Quasi-solid-state battery SOH-SOC joint estimation method for energy storage scenarios
Through the bidirectional jump TCN-BiGRU network combined with the SOH-SOC joint estimation model of the one-dimensional convolution module, the accuracy and computational complexity of SOH-SOC estimation of quasi-solid-state batteries are solved, and efficient and accurate joint estimation is achieved.
Patent Information
- Application Number
- CN202510473000.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2025-07-18
- Estimated Expiration
- 2045-04-16
AI Technical Summary
The prior art is difficult to accurately estimate the health status (SOH) and state of charge (SOC) of quasi-solid-state batteries, especially when considering their aging mechanisms that nonlinear responses and multiple coupling factors, resulting in increased computational complexity and reduced estimation accuracy.
The two-way jump TCN-BiGRU network is used to combine the one-dimensional convolution module and the SOH-SOC joint estimation model of the fully connected layer. By charging the encoder, SOH encoder and SOC decoder, data redundant input is reduced, local and global characteristics of the battery are captured, and joint estimation of SOH-SOC is realized.
It improves the accuracy and stability of SOH-SOC estimation of quasi-solid-state batteries, reduces the computational complexity, has high generalization capabilities and fast processing speed.
Smart Images
Figure CN119986407B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of joint estimation of the state of health (SOH) and state of charge (SOC) of quasi-solid-state batteries, and particularly to a method for jointly estimating SOH-SOC of quasi-solid-state batteries for energy storage scenarios. Background Art
[0002] With the rapid development of new energy vehicles and energy storage systems, quasi-solid-state batteries have become an important development direction for next-generation power batteries due to their high energy density, excellent safety performance, and low risk of electrolyte leakage. Accurately estimating the state of health (SOH) and state of charge (SOC) of quasi-solid-state batteries is a core challenge for new-generation technologies.
[0003] As key indicators characterizing the battery state, SOH and SOC directly affect the charge and discharge control, life prediction, and system safety of the battery. Traditional SOC estimation methods rely on the linear assumption of the battery model, but the solid-liquid mixed electrolyte characteristics of quasi-solid-state batteries lead to more significant polarization effects and interfacial impedance changes, making it difficult for traditional models to accurately capture the non-linear response under dynamic operating conditions. In addition, existing SOH estimations are mostly based on single-factor analysis of capacity or internal resistance, while the aging mechanism of quasi-solid-state batteries involves multiple coupled factors such as electrode-electrolyte interface degradation and solid electrolyte cracking, and single-parameter models are difficult to comprehensively reflect the capacity decay law. Moreover, the SOH estimation of quasi-solid-state batteries during constant current and constant voltage charging only depends on specific charging data within a predetermined range, while SOC estimation needs to be evaluated throughout the charging process. Previous studies used the estimation results of the SOH model as the input of the SOC model and adopted two independent models for estimation. However, these methods ignore the potential inherent correlation in the SOH and SOC charging data, resulting in increased computational complexity and reduced estimation accuracy. In addition, current convolutional models and deep network models ignore the internal connection between local features and global features, hindering the ability to capture time-series aging information. Therefore, there is an urgent need for a SOH-SOC joint estimation method that takes into account both model accuracy and computational efficiency and adapts to the characteristics of quasi-solid-state batteries to break through the bottleneck of existing technologies. Summary of the Invention
[0004] Aiming at the problems in the prior art, the present invention provides a method for jointly estimating SOH-SOC of quasi-solid-state batteries for energy storage scenarios, aiming to reduce redundant data input and improve the estimation efficiency of the battery state of health and the battery state of charge.
[0005] The method for jointly estimating SOH-SOC of quasi-solid-state batteries for energy storage scenarios includes the following steps:
[0006] Step 1: Obtain a quasi-solid-state battery data set, which includes all voltage data sets and specific voltage data sets;
[0007] Step 2: Establish an SOH-SOC joint estimation model, where the SOH-SOC joint estimation model includes a charging encoder, an SOH encoder, an SOH decoder, and an SOC decoder;
[0008] Step 3: Feed the specific voltage dataset into the charging encoder and obtain the charging encoded data;
[0009] Step 4: Feed the charging encoded data into the SOH encoder and obtain the specific SOH decoded data;
[0010] Step 5: Feed the specific SOH decoded data and the entire voltage dataset into the SOH encoder to obtain the entire SOH encoded data;
[0011] Step 6: Feed the entire SOH encoded data and the charging encoded data into the SOC decoder to obtain the SOC decoded data;
[0012] Step 7: Use the specific SOH decoded data and the SOC decoded data as the estimated values of the battery health state and the state of charge of the battery, respectively.
[0013] Furthermore: The quasi-solid-state battery dataset is a self-built dataset and includes the capacity and charge changes during the charging process of the entire life cycle of the quasi-solid-state battery; this dataset is the data collected by the Arbin BT2000 battery experimental system during continuous charging experiments, and all experimental quasi-solid-state batteries follow the same charging curve; the experiments are carried out in a controlled ambient temperature of 25°C, and the battery charging is carried out using a constant current-constant voltage charging protocol; during the constant current charging process, data is collected every 60s; four batteries are selected from the self-built quasi-solid-state battery dataset, named L11, L12, L13, and L14 respectively. These four batteries all use lithium iron phosphate as the positive electrode material, with a nominal capacity of 2AH and a nominal voltage of 4.2V; the dataset contains comprehensive information about the aging process, and the comprehensive information includes voltage, current, and charging time; L11, L12, and L13 are selected as the training dataset, while L14 is selected as the test dataset. During a complete constant current-constant voltage charging process, the initial charging currents are 0.25C (0.5A) and 2C (4A) respectively, and it transitions to a constant voltage of 4.2V for charging until the cut-off current.
[0014] Furthermore: The specific voltage dataset includes the data of voltage changes in the voltage range of [3.5, 4.2V] during the constant current charging stage.
[0015] Furthermore: The charging encoder includes a one-dimensional convolutional module (1D-CNN), a bidirectional jump TCN-BiGRU network module, a fully connected layer, and a dense layer. Step 3 specifically includes the following steps:
[0016] Step 3.1: The specific voltage dataset is converted by the 1D-CNN layer to obtain local feature data;
[0017] Step 3.2: The local feature data is processed by the bidirectional jump TCN-BiGRU network module to obtain model data;
[0018] Step 3.3: The model data is successively processed by a fully connected layer and a dense layer to obtain charging coding data.
[0019] Furthermore: The bidirectional jump TCN-BiGRU network module includes a forward temporal convolutional network, a reverse temporal convolutional network, a fully connected layer, and a bidirectional gated recurrent unit. The forward temporal convolutional network uses dilated causal convolution to process local feature data in chronological order to ensure that the time dependence conforms to causality; after the reverse temporal convolutional network flips the local feature data along the time axis, dilated causal convolution is performed to capture reverse time information; one branch of the outputs of both the forward temporal convolutional network and the reverse temporal convolutional network is sent to the bidirectional gated recurrent unit through the fully connected layer, and the other branch of their outputs is sent to the bidirectional gated recurrent unit through two skip connections respectively, and when the skip connection condition is met, the outputs of both the forward temporal convolutional network and the reverse temporal convolutional network are directly sent to the bidirectional gated recurrent unit through the skip connection.
[0020] Furthermore: The skip connection condition is that the output of the forward temporal convolutional network or the output of the reverse temporal convolutional network is inconsistent with the input dimension of the bidirectional gated recurrent unit, and the input dimension is the number of channels or the number of time steps.
[0021] Furthermore: The skip connection condition is that the input value of the specific voltage dataset is very large or very small, and the depth of the bidirectional gated recurrent unit exceeds 10 layers.
[0022] Furthermore: The skip connection condition is that the weight operations in the forward temporal convolutional network and the reverse temporal convolutional network judge whether the input of the specific voltage dataset is a high-value input. Among them, when the weight of the input of the specific voltage dataset is greater than 0.5, the input of the specific voltage dataset is recorded as a high-value input.
[0023] Furthermore: Both the SOH decoder and the SOC decoder include dense layers. The dense layer of the SOH decoding layer is used to extract long-term degradation features related to battery aging, capture low-frequency and slowly changing global trends, and perform a non-linear transformation on the charging coding data to obtain SOH decoding data; the dense layer of the SOC decoder is used to extract dynamic instantaneous features of the battery state, process high-frequency and rapidly changing local signals, and perform a non-linear transformation on the input to obtain SOC decoding data.
[0024] Furthermore: The SOH encoder includes a Gaussian noise layer, a padding layer, and a dense layer. The dense layer is used to process fine-grained feature relationships and perform a non-linear transformation on the SOH coding data as part of the input to the SOC decoder.
[0025] Advantages of the present invention: The bidirectional jump TCN-BiGRU is used to effectively reduce redundant data input, enabling the use of one model to solve the problem of joint estimation of SOH-SOC of quasi-solid-state battery data; the SOH-SOC joint estimation model has high accuracy, high stability and generalization ability, and its parallel processing speed is fast and the computing efficiency is increased. Brief Description of the Drawings
[0026] Figure 1 is a flowchart of the present invention;
[0027] Figure 2 is a structural block diagram of the SOH-SOC joint estimation model in the present invention;
[0028] Figure 3 is a structural block diagram of the bidirectional jump TCN-BiGRU network module in the present invention. Detailed Embodiment
[0029] The present invention will be described in detail below with reference to the drawings. The embodiments of the present invention are described in detail below. The examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention. The orientation terms such as left, middle, right, up, and down in the embodiments of the present invention are only relative concepts to each other or are referenced based on the normal use state of the product, and should not be considered as restrictive.
[0030] A method for joint estimation of SOH-SOC of quasi-solid-state batteries for energy storage scenarios, as Figure 1 shown, includes the following steps:
[0031] Step 1: Obtain a quasi-solid-state battery data set, which includes all voltage data sets and specific voltage data sets;
[0032] Specifically, the self-built quasi-solid-state battery data set includes all voltage data sets and specific voltage data sets. The data set self-built in the present invention includes the capacity and charge changes during the entire life cycle charging process. The data set used in the present invention is the data collected by the Arbin BT2000 battery experiment system during continuous charging experiments.
[0033] Two datasets of quasi-solid-state batteries were obtained from experiments. The experiments were conducted at a controlled ambient temperature of 25 °C, and the batteries were charged using a constant current-constant voltage charging protocol. During the constant current charging process, data were collected every 60 s. Four batteries were selected from the first condition of the self-built quasi-solid-state battery dataset and named L11, L12, L13, and L14. Eight batteries were selected from the second condition and named PH01 - PH08. The nominal capacity of the batteries in both conditions was 2 AH, and the nominal voltage was 3.6 V. The dataset contains comprehensive information about the aging process, including voltage, current, charging time, and other voltage-related parameters, providing a reference for the degradation of the battery throughout its life. For subsequent experiments, L11, L12, and L13 were selected as the training dataset, while L14 was selected as the test dataset, and PH01 - PH06 were selected as the training dataset, and PH07 and PH08 were selected as the test dataset to evaluate the performance of the training results. Analyzing the decay curves of quasi-solid-state batteries reveals three different characteristics. First, the capacity degradation process shows significant differences among different batteries. Second, the trend of capacity degradation is non-linear, and the degradation rate increases as the life cycle progresses. Third, the capacity degradation of the battery follows an oscillatory decline pattern rather than a monotonic decrease. In this invention, the State of Health (SOH) is defined as the ratio of the current capacity of the battery to the initial capacity, and the State of Charge (SOC) is defined as the ratio of the current charge capacity of the battery to the fully charged charge capacity. The definitions of SOH and SOC are shown in Formulas 1 and 2.
[0034] (1)
[0035] (2)
[0036] Wherein, represents the maximum remaining discharge capacity of the current battery, represents the initial capacity of the battery. When the State of Health drops below 80%, it indicates that the battery is approaching its upper life limit and should be replaced in a timely manner. represents the current charge capacity of the battery, represents the fully charged charge capacity.
[0037] Subsequent data analysis mainly focused on the lifespan within a specified time range. For the first operating condition, the initial charging current during the constant current and constant voltage charging process was 0.25C (0.5A), which then transitioned to a constant voltage of 4.2V for charging until the cut-off current. For the second operating condition, the initial charging current during the constant current and constant voltage charging process was 2C (4A), which then transitioned to a constant voltage of 4.2V for charging until the cut-off current. The duration of the constant current and constant voltage charging cycle in the battery dataset to reach the rated voltage decreased as the battery's life cycle progressed during the constant current charging process, thus supporting the relationship between the remaining capacity and the constant current charging. The voltage range for constant current charging mainly spanned [3.5V, 4.2V]. The voltage initially rose rapidly and then transitioned to a gradual increase, and the rate of increase became increasingly obvious. Previous studies used parameter variations during the charging process to estimate SOH. In this invention, data on voltage variations within a specific range during the constant current charging stage were used for SOH estimation, and the voltage range [3.5V, 4.2V] with a larger variation program was selected as the specific voltage dataset. Many studies have analyzed and verified that the voltage, current, and time variations of battery data are the most suitable input parameters for estimating SOC and SOH. Therefore, this invention does not further study these aspects. In this dataset, the data collection interval during the CC charging process was fixed. Here, the influence of the time parameter on the model was limited. Therefore, we chose to focus only on voltage and current as the input parameters for the model.
[0038] Step 2: Establish an SOH-SOC joint estimation model, as Figure 2 shown, the SOH-SOC joint estimation model includes a charging encoder, an SOH encoder, an SOH decoder, and an SOC decoder;
[0039] To better estimate SOC throughout the battery's life cycle, the SOH of the current battery must be obtained. Moreover, accurate estimation of SOH requires data collected within a specific voltage range during the charging process. Therefore, this invention creates an SOH-SOC joint estimation model based on a charging encoder. This model uses multiple data points from the specific voltage dataset in each charging cycle to jointly estimate SOC and SOH, and obtains an SOH decoder capable of estimating SOH. As the battery life decays, multiple observed variables appear at each time point. By arranging these variables in chronological order, a series of current-voltage characteristics are formed, making the current-voltage characteristics the input to the charging encoder.
[0040] Step 3: Feed the specific voltage dataset into the charging encoder and obtain the charging encoded data; where the charging encoder includes a one-dimensional convolutional module (1D-CNN), a bidirectional jump TCN-BiGRU network module, a fully connected layer, and a dense layer, specifically including the following steps:
[0041] Step 3.1: The specific voltage data set is converted by the 1D-CNN layer to obtain local feature data;
[0042] The first layer of the charging encoder consists of 1D-CNN. 1D-CNN has strong feature extraction capabilities and is particularly suitable for processing one-dimensional sequence data. In the joint estimation of SOH and SOC, parameters such as the current and voltage of the battery usually exist in the form of one-dimensional sequences. 1D-CNN can effectively extract local features from these parameters by means of the sliding window of the convolutional layer, thereby capturing the subtle information of the battery state change. It has two main purposes: dimension conversion and local feature extraction. It is used to capture the important features and local patterns of current and voltage features.
[0043] (3)
[0044] In the above equation, represents the output at time step , represents the input at time step , represents the weight, and represents the kernel size. One-dimensional convolution can enable parameter sharing and effective feature extraction from sequential data, which makes it particularly suitable for tasks such as time series analysis and natural language processing;
[0045] Step 3.2: The local feature data is processed by the bidirectional jump TCN-BiGRU network module to obtain model data; among them, as Figure 3 shown, the bidirectional jump TCN-BiGRU network module includes a forward temporal convolutional network, a reverse temporal convolutional network, a fully connected layer, and a bidirectional gated recurrent unit. The forward temporal convolutional network uses dilated causal convolution to process the local feature data in chronological order to ensure that the time dependence conforms to causality; the reverse temporal convolutional network flips the local feature data along the time axis and then performs dilated causal convolution to capture reverse time information; one branch of the outputs of both the forward temporal convolutional network and the reverse temporal convolutional network is sent to the bidirectional gated recurrent unit through the fully connected layer, and the other branch of their outputs is sent to the bidirectional gated recurrent unit through two skip connections respectively, and when the skip connection condition is met, the outputs of both the forward temporal convolutional network and the reverse temporal convolutional network are directly sent to the bidirectional gated recurrent unit through the skip connection; the conditions for enabling the skip connection are as follows:
[0046] (1) The output of the forward temporal convolutional network or the output of the reverse temporal convolutional network is inconsistent with the input dimension of the bidirectional gated recurrent unit. The input dimension is the number of channels or the number of time steps. The skip connection is enabled for dimension transformation to unify the input;
[0047] (2)When the input value of the specific voltage data set is very large or very small, the gradient will approach 0, resulting in gradient vanishing. Moreover, when the depth of the bidirectional gated recurrent unit exceeds 10 layers, gradient explosion may occur. Jumping connections are enabled to enhance the gradient backpropagation ability to solve the problems of gradient vanishing and explosion.
[0048] (3)The weight operations in the forward temporal convolutional network and the backward temporal convolutional network determine whether the input of the specific voltage data set is a high-value input. Among them, when the weight of the input of the specific voltage data set is greater than 0.5, the input of the specific voltage data set is recorded as a high-value input, and jumping connections are enabled to avoid feature loss when it is a high-value input.
[0049] Although the forward temporal convolutional network and the backward temporal convolutional network can capture recent trends and short-term dependencies, they are limited by the fixed-length historical information and cannot flexibly adapt to battery data of different lengths, resulting in limitations in the network when dealing with historical dependency problems. Secondly, in the deep structure, the distance of information transmission is limited, making it difficult to capture the correlation between distant time steps. To solve the problems of historical dependency and information transmission loss in battery data, the present invention uses a bidirectional gated recurrent unit to handle this problem. The bidirectional gated recurrent unit can simultaneously obtain the forward and backward information of the input data through the recurrent neural network structures in the forward and backward directions, thus better capturing the long-term dependencies in the data. During the degradation process of the battery, the changes in SOH and SOC are often affected by historical data. This ability of the bidirectional gated recurrent unit enables it to more accurately predict the future state of the battery. The bidirectional gated recurrent unit contains an update gate and a reset gate, which are used to control the flow of information. The role of the update gate is to determine how much information from the previous time step needs to be retained in the hidden state of the current time step. The output value of the update gate ranges between 0 and 1. The larger the value, the more past information is retained, and the smaller the value, the more it depends on the information of the current input. In addition, the receptive field determines the range of dependencies that the forward temporal convolutional network and the backward temporal convolutional network can capture. By using dilated convolutions, the forward temporal convolutional network and the backward temporal convolutional network can effectively model distant dependencies without significantly increasing the number of parameters. The key advantage of the forward temporal convolutional network and the backward temporal convolutional network lies in its ability to capture long-term dependencies while maintaining computational efficiency. The forward temporal convolutional network and the backward temporal convolutional network provide a powerful tool for analyzing sequential data. 1D-CNN allows local feature extraction, while the forward temporal convolutional network and the backward temporal convolutional network expand this concept to capture long-term dependencies..
[0050] Step 3.3: The model data is sequentially passed through a fully connected layer and a dense layer to obtain the charging coding data. The dense layer includes a weight matrix, a bias vector, and an activation function. The ReLU activation function is used to perform a non-linear transformation on the model data to extract high-order features, making the output dimension of the previous module match the input dimension of the next module, which can effectively fit the complex dynamic relationship during battery aging. It is also responsible for receiving more abstract global features in the model data and obtaining the charging coding data after performing a non-linear transformation on the model data. The fully connected layer can receive the model data processed by the previous bidirectional jump TCN-BiGRU network module and integrate these model data. In the SOH-SOC joint estimation task, various state information of the battery may be scattered in different model data. The fully connected layer can integrate these scattered features together to form a comprehensive feature representation, facilitating subsequent estimation tasks. Moreover, in the battery SOH-SOC joint estimation model, overfitting is a common problem. Adding a fully connected layer after the bidirectional jump TCN-BiGRU network module can further reduce overfitting.
[0051] Step 4: The charging coding data is fed into the SOH encoder to obtain specific SOH decoding data. The SOH decoder includes a dense layer. The dense layer of the SOH decoding layer is used to extract long-term degradation features related to battery aging, capture low-frequency, slowly varying global trends, and perform a non-linear transformation on the charging coding data to obtain the SOH decoding data.
[0052] Step 5: The specific SOH decoding data and the entire voltage data set are fed into the SOH encoder to obtain the entire SOH coding data. The SOH encoder includes a Gaussian noise layer, a padding layer, and a dense layer. The dense layer is used to process fine-grained feature relationships and perform a non-linear transformation on the SOH coding data as part of the input to the SOC decoder. In the SOH-SOC joint estimation task, various state information of the battery can be efficiently integrated and mapped through the dense layer into the SOH decoding data, providing strong support for subsequent estimation tasks.
[0053] Step 6: The entire SOH coding data and the charging coding data are fed into the SOC decoder to obtain the SOC decoding data. The SOC decoder includes a dense layer. The dense layer of the SOC decoder is used to extract dynamic instantaneous features of the battery state, process high-frequency, rapidly changing local signals, and perform a non-linear transformation on the input to obtain the SOC decoding data.
[0054] In the previous steps, a pre-trained charging encoder was adopted to encode the local feature data of the charging process, and a dedicated SOH encoder specifically for encoding SOH was combined. Subsequently, all the SOH encoded data and the charging encoded data were simultaneously input into the SOC decoder, thus achieving accurate decoding and estimation of SOC. It should be noted that in the actual estimation, the SOH decoded data represents an estimated value, which may have some inherent errors, different from the measured SOH value used in the training process. To mitigate overfitting, Gaussian noise was introduced into the SOH decoded data during the training process. According to the actual usage requirements, it is necessary to estimate SOC throughout the charging process over the entire life cycle. However, the initial stage of constant-current charging cannot provide a long enough sequence to be effectively input into the model. To overcome this limitation, a padding operation was introduced at the beginning of the specific SOH decoded data, that is, padding the entire voltage data set, which enables the model to accurately estimate SOC throughout the charging process. In addition, in the early stage of constant-current charging, the SOH decoded data is not yet clear because the model has no estimate. Therefore, padding is also applicable to the early stage of SOH input. It should be emphasized that the SOH input in the early stage of charging has little impact on the accuracy of SOC estimation. Therefore, padding the entire voltage data set at this stage will not significantly affect the accuracy of SOC estimation. The charging encoder and the SOH encoder are used to encode the charge sequence of multiple features and the current SOH respectively. Subsequently, the SOC decoder and the SOH decoder are used to estimate SOC and SOH respectively. In the early stage of each charging cycle, when the length of the entire voltage data set is less than the input sequence length of the model, a padding operation is applied to the sequence. In addition, when no specific voltage fluctuations are observed, it is not feasible to accurately estimate SOH. In this case, only SOC is estimated, and the entire voltage data set is padded. As the charging progresses and the voltage enters a specific range, the model can estimate both SOC and SOH. To better represent the estimated value of SOH, the final result of SOH is the average of the SOH estimated values of the known sequences.
[0055] Step 7: Use the specific SOH decoded data and the SOC decoded data as the estimated values of the state of health and the state of charge of the battery respectively.
[0056] In this study, we adopted multiple evaluation metrics to comprehensively evaluate the performance of the proposed encoder-decoder. These metrics help to objectively measure the estimation ability of the model and reveal its accuracy and robustness from different perspectives. The following are the five evaluation metrics we used:
[0057] 1. Mean Squared Error (MSE): Mean Squared Error is a commonly used metric to evaluate the difference between the model's estimated results and the true values; it calculates the average of the squared differences between the estimated values and the true values to measure the average deviation of the estimation;
[0058] (4)
[0059] Where n represents the number of samples, t represents the index of the sample, represents the estimated value of sample t, represents the true target value of sample t.
[0060] 2. Root Mean Square Error (RMSE): The root mean square error is the square root of the mean square error, which represents the average difference between the estimated value and the true value; RMSE is more sensitive to outliers and can be used to measure the precision of the model;
[0061] (5)
[0062] 3. Mean Absolute Error (MAE): The mean absolute error is the average of the absolute differences between the estimated value and the true value, measuring the average error of the estimation; different from MSE, MAE does not amplify the influence of large errors, so it can better reflect the overall accuracy of the estimation;
[0063] (6)
[0064] 4. Mean Absolute Percentage Error (MAPE): The mean absolute percentage error is the average of the relative differences between the estimated value and the true value, expressed as a percentage; it measures the relative error of the model within different data ranges and can reflect the relative accuracy of the estimation;
[0065] (7)
[0066] 5. Maximum Absolute Error (MAXE): The maximum absolute error is the maximum of the absolute differences between the estimated value and the true value, which identifies the estimation error of the model in the worst case; MAXE is particularly sensitive to outliers and helps to understand the maximum risk in the estimation of the model;
[0067] (8)
[0068] Where represents the operation of taking the maximum value.
[0069] In this paper, the quasi-solid-state batteries L11, L12, and L13 were used as the training dataset, L14 was selected as the test dataset, PH01 - PH06 as the training dataset, and PH07 and PH08 were selected as the test datasets. The mean absolute error (MAE) was used as the optimization objective to balance outlier sensitivity and convergence stability. The early stopping mechanism was triggered when the MAE of the validation set decreased by less than 0.1% for 10 consecutive epochs, and the minimum learning rate was set to 1% of the initial value. The parameter sensitivity experiment showed that when Batch size = 64, the stability of the gradient update direction was the best, and the MAE variance decreased by 28.7% and 45.2% compared to when Batch size was 32 and 16, respectively; the combination of Epoch = 100 and learning rate = 0.0025 reached the optimal balance point on the validation set, and continuing to increase the number of training epochs led to overfitting. The AdamW optimizer was used, and its core improvement lies in decoupling weight decay and adaptive learning rate.
[0070] Table 1 Device Configuration and Model Parameters
[0071]
[0072] To verify the effectiveness of the proposed estimation model, a series of experiments were conducted, specifically including:
[0073] To ensure the effectiveness of the experiment, we uniformly used the parameters set in Table 1 and conducted the experiment in the same environment. It can be clearly seen that the encoder-decoder achieved the best results in all five evaluation metrics. In the robustness experiment, we evaluated the impact of noise, aiming to explore the performance of the encoder-decoder under different noise levels. In the experiment, we introduced different amplitudes of noise (50mV, 100mV, and 150mV) to simulate the uncertainty of battery data in reality. The experimental results are shown in Table 2:
[0074] Table 2 Results of Introduced Noise
[0075]
[0076] It can be clearly observed from the results in the table that under all evaluation metrics, as the noise level increases, the performance of each model decreases. However, the encoder-decoder still maintains high estimation accuracy under different noise levels. From the test set L14, it can be seen that as the noise level increases from 50mV to 150mV, the MSE of the encoder-decoder only increases from 0.53% to 0.88%. This indicates that the encoder-decoder has strong robustness to noise and can resist the impact of data uncertainty to a certain extent.
[0077] In the comparative experiments, we further verified the accuracy of the proposed encoder-decoder by conducting comparative experiments with multiple open-source models (GRU, LSTM, and KAN), evaluating using the test set. In this series of comparative experiments, we maintained unified experimental parameter settings and the same environmental conditions to ensure the comparability of the experimental results. The results are shown in Table 3 as follows:
[0078] Table 3 Results of Comparative Experiments
[0079]
[0080] It can be clearly observed from the experimental results that the encoder-decoder shows significant advantages under all evaluation metrics. As seen from the test set L14, the encoder-decoder reduces by approximately 56%, 42%, 44%, 36%, and 51% respectively in the five evaluation metrics of MSE, RMSE, MAE, MAPE, and MAXE compared to the GRU, LSTM, and KAN models. It can be concluded that the encoder-decoder achieves more accurate and stable estimation results on the test set and has obvious advantages under multiple evaluation metrics compared to the open-source models GRU, LSTM, and KAN. This further proves the excellent performance of the encoder-decoder in the joint estimation of SOH-SOC of quasi-solid-state batteries, providing a powerful tool and guidance for the optimization of battery management and maintenance strategies.
[0081] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for jointly estimating the state of health (SOH) and state of charge (SOC) of a quasi-solid-state battery for energy storage scenarios, characterized in that: It includes the following steps: Step 1: Obtain a quasi-solid-state battery dataset, which includes all voltage datasets and specific voltage datasets; Step 2: Establish an SOH-SOC joint estimation model. The SOH-SOC joint estimation model includes a charging encoder, an SOH encoder, an SOH decoder, and an SOC decoder. Among them, the charging encoder includes a one-dimensional convolutional module, a bidirectional jump TCN-BiGRU network module, a fully connected layer, and a dense layer. The bidirectional jump TCN-BiGRU network module includes a forward temporal convolutional network, a backward temporal convolutional network, a fully connected layer, and a bidirectional gated recurrent unit. The forward temporal convolutional network uses dilated causal convolution to process local feature data in chronological order to ensure that the time dependence conforms to causality. The backward temporal convolutional network flips the local feature data along the time axis and then performs dilated causal convolution to capture reverse time information. One branch of the outputs of both the forward temporal convolutional network and the backward temporal convolutional network is sent to the bidirectional gated recurrent unit through the fully connected layer, and the other branch of their outputs is sent to the bidirectional gated recurrent unit through two skip connections respectively. When the skip connection conditions are met, the outputs of both the forward temporal convolutional network and the backward temporal convolutional network are directly sent to the bidirectional gated recurrent unit through the skip connection. The SOH encoder includes a Gaussian noise layer, a padding layer, and a dense layer. This dense layer is used to process fine-grained feature relationships and, after non-linearly transforming the SOH encoded data, serves as part of the input to the SOC decoder. Both the SOH decoder and the SOC decoder include dense layers. The dense layer of the SOH decoding layer is used to extract long-term degradation features related to battery aging, capture low-frequency and slowly varying global trends, and obtain SOH decoded data after non-linearly transforming the charging encoded data. The dense layer of the SOC decoder is used to extract dynamic instantaneous features of the battery state, process high-frequency and rapidly changing local signals, and obtain SOC decoded data after non-linearly transforming the input; Step 3: Send the specific voltage dataset into the charging encoder and obtain charging encoded data; Step 3 specifically includes the following steps: Step 3.1: The specific voltage dataset is converted by the 1D-CNN layer to obtain local feature data; Step 3.2: The local feature data passes through the bidirectional jump TCN-BiGRU network module to obtain model data; Step 3.3: The model data passes through the fully connected layer and the dense layer in sequence to obtain charging encoded data; Step 4: Send the charging encoded data into the SOH encoder and obtain specific SOH decoded data; Step 5: Send the specific SOH decoded data and all voltage datasets into the SOH encoder to obtain all SOH encoded data; Step 6: Send all SOH encoded data and charging encoded data into the SOC decoder to obtain SOC decoded data; Step 7: Use the specific SOH decoded data and the SOC decoded data as the estimated values of the battery health state and the battery charge state respectively.
2. The SOH-SOC joint estimation method for the quasi-solid-state battery for energy storage scenarios according to claim 1, wherein: The quasi-solid-state battery dataset is a self-built dataset and includes the capacity and charge changes during the charging process of the entire life cycle of the quasi-solid-state battery; this dataset is the data collected by the Arbin BT2000 battery experiment system during continuous charging experiments, and all experimental quasi-solid-state batteries follow the same charging curve; the experiments are carried out at a controlled ambient temperature of 25 °C, and the batteries are charged using a constant current and constant voltage charging protocol; during the constant current charging process, data is collected every 60 s; four batteries were selected from the self-built quasi-solid-state battery dataset, named L11, L12, L13, and L14 respectively. These four batteries all use lithium iron phosphate as the positive electrode material, with a nominal capacity of 1.1 AH and a nominal voltage of 4.2 V; The dataset contains comprehensive information about the aging process, and the comprehensive information includes voltage, current, and charging time; L11, L12, and L13 are selected as the training dataset, while L14 is selected as the test dataset. During a complete constant current and constant voltage charging process, the initial charging current is 0.55 A, and it transitions to a constant voltage of 4.2 V and charges until the cut-off current.
3. The SOH-SOC joint estimation method for the quasi-solid-state battery for energy storage scenarios according to claim 2, wherein: The specific voltage dataset includes data on voltage changes within the voltage range [3.7 V, 3.95 V] during the constant current charging stage.
4. The SOH-SOC joint estimation method for a quasi-solid-state battery for an energy storage scenario according to claim 1, characterized in that: The skip connection condition is that the output of the forward temporal convolutional network or the output of the reverse temporal convolutional network is inconsistent with the input dimension of the bidirectional gated recurrent unit, and the input dimension is the number of channels or the number of time steps.
5. The SOH-SOC joint estimation method for the quasi-solid-state battery for energy storage scenarios according to claim 1, characterized in that: The skip connection condition is that the weight operations in the forward temporal convolutional network and the reverse temporal convolutional network determine whether the input of the specific voltage dataset is a high-value input. Among them, when the weight of the input of the specific voltage dataset is greater than 0.5, the input of the specific voltage dataset is recorded as a high-value input.