A method and device for predicting the remaining life of lithium-ion batteries based on neural network
By using the interactive convolutional block-gate recurrent unit neural network (ICB-GRU) model to predict the remaining life of lithium batteries, the problems of low accuracy, weak data processing capabilities and insufficient prediction advances in the prior art are solved, and high-precision and efficient predictions within the long cycle range are achieved.
Patent Information
- Application Number
- CN202411807876.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2025-04-04
AI Technical Summary
The remaining life prediction model of existing lithium batteries is not accurate and has weak data processing capabilities, so it is impossible to achieve accurate prediction of long cycle ranges, and the prediction is insufficient in advance.
The interactive convolutional block-gate cyclic unit neural network (ICB-GRU) model is used to predict the remaining life of lithium batteries. By long-term rolling, the maximum discharge capacity-cyclic data of the battery is improved to improve the feature extraction effect and prediction accuracy.
It significantly improves the accuracy and versatility of the remaining life prediction of lithium batteries, and can achieve high-precision long-term rolling prediction with less data demand, reduces noise interference, and improves economic benefits and safety.
Smart Images
Figure CN119758095A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of battery health prediction, and in particular to a method and device for predicting the remaining life of a lithium-ion battery based on a neural network. Background Art
[0002] Lithium battery energy storage systems are widely used in the field of deep peak regulation after the access of new energy to the power system due to their high energy density and long cycle life. However, due to the safety hazards of lithium-ion batteries at the end of their life, timely and reliable life prediction methods are particularly important in the field of early warning.
[0003] Existing lithium battery life prediction methods have gradually shifted from the original single physical model to a multi-dimensional data-driven model, among which neural network training methods are the most eye-catching.
[0004] Literature 1 "Prediction of Remaining Service Life of Lithium-ion Batteries Based on CNN-Bi-LSTM Network" (Electric Power Automation Equipment, 2021, Issue 10, Pages 213-219) uses the equal current drop charging time S1 and the equal voltage drop charging time S2 of the charging process to predict the subsequent capacity changes and remaining life of the lithium battery. It shows a certain remaining service life (RUL) prediction ability in small sample data, which is better than the conventional bidirectional long short-term memory (Bi-LSTM) and convolutional neural network (CNN) models in comparison.
[0005] Document 2 "A lithium battery life prediction method based on modal decomposition and machine learning" (Energy Storage Science and Technology, 2022, Issue 12, Pages 3999-4009) uses the fully adaptive noise empirical mode decomposition method (CEEMDAN) to process the reference sample battery capacity data, and uses a simple encoder-decoder (SED) to strengthen the prediction dependency of the long short-term memory neural network (LSTM). The final performance evaluation indicators RMSE and MAPE are both less than 5%.
[0006] Document 3 "Lithium-ion Battery SOH Prediction Based on 1DCNN-LSTM" (Energy Storage Science and Technology, 2022, Issue 11, Pages 240-245) combines one-dimensional convolution (1DCNN) and long short-term memory neural network (LSTM) to predict battery health, and uses the collected multi-channel feature information of charge and discharge for subsequent health estimation. In the NASA data set, 50 cycles and 80 cycles are used as failure points. The final prediction error is less than 2 cycles, and the MAPE and MAE are as low as 1.12% and 0.01, respectively.
[0007] In general, the current practical prediction models for lithium batteries are generally not very accurate and are subject to many restrictions, making it difficult to achieve accurate predictions within a long cycle range. In addition, the data processing capabilities of the prediction models for the remaining life of lithium batteries are generally weak, and most of them are short-term mapping predictions that refer to multiple data indicators. Their actual application range is small, and they cannot achieve complete long-cycle predictions. Finally, the advance amount of the remaining life prediction of lithium batteries is generally insufficient. In order to achieve an accurate remaining life (RUL) prediction, it is usually necessary to cycle more than half of the time and reach the inflection point of lithium battery capacity decay before a more accurate prediction can be achieved. Summary of the invention
[0008] The purpose of the present invention is to propose a method and device for predicting the remaining life of a lithium-ion battery based on a neural network, so as to solve the technical problems that the existing lithium battery remaining life prediction model has low accuracy, weak data processing capability, cannot achieve accurate prediction over a long cycle range, and has insufficient prediction advance time.
[0009] In view of the above problems, the present invention adopts an equivalent model to replace the detailed model for simulation calculation, thereby simplifying the model, improving efficiency, and ensuring the accuracy of the analysis results when analyzing other problems.
[0010] The present invention specifically comprises the following steps:
[0011] S1: Construct a lithium battery remaining life prediction model based on interactive convolutional blocks and gated recurrent units;
[0012] S2: Obtain battery cycle data set;
[0013] S3: Using the data set to train the lithium battery remaining life prediction model to obtain a trained model;
[0014] S4: Obtain online data of lithium-ion batteries and input it into the trained model to obtain prediction results.
[0015] A storage medium stores instructions and data for implementing a method for predicting the remaining life of a lithium-ion battery based on a neural network.
[0016] A lithium-ion battery remaining life prediction device based on a neural network comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a lithium-ion battery remaining life prediction method based on a neural network.
[0017] The beneficial effects provided by the present invention are:
[0018] 1. At present, the practical prediction models of lithium batteries are generally not accurate and subject to many restrictions, making it difficult to achieve accurate predictions within a long cycle range. To address this technical problem, the present invention uses an interactive convolutional block-gated recurrent unit neural network (ICB-GRU) model to process the maximum discharge capacity-cycle data of the battery in a long-term rolling manner; since the interactive convolutional block can simultaneously match convolution kernels of different lengths, the feature extraction effect of this method is better than that of a conventional multi-layer convolutional neural network (CNN), and the final life prediction accuracy is also improved;
[0019] 2. Currently, the data processing capability of the remaining life prediction model of lithium batteries is generally weak, and most of them are short-term mapping predictions based on multiple data indicators. Their actual application scope is small, and they cannot achieve complete long-cycle predictions. In response to this technical problem, the fusion model used in the present invention is simple enough to be applicable to long-term rolling predictions, which provides great help in obtaining subsequent complete capacity changes;
[0020] 3. At present, the advance amount of the remaining life prediction of lithium batteries is generally insufficient. In order to achieve accurate remaining life (RUL) prediction, it is usually necessary to cycle more than half and reach the inflection point of lithium battery capacity decay before a relatively accurate prediction can be achieved. In the neural network model of the present invention, for long cycle conditions, the average data requirement is only 10-20%. In lithium battery cycle data sets such as MIT, the average required prediction data is about 100 cycles (the average total number of cycles is 600-1200 cycles). Under the same cycle starting point, the long-term rolling prediction iteration number of the model of the present invention is less than that of the basic baseline model, ensuring high accuracy while having a certain prediction speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic flow chart of the method of the present invention;
[0022] Figure 2 It is a diagram of the interactive convolutional block-gated recurrent unit (ICB-GRU) neural network structure;
[0023] Figure 3 is a schematic diagram of the gated recurrent unit architecture;
[0024] Figure 4 It is a schematic diagram of the model training process;
[0025] Figure 5 It is a schematic diagram of the prediction results of the long-term iteration model for three groups of batteries;
[0026] Figure 6 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION
[0027] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0028] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.
[0029] Please refer to Figure 1 , Figure 1 It is a schematic diagram of the process of the present invention.
[0030] The present invention provides a method for predicting the remaining life of a lithium-ion battery based on a neural network, comprising the following steps:
[0031] S1: Construct a lithium battery remaining life prediction model based on interactive convolutional blocks and gated recurrent units;
[0032] As an example, please refer to Figure 2 , Figure 2 This is the structure diagram of the interactive convolutional block-gated recurrent unit (ICB-GRU) neural network.
[0033] The lithium battery remaining life prediction model in step S1 includes: a first-layer normalization module, an interactive convolution module, a second-layer normalization module, a gated recurrent unit and an output layer.
[0034] Among them, the Interactive Convolution Block (ICB) is an application method of interactive convolutional networks proposed by Chen Zhenghua and others, the authors of the Time Series Lightweight Adaptive Network (TSLANet). This method can control convolutional neural networks of different kernel sizes with each other, significantly enhancing the model's ability to capture and interpret complex time patterns.
[0035] Specifically, the model includes two convolutional layer structures. The convolutional layers with two different kernels are independent and parallel. The result obtained by the activation function GELU is multiplied by the output value of another convolution layer. The final output values are added and passed through the third convolutional neural network before the corresponding operation results can be output. The structure is as follows Figure 2 shown.
[0036] Generally speaking, the kernel of the first convolution layer of this structure is small, which is used to capture small local details in the time series; in contrast, the kernel of the second convolution layer of this structure is large, which is designed to identify wider and longer-distance temporal dependencies. For the kernel of the third layer of the convolutional neural network, a smaller convolution kernel is usually used. If the overall amount of data is sufficient, it can also be set to a larger value. The relevant interactive convolution block formula is expressed as follows:
[0037] A1=φ(Conv1(S′))⊙Conv2(S′)#(1)
[0038] A2=φ(Conv2(S′))⊙Conv1(S′)#(2)
[0039] O ICB =Conv3(A1+A2)#(3)
[0040] Among them, S′ represents the input data of the model, Conv1, Conv2, and Conv3 represent three different convolutional layer structures (1D-CNN), and φ represents the GELU activation function used.
[0041] The Gated Recurrent Unit (GRU) is a variant of the Long Short-Term Memory (LSTM) neural network. It combines the forget gate and the input gate into an update gate and adds a reset gate to ensure that the model has fewer parameters and converges faster when processing features. This gating mechanism enables the recurrent network to not only remember past information, but also selectively forget some unimportant information and solve certain gradient vanishing problems. Compared with the Long Short-Term Memory Neural Network, the Gated Recurrent Unit (GRU) has higher training efficiency and slightly reduced corresponding training parameters. For the specific model architecture of the Gated Recurrent Unit and Long Short-Term Memory Neural Network, please refer to Figure 3 shown.
[0042] In addition, a layer normalization module (LN) is added before and after the interactive convolutional block structure. This structure is also called the horizontal normalization layer. Compared with conventional batch normalization, it does not require global mean variance statistics and does not require data format padding, so this structure is more suitable for dynamic networks and time series environments.
[0043] S2: Obtain battery cycle data set;
[0044] S3: Using the data set to train the lithium battery remaining life prediction model to obtain a trained model;
[0045] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the model training process.
[0046] The model training process in step S3 is as follows:
[0047] S31, inputting battery cycle data;
[0048] S32, performing Gaussian smoothing on the battery cycle data;
[0049] S33, dividing the smoothed data into windows to obtain training data and verification data;
[0050] S34. Use the training data to train the model, verify the data and evaluate it according to the evaluation indicators to obtain a trained model.
[0051] As an embodiment, in the present invention, by inputting the maximum discharge capacity-cycle curve of the lithium battery, the battery data before the starting point and the reference battery sample cycle data are input into the model, and then they need to be cut and reorganized according to the sliding window size.
[0052] The grid search method is used to find the optimal sliding window and adjust the training and validation data into a standard format.
[0053] In addition, due to the differences in battery capacity changes, the overall noise is very serious, and a relatively reliable smoothing tool must be used for processing before long-term rolling prediction can be performed. The Gaussian kernel smoothing formula used in the present invention is as follows:
[0054]
[0055] Among them, x i represents the center of the kernel function, (x * -x i ) 2 Represents the square of the Euclidean distance from any x to the center point. The σ parameter is related to the degree of smoothing required. The processing effect of Gaussian kernel smoothing needs to consider the actual data volume and error situation.
[0056] Since the data collection situations of different batteries vary greatly, certain model structure and model parameter adjustments should be made based on the actual battery data.
[0057] For the long-term rolling model, the main evaluation indicators of the present invention are as follows (RMSE is root mean square error, MAPE is absolute mean percentage error, R 2 For R 2 Coefficient of determination):
[0058]
[0059] Among them, M is the total number, y m is the true value, For the long-term rolling forecast model, the conventional verification accuracy is set to RMSE < 0.1, R 2 Coefficient of determination R 2 -square>0.9. When this preset condition is met, the preliminary model training is completed; the trained model is fully iterated until the final capacity data meets the battery degradation condition.
[0060] S4: Obtain online data of lithium-ion batteries and input it into the trained model to obtain prediction results.
[0061] As an embodiment, the present invention takes the MIT battery data set as an example to explain the method of the present invention in detail.
[0062] Dataset introduction: The MIT battery dataset includes full cycle data of 124 ternary lithium-ion batteries. The battery model is A123 APR18650M1A, and the default initial maximum capacity is 1.1Ah. The batteries in the MIT dataset all use step-by-step charging and constant-current discharge, and there is a corresponding constant-voltage trickle charge at the end of charging.
[0063] Step 101, establish and optimize the energy storage battery database, ensure that the overall battery data coverage is wide, and ensure the consistency and continuity of the cycle data changes. For data collection, it is necessary to ensure that the collected charge and discharge termination conditions are consistent. If an unexpected situation occurs and the data is interrupted in advance, human intervention should be performed to ensure the accuracy of the data. In the MIT data set, 26 groups of batteries with excessive noise were discarded according to various restrictions, and only 124 groups of battery data with less noise were optimized. In order to verify the model, the third batch of data was uniformly used for testing, and a total of 40 batteries were trained / verified.
[0064] Step 102, for the third batch of lithium-ion battery data, since its cycle range is about 800 to 1200 cycles, the starting point of the test setting is 100 cycles. According to the set starting point, the best sliding window size, learning rate, model kernel size and other parameters are found through grid search. After fine-tuning, the final unified model structure and corresponding parameters are shown in Table 1.
[0065] Table 1 Neural network model parameter settings (based on MIT dataset)
[0066]
[0067] Step 103, training the long-term rolling model. If the training time limit is taken into consideration, the number of iterations or the verification interval can be appropriately adjusted. If there is no time limit, more iterations can be appropriately performed and the corresponding preset standards can be improved. In addition, for the long-term rolling model of the present invention, it is necessary to ensure that the capacity of the starting position is basically corresponding to reduce the initial deviation, but it is also necessary to ensure the initial discrimination, and the setting of the Gaussian smoothing part should be reasonable. For single-iteration prediction training, the iteration standards set by the present invention are RMSE<0.1, R 2 >0.9. For the MIT battery dataset, the proportion of training batteries that reached the preset standard was 92.5%, and 3 batteries still failed to achieve the preset standard after the maximum iteration of 2000 times. For the entire battery set, the final average absolute remaining life error (RULae ) is 40.5 laps, and the average relative residual error (RUL re ) is 4.5%. Some long-term iterative prediction results are as follows Figure 5 shown.
[0068] Step 104, compare the model of this paper with other benchmark models and improved models, and the relevant model comparison results are shown in the following table. The ICB-GRU model selected by the present invention only requires the first 100 cycles of battery normal cycle data as input data, and does not require additional manual screening of feature parameters, but still achieves excellent performance of 40.5 cycles of absolute prediction error and 4.5% of relative remaining life error.
[0069] Table 2 Comparison of model prediction results
[0070]
[0071] See also Figure 6 , Figure 6 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a lithium-ion battery remaining life prediction device 401 based on a neural network, a processor 402 and a storage medium 403.
[0072] A lithium-ion battery remaining life prediction device 401 based on a neural network: The lithium-ion battery remaining life prediction device 401 based on a neural network implements the lithium-ion battery remaining life prediction method based on a neural network.
[0073] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the neural network-based lithium-ion battery remaining life prediction method.
[0074] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the neural network-based lithium-ion battery remaining life prediction method.
[0075] The beneficial effects of the present invention are as follows: the method proposed by the present invention for predicting the remaining life of lithium-ion batteries based on an interactive convolutional block-gated recurrent unit neural network (ICB-GRU) can significantly improve the accuracy and versatility of the model. While ensuring sufficient accuracy, it can be applied to energy storage battery systems with a larger number of samples and a wider cycle range. The life prediction method has an improved lead time relative to traditional methods, has better prediction accuracy, and can significantly reduce noise interference, thereby improving economic benefits and safety.
[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of a lithium-ion battery based on a neural network, characterized in that: Methods include: S1: Construct a lithium battery remaining life prediction model based on interactive convolutional blocks and gated recurrent units; S2: Obtain battery cycle data set; S3: Using the data set to train the lithium battery remaining life prediction model to obtain a trained model; S4: Obtain online data of lithium-ion batteries and input it into the trained model to obtain prediction results.
2. The method for predicting the remaining life of a lithium-ion battery based on a neural network as claimed in claim 1, characterized in that: The lithium battery remaining life prediction model in step S1 includes: a first-layer normalization module, an interactive convolution module, a second-layer normalization module, a gated recurrent unit and an output layer.
3. The method for predicting the remaining life of a lithium-ion battery based on a neural network as claimed in claim 2, characterized in that: The input data passes through the first-layer normalization module, the interactive convolution module, the second-layer normalization module, the gated recurrent unit and the output layer in sequence to obtain the prediction result.
4. The method for predicting the remaining life of a lithium-ion battery based on a neural network as claimed in claim 1, characterized in that: The evaluation indicators of the training model in step S3 include: root mean square error RMSE, absolute average percentage error MAPE, and determination coefficient R 2 .
5. The method for predicting the remaining life of a lithium-ion battery based on a neural network as claimed in claim 4, characterized in that: When the evaluation index is within the preset range or preset conditions, it means that the training of the lithium battery remaining life prediction model is completed.
6. A method for predicting the remaining life of a lithium-ion battery based on a neural network as claimed in claim 5, characterized in that: The model training process in step S3 is as follows: S31, inputting battery cycle data; S32, performing Gaussian smoothing on the battery cycle data; S33, dividing the smoothed data into windows to obtain training data and verification data; S34. Use the training data to train the model, verify the data and evaluate it according to the evaluation indicators to obtain a trained model.
7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a method for predicting the remaining life of a lithium-ion battery based on a neural network as described in any one of claims 1 to 6.
8. A device for predicting the remaining life of a lithium-ion battery based on a neural network, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a method for predicting the remaining life of a lithium-ion battery based on a neural network as described in any one of claims 1 to 6.