Battery capacity degradation curve prediction model based on genetic programming optimization encoder
Through genetic programming optimization encoder instead of variational autoencoder, combined with pruning algorithm and supervised learning, the traditional Seq2Seq model has solved the high parameter quantity and interpretability problems in lithium-ion battery capacity attenuation prediction, improving the accuracy and reliability of battery capacity prediction, and is suitable for new energy vehicles and power grid energy storage.
Patent Information
- Application Number
- CN202510620479.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-14
- Publication Date
- 2025-09-02
AI Technical Summary
The traditional Seq2Seq model has high parameters in the prediction of lithium-ion battery capacity decay, relies on a large amount of labeled data, and lacks interpretability, and cannot measure multiple possibilities and uncertainties, resulting in challenges in practical applications.
Genetic programming optimization encoder is used to replace the encoder part of the variational autoencoder, combine with pruning algorithm to optimize the tree structure, and use supervised learning to train the decoder to predict battery capacity changes.
It improves the interpretability of the model and the generalization of feature extraction, enhances the prediction ability of unknown batteries, provides a reliable basis for battery life management, and is suitable for new energy vehicles and power grid energy storage fields.
Smart Images
Figure CN120577699A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of battery technology, and in particular to a battery capacity degradation curve prediction model based on a genetic programming optimized encoder. Background Art
[0002] Lithium-ion batteries, with their long cycle life, compact size, and high energy density, are the most commonly used type of battery in energy storage packs. However, battery capacity degradation is a core issue affecting their long-term performance. In complex usage scenarios, power batteries can experience a rapid drop in capacity during operation due to issues such as unstable charging voltage, current interruptions, or abnormal temperatures. This can be accompanied by faults such as thermal runaway, physical failure, or internal short circuits. These issues can cause the device to go offline or even cause the battery to explode, seriously endangering personal and property safety. Therefore, monitoring and predicting changes in lithium-ion battery capacity is particularly important. Scientifically assessing capacity status can guide the rational use of batteries and maximize their energy storage performance. It also provides a clear basis for battery recycling and cascade utilization, and is a key link in promoting the development of energy storage technology.
[0003] Currently, there are numerous methods for predicting capacity decay curves. Sequence-to-Sequence (Seq2Seq) models have become a research hotspot due to their superior performance in time series modeling. Seq2Seq models capture key features of the input sequence through an encoder and generate an output sequence using a decoder, effectively handling complex nonlinear dynamic relationships. This demonstrates their broad potential for application in battery capacity prediction.
[0004] Traditional Seq2Seq models typically use neural networks (such as Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU)) as encoders, responsible for extracting features from the input sequence. However, this approach has some limitations. Neural network models such as LSTM and GRU typically have a high number of parameters, and their optimization process requires a large amount of labeled data to achieve stable performance, which poses a challenge in real-world application scenarios with limited data. In addition, the encoder of a traditional Seq2Seq model typically only generates a fixed numerical representation of the input sequence and is unable to measure the multiple possibilities or uncertainties of the prediction results. In addition, another important challenge is that both the encoder and the decoder are essentially black-box models, lacking interpretability, and cannot explicitly analyze the encoding and decoding process. This is also one of the core motivations for using the GP model instead of the encoder to improve the interpretability and transparency of the model.
[0005] In summary, the problems existing in the prior art are as follows:
[0006] Lithium-ion battery capacity degradation affects its long-term performance. In complex scenarios, unstable charging voltage may cause a rapid drop in capacity, accompanied by thermal runaway, physical failure, or internal short circuit failures, seriously endangering personal and property safety.
[0007] Although traditional sequence-to-sequence (Seq2Seq) models have superior performance in time series modeling, the neural networks (such as LSTM and GRU) used in these models have a high number of parameters and their optimization relies on large amounts of labeled data. This poses challenges in practical applications when data is limited.
[0008] The encoder of the traditional Seq2Seq model generates a fixed numerical value to represent the input sequence, which cannot measure the multiple possibilities or uncertainties of the prediction results;
[0009] The encoder and decoder of the traditional Seq2Seq model are black box models, lacking interpretability and making it impossible to explicitly analyze the encoding and decoding process. Summary of the Invention
[0010] To address these technical issues, we proposed an improved Seq2Seq model. This model replaces the encoder portion of a variational autoencoder (VAE) with genetic programming (GP). The tree structure optimized by GP is used to interpret features. The GP-generated tree is then pruned to improve the generalization of the feature extractor. The decoder uses an artificial neural network (ANN) to further predict future capacity trends.
[0011] In order to achieve the above-mentioned purpose, the technical solution adopted in this application is as follows: 1. A battery capacity degradation curve prediction model based on genetic programming optimization encoder, characterized in that it includes the following steps:
[0012] Step S1: Acquire multi-signal time series data during the battery charging and discharging process;
[0013] Step S2: Perform variational autoencoder training and establish and interpret the genetic programming GP encoder;
[0014] Step S3: Pruning and optimizing the tree structure generated by genetic programming GP;
[0015] Step S4: merging the genetically programmed encoders of each signal into a strong encoder;
[0016] Step S5: using the constructed genetic programming GP autoencoder to retrain the decoder to predict the battery degradation trajectory;
[0017] Step S6: Input the capacity data of a normal lithium battery into the trained model to make a prediction and generate a degradation curve of the battery capacity over time.
[0018] In order to better implement the present invention, further, in step S1, a cyclic charge and discharge test is performed on a specific battery under an experimental environment, and various signal data are recorded during the test, including: current data I(t), voltage data V(t), and temperature data T(t);
[0019] Store these signals in time series to form a data set Where c∈{1,2,…,n} is the charging cycle number and k is the signal type.
[0020] In order to better implement the present invention, further, in step S2, it specifically includes:
[0021] S2-1: Here, a variational autoencoder is used to reconstruct the acquired signal. This type of autoencoder can enhance the diversity of potential representation features.
[0022] Using I(t), V(t) and T(t) as input and output, we build autoencoders that can be used to reconstruct current, voltage and temperature respectively.
[0023] Assume that the encoder is represented as f(x k θ e ), then the potential output of the encoder is z k , which is obtained by the mean μ of the Gaussian distribution k and variance Sampling to obtain The mean and variance are obtained by mapping the input signal, which is expressed as follows
[0024]
[0025] in, is the corresponding encoder parameter, f k is the encoder of the corresponding signal; for the decoder, its goal is to k Reconstruction output It can be expressed as in is the parameter of the decoder;
[0026] S2-2: Discard the encoder of the trained autoencoder Keep only the decoder Using genetic programming (GP), the encoder part is reconstructed; for the encoder's latent variable output z k ={z k (1),…zk (n)}, parallel optimization and establishment of n genetic programming trees; for each GP tree Responsible for converting the input data x k Mapped to a dimension z in the latent space k (i), i∈(1,…,n), and the decoder retained by Reconstruct the original signal, that is
[0027]
[0028] S2-3: n GPs are selected, cross-optimized, and compiled simultaneously, generating variables that can be accurately embedded in the latent space generation by generation, and then the decoder reconstructs the original signal. The root mean square error is used as the fitness function in the optimization process.
[0029] In order to better implement the present invention, further, in step S3, it specifically includes:
[0030] S3-1: traverse each node v, calculate the contribution of each node to the tree output and the error change Δε after pruning; the preset threshold ε th resh old =0.01, indicating that the contribution C(v) is less than 1%. The contribution C(v) can be calculated using the following formula:
[0031]
[0032] in, is the tree output after removing node v, if C(v)<ε th resh old If Δε is small enough, the node is considered to contribute little to the overall output of the tree and can be pruned;
[0033] Starting from the leaf nodes of the tree, redundant nodes are pruned from bottom to top; if all child nodes of an internal node are pruned, the internal node is recursively deleted;
[0034] S3-2: After pruning is completed, the tree structure needs to be re-optimized to ensure that the performance of the model is not damaged; update the output expression T of the tree i (x); Evaluate the performance of the pruned model on the validation set and calculate the new mean square error ε and complexity (T i ); if pruning causes a significant performance degradation, some pruning operations are rolled back.
[0035] In order to better implement the present invention, further, in step S4, it specifically includes:
[0036] After pruning is completed in step S3, the GP encoder output T for each signal k (x), whose output is a low-dimensional representation where wk is the embedding dimension of the kth signal; since the latent variable distribution of VAE has been constrained by the KL divergence loss, the potential representations of each signal can be directly concatenated:
[0037] Z concat =[z1,z2,…,z k ].
[0038] In order to better implement the present invention, further, in step S5, the constructed GP encoder is used to retrain the decoder, switching to a supervised learning method to train an autoencoder model for predicting battery degradation trajectory; the specific steps are as follows:
[0039] S5-1: Before retraining, the decoder output needs to be dimensionally aligned with the degraded trajectory using the Piecewise Aggregation Approximation (PAA) method.
[0040] Assume that the trajectory of the training set is y and its length is L; the decoder neural network The output size is n, then the training data set can be reduced in length according to the following formula:
[0041]
[0042] Where n<L;
[0043] S5-2: New encoder optimized and pruned using GP Encode the input measurement signal, the encoder The latent variable Z output by the encoder is encoded; a degradation trajectory prediction model is constructed with the measurement signal as input and the degradation trajectory y as output, namely
[0044]
[0045] S5-3: Set the appropriate data batch size b and use the sample data set in the training set to train the autoencoder The training goal is to automatically adjust the parameters The capacity predicted by the model As close as possible to the true capacity y; the loss function used in the training process is as follows:
[0046]
[0047] In order to better implement the present invention, the variational autoencoder is further reconstructed and trained according to the following loss function:
[0048]
[0049] Among them, q(z k |x k ) is the potential distribution variable after signal k is encoded, p(z k ) is the prior distribution of the latent variable, p(x k |z k ) is the conditional distribution of data after a given latent variable z, β k is the hyperparameter that controls the composite loss function.
[0050] In order to better implement the present invention, further, it is assumed that the input signal x k The length of is p, that is, p=|x k |, then the final fitness function is expressed as follows:
[0051]
[0052] In order to better implement the present invention, further, in step S6,
[0053] The decoder is trained to use The degradation trajectory of lithium-ion batteries is predicted using the input measurable signal; it is expressed as follows:
[0054]
[0055] The technical solution provided by the present invention has the following beneficial effects compared with the prior art:
[0056] 1. The present invention extracts features from multiple signals (such as current, voltage, temperature, etc.) during the battery charging and discharging process, constructs an autoencoder corresponding to the multiple signals, and combines genetic programming to perform display conversion on the encoder, which can describe the attributes of the features from the side. In addition, the tree structure of the genetic programming solution set is utilized, and the pruning algorithm is used to improve the generalization and diversity of the overall model for feature extraction. After the encoder is determined, the output content of the training is adjusted, and the autoencoder is retrained using a supervised learning method to obtain the final battery degradation curve. This integrated method and training method can effectively utilize the interpretability and variability of genetic programming, improve the generalization of the feature extraction process, and thus enhance the prediction ability for unknown batteries. It provides a reliable basis for battery life prediction and health management. This method is applicable to new energy vehicles, grid energy storage and other fields, and has broad application prospects and industrial value. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative labor.
[0058] Figure 1 It is a flow chart of a battery capacity degradation curve prediction model based on genetic programming optimization encoder of the present invention;
[0059] Figure 2 It is a schematic diagram of establishing a variational autoencoder for each signal in the battery capacity degradation curve prediction model based on genetic programming optimization encoder of the present invention;
[0060] Figure 3 This is a schematic structural diagram of a battery capacity degradation curve prediction model based on genetic programming to optimize the encoder, in which the encoder is replaced by GP;
[0061] Figure 4 It is a schematic diagram of pruning optimization of the GP spanning tree in the battery capacity degradation curve prediction model based on genetic programming optimization encoder of the present invention;
[0062] Figure 5 It is a detailed structural diagram of the overall process in the battery capacity degradation curve prediction model based on genetic programming optimization encoder of the present invention. DETAILED DESCRIPTION
[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Generally, the components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.
[0064] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present application. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments in the present application without making any creative efforts shall fall within the scope of protection of the present application.
[0065] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.
[0066] In the description of this application, it should be noted that if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the accompanying drawings, or the orientation or position relationship in which the product of the application is usually placed when in use. It is only for the convenience of describing this application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, it cannot be understood as a limitation on this application. In addition, if the terms "first", "second", etc. appear in the description of this application, they are only used to distinguish the description and cannot be understood as indicating or implying relative importance.
[0067] Furthermore, the use of terms such as "horizontal" and "vertical" in the description of this application does not necessarily imply that a component must be absolutely horizontal or suspended, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical" and does not mean that the structure must be completely horizontal, but rather that it can be slightly tilted.
[0068] It should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood broadly. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal connections between two components. A person of ordinary skill in the art will understand the specific meanings of the above terms in this application based on the specific circumstances.
[0069] Example 1
[0070] The present invention provides a battery capacity degradation curve prediction model based on genetic programming optimization encoder, comprising the following steps:
[0071] S1. Acquire multi-signal timing data during battery charging and discharging;
[0072] S2, Variational Autoencoder Training and Genetic Programming (GP) Encoder Construction and Explanation;
[0073] S3, pruning and optimizing the tree structure generated by GP;
[0074] S4. Combine the genetically programmed encoders of each signal into a strong encoder
[0075] S5. Use the constructed GP autoencoder to retrain the decoder to predict the battery degradation trajectory.
[0076] S6. Input the capacity data of a normal lithium battery into the trained model to make a prediction and generate a degradation curve of the battery capacity over time.
[0077] Furthermore, the specific implementation of step S1 is as follows:
[0078] In an experimental environment, a specific battery is subjected to a cycle charge and discharge test. During the test, various signal data are recorded, including: current data I(t), voltage data V(t), and temperature data T(t). These signals are stored in time series to form a data set. Where c∈{1,2,…,n} is the charging cycle number and k is the signal type.
[0079] The specific implementation of step S2 is:
[0080] S2-1: Here, a variational autoencoder is used to reconstruct the acquired signal. This type of autoencoder can enhance the diversity of potential representation features. With I(t), V(t) and T(t) as input and output, autoencoders for reconstructing current, voltage and temperature are established respectively. Assume that the encoder is represented as f(x k θ e ), then the potential output of the encoder is z k , which is obtained by the mean μ of the Gaussian distribution k and variance Sampling to obtain The mean and variance are obtained by mapping the input signal, which is expressed as follows
[0081]
[0082] in, is the corresponding encoder parameter, f k is the encoder of the corresponding signal. As for the decoder, its goal is to k Reconstruction output It can be expressed as in Parameters for the decoder.
[0083] The variational autoencoder is trained for reconstruction using the following loss function:
[0084]
[0085] Among them, q(z k |x k ) is the potential distribution variable after signal k is encoded, p(z k ) is the prior distribution of the latent variable, p(x k |z k ) is the conditional distribution of data after a given latent variable z, β k is the hyperparameter that controls the composite loss function.
[0086] S2-3: Discard the encoder of the trained autoencoder Keep only the decoder Genetic programming (GP) is used to reconstruct the encoder part. Specifically, for the encoder’s latent variable output z k ={z k (1),…z k (n)}, parallel optimization and establishment of n genetic programming trees. For each GP tree Responsible for converting the input data x k Mapped to a dimension z in the latent space k (i), i∈(1,…,n), and the decoder retained by Reconstruct the original signal, that is
[0087]
[0088] S2-4: n GPs are selected, cross-optimized, and compiled simultaneously, generating variables that can be accurately embedded into the latent space generation by generation, and then reconstructed back to the original signal by the decoder. The root mean square error is used as the fitness function in the optimization process. Assume that the input signal x k The length of is p, that is, p=|x k |, then the final fitness function is expressed as follows:
[0089]
[0090] S2-5: Visualize the optimized n GP tree structures to observe and explain the structure and characteristics of latent space mappings in different dimensions.
[0091] Furthermore, the specific implementation of step S3 is as follows:
[0092] In the second step, the encoder is generated by multi-tree genetic programming (GP), resulting in a multi-tree structure consisting of w trees, each of which represents a function mapping from an input to a low-dimensional latent space. Next, the task of the fourth step is to prune and optimize these generated tree structures.
[0093] S3-1: Traverse each node v, calculate the contribution of each node to the tree output and the error change Δε after pruning. The preset threshold ε th resh old =0.01, indicating that the contribution C(v) is less than 1%. The contribution C(v) can be calculated using the following formula:
[0094]
[0095]
[0096] in is the tree output after removing node v, if C(v)<ε th resh old And if Δε is small enough, it is considered that the node contributes little to the overall output of the tree and can be pruned.
[0097] Starting from the leaf nodes of the tree, redundant nodes are pruned from bottom to top. If all the child nodes of an internal node are pruned, the internal node is recursively deleted.
[0098] S3-2: After pruning is completed, the tree structure needs to be re-optimized to ensure that the performance of the model is not damaged. Update the output expression T of the tree i (x). Evaluate the performance of the pruned model on the validation set and calculate the new mean square error ε and complexity (T i If pruning causes a significant performance degradation, some pruning operations are rolled back.
[0099] The specific steps of step S4 are:
[0100] After pruning is completed in step S3, the GP encoder output T for each signal k (x), whose output is a low-dimensional representation where w k is the embedding dimension of the kth signal. Since the latent variable distribution of VAE has been constrained by KL divergence loss, we can directly concatenate the potential representations of each signal:
[0101] Z concat =[z1,z2,…,z k ]
[0102] The specific steps of step S5 are: using the constructed GP encoder, retraining the decoder, switching to a supervised learning approach, and training the autoencoder model for predicting battery degradation trajectories. The specific steps are as follows:
[0103] S5-1: Before retraining, the decoder output needs to be aligned with the degraded trajectory in terms of dimension according to the Piecewise Aggregation Approximation (PAA) method. Assume that the trajectory of the training set is y and its length is L. Decoder Neural Network The output size is n, then the training data set can be reduced in length according to the following formula:
[0104]
[0105] Where n<L.
[0106] S5-2: New encoder optimized and pruned using GP Encode the input measurement signal, the encoder The latent variable Z output by the encoder is encoded. A degradation trajectory prediction model is constructed with the measurement signal as input and the degradation trajectory y as output, namely
[0107]
[0108] S5-3: Set the appropriate data batch size b and use the sample data set in the training set to train the autoencoder The training goal is to automatically adjust the parameters The capacity predicted by the model As close as possible to the true capacity y. The loss function used in the training process is as follows:
[0109]
[0110] The specific steps of step S6 are:
[0111] The decoder is trained to use The degradation trajectory of lithium-ion batteries is predicted using the input measurable signal. It is expressed as follows:
[0112]
[0113] The beneficial effects of the present invention are as follows: the present invention extracts features from multiple signals (such as current, voltage, temperature, etc.) during the battery charging and discharging process, constructs an autoencoder corresponding to the multiple signals, and combines genetic programming to perform display conversion on the encoder, which can describe the attributes of the features from the side. In addition, the tree structure of the genetic programming solution set is utilized, and the pruning algorithm is utilized to improve the generalization and diversity of the overall model for feature extraction. After the encoder is determined, the output content of the training is adjusted, and the autoencoder is retrained using a supervised learning method to obtain the final battery degradation curve. This integrated method and training method can effectively utilize the interpretability and variability of genetic programming, improve the generalization of the feature extraction process, and thus enhance the prediction ability for unknown batteries. Providing a reliable basis for battery life prediction and health management, this method is applicable to new energy vehicles, grid energy storage and other fields, and has broad application prospects and industrial value.
[0114] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.
Claims
1. A battery capacity degradation curve prediction model based on genetic programming optimized encoder, characterized by: The following steps are involved: Step S1: Acquire multi-signal time series data during the battery charging and discharging process; Step S2: Perform variational autoencoder training and establish and interpret the genetic programming GP encoder; Step S3: Pruning and optimizing the tree structure generated by genetic programming GP; Step S4: merging the genetically programmed encoders of each signal into a strong encoder; Step S5: using the constructed genetic programming GP autoencoder to retrain the decoder to predict the battery degradation trajectory; Step S6: Input the capacity data of a normal lithium battery into the trained model to make a prediction and generate a degradation curve of the battery capacity over time.
2. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to claim 1, characterized in that: In step S1, a specific battery is subjected to a cyclic charge and discharge test in an experimental environment, and various signal data are recorded during the test, including current data I(t), voltage data V(t), and temperature data T(t); Store these signals in time series to form a data set Where c∈{1, 2, ..., n} is the charging cycle number and k is the signal type.
3. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to claim 1, characterized in that: In step S2, it specifically includes: S2-1: Here, a variational autoencoder is used to reconstruct the acquired signal. This type of autoencoder can enhance the diversity of potential representation features. Using I(t), V(t) and T(t) as input and output, we build autoencoders that can be used to reconstruct current, voltage and temperature respectively. Assume that the encoder is represented as f(x k θ e ), then the potential output of the encoder is z k , which is obtained by the mean μ of the Gaussian distribution k and variance Sampling to obtain The mean and variance are obtained by mapping the input signal, which is expressed as follows in, is the corresponding encoder parameter, fk is the encoder of the corresponding signal; for the decoder, its goal is to calculate the latent variable z k Reconstruction output It can be expressed as in is the parameter of the decoder; S2-2: Discard the encoder of the trained autoencoder Keep only the decoder Using genetic programming (GP), the encoder part is reconstructed; for the encoder's latent variable output z k ={z k (1),...z k (n)}, parallel optimization and establishment of n genetic programming trees; for each GP tree Responsible for converting the input data x k Mapped to a dimension z in the latent space k (i), i∈(1,...,n), and the decoder retained by Reconstruct the original signal, that is S2-3: n GPs are selected, cross-optimized, and compiled simultaneously, generating variables that can be accurately embedded in the latent space generation by generation, and then the decoder reconstructs the original signal. The root mean square error is used as the fitness function in the optimization process.
4. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to claim 1, characterized in that: In step S3, it specifically includes: S3-1: traverse each node v, calculate the contribution of each node to the tree output and the error change Δε after pruning; the preset threshold ε threshold =0.01, indicating that the contribution C(v) is less than 1%. The contribution C(v) can be calculated using the following formula: in, is the tree output after removing node v, if C(v)<ε threshold If Δε is small enough, the node is considered to contribute little to the overall output of the tree and can be pruned; Starting from the leaf nodes of the tree, redundant nodes are pruned from bottom to top; if all child nodes of an internal node are pruned, the internal node is recursively deleted; S3-2: After pruning is completed, the tree structure needs to be re-optimized to ensure that the performance of the model is not damaged; update the output expression T of the tree i (x); Evaluate the performance of the pruned model on the validation set and calculate the new mean square error ε and complexity (T i ); if pruning causes a significant performance degradation, some pruning operations are rolled back.
5. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to claim 1, characterized in that: In step S4, it specifically includes: After pruning is completed in step S3, the GP encoder output T for each signal k (x), whose output is a low-dimensional representation where w k is the embedding dimension of the kth signal; since the latent variable distribution of VAE has been constrained by the KL divergence loss, the potential representations of each signal can be directly concatenated: WITH concat =[z1,z2,...,z k ]。 6. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to claim 1, characterized in that: In step S5, the constructed GP encoder is used to retrain the decoder, switching to a supervised learning approach to train an autoencoder model for predicting battery degradation trajectories; The specific steps are as follows: S5-1: Before retraining, the decoder output needs to be aligned with the degraded trajectory in terms of dimension using the segmented set approximation method; Assume that the trajectory of the training set is y and its length is L; the decoder neural network The output size is n, then the training data set can be reduced in length according to the following formula: Among them, n <L; S5-2: New encoder optimized and pruned using GP Encode the input measurement signal, the encoder The latent variable Z output by the encoder is encoded; a degradation trajectory prediction model is constructed with the measurement signal as input and the degradation trajectory y as output, namely S5-3: Set the appropriate data batch size b and use the sample data set in the training set to train the autoencoder The training goal is to automatically adjust the parameters The capacity predicted by the model As close as possible to the true capacity y; the loss function used in the training process is as follows:
7. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to any one of claims 1 to 6, characterized in that: The variational autoencoder is trained for reconstruction using the following loss function: Among them, q(z k |x k ) is the potential distribution variable after signal k is encoded, p(z k ) is the prior distribution of the latent variable, p(x k |z k ) is the conditional distribution of data after a given latent variable z, β k is the hyperparameter that controls the composite loss function.
8. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to any one of claim 3, characterized in that: Assume that the input signal x k The length of is p, that is, p=|x k |, then the final fitness function is expressed as follows:
9. The battery capacity degradation curve prediction model based on genetic programming optimized encoder according to any one of claims 1 to 6, characterized in that: In step S6, The decoder is trained to use The degradation trajectory of lithium-ion batteries is predicted using the input measurable signal; it is expressed as follows: