Park lithium battery SOC and SOH joint estimation method and device

Abstract

The invention belongs to the technical field of SOC and SOH estimation, and particularly relates to a park lithium battery SOC and SOH joint estimation method and device, and the method comprises the steps: collecting the multi-dimensional feature data of a lithium battery in the charging and discharging process, extracting a health factor, and carrying out the preprocessing of the data; constructing a CNN-BiLSTM model, combining the CNN-BiLSTM model with a LightGBM model, constructing a joint estimation model, and training the joint estimation model; and inputting the health factor into a CNN-BiLSTM model to estimate the SOH, and inputting the multi-dimensional feature data and an estimated value of the SOH into a LightGBM model to estimate the SOC, thereby obtaining a final SOC and SOH joint estimated value. According to the method, the time sequence features of the SOH are modeled through the BiLSTM and the CNN, and SOC estimation is performed in combination with the LightGBM, so that SOC calculation can dynamically consider the change of the SOH, and accumulation of SOC estimation errors is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of SOC and SOH estimation, and particularly relates to a method and device for jointly estimating the SOC and SOH of lithium batteries in a park. Background Art

[0002] As a bridge between energy production, transmission, and consumption, the energy storage system can balance the fluctuations in energy supply and demand and improve energy utilization efficiency. Lithium-ion batteries, with their advantages of high energy density, long life, fast response, and environmental friendliness, have become one of the core technologies of the energy storage system. In the battery management system, the state estimation of lithium batteries has important practical significance, mainly including the estimation of the state of charge (SOC) and the state of health (SOH). The SOC reflects the remaining charge of the battery at present and is a key parameter for evaluating the battery's endurance, optimizing energy scheduling, and ensuring the normal operation of equipment. The SOH, on the other hand, reflects the health status and degradation degree of the battery and is an important basis for predicting the battery life, formulating maintenance plans, and preventing potential safety risks.

[0003] With the use of lithium-ion batteries, the battery capacity decays, the internal resistance increases, and its dynamic performance also degrades. The SOC calculation is usually based on the rated capacity of the battery. If the capacity decay is not considered, it will lead to inaccurate SOC estimation. Therefore, in the process of estimating the SOC, it is necessary to dynamically adjust the actual capacity of the battery, that is, jointly estimate the SOH and SOC. Currently, there are mainly three types of methods for jointly estimating the SOC and SOH of lithium-ion batteries: model-driven methods, fusion-driven methods, and data-driven methods.

[0004] Model-driven methods (such as equivalent circuit models or electrochemical models) describe the internal mechanism of the battery through mathematical modeling and combine methods such as Kalman filtering for SOC / SOH estimation. However, this method relies on accurate battery parameters, is difficult to adapt to different battery types, has a high computational complexity, and is difficult to meet the real-time requirements of the BMS. In addition, the model parameters are easily affected by environmental factors, resulting in the accumulation of estimation errors.

[0005] Fusion-driven methods combine model-driven and data-driven methods. Usually, an equivalent circuit model is used to estimate the initial value, and then neural networks or machine learning methods are used for optimization. However, this method has a large computational overhead and is difficult to be deployed in real time in a resource-constrained BMS system. At the same time, the fusion of physical models and data models may lead to information inconsistency, affecting the estimation stability. In addition, the fusion method still relies on battery aging data for training, and its generalization ability is limited.

[0006] In contrast, the data-driven method is simpler and more flexible, capable of bypassing complex battery models and being applicable to a wider range of application scenarios through data-driven means. However, SOC and SOH are closely related. Traditional methods usually estimate the state of charge (SOC) and the state of health (SOH) separately, ignoring their mutual influence, which leads to error accumulation. For example, if SOH is not considered in the SOC calculation, when the battery degrades, there will be a systematic deviation in the SOC estimation. Especially in the battery aging stage, the SOC estimation error will continue to expand. Traditional SOC / SOH estimation methods usually rely on single-feature input (such as voltage or current), making it difficult for the model to adapt to complex battery operating conditions (such as different charge and discharge rates, ambient temperature changes, etc.), thus affecting the estimation accuracy. Although deep learning methods (such as RNN / LSTM) can learn complex time series relationships in SOC-SOH estimation, the computational cost is relatively high, making it difficult to meet the requirements of real-time monitoring. Summary of the Invention

[0007] Based on the above deficiencies in the prior art, the purpose of the present invention is to provide a method and device for jointly estimating the SOC and SOH of lithium batteries in a park based on LightGBM-CNN-BiLSTM. By using BiLSTM and CNN to model the time series characteristics of SOH and combining LightGBM for SOC estimation, the SOC calculation can dynamically consider the changes in SOH and reduce the accumulation of SOC estimation errors.

[0008] To achieve the above object, the present invention provides a method for jointly estimating the SOC and SOH of lithium batteries in a park, including the following steps: S1. Collect multi-dimensional characteristic data of the lithium battery during the charge and discharge process, extract health factors, and perform data preprocessing; S2. Construct a CNN-BiLSTM model; S3. Combine the CNN-BiLSTM model with the LightGBM model to construct a joint estimation model, where the SOH is estimated using the CNN-BiLSTM model and the SOC is estimated using the LightGBM model; S4. Train the CNN-BiLSTM model and the LightGBM model; S5. Based on the trained joint estimation model, input the health factors into the CNN-BiLSTM model to estimate SOH, input the multi-dimensional characteristic data and the estimated value of SOH into the LightGBM model to estimate SOC, and integrate the outputs of the CNN-BiLSTM model and the LightGBM model to obtain the final joint estimation values of SOC and SOH.

[0009] As a preferred solution of the present invention, in the above S1, the multi-dimensional characteristic data is the current [I of the lithium battery 1, I 2 , …, I n , voltage [U 1 , U 2 , …, U n , and temperature [T 1 , T 2 , …, T n , where n is the number of data points, i.e., the data corresponding to different time points; The extracted health factors are [HF 1 , HF 2 , …, HF m J , where m is the number of health factor data and J represents the number of types of health factors; The preprocessing includes missing value handling, outlier handling, data cleaning, and normalization.

[0010] As a preferred embodiment of the present invention, the health factors include: Charging stage characteristics, including constant current charging time, constant voltage stage current decay rate, charging voltage platform slope, charging termination voltage; Discharging stage characteristics, including discharging voltage drop gradient, effective discharging capacity; Resting stage characteristics, including relaxation voltage recovery rate, open circuit voltage decay; Temperature dynamic characteristics, including constant current charging temperature rise rate, local temperature difference; Differential / integral characteristics, including incremental capacity peak voltage, differential voltage curvature; Statistical distribution characteristics, including voltage fluctuation standard deviation, temperature-current covariance; Dynamic process characteristics, including pulse response recovery time, relaxation process time constant; Special working condition characteristics, including low temperature charging voltage overshoot, high temperature discharging cut-off advance; Time series correlation characteristics, including voltage-temperature hysteresis angle, capacity-temperature coupling coefficient.

[0011] As a preferred embodiment of the present invention, in the S2, the CNN-BiLSTM model consists of an input layer, a CNN convolutional block, a BiLSTM hidden layer, a fully connected layer, and an output layer, and the output layer is a regression layer, where: The CNN convolutional block contains a convolutional layer with a kernel size of 1×1, 32 convolutional kernels, the weight initialization method of the convolutional layer is He initialization, the padding mode is same, and the ELU activation function is adopted; after the convolutional layer, there is a batch normalization layer and an average pooling layer, and the pooling stride is set to FiltZise; ​The BiLSTM hidden layer includes a bidirectional LSTM layer with 128 nodes, and its weight initialization also adopts He initialization. The 128-node bidirectional LSTM layer is followed by a 32-node bidirectional LSTM layer.

[0012] As a preferred solution of the present invention, in the above S3, the LightGBM model consists of multiple decision trees based on gradient boosting, and continuously optimizes the fitting ability for complex features through the way of ensemble learning, so as to achieve efficient and accurate estimation; Set the number of leaf nodes of the LightGBM model to 4, adopt gradient boosting decision tree as the boosting type, and at the same time set the feature selection ratio to 0.9, the sample sampling ratio for tree building to 0.8, and set the frequency to 7, and the learning rate is set to 0.1.

[0013] As a preferred solution of the present invention, in the above S3, in the CNN-BiLSTM model, CNN extracts local features through convolution operations, and BiLSTM performs two LSTM processes on the time series in the forward and reverse directions. The final output is the concatenation of two hidden states: ; In the formula, is the output of the forward LSTM; is the output of the reverse LSTM; is the output after concatenation; Infer the SOH from the hidden state through the fully connected layer, which is expressed as: ; In the formula, is the estimated value of SOH; is the weight matrix; is the bias term; The LightGBM model integrates multiple weak learners, which is expressed as: ; In the formula, is the v-th weak learner, V is the number of weak learners, and each weak learner is a decision tree model; represents the output of the v-th weak learner; represents the final estimation result, that is, the sum of the outputs of all weak learners; is the set space of all weak learners.

[0014] As a preferred solution of the present invention, in the above S3, the LightGBM model consists of multiple decision trees based on gradient boosting. Under the framework of gradient boosting decision tree, the estimation process of the LightGBM model is as follows: The LightGBM model is initialized to a constant value, which is the mean of the historical battery SOC data. The initial LightGBM model is expressed as: ; In the formula, is the loss function; is the true SOC value of the lithium battery at the i-th moment, that is, the true SOC value of the i-th sample; N is the total number of samples; is the constant value for initialization; At the k-th iteration, the model makes an estimation, calculates the gap between the estimated value and the true value, that is, the residual, which is represented by the gradient of the loss function: ; In the formula, represents the gradient of the i-th sample at the k-th iteration; represents the true SOC value of the u-th sample; is the estimated SOC value of the i-th sample in the previous iteration; Using the residual as the target, construct a new decision tree to fit the current residual, and update the SOC estimation model according to the estimation result; Repeat the above process, continuously fit the residual with a new decision tree and update the model, and finally obtain the estimated value of SOC through the superposition of multiple rounds of iterations : ; In the formula, P is the total number of iterations; is the learning rate.

[0015] As a preferred solution of the present invention, in S4, model training is performed based on the multi-dimensional feature data and health factor data in S1. Before training, the data participating in the training is enhanced, including at least one of time series translation, time series scaling, noise addition, data truncation, data inversion, enhancement based on charge and discharge curves, enhancement based on generative adversarial network GAN, enhancement based on variational autoencoder VAE, enhancement based on time series generation model, and enhancement based on data interpolation, where: The enhancement based on the charge and discharge curve is to perform translation, scaling, and rotation transformations on the charge and discharge curve of the battery to generate new charge and discharge curve data; The enhancement based on GAN is to use GAN to generate new voltage, current, and temperature data, as well as health factor data; The enhancement based on VAE is to use VAE to generate new voltage, current, and temperature data, as well as health factor data; The enhancement based on the time series generation model is to use LSTM or Transformer to generate new time series data and simulate the performance changes of the battery during different charging and discharging processes; The enhancement based on data interpolation is to use linear interpolation or spline interpolation methods to generate new voltage, current, and temperature data.

[0016] As a preferred solution of the present invention, in S5, for the SOH estimation value and the SOC estimation value in the combined estimation value, weighted fusion is performed to obtain a fused estimation value for evaluating the overall state of the lithium battery. The method is to construct a dual-stream attention fusion network through dynamic attention fusion to dynamically generate the spatio-temporal adaptive weights of the SOH estimation value output by the CNN-BiLSTM model and the SOC estimation value output by the LightGBM model. The process is as follows: Step 1: Concatenate the hidden state vector of the last time step of the BiLSTM hidden layer of the CNN-BiLSTM model with the leaf node distribution vector of the final decision tree of the LightGBM model to form a cross-model feature interaction matrix , where is the hidden state dimension, is the leaf node dimension; Step 2: Construct an attention weight generation module that includes a time decay factor and spatial correlation: ; In the formula, represents the attention weight; is the query matrix, is the weight matrix of the query matrix; is the key matrix, is the weight matrix of the key matrix; is the dimension of the key matrix; is the learnable time decay coefficient; is the time difference between the current moment and the historical reference moment; Step 3: Calculate the probability distribution divergence of the outputs of the two models respectively: ; In the formula, is used to quantify the difference between the probability distributions of the outputs of the two models; represents the KL divergence; is the probability distribution output by the CNN-BiLSTM model; is the probability distribution output by the LightGBM model; Step 4: Construct a gated fusion mechanism to generate the final weight coefficient: ; ; In the formula, is the weight of the SOH estimated value; is the weight of the SOC estimated value; represents the sigmoid function; is the trainable parameter matrix; represents the Hadamard product; Step Five: Generate a fused estimated value using a weighted fusion method with residual connection : ; In the formula, is the residual coefficient, is the non-linear correction network composed of a multi-layer perceptron; is the estimated value of SOH; is the estimated value of SOC.

[0017] A device for jointly estimating the SOC and SOH of lithium batteries in a park, including a memory, a processor, and a computer program stored on the memory and capable of running on the processor. The above method is implemented by the processor executing the computer program.

[0018] The beneficial effects of the present invention are: By constructing a joint estimation model based on LightGBM-CNN-BiLSTM, the present invention fully utilizes the advantages of CNN in feature extraction and the long-term dependence ability of BiLSTM in processing time series data, making the model have higher accuracy in battery SOC and SOH estimation tasks. At the same time, LightGBM is introduced for SOC estimation, combined with the SOH estimation result, so that the SOC calculation can dynamically adapt to the influence of battery aging, thereby improving the accuracy of SOC estimation and reducing error accumulation. In addition, by screening key health factors through Kendall correlation coefficient analysis, the input of SOH estimation is optimized, further improving the generalization ability and stability of the model.

[0019] The present invention uses LightGBM for SOC estimation. Compared with the deep neural network model, LightGBM has higher computational efficiency and supports efficient parallel training, and can quickly respond and achieve real-time estimation. At the same time, when CNN-BiLSTM is used for SOH estimation, it combines the local feature extraction of CNN and the bidirectional time series learning ability of BiLSTM, reducing the computational overhead while improving the estimation accuracy. This combination significantly improves the computational efficiency of the entire joint estimation method on the premise of ensuring accuracy, meeting the requirements of real-time monitoring.

[0020] The present invention constructs a method for extracting health factors, screens key health factors from multi-dimensional data, and makes the SOH estimation more accurate. By adopting a data-driven approach to automatically learn the battery degradation mode, the adaptability of the model to different battery types and different usage conditions is improved. In addition, by combining LightGBM and CNN-BiLSTM to establish a lithium-ion battery SOC-SOH joint estimation model, the advantages of each algorithm are fully utilized. Therefore, when jointly estimating SOC and SOH, the estimation accuracy and generalization ability of the model are improved, enabling it to be applicable to embedded BMS systems and intelligent energy storage management platforms to achieve real-time SOC-SOH joint estimation. Description of the Drawings

[0021] Figure 1 is the process schematic diagram of the present invention; Figure 2 is the schematic diagram of the joint estimation model in the present invention; Figure 3 is the schematic diagram of the root mean square error of SOC estimation with and without considering SOH during the verification process of the present invention, Figure 3 in which (a) is the schematic diagram of the root mean square error of Cell3; Figure 3 in which (b) is the schematic diagram of the root mean square error of Cell6; Figure 4 is the schematic diagram of the maximum error of SOC estimation with and without considering SOH during the verification process of the present invention, Figure 4 in which (a) is the schematic diagram of the maximum error of Cell3; Figure 4 in which (b) is the schematic diagram of the maximum error of Cell6; Figure 5 is the schematic diagram of the SOC estimation value of Cell3 at the 100% stage of the SOH value during the charging cycle in the verification process of the present invention; Figure 6 is the schematic diagram of the SOC estimation value of Cell3 at the 85% stage of the SOH value during the charging cycle in the verification process of the present invention; Figure 7 is the schematic diagram of the SOC estimation value of Cell3 at the 75% stage of the SOH value during the charging cycle in the verification process of the present invention; Figure 8 is the schematic diagram of the SOH estimation accuracy of Cell3 in the verification process of the present invention. Detailed Embodiments

[0022] The following further describes the embodiments of the present invention with reference to the drawings: Embodiment 1: As Figure 1 and Figure 2 shown, a method for jointly estimating the SOC and SOH of lithium batteries in a park includes the following steps: S1. Collect multi-dimensional characteristic data of the lithium battery during charging and discharging, extract health factors, and perform data preprocessing; S2. Construct a CNN-BiLSTM model; S3. Combine the CNN-BiLSTM model with the LightGBM model to construct a joint estimation model, where the SOH is estimated using the CNN-BiLSTM model and the SOC is estimated using the LightGBM model; S4. Train the CNN-BiLSTM model and the LightGBM model; S5. Based on the trained joint estimation model, input the health factors into the CNN-BiLSTM model to estimate the SOH, input the multi-dimensional characteristic data and the estimated value of the SOH into the LightGBM model to estimate the SOC, and synthesize the outputs of the CNN-BiLSTM model and the LightGBM model to obtain the final joint estimated values of the SOC and SOH.

[0023] In S1, the multi-dimensional characteristic data are the current [I 1 , I 2 , …, I n , voltage [U 1 , U 2 , …, U n , and temperature [T 1 , T 2 , …, T n , where n is the number of data points, i.e., the data corresponding to different time points; The extracted health factors are [HF 1 , HF 2 , …, HF m J , where m is the number of health factor data and J represents the number of health factor types; The preprocessing includes missing value handling, outlier handling, data cleaning, and normalization.

[0024] The health factors include: Charging stage characteristics: Constant current charging time: Determine the time of the constant current charging stage by recording the current and voltage during charging; Constant voltage stage current decay rate: During the constant voltage charging stage, record the change of current with time and calculate the current decay rate; Charging voltage platform slope: During the constant current charging stage, record the change of voltage with time and calculate the slope of the voltage platform; Charging termination voltage: Record the voltage value at the end of charging.

[0025] Discharging stage characteristics:​ Discharge voltage drop gradient: During the discharge process, record the change of voltage over time and calculate the gradient of voltage drop; Effective discharge capacity: By recording the current and time during the discharge process, calculate the effective discharge capacity.

[0026] Rest stage characteristics: Relaxation voltage recovery rate: During the battery rest stage, record the change of voltage over time and calculate the voltage recovery rate; Open-circuit voltage attenuation: During the battery rest stage, record the change of open-circuit voltage over time and calculate the voltage attenuation rate.

[0027] Temperature dynamic characteristics: Constant current charging temperature rise rate: During the constant current charging stage, record the change of temperature over time and calculate the temperature rise rate; Local temperature difference: During the charging or discharging process, record the temperature difference at different positions.

[0028] Differential / integral characteristics: Incremental capacity peak voltage: By recording the voltage and capacity during the charge and discharge process, calculate the voltage corresponding to the incremental capacity peak; Differential voltage curvature: By recording the voltage and time during the charge and discharge process, calculate the curvature of the voltage curve.

[0029] Including incremental capacity peak voltage, differential voltage curvature; Statistical distribution characteristics: Standard deviation of voltage fluctuation: By recording the change of voltage over time, calculate the standard deviation of voltage fluctuation; Temperature-current covariance: By recording the change of temperature and current over time, calculate the covariance of temperature and current.

[0030] Dynamic process characteristics: Pulse response recovery time: By recording the voltage or current recovery time of the battery after pulse charge and discharge; Relaxation process time constant: By recording the change of voltage or current of the battery during the rest stage, calculate the time constant of the relaxation process.

[0031] Special working condition characteristics: Low-temperature charging voltage overshoot: Under low-temperature conditions, record the voltage overshoot during the charging process; High-temperature discharge cut-off advance: Under high-temperature conditions, record the cut-off voltage advance during the discharge process.

[0032] Time series correlation characteristics: Voltage-temperature hysteresis angle: By recording the change of voltage and temperature over time, calculate the hysteresis angle between voltage and temperature; Capacity-temperature coupling coefficient: By recording the changes of capacity and temperature over time, the coupling coefficient between capacity and temperature is calculated.

[0033] In S2, the CNN-BiLSTM model consists of an input layer, a CNN convolutional block, a BiLSTM hidden layer, a fully connected layer, and an output layer. The output layer is a regression layer, where: The CNN convolutional block contains a convolutional layer with a kernel size of 1×1 and 32 kernels. The weight initialization method of the convolutional layer is He initialization (setting a smaller initial weight value to help the network converge faster during training and avoid the problems of gradient vanishing or gradient explosion). The padding mode is same (i.e., the height and width of the output are the same as those of the input), and the ELU activation function is adopted. After the convolutional layer, there is a batch normalization layer and an average pooling layer, and the pooling stride is set to FiltZise; The BiLSTM hidden layer includes a bidirectional LSTM layer with 128 nodes, and its weight initialization also uses He initialization. The 128-node bidirectional LSTM layer is followed by a 32-node bidirectional LSTM layer.

[0034] In S3, the LightGBM model consists of multiple decision trees based on gradient boosting. By means of ensemble learning, it continuously optimizes the fitting ability for complex features, thus achieving efficient and accurate estimation; Set the number of leaf nodes of the LightGBM model to 4, use gradient boosting decision trees as the boosting type, and at the same time set the feature selection ratio to 0.9, the sample sampling ratio for tree building to 0.8, and the frequency to 7 (i.e., the model will re-sample the samples every 7 iterations to increase the generalization ability of the model). The learning rate is set to 0.1 to accelerate the model convergence process. The number of threads during training is 1, which ensures the stability and controllability in a single-threaded environment. To avoid overfitting, the model adds an early stopping strategy and sets it to 5. If there is no obvious performance improvement within 5 rounds, the model will stop training.

[0035] In S3, in the CNN-BiLSTM model, CNN extracts local features through convolution operations, and BiLSTM processes the time series twice through forward and backward LSTM. The final output is the concatenation of two hidden states: ; In the formula, is the output of the forward LSTM; is the output of the backward LSTM; is the concatenated output; Infer SOH from the hidden state through the fully connected layer, which is expressed as: ; Wherein, is the estimated value of SOH; is the weight matrix; is the bias term; The LightGBM model integrates multiple weak learners, expressed as: ; Wherein, is the v-th weak learner, V is the number of weak learners, and each weak learner is a decision tree model; represents the output of the v-th weak learner; represents the final estimated result, that is, the sum of the outputs of all weak learners; is the set space of all weak learners.

[0036] In S3, the LightGBM model consists of multiple decision trees based on gradient boosting. Under the framework of gradient boosting decision trees, the estimation process of the LightGBM model is as follows: The LightGBM model is initialized to a constant value, taking the mean of the historical battery SOC data. The initial LightGBM model is expressed as: ; Wherein, is the loss function; is the true value of the SOC of the lithium battery at the i-th moment, that is, the true value of the SOC of the i-th sample; N is the total number of samples; is the constant value used for initialization; At the k-th iteration, the model makes an estimation, calculates the gap between the estimated value and the true value, that is, the residual, represented by the gradient of the loss function: ; Wherein, represents the gradient of the i-th sample at the k-th iteration; represents the true value of the SOC of the u-th sample; is the estimated value of the SOC of the i-th sample in the previous iteration; Using the residual as the target, a new decision tree is constructed to fit the current residual, and according to the estimated result of, the SOC estimation model is updated; Repeat the above process, continuously fit the residual with a new decision tree and update the model, and finally obtain the estimated value of SOC through the superposition of multiple rounds of iterations : ; Wherein, P is the total number of iterations; is the learning rate.

[0037] In S4, model training is performed based on the multi-dimensional feature data and health factor data in S1. Before training, data augmentation is performed on the data participating in training, including at least one of time series translation, time series scaling, noise addition, data truncation, data inversion, augmentation based on charge-discharge curves, augmentation based on the generative adversarial network GAN, augmentation based on the variational autoencoder VAE, augmentation based on time series generation models, and augmentation based on data interpolation, where: The augmentation based on charge-discharge curves is to perform transformation such as translation, scaling, and rotation on the charge-discharge curves of the battery to generate new charge-discharge curve data; The augmentation based on GAN is to use GAN to generate new voltage, current, temperature data, and health factor data; The augmentation based on VAE is to use VAE to generate new voltage, current, temperature data, and health factor data; The augmentation based on time series generation models is to use LSTM or Transformer to generate new time series data to simulate the performance changes of the battery during different charge-discharge processes; The augmentation based on data interpolation is to use linear interpolation or spline interpolation methods to generate new voltage, current, and temperature data.

[0038] The verification process is as follows: Debug the activation functions, loss functions, and hyperparameters of each part to ensure the effective cooperation between modules of the model. Calculate the error between the estimated value and the measured value through the evaluation function, and quantitatively describe the error of the estimation method using the RMSE and maximum error MAXE (maxerror) of the estimation result.

[0039] Use the charging stage data of batteries No. 1 - 8 in the Oxford battery aging dataset for simulation experiments. After preliminary data analysis, this dataset is divided into two groups of battery data. The number of cycles of batteries No. 1, 2, 3, 7, and 8 is about 8000 times, and the number of cycles of batteries No. 4, 5, and 6 is about 5000 times. Because their charge-discharge conditions, models, and states are the same, battery No. 3 (Cell3) and battery No. 6 (Cell6) are selected for example analysis. The RMSE and MAXE of the estimated SOC are respectively as Figure 3 and Figure 4 shown.

[0040] From Figure 3It can be seen that when the SOH is relatively high, whether it is joint estimation or individual estimation, the RMSE of SOC estimation remains at a relatively low level. However, the error of joint estimation is always smaller than that of individual estimation. The average RMSE values of the joint estimation method and the individual estimation method adopted in this embodiment are as follows: for Cell3, they are 1.00% and 3.53% respectively; for Cell6, they are 1.20% and 4.00% respectively. This indicates that the joint estimation method based on LightGBM-CNN-BiLSTM proposed in this embodiment can provide more accurate SOC estimation compared to the individual estimation method.

[0041] As the SOH decreases, the error of individual estimation shows a significant upward trend, which indicates that the SOC estimation error becomes larger under low SOH conditions, verifying the correlation between SOC and SOH. While the joint estimation method can still maintain a relatively low RMSE under low SOH conditions, showing higher estimation accuracy.

[0042] Figure 4 The MAXE results in [relevant context] further confirm the above conclusion. Throughout the SOH decline cycle, the MAXE of individual SOC estimation is always greater than that of joint estimation. The maximum MAXE values of the joint estimation method and the individual estimation method are as follows: for Cell3, they are 6.18% and 13.46% respectively; for Cell6, they are 5.35% and 11.43% respectively. Especially after the SOH is lower than 90%, the error of individual estimation shows a significant upward trend, while the MAXE of joint estimation, although slightly fluctuating, generally remains at a relatively low level. This indicates that joint estimation performs more stably under extreme conditions, effectively reducing the peak error in SOC estimation.

[0043] It is worth noting that at certain SOH values (such as in the range of 80% to 90%), the MAXE of joint estimation shows a small increase. This may be due to the drastic change in the battery performance characteristics in this range, resulting in increased difficulty for the model to capture the battery state. However, compared with the increase in the error of individual estimation, the fluctuation range of joint estimation is still smaller, showing higher estimation accuracy.

[0044] Table 1 Comparison of Errors between Joint Estimation and Individual Estimation for Different Batteries

[0045] Table 1 presents the RMSE and MAXE results of the two estimation methods for each dataset respectively. It can be seen from Table 1 that: the mean RMSE of the joint estimation is 1.09%, and the mean RMSE of the individual estimation is 4.26%. This indicates that the error of the joint estimation is significantly smaller than that of the individual estimation overall, showing higher estimation accuracy. Among all battery cells, the average RMSE of the joint estimation is far lower than that of the individual estimation; the maximum MAXE of the joint estimation is 6.09%, and the maximum MAXE of the individual estimation is 12.60%. The MAXE results further prove the superiority of the joint estimation in reducing extreme errors. The maximum absolute error of the individual estimation is significantly higher, meaning that the individual estimation has a larger estimation bias at some data points.

[0046] To more intuitively reflect the superiority of the joint estimation method of this embodiment, the LSTM algorithm and the random forest algorithm (RF) which also uses decision trees as the basic building blocks are used as control experimental groups 1 and 2 respectively. The main experimental parameters of control experimental group 1 are set as: the number of trees is 800, and the minimum number of leaf nodes is 5; the main experimental parameters of control experimental group 2 are set as: the structure of the LSTM network consists of an input layer, 1 LSTM layer, ReLU activation layer, fully connected layer and regression layer. The Adam optimization algorithm is used for training parameters, the batch size is set to 30, the maximum number of iterations is 50, the initial learning rate is 0.01, and it is set that the learning rate decreases after every 800 iterations.

[0047] Figure 5 、 Figure 6 、 Figure 7 They are the SOC estimation values of Cell3 at three stages when the SOH value is 100%, 85%, and 75% during the charging cycle respectively. Figure 8 is the SOH estimation accuracy of Cell3. It can be clearly seen from Figures 5 - 8 that the joint estimation method based on CNN - BiLSTM - LightGBM used in this embodiment shows good estimation accuracy throughout the SOC estimation range. Especially when the SOC value is between 20% and 80%, its estimation results are closer to the actual values. This indicates that the joint estimation method adopted in this embodiment can effectively capture the dynamic changes of the battery SOC and reflects strong learning ability. However, there are certain fluctuations in the SOC estimation process of control experimental group 1 and control experimental group 2.

[0048] In control experiment group 1, when the SOC value was around 50%, the estimation error increased significantly. This might be because the RF model had limited ability to model non-linear relationships, especially in the case of complex changes in battery SOC, resulting in inaccurate estimation. In control experiment group 2, large error values occurred under different SOH states. This might be because the LSTM model needed a large amount of high-quality training data to perform well. If the data volume was insufficient or the data quality was poor (such as a lot of noise and missing values), the model performance would decrease significantly. At the same time, from Figure 8 it can be seen that the estimation error was larger when the number of cycles was small or large, and the corresponding SOC estimation error also increased simultaneously. When the number of cycles was medium, that is, when the SOH value was around 50%, the SOH estimation error was smaller, and its SOC estimation error was also smaller. This further verified the correlation between SOC and SOH mentioned in Section 2.

[0049] To further highlight the superiority of the SOC and SOH joint estimation method adopted in this embodiment, taking Cell3 as an example, when the model reached the best effect compared with the control group, the model training time T Train and the goodness of fit R 2 are shown in Table 2 as follows.

[0050] Table 2 Comparison of training time and goodness of fit of different algorithms

[0051] From the perspective of training time, the joint estimation method adopted in this embodiment significantly shortened the model training time, enabling rapid response and real-time estimation. This is crucial for the practical application of energy storage systems. Fast SOC estimation can reduce the charge and discharge delay of energy storage devices, avoid overuse or idleness of batteries caused by estimation delay, and thus improve the overall efficiency and economy of the system. In terms of the evaluation of the R2 value, the performance of the method in this embodiment is also better than that of the control group.

[0052] In summary, the SOC and SOH joint estimation method adopted in this embodiment not only performs excellently in estimation accuracy, but also is significantly higher in computational efficiency than the traditional RF and LSTM models, indicating its application potential in energy storage systems.

[0053] Example 2: Based on the previous example, in this example, for the SOH estimation value and SOC estimation value in the joint estimation value, weighted fusion is performed to obtain a fusion estimation value for evaluating the overall state of the lithium battery. The method is to construct a two-stream attention fusion network through dynamic attention fusion to dynamically generate the spatio-temporal adaptive weights of the SOH estimation value output by the CNN-BiLSTM model and the SOC estimation value output by the LightGBM model. The process is as follows: Step 1: Concatenate the hidden state vector of the last time step of the BiLSTM hidden layer of the CNN-BiLSTM model with the leaf node distribution vector of the final decision tree of the LightGBM model to form a cross-model feature interaction matrix , where is the hidden state dimension, is the leaf node dimension; Step 2: Construct an attention weight generation module that includes a time decay factor and spatial correlation: ; In the formula, represents the attention weight; is the query matrix, is the weight matrix of the query matrix; is the key matrix, is the weight matrix of the key matrix; is the dimension of the key matrix; is the learnable time decay coefficient; is the time difference between the current moment and the historical reference moment; log is the logarithm symbol; Step 3: Calculate the probability distribution divergence of the outputs of the two models respectively : ; In the formula, is used to quantify the difference between the probability distributions of the outputs of the two models; represents the KL divergence; is the probability distribution output by the CNN-BiLSTM model; is the probability distribution output by the LightGBM model; Step 4: Construct a gated fusion mechanism to generate the final weight coefficient: ; ; In the formula, is the weight of the SOH estimation value; is the weight of the SOC estimation value; represents the sigmoid function; is the trainable parameter matrix; represents the Hadamard product; Step 5: Generate a fusion estimation value using a weighted fusion method with residual connection : ; In the formula, is the residual coefficient, ​​It is a non - linear correction network composed of a multi - layer perceptron.

[0054] In the above process, by introducing cross - model feature interaction, it breaks through the traditional simple weighted fusion mode; designs a spatio - temporal attention mechanism to capture both time - decay and spatial - correlation features simultaneously; adds KL - divergence uncertainty calibration to quantify the degree of model prediction consistency; adopts a non - linear fusion structure with residuals to retain the dominant features of single models; and realizes end - to - end weight learning through a differentiable gating mechanism instead of manually setting rules.

[0055] The combined estimated values of SOH and SOC are weighted and output, and the weighted output value (fusion estimated value ) can be used as a comprehensive index to reflect both the state of health (SOH) and state of charge (SOC) of the battery. For example, the weighted output value can be used in a battery management system (BMS) to more comprehensively evaluate the overall state of the battery, so as to make more reasonable energy scheduling and maintenance decisions. The weighted output value can also be used to optimize the charge - discharge control strategy of the battery. For example, when the state of health of the battery is low, the charge - discharge rate of the battery can be reduced by adjusting the weighted output value to protect the battery; or it can be used for fault diagnosis and warning. For example, when the weighted output value exceeds the normal range, an alarm can be issued in advance to prompt the user to perform maintenance or replace the battery. The weighted output value can also serve as a method of data fusion and dimensionality reduction, integrating multiple state parameters into a single index to simplify the data processing and analysis process.

[0056] Embodiment 3: A device for jointly estimating the SOC and SOH of lithium batteries in a park, comprising a memory, a processor, and a computer program stored on the memory and capable of running on the processor. The method in Embodiment 1 or Embodiment 2 is implemented by the processor executing the computer program.

Claims

1. A method for jointly estimating SOC and SOH of lithium batteries in a park, characterized in that The following steps are involved: S1. Collect multi-dimensional feature data of lithium batteries during the charging and discharging process, extract health factors, and pre-process the data; S2, build CNN-BiLSTM model; S3. Combine the CNN-BiLSTM model with the LightGBM model to build a joint estimation model, in which the SOH estimation uses the CNN-BiLSTM model and the SOC estimation uses the LightGBM model; S4, training CNN-BiLSTM model and LightGBM model; S5. Based on the trained joint estimation model, the health factor is input into the CNN-BiLSTM model to estimate SOH, and the multi-dimensional feature data and the estimated value of SOH are input into the LightGBM model to estimate SOC. The outputs of the CNN-BiLSTM model and the LightGBM model are combined to obtain the final joint estimation value of SOC and SOH.

2. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In S1, the multi-dimensional feature data is the current of the lithium battery [I1, I2, ..., I n ], voltage [U1,U2,…,U n ] and the temperature [T1,T2,…,T n ], n is the number of data points, that is, the data corresponding to different time points; The extracted health factors are [HF1, HF2,…, HF m ] J , m is the number of health factor data, J represents the number of types of health factors; Preprocessing includes missing value processing, outlier processing, data cleaning and normalization.

3. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: Health factors include: Charging stage characteristics, including constant current charging time, current decay rate in constant voltage stage, charging voltage platform slope, and charging termination voltage; Discharge stage characteristics, including discharge voltage drop gradient and effective discharge capacity; Static phase characteristics, including relaxation voltage recovery rate and open circuit voltage decay; Temperature dynamic characteristics, including constant current charging temperature rise rate and local temperature difference; Differential / integral characteristics, including incremental capacity peak voltage, differential voltage curvature; Statistical distribution characteristics, including voltage fluctuation standard deviation and temperature-current covariance; Dynamic process characteristics, including impulse response recovery time and relaxation process time constant; Special operating conditions, including low-temperature charging voltage overshoot and high-temperature discharge cutoff advance; Timing correlation characteristics, including voltage-temperature hysteresis angle and capacity-temperature coupling coefficient.

4. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In S2, the CNN-BiLSTM model consists of an input layer, a CNN convolutional block, a BiLSTM hidden layer, a fully connected layer, and an output layer, where the output layer is a regression layer, wherein: The CNN convolution block contains a convolution layer with a convolution kernel size of 1×1 and 32 convolution kernels. The convolution layer weight initialization method is He initialization, the filling mode is same, and the ELU activation function is used. The convolution layer is followed by a batch normalization layer and an average pooling layer, and the pooling step size is set to FiltZise. The BiLSTM hidden layer includes a 128-node bidirectional LSTM layer, whose weight initialization is also He initialization. The 128-node bidirectional LSTM layer is followed by a 32-node bidirectional LSTM layer.

5. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In the above S3, the LightGBM model consists of multiple decision trees based on gradient boosting, which continuously optimizes the fitting ability of complex features through ensemble learning, thereby achieving efficient and accurate estimation; Set the number of leaf nodes of the LightGBM model to 4, use the gradient boosting decision tree as the boosting type, set the feature selection ratio to 0.9, the sample sampling ratio for tree building to 0.8, the frequency to 7, and the learning rate to 0.

1.

6. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In the CNN-BiLSTM model in S3, CNN extracts local features through convolution operations, and BiLSTM performs two LSTM processes on the time series, forward and reverse. The final output is the concatenation of two hidden states: ； In the formula, is the output of the forward LSTM; is the output of the reverse LSTM; is the output after splicing; Through the fully connected layer from the hidden state The SOH is inferred in , which is expressed as: ； In the formula, is the estimated value of SOH; is the weight matrix; is the bias term; The LightGBM model integrates multiple weak learners, expressed as: ； In the formula, is the vth weak learner, V is the number of weak learners, and each weak learner is a decision tree model; represents the output of the vth weak learner; Represents the final estimation result, which is the sum of the outputs of all weak learners; is the collection space of all weak learners.

7. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In the above S3, the LightGBM model consists of multiple decision trees based on gradient boosting. Under the framework of gradient boosting decision tree, the estimation process of the LightGBM model is as follows: The LightGBM model is initialized to a constant value, taking the mean of the battery SOC historical data, and the initial LightGBM model It is expressed as: ； In the formula, is the loss function; is the true SOC value of the lithium battery at the i-th moment, that is, the true SOC value of the i-th sample; N is the total number of samples; is the constant value used for initialization; At the kth iteration, the model makes an estimate and calculates the difference between the estimate and the true value, i.e. the residual, which is represented by the gradient of the loss function: ； In the formula, represents the gradient of the i-th sample in the k-th iteration; Represents the true value of SOC of the u-th sample; is the estimated SOC value of the i-th sample in the previous iteration; Using residuals As the goal, build a new decision tree Fit the current residuals, according to The estimation result is used to update the SOC estimation model; Repeat the above process, continuously fit the residuals through the new decision tree and update the model, and finally obtain the estimated value of SOC through the superposition of multiple rounds of iterations. : ； Where P is the total number of iterations; is the learning rate.

8. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In the above S4, model training is performed based on the multidimensional feature data and health factor data in S1. Before training, data enhancement is performed on the data involved in the training, including at least one of time series translation, time series scaling, noise addition, data truncation, data inversion, enhancement based on charge and discharge curves, enhancement based on generative adversarial networks GAN, enhancement based on variational autoencoders VAE, enhancement based on time series generation models, and enhancement based on data interpolation, wherein: Enhancement based on the charge and discharge curve is to translate, scale, and rotate the charge and discharge curve of the battery to generate new charge and discharge curve data; GAN-based enhancement uses GAN to generate new voltage, current, and temperature data, as well as health factor data; VAE-based enhancements use VAE to generate new voltage, current, and temperature data, as well as health factor data; The enhancement based on the time series generation model is to use LSTM or Transformer to generate new time series data to simulate the performance changes of the battery during different charging and discharging processes; Enhancements based on data interpolation use linear or spline interpolation methods to generate new voltage, current, and temperature data.

9. The method for jointly estimating SOC and SOH of a lithium battery in a park according to claim 1, characterized in that: In the above S5, the SOH estimate and the SOC estimate in the joint estimate are weightedly fused to obtain a fused estimate for evaluating the overall state of the lithium battery. The method is to construct a dual-stream attention fusion network through dynamic attention fusion, and dynamically generate the spatiotemporal adaptive weights of the SOH estimate output by the CNN-BiLSTM model and the SOC estimate output by the LightGBM model. The process is: Step 1: The hidden state vector of the last time step of the BiLSTM hidden layer of the CNN-BiLSTM model The leaf node distribution vector of the final decision tree of the LightGBM model Perform tensor splicing to form a cross-model feature interaction matrix ,in is the hidden state dimension, is the leaf node dimension; Step 2: Construct an attention weight generation module that includes time decay factor and spatial correlation: ； In the formula, represents the attention weight; is the query matrix, is the weight matrix of the query matrix; is the key matrix, is the weight matrix of the key matrix; is the dimension of the key matrix; is the learnable time decay coefficient; The time difference between the current time and the historical reference time; Step 3: Calculate the probability distribution divergence of the two model outputs respectively : ； In the formula, Used to quantify the difference between the probability distributions of two model outputs; represents KL divergence; The probability distribution of the output of the CNN-BiLSTM model; The probability distribution output by the LightGBM model; Step 4: Construct a gated fusion mechanism to generate the final weight coefficient: ； ； In the formula, is the weight of the SOH estimate; is the weight of the SOC estimate; Represents the sigmoid function; is the trainable parameter matrix; represents the Hadamard product; Step 5: Generate fusion estimate using weighted fusion with residual connection : ； In the formula, is the residual coefficient, A nonlinear correction network composed of a multi-layer perceptron; is the estimated value of SOH; is the estimated value of SOC.

10. A joint estimation device for SOC and SOH of lithium batteries in a park, characterized by: The method comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the method according to any one of claims 1 to 9 is implemented by executing the computer program by the processor.

Images