Fault diagnosis method for multi-stack system of all-vanadium redox flow battery

Abstract

The invention belongs to the technical field of battery management, particularly relates to a fault diagnosis method for a multi-stack system of an all-vanadium redox flow battery, and aims to accurately evaluate the health state of each stack in the multi-stack system. The method comprises the steps of collecting operation data of the all-vanadium redox flow battery system in real time and performing standardization processing. And constructing a neural network model, and training the neural network model by using the historical operation data. Inputting the system-level global features after real-time processing into a trained shared feature extraction module to output a shared feature vector; and fusing the shared feature vector and the real-time pile level local feature of each electric pile, inputting the fused feature vector and the real-time pile level local feature into corresponding branches in a multi-task variational reasoning module respectively, and outputting a state-of-charge point estimation value and estimation uncertainty of each electric pile in parallel. Calculating a system anomaly index and an individual anomaly index based on the charge state point estimation values and the estimation uncertainty of all the electric piles; and when any index continuously exceeds the corresponding preset threshold value, determining that a potential fault exists and generating an alarm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, and particularly relates to a fault diagnosis method for a multi-stack system of a full vanadium redox flow battery. BACKGROUND

[0002] As an excellent large-scale energy storage technology, the full vanadium redox flow battery has a broad application prospect in the fields of smart grids and renewable energy grid connection due to its advantages of independent design of power and capacity, long cycle life, high safety, etc. In practical applications, in order to meet the requirements of high power and high capacity, multiple stacks are usually connected in parallel or series to form a multi-stack system. However, this system structure also introduces new technical challenges.

[0003] At present, the fault diagnosis method for the multi-stack system of the full vanadium redox flow battery is still limited. The existing technology mainly relies on monitoring the terminal voltage of each stack and triggering a protection action when the voltage value reaches a preset fixed threshold. This method has obvious deficiencies. First, the terminal voltage of the stack is significantly affected by operating current, temperature, electrolyte flow and other working condition parameters, and changes slowly in the middle section of the state of charge platform, resulting in low sensitivity and high false alarm rate of the diagnosis method based on the fixed voltage threshold, and it is difficult to realize early warning of faults. Secondly, the traditional method cannot effectively distinguish the fault types, and it is difficult to distinguish whether the voltage anomaly is caused by the increase of the internal resistance of the stack itself or by the low actual state of charge. In addition, the existing state of charge estimation algorithm is mainly designed for single stacks, and cannot fully utilize the common operating information in the multi-stack system, and lacks collaborative modeling of the inconsistency between the stacks, resulting in insufficient estimation accuracy and robustness under complex operating conditions, and cannot provide reliable basis for accurate fault diagnosis and positioning. SUMMARY

[0004] The present application aims to provide a fault diagnosis method for a multi-stack system of a full vanadium redox flow battery, which can accurately evaluate the health status of each stack in the multi-stack system and effectively identify and warn at an early stage of fault occurrence.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: The application provides a fault diagnosis method for a multi-stack system of a vanadium redox flow battery, comprising the following steps: S1, collecting running data of the vanadium redox flow battery system in real time, wherein the running data comprises system-level global features and stack-level local features of each stack, and the collected data is subjected to standardization processing; S2, constructing an end-to-end neural network model, wherein the neural network model comprises a shared feature extraction module and a multi-task variational inference module connected in sequence; the neural network model is trained by using historical running data to optimize the parameters of the neural network model; S3, inputting the system-level global features processed in real time into the trained shared feature extraction module to output a shared feature vector; after the shared feature vector is fused with the real-time stack-level local features of each stack, the shared feature vector and the real-time stack-level local features of each stack are input into corresponding branches of the multi-task variational inference module respectively, and state of charge point estimates and estimation uncertainties of each stack are output in parallel; S4, calculating system abnormality indexes and individual abnormality indexes based on the state of charge point estimates and the estimation uncertainties of all the stacks output in step S3; when any index continuously exceeds a corresponding preset threshold value, it is determined that there is a potential fault and an alarm is generated.

[0006] In step S1, the system-level global features comprise total current, total voltage, total power, average temperature of electrolyte and speed of main circulating pump of the system; and the stack-level local features comprise end voltage, current, pressure difference between inlet and outlet and electrolyte flow of each stack.

[0007] In step S2, the shared feature extraction module is used for extracting common features of all the stacks from time series data of the system-level global features; the multi-task variational inference module is composed of a plurality of parallel variational autoencoder branches, the number of the branches is equal to the number of the stacks, each branch corresponds to one stack, and is used for outputting probability distribution of the state of charge of the stack based on the fused features.

[0008] The shared feature extraction module is a multi-layer one-dimensional convolutional neural network, comprising an input layer, a first convolution and pooling layer, a second convolution and pooling layer, and a flattening layer and a fully connected coding layer. The input layer receives a time series tensor constituted by the system-level global features of consecutive time steps. Wherein is the number of system-level features. The first convolution and pooling layer contains one-dimensional convolution kernels with a size of , adopts a ReLU activation function, and is followed by a one-dimensional maximum pooling layer with a step size of ; the first convolution and pooling layer outputs a feature map , is the length of the time dimension after convolution and pooling. The second convolution and pooling layer contains one-dimensional convolution kernels with a size of one-dimensional convolution kernel, adopts a ReLU activation function, and is followed by a one-dimensional max-pooling layer with a step size of ; the second convolution and pooling layer outputs a feature map . The flattening layer and the fully connected encoding layer flatten the feature map into a one-dimensional vector , and perform non-linear transformation and dimension reduction through one or more fully connected layers, and finally output a shared feature vector with a fixed dimension .

[0009] In step S3, the fusion is a vector splicing operation; each branch outputs an estimated value of the state of charge of the corresponding stack and an estimation uncertainty, which is used to represent the reliability of the estimated value.

[0010] In step S4, the system abnormality indicators include the state of charge difference degree between the stacks and the system average uncertainty, and the individual abnormality indicators include the state of charge deviation degree of the individual stacks.

[0011] When the state of charge difference degree between the stacks continuously exceeds a first preset threshold, a system-level consistency alarm is triggered, indicating that there is a fault causing inconsistent states of the stacks at the system level.

[0012] In step S2, when training the neural network model, the combined loss function takes into account both the accuracy of the estimated value of the state of charge and the difference between the probability distribution output by the multi-task variational inference module and the prior distribution.

[0013] Compared with the prior art, the present application has the beneficial effects of: 1. The full-vanadium redox flow battery multi-stack system fault diagnosis method provided by the embodiments of the present application combines system-level global features and stack-level local features and collects full-dimensional data, and combines standardization processing, thereby capturing both the operating conditions shared by multiple stacks, such as the total current and the average temperature of the electrolyte, and the exclusive states of a single stack, such as the terminal voltage and the inlet and outlet pressure difference, and providing data support for distinguishing between system-level faults and individual stack faults. At the same time, the shared feature extraction module mines the common rules of multiple stacks, and the independent branches of the multi-task variational inference module adapt to the individual characteristics of each stack, so that the SOC estimation accuracy of each stack is greatly improved compared with traditional single-stack algorithms.

[0014] 2、The method provided by the application breaks the limitation of traditional SOC estimation algorithm which is only designed for single stack and does not utilize the common information probability distribution of multiple stacks, realizes the collaborative estimation of multiple stack SOCs, and greatly improves the estimation accuracy compared with the traditional single stack algorithm. Meanwhile, the variable inference module can quantify the model uncertainty and data uncertainty by outputting the probability distribution of SOC estimation, solves the problem that the traditional algorithm only outputs a deterministic value and cannot evaluate the reliability of the result, and provides a key basis for subsequent fault diagnosis anti-interference. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of a full vanadium redox flow battery multiple stack system fault diagnosis method provided by the embodiment of the application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be clearly and completely described 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 a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0017] For example, referring to Figure 1 The embodiment of the application provides a full vanadium redox flow battery multiple stack system fault diagnosis method, which comprises the following steps. The embodiment of the application provides a full vanadium redox flow battery multiple stack system fault diagnosis method, which comprises the following steps. S1: Real-time acquisition of operation data of a full vanadium redox flow battery system, wherein the operation data comprises system-level global features and stack-level local features of each stack, and the acquired data is standardized.

[0018] For example, in step S1, the system-level global features comprise total current , total voltage , total power , average temperature of electrolyte , and main circulating pump speed ; and the stack-level local features comprise, for each stack, end voltage , current flowing through the stack , pressure difference between inlet and outlet , and electrolyte flow rate .

[0019] S2: constructing an end-to-end neural network model, the neural network model comprising a shared feature extraction module and a multi-task variational inference module connected in sequence; training the neural network model by using historical operation data to optimize parameters of the neural network model.

[0020] As a possible implementation, in step S2, the shared feature extraction module adopts a convolutional neural network structure, which is used to extract common features of all stacks from the time series data of the system-level global features; the multi-task variational inference module is composed of multiple parallel variational autoencoder branches, the number of branches is equal to the number of stacks, each branch corresponds to a stack, and is used to output the probability distribution of the state of charge of the stack based on the fused features.

[0021] For example, the shared feature extraction module is a multi-layer one-dimensional convolutional neural network (1D-CNN), and the specific structure includes: Input layer: receiving a time series tensor composed of continuous time steps of system-level global features , where is the number of system-level global features; First convolution and pooling layer: containing one-dimensional convolution kernels with a size of , using ReLU activation function, and followed by a one-dimensional maximum pooling layer with a step size of ; the layer outputs a feature map , where is the length of the time dimension after convolution and pooling; Second convolution and pooling layer: containing one-dimensional convolution kernels with a size of , using ReLU activation function, and followed by a one-dimensional maximum pooling layer with a step size of ; the layer outputs a feature map . Flattening layer and fully connected encoding layer: flattening the feature map into a one-dimensional vector , and performing nonlinear transformation and dimension reduction through one or more fully connected layers, finally outputting a shared feature vector with a fixed dimension of ; the mathematical expression of the process is: where, is the parameter of the fully connected layer, represents the trainable parameter set of all convolution kernel weights, biases and fully connected layer parameters of the shared feature extraction module.

[0022] As one possible implementation, when training the neural network model in step S2, the combined loss function simultaneously considers the accuracy of the charged state point estimate and the difference between the probability distribution output by the multi-task variational inference module and the prior distribution.

[0023] For example, the combined loss function It is obtained by weighted summation of the losses of each variational autoencoder branch, for the th Each branch, its loss Defined as: The weight loss uses mean squared error loss to measure the SOC prediction value of the neural network model. With real labels Differences between them: in, This is the training batch size.

[0024] KL divergence loss measures the distribution of latent variables generated by the inference network. With standard normal prior The difference between them is used as a regularization term: The total loss of the branch is the weighted sum of the above two items: in, This is a hyperparameter used to control the regularization strength.

[0025] The total global loss is all The sum of the losses of each branch: S3: Input the real-time processed system-level global features into the trained shared feature extraction module for forward propagation, outputting a shared feature vector. Then, fuse the shared feature vector with the real-time stack-level local features of each stack, and input them into the corresponding branches of the multi-task variational inference module. Each branch performs deterministic forward propagation, outputting in parallel the estimated state-of-charge (SOC) value and its uncertainty for each SOC, forming binary state information for fault diagnosis. .

[0026] For example, in step S3, the fusion is a vector concatenation operation; each branch outputs the estimated state of charge point of the corresponding fuel cell and the estimation uncertainty, and the estimation uncertainty is used to characterize the reliability of the estimated value of that point.

[0027] For example, the multi-task variational reasoning module consists of It consists of variational autoencoder branches with identical structures but independent parameters. For the number of fuel cells; for the first a variational autoencoder branch of the stack, the variational autoencoder branch performing the following operations: concatenating the shared feature vector output by the shared feature extraction module with the current stack level local feature vector of the stack to form a fusion input feature vector of the branch : the fusion feature vector passes through an inference network composed of at least one fully connected hidden layer, finally outputting the parameters of a latent space Gaussian distribution, i.e. mean and log-variance , where is the dimension of the latent space: where, is the specific parameter of the th branch inference network.

[0028] The latent variable is sampled from the distribution by the reparameterization trick, and then the state of charge (SOC) estimate is restored by a lightweight generation network: where, denotes element-wise multiplication, is the generation network parameter, and the Sigmoid function ensures that the output value is between 0 and 1; in the model inference stage, the sampling step is omitted, and is directly used to obtain the point estimate of SOC by the generation network , and the norm of or the variance obtained by the generation network propagation is used as the uncertainty measure S4: Based on the state of charge point estimates and estimated uncertainties of all stacks output by step S3, calculate the system abnormality index and individual abnormality index; when any index continuously exceeds its corresponding preset threshold, determine that there is a potential fault and generate an alarm.

[0029] For example, in step S4, the system abnormality index includes the state of charge difference degree between stacks and the system average uncertainty , and the individual abnormality index includes the state of charge deviation degree of the individual stack .

[0030] As one possible implementation, when the difference in state of charge between fuel cells continues to exceed a first preset threshold, a system-level consistency alarm is triggered, indicating that there is a fault at the system level that causes inconsistency in the state of the fuel cells.

[0031] As one possible implementation, when the deviation of the charge state of any individual stack exceeds a second preset threshold, an individual anomaly diagnosis procedure for that specific stack is triggered.

[0032] As one possible implementation, the individual anomaly diagnosis procedure includes: analyzing the internal resistance change and / or operating temperature of the individual fuel cell stack where the deviation of the individual state of charge continuously exceeds a second preset threshold, in order to diagnose whether there is an increase in internal resistance, abnormal flow, or flow channel blockage fault.

[0033] As one possible implementation, when the average uncertainty of the system continuously exceeds a third preset threshold, the observed data of the fuel cell stack is determined to be abnormal, indicating that there is a fault in the sensor or signal acquisition circuit of the fuel cell stack.

[0034] For example, the threshold for the difference in state of charge between fuel cells is set based on the 95th percentile of the system's healthy steady-state operation. That is, the first preset threshold, if Continue to exceed This triggers a system-level consistency alarm and initiates an auxiliary verification process, controlling the main pump speed to increase by N% in a stepwise manner while monitoring the readings of the main pipeline pressure sensor; if the pressure change is lower than the expected pressure threshold P... exp If the main pump's performance deteriorates, it is confirmed to be a fault. If the pressure change is normal, the agitator of the common storage tank should be started, and the differences in the flow meter readings of each branch should be monitored. If the deviation of the flow rate of any branch from the average value exceeds F... threshold This confirms a fault of uneven flow distribution in the main pipeline.

[0035] Relative deviation of individual charge state of fuel cell stack Quantizing a single fuel cell stack The formula for calculating the offset and direction relative to the system's average state is as follows: set up One-third of the threshold, i.e., the second preset threshold, is used to detect anomalies in individual fuel cells early and initiate individual anomaly diagnosis procedures: when When the battery is in use, the historical internal resistance data of the battery is retrieved, and the ratio of the current internal resistance to the average internal resistance is calculated. If the internal resistance deviates from the normal range, it is determined to be a local internal resistance change. If the internal resistance change is not significant, the operating temperature of the battery is compared with that of the adjacent battery. If the temperature is significantly higher, it is determined to be a low electrolyte flow rate or a slight blockage in the flow channel, and the opening degree K of the variable frequency valve in that branch is automatically increased.

[0036] when The offline internal short-circuit detection algorithm of the fuel cell stack is executed. If no short circuit is detected, the actual capacity is calculated by combining the ampere-hour integration method and the open-circuit voltage method to check the abnormal capacity decay.

[0037] The overall confidence level of the system state estimate by the system average uncertainty assessment model is calculated using the following formula: Introducing an uncertainty threshold That is, the third preset threshold, if when overtake If 85% of the data is within the range, the system should be prepared to switch to a backup estimation algorithm based on a simplified physical model, such as ampere-hour integration combined with open-circuit voltage correction, and the neural network model should be retrained based on historical data; if a certain fuel cell stack... estimation uncertainty Continue to exceed This indicates that the observed data of the fuel cell stack deviates significantly from the normal operating mode learned by the neural network model. This situation suggests significant sensor drift / failure, poor contact in the signal acquisition circuit, or sudden damage inside the fuel cell stack.

[0038] This application provides a fault diagnosis method for a multi-stack vanadium redox flow battery system. Through comprehensive data acquisition combining system-level global features and stack-level local features, along with standardized processing, it captures shared operating conditions across multiple stacks, such as total current and average electrolyte temperature, while also accurately characterizing the unique states of each individual stack, such as terminal voltage and inlet / outlet voltage difference. This provides data support for distinguishing between system-level faults and individual stack faults. Simultaneously, the shared feature extraction module uncovers common patterns across multiple stacks, and the independent branches of the multi-task variational inference module adapt to the individual characteristics of each stack, significantly improving the SOC estimation accuracy of each stack compared to traditional single-stack algorithms.

[0039] The method provided in this application, by quantifying the uncertainty of SOC estimation and combining the monitoring of inter-stack differences and individual deviations, can detect early anomalies caused by inconsistencies, such as slight SOC deviations and increased estimation uncertainty, before the stack voltage reaches the protection threshold. For example, when the SOC deviation of an individual stack continuously exceeds a second preset threshold, potential faults such as slight flow channel blockage or increased internal resistance can be identified in advance through internal resistance and temperature analysis. The warning time is more than one hour earlier than traditional methods, allowing sufficient time for maintenance personnel to handle the situation, preventing local faults from spreading into system-wide cascading damage, and significantly improving the operational safety of large-scale energy storage systems.

[0040] 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.

[0041] 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 fault diagnosis method for a multi-stack system of vanadium redox flow batteries, characterized in that, include: S1: Real-time acquisition of operational data from the vanadium redox flow battery system, including system-level global features and stack-level local features for each stack, and standardization processing of the acquired data; S2: Construction of an end-to-end neural network model, including a sequentially connected shared feature extraction module and a multi-task variational inference module; training the neural network model using historical operational data to optimize the neural network model parameters; S3: Inputting the real-time processed system-level global features into the trained shared feature extraction module, outputting a shared feature vector; fusing the shared feature vector with the real-time stack-level local features of each stack, and inputting them into the corresponding branches of the multi-task variational inference module, outputting the estimated state of charge (SOC) value and its estimation uncertainty for each stack in parallel; S4: Based on the estimated state of charge points of all fuel cells output in step S3 and the estimated uncertainty, calculate system anomaly and individual anomaly indicators; when any indicator continuously exceeds its corresponding preset threshold, determine that there is a potential fault and generate an alarm.

2. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 1, characterized in that, In step S1, the system-level global features include the system's total current, total voltage, total power, average electrolyte temperature, and main circulation pump speed; the stack-level local features for each stack include the stack's terminal voltage, flowing current, inlet and outlet pressure difference, and electrolyte flow rate.

3. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 1, characterized in that, In step S2, the shared feature extraction module is used to extract common features of all electric piles from the time series data of system-level global features; the multi-task variational inference module consists of multiple parallel variational autoencoder branches, the number of branches is equal to the number of electric piles, each branch corresponds to one electric pile, and is used to output the probability distribution of the charge state of the electric pile based on the fused features.

4. The fault diagnosis method for a multi-stack system of vanadium redox flow battery according to claim 3, characterized in that, The shared feature extraction module is a multi-layer one-dimensional convolutional neural network, including an input layer, a first convolutional and pooling layer, a second convolutional and pooling layer, a flattening layer, and a fully connected encoding layer; The input layer receives continuous The time series tensor composed of system-level global features at each time step ,in The number of system-level features; The first convolutional and pooling layer contains The size is A one-dimensional convolution kernel with ReLU activation function, followed by a stride of... A one-dimensional max pooling layer; the first convolutional and pooling layer outputs a feature map. , The length of the time dimension after convolution and pooling; The second convolutional and pooling layer contains The size is A one-dimensional convolution kernel with ReLU activation function, followed by a stride of... One-dimensional max pooling layer; second convolutional and pooling layer output feature map ; The flattening layer and the fully connected coding layer combine the feature maps Flattened into a one-dimensional vector It performs nonlinear transformation and dimensionality reduction through one or more fully connected layers, ultimately outputting a variable with a fixed dimension. Shared feature vectors .

5. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 1, characterized in that, In step S2, when training the neural network model, the combined loss function used simultaneously considers the accuracy of the estimated state of charge and the difference between the probability distribution output by the multi-task variational inference module and the prior distribution.

6. The fault diagnosis method for a multi-stack system of vanadium redox flow battery according to claim 3, characterized in that, In step S3, the fusion is a vector concatenation operation; each branch outputs the estimated state of charge point of the corresponding fuel cell and the estimation uncertainty, the estimation uncertainty being used to characterize the reliability of the estimated value at that point.

7. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 1, characterized in that, In step S4, the system anomaly indicators include the difference in state of charge between fuel cells and the system average uncertainty, and the individual anomaly indicators include the deviation of the state of charge of individual fuel cells.

8. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 7, characterized in that, When the difference in state of charge between the fuel cells continues to exceed a first preset threshold, a system-level consistency alarm is triggered, indicating that there is a fault at the system level that causes inconsistency in the state of the fuel cells.

9. The fault diagnosis method for a multi-stack vanadium redox flow battery system according to claim 7, wherein when the deviation of the state of charge of any individual stack exceeds a second preset threshold, an individual abnormality diagnosis procedure for that specific stack is triggered.

Images