Battery data enhancement and prediction method, device and equipment

Abstract

The invention discloses a battery data enhancement and prediction method and device, and belongs to the technical field of machine learning, and the method comprises the steps: carrying out the preprocessing of time series data, containing a plurality of features, of a lithium battery, taking a preset feature in the plurality of features as a target feature, and carrying out the slicing of the preprocessed data, so as to construct an original training set; constructing a DMnet diffusion model and performing training, and performing data enhancement processing on the original samples in the original training set by using the DMnet diffusion model after training to obtain enhanced samples; and constructing an STLnet prediction model, selecting the original sample and the enhanced sample as input data according to a preset proportion, carrying out regression training on the STLnet prediction model, and obtaining the trained STLnet prediction model to carry out regression prediction of the target features. According to the method, the STL decomposition, the DMnet diffusion model and the EDAnet are combined, the accuracy of lithium battery target feature prediction is remarkably improved, the requirement for large-scale annotation data is reduced, and the data acquisition cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and particularly to a battery data enhancement and prediction method and device. Background Art

[0002] At present, the new energy vehicle industry mainly dominated by electric vehicles is a national key emerging industry under development, and has also become the focus of enterprise competition and development. In the field of electric vehicles, lithium-ion batteries (lithium batteries) are the most commonly used power battery technologies. Their characteristics of high energy density and relatively fast charging speed make lithium batteries more competitive in the market. The remaining useful life (RUL) of a lithium battery refers to the proportion of the battery's initial capacity that can still be maintained after a certain number of charge and discharge cycles. To improve the economy and safety of lithium batteries, predicting the capacity degradation curve of lithium batteries is of great significance for studying their RUL.

[0003] In recent years, with the development of machine learning and deep learning technologies, data-driven methods have effectively replaced traditional statistical methods and empirical methods. The improvement of computer hardware computing power enables data-driven methods to be more efficiently applied in various fields and has become the mainstream method for time series prediction tasks such as lithium battery RUL prediction.

[0004] Among many methods, the gating mechanism enables RNN and its variants (LSTM and GRU) to outperform early neural networks such as MLP and CNN, playing an important role in data-driven time series prediction. Liu et al. proposed a climate time series prediction model combining STL decomposition and gated recurrent units, effectively improving the prediction accuracy for periodically varying temperature data. Meng et al. proposed a dynamic optimization algorithm for state of charge (SOC) estimation based on GRU-RNN, which effectively improves the speed and accuracy of lithium battery SOC estimation by compromising the current and historical gradient directions in the dynamic gradient method. Wan et al. proposed an improved whale optimization algorithm to optimize the LSTM model, achieving accurate SOC estimation for lithium-ion batteries. Ullah et al. proposed a CNN-LSTM hybrid model for short-term power load prediction. Wang et al. proposed a new LSTM-Informer model for long-term power load prediction, which can capture both short-term temporal correlations and long-term dependencies of data.

[0005] Since the attention mechanism (AT) was proposed, Transformer-based methods have emerged continuously and swept through the fields of computer vision and time series. Zhou et al. designed a Transformer for efficient long short-term time series forecasting (LSTF), which makes the Transformer have lower time complexity, more efficient memory usage, and faster inference speed. Han et al. developed a neural network model based on Transformer, which achieved accurate prediction of the remaining useful life (RUL) of lithium-ion batteries by processing the original data through a denoising autoencoder and capturing time information. Adithya et al. proposed a Transformer model combining double autoencoders and ensemble techniques for early prediction of the remaining useful life of lithium-ion batteries, improving the accuracy and efficiency of prediction through denoising and feature capture. Lim et al. proposed a new attention-based architecture, Temporal Fusion Transformers (TFT), which learns temporal relationships at different scales by combining recurrent layers and self-attention layers and uses gated layers to select relevant features to achieve interpretable multi-level time series prediction. Woo et al. proposed a Transformer based on exponential smoothing, which uses exponential smoothing attention and Fourier frequency attention mechanisms to extract growth, periodicity, and horizontal distribution features from data, achieving good prediction results. Fauzi et al. improved on the ETSformer and proposed a new LSTF method, using an exponential smoothing Transformer with periodic and growth embeddings to predict the health status of lithium batteries.

[0006] The above data-driven methods can effectively achieve time series prediction, but have high requirements for the data volume and data quality of the dataset. However, in the actual production and research of lithium batteries, high-quality battery data is very scarce, and data augmentation techniques can well solve this problem. Data augmentation is the process of expanding the dataset through various technical means, mainly including geometric transformation, noise perturbation and other methods. This process aims to improve the generalization ability of the model, reduce overfitting, and enhance its prediction performance for unseen data.

[0007] The Diffusion Denoising Probabilistic Model (DDPMs), also known as the Diffusion Model (DM), has sparked a craze in the field of computer vision since its emergence, especially in image generation. The basic principle of the diffusion model is to add noise to the original sample to obtain a noisy sample, and then use a noise learning network to learn the noisy sample to predict the added noise, and further reconstruct the original sample using the predicted noise, thereby learning the latent features of the sample. In recent years, researchers have gradually applied the diffusion model to the fields of time series prediction and data augmentation. Rasul et al. proposed an autoregressive model for multivariate probabilistic time series, estimating the gradient of the time step through the diffusion model and predicting samples in the data distribution. Li et al. proposed a bidirectional variational autoencoder that combines diffusion, denoising, and disentanglement, improving the accuracy and interpretability of prediction by increasing data and optimizing latent variables. Solis-Martin et al. confirmed the effectiveness of the diffusion model in data augmentation applications by comparing the denoising generation effects of the diffusion model and the autoencoder. Lee

[20] et al. constructed a prior distribution based on conditional information to reduce the burden of the diffusion model learning the reverse process.

[0008] Methods based on the gating mechanism and the attention mechanism can capture the time dependence of long-term and short-term time series well, but they all complete the prediction task through limited data and are prone to local information loss in the construction of cycle blocks. In the diffusion model, the gradually added noise usually destroys the independent components in the time series, and the data generation process lacks interpretability, requiring further feature extraction and decomposition of the data generated by diffusion. Methods based on signal decomposition (VMD and WT) are also widely used to solve the non-linearity and non-smoothness defects in time series, but their decomposed components often have different significance levels, resulting in information loss. The decomposition method using Loess (STL) can robustly, smoothly, and intuitively decompose the time series into two components: trend and cycle. Generally speaking, the above methods have certain limitations in the lithium battery capacitance prediction task. Summary of the Invention

[0009] The present invention provides a battery data augmentation and prediction method and device to solve at least one defect in the prior art.

[0010] In a first aspect, the present invention provides a battery data augmentation and prediction method, including:

[0011] Preprocess the time series data of the lithium battery containing multiple features, use a preset feature among the multiple features as the target feature, and perform slicing processing on the preprocessed data to construct an original training set;

[0012] Build the DMnet diffusion model and train it. Use the trained DMnet diffusion model to perform data augmentation on the original samples in the original training set to obtain augmented samples;

[0013] Build the STLnet prediction model. Select the original samples and augmented samples as input data according to a preset ratio, and perform regression training on the STLnet prediction model to obtain the trained STLnet prediction model for regression prediction of target features.

[0014] In a second aspect, the present invention also provides a battery data augmentation and prediction device, including:

[0015] An original training set construction module for preprocessing the time series data of a lithium battery containing multiple features, using a preset feature among the multiple features as the target feature, and performing slicing processing on the preprocessed data to construct an original training set;

[0016] A DMnet diffusion model generation module that uses a pre-trained DMnet diffusion model to perform data augmentation on the original samples in the original training set to obtain augmented samples;

[0017] An STLnet prediction model generation module for selecting the original samples and augmented samples as input data according to a preset ratio, performing regression training on the STLnet prediction model, and obtaining the trained STLnet prediction model for regression prediction of target features.

[0018] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the battery data augmentation and prediction method as described in any one of the above are implemented.

[0019] The battery data augmentation and prediction method and device provided by the present invention have the following beneficial effects compared with the prior art:

[0020] (1) The present invention combines STL decomposition, an attention mechanism encoding-decoding architecture, and data augmentation technology, significantly improving the accuracy of lithium battery target feature prediction and the generalization ability of the model, reducing the need for large-scale labeled data, and lowering the data acquisition cost.

[0021] (2) The present invention designs a noise learning network and uses STL decomposition for feature embedding to adjust the network's learning process of noise, and then proposes a diffusion model DMnet to perform data augmentation on the original features, thereby improving the generalization performance of the subsequent prediction model.

[0022] (3) The present invention introduces the STL decomposition method into the encoding and decoding attention model to better discover the feature dependencies of time series, designs the STLnet prediction model, and significantly improves the prediction accuracy of the target feature sequence.

[0023] (4) The present invention combines DMnet and STLnet, and uses the features extracted by STL to dynamically adjust the two processes of data augmentation and prediction, thereby designing a new integrated prediction model SDMnet for lithium battery capacity. Description of the Drawings

[0024] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0025] Figure 1 is one of the schematic flowcharts of the battery data augmentation and prediction method provided by the present invention;

[0026] Figure 2 is the second schematic flowchart of the battery data augmentation and prediction method provided by the present invention;

[0027] Figure 3 is the schematic diagram of the time series data of a lithium battery containing multiple features provided by the present invention;

[0028] Figure 4 is the schematic flowchart of the training process based on the diffusion model provided by the present invention;

[0029] Figure 5 is the schematic diagram of the noise learning network of the U-Net structure provided by the present invention;

[0030] Figure 6 is the schematic diagram of the generation process of DMnet provided by the present invention;

[0031] Figure 7 is the schematic structural diagram of the STLnet prediction model provided by the present invention;

[0032] Figure 8 is the experimental result comparison chart provided by the present invention;

[0033] Figure 9 is the schematic structural diagram of the electronic device provided by the present invention. Detailed Embodiments

[0034] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without making creative efforts fall within the protection scope of the present invention.

[0035] It should be noted that in the description of the embodiments of the present invention, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the presence of additional identical elements in the process, method, article or device including the element. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0036] The following Figures 1 - 9 describes the battery data enhancement and prediction methods and devices provided by the embodiments of the present invention.

[0037] Figure 1 is one of the flow diagrams of the battery data enhancement and prediction method provided by the present invention. As Figure 1 shown, it includes but is not limited to the following steps:

[0038] Step 101: Preprocess the time series data of the lithium battery including multiple features, use the preset feature among the multiple features as the target feature, and perform slicing processing on the preprocessed data to construct an original training set.

[0039] Among them, the original training set includes original samples and sample labels corresponding to the original samples; each original sample includes a feature sequence of multiple preset time steps; the sample label is the value of the target feature at the next time step corresponding to the original sample.

[0040] Step 102: Construct a DMnet diffusion model and train it, and use the trained DMnet diffusion model to perform data enhancement processing on the original samples in the original training set to obtain enhanced samples.

[0041] It can be understood that an enhanced training set corresponding to the original training set can be constructed based on the enhanced samples.

[0042] Step 103: Construct an STLnet prediction model. Select the original samples and augmented samples as input data according to a preset ratio, and perform regression training on the STLnet prediction model to obtain a trained STLnet prediction model for regression prediction of target features.

[0043] Among them, the STLnet prediction model introduces the STL decomposition method into the regression prediction model generated by the encoder-decoder attention model EDAnet.

[0044] It should be noted that STL (Seasonal-Trend Decomposition using Loess) is a time series decomposition method based on local weighted regression (Loess) proposed by Cleveland et al., which is used to decompose a time series into trend, seasonal, and residual components. Details are not elaborated here.

[0045] EDAnet is a neural network model based on encoder-decoder attention, which is used to model the trend and periodicity of wind power data for short-term and long-term wind power prediction. The EDAnet model processes and generates input and output sequences through an encoder and a decoder respectively, and incorporates an attention mechanism model to enable more focus on key information and time steps in the input data. Both the encoder and the decoder are composed of long short-term memory networks (LSTM), which are thus suitable for capturing long-term dependencies and improving the prediction accuracy of sequence data.

[0046] In addition, the present invention can also adjust the preset ratio of the original samples and the augmented samples, train the STLnet prediction model, and test the influence of the preset ratio on the prediction results; among them, the number of samples of the input data is the same as the number of samples in the original training set.

[0047] The present invention combines STL decomposition, attention mechanism encoding-decoding architecture, and data augmentation technology, significantly improving the accuracy of lithium battery target feature prediction and the generalization ability of the model, reducing the need for large-scale labeled data, and lowering the data acquisition cost.

[0048] Before further elaborating on the present invention, first perform mathematical modeling on the problems to be solved by the present invention to facilitate understanding of the work to be completed by the present invention:

[0049] Let X be a set of observed samples of lithium battery input time series subject to a probability distribution, which contains a sequence of N observed samples with m-dimensional features:

[0050]

[0051] Among them Denote the k - dimensional feature sequence, Denote the m - dimensional feature sequence and the target variable y of the i - th known observation sample i , where Y is the set of target variables y that have an implicit mapping relationship M with X.

[0052] The problem of predicting the lithium - battery capacity time series in the present invention is to model the set of feature sequences (X 1 , X 2 , X 3 , …, X N ) m+1 and then find the observed feature variables and the implicit mapping relationship M between the observed feature variables and the future target variable y j+1 , that is, to find the optimal fitting prediction model F for the implicit mapping function M(·).

[0053] To improve the generalization of the prediction model, an enhanced data sample X' is generated by constructing a diffusion model to expand the data volume and improve the data quality. The data enhancement problem in the present invention is to input the original set of feature sequences X into the diffusion model to generate new data X'=(X' 1 , X' 2 , X' 3 , …, X' N ) m+1 The process is as follows:

[0054] Y = M(H(X))≈F(H(X))

[0055] By introducing the conditional feature f = G(X) to participate in the data enhancement process, certain implicit features of the data can be better learned. The conditional feature f has a mapping relationship G with X, and the effect of model data enhancement can be better improved by learning the mapping function G(·). Accordingly, the problems of data enhancement and prediction in the present invention can be expressed as follows:

[0056] Y≈F(H({X, G(X)}, Ψ)).

[0057] Figure 2 is the second schematic flow diagram of the battery data enhancement and prediction method provided by the present invention. The following further describes Figure 2 Steps 101 to 103 above.

[0058] Step 101 is the process of data pre - processing:

[0059] The time - series data of the lithium - ion battery in the present invention, which contains multiple features, can be obtained from the fast - charge and discharge commercial lithium - ion battery dataset provided by the collaborative project between the Toyota Research Institute, Stanford University, and the Massachusetts Institute of Technology. This dataset is used to optimize the fast charging of lithium - ion batteries and contains data of 124 commercial lithium - ion batteries that have undergone fast - charge and discharge cycles until failure. The nominal capacity of the battery is 1.1 Ah, and the nominal voltage is 3.3 V. In the present invention, 625 charge - discharge cycle data of Battery No. 1 in the batch of May 12, 2017 in the dataset are selected for predicting the discharge capacity curve, which includes the time - series data of current I, voltage V, battery temperature T D , charge capacitance Q C and discharge capacity Q D . The sequence of the discharge capacity (i.e., the target feature) Q D is used as the target feature sequence. Figure 3 FIG. is a schematic diagram of the time - series data of the lithium - ion battery in the present invention, which contains multiple features. As Figure 3 shown, the time - series of five features of this dataset are given. It can be observed that there are obvious periodicities in all features except temperature T D . The discharge capacity Q D and charge capacitance Q C degrade continuously with the increase of the charge - discharge cycles, and the curves show a downward trend.

[0060] Optionally, pre - processing is performed on the time - series data of the lithium - ion battery containing multiple features, including: performing max - pooling processing on the time - series data of the lithium - ion battery containing multiple features; performing max - min normalization processing on the time - series data of each feature after max - pooling processing.

[0061] Specifically, since the data volume of the original dataset of a single charge - discharge cycle of the lithium - ion battery is large and contains a large amount of redundant data (such as consecutive 0 values). Therefore, first, max - pooling processing is performed on the original dataset to reduce the single - cycle data points of the dataset to 32 while maintaining the basic characteristics of the data. Then, max - min normalization processing is performed on each feature sequence of the dataset respectively.

[0062] Step 102 is a process of data augmentation based on the DMnet model (including DMnet training and DMnet generation):

[0063] The DMnet model can generate noisy sample sets, denoised sample sets, and predicted noise sets for each step through forward and reverse processes. The noise training network Ψ adjusts the corresponding parameters for network optimization in this process. Figure 4This is a schematic diagram of the training process based on the diffusion model provided by the present invention. The following will refer to Figure 4 to describe the training steps of the DMnet diffusion model of the present invention:

[0064] (1) Noise rate setting

[0065] To enhance the robustness of the model, 20 points are taken within the range of [2, 10] as the noise rate matrix β to amplify the added Gaussian noise. The specific setting of the noise rate matrix β is shown in Table 1:

[0066] Table 1 Noise rate matrix β

[0067]

[0068] (2) Sample forward denoising

[0069] In the forward process, at each diffusion step t, a denoised sample with the corresponding noise rate can be obtained:

[0070] X (t) = X (t-1) + β (t) ∈;

[0071] where the diffusion step t = 1, 2, 3,..., T, and T is a hyperparameter representing the maximum diffusion step; β (t) is the noise rate matrix, and ∈ represents random Gaussian noise (original noise), which follows the distribution

[0072] Thus, a set of denoised samples {X (1) , X (2) , X (3) ,..., X (t) , X (t+1) ,..., X (T)} is obtained. The noise sample at the t-th step can be regarded as the result of continuous addition of progressive noise:

[0073]

[0074] where X (0) is the original sample;

[0075] The denoised sample X (t) is input into the noise learning network Ψ for learning and training to obtain the predicted noise ∈', thus generating a set of predicted noises:

[0076] Z = {∈' (1) , ∈' (2) , ∈' (3) ,..., ∈' (t) , ∈' (t+1) ,..., ∈' (T)}}。

[0077] (3) Noise learning network

[0078] Traditional diffusion models focus on reconstructing the original samples. However, the higher the degree of approximation of the generated samples to the original samples does not necessarily mean it is more beneficial for prediction problems. According to the theory of conditional diffusion models, the original conditional feature f is provided to the noise learning network Ψ to guide and enhance the generation of data.

[0079] Figure 5 is a schematic diagram of the noise learning network of the U-Net structure provided by the present invention, as Figure 5 shown, its inputs are: the noisy sample X (t) , the original conditional feature f, and the diffusion step t. The original conditional feature is generated by performing STL decomposition on the target feature sequence of the original sample. At the same time, positional encoding is performed on t to retain its positional information. Both are added to X after corresponding transformation and activation (t) for layer normalization, and then passed through 4 convolutional layers with a kernel size of 1. After the first and second convolutional layers, the learned f and t are added again to obtain the noise prediction ∈'.

[0080] (4) Sample reverse denoising

[0081] During the reverse process, the diffusion model gradually reconstructs the sample by predicting the noise to obtain the denoised sample X' (t-1) , and finally obtains the generated sample X'. Among them, the denoised sample X' (t-1) is expressed as:

[0082]

[0083] The set {X' (1) , X' (2) , X' (3) ,..., X' (t-1) , X' (t) ,..., X' (T-1)} contains the denoised samples at each step during the denoising process. By comparing the sample losses of the corresponding steps t in the noisy sample and the denoised sample sets, as well as the noise losses between the predicted noise set Z and the original Gaussian noise ∈, the prediction ability of the network Ψ for noise can be further improved, and thus the reconstruction ability of DMnet for samples can be improved.

[0084] (5) STL feature decomposition

[0085] The original conditional feature is obtained by performing STL decomposition on the target feature sequence of the original sample (i.e., Figure 4 the original target feature X in (0) (Q D )); the original conditional feature is obtained by performing STL decomposition on the denoised sample X' (t-1)The target feature sequence in (i.e., Figure 4 The generated target feature X' in (t-1) (Q D )) is subjected to STL decomposition to obtain the generated conditional features; by comparing the two, the training of the network is guided.

[0086] (6) Loss function calculation

[0087] After the above process, the noisy sample X of the previous step can be obtained (t-1) , the denoised sample X' of the previous step (t-1) , the original conditional feature f, the generated conditional feature f', the original noise ∈, and the predicted noise ∈' (t) .

[0088] From this, the sample loss can be calculated based on the noisy sample X (t-1) and the denoised sample X' (t-1) . The feature loss is calculated based on the original conditional feature f and the generated conditional feature f′, and the noise loss is calculated based on the original noise ∈ and the predicted noise ∈'; during the training process, the KL divergence is introduced as the loss function to minimize the training loss, and weights are assigned to each loss item. Thus, the following total loss function definition can be obtained: (t) L = w1D

[0089] (X KL ||X′ (t-1) ) + w2D (t-1) (∈||∈′ KL ) + w3D (t) (f||f′); KL where D

[0090] (P||Q) is the KL divergence of the probability distributions P and Q, and w1, w2, and w3 are weight parameters. KL

[0091] Figure 6 is a schematic diagram of the generation process of DMnet provided by the present invention. The generation process of DMnet is similar to the training process, except that the loss does not need to be calculated and the maximum diffusion step is different. The following combines Figure 6 to illustrate the generation process of DMnet:

[0092] The noise rate r is obtained from the noise rate matrix β. In the forward process, based on each enhancement step k, the noisy sample X generated based on the noise rate r is obtained (k) . When k = 1, X (k) is obtained from the original sample X (0) . When k > 1, X (k) is obtained from the generated sample X′ of the previous step (k-1) , which is expressed as:

[0093]

[0094] Among them, ∈ is random Gaussian noise, and the data augmentation step k = 1, 2, 3,..., K, where K is a hyperparameter representing the maximum data augmentation step;

[0095] Input the noisy sample X (k) , the original conditional feature f, and the data augmentation step t into the already trained noise learning network Ψ to obtain the predicted noise ∈' at the k-th step (k) ; among them, the original conditional feature is generated by STL decomposition of the target feature sequence in the original sample;

[0096] In the reverse process, the diffusion model reconstructs the sample by predicting the noise to obtain the generated sample X' at the k-th step (k) , which is expressed as:

[0097] X′ (k) = X (k) - r∈′ (k) ;

[0098] Thus, a set of generated samples {X' (1) , X' (2) , X' (3) ,..., X' (k-1) , X' (k) ,..., X' (K)} is obtained, and finally the augmented samples are obtained.

[0099] When the noise learning network Ψ has been fully trained and the noise is small enough, the generated sample X' will approach the original sample X. However, under the influence of the original conditional feature f, X' may retain the hidden characteristics of X, and the features are generalized, so as to obtain the augmented samples in the augmented training set to participate in the training of the STLnet prediction model.

[0100] Step 103 is a process of training and regression prediction based on the STLnet prediction model:

[0101] Figure 7 is the structural schematic diagram of the STLnet prediction model provided by the present invention. The following combines Figure 7 to further illustrate step 103.

[0102] (1) STL feature decomposition

[0103] The augmented training set generated by DMnet and the original training set are mixed and input into the prediction model in a certain proportion. STLnet decomposes the target feature sequence (i.e., the Q D sequence) in the input data through STL decomposition to obtain a trend feature f T , periodic feature fS and the residual f R tensor

[0104] (2) Encoding process

[0105] The input data and the tensor respectively pass through the corresponding encoding layers to obtain the hidden states H t and h t at each moment. Then, the input data X t at each moment is input into the encoding block E1, and the conditional feature data at each moment is input into the encoding block E2, where E1 and E2 use the LSTM neural network:

[0106] H t = E1(X t );

[0107]

[0108] (3) Decoding process

[0109] The decoding layer uses temporal attention and LSTM to decode and output the prediction results for the hidden states H t , h t and the cell states S t , s t at different moments. In the traditional encoder-decoder structure, the encoding layer outputs the same context vector at all moments. However, not all moments of data and the hidden states of the hidden layer of the encoding layer contribute equally to the context vector. Therefore, a temporal attention mechanism is adopted to selectively focus on the relevant input sequences. The attention weight vector e i is calculated according to the following formula. This vector can be used to represent the unnormalized input importance:

[0110] e i = tanh(W d [H i ; s t-1 + b d );

[0111] where W d and U d are the weight parameters to be learned by the model. Normalization is performed through the following formula to obtain the attention probabilities of the input sequences at each moment:

[0112]

[0113] Then, the context vector at time t is obtained by weighted summation through the following formula to get the vector x t1 :

[0114]

[0115] Similarly, the vector can then be obtained according to the following formula to get the vector v for the final input to the decoding layer t1 , which can represent the importance of the encoded input variables at different times in predicting the output.

[0116]

[0117] Generally speaking, the input data and tensors respectively pass through the corresponding encoding blocks to obtain the hidden states at different times. The decoding layer uses temporal attention to obtain the inherent temporal correlation between the hidden states, and finally performs regression output to predict the result, and uses the mean absolute error (MAE) as the loss function for regression to train the model.

[0118] The technical solution of the present invention further includes: constructing an original test set; performing data augmentation on the original samples in the original test set; selecting the original samples and augmented samples as input data according to a preset ratio to test the STLnet prediction model.

[0119] Specifically, after the training of DMnet and STLnet is completed, the original samples in the test set data X test are input into DMnet to generate an augmented test set X t ′ est containing augmented samples, then STLnet is used to predict the target feature sequence, and finally evaluation metrics are used to evaluate the model.

[0120] In summary, the original data is augmented by DMnet and then regression is performed using STLnet to obtain the prediction result, thus forming the SDMnet model proposed by the present invention.

[0121] To verify the prediction effect of the present invention, multiple independent experiments were conducted, involving 13 comparison models, as shown in Table 2 below.

[0122] Among them, both ETSformer and SGEformer adopt the geometric transformation data augmentation method and the trend-seasonal feature decomposition method, EDAnet adopts the seasonal block embedding method, and DM_EDAnet adopts the DM data augmentation method. At the same time, the comparison models involve 4 seasonal and trend decomposition methods, 3 data augmentation methods (DM, geometric transformation, and SDMnet), etc. All experiments were completed on an Nvidia RTX 4060Ti 8GB GPU.

[0123] Table 2 Experimental comparison benchmark models

[0124]

[0125] The experiments designed in the present invention use the three evaluation metrics shown in Table 3 to conduct a comparative analysis of different methods. Among them, is the predicted value, and y(t) is the true value. is the mean of the true values, and n is the number of data. The larger the value of R 2 , the smaller the values of RMSE and MAE, indicating higher prediction accuracy.

[0126] Table 3 Error Evaluation Metrics for Prediction Results

[0127]

[0128]

[0129] Figure 8 is the experimental result comparison graph provided by the present invention. Figure 8 It gives the average prediction results of multiple models after 10 independent repeated experiments. Through analysis, it can be seen that:

[0130] 1) The prediction effect of STL-GRU is better than that of STL-LSTMnet. This is because GRU has a lower computational cost than LSTM during training and inference, and reduces the risk of overfitting. However, the prediction effect of EDAnet is better than that of STL-GRU, indicating that LSTM combined with encoder-decoder attention pays more effective attention to the change characteristics of time series, resulting in a lower risk of overfitting. 2) The prediction effect of EDAnet is better than that of models such as Transformer (M4, M5, M6, and M7), verifying the advantages of LSTM combined with encoder-decoder attention in the prediction task of lithium battery capacitance. 3) The prediction effect of STLnet is better than that of EDAnet. This is because STL decomposition accurately captures the periodic changes and trend directions in time series data, and then continuously pays attention to these potential features through encoder-decoder attention, reducing the learning cost of the model for sequence data and improving the prediction efficiency. 4) The prediction effect of DM_EDAnet is better than that of EDAnet, verifying the effectiveness of data augmentation using the diffusion model. 5) SDMnet is better than DM_EDAnet and STLnet, verifying that STL decomposition effectively improves the diffusion model, thereby enhancing the potential features of samples and further improving the generalization ability of STLnet. 6) The R2, RMSE, and MAE of SDMnet are all better than those of other models, indicating that the prediction accuracy of SDMnet is higher than that of other models.

[0131] On the other hand, the present invention also provides a battery data augmentation and prediction device, which includes:

[0132] The first processing module is used to preprocess the time series data of the lithium battery including multiple features, take the preset feature among the multiple features as the target feature, and perform slicing processing on the preprocessed data to construct the original training set;

[0133] The second processing module is used to construct and train the DMnet diffusion model, and use the trained DMnet diffusion model to perform data augmentation processing on the original samples in the original training set to obtain augmented samples;

[0134] The third processing module is used to construct the STLnet prediction model, select the original samples and augmented samples as input data according to a preset ratio, perform regression training on the STLnet prediction model, and obtain the trained STLnet prediction model for regression prediction of the target feature.

[0135] It should be noted that the battery data augmentation and prediction device provided by the embodiments of the present invention can execute the battery data augmentation and prediction method described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0136] The battery data augmentation and prediction method and device provided by the present invention have the following beneficial effects compared with the prior art:

[0137] (1) The present invention combines STL decomposition, attention mechanism encoding-decoding architecture and data augmentation technology, significantly improves the accuracy of lithium battery target feature prediction and the generalization ability of the model, reduces the demand for large-scale labeled data, and lowers the data acquisition cost.

[0138] (2) The present invention designs a noise learning network, uses STL decomposition for feature embedding to adjust the network's learning process of noise, and then proposes a diffusion model DMnet to perform data augmentation on the original features, thereby improving the generalization performance of the subsequent prediction model.

[0139] (3) The present invention introduces the STL decomposition method into the encoder-decoder attention model to better discover the feature dependencies of time series, designs the STLnet prediction model, and significantly improves the prediction accuracy of the target feature sequence.

[0140] (4) The present invention combines DMnet and STLnet, and uses the features extracted by STL to dynamically adjust the two processes of data augmentation and prediction, thereby designing a new lithium battery capacity integrated prediction model SDMnet.

[0141] Figure 9 It is a schematic structural diagram of the electronic device provided by the present invention, as Figure 9As shown, the electronic device may include: a processor 910, a communications interface 920, a memory 930, and a communication bus 940. Among them, the processor 910, the communications interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 may call logic instructions in the memory 930 to execute the battery data enhancement and prediction method.

[0142] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the battery data enhancement and prediction methods provided in the above embodiments.

[0143] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is configured to execute the battery data enhancement and prediction methods provided in the above embodiments.

[0144] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery data enhancement and prediction method, characterized in that: include: Preprocess the time series data of lithium batteries containing multiple features, use the preset features among the multiple features as the target features, and slice the preprocessed data to construct an original training set; Construct a DMnet diffusion model and train it, and use the trained DMnet diffusion model to perform data enhancement processing on the original samples in the original training set to obtain enhanced samples; Construct an STLnet prediction model, select original samples and enhanced samples as input data according to a preset ratio, perform regression training on the STLnet prediction model, and obtain the trained STLnet prediction model to perform regression prediction of the target features.

2. The battery data enhancement and prediction method according to claim 1, characterized in that: The training method of DMnet diffusion model includes: Construct the noise rate matrix β; In the forward process, based on each diffusion step t, a noise sample with a corresponding noise rate is obtained, and a noise sample set {X (1) ,X (2) ,X (3) ,...,X (t) ,X (t+1) ,...,X (T) }; X (t) =X (t-1) +b (t) ∈; Where,∈ is random Gaussian noise, diffusion step t = 1, 2, 3, ..., T, T is a hyperparameter, indicating the maximum diffusion step; Based on the noise sample set, the noise sample of step t is the result of continuous addition of progressive noise: Among them, X (0) is the original sample; The noise sample X (t) Input the noise learning network Ψ for learning and training to obtain the predicted noise ∈', then generate the predicted noise set: Z＝{∈' (1) ,∈' (2) ,∈' (3) ,...,∈' (t) ,∈' (t+1) ,...,∈' (T)}； Among them, the input of the noise learning network is the noise sample X (t) , original conditional features f and diffusion step t, where the original conditional features are generated by STL decomposition of the target feature sequence in the original sample; the noise learning network is a U-Net structure, which performs position encoding on the diffusion step t to retain its position information, and the original conditional features f and diffusion step t are added to X after corresponding transformation and activation. (t) In the layer normalization, after 4 convolution layers with a convolution kernel of 1, the learned f and t are added again after the first and second convolution layers to obtain the noise prediction value ∈'; In the reverse process, the diffusion model progressively reconstructs the sample by predicting the noise, and obtains the denoised sample X' (t-1) , expressed as: Among them, the set {X' (1) ,X' (2) ,X' (3) ,...,X' (t-1) ,X' (t) ,...,X' (T-1) The elements in} are the denoised samples at each step in the denoising process; Based on the noise sample X (t-1) , denoised sample X' (t-1) , original noise ∈, predicted noise ∈' (t) , the original conditional feature f and the generated conditional feature f', calculate the sample loss, noise loss and feature loss; among which, the generated conditional feature is the denoised sample X' (t-1) The target feature sequence in is generated through STL decomposition; According to the sample loss, noise loss and feature loss, the total loss function of the noise learning network is constructed, and the noise learning network is trained with the goal of minimizing the total loss function.

3. The battery data enhancement and prediction method according to claim 2, characterized in that: According to the sample loss, noise loss and feature loss, the total loss function of the noise learning network is constructed, which is: L=w1D KL (X (t-1) ||X′ (t-1) )+w2D KL (∈||∈′ (t) )+w3D KL (f||f′); Among them, D KL (P||Q) is the KL divergence of probability distributions P and Q, and w1, w2, and w3 are weight parameters.

4. The battery data enhancement and prediction method according to claim 1, characterized in that: The generation method of DMnet diffusion model includes: Obtain the noise rate r from the noise rate matrix β; In the forward process, based on each enhancement step k, the noise-added sample X generated based on the noise rate r is obtained. (k) : Where,∈ is random Gaussian noise, data enhancement step k=1, 2, 3, ..., K, K is a hyperparameter, indicating the maximum data enhancement step; The noise sample X (k) , the original conditional features f and the data enhancement step t are input into the trained noise learning network Ψ to obtain the predicted noise ∈' of the kth step (k) ; Among them, the original conditional features are generated by STL decomposition of the target feature sequence in the original sample; In the reverse process, the diffusion model reconstructs the sample by predicting the noise and obtains the generated sample X' in the kth step. ( k ) , expressed as: X′ (k) =X (k) -r∈′ (k) ; Thus, we obtain the generated sample set {X' (1) ,X' (2) ,X' (3) ,...,X' (k-1) ,X' (k) ,...,X' (K) }, and finally the enhanced sample is obtained.

5. The battery data enhancement and prediction method according to claim 1, characterized in that: Regression training of the STLnet prediction model includes: The target feature sequence in the input data is decomposed by STL to obtain a sequence containing trend features f T , periodic characteristics f S and the residual f R Tensor The input data and tensor are respectively passed through the encoding layer corresponding to the STLnet prediction model to obtain the hidden state H at each moment t and h t Then, the input data X at each moment t Input into the encoding block E1, and convert the conditional feature data at each moment Input to the encoding block E2, where E1 and E2 use LSTM neural network: H t =E1(X t ); The decoding layer uses temporal attention and LSTM to the hidden states H at different times t 、h t and the cell state S t 、s t Decode and output the prediction result; The mean square error loss is used as the regression loss function for regression training.

6. The battery data enhancement and prediction method according to claim 1, characterized in that: Preprocess the time series data of lithium batteries containing multiple features, including: Perform maximum pooling on the time series data of lithium batteries containing multiple features; The time series data of each feature after the maximum pooling process is subjected to maximum and minimum normalization respectively.

7. The battery data enhancement and prediction method according to claim 1, characterized in that: Also includes: Adjust the preset ratio of original samples to enhanced samples, train the STLnet prediction model, and test the impact of the preset ratio on the prediction results; The number of samples of the input data is the same as the number of samples of the original training set.

8. The battery data enhancement and prediction method according to claim 1, characterized in that: Also includes: Construct the original test set; Perform data augmentation on the original samples in the original test set; The original samples and enhanced samples are selected as input data according to the preset ratio to test the STLnet prediction model.

9. A battery data enhancement and prediction device, characterized in that: include: The first processing module is used to preprocess the time series data of the lithium battery containing multiple features, take a preset feature among the multiple features as a target feature, and perform slicing processing on the preprocessed data to construct an original training set; The second processing module is used to construct and train the DMnet diffusion model, and use the trained DMnet diffusion model to perform data enhancement processing on the original samples in the original training set to obtain enhanced samples; The third processing module is used to build an STLnet prediction model, select original samples and enhanced samples as input data according to a preset ratio, perform regression training on the STLnet prediction model, and obtain the trained STLnet prediction model to perform regression prediction of the target feature.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the battery data enhancement and prediction method according to any one of claims 1 to 8 are implemented.

Images