Battery remaining life prediction method based on multi-dimensional spatial characteristics
By constructing a convolutional neural network model of four-dimensional tensor input, considering the spatial connection of lithium battery characteristics, the problems of low computational efficiency and error accumulation of lithium battery life prediction methods in the existing technology are solved, and more accurate and efficient Lithium battery life prediction.
Patent Information
- Application Number
- CN202411260366.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-11-19
AI Technical Summary
The existing lithium battery residual life prediction method relies on empirical models and statistical analysis, lacks flexibility, and performs poorly when dealing with complex nonlinear degradation processes, ignoring the spatial connection of individual features, resulting in low computational efficiency and error accumulation. .
Using a prediction method based on multi-dimensional spatial features, the charging and discharging data of lithium batteries is obtained and preprocessed, and the four-dimensional tensor input is constructed. The model is built using a convolutional neural network to consider the spatial connection of individual features, and the feature data and residuals are directly established. The mapping relationship between life expectancy avoids complex calculations and error accumulation caused by iterative prediction.
It improves the accuracy and calculation efficiency of the remaining life prediction of lithium batteries, reduces error accumulation, enhances the prediction capability of the battery management system, extends the battery life and improves the energy use efficiency.
Smart Images

Figure CN118980959A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of battery remaining life prediction, and in particular to a battery remaining life prediction method based on multi-dimensional space characteristics. Background Art
[0002] In the current energy storage field, lithium batteries have become the mainstream choice due to their high energy density and long cycle life. However, the service life of lithium batteries is affected by many factors, such as the number of charge and discharge cycles, ambient temperature, and charging rate. The remaining useful life (RUL) prediction of batteries is of great significance for improving the efficiency and safety of battery management systems. Traditional prediction methods rely on empirical models and statistical analysis, which often lack flexibility and perform poorly for complex nonlinear degradation processes.
[0003] In recent years, deep learning technology has shown excellent performance in many fields due to its advantages in processing nonlinear and high-dimensional data. In particular, in terms of sequence data processing and time series prediction, deep learning models, such as recurrent neural networks (RNN) and long short-term memory networks (LSTM), have been shown to effectively learn the complex dynamic behavior of battery degradation from the historical operation data of the battery. Therefore, the application of deep learning to predict the remaining life of lithium batteries can not only improve the accuracy of the prediction, but also enhance the battery management system's ability to predict future states. This provides strong scientific and technological support for the maintenance and replacement of lithium batteries, helping to extend the service life of batteries and improve the efficiency of energy use.
[0004] However, this type of battery RUL prediction method that uses LSTM, RNN, etc., which considers time series relationships, mainly focuses on the temporal dependencies between a single or multiple features, or focuses on the spatial dependencies between different features, and based on these dependencies, uses a sliding time window for iterative prediction until the end of the battery life. This approach ignores the spatial dependencies of a single feature, and the time series iterative prediction operation is slow and the errors gradually accumulate, which has great limitations in practical applications. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a battery remaining life prediction method based on multi-dimensional spatial features, the purpose of which is to consider the spatial connection between single features themselves to improve calculation efficiency and accuracy.
[0006] The objective of the present invention is achieved through the following technical solutions:
[0007] A method for predicting remaining battery life based on multidimensional spatial features comprises the following steps:
[0008] S1. Obtain battery charge and discharge data, including the original characteristic data of each cycle;
[0009] S2. Data preprocessing, including cleaning abnormal data, interpolating data other than voltage based on the voltage upper and lower limits, and normalizing the data;
[0010] S3. Construct input and output, reconstruct voltage, current and other data into a four-dimensional tensor as input, and take the remaining life corresponding to each cycle as output;
[0011] S4. Divide the training set, validation set and test set;
[0012] S5. Build a convolutional neural network, use the training set data to optimize hyperparameters, use the validation set data to find the optimal prediction model, and test the prediction effect on the test set data.
[0013] Furthermore, step S1 includes the following sub-steps:
[0014] S101. According to the rated capacity of the battery, charge and discharge cut-off voltage and other nominal parameters, the battery is charged and discharged at room temperature. One cycle includes charging and discharging processes. Charging process: the battery is charged with a constant current or a variable current. When the battery voltage reaches the charging cut-off voltage, the battery enters the constant voltage charging stage until the current is less than a certain small current, and the battery stops charging; Discharging process: the battery is discharged with a constant current or a variable current, and the voltage gradually decreases. When the voltage drops to the discharge cut-off voltage, the discharge stops. Continue to cycle until the battery reaches the end of its service life point k. The end of the battery life is defined as the number of cycles corresponding to when the battery capacity reaches 80% of the rated capacity of the battery. There is the following relationship:
[0015]
[0016] Among them C k represents the capacity of the battery at the kth cycle, and C0 represents the rated capacity of the battery. This step must ensure that at least one of the charging and discharging strategies of different batteries is exactly the same.
[0017] S102. Record the original characteristic data such as current, voltage and temperature during each charge and discharge cycle of the battery. Convert the charge (discharge) current data into charge (discharge) capacity Q, the formula is as follows:
[0018]
[0019] Where Q represents the charge (discharge) capacity, I represents the charge (discharge) current, t represents time, and T represents the charge (discharge) duration.
[0020] The original characteristic data during the cycle should include but not be limited to current, voltage, and temperature.
[0021] Furthermore, the step S2 includes the following sub-steps:
[0022] S201. Select the segment data of the same strategy (assuming the same discharge strategy) in step S101 as feature data;
[0023] S202. Perform data cleaning on the feature data, mainly targeting outliers and missing values. Based on the specific data distribution, the cleaning strategies include but are not limited to moving mean filtering, moving median filtering, interpolation, etc.
[0024] S203. Since the degree of battery degradation may be different, even if the discharge strategy is the same, the duration may be different, but different batteries have the same upper and lower limit cut-off voltages. Therefore, the time data is discarded, and the voltage data is expanded from the original collected n samples to N samples within the upper and lower limits. Other feature data are interpolated based on the voltage data so that all feature data have N samples.
[0025] Furthermore, the step S3 includes the following sub-steps:
[0026] S301. Convert the voltage data of (N, 1) in each discharge cycle into a tensor form of (a, b), where:
[0027] N=ab (3)
[0028] The other feature data with N samples are also converted into the same matrix form, and the final model input is a four-dimensional tensor form of (k, a, b, p), where k represents the end of the battery life. The physical meaning of this dimension is the number of battery cycles, and the mathematical meaning is the index number of the (a, b, p) three-dimensional tensor feature. a and b represent the number of rows and columns after the original one-dimensional feature is converted into a two-dimensional tensor, and p represents the number of battery features. Each feature is a (a, b) two-dimensional tensor. This input form can take into account the spatial connection of the battery features themselves, including the connection between rows and rows, rows and columns, and columns in each (a, b) two-dimensional tensor.
[0029] S302. Assuming the current cycle number is m, the RUL corresponding to the current cycle number is
[0030]
[0031] Then RUL is a (k, 1) two-dimensional tensor, and the output of the model is the RUL of the battery.
[0032] S303. Normalize the input and output data. The normalization method is as follows:
[0033]
[0034] x is the original feature data, x' is the normalized data, max and min represent the maximum and minimum values in the feature data respectively.
[0035] Furthermore, the step S4 comprises the following steps:
[0036] S401. Assuming that there are a total of Z battery data, the input and output combinations of the model can be obtained according to step S3 as follows:
[0037]
[0038] where k i represents the end-of-life point of the ith battery. The battery contains K cycles of data in total, and the calculation formula for K is as follows:
[0039]
[0040] S402. Divide 80% of the battery data into a training set, 10% of the battery data into a validation set, and 10% of the battery data into a test set. The data set distribution is as follows:
[0041]
[0042] in
[0043] z=Z*0.8
[0044] z′=Z*0.1 (9)
[0045] Furthermore, the step S5 comprises the following steps:
[0046] S501. Construct a convolutional neural network (CNN), wherein the CNN includes an input layer, a convolution layer, a pooling layer, a fully connected layer, and a regression layer.
[0047] S502. The size of the input layer in S501 is (a, b, p), and the number of convolutional layers is The size of the convolutional layer The layer size is also The batch size for training is The maximum number of training iterations is The learning rate for training is
[0048] S503. Perform training on the training set in S4. After the training results converge, calculate the root mean square error (RMSE) on the validation set to evaluate the training effect. The calculation formula is as follows:
[0049]
[0050] RUL * Represents the model output value, and RUL represents the actual remaining service life of the battery.
[0051] S504. Taking the minimization of RMSE on the validation set as the optimization goal, The six hyperparameters are used as optimization quantities for hyperparameter optimization to obtain the optimal values of the hyperparameters and complete the establishment of the life prediction model.
[0052] S505. Verify the actual prediction accuracy of the established life prediction model on the test set.
[0053] The advantages and positive effects of the present invention are:
[0054] The method of the present invention takes into account the spatial connection of individual features themselves, and directly establishes a mapping relationship between the feature data of the current cycle and the remaining service life, thereby avoiding the complex calculation and error accumulation caused by iterative prediction using a time window. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 is a flow chart of the method of the present invention;
[0056] Figure 2 The data collected during the discharge process in the embodiment of the present invention;
[0057] Figure 3 Schematic diagram of spatial reconstruction of features in an embodiment of the present invention. DETAILED DESCRIPTION
[0058] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.
[0059] like Figure 1 As shown, a method for predicting the remaining battery life based on multidimensional space features specifically includes the following steps:
[0060] S1. Obtain battery charge and discharge data, including the original characteristic data of each cycle;
[0061] S101. Charge and discharge the battery at room temperature. Charging process: charge the battery with a constant current or a variable current. When the battery voltage reaches the charging cut-off voltage, the battery enters the constant voltage charging stage until the current is less than a certain small current, and the battery stops charging. Discharging process: the battery discharges at a constant current, and the voltage gradually decreases. When the voltage drops to the discharge cut-off voltage, the discharge stops. The cycle continues until the battery reaches the end of its service life. The discharge strategies of different batteries are exactly the same.
[0062] S102. Record the original characteristic data such as current, voltage and temperature during each charge and discharge cycle of the battery. Convert the discharge data into discharge capacity Q, the formula is as follows:
[0063]
[0064] Where Q represents the discharge capacity, I represents the discharge current, t represents the time, and T represents the discharge duration.
[0065] The discharge strategy in this embodiment remains consistent, and only voltage, current, and temperature characteristic data are collected. Those skilled in the art will appreciate that other charging and discharging strategies that meet the requirements, or the collection of more characteristic data such as internal resistance, can all be applied to the remaining service life prediction method provided by the present invention.
[0066] S2. Data preprocessing, including cleaning abnormal data, interpolating data other than voltage based on the voltage upper and lower limits, and normalizing the data;
[0067] S201. Select the same strategy segment data in step S101 as feature data, which is the discharge stage data in this embodiment;
[0068] S202. Perform data cleaning on the feature data, mainly targeting outliers and missing values. Based on the specific data distribution, the cleaning strategies include but are not limited to mean regression, median regression, and interpolation. In this embodiment, moving median filtering is used for outliers, and linear interpolation is used for missing values.
[0069] Those skilled in the art will appreciate that, depending on the specific data distribution, other data processing methods may be applicable to the remaining useful life prediction method provided by the present invention.
[0070] S203. Since the battery degradation degree may be different, even if the discharge strategy is the same, the duration may be different, but different batteries have the same upper and lower limit cut-off voltages. Therefore, the time data is discarded, and the voltage data is expanded from the original collected n sample data to N samples within the upper and lower limits. Other feature data are interpolated based on the voltage data so that all feature data have N samples. In this embodiment, n is 247, N is 800, and the feature data of the discharge process in a certain cycle after interpolation is as follows: Figure 2 Those skilled in the art will appreciate that N has no fixed value limit and may vary according to specific data distribution.
[0071] S3. Construct input and output, reconstruct voltage, current and other data into a four-dimensional tensor as input, and take the remaining life corresponding to each cycle as output;
[0072] S301. Convert the voltage data (800,1) in each discharge cycle into a tensor form of (40,20), and convert the other feature data with 800 samples into the same matrix form. Then the input of the final model is a four-dimensional tensor form of (k,40,20,3), where k represents the end point of the battery's life. The physical meaning of this dimension is the number of cycles of the battery, and the mathematical meaning is the index number of the (40,20,3) three-dimensional tensor feature. 40 and 20 represent the number of rows and columns after the original one-dimensional feature is converted into a two-dimensional tensor, and 3 represents the number of battery features. In this embodiment, it represents the three features of voltage, temperature, and discharge capacity. Each feature is a (40,20) two-dimensional tensor. This input form can take into account the spatial connection between the battery features themselves, including the connection between rows and rows, rows and columns, and columns and columns in each (40,20) two-dimensional tensor. The schematic diagram is as follows Figure 3 As shown. It can be understood by those skilled in the art that in the process of converting (N, 1) to (a, b), there is no fixed value restriction for a and b, and they may vary according to the specific data distribution.
[0073] S302. Assuming the current cycle number is m, the RUL corresponding to the current cycle number is
[0074]
[0075] Then RUL is a (k, 1) two-dimensional tensor, and the output of the model is the RUL of the battery.
[0076] S303. Normalize the input and output data. The normalization method is as follows:
[0077]
[0078] x is the original feature data, x' is the normalized data, max and min represent the maximum and minimum values in the feature data respectively.
[0079] S4. Divide the training set, validation set and test set;
[0080] S401. In this embodiment, there are a total of 40 battery data. According to step S3, the input and output combinations of the model can be obtained as follows:
[0081]
[0082] where k i represents the end-of-life point of the ith battery. The 40 batteries contain K cycles of data in total, and the calculation formula for K is as follows:
[0083]
[0084] In this embodiment, K is calculated to be 41240.
[0085] S402. Divide 80% of the battery data into a training set, 10% of the battery data into a validation set, and 10% of the battery data into a test set. The data set distribution is as follows:
[0086]
[0087] S5. Build a convolutional neural network, use the training set data to optimize hyperparameters, use the validation set data to find the optimal prediction model, and test the prediction effect on the test set data;
[0088] S501. Construct CNN, which includes: input layer, convolution layer, pooling layer, fully connected layer and regression layer. Those skilled in the art can understand that this embodiment only uses the basic CNN network structure, and techniques such as adding normalization layer, nonlinear layer, dropout layer, etc. to the network to improve computing efficiency are also within the framework of the life prediction algorithm proposed in this patent.
[0089] S502. The size of the input layer in S501 is (40, 20, 3). There can be multiple convolutional layers and pooling layers, which are arranged in order and finally connected to the fully connected layer. The number of convolutional layers is The size of the convolutional layer The pooling layer size is The batch size for training is The maximum number of training iterations is The learning rate for training is
[0090] S503. Perform training on the training set in S4. After the training results converge, calculate the root mean square error (RMSE) on the validation set to evaluate the training effect. The calculation formula is as follows:
[0091]
[0092] RUL * Represents the model output value, and RUL represents the actual remaining service life of the battery.
[0093] S504. Taking the minimization of RMSE on the validation set as the optimization goal, The six hyperparameters are used as optimization quantities for hyperparameter optimization. The grid search algorithm is used to obtain the optimal values of the hyperparameters and complete the establishment of the life prediction model. The optimal values are shown in the following table:
[0094]
[0095] Those skilled in the art can understand that the super-optimization algorithm is not limited to the grid search in this embodiment, and other super-optimization algorithms are also within the framework of the life prediction algorithm proposed in this patent.
[0096] S505. Check the actual prediction accuracy of the established life prediction model on the test set. In summary, a battery remaining life prediction method based on multi-dimensional spatial features takes into account the spatial connection of a single feature itself, and directly establishes a mapping relationship between this connection and the remaining service life, avoiding the complex calculation and error accumulation caused by the iterative prediction using time windows by RNN, LSTM and other methods.
[0097] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.
Claims
1. A method for predicting remaining battery life based on multidimensional spatial features, characterized in that: The following steps are involved: S1. Obtain battery charge and discharge data, including the original characteristic data of each cycle; S2. Data preprocessing, including cleaning abnormal data, interpolating data other than voltage based on the voltage upper and lower limits, and normalizing the data; S3. Construct input and output, reconstruct voltage, current and other data into a four-dimensional tensor as input, and take the remaining life corresponding to each cycle as output. The input of the model is in the form of a four-dimensional tensor of (k, a, b, p), where k represents the end point of the battery's life, a and b represent the number of rows and columns after the original one-dimensional features are converted into a two-dimensional tensor, and p represents the number of battery features. The output of the model is the RUL of the battery, which is a (k, 1) two-dimensional tensor; S4. Divide the training set, validation set and test set; S5. Build a convolutional neural network, use the training set data to optimize hyperparameters, use the validation set data to find the optimal prediction model, and test the prediction effect on the test set data.
2. The method for predicting remaining battery life based on multidimensional spatial features according to claim 1, characterized in that: Step S1 includes the following sub-steps: S101. According to the rated capacity of the battery, the charge and discharge cut-off voltage, the battery is charged and discharged at room temperature; S102. Record the original characteristic data of current, voltage and temperature during each charge and discharge cycle of the battery, and convert the charge and discharge current data into charge and discharge capacity.
3. The method for predicting remaining battery life based on multidimensional spatial features according to claim 2, characterized in that: The step S2 comprises the following sub-steps: S201. Select the same strategy segment data in step S101 as feature data; S202. Perform data cleaning on the feature data, mainly targeting outliers and missing values. Based on the specific data distribution, the cleaning strategies include moving mean filtering, moving median filtering, and interpolation; S203. Abandon the time data, expand the voltage data from the originally collected n sample data to N samples within the upper and lower bounds, and interpolate other feature data based on the voltage data so that all feature data have N samples.
4. The method for predicting remaining battery life based on multidimensional spatial features according to claim 3, characterized in that: The step S3 comprises the following sub-steps: S301. Convert the voltage data (N, 1) in each discharge cycle into a tensor form of (a, b), where N = ab. Convert other feature data with N samples into the same matrix form. Then the input of the final model is a four-dimensional tensor form of (k, a, b, p), where k represents the end point of the battery life. The physical meaning of this dimension is the number of cycles of the battery, and the mathematical meaning is the index number of the three-dimensional tensor feature (a, b, p). a and b represent the number of rows and columns after the original one-dimensional feature is converted into a two-dimensional tensor, and p represents the number of battery features. Each feature is a two-dimensional tensor of (a, b). This input form takes into account the spatial connection of the battery features themselves, including the connection between rows and rows, rows and columns, and columns in each (a, b) two-dimensional tensor; S302. Assuming the current cycle number is m, the RUL corresponding to the current cycle number is Then RUL is a (k, 1) two-dimensional tensor, and the output of the model is the RUL of the battery; S303. Normalize the input and output data.
5. The method for predicting remaining battery life based on multidimensional spatial features according to claim 4, characterized in that: The step S4 comprises the following steps: S401. Assuming that there are a total of Z battery data, the input and output combinations of the model can be obtained according to step S3 as follows: where k i represents the end-of-life point of the ith battery, then the battery contains a total of K cycles of data, and the calculation formula for K is as follows: S402. Divide 80% of the battery data into a training set, 10% of the battery data into a validation set, and 10% of the battery data into a test set. The data set distribution is as follows: in Z 2Z*0.8 z'=Z*0.1(9).
6. The method for predicting remaining battery life based on multidimensional spatial features according to claim 5, characterized in that: The step S5 comprises the following steps: S501. Construct a convolutional neural network CNN, which includes: an input layer, a convolution layer, a pooling layer, a fully connected layer and a regression layer; the size of the input layer is (a, b, p), the number of convolution layers is The size of the convolutional layer The pooling layer size is The batch size for training is The maximum number of training iterations is The learning rate for training is S502. Perform training on the training set in S4, and after the training results converge, calculate the root mean square error RMSE on the validation set to evaluate the training effect; S503. Taking the minimization of RMSE on the validation set as the optimization goal, The six hyperparameters are used as optimization quantities to perform hyperparameter optimization, obtain the optimal values of the hyperparameters, and complete the establishment of the life prediction model; S504. Verify the actual prediction accuracy of the established life prediction model on the test set.
