Battery health state estimation method based on white whale optimization algorithm improved LSTM network

Abstract

The invention relates to the field of lithium ion batteries, in particular to a battery health state estimation method based on a white whale optimization algorithm improved LSTM (Long Short Term Memory) network, which comprises the following steps of: 1, extracting battery aging characteristics and screening the battery aging characteristics; step 2, data preprocessing, including data normalization, time sequence data construction and VMD decomposition; step 3, constructing a data driving model, including constructing a battery SOH prediction model by using an LSTM neural network improved by a whale optimization algorithm; and 4, setting an evaluation index. According to the method, the DTV-VMD-LSTM network oriented to battery SOH estimation is constructed, the characteristic that the DTV is simple and easy to obtain the battery characteristics and the advantage of the LSTM network for capturing the long-term dependency relationship are combined, nonlinear factors in the battery can be considered, a large number of experiments do not need to be performed on the battery, the local characteristics and the overall trend in the battery aging process can be considered, and the battery SOH estimation accuracy is improved. And the method has high robustness while the prediction precision is higher.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery management, and in particular to a battery health state estimation method based on an improved LSTM network using a Beluga optimization algorithm. Background Art

[0002] In recent years, the electric vehicle industry has developed rapidly. As the main power battery for electric vehicles, the reliability of lithium-ion batteries has become a topic of great concern in recent years. During use, the various performances of lithium-ion batteries will continue to decline. When the battery life reaches a certain value, the continued use of the battery will bring great safety hazards and even cause serious accidents. Accurate prediction of the health status of the battery can effectively avoid such accidents, reduce the management cost of the system, and improve the safety and stability of the battery system.

[0003] There are currently many methods for predicting battery health status estimation, which can be mainly divided into four categories, namely experimental-based methods, model-based methods, data-driven methods and signal analysis-based methods.

[0004] The experimental method requires a large number of experiments to determine the relationship between battery health indicators (such as capacity, internal resistance) and external factors (such as current, temperature). According to the different battery aging pathways, this method can be divided into calendar aging model and cycle aging model. Battery calendar aging is the aging phenomenon caused by the influence of factors such as ambient temperature on the battery over a long period of time. Battery cycle aging involves complex external environmental factors, usually considering current, temperature and charge and discharge cycle range. This study couples the impedance-based electrothermal model with the calendar aging model to obtain a complete battery life model to simulate battery aging behavior. The principle of the experimental method is simple and easy to understand, but in actual applications, the use of batteries is very complex, and the characteristics of different batteries are not completely consistent, resulting in very large differences in the aging of different batteries. At the same time, battery aging is not completely linear. The existence of nonlinear factors greatly limits the applicability of models based on experimental methods, and simple models cannot accurately predict the health status of batteries.

[0005] Model-based methods predict the battery SOH (State of Health) by establishing an equivalent battery model, such as based on an electrochemical model and an equivalent circuit model. The electrochemical model can describe the internal and external characteristics of the battery. Among them, the computational cost of SPM is lower than that of P2D, but the prediction accuracy is lower than the latter. The equivalent circuit model establishes a mathematical model based on the electrical characteristics of the battery and reflects the health status of the battery through the changes in model parameters. The equivalent circuit model usually includes one or more resistors, capacitors, and inductors, which simulate the dynamic response and internal electrochemical reactions of the battery. The advantage of the model-based method lies in its mathematical simplicity and good simulation of the dynamic behavior of the battery, but it also relies on accurate model parameter identification and sufficient data to maintain the accuracy of its prediction. Building an accurate battery model is by no means easy, and usually requires a lot of time and computational cost. Therefore, the validity and feasibility of the established physical model are difficult to guarantee.

[0006] The data-driven SOH estimation method evaluates the health status of the battery through the real-time data or historical data of the battery. This method collects the data of the battery during the aging process and extracts features from it as the input database. Then, a machine learning algorithm is used to establish a SOH estimation model, and the battery SOH is estimated through the model, and the trend of battery health status changes is predicted. The data-driven method mainly includes support vector machine regression, Gaussian process regression, extreme learning machine, and long short-term memory network. The data-driven method does not rely on the physical and chemical model of the battery, but predicts SOH by analyzing historical or real-time data such as battery voltage, current, and temperature. It has strong flexibility and applicability. However, the prediction effect of this method is largely related to the input data of the model. The correlation between the input data and the battery capacity determines the establishment of the model's mapping relationship between the two. Therefore, how to optimize the input database is the research difficulty of this method at this stage.

[0007] The signal analysis-based method uses electrochemical analysis technology and signal processing technology to extract battery degradation characteristics with certain signal characteristics during the battery aging process, and conducts battery aging analysis based on these characteristics. Common characteristic signal analysis methods include electrochemical impedance spectroscopy, slow cyclic voltammetry, differential voltage analysis, incremental capacity analysis, etc. These are all in-situ non-destructive characterization methods. These methods draw curves by secondary processing of easily measured signals such as voltage, current, and temperature. The peak and valley values ​​and positions of these curves contain information about lithium ions participating in electrochemical reactions of the battery under different aging conditions. These methods can extract reliable battery degradation characteristics without destructive testing of the battery. It is a rapid electrochemical characterization method that has both engineering practicality and theoretical basis. The signal analysis-based method does not require in-depth analysis of the internal reaction process of lithium-ion batteries. It performs secondary processing on signals that are easy to measure during the battery aging process and draws curves. It extracts features through curves and uses them for information prediction. This method has good nonlinear characteristics and simple and easy data acquisition, but the generalization of this method is poor and requires high-quality training data to ensure the robustness and stability of the training model. Summary of the invention

[0008] To achieve the above object, the present invention provides the following technical solutions: A battery health state estimation method based on an improved LSTM network using a Beluga optimization algorithm, comprising: Step 1: battery aging feature extraction and battery aging feature screening; Step 2: Data preprocessing, including data normalization, time series data construction, and VMD decomposition; Step 3: Build a data-driven model, including using the LSTM neural network improved by the White Whale optimization algorithm to build a battery SOH prediction model; Step 4: Set evaluation indicators.

[0009] As a further solution of the present invention: in step 1, a lithium-ion battery charge and discharge experimental condition is set and the battery is charged and discharged, and the voltage, current, temperature data of the lithium-ion battery during the experiment and the battery capacity of each battery charge and discharge cycle are recorded. As the battery ages, the phase change inside the battery will cause a change in the system entropy, and this change will be reflected in the inflection point on the DTV curve. The peak position of the DTV curve can represent the peak potential during the charge and discharge stage, the change in the peak position can represent the change in impedance and stoichiometry, the peak height can represent the maximum heating rate, the peak width can represent the potential window of the electrode phase combination, and the peak area can provide information on the heat generated during the charge and discharge stage. Considering the entropy change coupling of the battery system, this study uses DTV as a battery aging analysis method. This method is very easy to implement and only requires the collection of two external parameters for calculation and research. The calculation of DTV can be expressed as: , where dT and dV are the changes of battery surface temperature and battery terminal voltage with sampling time, and dt is the differential term of time. The fluctuation of the DTV curve corresponds to the change of battery phase change. In order to remove noise interference in the present invention, the SG filtering method is used to filter the curve. The SG filter has excellent filtering performance in high-frequency digital signal filtering and is a local polynomial regression technology. The main advantage of the SG filter is that in the process of removing noise, it can effectively keep the basic form and key features of the signal unchanged. Therefore, it is very suitable for filtering DTV curves and can ensure that the filtered curve retains important data features. The structure of the SG filter is described as follows: ; in is the original input signal, The resulting output signal is is the coefficient given by the SG filter, N is the number of convolution integers, which is equal to the smoothing window size . In the DTV curve of the battery, the height and position of the peaks and troughs change with the increase in the number of battery cycles. Therefore, features can be extracted based on the height and position of the peaks and troughs to reflect the battery aging. In the present invention, the DTV values ​​and measured voltages at the peaks and troughs corresponding to the battery DTV curve are extracted as alternative features. In order to comprehensively and quantitatively study the influence of characteristic variables, reduce computational costs, and improve the robustness of the model, Pearson correlation analysis is used to quantitatively evaluate the correlation between the extracted characteristic variables and SOH, and select characteristic variable values ​​with higher correlation as prediction model input. The specific formula for calculating the correlation coefficient is as follows: ; Where n is the number of sample sequences, Each represents a variable. Respectively represent This method calculates the correlation coefficient of the input data and obtains a value between -1 and 1. According to the different correlation coefficients, the degree of correlation between the two variables can be quantified intuitively.

[0010] As a further solution of the present invention: Step 2 uses the maximum and minimum normalization technology, which can scale the data values ​​to the range of [0,1]. This process identifies the maximum and minimum values ​​in the data set and uses these extreme values ​​to adjust each data point to ensure that they reflect their relative position in the entire data set. In this way, the scale consistency of the data in subsequent processing is ensured, which is crucial to improving the training efficiency and prediction accuracy of the model. The normalization process for each data variable is as follows: ; in, is the normalized value, is the minimum value in the sequence, is the maximum value in the sequence, w is the original data, and in the present invention, the voltage, current and temperature values ​​are taken respectively.

[0011] After normalizing the data, we construct the time series data. First, we select the batch size, which is the amount of data contained in each batch after all the data are divided into a certain number of batches. Taking the No. 18 battery as an example, we select a batch size of 4, which means that the first batch contains the battery aging characteristics from the first battery aging cycle to the fourth battery aging cycle, that is, Figure 2 The F1, F3, and F5 feature values ​​and the corresponding residual values ​​corresponding to the first 4 rows of data. The second batch contains the battery aging features from the 5th battery aging cycle to the 8th battery aging cycle, and so on. Since the size of the No. 18 battery data set has a total of 132 battery cycles, the last batch contains the battery aging features from the 129th time point to the 132nd time point. After the batches are divided, a fixed-size window is used to divide the data. The window size is selected as 2. The first window contains the battery aging features from the 1st battery aging cycle to the 2nd battery aging cycle, that is Figure 2 The first two rows of data in the first batch in the middle bottom. The information in this window is used to predict the battery SOH value at the third time point after neural network learning. The remaining windows are all of size 2 and slide backward with a step size of 1. Time series analysis can reveal the evolution trend of data over time. It is based on historical data and helps to discover the causal relationship between variables, which is key to predicting future trends.

[0012] As a further solution of the present invention: VMD decomposition in step 2 is an advanced signal processing technology. VMD can effectively decompose complex nonlinear and non-stationary signals into several modal components with different center frequencies. These modal components are separated from each other in the frequency domain, and each modal component has a limited bandwidth. The schematic diagram of VMD decomposition effect is as follows Figure 3 As shown, Figure 3 The left side shows the battery capacity decline curve of No. 7 battery. The battery capacity decreases nonlinearly as the battery is continuously charged and discharged. Figure 3The right side is the VMD decomposition result diagram, from which we can see that the capacity curve of battery No. 7 is decomposed into a residual curve (i.e., the IMF1 curve) that can represent the overall downward trend of the curve and various component curves that can represent the local fluctuations of the battery (i.e., the curves other than IMF1). These curves have modal components with different center frequencies, and the original curve is obtained by linearly adding these curves. The present invention performs VMD decomposition on the battery SOH curve to obtain residuals and other components. The residuals represent the overall trend of battery aging, and the other components contain the local fluctuation information of the battery. Modeling and prediction of each component can well predict the capacity recovery phenomenon of the battery.

[0013] As a further solution of the present invention: the LSTM network in step 3 can solve the gradient problem and the long-term dependency problem well. LSTM contains three different types of gate structures, namely input gate, forget gate and output gate. The first step of LSTM is to use the forget gate to determine which information in the cell state is discarded and retained when the LSTM unit obtains data. The forget gate can be expressed as: ; in is the output of the forget gate, is the weight matrix, is the bias vector, is the hidden state of the neuron at the previous time step, is the input of the LSTM network at the current time step, is the sigmoid function.

[0014] In the second step, we decide what new information to store in the cell. This is done in two parts. First, we update the information in the cell through the input gate. Then, we use the tanh layer to get the new information. In the next step, we combine these two pieces of information to update the state. These two steps can be expressed as: ; ; in and is the weight matrix, and is the bias vector, is the output of the input gate at the current time step, is the newly generated memory state of the neuron, is the hyperbolic tangent function.

[0015] The third step is to integrate the data from the forget gate and the input gate. Multiply by the output of the forget gate , represents the process of partially discarding data, and then adding the output of the input gate Multiply by the new memory state of the neuron , changing the memory state of the neuron from Updated to , It will be passed to the next neuron, and the specific calculation process can be expressed as: , Finally, the output part is determined. The output gate is based on the unit state. , the previous cycle output and current input Generate a new hidden state The output gate can be expressed as: ; ; in and The values ​​are the weight matrix and bias vector, is the output of the output gate at the current time step, is the hidden state of the neuron at the current time step and is also the output of the current time step. Will make peace Pass it on to the next neuron.

[0016] As a further solution of the present invention: the present invention adopts RMSprop optimization algorithm and dropout technology to train LSTM network. In this study, dropout technology is used to randomly discard neurons to prevent any neuron from over-relying on the output of any other neuron, forcing it to rely on the group behavior of its input. Specifically, in the forward propagation process of the neural network, each neuron will be hidden with a certain probability, and the hidden neurons do not express any behavioral information. In the case that some neurons are hidden, a new neural network will be randomly generated on the basis of the original neural network. The model complexity of the network must be less than that of the original neural network, thereby reducing the complexity of the model. When solving the overfitting problem, the dropout strategy can reduce the number of learning parameters and increase the model convergence speed. In the present invention, three characteristic time series of the test data set are used as input. The constructed model includes two LSTM layers, two dropout layers and one dense layer. The architecture diagram of the prediction network is shown in the figure. Figure 4 The present invention adopts the White Whale optimization algorithm to optimize the hyperparameters of the LSTM neural network, and weaves the learning rate, sample sequence length, sample batch size, number of neurons in the LSTM layer and the discard rate of the dropout layer of the LSTM network into the initial population.

[0017] There are three stages in the algorithm optimization process, namely the exploration stage, the development stage and the whale fall stage. Each iteration will choose one of the two stages, depending on the balance factor. , its mathematical model is: , t is the current iteration, T is the maximum number of iterations, It changes randomly between (0,1) at each iteration. The exploration phase occurs when the balance factor , while the development phase occurs in . As the iteration T increases, The fluctuation range of decreases from (0,1) to (0,0.5), indicating that the probability of the exploration stage has decreased significantly, while the probability of the development stage increases with the continuous increase of iteration T.

[0018] The exploration phase is built by considering the swimming behavior of the beluga whale, and the position of the beluga whale is updated as follows: ; in, is the current iteration number, is the new position of the i-th beluga whale in the j-th dimension, is randomly selected from the d dimensions, is the i-th beluga whale in The position in dimension, and is the current position of the i-th and r-th beluga whales (representing randomly selected beluga whales), is a random number in the range (0,1) used to enhance the random operator in the exploration phase. and A random number used to average between fins.

[0019] The development phase is established by considering the predation behavior of beluga whales, and its mathematical model is expressed as: ; in, Indicates the current iteration number; and Respectively indicate the current position beluga whales and a randomly selected beluga whale; Indicates The first white whale updated its position; Indicates the optimal position of the white whale; and Indicates taking random numbers between (0,1) respectively; represents the random jump intensity of Levy flight, . is the Levy flight function, and the calculation formula is as follows: ; ; In the formula, and are normally distributed random numbers, is the step length of Levy flight; is a constant, , is the gamma function.

[0020] To simulate the whale falling behavior in each iteration, the probability of selecting a whale falling from the population is used to simulate the changes in the population. The beluga whales either move to other places or are shot down and fall into the deep sea. To ensure that the population size remains unchanged, the whale position is updated by using the beluga position and the whale falling step length. The formula is as follows: ; In the formula, and They are random numbers between (0,1); represents the step length of the whale's fall, namely: ; In the formula, is the step size factor related to the probability of whale decline and population size , and are the upper and lower bounds of the variables, respectively. It can be seen that the step size is affected by the bounds of the design variable, the number of iterations, and the maximum number of iterations.

[0021] In this model, the probability of a whale falling is is calculated as a linear function: .

[0022] The probability of a whale falling drops from 0.1 in the initial iteration to 0.05 in the final iteration, indicating that the danger to the beluga whale decreases as it gets closer to the food source during the optimization process. The overall model construction and prediction flow chart is shown in the figure below. Figure 5As shown in the figure, firstly, the battery temperature, measured voltage and battery capacity data are extracted from the battery data set, and the DTV curve is drawn for the battery temperature and measured voltage, and the feature extraction and screening are performed. The battery capacity data is decomposed by VMD to obtain the residual data reflecting the overall trend of the battery and the component data reflecting the local trend of the battery. A prediction model is established for the residual and each component respectively. For the residual prediction model, the extracted battery features are used as the input features, and the residual is used as the input label. The two are combined and converted into time series data, and then input into the network for prediction. For each component prediction model, the extracted battery features are used as the input features, and each component is used as the input label. The two are combined and converted into time series data, and then input into the network for prediction. The LSTM prediction network uses the White Whale optimization algorithm to optimize the network hyperparameters to improve the model prediction accuracy. After the prediction of each model is completed, the predicted values ​​of each model are linearly added to obtain the SOH prediction value.

[0023] As a further solution of the present invention: In step 4, two evaluation indicators are set to evaluate the performance of the model, namely, Mean Absolute Error (MAE) and Root Mean Square Error (RMSE): ; ; in for The true value of SOH at the moment, for The SOH forecast value at the moment. MAE is the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. It gives the same weight to all errors, the results are easy to understand, and it is insensitive to outliers. RMSE is the square root of the average of the squares of the differences between the predicted values ​​and the actual values. It gives greater weight to larger errors. Because the errors are squared, larger errors have a greater impact on RMSE. This makes RMSE more sensitive to outliers. The smaller its value, the better the prediction effect of the entire sequence.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs a DTV-VMD-LSTM network for battery SOH estimation. The network combines the simple and easy characteristics of DTV in acquiring battery features and the advantages of LSTM network in capturing long-term dependencies, and can achieve good results in battery SOH estimation.

[0025] 2. The present invention extracts features that can reflect the battery aging trend from the battery DTV curve, screens these features through Pearson correlation analysis, and selects features with strong correlation for battery SOH estimation. These features reflect the internal phase change of the battery and have a high correlation with the battery SOH curve.

[0026] 3. The present invention uses the VMD decomposition method to decompose the battery SOH curve to obtain a monotonic residual and several components, wherein the residual represents the overall trend of battery aging, and the remaining components contain the local fluctuation characteristics of the battery. By comprehensively considering both and modeling and predicting all components separately, the battery aging trend can be predicted while the capacity regeneration phenomenon in the battery degradation process can be well predicted.

[0027] 4. The present invention adopts the Beluga optimization algorithm to optimize the hyperparameters of the LSTM network. The algorithm simulates the migration and predation behavior of the beluga whale, and optimizes the network learning rate, sample sequence length, sample batch size, number of neurons in the LSTM layer, and the dropout rate of the dropout layer. By optimizing the hyperparameters, the model prediction error can be significantly reduced and the model performance can be improved.

[0028] 5. The present invention sets two indicators, the error value and the root mean square error value, to evaluate the performance of the model from a local and overall perspective respectively. By complementing each other with these two indicators, the performance of the model can be evaluated more comprehensively. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention is a flowchart of a method for estimating the battery health status of an LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of the time series data conversion process in the battery health state estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the VMD decomposition effect in the battery health state estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0032] Figure 4 This is an architecture diagram of the prediction network in the battery health state estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0033] Figure 5 This is a flowchart of model construction and prediction in a battery health state estimation method based on an improved LSTM network using the Beluga optimization algorithm in an embodiment of the present invention.

[0034] Figure 6This is a schematic diagram of the VMD decomposition results of AA batteries in the battery health status estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0035] Figure 7 This is a diagram of the prediction results of four types of batteries in the battery health status estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention.

[0036] Figure 8 These are four battery prediction error result diagrams in the battery health status estimation method based on the improved LSTM network based on the Beluga optimization algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The present invention is a SOH estimation method that integrates a data-driven method and a signal analysis method. Specifically, after the battery capacity curve is subjected to variable mode decomposition, it is input into the LSTM neural network together with the extracted battery aging features, and the LSTM network is optimized using the Beluga optimization algorithm. First, the DTV curve of the data set is drawn, the peak and trough values ​​in the DTV curve and their corresponding measured voltages are extracted as candidate features, and the battery aging features are further screened according to the correlation analysis. At the same time, the battery SOH curve is subjected to variational mode decomposition (VMD for short), and the decomposition result is a residual that can reflect the overall change trend of the battery and several components that can reflect the local information of the battery. Secondly, the time series data is constructed and the prediction model is built. For the residual and each component of the battery SOH decomposition, a prediction model is established respectively. For each model, the screened battery aging features are used as features, and the corresponding VMD component value of the battery SOH value is used as a label to construct the time series. The constructed time series is used as input and input into the LSTM network for result prediction. In the process of building the LSTM network, the hyperparameters of the network are optimized using the Beluga optimization algorithm to further reduce the prediction error. Finally, the prediction results of each model are linearly added to obtain the final SOH prediction value. The battery SOH model prediction in the present invention can take into account the nonlinear factors inside the battery, without the need for a large number of experiments on the battery to build the corresponding physical and chemical model, and can take into account the local characteristics and overall trends of the battery aging process, with higher prediction accuracy and good robustness.

[0039] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0040] Embodiment 1, a method for estimating the battery health state based on an improved LSTM network using a Beluga optimization algorithm is provided in an embodiment of the present invention, comprising the following steps: Step 1: Battery aging feature extraction and battery aging feature screening This experiment was verified on an open source dataset. The dataset used in the algorithm comes from the Center of Excellence for Prediction at NASA Ames Research Center. The entire dataset is synthesized by dynamic load curves consisting of different battery cycles and randomly selected charge and discharge currents. The dataset contains experimental data of various batteries under charging, discharging and electrochemical impedance spectroscopy measurement conditions. Therefore, the experimental data of batteries 5, 6, 7 and 18 under discharge conditions were selected for research. The lithium-ion battery tested is a commercially available 18650-sized lithium-ion battery produced by Idaho National Laboratory. The tested battery was subjected to degradation experiments in constant current constant voltage charging (CC-CV) mode and constant current discharge at room temperature (24°C). The charging current was maintained at 1.5A during constant current charging, and the charging cutoff voltage of the four batteries was 4.2V. Afterwards, during constant voltage charging, the charging process was stopped when the charging current was less than 20mA. The charging process of the four batteries was consistent. After charging, the four batteries are discharged at a current of 2A until the voltages of batteries No. 5, No. 6, No. 7, and No. 18 reach the preset values ​​of 2.7V, 2.5V, 2.2V, and 2.5V, respectively.

[0041] The present invention extracts the characteristics of battery aging based on the DTV curve. As the battery ages, the phase change inside the battery will cause a change in the system entropy, and this change will be reflected in the inflection point on the DTV curve. The peak position of the DTV curve can represent the peak potential during the charge and discharge phase, the change in the peak position can represent the change in impedance and stoichiometry, the peak height can represent the maximum heating rate, the peak width can represent the potential window of the electrode combination, and the peak area can provide information on the heat generated during the charge and discharge phase. Considering the entropy change coupling of the battery system, this study uses DTV as a battery aging analysis method. This method is easy to implement and only requires the collection of two external parameters for calculation and research.

[0042] The measured voltage and temperature data of the battery in the data set are selected, and the corresponding DTV value of the battery in each cycle is calculated according to the DTV calculation formula.

[0043] In order to remove noise interference in the present invention, the SG filtering method is used to filter the curve. The SG filter has excellent filtering performance in high-frequency digital signal filtering and is a local polynomial regression technology. The main advantage of the SG filter is that in the process of removing noise, it can effectively keep the basic form and key features of the signal unchanged, so it is very suitable for filtering the DTV curve, and can ensure that the filtered curve retains important data features. After filtering, a smooth battery DTV curve is obtained. After smoothing, there are two peaks and one trough on the DTV curve of each cycle of the battery. In order to comprehensively and quantitatively study the influence of characteristic variables, reduce computational costs, and improve the robustness of the model, Pearson correlation analysis is used to quantitatively evaluate the correlation between the extracted characteristic variables and SOH, and select characteristic variable values ​​with higher correlation as prediction model input. The correlation coefficient between each feature and the battery SOH is calculated, and the results are shown in Table 1.

[0044] It can be seen from Table 1 that compared with F2, F4, and F6, F1, F3, and F5 have a significantly stronger correlation with battery capacity decay. Therefore, this study establishes a battery aging model based on these three features.

[0045] Table 1. Schematic diagram of the correlation coefficients of the six extracted features

[0046] Step 2: Data Preprocessing Perform uniform maximum and minimum normalization calculations on the collected data. This technique can scale the data values ​​to the range of [0,1]. This process identifies the maximum and minimum values ​​in the data set and uses these extreme values ​​to adjust each data point to ensure that they reflect their relative position in the entire data set. In this way, the scale consistency of the data in subsequent processing is ensured, which is crucial to improving the training efficiency and prediction accuracy of the model.

[0047] After the data is normalized, it is converted from a one-dimensional vector to a two-dimensional matrix. The one-dimensional vector data of different cycles together constitute a two-dimensional matrix. In the two-dimensional matrix, each row of data is a one-dimensional vector, which contains three battery aging characteristic values ​​and a VMD component value or residual value of SOH. These data come from the peak and valley data in each corresponding cycle of the DTV curve and the battery capacity after the cycle.

[0048] The construction of the time series is divided into two steps. First, select the sequence length (N) and extract multiple matrix segments from the above two-dimensional matrix in turn. The shape of these matrix segments is N×4, the data interval between the matrix segments is 1, and there is overlapping data between different matrix segments, and the amount of overlapping data is N-1. Then, reconstruct these matrix segments into a three-dimensional tensor with a shape of M×N×4, where the size of M depends on the size of the above two-dimensional matrix. This three-dimensional tensor can meet the input data format requirements of the constructed battery degradation model.

[0049] The battery capacity curve is subjected to VMD decomposition to obtain residuals and components. Variational mode decomposition is an advanced signal processing technology. VMD can effectively decompose complex nonlinear and non-stationary signals into several modal components with different center frequencies. These modal components are separated from each other in the frequency domain, and each modal component has a limited bandwidth. The present invention performs VMD decomposition on the battery SOH curve to obtain residuals and other components. The residuals represent the overall trend of battery aging, and the other components contain local fluctuation information of the battery. Modeling and prediction of each component can well predict the battery capacity recovery phenomenon. The VMD decomposition result of battery No. 7 is shown as follows: Figure 7 shown.

[0050] Step 3: Use the LSTM neural network improved by the White Whale optimization algorithm to build a battery SOH prediction model For each residual and component, a prediction model is established respectively, and the network learning rate, sample sequence length, sample batch size, number of LSTM layer neurons and dropout layer discard rate of each neural network are optimized by the white whale optimization algorithm to obtain the optimal solution. The present invention adopts RMSprop optimization algorithm and dropout technology to train the LSTM network. In this study, the dropout technology is used to randomly discard neurons to prevent any neuron from over-relying on the output of any other neuron, forcing it to rely on the group behavior of its input. Specifically, in the forward propagation process of the neural network, each neuron will have a certain probability of being hidden, and the hidden neurons do not express any behavioral information. In the case where some neurons are hidden, a new neural network will be randomly generated on the basis of the original neural network, and the model complexity of the network must be less than the original neural network, thereby reducing the complexity of the model. When solving the overfitting problem, the dropout strategy can reduce the number of learning parameters and improve the model convergence speed. In the present invention, three characteristic time series of the test data set are used as input. The constructed model includes two LSTM layers, two dropout layers and one dense layer.

[0051] The first 50% of the data set is used as the training set, and the rest is used as the test set for model training and verification. Finally, the error value and root mean square error value are used as the result evaluation indicators. The prediction results of the four batteries are shown in Figure 7 shown. Figure 8 The error results of four types of batteries are shown, and the prediction errors of the four types of batteries are all around 1%.

[0052] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A battery health state estimation method based on the improved LSTM network using the White Whale optimization algorithm, characterized in that: The following steps are involved: Step 1: battery aging feature extraction and battery aging feature screening; Step 2: Data preprocessing, including data normalization, time series data construction, and VMD decomposition; Step 3: Build a data-driven model, including using the LSTM neural network improved by the White Whale optimization algorithm to build a battery SOH prediction model; Step 4: Set evaluation indicators.

2. The battery health state estimation method based on the improved LSTM network based on the White Whale optimization algorithm according to claim 1 is characterized in that: In step 1, the battery aging feature extraction is to extract the battery aging feature according to the DTV curve, and the DTV is calculated as follows: , dT and dV are the changes of battery surface temperature and battery terminal voltage with sampling time, dt is the differential term of time, and the SG filtering method is used to filter the DTV curve. The construction of the SG filter is described as follows: ； in is the original input signal, The resulting output signal is is the coefficient given by the SG filter, N is the number of convolution integers, which is equal to the smoothing window size .

3. The battery health state estimation method based on the improved LSTM network based on the Beluga optimization algorithm according to claim 2 is characterized in that: In step 1, the battery aging feature screening is to use Pearson correlation analysis to quantitatively evaluate the correlation between the extracted characteristic variables and SOH, and screen out the characteristic variable values ​​with a high degree of correlation as the prediction model input. The specific formula for calculating the correlation coefficient is as follows: ； Where n is the number of sample sequences, Each represents a variable. Respectively represent The mean of .

4. The battery health state estimation method based on the improved LSTM network based on the Beluga optimization algorithm according to claim 1 is characterized in that: The data normalization in step 2 adopts the maximum and minimum normalization technology. The normalization process for each data variable is as follows: ； in, is the normalized value, is the minimum value in the sequence, is the maximum value in the sequence, w is the original data, and in the present invention, the voltage, current and temperature values ​​are taken respectively.

5. The method for estimating the battery health status based on the improved LSTM network using the White Whale optimization algorithm according to claim 4 is characterized in that: The steps for constructing the time series data in step 2 are as follows: selecting the size of the data contained in each batch; using a fixed-size window to divide the data; VMD decomposition is to perform VMD decomposition on the battery capacity data to obtain residual data reflecting the overall trend of the battery and component data reflecting the local trend of the battery.

6. The battery health state estimation method based on the improved LSTM network based on the White Whale optimization algorithm according to claim 1 is characterized in that: The steps of constructing a battery SOH prediction model using the LSTM neural network improved by the White Whale optimization algorithm in step 3 are as follows: using a forget gate to determine the LSTM unit to obtain data, and the forget gate is expressed as: ； in is the output of the forget gate, is the weight matrix, is the bias vector, is the hidden state of the neuron at the previous time step, is the input of the LSTM network at the current time step, is a sigmoid function; the information in the unit is updated through the input gate, and the tanh layer is used to obtain new information. ; ； in and is the weight matrix, and is the bias vector, is the output of the input gate at the current time step, is the newly generated memory state of the neuron, is the hyperbolic tangent function; the old state Multiply by the output of the forget gate , represents the process of partially discarding data, and then adding the output of the input gate Multiply by the new memory state of the neuron , changing the memory state of the neuron from Updated to , It will be passed to the next neuron, and the calculation process is expressed as: , the output gate depends on the unit state , the previous cycle output and current input Generate a new hidden state , the output gate is expressed as: ； ； in and The values ​​are the weight matrix and bias vector, is the output of the output gate at the current time step, is the hidden state of the neuron at the current time step and is also the output of the current time step. Will make peace The White Whale optimization algorithm is used to optimize the hyperparameters of the LSTM neural network. The constructed model contains two LSTM layers, two dropout layers and one dense layer. The learning rate of the LSTM network, the sample sequence length, the sample batch size, the number of neurons in the LSTM layer and the dropout layer dropout rate are woven into the initial population. The hyperparameter optimization includes three stages, namely the exploration stage, the development stage and the whale drop stage. Each iteration will choose one of the exploration stage and the development stage, which depends on the balance factor. , its mathematical model is: , t is the current iteration, T is the maximum number of iterations, It changes randomly between (0,1) at each iteration, and the exploration phase occurs at the balance factor , while the development phase occurs in ,The exploration phase is established by considering the swimming behavior of the beluga whale, and the position of the beluga whale is updated as follows: ； in, is the current iteration number, is the new position of the i-th beluga whale in the j-th dimension, is randomly selected from the d dimensions, is the i-th beluga whale in The position in dimension, and is the current position of the i-th and r-th randomly selected beluga whales, is a random number in the range (0,1). and The random number used to average the fins was developed by considering the predation behavior of beluga whales, and its mathematical model is expressed as: ； in, Indicates the current iteration number; and Respectively indicate the current position beluga whales and a randomly selected beluga whale; Indicates The first white whale updated its position; Indicates the optimal position of the white whale; and Indicates taking random numbers between (0,1) respectively; represents the random jump intensity of Levy flight, , is the Levy flight function, and the calculation formula is as follows: ； ； In the formula, and are normally distributed random numbers, is the step length of Levy flight; is a constant, , is a gamma function; in the whale fall stage, the position of the white whale and the whale fall step length are used to update the whale position. The formula is as follows: ； In the formula, and They are random numbers between (0,1); represents the step length of the whale's fall, namely: ； In the formula, is the step size factor related to the probability of whale decline and population size , and are the upper and lower bounds of the variable, respectively, and the probability of a whale falling is calculated as a linear function: 。 7. The battery health state estimation method based on the improved LSTM network based on the White Whale optimization algorithm according to claim 1 is characterized in that: In step 4, the mean absolute error (MAE) and the root mean square error (RMSE) are set, and the calculation formula is as follows: ； ； in for The true value of SOH at the moment, for The predicted SOH value at the moment.

Images