Power battery SOC prediction method and device and vehicle

Abstract

The invention provides a power battery SOC prediction method and device and a vehicle, and aims to improve the estimation accuracy of the SOC of a power battery. The power battery SOC prediction method comprises the following steps: acquiring vehicle state data uploaded by a vehicle in a driving process; extracting at least one vehicle state variable having strong correlation with the SOC of the power battery from the vehicle state data; decomposing the SOC time sequence of the power battery, and extracting a plurality of intrinsic mode function component time sequences of the SOC of the power battery; combining the plurality of intrinsic mode parameter component time sequences with the at least one vehicle state variable time sequence to obtain a characteristic variable time sequence set; and inputting the characteristic variable time sequence set into a pre-constructed Informer prediction model to obtain the power battery SOC of the vehicle at a future preset moment.

Description

Technical Field

[0001] The present application relates to the field of power battery estimation, specifically to a power battery SOC prediction method, device and vehicle. Background Art

[0002] Lithium iron phosphate batteries are widely used in new energy vehicles. Their open-circuit voltage (OCV) versus state of charge (SOC) curve exhibits a nonlinear characteristic, with a flat center and steep slopes at both ends. Within the median range (20% < SOC < 80%), the battery's open-circuit voltage varies minimally, allowing the battery to stably output voltage and energy. However, at the extreme SOC ranges (SOC < 10% and SOC > 90%), the sensitivity of OCV to SOC changes increases significantly, exacerbating voltage fluctuations. Traditional SOC estimation methods primarily include the ampere-hour integration method and the open-circuit voltage method. The ampere-hour integration method requires high current measurement accuracy and suffers from the problem of cumulative error amplification in the initial SOC estimate. The open-circuit voltage method is significantly affected by environmental and internal parameters such as temperature and battery aging. Particularly in the low SOC range (< 10%), the nonlinearity of the voltage gradient increases, making the estimation accuracy insufficient for application requirements. Summary of the Invention

[0003] The present invention provides a power battery SOC prediction method, device and vehicle to improve the estimation accuracy of the power battery SOC.

[0004] The technical solution of the present invention is:

[0005] In one aspect, the present application provides a method for predicting SOC of a power battery, comprising:

[0006] Obtain vehicle status data uploaded by the vehicle during driving;

[0007] Extracting at least one vehicle state variable that is highly correlated with the power battery SOC from the vehicle state data;

[0008] Decompose the power battery SOC time series and extract multiple intrinsic mode function component time series of the power battery SOC;

[0009] Merging a plurality of intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set;

[0010] The characteristic variable time series set is input into the pre-built Informer prediction model to obtain the vehicle's power battery SOC at a preset time in the future.

[0011] Preferably, the step of extracting at least one vehicle state variable having a strong correlation with the power battery SOC from the vehicle state data includes:

[0012] The Spearman correlation coefficient is used as a feature selection tool to extract at least one vehicle state variable having a high Spearman correlation coefficient with the power battery SOC from the vehicle state data.

[0013] Preferably, before the step of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC, the method further includes:

[0014] Smoothing and denoising are performed on the power battery SOC time series and at least one vehicle state variable time series.

[0015] Preferably, the step of performing smoothing and noise reduction processing on the power battery SOC time series and at least one vehicle state variable time series includes:

[0016] For each time series, a sliding window of fixed length is constructed starting from the starting data point of the time series;

[0017] In the constructed sliding window, the least square method is used to perform polynomial fitting on all data points in the sliding window;

[0018] Estimate the value of each data point in the sliding window based on the fitted polynomial;

[0019] Slide the sliding window along the time series, moving one data point each time to construct a new sliding window;

[0020] In the new sliding window, the above polynomial fitting and data point value estimation operations are repeated until the entire time series is processed.

[0021] Preferably, the step of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC includes:

[0022] Through VMD variational mode decomposition technology, the power battery SOC time series is decomposed into multiple intrinsic mode function component time series with different frequencies.

[0023] Preferably, the step of decomposing the power battery SOC time series into a plurality of intrinsic mode function component time series of different frequencies by using VMD variational mode decomposition technology includes:

[0024] Perform VMD parameter setting; the VMD parameters include the number of intrinsic mode function components expected to be extracted;

[0025] The VMD algorithm is applied to optimize the power battery SOC time series, and the power battery SOC time series is decomposed into multiple original intrinsic mode function component time series;

[0026] Applying Hilbert transform to each original intrinsic mode function component time series to obtain an analytical signal of each intrinsic mode function component time series;

[0027] Extracting spectrum information from the analytical signal of each intrinsic mode function component time series;

[0028] Perform spectrum demodulation on the analytical signal of each intrinsic mode function component time series and demodulate its spectrum to the baseband;

[0029] The bandwidth of the analytical signal of each intrinsic mode function component time series is estimated by Gaussian smoothing method;

[0030] Using a variational constraint model, the bandwidth of the analytical signal of the time series of the intrinsic mode function components is minimized so that the sum of the bandwidths of the analytical signals of all the time series of the intrinsic mode function components is equal to the sum of the bandwidths of all the time series of the intrinsic mode function components;

[0031] Introducing a quadratic penalty factor and a Lagrangian multiplier into the variational constraint model to obtain an augmented Lagrangian function;

[0032] The alternating direction multiplier method is used to solve the augmented Lagrangian function to find the optimal time series and center frequency of the intrinsic mode function components.

[0033] Preferably, the step of inputting the characteristic variable time series set into a pre-built Informer prediction model to obtain the power battery SOC of the vehicle at a preset time in the future includes:

[0034] Map the feature vectors of each time step into a dense vector space of fixed dimension to obtain the embedded feature vectors of each time step;

[0035] The embedded feature vector of each time step is input into the encoder of the Informer model to extract key features;

[0036] The output of the encoder is input into the decoder of the Informer model to generate the SOC prediction value at the preset time.

[0037] Preferably, the feature vector after embedding each time step is input into the encoder of the Informer model, and the step of extracting key features includes:

[0038] Construct query, key, and value matrices based on the embedded feature vectors at each time step;

[0039] Using the multi-head probsparse self-attention mechanism, we calculate the sparsity evaluation of each query and select the most important top-order queries;

[0040] Calculate attention scores for the selected top predetermined queries and all keys;

[0041] Use the calculated attention scores to perform weighted summation on the value matrix to obtain the preliminary output feature map;

[0042] The self-attention distillation mechanism is used to extract the main features of the preliminary output feature map to obtain the optimized output feature map of the encoder.

[0043] On the other hand, the present application also provides a power battery SOC prediction method, comprising:

[0044] Obtaining a first power battery SOC prediction value sent by the cloud, where the first power battery SOC prediction value is obtained by the cloud according to the above-mentioned power battery SOC prediction method;

[0045] Obtain the SOC prediction value of the second power battery estimated by the vehicle itself;

[0046] A final power battery SOC prediction value is obtained according to the first power battery SOC prediction value and the second power battery SOC prediction value.

[0047] On the other hand, the present application also provides a power battery SOC prediction device, comprising:

[0048] The acquisition module is used to obtain the vehicle status data uploaded by the vehicle during driving;

[0049] a vehicle state variable extraction module, configured to extract at least one vehicle state variable having a strong correlation with the power battery SOC from the vehicle state data;

[0050] A decomposition module is used to decompose the power battery SOC time series and extract multiple intrinsic mode function component time series of the power battery SOC;

[0051] a merging module, configured to merge a plurality of intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set;

[0052] The prediction module is used to input the characteristic variable time series set into the pre-built Informer prediction model to obtain the vehicle's power battery SOC at a preset time in the future.

[0053] On the other hand, the present application also provides a vehicle, comprising the above-mentioned power battery SOC prediction device.

[0054] The beneficial effects of the present invention are:

[0055] By building a data-driven Informer prediction model based on the vehicle operating status data uploaded to the cloud in real time, we can fully exploit the status data of the vehicle during driving, effectively compensate for the estimation deviation of traditional evaluation methods, and improve the SOC prediction accuracy and robustness across the entire operating range. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the flow of the power battery SOC prediction method in an embodiment of the present application;

[0057] Figure 2 This is a structural block diagram of the power battery SOC prediction device in an embodiment of the present application. DETAILED DESCRIPTION

[0058] The present invention is further described below with reference to the following embodiments and accompanying drawings. This embodiment is based on the technical solution of the present invention and provides a detailed implementation method and specific operation process, but the scope of protection of the present invention is not limited to the following embodiments.

[0059] Reference Figure 1 The present application provides a method for predicting the SOC (state of charge) of a power battery. The method is applied in the cloud and includes:

[0060] S101, obtaining vehicle status data uploaded by the vehicle during driving;

[0061] S102, extracting at least one vehicle state variable that is highly correlated with the power battery SOC from the vehicle state data;

[0062] S103, decomposing the power battery SOC time series to extract multiple intrinsic mode function component time series of the power battery SOC;

[0063] S104, merging a plurality of intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set;

[0064] S105 , inputting the characteristic variable time series set into a pre-built Informer prediction model to obtain the power battery SOC of the vehicle at a preset time in the future.

[0065] According to relevant national standards (such as GB / T 32960, "Technical Specifications for Remote Service and Management Systems for Electric Vehicles"), real-time vehicle status data during driving must be uploaded to a cloud platform. This real-time vehicle status data includes, for example, battery SOC, total current, total voltage, vehicle speed, motor torque, motor speed, motor temperature, accumulated mileage, and insulation resistance. In this embodiment of the present application, this real-time vehicle status data uploaded to the cloud platform can be used to predict the vehicle's SOC.

[0066] Considering that there may be some missing data in the real-time uploaded data of the vehicle, it is necessary to pre-process these missing data before using them for SOC prediction. In the embodiment of the present application, in order to alleviate the impact of missing values ​​on the experimental results, for example, linear interpolation is used to fill in the missing data in the vehicle status data set. Its essence is to estimate the distance between the two adjacent values ​​on the left and right of the interpolation point by allocating the weight according to the distance between them. The formula is as follows:

[0067]

[0068] Y represents the estimated value of the interpolation point, that is, the missing value you want to calculate;

[0069] y0 represents the value of the known data point to the left of the interpolation point;

[0070] y1 represents the value of the known data point to the right of the interpolation point;

[0071] x represents the position of the interpolation point;

[0072] x0 represents the position of the known data point to the left of the interpolation point;

[0073] x1 represents the location of the known data point to the right of the interpolation point;

[0074] The basic principle of linear interpolation is to assume that the data change between two known data points is linear, that is, the rate of change is constant. Based on this assumption, the value y of any point x between the two known points (x0, y0) and (x1, y1) can be estimated.

[0075] The x1-x0 and y1-y0 parts of the formula calculate the slope between the two known points, which represents the rate at which the y value changes with the x value. This slope and the values ​​of the known points are used to estimate the y value at the x position.

[0076] Through the above-mentioned linear interpolation method, it is possible to complete missing data of the vehicle status data set uploaded by the vehicle in real time.

[0077] Since some data in the vehicle state data have little correlation with the power battery SOC, it is necessary to find some variables with high correlation with the power battery from all the vehicle state data and eliminate variables with low correlation. In this embodiment of the application, step S102 of extracting at least one vehicle state variable with high correlation with the power battery SOC from the vehicle state data includes:

[0078] S1021 , using the Spearman correlation coefficient as a feature selection tool, extracting at least one vehicle state variable having a high Spearman correlation coefficient with the power battery SOC from the vehicle state data.

[0079] Among them, the Spearman correlation coefficient p is an indicator to measure the dependence relationship between two variables. Its correlation can be expressed by a monotonic function. The formula is as follows:

[0080]

[0081] n is the number of data points;

[0082] d i is the ranking difference of the i-th pair of data points;

[0083] is the sum of squares of all ranking differences;

[0084] The standard Spearman correlation coefficient p is calculated by the following steps:

[0085] 1. Assign a rank to each data point of each variable (or the average rank if there are identical values).

[0086] 2. Calculate the difference d between the rankings i ;

[0087] 3. Calculate the difference d i The sum of squares

[0088] For each pair of observations (xi,yi), calculate their rank difference; if xi’s rank is Rx and yi’s rank is Ry, then di = Rx - Ry. Represents the sum of squares of all ranking differences, i.e. ∑(Rx-Ry) 2 .

[0089] In the embodiment of the present application, the Spearman correlation coefficient ranges from -1 to 1. If ρ = 1, it indicates a perfect positive monotonic relationship; if ρ = -1, it indicates a perfect negative monotonic relationship; if ρ = 0, it indicates no monotonic relationship.

[0090] By calculating the Spearman correlation coefficient between each variable in the vehicle state signal and the power battery SOC, vehicle state variables with a strong correlation with SOC can be screened out. For example, battery voltage and current typically have a strong monotonic relationship with the power battery SOC, while certain vehicle state variables (such as accumulated mileage and insulation resistance) may have a weaker relationship with the power battery SOC. By screening, vehicle state variables with a strong correlation with the power battery SOC are retained, while those with weak or irrelevant correlations are removed, thereby improving the input quality of the model.

[0091] In the embodiment of the present application, combined with the Spearman correlation coefficient p, the total current, total voltage, vehicle speed, motor speed, and motor temperature are finally selected as vehicle state variables that have a strong correlation with the power battery SOC.

[0092] Typically, the vehicle status signal collected in real time by the vehicle contains a certain degree of noise. Therefore, the vehicle status signal needs to be processed to reduce the noise in the vehicle status data before subsequent data processing. In the embodiment of the present application, before step S013 of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC, the method further includes:

[0093] S106: Perform smoothing and noise reduction processing on the power battery SOC time series and at least one vehicle state variable time series.

[0094] In the embodiment of the present application, for example, SG (Savitzky-Golay) filtering is used to reduce the noise in the sequence data. SG filtering can reduce the noise and fluctuation of the signal while ensuring that the signal trend remains unchanged, so that the potential trend of the signal can be more fully explored, thereby improving the accuracy of model prediction. As shown in the following fitting polynomial, SG filtering uses the least squares fitting principle and sliding window mechanism to calculate the vehicle data signal with x as the starting point. i As the center, select 2M+1 continuous collection points to build a sliding window for fitting operation, where n k is the collection point in the sliding window, a k is n in each window k The convolution coefficient of .

[0095]

[0096] M represents half of the sliding window, and N is the order of the polynomial; this polynomial expression indicates that within the sliding window, the fitting polynomial p(n) is obtained by adding the convolution coefficient a k It is obtained by multiplying and summing the k-th power of the position n in the sliding window.

[0097] The residuals from the least squares fit are given by the following formula:

[0098]

[0099] ε N represents the residual sum of squares, that is, the sum of squares of the differences between the fitted polynomial and the original data. By minimizing the residual sum of squares ε N , we can find the coefficients a of the best fitting polynomial k .

[0100] p(n) represents the value of the fitted polynomial at position n in the sliding window.

[0101] x[n] represents the value of the original data at position n.

[0102] M represents half of the sliding window size, and the sliding window contains a total of 2M+1 data points.

[0103] N represents the order of the fitting polynomial.

[0104] a k Indicates n in the fitting polynomial k The coefficient of .

[0105] n represents the position index in the sliding window, ranging from -M to M.

[0106] In the embodiment of the present application, step S106 of performing smoothing and noise reduction processing on the power battery SOC time series and at least one vehicle state variable time series specifically includes:

[0107] S1061: For each time series (including the power battery SOC time series and the vehicle state variable time series), a sliding window of fixed length is constructed starting from the starting data point of the time series. The length of this sliding window is pre-set and is used to select a group of continuous data points in the time series for subsequent processing.

[0108] S1062: Within the constructed sliding window, a polynomial is fitted to all data points within the sliding window using the least squares method. The least squares method is a commonly used mathematical technique that finds the best function matching the data by minimizing the sum of squared errors. Here, it is used to fit the data points within the sliding window to generate a polynomial function that can effectively describe the changing trend of the data within the window.

[0109] S1063, based on the fitted polynomial, estimate the value of each data point in the sliding window; calculate the estimated value of each data point through the polynomial function, and use these estimated values ​​to replace the value of the original data point, thereby achieving data smoothing and reducing noise interference.

[0110] S1064, slide the sliding window along the time series, moving one data point each time, and constructing a new sliding window; this means that the sliding window will gradually cover the next data point in the time series, while discarding the earliest data point, thereby gradually moving on the time series.

[0111] S1065: Repeat the above polynomial fitting and data point value estimation operations within the new sliding window until the entire time series has been processed. That is, the polynomial fitting is performed again using the least squares method within the new sliding window, and the value of each data point is estimated based on the fitting results. This process continues until the entire time series has been processed, that is, the sliding window covers all data points in the time series.

[0112] Through the above steps S1061-S1065, the power battery SOC time series and the vehicle state variable time series can be effectively smoothed and denoised, thereby improving the quality and availability of the data.

[0113] In the embodiment of the present application, the vehicle status data obtained in S102 is processed using a first-order polynomial and an SG filter with a seven-step filter window.

[0114] In the embodiment of the present application, step S103 of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC includes:

[0115] S1031 , using VMD (Variational Mode Decomposition) variational mode decomposition technology, decompose the power battery SOC time series into multiple intrinsic mode function component time series of different frequencies.

[0116] Using VMD variational mode decomposition technology, the SOC signal is decomposed into three IMF components of different frequencies {IMF1, IMF2, and IMF3}, effectively enriching vehicle data features while mitigating the nonlinearity and non-stationarity present in the sequence data. The core of the VMD algorithm lies in constructing and solving the variational problem, searching for the optimal solution through continuous iterative transformations, and then decomposing the non-stationary signal into a series of intrinsic mode functions (IMFs). Its principle is as follows:

[0117] In the embodiment of the present application, the SOC signal is decomposed into a series of IMF components, which are defined as follows:

[0118] u k (t) = A k (t)cos(φ k (t))

[0119] Among them, u k(t) represents the kth intrinsic mode function component, which is a component of the SOC signal. This component is a function of time t and represents the change of the signal over time.

[0120] A k (t) represents the envelope function, which describes the change of the amplitude of the signal uk(t) over time;

[0121] φ k (t) represents the phase of the signal, which describes the change of the frequency and phase of the signal uk(t) over time.

[0122] The specific implementation steps of the variational mode decomposition method are as follows:

[0123] The analytical signals of a series of intrinsic mode functions (IMFs) are obtained by Hilbert transform, and then the unilateral P-value is obtained and compared with Multiplication demodulates the spectrum to the baseband, as shown in the following equation:

[0124]

[0125] Where δ(t) represents the Dirac function, which is used to represent the impulse at t = 0 and plays a key role in the Hilbert transform;

[0126] j represents the imaginary unit, which is expressed as * represents the convolution operation; ω k represents the angular frequency of the kth IMF component. Represents the kernel function of the Hilbert transform, used to generate the analytical signal. Used to demodulate the spectrum to baseband.

[0127] The Hilbert transform is a signal processing technique used to generate an analytical representation of a signal; the analytical signal is the sum of the real signal and its Hilbert transform result, which contains the amplitude and phase information of the signal. A unilateral spectrum means that the spectrum of the signal contains only positive frequencies, which is a characteristic of the analytical signal obtained by the Hilbert transform. Multiplication can demodulate the signal's spectrum back to the baseband.

[0128] The bandwidth of each modal signal is estimated by Gaussian smoothing, and the total bandwidth is minimized. The variational constraint model is expressed as follows:

[0129]

[0130] in is the partial derivative about t, δ(t) represents the impulse function, u k are the k components obtained after decomposition, ω kis the center frequency of each mode, f represents the original signal, and K represents the total number of eigenmode function components.

[0131] By introducing the quadratic penalty factor α and the Lagrangian multiplication operator λ(t), the variational constraint model can be transformed into an unconstrained variational model. The following formula represents the augmented Lagrangian function:

[0132]

[0133] In the iterative search process, the ADMM alternating direction multiplier algorithm is used to calculate the component signals and the center frequency to obtain the optimal solution of the constrained variational model. The following formulas represent the update formulas of the center frequency and the modal components respectively. Iterative formula in the field:

[0134]

[0135] Based on the above principles, in the embodiment of the present application, step S1031 of decomposing the power battery SOC time series into multiple intrinsic mode function component time series of different frequencies by using VMD variational mode decomposition technology includes:

[0136] S10311, set the VMD parameters. One of the most important VMD parameters is the number of intrinsic mode function components (K) that you want to extract. This parameter determines the complexity of the decomposed signal and the accuracy of the decomposition. For example, if K = 3, the VMD algorithm will decompose the SOC time series into three intrinsic mode function components. The selection of an appropriate K value needs to be determined based on the actual application requirements and signal characteristics. If the K value is set too large, the decomposed components may be too complex and difficult to interpret; if the K value is set too small, some important signal features may be missed.

[0137] S10312, apply the VMD algorithm to optimize the power battery SOC time series, and decompose the power battery SOC time series into multiple original intrinsic mode function component time series. The core idea of ​​the VMD algorithm is to decompose the signal by optimizing a variational model. Specifically, VMD decomposes the signal into multiple modal components, and the bandwidth of each component is constrained to be minimized. At the same time, the sum of the bandwidths of all components is equal to the bandwidth of the original signal. This optimization problem can be solved by the alternating direction multiplier method (ADMM). When applying the VMD algorithm, the algorithm decomposes the SOC time series according to the set parameters (such as the K value) to generate multiple original intrinsic mode function component time series. These components can better reflect the inherent structure and change law of the original signal.

[0138] S10313, applying a Hilbert transform to each original intrinsic mode function component time series to obtain an analytical signal for each eigenmode function component time series. The Hilbert transform is a commonly used signal processing method that can convert a real signal into a complex analytical signal; the real part of the analytical signal is the original signal, and the imaginary part is the Hilbert transform of the original signal. The Hilbert transform can be used to extract the instantaneous frequency and amplitude information of the signal. For each original intrinsic mode function component time series, applying the Hilbert transform can obtain its corresponding analytical signal. The amplitude and phase information of the analytical signal can be used to further analyze the characteristics of the signal.

[0139] S10314, extracting spectral information from the analytical signal of each intrinsic mode function component time series. The spectral information of the analytical signal can be obtained through Fourier transform. Spectral information reflects the energy distribution of the signal at different frequencies. Extracting the spectral information of the analytical signal of each intrinsic mode function component can better understand the frequency characteristics of each component; for example, some components may be primarily concentrated in the low-frequency region, while others may contain high-frequency components. These frequency characteristics are very important for subsequent signal analysis and processing.

[0140] S10315 , performing spectrum demodulation on the analytical signal of each intrinsic mode function component time series, and demodulating the spectrum to a baseband.

[0141] Spectral demodulation is a signal processing technique that aims to shift a signal's spectrum to the baseband. In the analytical signal of an intrinsic mode function component, the spectra may be distributed across different frequency ranges. Spectral demodulation can shift these spectra to the baseband, facilitating subsequent analysis and processing. For example, if a component has a center frequency of fc, its spectrum can be shifted to the baseband through downconversion. The demodulated signal more intuitively reflects the signal's instantaneous frequency and amplitude changes.

[0142] S10316, performing bandwidth estimation on the analytical signal of each intrinsic mode function component time series by using a Gaussian smoothing method.

[0143] Bandwidth is a key signal characteristic, reflecting its width in the frequency domain. For each intrinsic mode function component of the analytical signal, the spectrum can be smoothed using Gaussian smoothing to more accurately estimate the signal's bandwidth. Gaussian smoothing is a commonly used smoothing technique that uses a Gaussian function to perform a weighted average of the spectrum, thereby reducing spectral noise. Estimating bandwidth provides a better understanding of the frequency range and signal characteristics of each component.

[0144] S10317, using a variational constraint model, minimize the bandwidth of the analytical signal of the time series of the intrinsic mode function components so that the sum of the bandwidths of the analytical signals of all the time series of the intrinsic mode function components is equal to the sum of the bandwidths of all the time series of the intrinsic mode function components.

[0145] In the VMD algorithm, the variational constraint model is a key optimization objective. This model aims to minimize the bandwidth of the analytical signal for each intrinsic mode function component, while ensuring that the sum of the bandwidths of all components equals the bandwidth of the original signal. This ensures that the decomposed components have a well-defined frequency range and can fully reconstruct the original signal. The variational constraint model optimization process can be solved using the alternating direction method of multipliers (ADMM), resulting in the optimal decomposition result.

[0146] S10318, introducing a quadratic penalty factor and a Lagrangian multiplier into the variational constraint model to obtain an augmented Lagrangian function.

[0147] In the optimization process, quadratic penalty factors and Lagrange multipliers are commonly used tools. The quadratic penalty factor is used to constrain the feasible solution of the optimization problem and ensure the stability and convergence of the optimization process. Lagrange multipliers are used to introduce constraints and transform the constrained optimization problem into an unconstrained optimization problem. By introducing the quadratic penalty factor and Lagrange multipliers, the augmented Lagrangian function can be obtained. The augmented Lagrangian function is an important form of optimization problem. It combines the objective function and constraints to facilitate the solution. S10319 uses the alternating direction multiplier method to solve the augmented Lagrangian function to find the optimal intrinsic mode function component time series and center frequency.

[0148] The alternating direction method of multipliers (ADMM) is an efficient optimization algorithm particularly suitable for solving large-scale optimization problems. ADMM gradually approaches the optimal solution by alternating between updating variables and Lagrangian multipliers. In this application, ADMM is used to solve the augmented Lagrangian function to find the optimal intrinsic mode function component time series and center frequency. Using the ADMM algorithm, the VMD model optimization problem can be efficiently solved to obtain the decomposed intrinsic mode function components.

[0149] Through the above steps, the power battery SOC time series can be effectively decomposed into multiple intrinsic mode function component time series with different frequency characteristics, thereby providing a more accurate basis for subsequent analysis and processing.

[0150] In an embodiment of the present application, the SOC signal is decomposed into three IMF components by applying the VMD variational mode decomposition algorithm, and the remaining variables are integrated as the final model input, which reduces the volatility of the time series, avoids the influence of mode mixing, and effectively improves the prediction effect of the model.

[0151] In the embodiment of the present application, step S105 of inputting the characteristic variable time series set into a pre-built informer prediction model to obtain the power battery SOC of the vehicle at a preset time in the future includes:

[0152] S1051, mapping the feature vectors of each time step into a dense vector space of fixed dimension to obtain the embedded feature vectors of each time step;

[0153] S1052: Input the embedded feature vector of each time step into the encoder of the Informer model to extract key features;

[0154] S1053: Input the output of the encoder into the decoder of the Informer model to generate the SOC prediction value at the preset time.

[0155] In this embodiment of the present application, the feature vector after embedding each time step is input into the encoder of the Informer model, and the step S1052 of extracting key features includes:

[0156] S10521, construct query, key and value matrices based on the embedded feature vectors at each time step;

[0157] S10522, using the multi-head probsparse self-attention mechanism, calculates the sparsity evaluation of each query and selects the most important top-order queries;

[0158] S10523, calculating attention scores for the selected first predetermined queries and all keys;

[0159] S10524, using the calculated attention score to perform weighted summation on the value matrix to obtain a preliminary output feature map;

[0160] S10525, using the self-attention distillation mechanism, extract the main features of the preliminary output feature map to obtain the optimized output feature map of the encoder.

[0161] The traditional attention mechanism consists of Query, Key, and Value, and the expression is as follows:

[0162]

[0163] Where d is the input dimension, Q, K, and V are query, key, and value matrices, respectively. LQ is the length of the query, and LK is the length of the key. Based on the above formula, the probsparse self-attention mechanism is introduced to filter important elements in the query Q for calculating the attention value.

[0164] The probsparse self-attention mechanism only needs to focus on the main query and key combination, which significantly reduces the amount of computation. The probability formula for the i-th query is as follows:

[0165]

[0166] Among them, q i , k i , v i Represents the i-th row of Q, K, and V respectively.

[0167] In the self-attention mechanism, the probability distribution has potential sparse rows. Therefore, the similarity and importance between the query and the key can be measured using the Kullback-Leibler Divergence divergence. The sparsity evaluation formula for the i-th query is as follows:

[0168]

[0169] The first item is the query q i and all keys k j The exponential sum of the dot products of , representing the query q i With all keys k j similarity.

[0170] The second item is the query q i With all keys k j The average similarity of

[0171] This formula is used to evaluate the query q i The sparsity of query q i With key k j The uniformity of the similarity distribution.

[0172] In the embodiment of the present application, the encoder part uses the multi-head probsparse self-attention mechanism to reduce the complexity of matrix calculation, and introduces the self-attention distillation mechanism to extract the main features of the time series, as shown in the following formula:

[0173]

[0174] Where Max pool is the maximum pooling operation, ELU is the activation function, and Convld represents a one-dimensional convolution that reduces the input length to half of the initial length.

[0175] The decoder part contains two identical multi-head self-attention layers, which are used to map the features extracted by the encoder back to the output space and obtain the corresponding prediction value, as shown in the following formula:

[0176] X de =Concar(X token ,X0)

[0177] After the processed vehicle state data is fed into the Informer model, it is processed using a unified input embedding method. The processed sequence is then fed back into the encoder. The encoder's multi-head sparse attention module calculates the attention of a small number of clicks with significant contributions, and then uses distillation to trim the input dimensions. In the decoder module, the input data is processed by the masked multi-head attention mechanism and, together with the encoder output, serves as the input for the next multi-head attention layer. The decoder then processes the data and generates the final prediction through a fully connected layer.

[0178] The GA genetic algorithm was used to optimize the vehicle's power battery SOC using the aforementioned Informer prediction model. The optimal model parameters were ultimately obtained, and MAE, RMSE, and MAPE were used as evaluation metrics. Experiments have shown that the Informer prediction model can be applied to evaluate the SOC of new energy vehicle batteries, achieving accurate battery SOC predictions and effectively alleviating vehicle owners' concerns about their vehicle's range.

[0179] Because VMD technology can extract key frequency components reflecting the battery's state from raw measurement data such as battery voltage and current, providing more accurate features for subsequent SOC predictions, the Informer prediction model uses the IMF components obtained from VMD as input and predicts the battery's future SOC value by learning the long-term dependencies between these components. Because the Informer prediction model can effectively process long sequences of data, it can capture both long-term trends and short-term fluctuations in battery SOC changes, thereby fully exploiting the vehicle's state data during driving, effectively compensating for the estimation bias of traditional evaluation methods, and improving the accuracy and robustness of SOC predictions across the entire operating range.

[0180] The present application also provides a method for predicting the SOC of a power battery applied to a vehicle, including:

[0181] Obtaining a first power battery SOC prediction value sent by the cloud, where the first power battery SOC prediction value is obtained by the cloud according to the power battery SOC prediction method in the above embodiment;

[0182] Obtain the SOC prediction value of the second power battery estimated by the vehicle itself;

[0183] A final power battery SOC prediction value is obtained according to the first power battery SOC prediction value and the second power battery SOC prediction value.

[0184] For vehicles, the second power battery SOC prediction value is determined based on conventional calculation logic. Because the SOC derived from this logic presents issues discussed in the background art, the embodiments of this application improve the vehicle's SOC prediction accuracy by comprehensively evaluating the first and second power battery SOC prediction values. In specific applications, the first and second power battery SOC prediction values ​​can be weighted to produce a final power battery SOC prediction value, improving the accuracy and robustness of SOC prediction across the entire operating range.

[0185] Reference Figure 2 , the embodiment of the present application further provides a power battery SOC prediction device, comprising:

[0186] The acquisition module 201 is used to acquire vehicle status data uploaded by the vehicle during driving;

[0187] A vehicle state variable extraction module 202 is configured to extract at least one vehicle state variable that is highly correlated with the power battery SOC from the vehicle state data;

[0188] A decomposition module 203 is used to decompose the power battery SOC time series and extract multiple intrinsic mode function component time series of the power battery SOC;

[0189] A merging module 204 is configured to merge the multiple intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set;

[0190] The prediction module 205 is used to input the characteristic variable time series set into a pre-built Informer prediction model to obtain the power battery SOC of the vehicle at a preset time in the future.

[0191] An embodiment of the present application also provides a vehicle, comprising the above-mentioned power battery SOC prediction device.

[0192] The vehicle may be, but is not limited to, a pure electric vehicle (Pure Electric Vehicle / Battery Electric Vehicle, PEV / BEV), a hybrid electric vehicle (Hybrid Electric Vehicle, HEV), a range extended electric vehicle (REEV), a plug-in hybrid electric vehicle (PHEV), a new energy vehicle (New Energy Vehicle), a fuel vehicle, etc.

[0193] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referenced to each other.

[0194] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0195] It should also be noted that, in this document, the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are for the purpose of facilitating the description of the present invention and simplifying the description, rather than indicating or implying that the devices or components referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention. In addition, relational terms such as "first" and "second" are used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any actual relationship or order between these entities or operations, nor should they be understood as indicating or implying relative importance. Moreover, the terms "comprises", "includes" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements does not include those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or terminal device comprising the element.

[0196] The technical solutions provided by the present invention are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only intended to help understand the present invention, and the contents of this specification should not be construed as limiting the present invention. At the same time, for those skilled in the art, according to the present invention, there may be various changes in the specific implementation methods and application scopes. It is not necessary and impossible to enumerate all implementation methods here, and obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A power battery SOC prediction method, characterized in that: include: Obtain vehicle status data uploaded by the vehicle during driving; Extracting at least one vehicle state variable that is highly correlated with the power battery SOC from the vehicle state data; Decompose the power battery SOC time series and extract multiple intrinsic mode function component time series of the power battery SOC; Merging a plurality of intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set; The characteristic variable time series set is input into the pre-built Informer prediction model to obtain the vehicle's power battery SOC at a preset time in the future.

2. The power battery SOC prediction method according to claim 1, characterized in that: Before the step of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC, the method further includes: Smoothing and denoising are performed on the power battery SOC time series and at least one vehicle state variable time series.

3. The power battery SOC prediction method according to claim 2, characterized in that: The step of performing smoothing and noise reduction processing on the power battery SOC time series and at least one vehicle state variable time series includes: For each time series, a sliding window of fixed length is constructed starting from the starting data point of the time series; In the constructed sliding window, the least square method is used to perform polynomial fitting on all data points in the sliding window; Estimate the value of each data point in the sliding window based on the fitted polynomial; Slide the sliding window along the time series, moving one data point each time to construct a new sliding window; In the new sliding window, the above polynomial fitting and data point value estimation operations are repeated until the entire time series is processed.

4. The power battery SOC prediction method according to claim 1, characterized in that: The steps of decomposing the power battery SOC time series and extracting multiple intrinsic mode function component time series of the power battery SOC include: Through VMD variational mode decomposition technology, the power battery SOC time series is decomposed into multiple intrinsic mode function component time series with different frequencies.

5. The power battery SOC prediction method according to claim 4, characterized in that: The steps of decomposing the power battery SOC time series into multiple intrinsic mode function component time series of different frequencies using VMD variational mode decomposition technology include: Perform VMD parameter setting; the VMD parameters include the number of intrinsic mode function components expected to be extracted; The VMD algorithm is applied to optimize the power battery SOC time series, and the power battery SOC time series is decomposed into multiple original intrinsic mode function component time series; Applying Hilbert transform to each original intrinsic mode function component time series to obtain an analytical signal of each intrinsic mode function component time series; Extracting spectrum information from the analytical signal of each intrinsic mode function component time series; Perform spectrum demodulation on the analytical signal of each intrinsic mode function component time series and demodulate its spectrum to the baseband; The bandwidth of the analytical signal of each intrinsic mode function component time series is estimated by Gaussian smoothing method; Using a variational constraint model, the bandwidth of the analytical signal of the time series of the intrinsic mode function components is minimized so that the sum of the bandwidths of the analytical signals of all the time series of the intrinsic mode function components is equal to the sum of the bandwidths of all the time series of the intrinsic mode function components; Introducing a quadratic penalty factor and a Lagrangian multiplier into the variational constraint model to obtain an augmented Lagrangian function; The alternating direction multiplier method is used to solve the augmented Lagrangian function to find the optimal time series and center frequency of the intrinsic mode function components.

6. The power battery SOC prediction method according to claim 1, characterized in that: The steps of inputting the characteristic variable time series set into a pre-built Informer prediction model to obtain the vehicle's power battery SOC at a preset time in the future include: Map the feature vectors of each time step into a dense vector space of fixed dimension to obtain the embedded feature vectors of each time step; The embedded feature vector of each time step is input into the encoder of the Informer model to extract key features; The output of the encoder is input into the decoder of the Informer model to generate the SOC prediction value at the preset time.

7. The power battery SOC prediction method according to claim 6, characterized in that: The embedded feature vector of each time step is input into the encoder of the Informer model. The steps to extract key features include: Construct query, key, and value matrices based on the embedded feature vectors at each time step; Using the multi-head probsparse self-attention mechanism, we calculate the sparsity evaluation of each query and select the most important top-order queries; Calculate attention scores for the selected top predetermined queries and all keys; Use the calculated attention scores to perform weighted summation on the value matrix to obtain the preliminary output feature map; The self-attention distillation mechanism is used to extract the main features of the preliminary output feature map to obtain the optimized output feature map of the encoder.

8. A power battery SOC prediction method, characterized in that: include: Obtaining a first power battery SOC prediction value issued by the cloud, where the first power battery SOC prediction value is obtained by the cloud using the power battery SOC prediction method according to any one of claims 1 to 7; Obtain the SOC prediction value of the second power battery estimated by the vehicle itself; A final power battery SOC prediction value is obtained according to the first power battery SOC prediction value and the second power battery SOC prediction value.

9. A power battery SOC prediction device, characterized in that: include: The acquisition module is used to obtain the vehicle status data uploaded by the vehicle during driving; a vehicle state variable extraction module, configured to extract at least one vehicle state variable having a strong correlation with the power battery SOC from the vehicle state data; A decomposition module is used to decompose the power battery SOC time series and extract multiple intrinsic mode function component time series of the power battery SOC; a merging module, configured to merge a plurality of intrinsic modal parameter component time series with at least one vehicle state variable time series to obtain a characteristic variable time series set; The prediction module is used to input the characteristic variable time series set into the pre-built Informer prediction model to obtain the vehicle's power battery SOC at a preset time in the future.

10. A vehicle, characterized in that: Including the power battery SOC prediction device of claim 9.

Images