Lithium ion battery state of charge estimation method and system fusing time scale features
By fusing time-scale features and utilizing the TCN and MSC structures and gating fusion mechanism, the adaptability and stability issues of lithium-ion battery state-of-charge estimation under complex operating conditions were solved, achieving accurate SOC estimation under multiple operating conditions and temperatures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAMEN INST OF RARE EARTH MATERIALS
- Filing Date
- 2026-02-03
- Publication Date
- 2026-06-12
AI Technical Summary
Existing methods for estimating the state of charge of lithium-ion batteries lack adaptability and stability under complex operating conditions, making it difficult to accurately reflect the remaining usable capacity of the battery.
A method for fusing time-scale features is adopted, which captures long-term temporal dependencies and short-term high-frequency features through a dual-branch structure of TCN and MSC respectively. The contribution ratio of different features is dynamically balanced through a gating fusion mechanism. Combined with parameterized hyperbolic tangent linear unit and smooth quadratic loss function, the weighted fusion of charge state is realized.
It significantly improves the accuracy and robustness of lithium-ion battery state of charge estimation, and can effectively estimate the battery state of charge under multiple temperature and operating conditions, making it suitable for battery management systems.
Smart Images
Figure CN122193918A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lithium-ion battery management technology, and in particular to a method and system for estimating the state of charge of lithium-ion batteries by incorporating time-scale characteristics. Background Technology
[0002] Lithium-ion batteries are widely used in electric vehicles and energy storage systems due to their high energy density and long cycle life. To ensure the safe operation and performance of lithium-ion batteries, it is usually necessary to monitor and control the battery status in real time within the battery management system. Among these parameters, SOC (State of Charge) is an important parameter reflecting the remaining usable capacity of the battery and is one of the key state variables in the battery management system. Accurate SOC estimation is of great significance for range assessment, charge and discharge control, and safety protection.
[0003] Since the state of charge (SOC) cannot be directly measured by sensors, existing technologies typically estimate the SOC indirectly based on measurable parameters such as battery voltage, current, and temperature, using specific algorithms or models. In related research and engineering applications, various SOC estimation schemes have been proposed to predict the battery's SOC under different operating conditions and environmental environments.
[0004] However, in practical applications, the battery's operating state is affected by various factors such as changes in operating conditions, temperature, and the battery's own state, resulting in a highly dynamic and nonlinear state estimation process.
[0005] With the improvement of computing power, data-driven state of charge estimation methods have been gradually applied. These methods estimate the state of charge by modeling a large amount of time-series data collected during battery operation.
[0006] However, under complex operating conditions, battery operating data simultaneously contains features at different time scales, such as both slowly changing long-term features and rapidly changing transient features. These real-world features at different time scales pose significant challenges and difficulties for estimating the state of charge (SOC) of lithium-ion batteries in practical applications. Existing SOC estimation methods for lithium-ion batteries suffer from insufficient adaptability and stability under complex operating conditions. Summary of the Invention
[0007] The purpose of this invention is to address the problems and shortcomings of existing technologies by providing a method and system for estimating the state of charge (SOC) of lithium-ion batteries that integrates time-scale features, thereby improving the accuracy and robustness of SOC estimation under different operating conditions and environmental conditions.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] In a first aspect, the present invention provides a method for estimating the state of charge of a lithium-ion battery by incorporating time-scale features, comprising the following steps:
[0010] At least two battery operating parameters are collected in the corresponding time series during the operation of lithium-ion batteries. The battery operating parameters are preprocessed to construct derived dynamic features and form an input feature sequence. The long-term trend features and short-term dynamic features during the battery operation are modeled separately, and corresponding feature extraction networks are constructed to extract the long-term trend features and short-term dynamic features respectively.
[0011] An adaptive fusion mechanism is constructed to dynamically generate fusion weights based on the battery's operating state. Based on the fusion weights, long-term trend features and short-term dynamic features are weighted and fused to obtain a fusion feature representation. Based on the fusion feature representation, the estimated state of charge of the lithium-ion battery at the corresponding time step is output.
[0012] As a preferred implementation, the process involves collecting at least two battery operating parameters at corresponding time series during the operation of the lithium-ion battery, preprocessing the battery operating parameters to construct derived dynamic features, and forming an input feature sequence. This specifically includes the following steps:
[0013] At least two battery operating parameters of lithium-ion batteries under different temperature conditions and different driving cycle conditions were collected. Outlier removal and missing value completion were performed on the raw data. Min-max normalization was used to map all features to the interval [-1, 1] to eliminate the influence of dimensional differences and form a standardized feature sequence.
[0014] As a preferred implementation, the battery operating parameters include at least: voltage, current, temperature, and one or more of the following derived features: voltage change rate, current change rate, cumulative charge / discharge capacity, or equivalent integral capacity. The battery operating parameters are derived from experimental or operational data under various operating conditions and ambient temperatures. Specific operating conditions include: constant current operating conditions and dynamic operating conditions. Specific ambient temperatures include: room temperature and non-room temperature conditions.
[0015] As a preferred implementation, the step of modeling the long-term trend features and short-term dynamic features during battery operation, constructing corresponding feature extraction networks, and extracting the long-term trend features and short-term dynamic features specifically includes the following steps:
[0016] The standardized feature sequences are input in parallel into two complementary feature extraction networks for synchronous processing:
[0017] The long-term trend feature extraction network adopts the temporal convolutional network (TCN) structure. By stacking dilated causal convolution and residual connections, the temporal receptive field is expanded exponentially, accurately capturing the long-term temporal dependencies in the evolution of the state of charge of lithium-ion batteries and characterizing the slowly changing long-term trend features.
[0018] The short-term dynamic feature extraction network adopts a multi-scale convolutional MSC structure. Through parallel convolutional kernels of different sizes, it simultaneously extracts high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
[0019] As a preferred implementation, the adaptive fusion mechanism dynamically generates fusion weights based on the battery operating state; and weights the long-term trend features and short-term dynamic features according to the fusion weights to obtain a fused feature representation, specifically including the following steps:
[0020] Based on the gated fusion mechanism, an adaptive fusion module is constructed. The TCN output features are encoded through a one-dimensional convolutional layer and dynamically generated by the Sigmoid function. According to the current operating state of the battery, the contribution ratio of long-term trend features and short-term dynamic features is adaptively adjusted, and the long-term trend features and short-term dynamic features are weighted and fused to obtain a fused feature representation that combines global trends and local details.
[0021] As a preferred implementation, the gating fusion mechanism specifically includes:
[0022] Based on long-term trend features and short-term dynamic features, a gate value s is generated, which is a weight vector or weight scalar between 0 and 1. According to the gate value s, the long-term trend features and short-term dynamic features are weighted and combined to obtain a fused feature representation, which satisfies: fused feature representation = s ⊙ long-term trend features + (1-s) ⊙ short-term dynamic features, where ⊙ represents element-wise multiplication or channel-wise multiplication.
[0023] As a preferred implementation, based on fused feature representation, the estimated state of charge of the lithium-ion battery at the corresponding time step is output, specifically including the following steps:
[0024] The fused feature representation is input into the fully connected layer, and the nonlinear expressive power is enhanced by the parameterized hyperbolic tangent linear unit PTeLU activation function. Finally, after nonlinear mapping, the estimated state of charge of the lithium-ion battery at the corresponding time step is output.
[0025] During model training, a smoothed quadratic loss function is used to constrain the prediction error. The smoothed quadratic loss function applies a quadratic penalty to small error intervals and a linear or bounded growth penalty to large error intervals.
[0026] Secondly, the present invention also provides a lithium-ion battery state-of-charge estimation system that integrates time-scale features, the system comprising:
[0027] The data acquisition module is used to collect the operating parameters of lithium-ion batteries and generate a time series.
[0028] The preprocessing module is used to: perform anomaly handling and normalization on the running parameters to form the input feature sequence;
[0029] The feature extraction module is used to: model the long-term trend features and short-term dynamic features during battery operation, construct corresponding feature extraction networks, and extract the long-term trend features and short-term dynamic features respectively;
[0030] The adaptive fusion module is used to: construct an adaptive fusion mechanism, dynamically generate fusion weights based on the battery operating status, and perform weighted fusion of long-term trend features and short-term dynamic features based on the fusion weights to obtain a fused feature representation;
[0031] The state of charge output module is used to: output the estimated state of charge of the lithium-ion battery at the corresponding time step based on the fused feature representation.
[0032] As a preferred implementation, the feature extraction module is specifically used to: input the standardized feature sequence in parallel into two complementary feature extraction networks for synchronous processing.
[0033] The long-term trend feature extraction network adopts the temporal convolutional network (TCN) structure. By stacking dilated causal convolution and residual connections, the temporal receptive field is expanded exponentially, accurately capturing the long-term temporal dependencies in the evolution of the state of charge of lithium-ion batteries and characterizing the slowly changing long-term trend features.
[0034] The short-term dynamic feature extraction network adopts a multi-scale convolutional MSC structure. Through parallel convolutional kernels of different sizes, it simultaneously extracts high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
[0035] As a preferred implementation, the long-term trend feature extraction network includes: a residual connection structure and a normalization structure; the residual connection structure is used to: superimpose the input and output of the convolutional unit residuals; the normalization structure is layer normalization or batch normalization.
[0036] The short-term dynamic feature extraction network includes at least two parallel convolutional sub-branches, each of which uses different convolutional kernel sizes or different pooling scales to extract local features at different time scales.
[0037] Compared with the prior art, the present invention has at least the following beneficial effects:
[0038] (1) The present invention adopts a dual-branch structure of TCN and MSC to capture long-term time-series dependence and short-term high-frequency features respectively, so as to achieve comprehensive coverage of features at multiple time scales and greatly improve the adaptability to complex working conditions.
[0039] (2) The present invention adopts a gated fusion mechanism to dynamically balance the contribution ratio of different features. The GTM-PS (Gated Temporal Convolutional Network with Multi-Scale Convolution using Parameterized Tanh Linear Unit and Smoothed Quadratic Loss) model adopts a parameterized hyperbolic tangent linear unit activation function and a smooth quadratic loss function, which can be flexibly adjusted according to the real-time operating status of the battery, significantly improving the stability and consistency of SOC estimation.
[0040] (3) The present invention uses an optimized activation function and loss function to enhance the nonlinear expression ability and noise resistance of the model, and ensures high robustness under multiple temperature and working conditions.
[0041] (4) The voltage, current, and temperature data of lithium-ion batteries can be directly measured, and SOC estimation can be achieved based on derived features. This invention models the multi-source time-series features during battery operation and combines them with a reasonable feature fusion mechanism to achieve effective estimation of the state of charge of lithium-ion batteries under different operating conditions and ambient temperatures, exhibiting good robustness and engineering applicability. The technical solution of this invention does not require modeling of the complex internal mechanisms of lithium-ion batteries, has high computational efficiency, is easy to implement in automotive BMS (Battery Management System), and has broad application prospects. Attached Figure Description
[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0043] Figure 1 This is a flowchart of the lithium-ion battery state-of-charge estimation method that integrates time-scale features in Embodiment 1 of the present invention;
[0044] Figure 2 This is a block diagram of the lithium-ion battery state-of-charge estimation system that integrates time-scale features in Embodiment 2 of the present invention.
[0045] Figure 3 This is a structural block diagram of the GTM-PS model in Embodiment 3 of the present invention;
[0046] Figure 4 This is a schematic diagram of the overall framework of the GTM-PS model in Embodiment 3 of the present invention, showing the complete process of data processing, model training, SOC prediction and error analysis;
[0047] Figure 5 for Figure 4 A schematic diagram of the data processing section;
[0048] Figure 6 for Figure 4 A schematic diagram of the model training section;
[0049] Figure 7 for Figure 4 A schematic diagram of the SOC prediction and error analysis section;
[0050] Figure 8 This is a schematic diagram of the core structure of TCN and MSC in an embodiment of the present invention, wherein (a) is a dilated causal convolution structure, (b) is a residual block structure, and (c) is a multi-scale convolution structure;
[0051] Figure 9 for Figure 8 (a) Schematic diagram of dilated causal convolution structure;
[0052] Figure 10 for Figure 8 (b) Schematic diagram of the residual block structure;
[0053] Figure 11 for Figure 8 (c) Schematic diagram of multi-scale convolution structure;
[0054] Figure 12 This is a schematic diagram illustrating the change of SOC estimation results over time under LA92 operating conditions and 25°C conditions according to an embodiment of the present invention.
[0055] Figure 13 This is a schematic diagram illustrating the change of SOC estimation error over time under LA92 operating conditions and 25°C conditions according to an embodiment of the present invention.
[0056] Figure 14 This is a schematic diagram illustrating the SOC estimation accuracy under different operating conditions according to an embodiment of the present invention;
[0057] Figure 15 This is a schematic diagram illustrating the SOC estimation accuracy under different ambient temperature conditions according to an embodiment of the present invention. Detailed Implementation
[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0059] In the description of this invention, it should be understood that the orientation or positional relationship indicated by terms such as "above" is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use, or the orientation or positional relationship in which those skilled in the art are usually understood. It is only for the convenience of describing this invention and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.
[0060] The order in which the embodiments are described below is not intended to limit the preferred order of the embodiments. Furthermore, in the description of this application, the term "comprising" means "including but not limited to". Various embodiments of the invention may exist in the form of a range; it should be understood that the description in the form of a range is merely for convenience and brevity and should not be construed as a rigid limitation on the scope of the invention; therefore, it should be considered that the range description has specifically disclosed all possible sub-ranges and single numerical values within that range. For example, it should be considered that the range description from 1 to 6 has specifically disclosed sub-ranges, such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and single numbers within the range, such as 1, 2, 3, 4, 5, and 6, regardless of the range. Additionally, whenever a numerical range is indicated herein, it means including any referenced number (fraction or integer) within the indicated range.
[0061] Example 1: A method for estimating the state of charge of lithium-ion batteries by incorporating time-scale features
[0062] Given that existing lithium-ion battery state-of-charge (SOC) estimation methods suffer from insufficient adaptability and stability under complex operating conditions, this invention provides a lithium-ion battery SOC estimation method that integrates time-scale features. (See [link to relevant documentation]). Figure 1 As shown, it includes the following steps:
[0063] S1. Collect at least two battery operating parameters in the corresponding time series during the operation of the lithium-ion battery, preprocess the battery operating parameters, construct derived dynamic features, and form an input feature sequence;
[0064] S2. Model the long-term trend features and short-term dynamic features during battery operation, construct corresponding feature extraction networks, and extract the long-term trend features and short-term dynamic features respectively.
[0065] S3. Construct an adaptive fusion mechanism to dynamically generate fusion weights based on the battery operating status; and perform weighted fusion of long-term trend features and short-term dynamic features based on the fusion weights to obtain a fusion feature representation.
[0066] S4. Based on the fusion feature representation, output the estimated state of charge of the lithium-ion battery at the corresponding time step.
[0067] As a preferred implementation, step S1 specifically includes the following steps:
[0068] At least two battery operating parameters of lithium-ion batteries under different temperature conditions and different driving cycle conditions were collected. Outlier removal and missing value completion were performed on the raw data. Min-max normalization was used to map all features to the interval [-1, 1] to eliminate the influence of dimensional differences and form a standardized feature sequence.
[0069] Battery operating parameters include at least: voltage, current, temperature, and one or more of the following derived characteristics: rate of change of voltage, rate of change of current, cumulative charge and discharge capacity, or equivalent integral capacity.
[0070] The battery operating parameters are derived from experimental or operational data under various operating conditions and ambient temperatures. Specific operating conditions include constant current conditions and dynamic operating conditions, and specific ambient temperatures include room temperature and non-room temperature conditions.
[0071] As a preferred implementation, step S2 specifically includes the following steps:
[0072] Based on the input feature sequence, feature representations with different time scales or different dynamic characteristics are extracted through at least two feature extraction networks.
[0073] At least one feature extraction network is used to: model the long-term temporal evolution characteristics of the battery during operation.
[0074] At least one other feature extraction network is used to: model the short-term dynamic changes in battery operation;
[0075] For ease of description below, these two feature extraction networks will be named: Long-term trend feature extraction network and Short-term dynamic feature extraction network, respectively.
[0076] The standardized feature sequences are input in parallel into two complementary feature extraction networks for synchronous processing:
[0077] The long-term trend feature extraction network adopts the TCN (Temporal Convolutional Network) structure. By stacking dilated causal convolutions and residual connections, it exponentially expands the temporal receptive field, accurately captures the long-term temporal dependencies in the evolution of the state of charge of lithium-ion batteries, and characterizes the slowly changing long-term trend features.
[0078] The short-term dynamic feature extraction network adopts the MSC (Multi-Scale Convolutional) structure. Through parallel convolution kernels of different sizes, it simultaneously extracts high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
[0079] As a preferred implementation, step S3 specifically includes the following steps:
[0080] Based on the gated fusion mechanism, an adaptive fusion module is constructed. The TCN output features are encoded through a one-dimensional convolutional layer and dynamically generated by the Sigmoid function. According to the current operating state of the battery, the contribution ratio of long-term trend features and short-term dynamic features is adaptively adjusted, and the long-term trend features and short-term dynamic features are weighted and fused to obtain a fused feature representation that combines global trends and local details.
[0081] The gating fusion mechanism specifically includes:
[0082] Based on long-term trend features and short-term dynamic features, a gate value s is generated, which is a weight vector or weight scalar between 0 and 1. According to the gate value s, the long-term trend features and short-term dynamic features are weighted and combined to obtain a fused feature representation, which satisfies: fused feature representation = s ⊙ long-term trend features + (1-s) ⊙ short-term dynamic features, where ⊙ represents element-wise multiplication or channel-wise multiplication.
[0083] As a preferred implementation, step S4 specifically includes the following steps:
[0084] The fused feature representation is input into the fully connected layer, and the nonlinear expression capability is enhanced by the PTeLU (Parameterized Tanh Linear Unit) activation function. Finally, after nonlinear mapping, the estimated state of charge of the lithium-ion battery at the corresponding time step is output.
[0085] During model training, a smoothed quadratic loss function is used to constrain the prediction error. The smoothed quadratic loss function applies a quadratic penalty to small error intervals and a linear or bounded growth penalty to large error intervals.
[0086] Example 2: Lithium-ion battery state-of-charge estimation system integrating time-scale features
[0087] Based on the same inventive concept, embodiments of the present invention provide a lithium-ion battery state-of-charge estimation system that integrates time-scale features, see [link to relevant documentation]. Figure 2 As shown, the system includes:
[0088] The data acquisition module is used to collect the operating parameters of lithium-ion batteries and generate a time series.
[0089] The preprocessing module is used to: perform anomaly handling and normalization on the running parameters to form the input feature sequence;
[0090] The feature extraction module is used to: model the long-term trend features and short-term dynamic features during battery operation, construct corresponding feature extraction networks, and extract the long-term trend features and short-term dynamic features respectively;
[0091] The adaptive fusion module is used to: construct an adaptive fusion mechanism, dynamically generate fusion weights based on the battery operating status, and perform weighted fusion of long-term trend features and short-term dynamic features based on the fusion weights to obtain a fused feature representation;
[0092] The state of charge output module is used to: output the estimated state of charge of the lithium-ion battery at the corresponding time step based on the fused feature representation.
[0093] As a preferred implementation, the feature extraction module is specifically used to: input the standardized feature sequences in parallel into two complementary feature extraction networks for synchronous processing.
[0094] The long-term trend feature extraction network adopts the TCN structure. Through stacked dilated causal convolution and residual connections, it exponentially expands the temporal receptive field, accurately captures the long-term temporal dependence in the evolution of the state of charge of lithium-ion batteries, and characterizes the slowly changing long-term trend features.
[0095] The short-term dynamic feature extraction network adopts the MSC structure and uses parallel convolution kernels of different sizes to simultaneously extract high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
[0096] The long-term trend feature extraction network specifically includes a convolution-based temporal modeling structure, which includes dilated convolutions or convolutional units that expand the receptive field and satisfy temporal causal constraints to avoid using information from future moments; wherein, the dilation factor increases layer by layer, and the increasing method is doubling or exponential.
[0097] Furthermore, the long-term trend feature extraction network includes: a residual connection structure and a normalization structure; the residual connection structure is used to: superimpose the input and output of the convolutional unit residuals; the normalization structure is layer normalization or batch normalization to improve training stability and suppress gradient vanishing or gradient exploding.
[0098] Furthermore, the short-term dynamic feature extraction network includes at least two parallel convolutional sub-branches, each employing different kernel sizes or pooling scales to extract local features at different time scales.
[0099] Specifically, battery operating parameters include at least voltage, current, and temperature; and further include one or more of the following derived characteristics: voltage change rate dV / dt, current change rate dI / dt, cumulative charge / discharge capacity, or equivalent integral capacity.
[0100] The battery operating parameters are derived from experimental or operational data under various operating conditions and ambient temperatures. Specific operating conditions include constant current conditions and dynamic operating conditions, and specific ambient temperatures include room temperature and non-room temperature conditions.
[0101] The adaptive fusion mechanism in this embodiment is a gated fusion mechanism, which specifically includes:
[0102] Based on long-term trend features and short-term dynamic features, a gate value s is generated, where s is a weight vector or weight scalar between 0 and 1. According to the gate value s, the long-term trend features and short-term dynamic features are weighted and combined to obtain a fused feature representation, which satisfies: Fusion feature representation = s ⊙ Long-term features + (1-s) ⊙ Short-term features.
[0103] Here, ⊙ represents element-wise multiplication or channel-wise multiplication.
[0104] In step S4, the nonlinear mapping of the fused feature representation is applied using a nonlinear activation function PTeLU with learnable parameters. The parameters of the nonlinear activation function PTeLU are used to control the negative half-axis response and nonlinear curvature to enhance the ability to express the nonlinear dynamics of the battery.
[0105] During model training, the SQL (Smooth Quadratic Loss) function is used to constrain the prediction error.
[0106] The smoothed quadratic loss function employs an approximate quadratic penalty for small error intervals and an approximate linear or bounded growth penalty for large error intervals to reduce the impact of abnormal fluctuations, noise points, or abrupt changes on model training.
[0107] Example 3: SOC Estimation Method Based on GTM-PS Model
[0108] The method of this application is further illustrated below with specific embodiments. This section further illustrates the content of the present invention in conjunction with specific embodiments, but should not be construed as limiting the present invention. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art. To make the technical solution of the present invention clearer and easier to understand, the lithium-ion battery SOC estimation method of the present invention is described in detail below with reference to specific experimental data, model parameters and operating steps.
[0109] This embodiment uses the LG18650 lithium-ion battery dataset, which has a rated capacity of 3Ah and an operating voltage range of 2.8V-4.2V, perfectly matching the typical application scenarios of electric vehicles and energy storage systems.
[0110] During the data acquisition process, three constant temperature gradients were set: 0℃, 10℃, and 25℃, covering both normal and low temperature application scenarios. Temperature stability was maintained through a constant temperature chamber, with fluctuation errors controlled within ±0.2℃.
[0111] In terms of operating conditions, there are four typical types: standard operating conditions US06 (high-speed acceleration and deceleration), UDDS (urban low-speed), and LA92 (combined urban and suburban operating conditions). The mixed operating condition is Mixeds (a complex dynamic operating condition that combines the above standard operating condition segments), which fully simulates the diverse operating states of electric vehicles in actual driving.
[0112] See Figure 3 As shown, the GTM-PS model constructed in this embodiment generally includes three parts: data processing, model training, and SOC estimation and error analysis. The specific implementation plan is as follows: Figure 4 As shown.
[0113] See Figure 4 and Figure 5 As shown, to enhance the model's ability to capture the instantaneous electrochemical response of the battery, the voltage-time derivative dV / dt is calculated based on the collected voltage data, using the following formula: , where Δt is the difference between the current time (t) and the previous sampling time (t-1).
[0114] See Figure 4 and Figure 5 As shown, min-max normalization is finally used to map the four types of feature data—voltage, current, temperature, and dV / dt—to the interval [-1, 1]. The normalization formula is: , where x is the original eigenvalue, x min x max These are the minimum and maximum values of the feature, respectively, to ensure that each feature has equal weight during model training.
[0115] The dataset was divided according to a scientific and reasonable proportion principle. The training set was selected from Mixed Cycles 2, 4, 5, 7, and 8 at three different temperatures for model parameter learning.
[0116] The validation set uses data from Mixed Cycle 1 at three temperatures to monitor overfitting during training, and an early stopping mechanism (patience value 50) is set.
[0117] The test set uses data from UDDS, US06, LA92, and Mixed6 at three temperatures to comprehensively evaluate the model's generalization ability.
[0118] See Figure 4 and Figure 6 As shown, the GTM-PS model constructed in this embodiment of the invention includes: TCN structure, MSC structure, GF mechanism, PTeLU activation function and SQL loss function. The parameter settings and working principles of each component are as follows:
[0119] See Figure 6 As shown, the TCN structure serves as the core for capturing long-term temporal dependencies. It consists of four convolutional layers, each employing dilated causal convolution with dilation coefficients of 1, 2, 4, and 8, respectively. This achieves an exponential expansion of the receptive field, ensuring the ability to capture long-term trend information during the battery SOC evolution process.
[0120] The convolutional kernel size is set to 3×1 to ensure the causality of temporal data and prevent future information leakage. Residual connections are added after each convolutional layer to alleviate the gradient vanishing problem in deep networks. The activation function is PTeLU, and the output dimension is 128. The specific structures of TCN and MSC are as follows: Figure 6 As shown.
[0121] See Figure 6 As shown, the MSC structure employs three parallel convolutional branches with kernel sizes of 3, 5, and 7, respectively, to extract features corresponding to different local time windows. This enables the accurate capture of short-term, high-frequency transient information such as current spikes and voltage surges. The outputs of each branch are concatenated and then subjected to channel weight calibration through a squeezing-excitation layer to enhance the response intensity of effective features. The final output is a 128-dimensional short-term feature vector, which complements the output of the TC module.
[0122] See Figure 6 As shown, the GF mechanism dynamically balances the contribution ratio of long-term trend features and short-term dynamic features through a data-driven approach. First, it uses a 1×1 convolutional layer to process the TCN output feature h. t Encode the data and generate a gate value s using the Sigmoid function. t The calculation formula is: Then follow the formula: To complete feature fusion, where b t For the output of the MSC module, ⊙ indicates element-wise multiplication or channel-wise multiplication, which enhances the weight of long-term features when the battery is running smoothly and improves the contribution of short-term features when the operating conditions change abruptly.
[0123] See Figure 6 As shown, the activation function used in this embodiment is PTeLU. By introducing learnable parameters α and β to adaptively adjust the nonlinear mapping curve, the gradient sparsity problem of the traditional ReLU function in the negative region is solved. The formula is as follows: The initial value of α is set to 1.5, and the initial value of β is set to 1.2. They are adaptively updated as the model is optimized during training.
[0124] See Figure 6 As shown, the loss function used in this embodiment is the SQL (Smooth Quadratic Loss) function, and the formula is: Where y is the true value of SOC. For the predicted value, a double penalty is applied to small errors to ensure estimation accuracy; for large errors, a linear penalty is applied to reduce outlier interference and improve training stability and noise resistance.
[0125] See Figure 6 As shown, the GTM-PS model training in this embodiment uses the Adam optimizer, with specific parameter settings as follows:
[0126] The learning rate is set to 5e-4, and the numerical stability parameter is set to 1e-5. The batch size is set to 512 to balance training efficiency and memory usage. The sequence length is set to 96, meaning that each input sample contains feature data for 96 consecutive time steps, covering 1.6 minutes of battery operation. The number of iterations is set to 100, combined with an early stopping mechanism (patience value of 50), which stops training and saves the current optimal model parameters when the validation set metric shows no improvement for 50 consecutive iterations.
[0127] During training, the preprocessed training set data is sliced according to a sequence length of 96 to generate an input feature matrix (dimension: N×96×4, where N is the number of samples and 4 is the number of features). After initializing the model parameters (following a normal distribution N(0,0.01)), the feature matrix is input into the TCN and MSC structures in parallel, and the dual-branch output is fused through the GF mechanism.
[0128] See Figure 7As shown, the model is mapped to SOC predicted values (dimension: N×1) through a fully connected layer. After calculating the SQL loss value, the backpropagation algorithm is used to update all model parameters (including the learnable parameters of TCN, MSC, GF layers and PTeLU). In each iteration, the RMSE (Root Mean Squared Error) and MAE (Mean Squared Error) indices are calculated using the validation set to ensure the model's generalization ability.
[0129] This embodiment provides a method for estimating the state of charge of a lithium-ion battery based on the GTM-PS model, including the following steps:
[0130] Step 1: Data Acquisition and Preprocessing
[0131] See Figure 5 As shown, time series data of voltage, current and temperature of lithium-ion batteries were obtained under different temperature conditions (0℃, 10℃, 25℃, etc.) and different driving cycle conditions (US06, UDDS, LA92, mixed conditions, etc.).
[0132] Based on voltage data, the voltage time derivative (dV / dt) is calculated as a dynamic supplementary feature;
[0133] The original data is processed by outlier removal and missing value completion. Min-max normalization is used to map all features to the interval [-1, 1] to eliminate the influence of dimensional differences and form a standardized feature sequence.
[0134] Step 2: Multi-timescale feature extraction
[0135] See Figure 6 As shown, the standardized feature sequences are input in parallel into two complementary feature extraction networks for synchronous processing:
[0136] The long-term trend feature extraction network adopts the TCN structure. Through stacked dilated causal convolution and residual connections, it exponentially expands the temporal receptive field, accurately captures the long-term temporal dependence in the evolution of the state of charge of lithium-ion batteries, and characterizes the slowly changing long-term trend features.
[0137] The short-term dynamic feature extraction network adopts the MSC structure and uses parallel convolution kernels of different sizes (3, 5, 7, etc.) to simultaneously extract high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features such as current fluctuations and voltage transient responses.
[0138] Figure 8 The diagram shows the core structures of TCN and MSC in this embodiment of the invention, where (a) is a dilated causal convolution structure, (b) is a residual block structure, and (c) is a multi-scale convolution structure. Figure 9for Figure 8 (a) Schematic diagram of dilated causal convolution structure; Figure 10 for Figure 8 (b) Schematic diagram of the residual block structure.
[0139] See Figure 8 , Figure 9 , Figure 10 As shown, the TCN structure in this embodiment adopts a causal convolution design with increasing dilation factor, which expands the receptive field and efficiently captures long-term dependencies across time scales without increasing computational complexity.
[0140] Figure 11 for Figure 8 (c) Schematic diagram of multi-scale convolution structure.
[0141] See Figure 11 As shown, the MSC structure in this embodiment uses parallel computation of multi-size convolution kernels to simultaneously extract high-frequency features at different local time scales, enhancing the ability to capture key dynamic information such as instantaneous current spikes and voltage changes, thus forming a functional complementarity.
[0142] Step 3, Adaptive Fusion Step
[0143] See Figure 6 As shown, an adaptive fusion module is constructed based on the GF (Gated Fusion) mechanism. The TCN output features are encoded through the Conv1d layer (one-dimensional convolutional layer), and the gating weights are dynamically generated through the Sigmoid function.
[0144] The adaptive fusion mechanism in this embodiment is a gated fusion mechanism. By gating the features output by different feature extraction networks, the model can adaptively adjust between long-term trend features and short-term dynamic features according to the battery's operating status, thereby obtaining a more stable and reliable SOC estimation result.
[0145] Based on the battery's current operating state, such as current change rate and temperature fluctuation, the contribution ratio of long-term trend characteristics and short-term dynamic characteristics is adaptively adjusted, specifically according to the formula: The long-term trend characteristics and short-term dynamic characteristics are weighted and fused, where z t h represents the features after gating fusion. t For TCN output, b t For MSC output, s t The value is the gate value, and ⊙ indicates element-wise multiplication or channel-wise multiplication.
[0146] By weighted fusion of long-term trend features and short-term dynamic features, a fused feature representation that combines global trends with local details is obtained.
[0147] Step 4: SOC Estimation Output Step
[0148] See Figure 6 and Figure 7 As shown, the fused feature representation is input into the fully connected layer, and the nonlinear expressive power is enhanced by the PTeLU activation function. Finally, after nonlinear mapping, the SOC estimate corresponding to the time step is output.
[0149] The PTeLU activation function, by introducing learnable parameters α and β, adaptively adjusts the nonlinear mapping curve, solving the problem of gradient sparsity in the negative interval of the traditional ReLU (Rectified Linear Unit) function, and enhancing the model's ability to express the nonlinear electrochemical characteristics of the battery.
[0150] See Figure 6 As shown, in the model training process of this embodiment, SQL functions are used to constrain training errors, balance the penalty weights of errors at different scales, and improve training stability and noise resistance.
[0151] See Figure 6 As shown, the SQL loss function employs a quadratic penalty for small errors to ensure estimation accuracy, and a linear penalty for large errors to reduce outlier interference, significantly improving the training stability and estimation reliability of the model under noisy data and complex conditions.
[0152] The method for estimating the state of charge of a lithium-ion battery according to the present invention will be further described below with reference to specific embodiments and accompanying drawings. It should be understood that the following embodiments are only used to explain the technical solutions of the present invention, and not to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the concept of the present invention should all fall within the scope of protection of the present invention.
[0153] In this embodiment, the trained SOC estimation model is applied to test data under typical dynamic operating conditions to illustrate the SOC estimation results.
[0154] See Figure 12 As shown, Figure 12 The diagram illustrates the changes in SOC estimation results and SOC reference values over time under LA92 operating conditions and an ambient temperature of 25°C.
[0155] from Figure 12As can be seen, the SOC shows a trend of gradual change over time during the entire operation. The method of this embodiment can track the overall trend of SOC change well. During the stage of frequent switching of operating conditions and current changes, the SOC estimation results remain continuous and smooth, without obvious sudden changes or divergence.
[0156] The results show that the SOC estimation method in this embodiment of the invention can effectively model the long-term changing trend of the battery during operation and is suitable for actual working conditions with complex dynamic characteristics.
[0157] To further illustrate the stability of the SOC estimation process, examples are provided to illustrate the variation of the SOC estimation error under the same operating conditions and temperature conditions described above.
[0158] like Figure 13 As shown, Figure 13 The diagram illustrates the change of SOC estimation error over time for the SOC estimation method in this embodiment of the invention under LA92 operating conditions and an ambient temperature of 25°C.
[0159] from Figure 13 It can be seen that throughout the entire operation, the SOC estimation error remained within a small range and did not accumulate or increase significantly when the operating conditions changed.
[0160] The results show that the SOC estimation method in this embodiment of the invention has good stability under dynamic operating conditions, can effectively suppress the spread of error over time, and meets the requirements of the battery management system for the stability of online SOC estimation.
[0161] In this embodiment, to illustrate the adaptability of the method of the present invention under different working conditions, statistical analysis is performed on the SOC estimation results based on data from various typical working conditions.
[0162] like Figure 14 As shown, Figure 14 The diagram illustrates the SOC estimation method in this embodiment of the invention, showing the SOC estimation accuracy under different operating conditions, including LA92, US06, UDDS, and mixed operating conditions.
[0163] The accuracy indicators in the embodiments of the present invention include mean absolute error, root mean square error, and coefficient of determination.
[0164] from Figure 14 It can be seen that under different working conditions, the SOC estimation method in the embodiments of the present invention can obtain stable SOC estimation results and the accuracy index remains within a reasonable range, indicating that the SOC estimation method in the embodiments of the present invention has good adaptability and consistency for different types of working conditions.
[0165] Figure 14 To verify the applicability of the SOC estimation method in the embodiments of the present invention under different ambient temperature conditions, examples of SOC estimation results under different temperature conditions are provided.
[0166] like Figure 15 As shown, Figure 15 The diagram illustrates the SOC estimation accuracy under different ambient temperature conditions according to the SOC estimation method in this embodiment of the invention, where the ambient temperature includes 25℃, 10℃ and 0℃.
[0167] from Figure 15 It can be seen that, under different temperature conditions, the SOC estimation method in the embodiments of the present invention can maintain stable SOC estimation performance and does not deteriorate significantly due to temperature changes.
[0168] The results show that the SOC estimation method in this embodiment of the invention has good robustness to changes in ambient temperature and can be applied to battery operation scenarios under multiple temperature conditions.
[0169] Combination Figures 12 to 15 The results show that the SOC estimation method in this embodiment of the invention, by modeling the characteristics of different time scales during battery operation and fusing the relevant features, can achieve stable and reliable SOC estimation under complex operating conditions and different ambient temperatures, and has good engineering application feasibility.
[0170] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for estimating the state of charge (SOC) of a lithium-ion battery by incorporating time-scale features, characterized in that, Includes the following steps: Collect at least two battery operating parameters in the corresponding time series during the operation of lithium-ion batteries, preprocess the battery operating parameters, construct derived dynamic features, and form an input feature sequence; The long-term trend characteristics and short-term dynamic characteristics of the battery operation process are modeled separately, and corresponding feature extraction networks are constructed to extract the long-term trend characteristics and short-term dynamic characteristics respectively. An adaptive fusion mechanism is constructed to dynamically generate fusion weights based on the battery operating status; Based on the fusion weights, long-term trend features and short-term dynamic features are weighted and fused to obtain a fused feature representation; Based on the fused feature representation, the estimated state of charge of the lithium-ion battery at the corresponding time step is output.
2. The lithium-ion battery state-of-charge estimation method integrating time-scale features as described in claim 1, characterized in that, The process involves collecting at least two battery operating parameters from the lithium-ion battery during operation, preprocessing these parameters to construct derived dynamic features, and forming an input feature sequence. Specifically, this includes the following steps: At least two battery operating parameters of lithium-ion batteries under different temperature conditions and different driving cycle conditions were collected. Outlier removal and missing value completion were performed on the raw data. Min-max normalization was used to map all features to the interval [-1, 1] to eliminate the influence of dimensional differences and form a standardized feature sequence.
3. The lithium-ion battery state-of-charge estimation method integrating time-scale features as described in claim 2, characterized in that, The battery operating parameters include at least: voltage, current, temperature, and one or more of the following derived characteristics: voltage change rate, current change rate, cumulative charge / discharge capacity, or equivalent integral capacity. The battery operating parameters are derived from experimental or operational data under various operating conditions and ambient temperatures. Specific operating conditions include: constant current operating conditions and dynamic operating conditions. Specific ambient temperatures include: room temperature and non-room temperature conditions.
4. The lithium-ion battery state-of-charge estimation method integrating time-scale features as described in claim 2, characterized in that, The process of modeling the long-term trend characteristics and short-term dynamic characteristics during battery operation, constructing corresponding feature extraction networks, and extracting long-term trend characteristics and short-term dynamic characteristics specifically includes the following steps: The standardized feature sequences are input in parallel into two complementary feature extraction networks for synchronous processing: The long-term trend feature extraction network adopts the temporal convolutional network (TCN) structure. By stacking dilated causal convolution and residual connections, the temporal receptive field is expanded exponentially, accurately capturing the long-term temporal dependencies in the evolution of the state of charge of lithium-ion batteries and characterizing the slowly changing long-term trend features. The short-term dynamic feature extraction network adopts a multi-scale convolutional MSC structure. Through parallel convolutional kernels of different sizes, it simultaneously extracts high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
5. The lithium-ion battery state-of-charge estimation method integrating time-scale features as described in claim 4, characterized in that, The adaptive fusion mechanism is constructed by dynamically generating fusion weights based on the battery operating status. Based on the fusion weights, long-term trend features and short-term dynamic features are weighted and fused to obtain a fused feature representation, which specifically includes the following steps: Based on the gated fusion mechanism, an adaptive fusion module is constructed. The TCN output features are encoded through a one-dimensional convolutional layer and the gate weights are dynamically generated by the Sigmoid function. Based on the current operating state of the battery, the contribution ratio of long-term trend features and short-term dynamic features is adaptively adjusted, and the long-term trend features and short-term dynamic features are weighted and fused to obtain a fused feature representation that combines global trends and local details.
6. The lithium-ion battery state-of-charge estimation method based on fused time-scale features as described in claim 5, characterized in that, The gating fusion mechanism specifically includes: Based on long-term trend features and short-term dynamic features, a gate value s is generated, which is a weight vector or weight scalar between 0 and 1. According to the gate value s, the long-term trend features and short-term dynamic features are weighted and combined to obtain a fused feature representation, which satisfies: fused feature representation = s ⊙ long-term trend features + (1-s) ⊙ short-term dynamic features, where ⊙ represents element-wise multiplication or channel-wise multiplication.
7. The lithium-ion battery state-of-charge estimation method integrating time-scale features as described in claim 5, characterized in that, Based on the fused feature representation, the estimated state of charge of the lithium-ion battery at the corresponding time step is output, specifically including the following steps: The fused feature representation is input into the fully connected layer, and the nonlinear expressive power is enhanced by the parameterized hyperbolic tangent linear unit PTeLU activation function. Finally, after nonlinear mapping, the estimated state of charge of the lithium-ion battery at the corresponding time step is output. During model training, a smoothed quadratic loss function is used to constrain the prediction error. The smoothed quadratic loss function applies a quadratic penalty to small error intervals and a linear or bounded growth penalty to large error intervals.
8. A lithium-ion battery state-of-charge estimation system incorporating time-scale features, characterized in that, The system includes: The data acquisition module is used to collect the operating parameters of lithium-ion batteries and generate a time series. The preprocessing module is used to: perform anomaly handling and normalization on the running parameters to form the input feature sequence; The feature extraction module is used to: model the long-term trend features and short-term dynamic features during battery operation, construct corresponding feature extraction networks, and extract the long-term trend features and short-term dynamic features respectively; The adaptive fusion module is used to: construct an adaptive fusion mechanism, dynamically generate fusion weights based on the battery operating status, and perform weighted fusion of long-term trend features and short-term dynamic features based on the fusion weights to obtain a fused feature representation; The state of charge output module is used to: output the estimated state of charge of the lithium-ion battery at the corresponding time step based on the fused feature representation.
9. The lithium-ion battery state-of-charge estimation system integrating time-scale features as described in claim 8, characterized in that, The feature extraction module is specifically used to: input standardized feature sequences in parallel into two complementary feature extraction networks for synchronous processing. The long-term trend feature extraction network adopts the temporal convolutional network (TCN) structure. By stacking dilated causal convolution and residual connections, the temporal receptive field is expanded exponentially, accurately capturing the long-term temporal dependencies in the evolution of the state of charge of lithium-ion batteries and characterizing the slowly changing long-term trend features. The short-term dynamic feature extraction network adopts a multi-scale convolutional MSC structure. Through parallel convolutional kernels of different sizes, it simultaneously extracts high-frequency transient features within different local time windows, accurately characterizing short-term dynamic features.
10. The lithium-ion battery state-of-charge estimation system integrating time-scale features as described in claim 9, characterized in that, The long-term trend feature extraction network includes: a residual connection structure and a normalization structure; the residual connection structure is used to: superimpose the input and output of the convolutional unit residuals; the normalization structure is layer normalization or batch normalization; The short-term dynamic feature extraction network includes at least two parallel convolutional sub-branches, each of which uses different convolutional kernel sizes or different pooling scales to extract local features at different time scales.