Lithium battery remaining service life prediction method based on TVF-EMD model and BWO-CNN-ONLSTM model

Abstract

The invention discloses a lithium battery remaining service life prediction method based on a TVF-EMD model and a BWO-CNN-ONLSTM model. The lithium battery remaining service life prediction method comprises the following steps of 1, obtaining a capacity degradation sequence of a lithium battery of which the service life is to be predicted; 2, adaptively decomposing the capacity degradation sequence of the lithium battery of which the service life is to be predicted into a plurality of intrinsic mode components and residual errors by adopting a time variation filtering empirical mode decomposition method; 3, dividing the decomposed intrinsic mode component into a high-frequency component and a low-frequency component according to the zero-crossing rate, and reconstructing the high-frequency component and the low-frequency component into a high-frequency sequence and a low-frequency sequence respectively; step 4, inputting the high-frequency sequence into the trained BWO-CNN-ONLSTM model for the high-frequency sequence to obtain a high-frequency prediction result; the low-frequency sequence is input into a trained BWO-CNN-ONLSTM model for the low-frequency sequence, and a low-frequency prediction result is obtained; and 5, fusing the high-frequency prediction result and the low-frequency prediction result to obtain a final prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a method for predicting the remaining service life of a lithium battery based on a TVF-EMD model and a BWO-CNN-ONLSTM model. Background Art

[0002] With the rapid development of science and technology, lithium-ion batteries have been deeply integrated into many key areas of modern society. For example, smart devices provide people with a convenient mobile interactive experience, electric vehicles become the mainstream choice for green travel, microgrids ensure the stability of power supply in local areas, and smart power systems optimize energy distribution. All of this relies on the excellent performance of lithium batteries. Their high energy density can store a large amount of electricity in a limited space, meeting the needs of long-term equipment operation; their excellent low-temperature performance enables them to maintain good working condition in cold environments; their long service life reduces the cost and resource consumption of frequent battery replacement; and their low self-discharge rate ensures that the battery power is effectively preserved during idle periods.

[0003] However, in actual use, lithium-ion batteries face a serious problem: with continuous charge and discharge cycles, a series of complex chemical reactions and physical changes occur within them that are detrimental to maintaining performance. On the one hand, the electrolyte gradually decomposes. As the key medium for ion transport within the battery, the electrolyte's chemical composition changes after decomposition, affecting its conductivity and hindering the smooth migration of ions. On the other hand, a passivation film forms inside the battery. Although this film can protect the electrodes to a certain extent, it also increases the resistance to the electrode reaction. In addition, the electrode active material will undergo irreversible dissolution, resulting in a continuous decrease in the effective components participating in the electrochemical reaction. The combined effect of these factors inevitably leads to a gradual decrease in battery capacity, greatly reducing the battery's energy storage capacity, and thus posing a major threat to the safety and reliability of the entire energy system. Once the battery capacity decays to an excessively low level, it may not be able to provide sufficient power support at critical moments, causing system failures or even safety accidents. Currently, the industry generally uses the fact that a lithium-ion battery's capacity has dropped to 70%-80% of its rated capacity as an important basis for judging that it has reached the end of its service life. If it continues to be used at this stage, the potential risks are extremely high. Therefore, accurately predicting the remaining service life of lithium-ion batteries has become a key technical problem that needs to be solved urgently. It is of great significance for ensuring the stable and safe operation of various application systems.

[0004] In-depth research into the aging process of lithium-ion batteries reveals that their capacity decay does not follow a simple linear pattern, but rather exhibits a special capacity regeneration phenomenon. This phenomenon stems from the synergistic effect of multiple complex chemical and physical dynamic changes within the battery. Specifically, these include the reconstruction of the solid electrolyte interface (SEI) membrane, where the properties of the newly formed SEI membrane differ from those of the initial state, affecting ion transmission efficiency; the deposition and dissolution of metallic lithium within the battery, which changes the lithium concentration distribution near the electrodes and thus interferes with the electrochemical reaction process; the dynamic changes in the degree of homogenization of the electrolyte itself, which affects the ion diffusion rate; and the continuous rearrangement of the battery's internal microstructure, where changes in structural parameters such as the electrode material's particle spacing and porosity affect the reaction active sites. The capacity regeneration phenomenon causes the capacity degradation curve of lithium-ion batteries to exhibit significant volatility and highly nonlinear characteristics, which greatly increases the difficulty of accurately predicting their remaining service life. Traditional prediction methods are difficult to adapt to this complex change pattern and are prone to large deviations.

[0005] At present, in order to overcome the problem of predicting the remaining service life of lithium-ion batteries, researchers have developed three main methods: model-based methods, data-driven methods and hybrid methods.

[0006] Model-based approaches are subdivided into physical models and mathematical models. Physical models aim to construct mathematical descriptions based on the actual chemical and physical reaction mechanisms occurring within the battery, employing a large number of algebraic and differential equations to accurately characterize the battery aging process. Typical examples include pseudo-two-dimensional models, which can deeply reflect the dynamic details of material transport and electrochemical reactions within the battery. However, the model construction process is extremely complex, requiring in-depth expertise and extensive experimental data. This consumes significant human, material, and time resources, and is difficult to maintain and optimize. Mathematical models take a different approach, employing statistical or probabilistic methods, such as particle filters (PFs), to summarize aging patterns through statistical analysis of large amounts of battery operating data. While these models avoid the complex derivation of physical mechanisms to a certain extent, they suffer from poor stability and are easily affected by factors such as temperature and charge / discharge rate in the actual operating environment of the battery. This leads to large fluctuations in prediction results, limited generalizability, and difficulty in ensuring accurate predictions under different operating conditions.

[0007] In contrast, data-driven approaches abandon reliance on precise physical models of the battery's internal structure. Instead, they leverage the powerful data processing capabilities of modern machine learning to directly mine potential aging patterns from large amounts of real-world measurement data and construct predictive models. For example, the Joint Prediction Model (JPM) proposed by Gao et al. utilizes a Bayesian model to efficiently integrate multi-sensor data and combines it with an artificial neural network (NN) for RUL prediction. Ma et al. employed a convolutional neural network (CNN) to focus on battery cycle life estimation and combined it with Gaussian process regression (GPR) to achieve RUL prediction. Li et al. creatively integrated deep learning with the Attention Mechanism (AM) to accurately capture capacity regeneration and aid RUL prediction. While data-driven approaches demonstrate a degree of flexibility and efficiency, they rely solely on historical data. Li-ion batteries are subject to significant data fluctuations due to capacity regeneration. In practice, data collection is plagued by numerous factors, including sensor accuracy limitations, external electromagnetic interference, and individual battery differences. Consequently, the collected data is often inaccurate and incomplete. This significantly reduces the accuracy of predictive models built based on this flawed data in practical applications, making it difficult to meet the requirements for high-precision predictions.

[0008] Hybrid methods have emerged, aiming to combine the strengths of the first two approaches, address their respective shortcomings, and improve the accuracy and reliability of predictions. Numerous research teams have continued to explore this direction. For example, Chang et al. first used an unscented Kalman filter (UKF) to obtain the original error sequence, then used a relevance vector machine (RVM) regression model to predict the error based on the reconstructed sequence, thereby correcting the UKF results. Zhang et al. used empirical mode decomposition (EMD) and kernel principal component analysis (KPCA) to analyze the high- and low-frequency components of battery capacity, reconstructed the sequence, and combined long-short-term memory networks (LSTMs) with transfer learning to achieve prediction. Although hybrid methods have made some progress, they still face many challenges in practical applications: First, the complexity of the capacity degradation curve caused by CR still poses a huge obstacle to the model's prediction accuracy. Even if multiple methods are combined, the handling of nonlinear problems is still difficult. Second, in real engineering scenarios, it is necessary to pursue prediction accuracy while taking into account limited computing resources. How to find an accurate balance between the two and coordinate the relationship between model complexity and accuracy is a major difficulty. Finally, due to the limitations of sample collection costs, time, and interference from the actual operating environment, the number of available samples is often insufficient, and the data quality is worrying, with a large amount of noise and outliers. This undoubtedly poses severe challenges to the robustness and generalization ability of the model, hindering its large-scale application in actual industrial production.

[0009] In summary, the existing remaining service life prediction technology of lithium-ion batteries still has many defects and shortcomings when facing complex and changeable actual application scenarios, and it is difficult to meet the growing demand for high-performance and high-reliability lithium-ion battery prediction. There is an urgent need for a new and more effective prediction method and technical solution to break the deadlock and overcome this technical difficulty. Summary of the Invention

[0010] Purpose of the invention: To solve the problem that the existing remaining service life prediction technology of lithium-ion batteries still has many defects and shortcomings when facing complex and changeable actual application scenarios, and is difficult to meet the growing demand for high-performance and high-reliability lithium-ion battery prediction, the present invention proposes a lithium battery remaining service life prediction method based on the TVF-EMD model and the BWO-CNN-ONLSTM model, which can achieve higher-precision lithium battery remaining service life prediction.

[0011] Technical solution: A method for predicting the remaining service life of lithium batteries based on the TVF-EMD model and the BWO-CNN-ONLSTM model, including the following steps:

[0012] Step 1: Obtain the capacity degradation sequence of the lithium battery whose service life is to be predicted;

[0013] Step 2: Use the time-varying filtered empirical mode decomposition method to adaptively decompose the capacity degradation sequence of the lithium battery whose service life is to be predicted into multiple intrinsic mode components and residuals;

[0014] Step 3: According to the zero-crossing rate, the decomposed intrinsic modal components are divided into high-frequency components and low-frequency components, and reconstructed into high-frequency sequences and low-frequency sequences respectively;

[0015] Step 4: Input the high-frequency sequence into the BWO-CNN-ONLSTM model trained for high-frequency sequences to obtain high-frequency prediction results; input the low-frequency sequence into the BWO-CNN-ONLSTM model trained for low-frequency sequences to obtain low-frequency prediction results;

[0016] Step 5: Fuse the high-frequency prediction results and the low-frequency prediction results to obtain the final prediction results;

[0017] The BWO-CNN-ONLSTM model is obtained by optimizing the number of neurons in the ONLSTM layer of the CNN-ONLSTM, the number of neurons in the CNN layer of the CNN-ONLSTM, and the learning rate of the CNN-ONLSTM using the BWO algorithm.

[0018] The CNN-ONLSTM includes: a 1D convolution layer, a first ONLSTM layer, a first Dropout layer, a second ONLSTM layer, a second Dropout layer, a third ONLSTM layer, a third Dropout layer, a fourth ONLSTM layer and a Flatten layer.

[0019] Furthermore, in step 2, the time-varying filtered empirical mode decomposition method is used to adaptively decompose the capacity degradation sequence of the lithium battery whose service life is to be predicted into multiple intrinsic mode components and residuals. The specific operations include:

[0020] Step 2-1: Detect all the maximum points from the capacity degradation sequence x(t) of the lithium battery whose service life is to be predicted, and represent these maximum points as a set u i , where i = 1, 2, 3, ...;

[0021] Step 2-2: Determine the signal interruption point e j , where j = 1, 2, 3...; when e j =u i When , there exists a critical value ρ that satisfies the following formula:

[0022]

[0023] Where ρ represents the preset critical value of the frequency change rate between two consecutive maximum values, represents the frequency of bisection. Then e j On the rising edge, is the minimum value; if Then e j On the falling edge, is the minimum value. The rest of can be considered as the peak;

[0024] Step 2-3: Interpolate between peaks to obtain the local cutoff frequency Adjustment To improve modal aliasing;

[0025] Step 2-4: Reconstruct and obtain the new signal h(t);

[0026]

[0027] Step 2-5: The frequency information of the original signal is obtained from the new signal h(t). Based on this information, the battery capacity discharge sequence x(t) is filtered using a B-spline approximate filter to obtain a local mean function, which is equivalent to a low-frequency component. The local mean function is subtracted from the battery capacity discharge sequence x(t) to obtain a high-frequency component, which is then tested in Step 2-6.

[0028] Step 2-6: Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard. If not, repeat Step 2-2 to Step 2-5:

[0029]

[0030] Where B Loughlin (t) represents the instantaneous bandwidth of Loughlin, Represents the weighted average instantaneous frequency. There is a bandwidth threshold ε. If θ(t)≤ε, the signal is an IMF component signal.

[0031] Therefore, x(t) is decomposed into a finite number of IMF component signals and a residual r(t):

[0032]

[0033] Furthermore, in step 3, the decomposed intrinsic modal components are divided into high-frequency components and low-frequency components according to the zero-crossing rate, and reconstructed into high-frequency sequences and low-frequency sequences respectively. The specific operations include:

[0034] Step 3-1: According to the zero-crossing rate, the decomposed natural modal components are divided into high-frequency components and low-frequency components, which can be expressed as:

[0035]

[0036] Where, P zero represents the zero-crossing rate, n zero Represents the number of points where the signal crosses zero when changing from positive to negative or from negative to positive. N is the total number of sample points.

[0037] Step 3-2: If P zero If it is less than 0.01, it is classified as a low-frequency component; otherwise, it is classified as a high-frequency component.

[0038] Furthermore, the BWO-CNN-ONLSTM model is obtained by optimizing the number of neurons in the ONLSTM layer of CNN-ONLSTM, the number of neurons in the CNN layer of CNN-ONLSTM, and the learning rate of CNN-ONLSTM using the BWO algorithm. The specific operations include:

[0039] Step 4-1: Set the search ranges for the three parameters: the number of neurons in the ONLSTM layer, the number of neurons in the CNN layer, and the learning rate of CNN-ONLSTM, and determine the maximum number of iterations and the number of white whales;

[0040] Step 4-2: Input the high-frequency and low-frequency signals in the reconstructed high-frequency and low-frequency sequences into the CNN-ONLSTM respectively, and use randomly determined parameters to initialize the initial positions of the white whales in the search space [k, β]. Each white whale corresponds to a set of [k, α], where k and α represent the coordinates of the search space.

[0041] Step 4-3: Calculate the fitness of each beluga whale individual, and use the adaptive mechanism to update the balance factor B of each beluga whale individual according to the fitness of each beluga whale individual f and whale fall probability W f , expressed as:

[0042]

[0043]

[0044] Where T represents the current number of iterations, T max Indicates the maximum number of iterations, B0 is a random number between (0,1);

[0045] Step 4-4: Based on the results of the exploration and development phases of the beluga whale, if the calculated value is better than the previous result, the position of the beluga whale is updated according to the formula, otherwise it remains unchanged, which is expressed as:

[0046]

[0047] Where, γ5, γ6, γ7 are random numbers between (0, 1), represents the position of i beluga whale individuals after the T+1th iteration, represents the position of r beluga whales after the Tth iteration, r is a randomly selected beluga whale, X step Represents the step size of the whale's movement, expressed as:

[0048] X step =(u b -l b )exp(-C2T / T max )

[0049] Where C2 is the step size factor related to the whale fall probability and population size, defined as C2 = 2W f *n;u b and l b Represents the upper and lower bounds of a variable;

[0050] Step 4-5: Repeat Step 4-4 until the termination condition is met. After the iteration is completed, the optimal solution calculated during the iteration process is retained;

[0051] Step 4-6: Bring the optimal parameter combination corresponding to the optimal solution into CNN-ONLSTM.

[0052] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0053] (1) To address the problem that the capacity degradation curve of lithium batteries caused by the capacity regeneration phenomenon has strong volatility and nonlinearity, the present invention adopts the TVF-EMD adaptive decomposition method to extract the fluctuation component (high-frequency component) caused by the capacity regeneration phenomenon, while also solving the problem of data complexity;

[0054] (2) The present invention proposes a neural network that combines a convolutional neural network optimized by Beluga Optimization (BWO) with an ordered neuron long short-term memory network (CNN-ONLSTM), which greatly improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flowchart of the method for predicting the remaining useful life of a lithium battery based on the TVF-EMD model and the BWO-CNN-ONLSTM model proposed in the present invention;

[0056] Figure 2 Schematic diagram of degradation curves of four batteries in the NASA lithium battery aging dataset proposed in the present invention;

[0057] Figure 3 This is the result of TVF-EMD decomposition of the B0005 battery set in the NASA lithium battery aging data proposed in this invention;

[0058] Figure 4 A comparison chart of the BWO optimization algorithm proposed in the present invention and other optimization algorithms;

[0059] Figure 5 The prediction results of four batteries in NASA lithium battery aging data proposed by the present invention are shown, as well as a comparison chart with other methods;

[0060] Figure 6 This is a comparison chart of the prediction errors of the method proposed in this invention for four batteries in NASA lithium battery aging data and other methods. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will now be further described with reference to the accompanying drawings and embodiments.

[0062] Example 1:

[0063] like Figure 1 As shown, this embodiment proposes a method for predicting the remaining service life of a lithium battery based on the TVF-EMD model and the BWO-CNN-ONLSTM model, which mainly includes the following steps:

[0064] Step 1: Extract the capacity degradation sequence of the lithium-ion battery from the lithium battery aging dataset. The capacity degradation sequence of the lithium-ion battery contains a complete record of the battery from the initial state to the degradation stage; select a prediction starting point T, the battery discharge data before the prediction starting point T as training data, and the battery discharge data after the prediction starting point T as test data; determine the battery failure threshold (EOL).

[0065] Step 2: Adopt the time-varying filter empirical mode decomposition (TVF-EMD) method to adaptively decompose the capacity degradation sequence of the lithium-ion battery into multiple intrinsic mode functions (IMFs) and residuals, reducing the complexity of the sequence while retaining key characteristics. The specific operations include:

[0066] Step 2-1: Detect all the maximum points from the capacity degradation sequence x(t) of the lithium-ion battery and represent these maximum points as a set u i , where i = 1, 2, 3...

[0067] Step 2-2: Determine the signal interruption point e j , where j = 1, 2, 3... When e j =u i When , there exists a critical value ρ that satisfies the following formula:

[0068]

[0069] Where ρ represents the preset critical value of the frequency change rate between two consecutive maximum values, represents the frequency of bisection. Then e j On the rising edge, is the minimum value; if Then e j On the falling edge, is the minimum value. The rest of can be considered as the peak.

[0070] The critical value is used to determine the rising and falling edges of the interruption point and is an important basis for determining the next local cutoff frequency. Under different rising and falling edge properties, the location of the local minimum of the signal is clear. The interpolation operation is performed based on this precise position information, making the local cutoff frequency more consistent with the instantaneous characteristics of the signal, thereby optimizing the time-varying filtering process and reducing modal aliasing.

[0071] Step 2-3: Find the local cutoff frequency Interpolation between peaks yields Adjustment Dynamically adjust the time-varying filtering characteristics to improve modal aliasing and define more precise boundaries for different frequency components.

[0072] Step 2-4: Reconstruction is performed to obtain a new signal h(t). The original battery sequence is mixed with various frequency changes, such as the overall downward trend and fluctuations caused by capacity regeneration. Integrating the cutoff frequency is phase accumulation in a physical sense. The role of taking cos is to convert the frequency information into an oscillation signal in the time domain. The frequency information of the original signal can be seen from h(t), which provides a basis for filtering in Step 2-5.

[0073]

[0074] Step 2-5: Use a B-spline approximate filter to filter the battery capacity discharge sequence x(t) to obtain the local mean function. The purpose of filtering is to extract the local trend term of the signal, which is the local mean function, which is equivalent to the low-frequency component. Subtracting the local mean function from the original signal, the remaining high-frequency component is then tested in step 6. If it does not meet the requirements, repeat steps 2-2 to 2-5.

[0075] Step 2-6: Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard.

[0076]

[0077] Where B Loughlin (t) represents the instantaneous bandwidth of Loughlin, Represents the weighted average instantaneous frequency. There is a bandwidth threshold ε. If θ(t)≤ε, the signal is an IMF component signal.

[0078] x(t) is decomposed into a finite number of IMF component signals and a residual r(t).

[0079]

[0080] Step 3: Divide the decomposed modal components into high-frequency components and low-frequency components according to the zero-crossing rate, and reconstruct them into high-frequency sequences and low-frequency sequences respectively; the specific operations include:

[0081] Step 3-1: The IMF component signal obtained by adaptively decomposing the battery discharge capacity sequence x(t) through TVF-EMD is classified according to the formula of zero-crossing rate, which is expressed as:

[0082]

[0083] Where, P zero represents the zero-crossing rate, n zero Represents the number of points where the signal crosses zero when changing from positive to negative or vice versa. These points can be determined by examining the signs of adjacent samples. N is the total number of sample points.

[0084] Step 3-2: If P zero If it is less than 0.01, it is classified as a low-frequency component; otherwise, it is classified as a high-frequency component.

[0085] Step 4: Use the optimized ordered neuron long short-term memory convolutional neural network (BWO-CNN-ONLSTM) to model low-frequency sequences and high-frequency sequences respectively, and optimize the model parameters using the Beluga Optimization Algorithm (BWO) to improve prediction accuracy. The specific operations include:

[0086] Step 4-1: Convert the high-frequency and low-frequency signals in the reconstructed high-frequency and low-frequency sequences to floating-point data and normalize them to the range [0, 1]. Use the sliding window method to create a time series dataset and construct the feature sequence and target sequence.

[0087] Step 4-2: Build a time series prediction model (CNN-ONLSTM). This model consists of Conv1D and ONLSTM. The convolutional layer (Conv1D) is used to extract local features in the time series. The ordered neuron long short-term memory layer (ONLSTM) is used to capture long-term dependencies in the time series.

[0088] Step 4-3: Set the search ranges for the three parameters in the BWO algorithm: the number of neurons in the ONLSTM layer, the number of neurons in the CNN layer, and the learning rate of the CNN-ONLSTM layer, and determine the maximum number of iterations and the number of white whales.

[0089] Step 4-4: Input the high-frequency and low-frequency signals in the reconstructed high-frequency and low-frequency sequences into the model respectively, and use randomly determined parameters to initialize the initial position [k, α] of the beluga whale in the search space. Each beluga whale corresponds to a set of [k, α], where k and α represent the coordinates of the search space.

[0090] Step 4-5: Calculate the fitness of each beluga whale individual, and use the adaptive mechanism to update the balance factor B of each beluga whale individual according to the fitness of each beluga whale individual f and whale fall probability W f , expressed as:

[0091]

[0092] Where T represents the current number of iterations, T max Indicates the maximum number of iterations, and B0 is a random number between (0,1).

[0093] Step 4-6: Based on the results of the exploration and development phases of the beluga whale individual, if the calculated value is better than the previous result, the position of the beluga whale individual is updated according to the formula, otherwise it remains unchanged.

[0094]

[0095] Where, γ5, γ6, γ7 are random numbers between (0, 1), represents the position of i beluga whale individuals after the T+1th iteration, represents the position of r beluga whales after the Tth iteration, r is a randomly selected beluga whale, X step Represents the step size of the whale's movement, expressed as:

[0096] X step =(u b -l b )exp(-C2T / T max )

[0097] Where C2 is the step size factor related to the whale fall probability and population size, defined as C2 = 2W f *n.u b and l b Represents the lower and upper bounds of a variable.

[0098] Steps 4-7: Check whether the termination condition has been met, such as finding a satisfactory solution or reaching the maximum number of iterations. Repeat Steps 4-6 until the termination condition is met. After the iteration is complete, retain the optimal solution calculated during the iteration.

[0099] Step 4-8: Bring the optimal parameter combination corresponding to the optimal solution into CNN-ONLSTM to predict the low-frequency component and high-frequency component respectively.

[0100] Among them, the network structure and layer structure of CNN-ONLSTM are shown in Table 1:

[0101] Table 1 Network structure of CNN-ONLSTM

[0102] Layer Name Output shape Activation Function 1D convolutional layer Convolution kernel: 49 ReLU ONLSTM layer Number of neurons: 50 Tanh / Sigmoid Dropout layer Pickup rate: 0.03 N / A ONLSTM layer Number of neurons: 50 Tanh / Sigmoid Dropout layer Pickup rate: 0.03 N / A ONLSTM layer Number of neurons: 50 Tanh / Sigmoid Dropout layer Pickup rate: 0.03 N / A ONLSTM layer Number of neurons: 50 Tanh / Sigmoid Flatten layer Expanded length: 430 N / A

[0103] Figure 4The figure shows the comparison between the BWO optimization algorithm and other optimization algorithms. The experiment selected 6 benchmark functions, and their optimal solutions were all 0. To verify the effectiveness of the BWO algorithm, five mainstream intelligent optimization algorithms, including Grey Wolf Optimizer (GWO), Northern Goshawk Optimization (NGO), Whale Optimization (WOA), Sparrow Search Algorithm (SSA), and Particle Swarm Optimization (PSO), were selected as the control group. The unified parameter configuration was set: the population size was 30 individuals, the maximum number of iterations was 1000 times, and the convergence trajectory of the optimization process was shown in Figure 2. Figure 4 The experimental results show that in the F1, F2, and F4-F6 test functions, the BWO algorithm successfully locates the global optimal solution and converges significantly faster than the control algorithm, while the control algorithm generally suffers from premature convergence. For the F3 function, although both BWO and NGO reach the optimal solution, BWO still has a clear advantage in convergence efficiency.

[0104] Step 5: Add the prediction results of the high-frequency sequence and the low-frequency sequence to obtain the final prediction result of the remaining service life of the lithium-ion battery, and verify the prediction result using the test set. The specific operations include:

[0105] Step 5-1: Fusion the prediction results of the high-frequency and low-frequency components of the test set to obtain the final prediction result x'(t).

[0106] Step 5-2: Calculate the root mean square error (RMSE), mean absolute error (MAE), and relative error (AE) between x'(t) and x(t) to determine the accuracy of the model prediction.

[0107]

[0108]

[0109] AE=|RUL PR -RUL ac |

[0110] Where C j represents the jth true value in the battery capacity degradation sequence x(t), Represents the j predicted values ​​in the test set prediction result x'(t); RUL PR Represents the number of cycles until the predicted value reaches EOL, RUL ac Indicates the number of cycles the actual value reaches EOL.

[0111] Example 2:

[0112] A dataset obtained from the NASA Ames Prediction Center of Excellence (PCOE) records experimental aging data for 18650 lithium-ion batteries, providing critical data support for battery performance research. The dataset accurately captures the three key processes of battery charging, discharging, and impedance measurement, conducted under a constant 24°C experimental environment.

[0113] Using the prediction method disclosed in Example 1, we conducted an in-depth study of four lithium-ion batteries: B0005, B0006, B0007, and B0018. During charging, all four batteries followed a unified charging protocol: The charging process began in constant current (CC) mode at 1.5A. As the battery terminal voltage climbed to 4.2V as the charge accumulated, the charging mode switched to constant voltage (CV) mode. The charging process continued until the charging current decayed to 20mA, ensuring that the batteries reached a fully saturated state.

[0114] During the discharge phase, all four lithium-ion batteries maintained consistent discharge settings, using a constant current (CC) discharge of 2A. However, due to subtle differences in the characteristics of each battery, the discharge cut-off voltage varied. Discharge ended when the voltage of battery B0005 dropped to 2.7V, the voltages of batteries B0006 and B0018 dropped to 2.5V, and the voltage of battery B0007 dropped to 2.2V.

[0115] Figure 2 Figure 2 shows a schematic diagram of the degradation curves of the four batteries; Table 2 records in detail the more detailed parameter information of the four lithium-ion batteries selected this time, providing clear guidance for subsequent analysis and research based on this dataset.

[0116] Table 2 NASA lithium-ion battery parameters

[0117] Battery Model Minimum charging current Constant discharge current Rated capacity Charge / discharge cut-off voltage B0005 20mA 2A 2Ah 4.2V / 2.7V B0006 20mA 2A 2Ah 4.2V / 2.5V B0007 20mA 2A 2Ah 4.2V / 2.2V B0018 20mA 2A 2Ah 4.2V / 2.5V

[0118] The above four batteries were selected to evaluate the prediction method disclosed in Example 1 to test the effectiveness and generalizability of the prediction method. Figure 3 The results of TVF-EMD decomposition of the B0005 battery set are shown; the prediction results for the four batteries are shown in Table 3, which are compared with other prediction methods. It can be seen that the prediction method disclosed in Example 1 has a high accuracy in predicting lithium-ion batteries, can effectively judge their future working capacity, detect problems in time, and avoid losses caused by the battery reaching the end of its life.

[0119] Table 3 Errors of different prediction methods on NASA dataset

[0120]

[0121] Four algorithms, BWO-CNN-ONLSTM, VMD-BWO-CNN-ONLSTM, EMD-CNN-ONLSTM and TVF-EMD-CNN-ONLSTM, are selected for comparison. The experimental results are shown in Figure 5 and Figure 6 , it can be seen that the method proposed in Example 1 is effective and accurate.

Claims

1. A method for predicting the remaining useful life of a lithium battery based on the TVF-EMD model and the BWO-CNN-ONLSTM model, characterized by: The following steps are involved: Step 1: Obtain the capacity degradation sequence of the lithium battery whose service life is to be predicted; Step 2: Use the time-varying filtered empirical mode decomposition method to adaptively decompose the capacity degradation sequence of the lithium battery whose service life is to be predicted into multiple intrinsic mode components and residuals; Step 3: According to the zero-crossing rate, the decomposed intrinsic modal components are divided into high-frequency components and low-frequency components, and reconstructed into high-frequency sequences and low-frequency sequences respectively; Step 4: Input the high-frequency sequence into the BWO-CNN-ONLSTM model trained for high-frequency sequences to obtain high-frequency prediction results; input the low-frequency sequence into the BWO-CNN-ONLSTM model trained for low-frequency sequences to obtain low-frequency prediction results; Step 5: Fuse the high-frequency prediction results and the low-frequency prediction results to obtain the final prediction results; The BWO-CNN-ONLSTM model is obtained by optimizing the number of neurons in the ONLSTM layer of the CNN-ONLSTM, the number of neurons in the CNN layer of the CNN-ONLSTM, and the learning rate of the CNN-ONLSTM using the BWO algorithm. The CNN-ONLSTM includes: a 1D convolution layer, a first ONLSTM layer, a first Dropout layer, a second ONLSTM layer, a second Dropout layer, a third ONLSTM layer, a third Dropout layer, a fourth ONLSTM layer and a Flatten layer.

2. A lithium battery remaining service life prediction method based on the TVF-EMD model and the BWO-CNN-ONLSTM model according to claim 1, characterized in that: In step 2, the time-varying filtered empirical mode decomposition method is used to adaptively decompose the capacity degradation sequence of the lithium battery whose service life is to be predicted into multiple intrinsic mode components and residuals. The specific operations include: Step 2-1: Detect all the maximum points from the capacity degradation sequence x(t) of the lithium battery whose service life is to be predicted, and represent these maximum points as a set u i , where i = 1, 2, 3, ...; Step 2-2: Determine the signal interruption point e j , where j = 1, 2, 3...; when e j =u i When , there exists a critical value ρ that satisfies the following formula: Where ρ represents the preset critical value of the frequency change rate between two consecutive maximum values, represents the frequency of bisection. Then e j On the rising edge, is the minimum value; if Then e j On the falling edge, is the minimum value. The rest of can be considered as the peak; Step 2-3: Interpolate between peaks to obtain the local cutoff frequency Adjustment To improve modal aliasing; Step 2-4: Reconstruct and obtain the new signal h(t); Step 2-5: The frequency information of the original signal is obtained from the new signal h(t). Based on this information, the battery capacity discharge sequence x(t) is filtered using a B-spline approximate filter to obtain a local mean function, which is equivalent to a low-frequency component. The local mean function is subtracted from the battery capacity discharge sequence x(t) to obtain a high-frequency component, which is then tested in Step 2-6. Step 2-6: Calculate the criterion value θ(t) to determine whether the residual signal meets the cutoff standard. If not, repeat Step 2-2 to Step 2-5: Where B Loughlin (t) represents the instantaneous bandwidth of Loughlin, Represents the weighted average instantaneous frequency. There is a bandwidth threshold ε. If θ(t)≤ε, the signal is an IMF component signal. Therefore, x(t) is decomposed into a finite number of IMF component signals and a residual r(t):

3. The method for predicting the remaining useful life of a lithium battery based on the TVF-EMD model and the BWO-CNN-ONLSTM model according to claim 1, wherein: In step 3, the decomposed intrinsic modal components are divided into high-frequency components and low-frequency components according to the zero-crossing rate, and reconstructed into high-frequency sequences and low-frequency sequences respectively. The specific operations include: Step 3-1: According to the zero-crossing rate, the decomposed natural modal components are divided into high-frequency components and low-frequency components, which can be expressed as: Where, P zero represents the zero-crossing rate, n zero Represents the number of points where the signal crosses zero when changing from positive to negative or from negative to positive. N is the total number of sample points. Step 3-2: If P zero If it is less than 0.01, it is classified as a low-frequency component; otherwise, it is classified as a high-frequency component.

4. The method for predicting the remaining useful life of a lithium battery based on the TVF-EMD model and the BWO-CNN-ONLSTM model according to claim 1, wherein: The BWO-CNN-ONLSTM model is obtained by optimizing the number of neurons in the ONLSTM layer of CNN-ONLSTM, the number of neurons in the CNN layer of CNN-ONLSTM, and the learning rate of CNN-ONLSTM using the BWO algorithm. The specific operations include: Step 4-1: Set the search ranges for the three parameters: the number of neurons in the ONLSTM layer, the number of neurons in the CNN layer, and the learning rate of CNN-ONLSTM, and determine the maximum number of iterations and the number of white whales; Step 4-2: Input the high-frequency and low-frequency signals in the reconstructed high-frequency and low-frequency sequences into the CNN-ONLSTM respectively, and use randomly determined parameters to initialize the initial position of the white whale in the search space [k, α]. Each white whale corresponds to a set of [k, α], where k and α represent the coordinates of the search space. Step 4-3: Calculate the fitness of each beluga whale individual, and use the adaptive mechanism to update the balance factor B of each beluga whale individual according to the fitness of each beluga whale individual f and whale fall probability W f , expressed as: Where T represents the current number of iterations, T max Indicates the maximum number of iterations, B0 is a random number between (0,1); Step 4-4: Based on the results of the exploration and development phases of the beluga whale, if the calculated value is better than the previous result, the position of the beluga whale is updated according to the formula, otherwise it remains unchanged, which is expressed as: Where, γ5, γ6, γ7 are random numbers between (0, 1), represents the position of i beluga whale individuals after the T+1th iteration, represents the position of r beluga whales after the Tth iteration, r is a randomly selected beluga whale, X step Represents the step size of the whale's movement, expressed as: X step =(u b -l b )exp(-C2T / T max ) Where C2 is the step size factor related to the whale fall probability and population size, defined as C2 = 2W f *n;u b and l b Represents the upper and lower bounds of a variable; Step 4-5: Repeat Step 4-4 until the termination condition is met. After the iteration is completed, the optimal solution calculated during the iteration process is retained; Step 4-6: Bring the optimal parameter combination corresponding to the optimal solution into CNN-ONLSTM.

Images