multiply the memory cell state and the query vector composed of the terminal voltage prediction error by the corresponding learnable weight matrix to obtain the query matrix, key matrix and value matrix.

[0038] In the multi-head attention layer, for the i-th attention (i=1,2,...,H), the query, key, and value matrix is ​​represented by the memory units of S22. Multiplying by the corresponding weight matrix yields: in: For query vectors, For key vectors, It is a value vector; , , These are the learnable weight matrices for the i-th attention head.

[0039] S32, Attention Fusion Sub-step: Calculate the scaled dot product of the query matrix and the key matrix, obtain the attention weights after normalization, and use the attention weights to perform a weighted summation on the value matrix to obtain the fused feature representation.

[0040] Attention scores are calculated by scaling the dot product of the query and the key, and weights are obtained after softmax activation. in: For query vectors, For key vectors, It is the dimension of the key vector (used for normalization to avoid excessively large dot product results); the softmax function converts the scores into weights of 0 to 1, making the sum equal to 1.

[0041] The output of each attention head is a weighted sum of the attention score and the value matrix: S33, Charge State Prediction Output Sub-step: Input the fused feature representation into the fully connected layer, and output the final charge state prediction value through nonlinear mapping.

[0042] The outputs of all attention heads are aggregated through a fusion layer and weighted summation is used to obtain the final features: Where: H is the total number of attention heads; It is the weight of the i-th attention head output.

[0043] In sub-step S31, the query vector consists of the error between the real-time terminal voltage and the predicted terminal voltage value in step S4. The fused features are then used to generate a SOC estimate through a fully connected layer. in: and These are the weights and biases of the fully connected layer; It is the SOC estimate at time step t.

[0044] S4: Based on the error between the real-time terminal voltage and the predicted value of the memory cell terminal voltage in step S3, determine whether to trigger the update of the circuit parameters in the physical information long short-term memory network.

[0045] As one possible implementation, the vanadium redox flow battery SOC prediction system also includes an event triggering condition module connected to the multi-head attention module, used to execute the contents of step S4. The method provided in this application introduces an event triggering mechanism, which determines whether to trigger the Physical Information Long Short-Term Memory (PI-LSTM) network update by setting a time window h and performing a weighted average of the voltage prediction error within the window.

[0046] Step S4 specifically includes: S41, Triggering Condition Judgment Sub-step: Calculate the weighted average of the error between the real-time terminal voltage and the predicted terminal voltage within the preset time window; if the weighted average exceeds the preset threshold, trigger parameter update.

[0047] For example, based on the weighted average voltage prediction error: SOC estimation performance is achieved through S31. and The event triggering conditions were designed based on the weighted average voltage error and the one-step prediction error between the two steps. in: Represents absolute value; The weighting parameter is h; h is the average time range before time k; the trigger threshold is... This is achieved through experience; as one possible implementation, the weighting coefficients used in the weighted average calculation increase as time approaches the current moment, that is... The closer to the current time k, the more parameters The weight can be designed to be larger.

[0048] when When =0, continue using the current physical information Long Short-Term Memory network. When =1, it means that the set "error weighted average is less than or equal to the threshold" "The conditions are not met. The estimation bias of the physical information long short-term memory network is too large and needs to be updated. That is, proceed to step S42 to update the RC parameters in the PI-LSTM network."

[0049] S42, Parameter Update Sub-step: When the triggering condition is met, the circuit parameters in the physical information long short-term memory network are jointly fine-tuned based on the residual of the polarization voltage mechanism equation and the terminal voltage prediction error.

[0050] As one possible implementation, in sub-step S42, the residual of the mechanistic equation is determined by comparing the difference between the theoretical derivative of the polarization voltage and the model prediction derivative calculated based on the memory cell state.

[0051] For example, in the RC model, the dynamic behavior of the polarization voltage is described by a first-order differential equation: Among them U i It is the polarization voltage of the RC branch, and I is the charging and discharging current. R1 and R2 are both polarization resistors, where R is the time constant.

[0052] Its discretized recursive solution is: For the calculation of RC dynamic residuals, the mechanism equation starts from the differential equation, and the theoretical derivative of the polarization voltage should satisfy: In the actual calculations within the code, the derivative is approximated using discrete differences: Therefore, residual is defined as the difference between the mechanistic value and the model prediction: The method provided in this application integrates physical mechanisms and data-driven techniques, constructing a physical information long short-term memory network embedded in the differential equations of a second-order equivalent circuit model. It dynamically adjusts the contribution weights of polarization voltage and state of charge (SOC) to terminal voltage prediction through a multi-head attention mechanism, and combines an event-triggered mechanism based on weighted average voltage error to achieve adaptive updates of circuit parameters. Simultaneously, it utilizes a Gaussian process regression model to establish a nonlinear mapping between open-circuit voltage and SOC to correct terminal voltage distortion caused by polarization effects. This effectively solves the problem of insufficient model stability caused by dynamic changes in battery parameters under dynamic charge-discharge conditions, significantly improving the accuracy and robustness of SOC prediction for vanadium redox flow batteries, reducing dependence on massive training and initial calibration data, and the event-triggered mechanism balances computational efficiency and engineering practicality, facilitating deployment in embedded battery management systems.

[0053] 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.

[0054] 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 method for predicting the state of charge (SOC) of a vanadium redox flow battery, comprising: S1: collecting real-time current and real-time terminal voltage in the operation process of the all-vanadium redox flow battery, and performing normalization processing on the real-time current and the real-time terminal voltage to obtain a normalized current sequence and a normalized voltage sequence; S2: inputting the normalized current sequence into a physical information long short-term memory network, the physical information long short-term memory network updating an internal memory cell state based on a physical law of a battery equivalent circuit model, the memory cell state including a first polarization voltage state, a second polarization voltage state and a state of charge; S3: based on the memory cell state, using an error between a real-time terminal voltage and a memory cell terminal voltage prediction value, dynamically adjusting a contribution weight of the first polarization voltage state, the second polarization voltage state and the state of charge to terminal voltage prediction through a multi-head attention mechanism, generating a fused feature representation and inputting it into a fully connected layer, and obtaining a prediction value of the state of charge through nonlinear mapping; S4: determining whether to trigger an update of circuit parameters in the physical information long short-term memory network according to an error between the real-time terminal voltage and the memory cell terminal voltage prediction value in step S3.

2. A method of predicting SOC of a vanadium redox flow battery according to claim 1, characterized in that, In step S1, a step of establishing a nonlinear mapping relationship between open circuit voltage and state of charge is further included; the nonlinear mapping relationship is established by a Gaussian process regression model, and the Gaussian process regression model is trained by using open circuit voltage data measured at different state of charge calibration points.

3. A method of predicting SOC of a vanadium redox flow battery according to claim 2, characterized in that, In step S2, gate parameters of the physical information long short-term memory network are determined by circuit parameters of the battery equivalent circuit model; wherein, a forgetting gate parameter is determined by a time constant of an RC circuit in the equivalent circuit model; an input gate parameter is determined by a polarization resistance, a time constant, a coulomb efficiency, a capacity and a sampling interval of the battery in the equivalent circuit model.

4. A method of predicting SOC of a vanadium redox flow battery according to claim 3, characterized in that, In step S2, the output gate of the physical information long short-term memory network is configured as: Combining the updated memory cell state and the input current, the memory cell state is multiplied with a parameter matrix in the equivalent circuit model to obtain a linear voltage component of the sum of two polarization voltage values; the input current is multiplied with a parameter matrix in the equivalent circuit model to obtain an ohmic voltage component; meanwhile, a state variable in the memory cell state is substituted into the Gaussian process regression model to obtain a nonlinear open circuit voltage component; finally, the linear voltage component of the sum of the two polarization voltage values, the ohmic voltage component and the nonlinear open circuit voltage component are superimposed to obtain the terminal voltage prediction value. Combining the updated memory cell state and the input current, the memory cell state is multiplied with a parameter matrix in the equivalent circuit model to obtain a linear voltage component of the sum of two polarization voltage values; the input current is multiplied with a parameter matrix in the equivalent circuit model to obtain an ohmic voltage component; meanwhile, a state variable in the memory cell state is substituted into the Gaussian process regression model to obtain a nonlinear open circuit voltage component; finally, the linear voltage component of the sum of the two polarization voltage values, the ohmic voltage component and the nonlinear open circuit voltage component are superimposed to obtain the terminal voltage prediction value.

5. A method for predicting the state of charge (SOC) of an all-vanadium redox flow battery according to claim 1, characterized in that, In step S3, it specifically includes: S31, a query-key-value calculation sub-step: multiplying the memory cell state and a query vector composed of terminal voltage prediction errors by corresponding learnable weight matrices respectively to obtain a query matrix, a key matrix and a value matrix; S32, an attention fusion sub-step: calculating a scaled dot product of the query matrix and the key matrix, obtaining an attention weight after normalization, and performing weighted summation on the value matrix using the attention weight to obtain a fused feature representation; S33, a state of charge prediction output sub-step: inputting the fused feature representation into a fully connected layer, and outputting a final state of charge prediction value through nonlinear mapping.

6. A method of predicting SOC of a vanadium redox flow battery according to claim 5, characterized in that, In sub-step S31, the query vector is composed of an error between the real-time terminal voltage and the terminal voltage prediction value in step S4.

7. A method for predicting the state of charge (SOC) of an all-vanadium redox flow battery according to claim 1, characterized in that, In step S4, specifically comprising: S41, trigger condition judgment sub-step: calculate the weighted average of the error between the real-time terminal voltage and the terminal voltage prediction value within the preset time window; if the weighted average exceeds the preset threshold, trigger parameter update; S42, parameter updating sub-step: when the trigger condition is met, based on the mechanism equation residual of the polarization voltage and the terminal voltage prediction error, the circuit parameters in the physical information long short-term memory network are jointly fine-tuned.

8. A method of predicting SOC of a vanadium redox flow battery according to claim 7, characterized in that, In sub-step S41, the weight coefficient used in the calculation of the weighted average increases with time approaching the current time.

9. A method of predicting SOC of a vanadium redox flow battery according to claim 7, characterized in that, In sub-step S42, the mechanism equation residual is determined by comparing the difference between the theoretical derivative of the polarization voltage and the model predicted derivative calculated based on the memory cell state.

Images