A secondary battery SOH prediction model based on Adamom optimizer
The SOH prediction method, which combines the Adamom optimizer and the sample convolution model, solves the problems of model complexity and high computational cost in existing technologies, and achieves efficient and accurate lithium-ion battery health state estimation, which is applicable to fields such as electric vehicles.
Patent Information
- Application Number
- CN202411838680.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-10-28
- Estimated Expiration
- 2044-12-13
AI Technical Summary
Existing technologies for predicting SOH in lithium-ion batteries suffer from problems such as complex models, large computational load, and reliance on prior knowledge of battery degradation. Furthermore, existing optimizers are prone to instability or slow convergence when parameters are not set properly.
A secondary battery SOH prediction model based on the Adamom optimizer is adopted. By extracting influencing factors from direct observation data, combining sample convolution model and Pearson correlation analysis, the model parameters are optimized using an improved Adamom optimizer, and a momentum mechanism is introduced to accelerate convergence and reduce oscillations.
It significantly improves the accuracy and efficiency of SOH prediction for lithium-ion batteries, reduces computational complexity, adapts to different operating conditions, reduces reliance on prior knowledge, and improves the optimization and safety of the battery management system.
Smart Images

Figure CN119667490B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of secondary battery technology, specifically to a secondary battery SOH prediction model based on the Adamom optimizer. Background Technology
[0002] Lithium-ion batteries (LiBs) are widely used in electric vehicles, mobile electronic products, large-scale energy storage systems, aerospace, and other fields due to their low self-discharge rate, long cycle life, high energy density, high power density, absence of memory effect, and environmental friendliness. However, under long-term cycling, the irreversible consumption of lithium ions and the loss of positive / negative electrode active materials lead to a certain degree of performance degradation, mainly manifested as capacity decay, increased internal resistance, and decreased power. This performance degradation can lead to system failures or even serious accidents, such as fires or explosions. State of Balance (SOH) represents the ratio of a battery's current maximum capacity to its rated capacity, reflecting the battery's current ability to store and supply energy relative to its initial capacity at the beginning of its lifespan. It is a key indicator of battery degradation. Accurate SOH estimation helps in proactive maintenance, optimal utilization, and effective planning and management of batteries, thereby improving the long-term sustainability of lithium-ion batteries in electric vehicles.
[0003] In recent years, existing technologies have conducted extensive and in-depth research on the prediction of the state of health (SOH) of LiBs, and various real-time methods have been proposed, mainly including model-based methods, data-driven methods, and fusion methods. Model-based methods, by establishing battery models (such as equivalent circuit models and electrochemical models), can directly monitor current, voltage, and battery surface temperature in real time, and combine this with the material and electrochemical characteristics of lithium-ion batteries to achieve accurate prediction of the battery's state of health. Electrochemical models, in particular, simulate the physical and electrochemical reaction processes at various stages within the battery by constructing nonlinear coupled partial differential equations, including material changes, temperature distribution, stress changes, and lithium-ion migration, thereby predicting the battery's SOH. However, in practical applications, this is limited by the model's versatility and the large number of parameters required. Equivalent circuit models use a combination of electronic components such as voltage sources, capacitors, and resistors to simulate the charging and discharging processes within the battery. Compared to electrochemical models, equivalent circuit models have a simpler architecture, lower computational complexity, and do not require a large amount of computational resources. However, equivalent circuit models cannot describe all aging dynamics, and their accuracy depends on the constructed model and the determined parameters. For the complex electrochemical reactions and physical processes in lithium-ion batteries, model-based methods are often complex and difficult to implement. In contrast, data-driven methods do not need to consider the internal electrochemical reactions and failure mechanisms of lithium-ion batteries and can directly estimate the battery's state of harm (SOH) using observational data. This characteristic has made it one of the more popular research methods in recent years. Although data-driven methods can automatically learn and adapt to complex battery degradation processes, they also have some significant drawbacks, such as high dependence on the quality and quantity of data, high computational costs when the model is complex, and lack of interpretability.
[0004] Therefore, understanding how to directly characterize the health status of a battery through online monitoring data is of great guiding significance for secondary battery management systems. Summary of the Invention
[0005] To address the aforementioned shortcomings of existing technologies, this application aims to provide a secondary battery SOH prediction model based on the Adamom optimizer. This model analyzes and extracts influencing factors characterizing battery capacity degradation from directly observed data, verifies the effectiveness of health factors through correlation analysis, and applies the Adamom optimizer and sample convolution model to estimate the state of health of lithium batteries. This model can significantly improve prediction accuracy and efficiency, and the health factor sequence data comes from data directly monitored by the BMS, effectively solving the problems of complex model structure, large computational load, and reliance on prior knowledge of battery degradation in existing technologies.
[0006] To solve the above-mentioned technical problems, this application adopts the following technical solution:
[0007] A secondary battery SOH prediction model based on the Adamom optimizer specifically includes the following steps:
[0008] Step 1: Obtain the feature sequence and capacity sequence of the secondary battery, and then normalize them respectively; wherein, the feature sequence includes constant current charging time, constant voltage charging time, and voltage rise value within the same time period;
[0009] Step 2: Perform correlation analysis on the normalized feature sequence and capacity sequence, select the feature sequence with strong correlation to the capacity sequence as the health factor sequence, and then divide the obtained health factor sequence into training set, validation set and test set according to a certain ratio.
[0010] Step 3: Construct a secondary battery SOH prediction model based on the sample convolution model, and train the prediction model using the training set obtained in Step 2; during the training process, use the Adamom optimizer to optimize the parameters of the prediction model, and use the validation set obtained in Step 2 to validate the model, and select the parameter combination with the smallest error as the final model parameters.
[0011] Step 4: Test the prediction model obtained in Step 3 using the test set obtained in Step 2 to obtain the estimated value of SOH. Then, inverse normalize the estimated value of the cycle capacity to obtain the predicted value of the cycle capacity.
[0012] Step 5: Determine the neural network prediction results based on the root mean square error (RMSSE), and then output, optimize, and verify them.
[0013] Preferably, the Adamom optimizer is obtained by improving the Adadelta optimizer. At time t, the Adamom optimizer performs the following update calculation, with the specific steps as follows:
[0014] (1) Initialize the cumulative variables:
[0015] E[g 2 ]0=0 (4)
[0016] E[Δx 2 ]0=0 (5)
[0017] Among them, E[g 2 ]0 represents the cumulative squared gradient at the initial time; E[Δx 2 ]0 represents the cumulative squared parameter updates at the initial time step;
[0018] (2) For each parameter, at time t, the exponentially weighted average of the squared gradient is calculated using the decay factor α.
[0019]
[0020] Where α is the attenuation factor, with a default value of 0.9;
[0021] (3) For each parameter, at time t, the exponentially weighted average of the parameter update squares is calculated using the decay factor β;
[0022]
[0023] β is the decay factor, with a default value of 0.9;
[0024] (4) Update the momentum by adding some information from the current gradient based on the previous time step:
[0025] v t =γv t-1 +(1-γ)g t (8)
[0026] In equation (8), v t-1 and v t These represent the momentum at the previous and current moments, respectively; γ represents the decay factor, with a default value of 0.9; g t The gradient represents the current time step;
[0027] (5) Using the previously calculated E[g 2 ] t and E[Δx 2 ] t To calculate the parameter update amount for this iteration:
[0028]
[0029] x t+1 =x t +Δx t (10)
[0030] In equations (9) to (10), x t+1 and x t These represent the weights of the next time step and the current time step, respectively; the default value is 1e. -6 .
[0031] Preferably, in step 1, the normalization calculation formula for the feature sequence and capacity sequence is as follows:
[0032] Formula 1: N(x)=(x-min(x)) / (max(x)-min(x));
[0033] In Equation 1: x represents the sequence data, min(x) represents the minimum value in the sequence, and max(x) represents the maximum value in the sequence.
[0034] Preferably, in step 2, the correlation analysis is performed using the Pearson correlation coefficient, and a correlation is considered strong when the absolute value is greater than or equal to 0.8.
[0035] Preferably, the formula for calculating the Pearson correlation coefficient is:
[0036] Formula 2:
[0037] In Equation 2: R(x,y) represents the magnitude of the Pearson coefficient between x and y; and R(x,y) represents the average values of x and y, respectively. The range of R(x,y) is between -1 and 1. The closer the absolute value of R(x,u) is to 1, the stronger the linear correlation between x and y.
[0038] Preferably, the process of establishing the secondary battery SOH prediction model is as follows: constructing a sample convolutional neural network model and an Adamom optimizer, and inputting the training set into the model for training.
[0039] Preferably, the ratio of the data volume of the training set, validation set, and test set is 5:2:3, 6:2:2, and 7:2:1; the training set, validation set, and test set need to undergo format processing before being input into the model, and the format processing process is as follows:
[0040] Let the sliding window size be L. Take the health factor sequence from the 1st to the Lth cycle as input to obtain the data for the (L+1)th cycle. Then take the health factor sequence from the 2nd to the (L+1)th cycle as input to the model to predict the data for the (L+2)th cycle, until the entire training or test set has been processed.
[0041] Preferably, the format of the input samples for the training set, validation set, and test set convolutional neural network is as follows:
[0042] Equation 3: x i =[x i ,x i+1 ,x i+2 …x i+L ], i = 1, 2, 3…NL;
[0043] In Equation 3: N is the total number of samples in the dataset, and L is the input length.
[0044] Preferably, the calculation process of the root mean square error (RMSE) is as follows:
[0045] Formula 4:
[0046] In Equation 4: c i This represents the actual value from the capacity test. Here, N represents the predicted capacity value, and N is the total number of data samples.
[0047] Preferably, the process of judging the neural network prediction results based on the root mean square error (RMSE) and outputting, optimizing and verifying them is as follows: the hyperparameters of the sample convolutional neural network are adjusted using the control variable method, steps 3 and 4 are repeated, and the set of hyperparameters with the lowest root mean square error is selected as the hyperparameters of the model.
[0048] This application also proposes an application of a secondary battery SOH prediction model based on the Adamom optimizer, which is used to detect the health status of secondary batteries.
[0049] Compared with the prior art, this application has the following beneficial effects:
[0050] 1. The secondary battery SOH prediction model provided in this application analyzes and extracts influencing factors that can characterize battery capacity degradation from directly observable data, verifies the effectiveness of health factors through Pearson correlation analysis, and applies the sample convolution model to estimate the health status of lithium batteries. This model can significantly improve the accuracy and efficiency of prediction, and the health factor sequence data comes from data directly monitored by the battery management system, which can effectively solve the problems of complex model structure, large amount of computation, and the need to use prior knowledge of battery degradation in the prior art.
[0051] 2. The technical solution provided in this application improves upon the existing Adadelta optimizer, combining the advantages of Adadelta and Momentum to design a new optimizer: the Adamom optimizer. This optimizer optimizes the parameters of the sample convolutional model and introduces a momentum mechanism during parameter updates to reduce oscillations on the optimization path and accelerate convergence. The Adamom optimizer designed in this application has advantages such as not requiring manual learning rate settings, being insensitive to hyperparameters, and providing independent dynamic learning rates for each dimension. Compared with the existing Adadelta optimizer, the Adamom optimizer can more effectively guide the model to smaller local optima.
[0052] 3. The Adamom optimizer and sample convolution model provided in this application can significantly improve the accuracy and efficiency of SOH prediction for lithium-ion batteries, thereby providing strong technical support for the optimization of battery management systems, extending battery life and improving energy use safety, and has important application significance. Attached Figure Description
[0053] Figure 1 This is a flowchart of the technical solution of the model provided in Embodiment 1 of this application.
[0054] Figure 2 This is a network structure diagram of the model provided in Embodiment 1 of this application.
[0055] Figure 3 This is the default architecture of the convolution module in the sample convolution model used in Embodiment 1 of this application.
[0056] Figure 4 The Adamom optimizer is updated at time t in Embodiment 1 of this application.
[0057] Figure 5 The data sets B0005 and B0006 used in Embodiment 1 of this application are shown as the decay curves of the true health status (SOH) with the number of cycles.
[0058] Figure 6 In Embodiment 1 of this application, the loss function (mean squared error MSE) changes with the number of training iterations during the training process of the sample convolutional neural networks based on different optimizers after the B0006 dataset is divided in a 6:2:2 ratio. Detailed Implementation
[0059] This application will clearly and completely describe the technical solutions of the embodiments of this application with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.
[0060] Unless otherwise specified in the specific circumstances, the numerical ranges listed herein include upper and lower limits, as well as all integers and fractions within that range, but are not limited to the specific values listed when the range is defined.
[0061] I. A secondary battery SOH prediction model based on Adamom optimizer
[0062] Step 1: Obtain the feature sequence and capacity sequence of the secondary battery, and then normalize them respectively; wherein, the feature sequence includes constant current charging time, constant voltage charging time, and voltage rise value within the same time period;
[0063] Step 2: Perform correlation analysis on the normalized feature sequence and capacity sequence, select the feature sequence with strong correlation to the capacity sequence as the health factor sequence, and then divide the obtained health factor sequence into training set, validation set and test set according to a certain ratio.
[0064] Step 3: Construct a secondary battery SOH prediction model based on the sample convolution model, and train the prediction model using the training set obtained in Step 2; during the training process, use the Adamom optimizer to optimize the parameters of the prediction model, and use the validation set obtained in Step 2 to validate the model, and select the parameter combination with the smallest error as the final model parameters.
[0065] Step 4: Test the prediction model obtained in Step 3 using the test set obtained in Step 2 to obtain the estimated value of SOH. Then, inverse normalize the estimated value of the cycle capacity to obtain the predicted value of the cycle capacity.
[0066] Step 5: Determine the neural network prediction results based on the root mean square error (RMSSE), and then output, optimize, and verify them.
[0067] In some embodiments of this application, it was found during the research that directly using existing optimizers presents problems. For example, Adadelta can address the monotonically decreasing learning rate problem in Adagrad by limiting the window size of the accumulated gradient. However, Adadelta still relies on a preset learning rate, and its performance largely depends on the selection of hyperparameters, such as the setting of the decay factor. If these parameters are set improperly, the optimization process may become unstable or converge slowly. Furthermore, while Adadelta can use an exponentially weighted moving average to calculate the accumulated gradient, discarding distant gradient history information and thus avoiding the problems of premature and excessive reduction of the learning rate in AdaGrad, it may still experience slow convergence when dealing with complex optimization problems, especially when facing highly non-convex loss functions, easily getting trapped in local optima. Moreover, although Adadelta introduces an exponentially weighted average of the squared gradient, it does not directly address the fluctuations in the gradient itself. Therefore, Adadelta's performance is unsatisfactory when facing drastic gradient changes. Building upon this foundation, this application redesigns the optimizer and proposes an Adamom optimizer, an adaptive learning rate optimization algorithm designed to address the learning rate selection problem in the traditional Adam optimizer and further improve optimization efficiency and convergence speed. The Adamom optimizer is an improvement upon the Adadelta optimizer and also incorporates the advantages of Momentum. This application further introduces a momentum mechanism, which, by incorporating a weighted average of past gradients, allows the model weight updates to consider gradient information over a longer time frame, thereby mitigating the impact of gradient fluctuations, avoiding getting trapped in local minima, and making the optimization process more stable and converging faster. Therefore, in this application, the Adamom optimizer is obtained by improving the Adadelta optimizer. At time t, the Adamom optimizer performs the following update calculation, with the specific steps as follows:
[0068] (1) Initialize the cumulative variables:
[0069] E[g 2 ]0=0 (4)
[0070] E[Δx 2 ]0=0 (5)
[0071] Among them, E[g 2 ]0 represents the cumulative squared gradient at the initial time; E[Δx 2 ]0 represents the cumulative squared parameter updates at the initial time step;
[0072] (2) For each parameter, at time t, the exponentially weighted average of the squared gradient is calculated using the decay factor α.
[0073]
[0074] Among them, E[g 2 ] t-1 and E[g 2 ] t These represent the cumulative squared gradients at the previous and current time steps, respectively. This represents the squared gradient at the current time, where α is the decay factor, with a default value of 0.9.
[0075] (3) For each parameter, at time t, the exponentially weighted average of the parameter update squares is calculated using the decay factor β;
[0076]
[0077] Where, E[Δx 2 ] t-2 and E[Δx 2 ] t These represent the cumulative squares of the parameter changes at the previous and current times, respectively. This represents the square of the parameter change at the previous moment, where β is the decay factor with a default value of 0.9.
[0078] (4) Update the momentum by adding some information from the current gradient based on the previous time step:
[0079] v t =γv t-1 +(1-γ)g t (8)
[0080] In equation (8), v t-1 and v t These represent the momentum at the previous and current moments, respectively; γ represents the decay factor, with a default value of 0.9; g t The gradient represents the current time step;
[0081] (5) Using the previously calculated E[g 2 ] t and E[Δx 2 ] t To calculate the parameter update amount for this iteration:
[0082]
[0083] x t+1 =x t +Δx t (10)
[0084] In equations (9) to (10), x t+1 and x t These represent the weights of the next time step and the current time step, respectively; the default value is 1e. -6 .
[0085] In some embodiments of this application, in step 1, the calculation formula for the normalization processing of the feature sequence and the capacity sequence is as follows:
[0086] Formula 1: N(x)=(x-min(x)) / (max(x)-min(x));
[0087] In Equation 1: x represents the sequence data, min(x) represents the minimum value in the sequence, and max(x) represents the maximum value in the sequence.
[0088] In some embodiments of this application, in step 2, the correlation analysis uses the Pearson correlation coefficient. A correlation is considered strong when its absolute value is greater than or equal to 0.8. The formula for calculating the Pearson correlation coefficient is:
[0089] Formula 2:
[0090] In Equation 2: R(x,y) represents the magnitude of the Pearson coefficient between x and y; and R(x,y) represents the average values of x and y, respectively. The range of R(x,y) is between -1 and 1. The closer the absolute value of R(x,u) is to 1, the stronger the linear correlation between x and y.
[0091] In some embodiments of this application, the process of establishing a secondary battery SOH prediction model is as follows: constructing a sample convolutional neural network model and an Adamom optimizer, and inputting the training set into the model for training. This application found that in the prior art, there are three main types of deep neural networks used for sequence modeling, all of which are applied to time series prediction (TSF): (i) recurrent neural networks (RNNs) and their variants; (ii) Transformer models; and (iii) temporal convolutional networks (TCNs). Recently, one-dimensional convolutional neural networks (1D-CNNs) have gradually become a replacement for long short-term memory (LSTM) networks in the field of time series prediction. This specific network architecture can effectively extract local features from continuous data and capture temporal dependencies by using sliding convolutional kernels. Compared with Transformer and TCN models, 1D-CNNs are more computationally efficient and can better map input data to output data by fully utilizing the temporal features in the time series through convolutional kernels. In traditional typical one-dimensional convolutional neural network architectures, only one type of convolutional kernel is usually used in the same layer, tending to extract average features. However, in complex time-series data, a large number of different temporal features need to be extracted, and the traditional convolutional neural network architecture in existing technologies cannot meet this requirement. Therefore, this application proposes a novel one-dimensional convolutional neural network model—the sample convolution model. The sample convolution model adopts an encoder-decoder architecture, using convolution kernels of different sizes to analyze the charging and discharging data of secondary batteries, thereby capturing the different time-scale features of battery state changes and improving the accuracy of SOH prediction.
[0092] In some embodiments of this application, the data ratio of the training set, validation set, and test set is 5:2:3, 6:2:2, and 7:2:1; the training set, validation set, and test set need to undergo format processing before being input into the model, and the format processing process is as follows:
[0093] Let the sliding window size be L. Take the health factor sequence from the 1st to the Lth cycle as input to obtain the data for the (L+1)th cycle. Then take the health factor sequence from the 2nd to the (L+1)th cycle as input to the model to predict the data for the (L+2)th cycle, until the entire training or test set has been processed.
[0094] The training set, validation set, and test set input samples of the convolutional neural network are in the following format:
[0095] Equation 3: x i =[x i ,x i+1 ,x i+2 …x i+L ], i = 1, 2, 3…NL;
[0096] In Equation 3: N is the total number of samples in the dataset, and L is the input length.
[0097] Preferably, the calculation process of the root mean square error (RMSE) is as follows:
[0098] Formula 4:
[0099] In Equation 4: c i This represents the actual value from the capacity test. Here, N represents the predicted capacity value, and N is the total number of data samples.
[0100] The process of judging the neural network prediction results based on the root mean square error (RMSE) and outputting, optimizing and verifying them is as follows: the hyperparameters of the sample convolutional neural network are adjusted using the control variable method, steps 3 and 4 are repeated, and the set of hyperparameters with the lowest RMSE is selected as the hyperparameters of the model.
[0101] II. A method for predicting the health status of lithium-ion batteries
[0102] The model described in this application is used for predicting the health status of secondary batteries, and is not limited to lithium-ion batteries. Lithium-ion batteries are used as an example here for illustration.
[0103] like Figure 1 As shown, step S1 involves analyzing and extracting the constant current charging time CCCT = {ccct(k) | k = 1, 2, 3…n}, constant voltage charging time CVCT = {cvct(k) | k = 1, 2, 3…n}, voltage rise value RVCT = {rvct(k) | k = 1, 2, 3…n}, and lithium battery capacity sequence C = {c(k) | k = 1, 2, 3…n} from the battery charging and discharging data collected by the Battery Management System (BMS). Here, n is the total number of charge-discharge cycles of the lithium-ion battery. To unify the dimensions of different features and avoid the impact of dimensional differences on model training, the feature sequence and capacity sequence need to be normalized and scaled within the range of 0 to 1.
[0104] N(x)=(x-min(x)) / (max(x)-min(x)) (1)
[0105] In the formula, x represents the sequence data, min(x) represents the minimum value in the sequence, and max(x) represents the maximum value in the sequence.
[0106] Step S2: Perform Pearson correlation analysis on the normalized constant current charging time, constant voltage charging time, voltage rise within the same time period, and battery capacity sequences. Select the feature sequences with strong correlation to the battery capacity sequence as the normalized health factor sequences. Divide the normalized capacity sequence and normalized health factor sequence data into training datasets, validation datasets, and test datasets for health status prediction according to a certain ratio.
[0107] The Pearson correlation coefficient is calculated according to formula (2):
[0108]
[0109] In the formula, R(x,y) represents the magnitude of the Pearson coefficient between x and y; and Let x and y represent the average values of x and y, respectively. The value of R(x,y) ranges from [-1,1]. The closer the absolute value of R(x,u) is to 1, the stronger the linear correlation between x and y.
[0110] Step S3: Construct a prediction model based on the sample convolution model and train it using the training dataset. During training, the Adamom optimizer is used to optimize the model parameters. Simultaneously, the model performance is evaluated using the validation dataset. By comparing the root mean square error (RMSE) of different parameter combinations, the parameter combination with the smallest error is selected as the optimal parameters for the final model.
[0111] The root mean square error is calculated according to formula (3);
[0112]
[0113] In the formula, x i This is the actual capacity value. Here, N represents the predicted capacity value, and N is the total number of data samples.
[0114] Step S31: Construct the sample convolution model. The overall structure of the sample convolution model is as follows: Figure 2 As shown.
[0115] The sample convolutional neural network employs an encoder-decoder architecture. The encoder consists of a hierarchical convolutional module and a fully connected layer. The input feature sequence is processed by the convolutional module and then added to the original sequence to avoid gradient vanishing during backpropagation. Placing the fully connected layer after the convolutional module enhances the nonlinear feature transformation capability and global feature integration capability. The convolutional module primarily focuses on extracting long-term, medium-term, and short-term temporal dependencies from the input feature sequence, while the fully connected layer further processes and transforms these local features through linear transformations, thereby capturing more complex feature relationships. This combination not only improves the model's expressive power but also more effectively extracts and integrates temporal features from time series, thus improving the model's predictive performance and generalization ability. In the decoder, the design of adding two fully connected layers after the flattening layer aims to effectively flatten high-dimensional features and perform complex feature transformations and combinations. Compared to a single fully connected layer, the advantage of a double fully connected layer is that the first layer can learn richer intermediate feature representations, and the second layer further combines these features and maps them to the final prediction target, thereby improving the model's expressive power and generalization ability, and enhancing the accuracy and stability of predictions. This architectural configuration not only helps in exploring the complex relationships inherent in the data, but also effectively addresses various challenges in practical applications. The steps for building the sample convolutional model described above are applicable to all secondary batteries; lithium-ion batteries are used as an example here.
[0116] Step S32: Construct the Adamom optimizer, and update the Adamom optimizer at time t as follows: Figure 4 As shown.
[0117] Initialize the cumulative variable:
[0118] E[g 2 ]0=0 (4)
[0119] E[Δx 2 ]0=0 (5)
[0120] Calculate the cumulative squared gradient of each parameter:
[0121]
[0122] Cumulative squared change of each parameter:
[0123]
[0124] In the formula, E[g 2 E[Δx] represents the cumulative squared gradient at the current moment, reflecting the influence of historical gradients; 2 ]0 represents the cumulative square of parameter updates, reflecting the impact of historical updates on the current parameter update; α and β are decay factors, with default values of 0.9 and 0.9 respectively; g tΔx and Δx represent the gradient and weight update size at the current time step, respectively.
[0125] To accelerate the convergence of model parameters and reduce oscillations, "momentum" is introduced. With each parameter update, momentum is added to the previous time step's value along with some information from the current gradient.
[0126] v t =γv t-1 +(1-γ)g t (8)
[0127] In the formula, v t-1 and v t These represent the momentum at the previous and current moments, respectively; γ represents the decay factor, with a default value of 0.9; g t This represents the gradient at the current moment.
[0128] Calculation update:
[0129]
[0130] x t+1 =x t +Δx t (10)
[0131] In the formula, x t+1 and x t These represent the weights for the next and current time steps, respectively; ∈ has a default value of 1e to avoid division by zero. -6 .
[0132] Step S33: The current SOH value is not only related to the current state information, but also significantly related to historical state information. Incorporating historical observation data into the model to represent historical SOH values can improve the estimation accuracy of the current SOH value. Selecting a health state prediction input length of L, using the health factors from the 1st to the Lth cycle as input, yields the predicted battery health state for the (L+1)th cycle; then using the health factors from the 2nd to the (L+1)th cycle as input to predict the battery health state for the (L+2)th cycle. This process continues until the entire training or test set for health state prediction is input into the model for training. Therefore, the model's input vector is ψ. t =[X t-L ,X t-L+1 … X t ], where t represents the cycle period at the current moment, ψ t It is the input vector of the model at time t, which contains all the observation data from t to tL.
[0133] Step S33: During training, the sample convolutional model is trained on the training set, and the Adamom optimizer is used to optimize the model parameters. Simultaneously, the model performance is evaluated using the validation dataset, and the root mean square error (RMSE) under different parameter combinations is compared. Finally, the parameter combination with the smallest error is selected as the optimal model parameters.
[0134] Step S4: Test the trained model using the test set to obtain an estimate of the battery's state of health (SOH). Inversely normalize the predicted SOH value to obtain the actual predicted capacity of the battery.
[0135] Step S5: The model's predictive performance is evaluated by calculating the root mean square error (RMSE), and the model's hyperparameters are adjusted based on the magnitude of the RMSE value to optimize model performance.
[0136] III. Examples
[0137] Example 1
[0138] The experimental data used in this embodiment comes from the NASA PCoE website. The basic information of the selected battery is shown in Table 1.
[0139] Table 1 Basic Information about Batteries
[0140] model nominal voltage Rated capacity Cut-off voltage Operating temperature 18650 3.6V 2Ah 3.2V / 4.2V 0~60℃
[0141] This embodiment selects test data from batteries B0005 and B0006 for research. The test information is as follows: ambient temperature: 24℃, charging current: 2A, discharging current: 2A, and discharge cutoff voltage: 2.7V (B0005) and 2.5V (B0006).
[0142] The voltage difference sequence at 200s was selected as the characteristic sequence of voltage rise over the same time (RVCT). The CCCT, CVCT, RVCT characteristic sequences and capacity sequences were extracted from batteries B05 and B06, respectively, and then normalized.
[0143] According to step S2, correlation analysis is performed on the three sets of feature sequences.
[0144] Table 2 Correlation between CCCT, CVCT, RVCT and volume
[0145]
[0146] As can be seen from Table 2, only the constant current charging time (CCCT) and battery capacity have a strong Pearson correlation, so CCCT was selected as the health factor.
[0147] According to step S2, the health factor sequence and the normalized capacity sequence are divided into training set, validation set and test set in the proportions of 5:2:3, 6:2:2 and 7:2:1.
[0148] Based on steps S3, S4, and S5, a sample convolutional model is constructed with a batch size of 32, a sliding window size of 24, and a random seed of 42. The RMSE errors of the following models—a one-dimensional convolutional neural network based on the Adam optimizer (Adam-CNN), a sample convolutional model based on the Adam optimizer (Adam-SCM, where the Adam optimizer is a conventional optimizer in the prior art, not the Adamom optimizer mentioned in this application), a sample convolutional model based on the Adadelta optimizer (Adadelta-SCM), and a sample convolutional model based on the Adamom optimizer (Adamom-SCM, i.e., this application)—for predicting the SOH of batteries B05 and B06 are compared to demonstrate the different impacts of different optimizers and convolutional neural networks on the prediction results. The results are shown in Table 3. The curves showing the change of the loss function (mean squared error MSE) with the increase of training iterations during the training process of the sample convolutional neural networks based on different optimizers after the B0006 dataset was divided in a 6:2:2 ratio are shown in Table 3. Figure 6 As shown.
[0149] Table 3. RMSE Errors of CNN, SCM, and Adamom-SCM Models in Predicting SOH of Lithium-ion Batteries
[0150]
[0151] Depend on Figure 6 It can be seen that the Adamdelta optimizer "jitters" around a local minimum during training, while the Adamom optimizer enters a smaller local minimum without exhibiting this "jittering" phenomenon. This indicates that compared to the Adamdelta optimizer, the Adamom optimizer can mitigate the impact of gradient fluctuations, avoid getting trapped in local minima, and make the optimization process more stable and converge faster. Table 3 shows that the sample convolutional model based on the Adamom optimizer has the highest prediction accuracy for battery SOH. This indicates that the sample convolutional model can extract more detailed time dependencies than the traditional one-dimensional convolutional model to achieve accurate prediction of battery SOH. Furthermore, the Adamom optimizer, compared to the Adam and Adadelta optimizers, allows the model to enter more local minima, thus improving the prediction accuracy of the sample convolutional model for battery SOH.
[0152] In summary, this application fully leverages the feature mining capabilities of the sample convolutional model for time-series information, eliminating the need to consider prior knowledge such as battery degradation, and instead observing battery SOH through directly observable data. Compared to traditional Long Short-Term Memory (LSTM) neural networks and other deep learning models, the sample convolutional model has fewer parameters, lower computational cost, and lower requirements for the configuration of electric vehicle battery management systems. Furthermore, this model can adapt to different charging and discharging conditions, exhibiting good versatility. In addition, the Adamom optimizer addresses the learning rate adjustment issue compared to the traditional Adam optimizer. While inheriting the advantages of the Adadelta optimizer, it further improves the model training convergence speed and effectively avoids oscillation problems. This method can accurately predict battery SOH, ensuring the safe operation of the battery system and reducing operating costs.
[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of this application without departing from the spirit and scope of this technical solution should be covered within the scope of the claims of this application.
Claims
1. A secondary battery SOH prediction model based on the Adamom optimizer, characterized in that, Specifically, the steps include the following: Step 1: Obtain the feature sequence and capacity sequence of the secondary battery, and then normalize them respectively; wherein, the feature sequence includes constant current charging time, constant voltage charging time, and voltage rise value within the same time period; Step 2: Perform correlation analysis on the normalized feature sequence and capacity sequence, select the feature sequence with strong correlation to the capacity sequence as the health factor sequence, and then divide the obtained health factor sequence into training set, validation set and test set according to a certain ratio. Step 3: Construct a secondary battery SOH prediction model based on the sample convolution model, and train the prediction model using the training set obtained in Step 2; during the training process, use the Adamom optimizer to optimize the parameters of the prediction model, and use the validation set obtained in Step 2 to validate the model, and select the parameter combination with the smallest error as the final model parameters. Step 4: Test the prediction model obtained in Step 3 using the test set obtained in Step 2 to obtain the estimated value of SOH. Then, inverse normalize the estimated value of the cycle capacity to obtain the predicted value of the cycle capacity. Step 5: Determine the neural network prediction results based on the root mean square error (RMSSE), and then output, optimize, and verify them. The Adamom optimizer is obtained by improving the Adadelta optimizer. At time t, the Adamom optimizer is updated and calculated as follows, with specific steps as follows: (1) Initialize the cumulative variables; (4) (5) in, This represents the cumulative squared gradient at the current time step; This represents the cumulative square of the parameter updates at the current time. (2) For each parameter, at time t, the exponentially weighted average of the squared gradient is calculated using the decay factor α; (6) in, This is the attenuation factor, with a default value of 0.9; (3) For each parameter, at time t, the exponentially weighted average of the squared parameter updates is calculated using the decay factor β; (7) in, This is the attenuation factor, with a default value of 0.9; (4) Update the momentum by adding some information from the current gradient based on the previous time step: (8) In equation (8), and These represent the momentum at the previous moment and the current moment, respectively; This represents the attenuation factor, with a default value of 0.
9. The gradient represents the current time step; (5) Using the previously calculated results and To calculate the parameter update amount for this iteration: (9) (10) In equations (9) to (10), and These represent the weights of the next time step and the current time step, respectively. The default value is 1e -6 ; In step 2, the correlation analysis is performed using the Pearson correlation coefficient. When the absolute value is greater than or equal to 0.8, the correlation is considered to be strong.
2. The prediction model according to claim 1, characterized in that, In step 1, the normalization calculation formulas for the feature sequence and capacity sequence are as follows: Formula 1: ; In Equation 1: x represents the sequence data, min(x) represents the minimum value in the sequence, and max(x) represents the maximum value in the sequence.
3. The prediction model according to claim 1, characterized in that, The formula for calculating the Pearson correlation coefficient is: Formula 2: ; In Equation 2: This represents the magnitude of the Pearson coefficient between x and y; and These represent the average values of x and y, respectively. The value of is in the range of [-1, 1].
4. The prediction model according to claim 1, characterized in that, The data volume ratio of the training set, validation set, and test set is 5:2:3, 6:2:2, and 7:2:1; the training set, validation set, and test set need to undergo format processing before being input into the model. The format processing process is as follows: Let the sliding window size be L. Take the health factor sequence from the 1st to the Lth cycle as input to obtain the data for the (L+1)th cycle. Then take the health factor sequence from the 2nd to the (L+1)th cycle as input to the model to predict the data for the (L+2)th cycle, until the entire training or test set has been processed.
5. The prediction model according to claim 4, characterized in that, The format of the input samples for the training set, validation set, and test set of the convolutional neural network is as follows: Formula 3: ; In Equation 3: N is the total number of samples in the dataset, and L is the input length.
6. The prediction model according to claim 1, characterized in that, The calculation process for the root mean square error (RMSE) is as follows: Formula 4: ; In Equation 4: This represents the actual value from the capacity test. Here, N represents the predicted capacity value, and N is the total number of data samples.
7. The prediction model according to claim 6, characterized in that, The process of judging the neural network prediction results based on the root mean square error (RMSE) and outputting, optimizing and verifying them is as follows: the hyperparameters of the sample convolutional neural network are adjusted using the control variable method, steps 3 and 4 are repeated, and the set of hyperparameters with the lowest RMSE is selected as the hyperparameters of the model.
8. An application of a secondary battery SOH prediction model based on the Adamom optimizer, characterized in that, The prediction model described in any one of claims 1 to 7 is used to detect the health status of a secondary battery.
Citation Information
Patent Citations
Combined power load prediction method based on IVMD-ACMPSO-CSLSTM
CN112270442A
Internet of Things intrusion detection system and method based on hybrid neural network model optimization
CN115412332A