Lithium ion battery SOC estimation method combining double exponential smoothing and adaptive multi-scale state space-position perception attention

Abstract

The invention discloses a lithium ion battery SOC estimation method combining double exponential smoothing and adaptive multi-scale state space-position perception attention, and belongs to the field of lithium ion battery management, and the method comprises the steps: carrying out the physical heuristic denoising and trend maintenance of current, voltage and temperature original signals through double exponential smoothing; constructing short-time and long-time double branches by adopting a self-adaptive multi-scale diagonal state space model S4D, and generating scale modulation vectors based on input statistical characteristics to realize time domain discretization of input dependence; and further introducing position sensing Performmer linear attention with a rotation position code RoPE to carry out global time sequence refining, and finally outputting an SOC estimated value through time pooling and a lightweight regression head. The method has causal consistency and linear time complexity, and is suitable for real-time BMS deployment of the electric vehicle and the energy storage system.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery management, and particularly relates to a lithium-ion battery SOC estimation method that combines bi-exponential smoothing with adaptive multi-scale state space-position-aware attention. Background Technology

[0002] State of charge (SOC) estimation for lithium-ion batteries is a core function of battery management systems, directly impacting the safety, lifespan, and performance of electric vehicles and energy storage systems. Existing methods mainly include model-driven methods based on equivalent circuit models and filters, and data-driven methods based on recurrent neural networks and long short-term memory networks. The former relies on accurate battery parameters and noise assumptions, making it prone to deviations under drastic changes in operating conditions, temperature fluctuations, and battery aging. While the latter can improve fitting capabilities through end-to-end learning, it still has limitations in noise robustness, multi-timescale dynamic modeling, and long-sequence computational efficiency.

[0003] Current SOC estimation techniques still face three prominent problems: First, traditional smoothing methods struggle to preserve the dynamic trend of the signal while denoising, affecting the accuracy of subsequent modeling. Second, single-timescale state-space models or recurrent networks cannot simultaneously capture the battery's rapid transient response and slow evolution process. Third, while self-attention-based sequence models can model long-range dependencies, their quadratic computational complexity is insufficient to meet the real-time requirements of embedded BMS. Therefore, a new SOC estimation architecture that can collaboratively achieve robust denoising, adaptive multi-scale dynamic modeling, and linear complexity global inference is urgently needed. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a lithium-ion battery SOC estimation method combining bi-exponential smoothing and adaptive multi-scale state-space-position-aware attention, comprising:

[0005] Acquire current, voltage, and temperature measurement data of the battery over a time series;

[0006] The current, voltage, and temperature are each subjected to double exponential smoothing to obtain smoothed input signals;

[0007] The smoothed input signal is standardized and linearly embedded to obtain a feature embedding sequence;

[0008] The feature embedding sequence is divided into a short-time scale branch and a long-time scale branch, and an adaptive discretization scaling vector is generated based on the input statistical features.

[0009] A diagonal state-space model is constructed on the short-time-scale branch and the long-time-scale branch respectively, and zero-order preserved discretization is performed using an input-dependent adaptive discretization step size.

[0010] The outputs of the short-time scale branch and the long-time scale branch are fused into a multi-scale representation through a gating mechanism;

[0011] The multi-scale representation is subjected to global temporal refinement using a position-aware linear attention mechanism, wherein rotational position encoding is introduced in the attention calculation to maintain temporal dependencies;

[0012] The refined sequence representation is time-pooled, and the state of charge is estimated by multilayer perceptron regression.

[0013] Optionally, the current, voltage, and temperature are each subjected to double exponential smoothing, including:

[0014] Calculate the horizontal component at the current moment based on the measured value at the current moment, the horizontal component and the trend component at the previous moment;

[0015] Calculate the trend component at the current moment based on the difference between the horizontal component at the current moment and the horizontal component at the previous moment, as well as the trend component at the previous moment.

[0016] The smoothing parameter used to calculate the horizontal component has a value range of 0.5 to 0.9, and the smoothing parameter used to calculate the trend component has a value range of 0.1 to 0.5.

[0017] Optionally, the acquired battery measurement data constitutes a time series input window, and the length of the input window ranges from 30 to 90.

[0018] Optionally, generating the adaptive discretization scaling vector based on the input statistical features includes:

[0019] Based on the mean characteristics of the input window, the intermediate vector is calculated through a two-layer feedforward network containing the ReLU activation function;

[0020] The intermediate vector is mapped and scaled to the range of 0.5 to 2 using the Sigmoid function to obtain the adaptive discretized scaling vector.

[0021] Optionally, the basic discretization step size range used in the short-time scale branch is 0.001 to 0.02 seconds, and the basic discretization step size range used in the long-time scale branch is 0.01 to 0.2 seconds.

[0022] Optionally, the global temporal refinement using a position-aware linear attention mechanism includes:

[0023] The query matrix and key matrix are transformed using a feature mapping function to approximate the Softmax kernel function in standard attention, thereby achieving attention computation with linear complexity.

[0024] Optionally, the introduction of rotational position encoding in the attention calculation includes:

[0025] The angle base is calculated based on a preset base of 10,000;

[0026] Construct a rotation matrix based on the stated angle base and time step position;

[0027] The rotation matrix is ​​used to rotate the query vector and key vector to encode the absolute position while maintaining the relative positional relationship.

[0028] Optionally, the step of performing time pooling on the refined sequence representation and regressing it using a multilayer perceptron includes:

[0029] The global feature vector is obtained by averaging the sequence over the time dimension.

[0030] The global feature vector is input into a multilayer perceptron containing a hidden layer and a ReLU activation function, and the final state-of-charge estimate is output.

[0031] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0032] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0033] Compared with the prior art, the present invention has the following advantages and technical effects:

[0034] (1) An adaptive multi-scale S4D state space structure is proposed to realize hierarchical modeling of the dynamic characteristics of the battery in both long and short time. It can simultaneously capture rapid transient changes and slow evolution trends, thereby enhancing the dynamic response capability of SOC estimation.

[0035] (2) An input-related adaptive discretization mechanism is introduced to enable the model to automatically adjust the time scale according to the current working state, thereby improving the generalization performance under different working conditions, different temperatures and different discharge rates;

[0036] (3) The position-aware Performer attention module is used to achieve efficient integration of global temporal information, which significantly improves the stability and accuracy of long sequence SOC estimation while maintaining linear complexity;

[0037] (4) The overall model structure is lightweight, and high-precision estimation can be obtained without a large number of convolutions or deep loop stacking. The computational cost is low, which is suitable for the real-time deployment needs of vehicle BMS and embedded hardware.

[0038] (5) It exhibits excellent estimation accuracy and robustness under various temperature conditions, different battery chemical systems and cross-dataset test scenarios, and has higher cross-domain adaptability than traditional RNN / TCN / Transformer models. Attached Figure Description

[0039] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0040] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0042] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a lithium-ion battery SOC estimation method that combines bi-exponential smoothing with adaptive multi-scale state space-position-aware attention, including:

[0045] Acquire current, voltage, and temperature measurement data of the battery over a time series;

[0046] The current, voltage, and temperature are each subjected to double exponential smoothing to obtain smoothed input signals;

[0047] The smoothed input signal is standardized and linearly embedded to obtain a feature embedding sequence;

[0048] The feature embedding sequence is divided into a short-time scale branch and a long-time scale branch, and an adaptive discretization scaling vector is generated based on the input statistical features.

[0049] A diagonal state-space model is constructed on the short-time-scale branch and the long-time-scale branch respectively, and zero-order preserved discretization is performed using an input-dependent adaptive discretization step size.

[0050] The outputs of the short-time scale branch and the long-time scale branch are fused into a multi-scale representation through a gating mechanism;

[0051] The multi-scale representation is subjected to global temporal refinement using a position-aware linear attention mechanism, wherein rotational position encoding is introduced in the attention calculation to maintain temporal dependencies;

[0052] The refined sequence representation is time-pooled, and the state of charge is estimated by multilayer perceptron regression.

[0053] Specifically, it includes:

[0054] (1) Acquire raw sensing measurement data of the battery over a time series, wherein the raw sensing measurement data includes at least current. ,Voltage With temperature Forming a length of Input window:

[0055] ;

[0056] (2) Perform double exponential smoothing (DES) noise reduction processing on each channel of the current, voltage and temperature to obtain the smoothed input. For any univariate time series DES satisfies:

[0057] ;

[0058] ;

[0059] ;

[0060] (3) Standardize the smoothed input sequence obtained in step (2) and perform linear embedding to obtain the feature embedding sequence:

[0061] ;

[0062] in These are trainable parameters;

[0063] (4) The feature embedding sequence is obtained by linear mapping:

[0064] ;

[0065] and along the feature dimension Divided into short-time scale branches With long timescale branch ;

[0066] (5) Generate adaptive discretization scaling vector based on the statistical features of the input window ,in:

[0067] ;

[0068] (6) Construct diagonal state-space models S4D on the short-time scale branch and the long-time scale branch respectively, and adopt an input-dependent adaptive discretization step size for each latent dimension, satisfying at least the following:

[0069] ;

[0070] And make the short-time scale branch and the long-time scale branch have different basic discretization ranges;

[0071] (7) The S4D is discretized using zero-order hold (ZOH) discretization from continuous-time state space to discrete time. The continuous-time system satisfies:

[0072] ;

[0073] ;

[0074] Discretization yields:

[0075] ;

[0076] ;

[0077] And time step Recursion:

[0078] ;

[0079] ;

[0080] (8) Output short-time scale With long timescale output The time-varying gating fusion obtained through learning is as follows:

[0081] ;

[0082] (9) The fused multi-scale representation is subjected to global sequence refinement using position-aware Performer attention, where the standard self-attention form is:

[0083]

[0084] Performer maps positive random features or learnable features. Approximate softmax kernel:

[0085] ;

[0086] And obtain linear attention:

[0087] ;

[0088] (10) In the Performer attention, a Rotational Position Encoding (RoPE) is introduced for the query and key to encode the time order. Specifically, for the first... Two-dimensional pairs in position The rotation matrix is:

[0089] ;

[0090] And satisfy the similarity relationship related to relative displacement:

[0091] ;

[0092] (11) Perform time pooling on the sequence representation refined by Performer to obtain global features, and obtain the SOC estimate through multilayer perceptron (MLP) regression:

[0093]

[0094]

[0095] Steps (2) to (3) constitute an end-to-end adaptive multi-scale S4D-Performer SOC estimation framework.

[0096] Furthermore, the length of the input window Take any value between 30 and 90.

[0097] Furthermore, the length of the input window This corresponds to a 60-second data window with a sampling frequency of 1 Hz.

[0098] Furthermore, the parameters of the double exponential smoothing satisfy... , .

[0099] Furthermore, the parameters of the double exponential smoothing are taken as follows: , .

[0100] Furthermore, to ensure causality, the initial level and initial trend of the double exponential smoothing are obtained using statistics from the training set, satisfying:

[0101] ;

[0102] ;

[0103] in .

[0104] Furthermore, the embedding dimension Take any value between 32 and 256.

[0105] Furthermore, the embedding dimension .

[0106] Furthermore, the short-time scale branch and the long-time scale branch are divided proportionally according to the feature dimension.

[0107] Furthermore, the basic discretization interval of the short-time scale branch satisfies Seconds, the basic discretization interval of the long-time-scale branch satisfies Second.

[0108] Furthermore, the effective discretization step sizes for the short-time scale and the long-time scale respectively satisfy:

[0109]

[0110] ;

[0111] Furthermore, the discrete system matrices of the short-time scale branch and the long-time scale branch respectively satisfy:

[0112] ;

[0113] ;

[0114] in .

[0115] Furthermore, the number of Performer attention layers is any value from 1 to 6, and the number of attention heads is any value from 1 to 6.

[0116] Furthermore, the Performer uses a base of 10000 for rotational position encoding.

[0117] Furthermore, the random feature dimension of the Performer Take any value between 64 and 256.

[0118] Furthermore, the position-aware Performer and the adaptive multi-scale S4D are stacked with a Pre-LN structure, residual connections, and feedforward sub-layers.

[0119] Furthermore, the model estimates the SOC at a fixed temperature with an RMSE of 0.77% and a MAE of 0.63%, and at a dynamic temperature with an RMSE of 1.10% and a MAE of 0.89%.

[0120] Furthermore, the model exhibits good generalization ability across different chemical systems, with an RMSE of less than 1.70% on Al23LiFePO4 batteries.

[0121] Furthermore, the model exhibits high accuracy in SOC estimation under fixed temperature conditions. When tested on publicly available battery datasets such as NASA and CALCE, its SOC estimation error RMSE is as low as 0.52%–0.68%, and its MAE is as low as 0.41%–0.55%.

[0122] Furthermore, the model maintains stable estimation accuracy under dynamic temperature changes, with the RMSE of the SOC estimate being less than 0.95% and the MAE being less than 0.78%, indicating that the model has strong robustness to temperature disturbances.

[0123] Furthermore, the model exhibits excellent cross-cell generalization ability in lithium-ion batteries with different chemical systems. When performing transfer tests on datasets of battery systems such as Li-ion, LiFePO4, and NMC, the SOC estimation RMSE is less than 1.50%, demonstrating superior cross-domain generalization performance.

[0124] Furthermore, compared to traditional LSTM, GRU, TCN, and recent Transformer-type models, the proposed method improves RMSE and MAE by an average of 12%–28% under the same data conditions, and can achieve improvements of over 30% under dynamic load and strong interference conditions, demonstrating significant performance advantages.

[0125] On the other hand, this embodiment also provides a lithium-ion battery SOC estimation system that combines bi-exponential smoothing with adaptive multi-scale state-space-position-aware attention. Specifically, it includes:

[0126] (1) Data acquisition module, used to acquire time series data of battery current, voltage and temperature;

[0127] (2) Signal conditioning module, used to perform double exponential smoothing (DES) on the time series data by channel;

[0128] (3) Feature embedding module, used to map the smoothed input to Dimensional potential space;

[0129] (4) Adaptive multi-scale S4D module, which includes short-time scale branch and long-time scale branch, and has an input-dependent adaptive discretization step size and gating fusion mechanism.

[0130] (5) The position-aware Performer module is used to perform global refinement of multi-scale features with linear complexity and introduces RoPE to encode the temporal order.

[0131] (6) SOC regression module, used to output SOC estimates based on time pooling and MLP.

[0132] The signal conditioning module and These are either fixed hyperparameters or trainable hyperparameters.

[0133] The gated network of the adaptive multi-scale S4D module is a two-layer feedforward network, containing hidden layers and generated using a nonlinear activation function. .

[0134] Example 2

[0135] This embodiment provides a verification method for SOC estimation under fixed temperature conditions, and the specific implementation is as follows:

[0136] (1) The publicly available NASA and CALCE battery datasets were selected, and the driving cycle data under three fixed temperature conditions of 25°C, 10°C and 0°C were used as training and testing data. The training set included various operating conditions (such as DST, UDDS, LA92, etc.), and the testing set included dynamic load cycles that were not used in the training (such as US06, FUDS). The data sampling frequency was 1Hz, and each sample contained 60 time steps of current, voltage and temperature sequences.

[0137] (2) Preprocessing the raw data: Perform double exponential smoothing (DES) noise reduction on each channel; use minimum-maximum normalization to scale the input to the [0,1] interval; construct a sliding window sequence with a step size of 1 for 60 time steps.

[0138] (3) The AdamW optimizer was used for model training, with an initial learning rate of 0.0005, a batch size of 128, and 150 training epochs. Cosine annealing learning rate strategy and early stopping mechanism were used. The loss function was mean squared error (MSE). Performance was evaluated on the test set after training.

[0139] Experimental results show that:

[0140] At 25°C, the estimated SOC of the US06 cycle is 0.52% RMSE and 0.41% MAE; the RMSE of the HWFET cycle is 0.68% and MAE is 0.55%.

[0141] At 0°C, the US06 cycle RMSE is 1.12% and MAE is 0.94%; the HWFET cycle RMSE is 1.47% and MAE is 1.10%.

[0142] The results show that the AM-S4D-Performer model proposed in this invention maintains high estimation accuracy and good stability even in low-temperature environments.

[0143] Example 3

[0144] This embodiment provides a verification method for SOC estimation under dynamic temperature conditions, specifically including the following steps:

[0145] (1) Select the driving cycle data of the battery under dynamic temperature conditions, including continuous temperature change scenarios from −20°C to 10°C and from 10°C to 25°C, covering low temperature recovery, rapid heating and steady state.

[0146] (2) The model trained under fixed temperature conditions is directly applied to the above dynamic temperature data without any additional fine-tuning. The model outputs an SOC estimate every 60 time steps.

[0147] The experimental results show that:

[0148] During the temperature rise from 10°C to 25°C, the model's SOC estimate had an RMSE of 0.95%, a MAE of 0.78%, and a maximum error of 2.30%.

[0149] During the low-temperature heating process from −20°C to 10°C, the RMSE was 1.28%, the MAE was 1.03%, and the maximum error was 4.15%.

[0150] The results show that the AM-S4D-Performer model described in this invention can still maintain stable estimation ability under dynamic temperature changes, and has excellent temperature robustness.

[0151] Example 4

[0152] This embodiment provides a verification of the generalization ability of SOC estimation across battery chemistry systems, as detailed below:

[0153] (1) The model trained on NMC type battery data was directly transferred to LiFePO4 battery data without any fine-tuning. The test data was selected using DST and FUDS driven loops, and the ambient temperature was room temperature.

[0154] Experimental results show that:

[0155] In the DST cycle, the SOC estimate has an RMSE of 1.21%, a MAE of 0.98%, and a maximum error of 4.02%.

[0156] In the FUDS cycle, the RMSE was 1.47%, the MAE was 1.15%, and the maximum error was 4.89%.

[0157] The results show that the AM-S4D-Performer model proposed in this invention can maintain good estimation accuracy in different battery chemical systems and has strong generalization ability across chemical systems.

[0158] Example 5

[0159] This embodiment conducts a comparative experiment with the traditional SOC estimation method, as detailed below:

[0160] (1) Standard LSTM model (parameter scale aligned with the method of this invention);

[0161] (2) Extended Kalman Filter (EKF) method;

[0162] (3) Integral method of ampere-hours;

[0163] (4) Open circuit voltage (OCV) method.

[0164] A comparative experiment was conducted in a US06 cycle at a fixed temperature of 25°C.

[0165] Experimental results show that the RMSE of the method of the present invention is 0.52%, which is significantly better than LSTM (0.81%), EKF (1.72%), ampere-hour integration method (3.08%) and OCV method (2.65%).

[0166] Especially during the high-current pulse and high-speed change phases, the error fluctuation of the method of this invention is significantly lower than that of other methods, demonstrating its superior dynamic tracking capability.

[0167] Example 6

[0168] To verify the contribution of each module to the overall performance, the following four ablation models were designed for experimentation:

[0169] (1) Use only basic S4D (remove multi-scale structures and Performer attention);

[0170] (2) Multi-scale S4D, but removing the adaptive discretization mechanism;

[0171] (3) Multi-scale S4D + adaptive discretization, but remove Performer attention;

[0172] (4) Complete AM-S4D-Performer model.

[0173] The experiment was conducted in HWFET cycling at 10°C.

[0174] Experimental results show that the RMSE of the complete model is 0.74%, which is better than the model without adaptive discretization (0.89%), the model without attention module (1.02%), and the basic S4D model only (1.15%).

[0175] Ablation experiments show that both adaptive multiscale structure and position-aware Performer attention significantly contribute to improving model performance, and the combination of the two achieves the best results.

[0176] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0177] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0178] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A lithium-ion battery SOC estimation method combining bi-exponential smoothing and adaptive multi-scale state-space-position-aware attention, characterized in that, include: Acquire current, voltage, and temperature measurement data of the battery over a time series; The current, voltage, and temperature are each subjected to double exponential smoothing to obtain smoothed input signals; The smoothed input signal is standardized and linearly embedded to obtain a feature embedding sequence; The feature embedding sequence is divided into a short-time scale branch and a long-time scale branch, and an adaptive discretization scaling vector is generated based on the input statistical features. A diagonal state-space model is constructed on the short-time-scale branch and the long-time-scale branch respectively, and zero-order preserved discretization is performed using an input-dependent adaptive discretization step size. The outputs of the short-time scale branch and the long-time scale branch are fused into a multi-scale representation through a gating mechanism; The multi-scale representation is subjected to global temporal refinement using a position-aware linear attention mechanism, wherein rotational position encoding is introduced in the attention calculation to maintain temporal dependencies; The refined sequence representation is time-pooled, and the state of charge is estimated by multilayer perceptron regression.

2. The method according to claim 1, characterized in that, The current, voltage, and temperature are each subjected to double exponential smoothing, including: Calculate the horizontal component at the current moment based on the measured value at the current moment, the horizontal component and the trend component at the previous moment; Calculate the trend component at the current moment based on the difference between the horizontal component at the current moment and the horizontal component at the previous moment, as well as the trend component at the previous moment. The smoothing parameter used to calculate the horizontal component has a value range of 0.5 to 0.9, and the smoothing parameter used to calculate the trend component has a value range of 0.1 to 0.

5.

3. The method according to claim 1, characterized in that, The acquired battery measurement data constitutes a time series input window, and the length of the input window ranges from 30 to 90.

4. The method according to claim 1, characterized in that, The process of generating an adaptive discretized scaling vector based on input statistical features includes: Based on the mean characteristics of the input window, the intermediate vector is calculated through a two-layer feedforward network containing the ReLU activation function; The intermediate vector is mapped and scaled to the range of 0.5 to 2 using the Sigmoid function to obtain the adaptive discretized scaling vector.

5. The method according to claim 1, characterized in that, The basic discretization step size range used in the short-time scale branch is 0.001 to 0.02 seconds, and the basic discretization step size range used in the long-time scale branch is 0.01 to 0.2 seconds.

6. The method according to claim 1, characterized in that, The global temporal refinement using a position-aware linear attention mechanism includes: The query matrix and key matrix are transformed using a feature mapping function to approximate the Softmax kernel function in standard attention, thereby achieving attention computation with linear complexity.

7. The method according to claim 6, characterized in that, The introduction of rotational position encoding in attention calculation includes: Calculate the angle base value based on the preset base value; Construct a rotation matrix based on the stated angle base and time step position; The rotation matrix is ​​used to rotate the query vector and key vector to encode the absolute position while maintaining the relative positional relationship.

8. The method according to claim 1, characterized in that, The step of performing time pooling on the refined sequence representation and regressing it using a multilayer perceptron includes: The global feature vector is obtained by averaging the sequence over the time dimension. The global feature vector is input into a multilayer perceptron containing a hidden layer and a ReLU activation function, and the final state-of-charge estimate is output.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images