SOH prediction method based on filter and DLinear model

Abstract

The invention discloses an SOH prediction method based on a filter and a DLi near model, and relates to the technical field of battery life prediction, and the method comprises the steps: S1, data preprocessing, S2, data standardization, S3, wavelet denoising, and S4, DLi near model prediction. According to the method, effective components and high-frequency noise of the signals can be accurately distinguished through multi-scale decomposition, nonlinear threshold processing and wavelet denoising; specifically, the low-frequency component is reserved to reflect the long-term attenuation trend of the battery capacity, and the high-frequency noise component is effectively eliminated through threshold screening; compared with a traditional Fourier transform or linear filtering method, the method has the advantages that global features and local abrupt change details of signals can be captured at the same time due to the tight supporting performance and the high vanishing moment characteristic of the wavelet basis function, and the problem that the signals are fuzzy or key information is lost is avoided; according to the design, the data quality is improved, a purer input signal is provided for a subsequent model, and therefore interference of noise on prediction precision is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery life prediction, and in particular relates to a SOH prediction method based on a filter and a DLinear model. Background Art

[0002] Lithium-ion batteries, as core energy storage units in the new energy sector, are widely used in electric vehicles, energy storage systems, consumer electronics, and other fields. Battery state of health (SOH) is a key indicator for measuring battery performance degradation and remaining life. Accurate SOH prediction is crucial to the safety, reliability, and cost-effectiveness of battery management systems (BMS).

[0003] Current SOH prediction methods have limitations when dealing with non-stationary time series data. Traditional statistical methods struggle to effectively capture the nonlinear characteristics and short-term fluctuations of battery capacity decay, resulting in insufficient prediction accuracy. Furthermore, battery data often contains significant noise, further interfering with model performance. While methods such as Fourier transforms and linear filtering have been attempted for denoising, they are unable to simultaneously preserve both global trends and local details in the signal, thus compromising the reliability of prediction results. To address these issues, the following solutions are proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a SOH prediction method based on a filter and a DLinear model, which can simultaneously capture the global characteristics and local mutation details of the signal through the compact support and high vanishing moment characteristics of the wavelet basis function, thereby solving the problem of signal ambiguity or loss of key information in the prior art.

[0005] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0006] The present invention is a SOH prediction method based on a filter and a DLinear model, comprising:

[0007] Step S1, data preprocessing: preprocess the initial battery data (such as current, voltage, SOC, etc.) to obtain the SOH (State of Health) data corresponding to the current time point of the battery data;

[0008] Step S2, data standardization: Use StandardScaler to standardize the SOH time series data to eliminate dimensional differences and make the data mean 0 and variance 1;

[0009] Step S3, wavelet denoising: putting the normalized data into wavelet denoising to remove noise;

[0010] Step S4, DLinear model prediction: decompose the time series data into trend terms and seasonal terms, use linear layer fitting respectively, and dynamically input historical data through the sliding window mechanism to predict the SOH value of the next n cycles.

[0011] Furthermore, the data preprocessing step S1 specifically includes the following steps:

[0012] Step S11, raw data conversion: converting the raw charge and discharge data of the lithium-ion battery into SOH data. The specific formula is:

[0013]

[0014] Where Δt is the time interval, I is the current during charging, t1 and t2 represent the time of starting and ending charging respectively, and C n is the current actual capacity of the lithium-ion battery, and C0 represents the rated capacity of the battery;

[0015] Step S12, constructing time series data: arranging the converted SOH data in chronological order to form an SOH time series sequence.

[0016] Furthermore, in step S2, the normalization formula for data normalization is:

[0017]

[0018] Where z is the standardized data value, x is the original SOH data value, μ is the mean, and σ is the standard deviation.

[0019] Furthermore, the step S3, wavelet denoising, specifically includes the following steps:

[0020] Step S31, wavelet basis function selection: selecting a wavelet basis function with compact support and high vanishing moment (such as Daubechies series) to adapt to the non-stationary characteristics of lithium-ion battery SOH data;

[0021] Step S32, wavelet decomposition: perform multi-layer wavelet decomposition on the standardized SOH data to decompose the signal into low-frequency (trend component) and high-frequency components (noise and detail components). The decomposition formula is:

[0022]

[0023] Where w j,k is the wavelet coefficient, ψ j,k (t) is the wavelet basis function, j and k are scale and translation parameters respectively, and f(t) is the array for wavelet transform;

[0024] Step S33, threshold denoising: Calculate the threshold and use the general threshold formula to determine the noise threshold T, which is:

[0025]

[0026] Where cD1 is the first layer wavelet coefficient, N is the data length;

[0027] The wavelet coefficients (signal components) above the threshold are retained, and the coefficients (noise components) below the threshold are discarded;

[0028] Step S34, wavelet reconstruction: performing inverse wavelet transform on the denoised low-frequency and high-frequency coefficients to reconstruct the denoised SOH time series data.

[0029] Furthermore, the step S4, DLinear model prediction specifically includes the following steps:

[0030] Step S41, sequence decomposition: decompose the denoised SOH time series data into trend terms and seasonal terms, and use a sequence decomposition module (such as a moving average method) to separate long-term trends and short-term fluctuations;

[0031] Step S42: Model construction: specifically:

[0032] Trend item prediction: Apply linear layer fitting to the trend item, the formula is:

[0033]

[0034] Where, is the predicted output of the trend term, W trend is the weight matrix of the linear layer, x trend Enter data for the trend item obtained after series decomposition;

[0035] Seasonal term prediction: Another linear layer is applied to fit the seasonal term, the formula is:

[0036]

[0037] Where, is the forecast output of seasonal terms, W seasonal is the weight matrix of the linear layer, x seasonal Enter data for the seasonal terms obtained after series decomposition;

[0038] Result fusion: Add the forecast results of trend item and seasonal item to get the final forecast value:

[0039]

[0040] Where, is the final predicted SOH value;

[0041] Step S43, sliding window mechanism: Set the sliding window length (e.g., 96 cycles), and each time historical data in the window is input, predict the SOH value of the next n cycles. After the prediction is completed, the window slides backward by a fixed step size (e.g., 25 cycles) to cover the new data and repeat the prediction;

[0042] Step S44, post-processing of the prediction results: averaging the multiple sliding window prediction results at the same time point to eliminate the random error of a single prediction, and converting the normalized prediction results back to the original dimension to obtain the final SOH prediction value.

[0043] The present invention has the following beneficial effects:

[0044] 1. The present invention demonstrates advantages in noise processing of non-stationary time series data by adopting wavelet denoising technology. Through multi-scale decomposition and nonlinear threshold processing, wavelet denoising can accurately distinguish the effective components of the signal from high-frequency noise. Specifically, low-frequency components are retained to reflect the long-term attenuation trend of battery capacity, while high-frequency noise components are effectively eliminated through threshold screening. Compared with traditional Fourier transform or linear filtering methods, the compact support and high vanishing moment characteristics of the wavelet basis function enable it to simultaneously capture the global characteristics and local mutation details of the signal, avoiding the problems of signal blur or loss of key information. This design improves data quality and provides a purer input signal for subsequent models, thereby reducing the interference of noise on prediction accuracy.

[0045] 2. The present invention decomposes time series data into trend terms and seasonal terms through the DLinear model, realizing multi-dimensional modeling of battery health status; the linear fitting of trend terms focuses on the long-term law of capacity attenuation, while the seasonal terms capture the short-term fluctuation characteristics. The combination of the two takes into account both the global and local characteristics of time series data; in addition, the introduction of the sliding window mechanism enables the model to dynamically update the input data and adapt to the gradual change of battery performance in real time; through the rolling input of historical data and the gradual correction of the prediction results, the model effectively alleviates the prediction deviation of traditional static models caused by data lag while maintaining sensitivity to the current state; this design not only improves the model's ability to parse complex time series patterns, but also enhances its stability in long-term prediction tasks.

[0046] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 Schematic diagram of a flow chart of a SOH prediction method based on a filter and a DLinear model according to the present invention;

[0049] Figure 2 is the SOH curve and absolute error of the short time series of the present invention;

[0050] Figure 3 is the SOH curve and absolute error of the long time series of the DLinear model of the present invention;

[0051] Figure 4 This is a flow chart of wavelet denoising of the present invention;

[0052] Figure 5 This is the original data diagram of the present invention;

[0053] Figure 6 1 is a comparison diagram of the denoising results of the SOH after denoising (a) and the SOH before denoising (b). DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0055] See also Figure 1-6 As shown, the present invention is a SOH prediction method based on a filter and a DLinear model, comprising:

[0056] Step S1: Data preprocessing:

[0057] The raw charge and discharge data of lithium-ion batteries are converted into SOH data. The specific formula is:

[0058]

[0059] Where Δt is the time interval, I is the current during charging, t1 and t2 represent the time of starting and ending charging respectively, and C n is the current actual capacity of the lithium-ion battery, and C0 represents the rated capacity of the battery;

[0060] Arrange the converted SOH data in chronological order to form an SOH time series;

[0061] Step S2: Data standardization:

[0062] The preprocessed data is standardized to effectively reduce the impact of changes in the input data distribution on the neural network and improve the accuracy of the model. The StandardScaler tool is used to standardize the data. For each attribute, all data are clustered around 0 with a standard deviation of 1, so that the variance of the standardized data set is 1 and the mean is 0:

[0063]

[0064] Where z is the standardized data value, x is the original SOH data value, μ is the mean, and σ is the standard deviation;

[0065] Step S3, wavelet denoising:

[0066] Perform wavelet transform on input data. The introduction of wavelet denoising is an important breakthrough in the field of signal processing. Its development combines the progress of mathematical theory, engineering needs and computing technology. Before wavelet denoising, denoising operations mainly relied on Fourier transform and linear filtering. However, these methods have obvious defects. Fourier transform is good at processing stationary signals, but this method cannot capture local time-frequency characteristics such as signal mutations and transient noise. The linear filtering method will cause the high-frequency details in the signal to be blurred while removing noise, resulting in information loss. This makes these methods unable to complete the filtering work normally when used as filters for denoising. The wavelet denoising method can capture the global characteristics and local details of the signal at the same time through basis functions of different scales. At the same time, the multi-resolution characteristics of wavelet denoising enable it to better analyze non-stationary signals when used as a filter for denoising. It uses the time-frequency localization ability of wavelet transform to distinguish the characteristics of signal and noise at different scales. Its advantages are:

[0067] Multi-scale analysis: Noise is usually distributed in high-frequency details, while signals have structural characteristics at different scales;

[0068] Nonlinear threshold: retains important coefficients and removes noise coefficients through threshold processing, which is better than traditional linear filtering methods;

[0069] Therefore, in the filter selection stage, the present invention selects the wavelet denoising method to perform the denoising work of the lithium-ion battery SOH; the key role of this step is to perform wavelet denoising on the standardized data to improve the data quality, thereby improving the prediction efficiency of subsequent model learning; wavelet denoising is a signal processing technology based on wavelet transform, which is widely used in the field of noise suppression of non-stationary signals; it is mainly divided into three steps: wavelet decomposition, threshold denoising and wavelet reconstruction; its core idea is to use the multi-resolution analysis characteristics of wavelet transform to decompose the signal into high-frequency details and low-frequency approximate components at different scales, through Threshold processing is used to separate noise from valid signals. Specifically, wavelet denoising is performed on the standardized SOH data. The wavelet transform replaces the basis functions in the Fourier transform with wavelet functions, converting an infinitely long sine function into a finite, decaying wavelet function. In this transform, the degree of stretching reflects the frequency content of the signal, with stretching corresponding to low-frequency components and compression corresponding to high-frequency components. By translating and scaling the signal, the transform identifies the frequency components at specific locations in the signal. This process generates a new wavelet coefficient array from the SOH data, which facilitates more accurate analysis. First, a wavelet transform is performed on the input data:

[0070]

[0071] Where w j,k is the wavelet coefficient, ψ j,k (t) is the wavelet basis function, j and k are scale and translation parameters respectively, and f(t) is the array for wavelet transform;

[0072] The selection of wavelet basis function should be analyzed from the perspective of orthogonality, high vanishing moment, compact support, symmetry and antisymmetry, and the wavelet basis function that meets all the conditions should be selected; but in actual application, it is impossible for a wavelet basis function to have all the above characteristics at the same time. Therefore, when applying wavelet transform, selecting an appropriate wavelet basis function is also a key factor in enhancing the wavelet denoising effect. When applying, a wavelet with compact support should be selected and a more appropriate wavelet basis function should be selected according to the characteristics of the signal.

[0073] The wavelet-transformed data undergoes multi-layer decomposition, separating the signal into superimposed components with different frequencies, divided into low-frequency (LF) and high-frequency (HF) signals. The low-frequency component reflects the overall trend, while the high-frequency component reflects fine-grained changes. In practice, low-frequency components are more important than high-frequency components. Therefore, high-pass and low-pass filters decompose the data into low-frequency and high-frequency components. The extracted LF undergoes further multi-level decomposition. At each layer, the LF and HF components are wavelet transformed separately to obtain the corresponding wavelet coefficients, thereby identifying the characteristics of the LF and HF at each decomposition stage. After decomposition, the wavelet coefficients are denoised. Important signal features correspond to larger wavelet coefficients, while noise is associated with smaller wavelet coefficients.

[0074] The threshold is determined according to the specific internal calculation rules of the wavelet. The coefficients exceeding this threshold are retained as meaningful data, while the coefficients below this threshold are identified as noise. The threshold is also an important factor that directly affects the denoising effect. Selecting different thresholds for denoising may result in completely different denoising effects. If the threshold is too small, the noise will remain, and if the threshold is too large, the signal characteristics will be lost. The principle and calculation method of the threshold are as follows, where the formula for the general threshold is:

[0075]

[0076] Where cD1 is the first layer wavelet coefficient, N is the data length;

[0077] This threshold selection method is simple and efficient, and is suitable for Gaussian white noise. However, it is too conservative and may lead to over-smoothing and loss of signal details. It is also suitable for situations where the signal length is long and the noise intensity is high.

[0078] Step S4, DLinear model prediction:

[0079] Put in the DLinear model for prediction and set different output lengths:

[0080] The LTSF-Linear model is a linear model developed and developed based on the Transformer model developed for the long time series prediction (LTSF) task; this model questions the effectiveness of the Transformer in this work; specifically, the Transformer model can be said to be the most successful model for extracting semantic relevance between elements in long sequences; however, in the process of modeling time series data, it is necessary to extract time-related data points from the time series data, and these data points must be ordered and continuous; although the Transformer uses orthogonal encoding to mark some of the subsequences, thereby retaining medical order information, Tran The self-attention mechanism in the Transformer does not change the arrangement of the sequence, so it will inevitably lead to the loss of some useful time information. The main working ability of the Transformer comes from its multi-head self-attention mechanism, which can extract the semantic correlation between elements in long time series. However, the self-attention mechanism in the Transformer is somewhat different and out of order. Although various types of positional encoding techniques can retain some arrangement information, they will inevitably lose information after self-attention is applied to them. The LTSF-Linear model is a simple linear model proposed based on the assumption that long-term prediction is only applicable to time series with relatively clear trends and periodicities.

[0081] In the current experiments on long-term prediction solutions based on Transformer, there is a common problem: the technologies used in these solutions all have the problem of error accumulation, and the baselines for these technologies are all IMS prediction technologies. By assuming that the DMS strategy leads to the improvement of the prediction accuracy of the models in these studies, a simple DMS model called Linear is proposed. The core of the Linear model is to directly regress the historical time series through weighted sum operations to achieve future predictions. The mathematical expression is Where W is the linear layer along the time axis; and X i is the prediction and input of each i variable; Linear is a set of linear models. In order to span time series in different fields, variant models DLinear and NLinear with two preprocessing methods are proposed;

[0082] Specifically, the sequence decomposition module in Autoformer is retained, and this internal module is transplanted into the FEDformer model, and a simple linear combination of the two is the DLinear model; like Autoformer, it gradually decomposes long time series data into trend data and seasonal data through the sequence decomposition module; then, the two types of data are processed and learned by a single layer of linear layer respectively, and finally the two types of data are combined to obtain the final prediction result; by performing linear layer processing on the trend of the data, the DLinear model can fit the data with strong trend well during prediction and improve its prediction performance compared with the Linear model; at the same time, when the Linear model processes distributed data sets, in order to improve its prediction accuracy and efficiency, it is improved and the NLinear model is proposed; it first discards the last value in the input sequence; then, the normal Linear model-related processing is performed, and after all processing steps are completed, the subtracted last value is added back as the final prediction result; the subtraction and addition in NLinear is actually a simple normalization of the input sequence.

[0083] A specific application of this embodiment is:

[0084] In this embodiment, 70% of the data from each cell is used for training, 10% for validation, and 20% for testing; then, the test set is put into the trained model to obtain a prediction output with a prediction length of 96, and the input sequence in the test set is compared with the real data with a past data length of 96, and the MAE and MSE are calculated; the subsequent related battery SOH data curve is predicted based on the historical data set; and a comparative experiment is conducted on the same series of autoformer, nlinear and linear models; first, more than 160,000 test sets are grouped into 96 data sets; then, each set of data is put into the trained model to predict the 96 data after the data set; finally, each set of predicted 96 data sets is combined to obtain a prediction line chart, reference Figure 2 , (a) and (b) are the ev6SOH prediction curves and errors, (c) and (d) are the EV 8SOH prediction curves and errors;

[0085]

[0086] Table 1 Prediction results

[0087] At the same time, by maintaining the input sequence length of 96 and gradually increasing the output prediction sequence, the feasibility of the model's long-term time series prediction is verified; the output sequence is set to 192 and 336, and then input into the model for prediction; the five electric vehicle data sets are grouped according to different prediction lengths, and then the lithium-ion battery SOH is predicted; with the increase of the prediction length, the degree of fitting of the model to the SOH curve does not decrease much, and it can also fit the real curve of the SOH data well; that is, when the prediction sequence increases, the maximum absolute error of the prediction result is smaller, but the mean absolute error increases; among them, for the average MSE, MAE and maximum absolute error, the prediction results of 96-192 are 0.37%, 1.77% and 1.90% respectively, and the prediction results of 96-336 are 0.29%, 1.95% and 1.62% respectively; when the prediction length is extended, the maximum increase in MSE is 0.90%, and the maximum increase in MAE is 1.0%; on the contrary, the maximum absolute error is reduced by at least 0.3%, for specific reference Figure 3 , (a) and (b) are the predictions and absolute errors of 96~192SOH, (c) and (d) are the predictions and absolute errors of 96~336SOH.

[0088]

[0089] Table 2 Prediction results of different time series

[0090] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0091] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A SOH prediction method based on a filter and a DLinear model, characterized in that: The prediction method comprises the following steps: Step S1, data preprocessing: preprocessing the initial battery data to obtain the SOH data corresponding to the current time point of the battery data; Step S2, data standardization: use StandardScaler to standardize the SOH time series data to eliminate dimensional differences; Step S3, wavelet denoising: putting the normalized data into wavelet denoising to remove noise; Step S4, DLinear model prediction: decompose the time series data into trend terms and seasonal terms, use linear layer fitting respectively, and dynamically input historical data through the sliding window mechanism to predict the SOH value of the next n cycles.

2. The SOH prediction method based on a filter and a DLinear model according to claim 1, characterized in that: The data preprocessing step S1 specifically includes the following steps: Step S11, raw data conversion: converting the raw charge and discharge data of the lithium-ion battery into SOH data. The specific formula is: Where Δt is the time interval, I is the current during charging, t1 and t2 represent the time of starting and ending charging respectively, and C n is the current actual capacity of the lithium-ion battery, and C0 represents the rated capacity of the battery; Step S12, constructing time series data: arranging the converted SOH data in chronological order to form an SOH time series sequence.

3. The SOH prediction method based on a filter and a DLinear model according to claim 1, characterized in that: In step S2, the normalization formula for data normalization is: Where z is the standardized data value, x is the original SOH data value, μ is the mean, and σ is the standard deviation.

4. The SOH prediction method based on a filter and a DLinear model according to claim 1, characterized in that: Said step S3, wavelet denoising specifically comprises the following steps: Step S31, wavelet basis function selection: selecting a wavelet basis function with compact support and high vanishing moment; Step S32, wavelet decomposition: perform multi-layer wavelet decomposition on the standardized SOH data to decompose the signal into low-frequency and high-frequency components. The decomposition formula is: Where w j,k is the wavelet coefficient, ψ j,k (t) is the wavelet basis function, j and k are scale and translation parameters respectively, and f(t) is the array for wavelet transform; Step S33, threshold denoising: Calculate the threshold and use the general threshold formula to determine the noise threshold T, which is: Where cD1 is the first layer wavelet coefficient, N is the data length; The wavelet coefficients above the threshold are retained, and the coefficients below the threshold are discarded; Step S34, wavelet reconstruction: performing inverse wavelet transform on the denoised low-frequency and high-frequency coefficients to reconstruct the denoised SOH time series data.

5. The SOH prediction method based on filter and DLinear model according to claim 1, characterized in that: The step S4, DLinear model prediction specifically includes the following steps: Step S41, sequence decomposition: decompose the denoised SOH time series data into trend terms and seasonal terms, and use the sequence decomposition module to separate long-term trends and short-term fluctuations; Step S42: Model construction: specifically: Trend item prediction: Apply linear layer fitting to the trend item, the formula is: Where, is the predicted output of the trend term, W trend is the weight matrix of the linear layer, x trend Enter data for the trend item obtained after series decomposition; Seasonal term prediction: Another linear layer is applied to fit the seasonal term, the formula is: Where, is the forecast output of seasonal terms, W seasonal is the weight matrix of the linear layer, x seasonal Enter data for the seasonal terms obtained after series decomposition; Result fusion: Add the forecast results of trend item and seasonal item to get the final forecast value: Where, is the final predicted SOH value; Step S43, sliding window mechanism: Set the sliding window length, input historical data in the window each time to predict the SOH value of the next n cycles. After the prediction is completed, the window slides backward by a fixed step length to cover the new data and repeat the prediction; Step S44, post-processing of the prediction results: averaging the multiple sliding window prediction results at the same time point to eliminate the random error of a single prediction, and converting the normalized prediction results back to the original dimension to obtain the final SOH prediction value.

Images