All-vanadium redox flow battery state-of-charge prediction method based on hierarchical physical prior strengthening

Abstract

The invention relates to the technical field of battery management systems, discloses an all-vanadium redox flow battery state-of-charge prediction method based on hierarchical physical prior strengthening, and aims to realize accurate prediction of the battery state-of-charge. The method comprises the following steps: collecting an endogenous time sequence and an exogenous time sequence of all-vanadium redox flow battery operation; respectively performing embedding processing on the endogenous time sequence and the exogenous time sequence to obtain an endogenous feature marker and an exogenous feature marker; wherein the endogenous time sequence is a charge state sequence, and the exogenous time sequence comprises current, terminal voltage, open-circuit voltage and temperature. And sequentially carrying out local time sequence feature extraction, physical memory state tracking, electrochemical dynamic state correction, endogenous and endogenous feature fusion and thermodynamic constraint enhancement processing on the endogenous feature mark and the exogenous feature mark, and outputting a feature vector after physical consistency enhancement. And performing constraint adjustment on the feature vector after physical consistency enhancement based on the law of conservation of charge, and outputting a final state of charge prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management system, in particular to a full vanadium redox flow battery state of charge prediction method based on layered physical prior reinforcement. BACKGROUND

[0002] As a kind of grid-level energy storage technology with intrinsic safety, capacity scalability and long cycle life, full vanadium redox flow battery shows broad application prospects in smoothing new energy fluctuations and peak shaving. Its state of charge is the core parameter of battery management system for energy scheduling, health management and safety warning. Achieving high-precision and strong-robustness estimation of SOC is the key to ensuring stable and efficient operation of the entire energy storage system.

[0003] Currently, the estimation methods of SOC can be divided into three categories: direct measurement-based methods, model-based methods and data-driven methods. Although the ampere-hour integral method, as a typical direct measurement method, is simple and easy to implement, it has problems such as initial error accumulation and measurement noise amplification, making it difficult to maintain accuracy in long-term dynamic operation. Model-based methods, such as combining Kalman filter with equivalent circuit model or electrochemical model, can improve estimation accuracy through feedback mechanism, but their performance is heavily dependent on the accuracy of the selected model. There are complex multi-physical field coupling effects inside the full vanadium redox flow battery, and it is extremely challenging to build a general model that is accurate enough under various dynamic operating conditions, resulting in a complex and less universal method. Pure data-driven methods, such as using various neural networks, can learn the complex nonlinear mapping relationship between input and output from historical data, but this kind of method is essentially a "black box", and its prediction results lack clear physical mechanism constraints. In operating conditions not covered by the training data, such as sudden current changes and extreme temperatures, the model is prone to output results that violate electrochemical laws, with poor physical consistency and potential safety hazards. SUMMARY

[0004] The present application aims to provide a full vanadium redox flow battery state of charge prediction method based on layered physical prior reinforcement to solve the problems raised in the background.

[0005] In order to solve the above technical problems, the present application provides the following technical solutions: a vanadium redox flow battery state of charge prediction method based on layered physical prior reinforcement. It comprises the following steps: S1: collecting the endogenous time series and exogenous time series of the vanadium redox flow battery operation, and embedding the endogenous time series and exogenous time series respectively to obtain endogenous feature markers and exogenous feature markers; wherein the endogenous time series is the state of charge sequence, and the exogenous time series includes current, terminal voltage, open circuit voltage and temperature; S2: sequentially performing local time sequence feature extraction, physical memory state tracking, electrochemical dynamics state correction, endogenous and exogenous feature fusion and thermodynamic constraint reinforcement processing on the endogenous feature markers and exogenous feature markers obtained in step S1, and outputting a feature vector with enhanced physical consistency; S3: based on the law of conservation of charge, the feature vector with enhanced physical consistency is adjusted, and the final state of charge prediction result is output.

[0006] Step S1 specifically includes: dividing the endogenous time series into multiple time segments according to the charge and discharge period, converting the features of each time segment and adding position information to form endogenous segment markers, and introducing a global marker to integrate the overall features of the sequence; the current, terminal voltage, open circuit voltage and temperature in the exogenous time series are respectively divided into segments according to the same charge and discharge period as the endogenous time series, the variable level feature conversion is performed on the period segments corresponding to the current, terminal voltage, open circuit voltage and temperature, the independent variable markers dedicated to each variable are generated, and the variable markers of all types constitute the exogenous feature markers, wherein the conversion process of the exogenous feature markers is related to the physical relationship between the variables in the exogenous time series and the state of charge.

[0007] The local time sequence feature extraction in step S2 includes: performing convolution operation on the endogenous segment markers and the global marker, combining residual connection and pooling operation, and outputting the extracted sequence features and updated global features.

[0008] The physical memory state tracking in step S2 includes: based on the law of conservation of mass, taking the updated global feature as the initial state, combining the current information in the exogenous time series, and using the gated recurrent unit to recursively update a physical memory state representing the charge accumulation effect, and outputting a physical memory sequence.

[0009] The electrochemical dynamics state correction in step S2 includes: based on the electrode reaction kinetics equation, using the actual measured current to calibrate the physical memory sequence, and outputting the corrected global feature conforming to the electrode reaction kinetics law.

[0010] The internal and external feature fusion and thermodynamic constraint strengthening processing in step S2 includes: fusing the modified global feature and the external feature label, and outputting the fused global feature label; based on the Nernst equation, the fused global feature label and temperature information are used to thermodynamically correct the features, and a physical consistency enhanced feature vector is output.

[0011] In step S3, the physical consistency enhanced feature vector is subjected to linear projection and activation function processing to obtain a preliminary state of charge prediction value, and the current in the external time sequence is introduced to correct the preliminary state of charge prediction value based on the law of conservation of charge; the difference between the corrected state of charge prediction result and the reference true value is calculated, and the prediction result is optimized according to the difference, and the final state of charge prediction result is output.

[0012] The correction of the preliminary state of charge prediction value based on the law of conservation of charge specifically includes: introducing a charge loss term and an electrolyte volume factor, calculating an effective charge amount, and correcting the preliminary state of charge prediction value based on the relationship between the effective charge amount and the rated capacity of the battery.

[0013] The electrode reaction kinetics equation is the Butler-Volmer equation.

[0014] Compared with the prior art, the beneficial effects achieved by the present application are: First, the method provided by the present application adopts a hierarchical embedding design, divides the endogenous SOC according to the charge and discharge cycle and adds a global label to ensure the time sequence integrity; the exogenous sequence is divided according to the same cycle, and the physical relationship between the exogenous variable and the SOC is learned through the linear projector parameter to associate the multi-source data from statistical correlation to physical correlation, providing bottom support for subsequent feature enhancement and prediction accuracy.

[0015] Secondly, the application is based on a physical memory state tracking mechanism, taking the updated global feature as the initial state, combining with the exogenous current information, recursively updating the physical memory state through the gated recurrent unit and following the law of conservation of mass, realizing the accurate timing coding of the charge accumulation effect, and making the feature have the charge change memory ability conforming to the physical nature, avoiding the charge logic contradiction problem prone to occur in pure data driven models. With the aid of electrochemical kinetics state correction means, based on the Butler-Volmer equation, the actual measured current is used to calibrate the physical memory sequence, so that the corrected global feature strictly conforms to the kinetics mechanism of the electrode reaction, solving the pain point that the feature deviates from the electrochemical law under the condition of severe current mutation of the traditional model. In addition, through the internal and external feature fusion and thermodynamic constraint strengthening processing, the fusion of the corrected global feature and the exogenous feature label is completed first, and then the thermodynamic correction is implemented based on the Nernst equation combined with the temperature information, and finally the feature vector with enhanced physical consistency is output. The vector not only retains the ability of data driven methods to capture complex nonlinear relationships, but also integrates physical constraints at the thermodynamic level, and can adapt to complex working conditions such as extreme temperature. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a vanadium redox flow battery state of charge prediction method based on hierarchical physical prior reinforcement provided by an embodiment of the application is shown in the figure. Figure 2 A schematic diagram of a physical feature enhancement cascade module provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] It should be understood that the method provided by the embodiments of the application is realized based on a prediction model, and the prediction model includes a hierarchical embedding module, a physical feature enhancement cascade module and a conservation driven module.

[0019] The core function of the hierarchical embedding module is to collect multi-source time series in the operation process of the vanadium redox flow battery, and to perform targeted embedding processing on the physical attribute differences between the endogenous time series and the exogenous time series, to provide basic features with physical correlation for subsequent feature enhancement.

[0020] Specifically, the hierarchical embedding module collects two types of key time series: one is the endogenous time series, i.e., the state of charge (SOC) sequence directly reflecting the state of battery energy storage; the other is the exogenous time series, i.e., the external operating parameter sequence affecting the change of the state of charge, including the current sequence, the terminal voltage sequence, the open circuit voltage sequence, and the stack temperature sequence. In the embedding processing stage, the hierarchical embedding module performs adaptive operations on the two types of sequences: for the endogenous time series, the hierarchical embedding module divides it into multiple independent time segments (each segment corresponds to a complete charging cycle or discharging cycle) according to the battery charging and discharging cycle, performs feature conversion on each time segment and adds position information to retain the time sequence logic, forming the endogenous segment marker; at the same time, a global marker is introduced to integrate the overall features of all endogenous segment markers, and finally combined to form the endogenous feature marker. The current, terminal voltage, open circuit voltage, and temperature in the exogenous time series are respectively divided into segments according to the same charging and discharging cycle as the endogenous time series. The variable-level feature conversion is performed on the period segments corresponding to the current, terminal voltage, open circuit voltage, and temperature, generating independent variable markers exclusive to each variable, and the variable markers of all types jointly constitute the exogenous feature marker. In the feature conversion process, the physical relationship between each variable (such as current, temperature) in the exogenous time series and the state of charge (such as the coupling relationship between the current change rate and the state of charge change trend) is learned through parameters, and finally the exogenous feature marker is generated.

[0021] The physical feature enhancement cascade module is connected with the hierarchical embedding module, including the endogenous sequence convolution module, the physical memory unit, the electrochemical dynamics constraint module, the feature fusion layer, and the Nernst drive feedforward layer connected in turn. The physical feature enhancement cascade module is the core of the prediction method of the present application, its function is to receive the endogenous feature marker and the exogenous feature marker output by the hierarchical embedding module, through multiple steps of physical prior constraint and data feature extraction, gradually strengthen the physical consistency of the features, and finally output the feature vector after the physical consistency is enhanced. Corresponding to the local time sequence feature extraction, the physical memory state tracking, the electrochemical dynamics state correction, and the internal and external feature fusion and thermodynamic constraint enhancement processing in the method steps.

[0022] For local time sequence feature extraction, the endogenous segment marker and the global marker in the endogenous feature marker are processed through convolution operation combined with residual connection and pooling operation. Convolution operation accurately captures the local dynamic correlation of endogenous time series, such as the mutation characteristics of state of charge at charging and discharging switching, residual connection ensures the continuity of deep feature transmission to avoid gradient disappearance, and pooling operation aggregates multi-time step features to filter local noise, finally outputs the extracted sequence features and updated global features.

[0023] For physical memory state tracking, based on the law of conservation of mass, the updated global features output by the local temporal feature extraction are taken as the initial state, combined with the current information in the exogenous time sequence, and a physical memory state representing the cumulative effect of charge is recursively updated using a gated recurrent unit, which encodes the dynamic influence of current integration on the state of charge in real time, and finally outputs the physical memory sequence.

[0024] For electrochemical kinetics state correction, based on the Butler-Volmer equation, the physical memory sequence output by the physical memory state tracking is calibrated for physical consistency using the actual measured current in the exogenous time sequence, ensuring that the calibrated features conform to the kinetics of the battery electrode reaction, and finally output the corrected global features.

[0025] For internal and external feature fusion and thermodynamic constraint reinforcement processing, first, the corrected global features output by the electrochemical kinetics state correction are connected with the exogenous feature markers output by the hierarchical embedding module, and the fused global feature markers are output through fusion processing; then, based on the Nernst equation, combined with the stack temperature information in the exogenous time sequence, the fused global feature markers are thermodynamically corrected to quantify the thermodynamic correlation between the state of charge and temperature and electrode potential, and finally output the feature vector with enhanced physical consistency.

[0026] The conservation driven module is connected with the physical feature enhancement cascade module, and the conservation driven module includes a charge conservation activation layer that receives the physical consistency enhanced feature vector output by the physical feature enhancement cascade module, and adjusts the features based on the law of conservation of charge through multiple rounds of correction and optimization, and finally outputs the accurate state of charge prediction result of the vanadium redox flow battery. The specific processing process of the conservation driven module is as follows: first, the physical consistency enhanced feature vector is subjected to linear projection and activation function processing to obtain a preliminary state of charge prediction value; then, the current information in the exogenous time sequence is introduced, combined with the non-ideal factors in the actual operation of the battery, such as self-discharge, side reactions, leakage current, charge loss term and electrolyte volume factors are introduced, based on the law of conservation of charge, the effective charge amount participating in the effective electrochemical reaction is calculated, and the ratio of the effective charge amount to the rated capacity of the battery is used to correct the preliminary state of charge prediction value; finally, the difference between the corrected state of charge prediction result and the state of charge reference true value is calculated, and the prediction result is optimized again according to the difference, and finally the final state of charge prediction result meeting the engineering actual demand is output.

[0027] The prediction model provided in the application realizes the improvement of the accuracy and reliability of the state of charge prediction of the all-vanadium redox flow battery under the dynamic charging and discharging working condition by the synergistic work of the three modules, the deep fusion of the data-driven feature extraction capability and the physical priori law (conservation of mass, electrochemical kinetics, thermodynamics, conservation of charge), avoiding the problem of insufficient physical consistency of the traditional pure data-driven model, and solving the defect of weak dynamic adaptability of the traditional physical model.

[0028] With reference to Figure 1 The embodiment of the application provides a state of charge prediction method for an all-vanadium redox flow battery based on hierarchical physical priori reinforcement. The method comprises the following steps: Step S1: collecting endogenous time series and exogenous time series of the operation of the all-vanadium redox flow battery, and respectively performing embedding processing on the endogenous time series and the exogenous time series to obtain endogenous feature labels and exogenous feature labels. The endogenous time series is a state of charge sequence, and the exogenous time series comprises current I, terminal voltage U, open circuit voltage OCV and temperature T.

[0029] For example, step S1 specifically comprises: dividing the endogenous time series into a plurality of time segments according to the charging and discharging period, each segment corresponding to a complete charging or discharging cycle, so as to ensure that the segment contains a complete SOC change trend. A trainable linear projector (PatchEmbed) is used to perform feature conversion on each time segment and add position information. First, the segment is added with position embedding to retain the time sequence information, and then mapped into a fixed-dimensional vector (D-dimensional) to form endogenous segment labels (patch_tokens), and a global label (global_token) is introduced to integrate the overall features of the sequence. The global label learns the macroscopic change law of the endogenous time series in the entire operation period, and is input into the processing unit responsible for the preliminary integration of the endogenous features in the physical feature enhancement cascade module together with the segment labels.

[0030] The hierarchical embedding module associates all segment markers with the information of the global marker when processing the endogenous time series. For example, when a segment marker reflects that the current segment is a charging phase (SOC from 30% to 80%), the global marker will integrate the information of other segment markers at the same time (such as the previous segment is a discharging phase (SOC from 80% to 30%), and the next segment will continue to charge to 100%). Through training, gradually learn to identify and remember the macroscopic law of state of charge sequence, and finally become a carrier that can represent the global characteristics of the whole SOC sequence. Each segment marker corresponds to the local characteristics of a complete charging and discharging cycle, such as the rate of SOC from 20% to 95% during charging, and the sudden drop amplitude of SOC due to current surge during discharging. These local details are the key to capturing the dynamic fluctuations of state of charge. The global trend carried by the global marker will form a physically reasonable constraint on the local characteristics of the segment marker, avoiding interference of subsequent feature processing by local anomalies.

[0031] The four types of variables in the exogenous time series, current, terminal voltage, open circuit voltage and temperature, are divided into segments according to the same charging and discharging cycle as the endogenous time series. A trainable linear projector is configured for each type of exogenous variable to convert the variable-level features of the corresponding cycle segments and generate independent variable markers specific to each variable. The variable markers of each type form the exogenous feature markers; the conversion process of the exogenous feature markers is related to the physical relationship between the variables in the exogenous time series and the state of charge, such as the coupling relationship between current and SOC change rate.

[0032] The combined sensor unit is composed of current sensor, voltage sensor and temperature sensor, which can receive the exogenous time series (current I, terminal voltage U, open circuit voltage OCV, stack temperature T) of the all-vanadium redox flow battery in real time, and obtain the reference true value of the endogenous time series through experimental measurement method.

[0033] Step S2: sequentially performing local time sequence feature extraction, physical memory state tracking, electrochemical dynamics state correction, endogenous and exogenous feature fusion, and thermodynamic constraint strengthening processing on the endogenous feature markers and exogenous feature markers obtained in step S1, and outputting the feature vectors with enhanced physical consistency.

[0034] Reference Figure 2 As a possible implementation, the local time sequence feature extraction in step S2 includes: performing convolution operation on the endogenous segment markers and the global marker, and combining residual connection and pooling operation to output the extracted sequence features and the updated global features.

[0035] The endogenous sequence convolution module in the physical feature enhancement cascade module is used to process the endogenous segment markers and the global marker. The processing of the endogenous sequence convolution module includes an initialization stage and a forward propagation stage. In the initialization stage, the endogenous sequence convolution module completes the construction of three core components: First, the convolutional layer stack is performed, creating a 2-layer one-dimensional convolutional structure, each layer consisting of a one-dimensional convolutional layer (Conv1d), an activation function (ReLU), and batch normalization (BatchNorm); the padding filling strategy is used to ensure that the input and output sequence lengths are consistent, and the first layer uses in_dim as the input dimension, and the subsequent layers use out_dim dimensions.

[0036] Residual connection preparation: when the input dimension and the output dimension are inconsistent, a 1x1 convolutional layer is automatically created to align the dimensions, ensuring the feasibility of residual connection and ensuring that the input and convolutional output can be directly added.

[0037] Normalization and pooling: LayerNorm is used to normalize the features, and global average pooling is used to prepare for the subsequent generation of global feature representation.

[0038] In the forward propagation phase, the endogenous sequence convolutional module completes feature processing and extraction in five steps: Dimension adjustment: the endogenous features output by step S1 with a shape of [B, N, D] are exchanged in the second and third dimensions through dimension transposition (transpose), where B is the batch size, N is the number of patches, and D is the embedding dimension.

[0039] Convolution processing: the input after dimension adjustment is fed into the convolutional layer stack for feature extraction, and the output still maintains the shape of [B, N, D].

[0040] Residual connection: if there is a difference between the input and output dimensions, the residual convolution (res_conv) is used to adjust the residual dimension, and then the convolutional output and the residual term are added to complete the residual connection operation, ensuring smooth information flow in deep networks and alleviating the gradient vanishing problem.

[0041] Dimension calibration and layer normalization: the dimensions are exchanged again to restore the original sequence shape of [B, N, D], and LayerNorm is applied to further stabilize the feature distribution.

[0042] Global feature extraction: the dimensions are adjusted again to [B, N, D] to adapt to the pooling operation, and then all time steps of each feature dimension are aggregated into a single value through global average pooling, finally outputting the extracted sequence features with a shape of [B, N, D] and the updated global features with a shape of [B, D].

[0043] As a possible implementation, the physical memory state tracking in step S2 includes: based on the law of conservation of mass, taking the updated global features as the initial state, combining the current information in the exogenous time sequence, and using the gated recurrent unit to recursively update a physical memory state representing the cumulative effect of electric charge, outputting the physical memory sequence.

[0044] The physical memory unit in the physical feature enhancement cascade module receives the updated global features and the current term in the exogenous time series. The physical memory unit for explicitly tracking the cumulative effect of SOC is based on the law of conservation of mass, updates the physical memory state in real time, encodes the dynamic influence of current integration on SOC, combines the updated global features and current information at each time step, recursively updates the physical state, and provides the network with physically consistent time series memory features.

[0045] In the physical memory unit, the input layer receives the output of the endogenous sequence convolution module, the hidden layer takes the global label as the initial hidden state, and the current Physical normalization At the same time, the current hidden state is spliced with the normalized current and input into the gated recurrent unit to obtain a new hidden state, which is saved in the sequence, and finally output by the output layer.

[0046] For example, the design formula of the physical memory unit is: wherein, is the reset gate, is the update gate, which controls the degree of "forgetting" and "updating" respectively, is the candidate hidden state, is the final hidden state, is the hidden state of the previous time step, is the current input, which includes the state of the previous time step and the normalized current, is the Sigmoid function, tanh is the hyperbolic tangent function, W is the input weight matrix, U is the weight matrix of the hidden state, and b represents the bias term, which is a learnable parameter for linear transformation in neural networks, is the element-wise multiplication operation.

[0047] As a possible implementation, the electrochemical kinetics state correction in step S2 includes: based on the electrode reaction kinetics equation, the physical memory sequence is calibrated using the actual measured current, and the corrected global features conforming to the electrode reaction kinetics law are output.

[0048] In some embodiments, the electrode reaction kinetics equation is the Butler-Volmer equation.

[0049] The input of the electrochemical kinetics constraint module is the physical memory sequence output by the physical memory unit, and the theoretical current value is calculated by the Butler-Volmer equation.

[0050] The Butler-Volmer calculation formula is as follows: wherein F is a Faraday constant, and takes 96500 C / mol, is an activation overpotential, and is calculated by an electrode potential E and an equilibrium potential , R is a gas constant, and takes 8.314 J / (K·mol), is an exchange current, and is a transfer coefficient, T is a temperature, and F is a Faraday constant, F=96485 C / mol. Finally, the hidden state is scaled in physical consistency with the actual current I, so that the hidden state conforms to the electrode reaction kinetics.

[0051] The physical consistency scaling result is as follows: wherein is a hidden state output by a physical memory unit, I is an actual current, is a theoretical current calculated by the above Butler-Volmer calculation, and the physical calibration of the hidden state is achieved by the ratio of the actual current to the theoretical current, so that the hidden state conforms to the electrode reaction kinetics.

[0052] As a possible implementation manner, the internal and external feature fusion and thermodynamic constraint strengthening processing in step S2 includes: fusing the modified global feature and the external feature label to output the fused global feature label. Based on the Nernst equation, the features are thermodynamically corrected based on the fused global feature label and temperature information, and a feature vector after physical consistency enhancement is output.

[0053] The input of the feature fusion layer is the feature vector after physical consistency enhancement and the external feature label. After the two are connected, they are fused through a fully connected layer, and finally the fused global feature label is output. The input of the Nernst driven feedforward layer is the fused global feature label and temperature T, wherein the fused global feature label implicitly contains the SOC related features. The implicit value of the SOC is obtained by decoding from the global feature label, and is substituted into the Nernst equation to calculate the electrode potential correction term (ΔE). Finally, the feature vector after correction and physical consistency enhancement is obtained.

[0054] As an example, for a full vanadium redox flow battery, the Nernst equation is as follows: After moving the equal sign to the left and right, the following formula is obtained: wherein F is Faraday's constant of 96485 C / mol, R is the gas constant of 8.314 J / (K mol), is the standard electrode potential. The number of transferred electrons n = 1, and the correction term directly quantifies the influence of SOC change on electrode potential. The Nernst-driven feedforward layer realizes physical enhancement of features by injecting the correction term AE into the global feature vector x. The specific correction formula is: wherein W is a trainable weight matrix for mapping the scalar correction term to the D-dimensional feature space; b is a bias term for fine-tuning the corrected feature distribution to avoid excessive correction that may mask the data-driven features; and y is the corrected global feature vector that retains the learned correlations in x and encodes the thermodynamic constraints of the Nernst equation.

[0055] Step S3: Based on the charge conservation law, the physically consistent enhanced feature vector is adjusted to output the final state of charge prediction result.

[0056] For example, in step S3, the physically consistent enhanced feature vector is subjected to linear projection and activation function processing to obtain a preliminary state of charge prediction value, and the current in the exogenous time sequence is introduced. Based on the charge conservation law, the preliminary state of charge prediction value is corrected. The difference between the corrected state of charge prediction result and the reference true value is calculated, and the prediction result is optimized according to the difference to output the final state of charge prediction result.

[0057] Ideally, the total electric quantity passing through the external circuit is equal to the electric quantity of the internal electrochemical reaction, but in practice, the battery has charge loss (such as self-discharge, side reaction, leakage current, etc.), and the loss degree is related to the electrolyte volume, so the ideal electric quantity needs to be introduced into the loss correction to obtain the effective charge quantity, and finally associate the SOC.

[0058] As a possible implementation manner, the correction of the preliminary state of charge prediction value based on the charge conservation law specifically includes: introducing a charge loss term and an electrolyte volume factor, calculating an effective charge quantity, and correcting the preliminary state of charge prediction value based on the relationship between the effective charge quantity and the rated capacity of the battery.

[0059] The core logic of charge conservation is that the change rate of total charge quantity is equal to the applied current minus the loss current, which can be expressed as: wherein is the effective charge quantity, is the charging and discharging current, is the loss current, representing the charge rate that does not participate in the effective reaction due to non-ideal factors, such as self-discharge, side reaction, leakage current, etc.

[0060] In the formula, is the charge loss resistance, is the electrolyte volume.

[0061] The core definition of the state of charge is the ratio of the effective charge amount to the rated capacity of the battery, that is: Wherein, is the amount of charge participating in the effective electrochemical reaction, C is the rated capacity of the battery, and the derivative of the above equation on both sides can obtain the rate of change of SOC with time: Integrating the above formula can obtain the correction amount of SOC, that is: The mathematical expression of the corrected state of charge is: Wherein, is a Sigmoid function, used to map the SOC value to the range of 0-1, and C is the battery capacity.

[0062] As a possible implementation, a data loss function is set to make a residual between the SOC prediction result output by the prediction model and the true value, correct the prediction result, and obtain the final prediction result, wherein the data loss function is as follows: Wherein, represents the prediction value of the state of charge of the battery at time i, represents the true value of the state of charge of the battery at time i, and N refers to the number of samples participating in the loss calculation.

[0063] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing embodiments of the present application have been described in detail, for those skilled in the art, it still can be modified to the technical solutions recorded in the foregoing embodiments, or equivalent replacement of part of the technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for state of charge prediction of all-vanadium redox flow battery based on hierarchical physical prior reinforcement, characterized in that, The method comprises the following steps: S1, collecting an endogenous time series and an exogenous time series of a full vanadium flow battery, and embedding the endogenous time series and the exogenous time series to obtain endogenous feature labels and exogenous feature labels; wherein the endogenous time series is a state of charge sequence, and the exogenous time series comprises current, terminal voltage, open circuit voltage and temperature; S2, sequentially performing local time sequence feature extraction, physical memory state tracking, electrochemical dynamics state correction, endogenous and exogenous feature fusion and thermodynamic constraint strengthening processing on the endogenous feature labels and the exogenous feature labels obtained in step S1 to output a feature vector with enhanced physical consistency; and S3, performing constraint adjustment on the feature vector with enhanced physical consistency based on the law of conservation of charge to output a final state of charge prediction result.

2. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 1, characterized in that, Step S1 specifically comprises: dividing the endogenous time series into multiple time segments according to the charge and discharge cycle, converting the features of each time segment and adding position information to form endogenous segment labels, and introducing a global label to integrate the overall features of the sequence; the current, terminal voltage, open circuit voltage and temperature in the exogenous time series are divided into segments according to the same charge and discharge cycle as the endogenous time series, the variable level feature conversion is performed on the period segments corresponding to the current, terminal voltage, open circuit voltage and temperature, the independent variable labels specific to each variable are generated, and the variable labels of all types jointly constitute the exogenous feature labels; wherein the conversion process of the exogenous feature labels is related to the physical relationship between the variables in the exogenous time series and the state of charge.

3. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 2, characterized in that, The local time sequence feature extraction in step S2 comprises: performing convolution operation on the endogenous segment labels and the global label, combining residual connection and pooling operation, and outputting the extracted sequence features and the updated global features.

4. The method of state of charge prediction for a vanadium redox flow battery based on hierarchical physical prior reinforcement according to claim 3, characterized in that, The physical memory state tracking in step S2 comprises: based on the law of conservation of mass, taking the updated global features as the initial state, combining the current information in the exogenous time series, updating a physical memory state representing the charge accumulation effect by using a gated recurrent unit, and outputting a physical memory sequence.

5. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 4, characterized in that, The electrochemical dynamics state correction in step S2 comprises: based on the electrode reaction kinetics equation, calibrating the physical memory sequence by using the actual measured current, and outputting a corrected global feature conforming to the electrode reaction kinetics law.

6. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 5, characterized in that, The endogenous and exogenous feature fusion and thermodynamic constraint strengthening processing in step S2 comprises: fusing the corrected global feature and the exogenous feature label to output a fused global feature label; based on the Nernst equation, thermodynamically correcting the features by using the fused global feature label and temperature information to output the feature vector with enhanced physical consistency.

7. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 1, characterized in that, In step S3, the physical consistency enhanced feature vector is subjected to linear projection and activation function processing to obtain a preliminary state of charge prediction value, and the current in the exogenous time sequence is introduced to correct the preliminary state of charge prediction value based on the law of conservation of charge; the difference between the corrected state of charge prediction result and a reference true value is calculated, and the prediction result is optimized according to the difference to output a final state of charge prediction result.

8. The hierarchical physical prior reinforcement-based state of charge prediction method for a vanadium redox flow battery according to claim 7, characterized in that, The correction of the preliminary state of charge prediction value based on the law of conservation of charge specifically includes: introducing a charge loss term and an electrolyte volume factor, calculating an effective charge amount, and correcting the preliminary state of charge prediction value based on the relationship between the effective charge amount and the rated capacity of the battery. 9.The method of state-of-charge prediction for all-vanadium redox flow battery based on hierarchical physical prior reinforcement according to claim 5, characterized in that, The electrode reaction kinetics equation is a Butler-Volmer equation.

Images