Flow battery state-of-charge estimation method based on TCN-BiLSTM neural network and EIS data

Abstract

The invention discloses a flow battery state-of-charge estimation method based on a TCN-BiLSTM neural network and EIS data, and belongs to the technical field of flow battery management. According to the method, temperature, flow velocity, frequency, real part impedance and imaginary part impedance are extracted as input characteristics by measuring EIS data of the flow battery under different SOCs, and a hybrid model fusing a time convolutional network (TCN) and a bidirectional long-short-term memory network (BiLSTM) is constructed. The TCN captures a local time sequence dependency relationship through multi-layer expansion convolution and residual connection, the BiLSTM learns global sequence features by using a bidirectional gating mechanism, and finally, the global sequence features are mapped into SOC values through a full connection layer. According to the model, Min-Max normalization preprocessing data is adopted, a mean square error is used as a loss function and is combined with an Adam optimizer for training, and the estimation precision is evaluated through MAE, RMSE and R2 indexes. According to the method, the TCN-BiLSTM neural network model is trained based on the EIS data, the defects of a traditional SOC estimation method under the dynamic operation condition can be overcome, and the method has the advantages of being high in precision, high in robustness and wide in applicability.

Description

Technical Field

[0001] The present invention belongs to the technical field of liquid flow battery management, and in particular relates to a liquid flow battery state of charge estimation method based on a TCN-BiLSTM neural network and EIS data. Background Art

[0002] With the rapid development of new energy sources and the construction of new power systems, flow batteries, as a new energy storage technology, can independently adjust battery energy and power density through external storage tanks. They have the characteristics of low cost, high safety, and environmental friendliness. They are considered to be one of the most promising medium and large-scale energy storage technologies. In the flow battery management system, the state of charge (SOC) is the core parameter for evaluating the remaining battery power, which directly affects the safe operation and cycle life of the battery. SOC is defined as the ratio of the current available battery capacity to the rated maximum capacity, expressed as:

[0003]

[0004] Where Q available is the current available battery capacity, Q rated is the rated maximum capacity.

[0005] Traditional SOC estimation methods include coulomb counting, open circuit voltage and Kalman filtering. The accuracy and robustness of traditional estimation methods are limited under dynamic operating conditions of the battery, and they are unable to quickly and accurately estimate the battery SOC. Electrochemical impedance spectroscopy (EIS) is an important technology that studies the characteristics of an electrochemical system by measuring its impedance response under the excitation of AC signals of different frequencies. Due to its non-invasiveness and high sensitivity, it has become an important tool for characterizing the internal state of the battery and can be used as a measurement method for SOC estimation. Since EIS has high requirements for system stability and is very sensitive to noise, there is a large error in estimating SOC using only EIS. Neural network (NN) is a model of machine learning. Inspired by the biological nervous system, it simulates the structure and function of neurons in the human brain. Through machine learning using the obtained data, the neural network is taught to discover the complex relationship between different battery indicators and SOC, and can provide reliable SOC results even in the face of noise and uncertainty.

[0006] Therefore, in view of the shortcomings of the existing flow battery SOC estimation technology, there is an urgent need for a flow battery state of charge estimation method based on TCN-BiLSTM neural network and EIS data to achieve accurate estimation of the flow battery SOC. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a flow battery SOC estimation method based on a fusion neural network and EIS data using a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM). By using EIS testing technology to measure the flow battery at different SOCs, and relying on the measured data to train a TCN-BiLSTM neural network model, this method can achieve rapid and accurate estimation of the flow battery SOC under different operating conditions. This method can overcome the shortcomings of traditional SOC estimation methods under dynamic operating conditions and improve the accuracy and robustness of SOC estimation.

[0008] The specific technical solutions adopted in the present invention are as follows:

[0009] The present invention provides a flow battery state of charge estimation method based on TCN-BiLSTM neural network and EIS data, which is as follows:

[0010] S1. Measure the electrochemical impedance spectrum of the flow battery at different states of charge to obtain raw data including temperature, flow rate, frequency, real impedance and imaginary impedance;

[0011] S2, performing normalization preprocessing on the raw data, and then extracting characteristic parameters;

[0012] S3. Using the results obtained in S2 as input features of the neural network model, fusing the output features of the temporal convolutional network and the bidirectional long short-term memory network, and mapping the output features to the state of charge value through a fully connected layer to construct a TCN-BiLSTM neural network model;

[0013] S4. Using the training data set to train the TCN-BiLSTM neural network model, adjusting the model parameters through the optimization algorithm to minimize the prediction error;

[0014] S5. Use mean absolute error, root mean square error, and coefficient of determination to evaluate the training accuracy of the model obtained in S4, and obtain a TCN-BiLSTM neural network model that meets the evaluation standards.

[0015] S6. Use the obtained TCN-BiLSTM neural network model to estimate and evaluate the state of charge of the flow battery.

[0016] Preferably, the S1 step is as follows:

[0017] S11. A sinusoidal voltage signal with an amplitude of 10 mV was selected as the excitation source, and the sweep frequency range was from 0.1 Hz to 10 kHz to cover the time scales corresponding to typical processes in the flow battery, including charge transfer, electrode reaction, and material diffusion.

[0018] S12, under the set temperature and flow rate working conditions, using a constant current charge and discharge mode to increase the state of charge of the flow battery to a first measurement value;

[0019] S13. After the charge and discharge are completed, the peristaltic pump continues to circulate the electrolyte for 1 minute to achieve uniform mixing of the electrolytes, and the flow battery is allowed to stand to enter a steady state. After the flow battery stabilizes, an electrochemical impedance spectroscopy measurement is performed at the current state of charge level to obtain corresponding impedance response spectrum data.

[0020] S14, then gradually increasing the state of charge of the flow battery to other required measurement values ​​according to the same constant current charge and discharge mode as S12, repeating the process of S13 each time until the impedance response spectrum data of the complete state of charge range is obtained;

[0021] S15. Change the temperature and flow rate working conditions, and repeat S12 to S14 until all the original data under the required conditions are obtained.

[0022] Preferably, in S2, the normalization preprocessing adopts the Min-Max normalization method, which is expressed as:

[0023]

[0024] Among them, X norm is the normalized data value, X is the original data value, and X min is the minimum data value, X max is the maximum data value.

[0025] Preferably, in S3, the method for fusing the output features of the temporal convolutional network and the bidirectional long short-term memory network is as follows:

[0026] S31. Construct the temporal convolutional network layer as follows:

[0027] The temporal convolutional network is composed of three stacked temporal convolutional blocks. Each temporal block contains a causal convolutional layer, a dilated convolutional layer, a residual connection, a regularization function, and an activation function to capture local temporal dependencies. The number of input channels of each temporal convolutional block is 64, the convolution kernel size is 3, and the dilation rates are 1, 2, and 4 respectively.

[0028] S32. Based on the results obtained in S31, a bidirectional long short-term memory network layer is constructed as follows:

[0029] The bidirectional long short-term memory network captures the bidirectional dependency of sequence data by combining forward and backward long short-term memory networks. It consists of a forget gate, an input gate, an update gate, and an output gate. Each gated unit contains a sigmoid activation function and a dot product operation. The specific input size is 64, the hidden layer size is 32, the number of layers is 2, a bidirectional structure is adopted, and the dropout probability is 0.4.

[0030] Preferably, the S4 is as follows:

[0031] In model training, mean square error is used as the loss function, and the Adam optimizer is used for optimization. The early stopping mechanism and Dropout technology are used to prevent overfitting and improve the generalization ability of the model. The specific number of training rounds is 100, the batch size is 32, and the learning rate is 0.001.

[0032] As an advantage, in said S5, the mean absolute error MAE, the root mean square error RMSE and the coefficient of determination R 2 The indicators are expressed as:

[0033]

[0034] Among them, SOC real is the actual value of the state of charge, SOC i is the estimated value of state of charge, N is the total number of samples, is the average state of charge.

[0035] As a preference, in said S5, when the mean absolute error MAE is less than 2%, the root mean square error RMSE is less than 2% and the coefficient of determination R 2 When both S and S are greater than 0.9, the evaluation considers that the model training accuracy meets the standard; otherwise, adjust the model hyperparameters and optimize the model structure according to S4 and re-optimize until the evaluation meets the standard.

[0036] Preferably, in S6, the model evaluates the estimated predicted value of the state of charge of the flow battery using mean absolute error and root mean square error indicators.

[0037] Preferably, the flow battery includes an all-vanadium flow battery, a zinc-iron flow battery, a zinc-bromine flow battery, an organic flow battery and a lead-acid flow battery.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1) The present invention combines the advantages of temporal convolutional networks (TCN) and bidirectional long short-term memory networks (BiLSTM) to construct a TCN-BiLSTM network model. TCN can effectively capture local temporal dependencies in EIS data, while BiLSTM can capture bidirectional dependencies in sequence data. By fusing the output features of the two, the present invention can more accurately learn the complex relationship between the internal state of the battery and SOC, thereby improving the accuracy and robustness of SOC estimation. Especially under dynamic operating conditions of the battery, the present invention can overcome the shortcomings of traditional SOC estimation methods and provide more reliable SOC estimation results.

[0040] 2) This method leverages the rich information in the EIS data by measuring the EIS of a flow battery at different SOCs and extracting characteristic parameters such as temperature, flow rate, frequency, and real and imaginary impedance as inputs to a neural network. This approach not only improves the accuracy of SOC estimation but also enables a more comprehensive understanding of the battery's internal state, providing stronger support for battery management and maintenance.

[0041] 3) The method of the present invention is applicable to different types of flow battery systems. By adjusting the parameters of the TCN-BiLSTM neural network model, it can be easily adapted to different battery systems, showing wide applicability and scalability. This versatility makes the present invention more versatile and valuable for promotion in energy storage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A framework diagram of a flow battery state of charge estimation method based on a TCN-BiLSTM neural network and EIS data provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to more clearly illustrate the objectives, technical paths and advantages of the present invention, the following is a detailed description with reference to the accompanying drawings and specific embodiments. It should be understood that the description of the embodiments herein only covers some application scenarios of the present invention, rather than all embodiments. Therefore, without departing from the spirit and scope of the present invention, other embodiments derived by those skilled in the art without performing creative work all fall within the scope of protection of the present invention. The following examples will be used to further illustrate the method and effects of the present invention.

[0044] The present invention provides a flow battery state of charge estimation method based on a TCN-BiLSTM neural network and EIS data. The method is as follows:

[0045] S1. Measure EIS experimental data: Measure the electrochemical impedance spectrum of the flow battery at different states of charge to obtain raw data including temperature, flow rate, frequency, real impedance and imaginary impedance.

[0046] As a preferred embodiment of the present invention, the steps are as follows:

[0047] S11. A sinusoidal voltage signal with an amplitude of 10 mV was selected as the excitation source, and the sweep frequency range was 0.1 Hz to 10 kHz to cover the time scales corresponding to typical processes in the flow battery, including charge transfer, electrode reaction, and material diffusion.

[0048] S12, under the set temperature and flow rate working conditions, using a constant current charge and discharge mode to increase the state of charge of the flow battery to a first measurement value;

[0049] S13. After the charge and discharge are completed, the peristaltic pump continues to circulate the electrolyte for 1 minute to achieve uniform mixing of the electrolytes, and the flow battery is allowed to stand to enter a steady state. After the flow battery stabilizes, an electrochemical impedance spectroscopy measurement is performed at the current state of charge level to obtain corresponding impedance response spectrum data.

[0050] S14, then gradually increase the state of charge of the flow battery to other required measurement values ​​according to the same constant current charge and discharge mode as S12, and repeat the process of S13 (including electrolyte circulation, static stabilization and EIS test process) each time until the impedance response spectrum data of the complete state of charge range is obtained;

[0051] S15. Change the temperature and flow rate working conditions, and repeat S12 to S14 until all the original data under the required conditions are obtained.

[0052] S2. Data preprocessing and feature extraction: The raw data obtained in S1 is normalized and preprocessed, and then characteristic parameters including temperature, flow rate, frequency, real impedance and imaginary impedance are extracted.

[0053] As a preferred embodiment of the present invention, in this step, the normalization preprocessing adopts the Min-Max normalization method, which is expressed as:

[0054]

[0055] Among them, X norm is the normalized data value, X is the original data value, and X min is the minimum data value, X max is the maximum data value.

[0056] S3. Build a TCN-BiLSTM neural network model: Use the results obtained in S2 as the input features of the neural network model, fuse the output features of the temporal convolutional network (TCN) and the bidirectional long short-term memory network (BiLSTM), combining the advantages of both. Then, use a fully connected layer to map the above output features to the state of charge value to build the TCN-BiLSTM neural network model.

[0057] As a preferred embodiment of the present invention, in this step, the input features of the neural network include five characteristic parameters: temperature, flow rate, frequency, real impedance and imaginary impedance.

[0058] As a preferred embodiment of the present invention, in this step, the method of fusing the output features of the temporal convolutional network and the bidirectional long short-term memory network is as follows:

[0059] S31. Construct the temporal convolutional network layer as follows:

[0060] The temporal convolutional network is composed of three stacked temporal convolutional blocks. Each temporal block contains a causal convolutional layer, a dilated convolutional layer, a residual connection, a regularization function, and an activation function to capture local temporal dependencies. The number of input channels of each temporal convolutional block is 64, the convolution kernel size is 3, and the dilation rates are 1, 2, and 4 respectively.

[0061] Of course, in actual use, the number of temporal convolution blocks and the above-mentioned specific parameter values ​​can be adjusted as needed.

[0062] S32. Based on the results obtained in S31, a bidirectional long short-term memory network layer is constructed as follows:

[0063] The bidirectional long short-term memory network captures the bidirectional dependencies of sequence data by combining two long short-term memory networks (LSTMs) with forward and backward passes. It consists of a forget gate, an input gate, an update gate, and an output gate. Each gated unit contains a sigmoid activation function and a dot product operation. The specific input size is 64, the hidden layer size is 32, the number of layers is 2, a bidirectional structure is used, and the dropout probability is 0.4.

[0064] Of course, the above specific parameter values ​​can be adjusted as needed during actual use.

[0065] S4, model training and optimization: Use the training data set to train the TCN-BiLSTM neural network model obtained in S3, adjust the model parameters through the optimization algorithm to minimize the prediction error;

[0066] As a preferred embodiment of the present invention, the steps are as follows:

[0067] In model training, the mean square error (MSE) is used as the loss function, and the Adam optimizer is used for optimization. The early stopping mechanism and Dropout technology are used to prevent overfitting and improve the generalization ability of the model. The specific number of training rounds is 100, the batch size is 32, and the learning rate is 0.001.

[0068] Of course, in actual use, the number of temporal convolution blocks and the above-mentioned specific parameter values ​​can be adjusted as needed.

[0069] S5. Model accuracy evaluation: The training accuracy of the model obtained in S4 is evaluated using mean absolute error, root mean square error, and coefficient of determination indicators, and a TCN-BiLSTM neural network model that meets the evaluation standards is obtained.

[0070] As a preferred embodiment of the present invention, in this step, the mean absolute error MAE, the root mean square error RMSE and the determination coefficient R 2 The indicators are expressed as:

[0071]

[0072] Among them, SOC real is the actual value of the state of charge, SOC i is the estimated value of state of charge, N is the total number of samples, is the average state of charge.

[0073] As a preferred embodiment of the present invention, in this step, when the mean absolute error MAE is less than 2%, the root mean square error RMSE is less than 2% and the determination coefficient R 2 >0.9 When all three are met at the same time, the evaluation considers that the model training accuracy meets the standard; otherwise, adjust the model hyperparameters and optimize the model structure according to S4 and re-optimize until the evaluation meets the standard.

[0074] S6. SOC estimation and evaluation: The obtained TCN-BiLSTM neural network model is used to estimate and evaluate the state of charge of the flow battery.

[0075] As a preferred embodiment of the present invention, in this step, the model evaluates the estimated predicted value of the flow battery state of charge using the mean absolute error (MAE) and root mean square error (RMSE) indicators.

[0076] As a preferred embodiment of the present invention, the method of the present invention is applicable to a variety of liquid flow battery systems, including all-vanadium liquid flow batteries, zinc-iron liquid flow batteries, zinc-bromine liquid flow batteries, organic liquid flow batteries and lead-acid liquid flow batteries.

[0077] The method and effects of the present invention will be further illustrated below by way of examples.

[0078] Example

[0079] This embodiment provides a method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data. The method block diagram is as follows: Figure 1 As shown in Figure 2, the estimation method mainly includes the following steps:

[0080] Step S1, measuring EIS experimental data: measuring the EIS of the flow battery at different SOCs to obtain raw data including temperature, flow rate, frequency, real impedance and imaginary impedance.

[0081] In this embodiment, the negative electrode electrolyte used is 80ml 0.1mol / L and 3 mol / LOH - Mixed solution, positive electrode consists of 80ml 0.2mol / L and 3 mol / LOH - A zinc-based flow battery with a single electrode effective reaction area of ​​5cm*5cm composed of a mixed solution was used as the experimental object.

[0082] Under the set temperature and flow rate conditions, the constant current charging mode is used to increase the battery SOC to 20%. After charging and discharging are completed, the peristaltic pump continues to circulate the electrolyte for 1 minute to achieve uniform mixing of the electrolyte and allow the system to enter a steady state. After the system stabilizes, EIS measurements are performed at the current SOC level to obtain the corresponding impedance response spectrum. Subsequently, the SOC is gradually increased to 30%, 40%, 50%, 60%, 70% and 80% according to the same constant current strategy. For each level increase, the electrolyte circulation, static stabilization and EIS test process are repeated until the impedance data of the complete charging range is obtained. Finally, the temperature and flow rate conditions are changed. The temperature is set to 20℃, 25℃, 30℃, 35℃ and 40℃, and the flow rate is set to 60mL / min, 70mL / min, 80mL / min, 90mL / min, and 100mL / min, and the above steps are repeated.

[0083] Step S2: Data preprocessing and feature extraction: The collected EIS data is normalized and preprocessed to extract characteristic parameters such as temperature, flow rate, frequency, real impedance, and imaginary impedance, which are used as input features of the neural network.

[0084] In this embodiment, temperature, flow rate, frequency, real impedance and imaginary impedance characteristic parameters are extracted as input features, and each feature is normalized to the interval [0.1, 0.9] by Min-Max, so as to avoid the numerical range of some features being too large or too small and the model output boundary values ​​(such as 0 or 1), thereby improving the stability and convergence speed of model training.

[0085] Step S3: Construct a TCN-BiLSTM neural network model: Combining the advantages of the temporal convolutional network (TCN) and the bidirectional long short-term memory network (BiLSTM), construct a TCN-BiLSTM neural network model, fuse the output features of TCN and BiLSTM, and map the features to SOC values ​​through a fully connected layer.

[0086] In this embodiment, the neural network model adopts a TCN-BiLSTM hybrid architecture. First, multi-scale temporal features are extracted through a three-layer temporal convolutional network (TCN): each TCN layer contains two dilated convolutions (the convolution kernel size is 3, and the dilation rates are 1, 2, and 4, respectively, to exponentially expand the receptive field), with batch normalization and ReLU activation, and residual connections (1x1 convolution is used to adjust the dimension when the input and output channels are inconsistent) to ensure gradient stability. The output channels of all TCN layers are unified to 64; then, the 64-channel temporal features output by the TCN are input into a bidirectional LSTM module. The module adopts a two-layer structure (32 hidden units per layer). The bidirectional design enables it to simultaneously capture forward and backward temporal dependencies, and a 40% inter-layer dropout is set to prevent overfitting. Finally, the bidirectional hidden state (dimension 64) at the end of the sequence is extracted; finally, feature mapping is implemented through a regression head containing two fully connected layers (64→64→1). The first layer is equipped with ReLU activation and 40% dropout, and the second layer directly outputs the SOC prediction value.

[0087] Step S4, model training and optimization: Use the training data set to train the TCN-BiLSTM neural network model, adjust the model parameters through the optimization algorithm, and minimize the prediction error. The mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 ) indicators to evaluate the model training accuracy.

[0088] In this embodiment, mean square error (MSE) is used as the loss function in model training, and Adam optimizer is used for optimization. Early stopping mechanism and Dropout technology are used to prevent overfitting and improve the generalization ability of the model. The specific number of training rounds is 100, the batch size is 32, and the learning rate is 0.001. After training, the loss function Loss is less than 0.0015, MAE=0.42%, RMSE=0.61%, R 2 =0.999, the model training accuracy meets the standard, and the model is saved for SOC estimation.

[0089] Step S5, SOC estimation and evaluation: Use the trained model to estimate the SOC of the flow battery and evaluate the estimation accuracy of the SOC.

[0090] As a preferred embodiment of the present invention, this step estimates the battery SOC under three operating conditions: 20°C at 80 ml / min, 30°C at 100 ml / min, and 40°C at 60 ml / min. The results are shown in Tables 1-3. The estimated MAE and RMSE under all three conditions are less than 1%, demonstrating that the model exhibits high accuracy and robustness under complex operating conditions such as varying temperatures and flow rates, significantly outperforming traditional models.

[0091] Table 1 SOC estimation and index analysis of TCN-BiLSTM model under 20℃ and 80ml / min conditions

[0092] SOC true value SOC estimation 20％ 20.30％ 30％ 29.76％ 40％ 40.25％ 50％ 50.44％ 60％ 60.43％ 70％ 70.94％ 80％ 80.07％ MAE 0.38％ RMSE 0.46％

[0093] Table 2 SOC estimation and index analysis of TCN-BiLSTM model under 30℃ and 100ml / min conditions

[0094] SOC true value SOC estimation 20％ 20.34％ 30％ 29.38％ 40％ 40.14％ 50％ 50.34％ 60％ 60.51％ 70％ 70.80％ 80％ 80.43％ MAE 0.45％ RMSE 0.50％

[0095] Table 3 SOC estimation and index analysis of TCN-BiLSTM model under 40℃ and 60ml / min conditions

[0096] SOC true value SOC estimation 20％ 20.45％ 30％ 29.71％ 40％ 40.01％ 50％ 50.45％ 60％ 60.48％ 70％ 71.12％ 80％ 79.90％ MAE 0.41％ RMSE 0.53％

[0097] This paper combines the advantages of a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM) to construct a TCN-BiLSTM network model. Furthermore, by combining EIS data, characteristic parameters such as temperature, flow rate, frequency, real impedance, and imaginary impedance are extracted as neural network inputs. The TCN-BiLSTM neural network model is trained to achieve rapid and accurate estimation of the SOC of the flow battery under different operating conditions. This method can overcome the shortcomings of traditional SOC estimation methods under dynamic operating conditions, improve the accuracy and robustness of SOC estimation, and is suitable for flow battery systems under different operating conditions.

[0098] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data, characterized in that: The details are as follows: S1. Measure the electrochemical impedance spectrum of the flow battery at different states of charge to obtain raw data including temperature, flow rate, frequency, real impedance and imaginary impedance; S2, performing normalization preprocessing on the raw data, and then extracting characteristic parameters; S3. Using the results obtained in S2 as input features of the neural network model, fusing the output features of the temporal convolutional network and the bidirectional long short-term memory network, and mapping the output features to the state of charge value through a fully connected layer to construct a TCN-BiLSTM neural network model; S4. Using the training data set to train the TCN-BiLSTM neural network model, adjusting the model parameters through the optimization algorithm to minimize the prediction error; S5. Use mean absolute error, root mean square error, and coefficient of determination to evaluate the training accuracy of the model obtained in S4, and obtain a TCN-BiLSTM neural network model that meets the evaluation standards. S6. Use the obtained TCN-BiLSTM neural network model to estimate and evaluate the state of charge of the flow battery.

2. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: The S1 step is specifically as follows: S11. A sinusoidal voltage signal with an amplitude of 10 mV was selected as the excitation source, and the sweep frequency range was from 0.1 Hz to 10 kHz to cover the time scales corresponding to typical processes in the flow battery, including charge transfer, electrode reaction, and material diffusion. S12, under the set temperature and flow rate working conditions, using a constant current charge and discharge mode to increase the state of charge of the flow battery to a first measurement value; S13. After the charge and discharge are completed, the peristaltic pump continues to circulate the electrolyte for 1 minute to achieve uniform mixing of the electrolytes, and the flow battery is allowed to stand to enter a steady state. After the flow battery stabilizes, an electrochemical impedance spectroscopy measurement is performed at the current state of charge level to obtain corresponding impedance response spectrum data. S14, then gradually increasing the state of charge of the flow battery to other required measurement values ​​according to the same constant current charge and discharge mode as S12, repeating the process of S13 each time until the impedance response spectrum data of the complete state of charge range is obtained; S15. Change the temperature and flow rate working conditions, and repeat S12 to S14 until all the original data under the required conditions are obtained.

3. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: In S2, the normalization preprocessing adopts the Min-Max normalization method, which is expressed as: Among them, X norm is the normalized data value, X is the original data value, and X min is the minimum data value, X max is the maximum data value.

4. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: In S3, the method for fusing the output features of the temporal convolutional network and the bidirectional long short-term memory network is as follows: S31. Construct the temporal convolutional network layer as follows: The temporal convolutional network is composed of three stacked temporal convolutional blocks. Each temporal block contains a causal convolutional layer, a dilated convolutional layer, a residual connection, a regularization function, and an activation function to capture local temporal dependencies. The number of input channels of each temporal convolutional block is 64, the convolution kernel size is 3, and the dilation rates are 1, 2, and 4 respectively. S32. Based on the results obtained in S31, a bidirectional long short-term memory network layer is constructed as follows: The bidirectional long short-term memory network captures the bidirectional dependency of sequence data by combining forward and backward long short-term memory networks. It consists of a forget gate, an input gate, an update gate, and an output gate. Each gated unit contains a sigmoid activation function and a dot product operation. The specific input size is 64, the hidden layer size is 32, the number of layers is 2, a bidirectional structure is adopted, and the dropout probability is 0.

4.

5. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: The S4 is specifically as follows: In model training, mean square error is used as the loss function, and the Adam optimizer is used for optimization. The early stopping mechanism and Dropout technology are used to prevent overfitting and improve the generalization ability of the model. The specific number of training rounds is 100, the batch size is 32, and the learning rate is 0.

001.

6. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: In S5, the mean absolute error MAE, root mean square error RMSE and determination coefficient R 2 The indicators are expressed as: Among them, SOC real is the actual value of the state of charge, SOC i is the estimated value of state of charge, N is the total number of samples, is the average state of charge.

7. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: In S5, when the mean absolute error MAE is less than 2%, the root mean square error RMSE is less than 2% and the coefficient of determination R 2 When both S and S are greater than 0.9, the evaluation considers that the model training accuracy meets the standard; otherwise, adjust the model hyperparameters and optimize the model structure according to S4 and re-optimize until the evaluation meets the standard.

8. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: In S6, the model evaluates the estimated predicted value of the flow battery state of charge using mean absolute error and root mean square error indicators.

9. The method for estimating the state of charge of a flow battery based on a TCN-BiLSTM neural network and EIS data according to claim 1, wherein: The flow batteries include all-vanadium flow batteries, zinc-iron flow batteries, zinc-bromine flow batteries, organic flow batteries and lead-acid flow batteries.

Images