Informer battery SOC prediction method based on bidirectional calibration

Abstract

The invention relates to the technical field of electric vehicle battery management systems, in particular to an Informer battery SOC prediction method based on bidirectional calibration, which comprises the following steps: collecting battery key parameter data in the running process of an electric vehicle, generating a dynamic characteristic matrix and inputting the dynamic characteristic matrix into an Informer encoder, and optimizing input characteristic weight distribution through a learnable characteristic space projection operator. According to the method, online adaptive weight distribution of input features is realized through a dynamic feature re-calibration mechanism, dependence on a prior physical model is not needed, the long-range time sequence correlation modeling capability is enhanced in combination with a double-path residual structure, gradient conflicts among modules are avoided through a gradient decoupling alternating optimization strategy, training oscillation is reduced, and the convergence speed is increased; and meanwhile, through modular integration and structural optimization, prediction precision, speed and robustness are considered, the defects of a traditional method are overcome, and more reliable technical support is provided for a vehicle-mounted battery management system.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle battery management system technology, specifically to the Informer battery SOC prediction method based on bidirectional calibration. Background Technology

[0002] The state of charge (SOC) of an electric vehicle battery is a core parameter of the battery management system, and its accurate prediction is crucial for ensuring vehicle safety and extending battery life. However, battery systems exhibit strong nonlinearity and long-term time-series dependence, and their characteristics are affected by complex operating conditions such as temperature fluctuations, current disturbances, and battery aging, making it difficult to guarantee the accuracy of SOC prediction.

[0003] Among existing SOC prediction methods, the Kalman filter series of algorithms have good real-time performance and interpretability, but their performance is heavily dependent on an accurate battery equivalent circuit model. Their accuracy drops significantly in scenarios of model mismatch, such as battery aging and drastic temperature changes. Methods based on recurrent neural networks such as LSTM, GRU, and CNN can capture temporal dependencies, but they are limited by the gradient vanishing problem and have difficulty effectively modeling long-range correlations across charge and discharge cycles. Summary of the Invention

[0004] The purpose of this invention is to provide an Informer battery SOC prediction method based on bidirectional calibration to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an Informer battery SOC prediction method based on bidirectional calibration, comprising the following steps:

[0006] S1. Data preprocessing: Collect key battery parameter data during the operation of electric vehicles, clean and select features to generate a dynamic feature matrix, and realize adaptive calibration of the input feature weights.

[0007] S2, Encoder Optimization: Input the dynamic feature matrix generated in step S1 into the Informer encoder, optimize the input feature weight allocation through a learnable feature space projection operator; adopt a dual-path residual structure to fuse the feature outputs of the attention residual path and the convolution enhancement path; introduce a gradient decoupling mechanism to block the gradient backpropagation of the dynamic feature matrix in stages and suppress parameter oscillations in the early stage of training.

[0008] S3. Decoder optimization: The dynamic feature matrix is ​​embedded in the Informer model decoder to achieve feature recalibration; the relationship between encoder output and decoder input is established through cross-attention mechanism; a hierarchical parameterized residual learning strategy is adopted; hierarchical learnable health factors and channel-level factors are introduced to optimize feature fusion and improve model stability.

[0009] S4. Model Training and Prediction: The optimized encoder output is fed into the decoder. After feature fusion and mapping, the predicted battery SOC value is output. The prediction performance is evaluated by the coefficient of determination R², mean square error MSE, and mean absolute error MAE.

[0010] Preferably, the generation of the dynamic feature matrix in step S1 includes:

[0011] S11. Calculate the Pearson correlation coefficient between the characteristics and SOC. The formula is:

[0012] ;

[0013] in, These are the observed values ​​of features and SOC, respectively. , respectively, are the average values ​​of the features and SOC, where n is the number of samples;

[0014] S12. Use the K-means clustering algorithm to partition dimensions. The objective function is:

[0015] ;

[0016] Where K is the number of clusters, It is the center of the k-th cluster. These are data points;

[0017] S13. Calculate feature weights using the normalized correlation coefficient. The formula is:

[0018] ;

[0019] in, is the correlation coefficient between the j-th feature and SOC, and m is the total number of features;

[0020] S14. Calculate the correlation matrix between features. Generate dynamic feature matrix The formula is:

[0021] ;

[0022] ;

[0023] in, It is the correlation coefficient between the i-th feature and the j-th feature. It is the feature index of the j-th feature under the i-th working condition.

[0024] Preferably, the dual-path residual structure of the encoder in step S2 satisfies:

[0025] Attention residual path: Stable updates are achieved through layer normalization and random deactivation, as shown in the formula:

[0026] ;

[0027] in, , D is the feature dimension. Let h be a small value with a denominator of zero, h be the single-layer output vector, and m be the random mask vector. This represents the probability of inactivation. For the first The output of each attention head, It is a dynamic feature matrix;

[0028] Convolution Enhancement Path: Dilated convolution and a non-linear activation function are used to capture local patterns. The formula is as follows:

[0029] ;

[0030] Where Conv represents the convolution operation, and the convolution kernel is... Step length ,filling α is the activation function parameter;

[0031] Dual-output weighted fusion:

[0032] ;

[0033] in These are adaptive weight parameters.

[0034] Preferably, the gradient decoupling mechanism in step S2 is implemented through alternating optimization, decomposing the objective function into two sub-problems:

[0035] ;

[0036] ;

[0037] in To fix the dynamic feature matrix parameters, To optimize attention / convolution parameters, To fix the attention / convolution parameters, To optimize the dynamic feature matrix parameters, To determine the number of training iterations, This is the loss function.

[0038] Preferably, the feature recalibration operation of the decoder in step S3 is as follows:

[0039] ;

[0040] in For decoder input, For batch size, The length of the decoded sequence. For feature dimensions;

[0041] It is a trainable parameter matrix;

[0042] This is the bias vector.

[0043] Preferably, the hierarchical parameterized residual learning strategy in step S3 satisfies:

[0044] Hierarchical feature update:

[0045] ;

[0046] Output feature calibration:

[0047] ;

[0048] in, Here, ⊙ represents the sigmoid function, ⊙ denotes channel-by-channel product, and N is the total number of decoder layers. For hierarchical learnable health factors, As a channel-level factor, Here, d represents the transformation function of the current layer of the decoder, and d represents the feature dimension.

[0049] when As the network approaches zero, it degenerates into an identity mapping, requiring optimization of gradient flow.

[0050] when When the value approaches 1, the transformation of the current layer is strengthened.

[0051] Preferably, the calculation method for the cross-attention mechanism in step S3 is as follows;

[0052] ;

[0053] ;

[0054] in, For decoder input, Encoder output, The length of the encoded sequence. Let be the projection matrix. For feature dimension, is the head dimension, and Softmax is the normalization function.

[0055] Preferably, the key battery parameters in step S1 include at least one of total voltage, total current, ambient temperature, vehicle speed, and motor temperature.

[0056] Preferably, the formulas for calculating the coefficient of determination R², mean square error MSE, and mean absolute error MAE in step S4 are as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] in, Represents the actual observed value. This represents the model's predicted value. It is the average of the actual values.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] This invention achieves online adaptive weight allocation of input features through a dynamic feature recalibration mechanism. It can proactively adapt to complex operating conditions such as battery temperature fluctuations, aging, and high-rate charging and discharging without relying on prior physical models. Combined with a dual-path residual structure, it enhances the ability to model long-range time-series correlations. Through gradient decoupling and alternating optimization strategies, it avoids gradient conflicts between modules, reduces training oscillations, and accelerates convergence. At the same time, through modular integration and structural optimization, it balances prediction accuracy, speed, and robustness, effectively making up for the shortcomings of various traditional methods and providing more reliable technical support for vehicle battery management systems. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the overall process of the Informer battery SOC prediction method based on bidirectional calibration of the present invention;

[0064] Figure 2 This is a schematic diagram of SOC prediction when the battery temperature is -10~10℃ in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram illustrating SOC prediction when the battery temperature is 10~20℃ in an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram illustrating SOC prediction when the battery temperature is 20~30℃ in an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram illustrating SOC prediction when the battery temperature is 30~45℃ in an embodiment of the present invention;

[0068] Figure 6This is a schematic diagram of SOC prediction for vehicle 7 in an embodiment of the present invention;

[0069] Figure 7 This is a schematic diagram of SOC prediction for vehicle 20 in an embodiment of the present invention;

[0070] Figure 8 This is a schematic diagram illustrating the SOC prediction of an electric vehicle equipped with a lithium iron phosphate battery in an embodiment of the present invention;

[0071] Figure 9 This is a schematic diagram illustrating the SOC prediction of an electric vehicle equipped with a lithium manganese oxide battery in an embodiment of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Informer is a self-attention prediction model. As an improved prediction model, it leverages its ProbSparse self-attention mechanism and distillation encoder structure to reduce computational complexity to O(LlogL) while maintaining the powerful representational capabilities of the Transformer model, significantly improving the efficiency of long sequence processing. Its generative decoder outputs all prediction points in a single forward propagation, effectively avoiding the error accumulation problem in the autoregressive process. More importantly, this architecture does not rely on any prior physical model and can directly learn the complex nonlinear mapping relationship between voltage, current, temperature, and SOC from historical data, exhibiting stronger generalization ability for operating conditions and battery aging states not encountered during training.

[0074] Current SOC prediction technologies are mainly divided into three categories, and each category has significant shortcomings:

[0075] Traditional physical modeling methods, represented by Kalman filtering algorithms, rely on accurate battery equivalent circuit models to construct state equations. Their advantages lie in strong real-time performance and good interpretability, but their core drawback is their extremely high dependence on model parameters. When batteries face complex operating conditions such as aging, drastic temperature fluctuations, and high-rate charging and discharging, the model is prone to mismatch, leading to a significant decrease in SOC estimation accuracy and making it difficult to meet practical application requirements.

[0076] Recurrent neural network methods, including models such as LSTM and GRU, capture dependencies in battery time-series data through network structures. However, these models are limited by the vanishing gradient problem and cannot effectively model long-range time-series correlations across charge-discharge cycles, and are insufficiently adaptable to state changes throughout the battery's entire lifespan.

[0077] Self-attention prediction model method: The Informer model is a typical example. Through the ProbSparse self-attention mechanism and distillation encoder structure, the computational complexity of long sequence processing is reduced to O(LlogL), significantly improving processing efficiency. Simultaneously, this model does not rely on prior physical models and can directly learn the nonlinear mapping relationship between voltage, current, temperature, and SOC from historical data, exhibiting strong generalization ability. However, the traditional Informer model still has inherent limitations:

[0078] Static feature processing: The standardized input features are directly encoded, and all features have the same initial weight. There is no mechanism to dynamically adjust the feature weights according to the real-time state of the battery (such as temperature, current changes, and aging degree). Sensor noise and random outliers can easily enter the core calculation module, affecting the prediction accuracy.

[0079] Passive adaptation to operating conditions: After the network parameters are trained, they remain fixed. The adaptation to operating conditions such as battery aging, extreme temperatures, and high-rate charging and discharging relies entirely on the model's generalization ability, which is a passive adaptation mode. For example, during high-rate charging and discharging, the current dynamics change drastically, but the model's weight allocation for current features is consistent with the static voltage, failing to focus on key signals and leading to increased instantaneous errors. In the later stages of battery aging, the characteristics of the terminal voltage curve change, and the model cannot adaptively adjust the voltage signal weights, causing systematic biases.

[0080] The training process is unstable: the encoder and decoder adopt independent architectures and lack interaction mechanisms. During training, the gradients of different modules (such as feature extractor and sequence modeler) are prone to conflict or cancellation, resulting in severe training oscillations, slow convergence speed, difficulty in reaching the global optimum, and indirectly limiting the upper limit of prediction accuracy.

[0081] With the increasing driving range of electric vehicles and the development of fast charging technology, batteries are facing increasingly complex operating conditions. Scenarios such as wide temperature range (-10℃~45℃), high-rate charging and discharging, and long-life-cycle aging place new demands on the accuracy of SOC prediction. The static feature processing, passive operating condition adaptation, and training instability issues of traditional Informer models can no longer meet the high accuracy and robustness requirements of practical applications. Therefore, this paper proposes an Informer battery SOC prediction model design based on encoder-decoder bidirectional calibration.

[0082] Please see Figure 1-9This invention provides a technical solution: an Informer battery SOC prediction method based on bidirectional calibration, comprising the following steps:

[0083] S1. Data preprocessing: Collect key battery parameter data during the operation of electric vehicles. Key battery parameters include at least one of total voltage, total current, ambient temperature, vehicle speed, and motor temperature. After cleaning and feature selection, a dynamic feature matrix is ​​generated.

[0084] Specifically:

[0085] Data collection: In this stage, a dedicated data collection module is used to acquire key parameters of the electric vehicle during operation, including total voltage, total current, ambient temperature, vehicle speed, and motor temperature.

[0086] Data cleaning aims to remove null values, outliers, and noise from the data, thereby improving the overall quality and reliability of the data. This step ensures the accuracy of the analysis and the effectiveness of the model.

[0087] Feature selection: Important features are selected from the data based on their correlation with the battery's state of charge (SOC). This process helps to focus on the data features most critical to SOC prediction.

[0088] Matrix operations: First, the matrix elements are determined based on the correlation between features and the battery's state of charge (SOC). The correlation coefficient between each feature and SOC is calculated to assess their impact on SOC. The generation of the dynamic feature matrix includes:

[0089] S11. Calculate the Pearson correlation coefficient between the characteristics and SOC. The formula is:

[0090] (1);

[0091] in, These are the observed values ​​of features and SOC, respectively. , respectively, are the average values ​​of the features and SOC, where n is the number of samples;

[0092] S12. Dimensional partitioning is performed using the K-means clustering algorithm. Dimensional partitioning assigns features to different dimensions based on their importance and interrelationships. This helps to better capture the interactions between features in matrix calculations. Dimensional partitioning can be achieved through clustering algorithms, with the objective function being:

[0093] (2);

[0094] Where K is the number of clusters, It is the center of the k-th cluster. These are data points;

[0095] S13. Calculate feature weights using the normalized correlation coefficient. Weight calculation is based on the correlation between features and SOC, determining the weight of each feature's impact on SOC. It can be calculated using the normalized correlation coefficient:

[0096] (3);

[0097] in, is the correlation coefficient between the j-th feature and SOC, and m is the total number of features;

[0098] S14. Perform dimensionality calculation. Dimensionality calculation assesses the dimensions of the interactions between different features. This can be achieved by calculating the correlation matrix between features:

[0099] (4);

[0100] in, It is the correlation coefficient between the i-th feature and the j-th feature.

[0101] Generate the feature matrix. Integrate the weight and dimension calculation results into the feature matrix. Feature matrix It can be represented as:

[0102] (5);

[0103] in, It is the feature index of the j-th feature under the i-th working condition, which can be determined by the weight and dimension calculation results.

[0104] By following the steps above, a matrix that comprehensively reflects the characteristic state can be generated, providing important input for subsequent battery SOC prediction.

[0105] S2. Encoder Optimization: The encoder suffers from two main problems in SOC prediction: first, the lack of a dynamic adjustment mechanism in the feature space leads to insufficient sensitivity to key features; second, feature fusion between layers relies solely on residual connections, failing to establish cross-layer feature associations. To address these issues, an improved scheme integrating a dynamic feature matrix and structural optimization is proposed. By introducing a learnable feature space projection operator, enhancing the residual learning mechanism, and integrating modular components, the encoder's representational ability and computational efficiency are significantly improved. Specifically, the dynamic feature matrix generated in step S1 is input into the Informer encoder, and the input feature weight allocation is optimized through a learnable feature space projection operator; a dual-path residual structure is adopted, fusing the feature outputs of the attention residual path and the convolution enhancement path; a gradient decoupling mechanism is introduced to periodically block the gradient backpropagation of the dynamic feature matrix, suppressing parameter oscillations in the early stages of training.

[0106] The encoder incorporates a dynamic feature matrix module at the input for weight optimization; its core operation can be described as follows:

[0107] (6);

[0108] In the formula, B is the batch size, L is the sequence length, and d is the feature dimension. Let be the trainable parameter matrix. This linear transformation enables the model to adaptively enhance important feature channels and suppress noise by adjusting the feature space basis. In the multi-head attention mechanism, let each head perform independent scaled dot product attention. The combination of dynamic feature matrices and query, key, and value projections can be expanded as follows:

[0109] (7);

[0110] (8);

[0111] in, These are the projection matrices corresponding to the i-th attention head. This is the head dimension. The dynamic feature matrix acts as a feature preprocessor here, enabling each attention head to compute similarity within the optimized feature subspace, thus enhancing the ability to capture cross-dimensional interaction patterns.

[0112] Secondly, at the model architecture level, the original encoder's sequential stacking of attention layers and convolutional layers suffers from a single gradient path and insufficient local context modeling. The improved method employs a dual-path residual structure in the feature processing unit, where the forward propagation process is decomposed into a fusion of the attention residual path and the convolutional enhancement path. The attention residual path achieves stable updates through layer normalization and a random deactivation mechanism.

[0113] (9);

[0114] (10);

[0115] (11);

[0116] in, , D is the feature dimension. Let h be a small value with a denominator of zero, h be the single-layer output vector, and m be the random mask vector. This represents the probability of inactivation. For the first The output of each attention head, It is a dynamic feature matrix;

[0117] The convolution enhancement path uses dilated convolution and nonlinear activation functions to capture local patterns.

[0118] (12) ;

[0119] Where Conv represents the convolution operation, and the convolution kernel is... Step length ,filling α is the activation function parameter;

[0120] The two path outputs are weighted and fused to form the final feature representation:

[0121] (13);

[0122] in For adaptive weight parameters;

[0123] To address the gradient degradation problem in deep encoders, a gradient decoupling mechanism is introduced during the encoder's forward propagation to prevent the backpropagation of gradients from the dynamic feature matrix. This is equivalent to... Apply a phased parameter freeze. The theoretical explanation for this strategy can be compared to an alternating optimization algorithm, which decomposes the objective function into two subproblems that are optimized alternately:

[0124] ;

[0125] ;

[0126] in To fix the dynamic feature matrix parameters, To optimize attention / convolution parameters, To fix the attention / convolution parameters, To optimize the dynamic feature matrix parameters, To determine the number of training iterations, The loss function is used to effectively alleviate the parameter coupling problem in complex models through the above alternating optimization.

[0127] At the engineering implementation level, the improved solution encapsulates components by building an integrated encoder module, integrating components such as dynamic feature matrices, multi-head attention, and convolutional layers into a unified interface. Its computational flow can be abstracted as follows:

[0128] (14);

[0129] ;

[0130] This modular design not only improves code reusability but also reduces memory usage by shortening data paths.

[0131] S3. Decoder Optimization: A systematic optimization of the feature fusion and cross-modal interaction mechanisms of the decoder module in sequence-to-sequence architectures is proposed, using an improved scheme based on dynamic feature matrices and multi-granularity adaptive residual learning. The original decoder suffers from limitations such as fixed feature channel weights and insufficient local context modeling when handling long-range dependencies and cross-modal alignment. The improved scheme significantly enhances the decoder's representational capabilities by introducing a differentiable feature recalibration mechanism and a hierarchical parameterized adjustment strategy.

[0132] Feature recalibration is achieved by embedding the dynamic feature matrix into the Informer model decoder; the correlation between encoder output and decoder input is established through a cross-attention mechanism; a hierarchical parameterized residual learning strategy is adopted, and hierarchical learnable health factors and channel-level factors are introduced to optimize feature fusion, thereby improving model stability. Specifically:

[0133] A dynamic feature matrix is ​​embedded during feature transfer to achieve feature recalibration. Its operational form is defined as follows:

[0134] (15);

[0135] in For decoder input, For batch size, The length of the decoded sequence. For feature dimensions;

[0136] It is a trainable parameter matrix;

[0137] It is the bias vector;

[0138] This transformation works by maximizing conditional mutual information. Feature selection is achieved, where Y is the target output. In the cross-attention mechanism, the adjustment of the dynamic feature matrix to the query vector can be formalized as:

[0139] (16);

[0140] (17);

[0141] Here, For decoder input, Encoder output, The length of the encoded sequence. Let be the projection matrix. For feature dimension, is the head dimension, and Softmax is the normalization function.

[0142] At the residual connection design level, the fixed weight addition of the original decoder limits the flexibility of feature fusion. The improved decoder introduces a hierarchical learnable health factor. With channel level factor Its update strategy:

[0143] (18);

[0144] (19);

[0145] in, Here, ⊙ represents the sigmoid function, ⊙ denotes channel-by-channel product, and N is the total number of decoder layers. For hierarchical learnable health factors, As a channel-level factor, Here, d represents the transformation function of the current layer of the decoder, and d represents the feature dimension.

[0146] when As the network approaches zero, it degenerates into an identity mapping, requiring optimization of gradient flow.

[0147] when When the value approaches 1, the transformation of the current layer is strengthened.

[0148] Combining steps S2 and S3, optimizing the encoder can significantly speed up the Informer's prediction speed, while optimizing the decoder can significantly improve its prediction accuracy. These two optimization measures are independent of each other and do not conflict.

[0149] By employing the above encoder-decoder joint optimization strategy, both the speed and accuracy of Informer prediction can be improved simultaneously.

[0150] S4. Model Training and Prediction: The optimized encoder output is fed into the decoder. After feature fusion and mapping, the predicted battery SOC value is output. The prediction performance is evaluated by the coefficient of determination R², mean square error MSE and mean absolute error MAE.

[0151] The formulas for calculating the coefficient of determination R², mean squared error (MSE), and mean absolute error (MAE) are as follows:

[0152] ;

[0153] ;

[0154] ;

[0155] in, Represents the actual observed value. This represents the model's predicted value. It is the average of the actual values.

[0156] The goodness of fit of a model is measured by its value; the closer the value is to 1, the more accurately the model fits the data. MSE and MAE are used to assess the deviation between predicted and actual values; the closer these two metrics are to 0, the higher the stability and the smaller the fluctuation in the model's predictions. Improvements in prediction accuracy are mainly achieved through… This growth is reflected in [the data / information].

[0157] Example

[0158] The simulation test platform consisted of an AMD Ryzen 5 5600H CPU with Radeon Graphics, an NVIDIA GeForce GTX1650, 16GB DDR4 RAM, a 477GB WDC HDD, and Windows 11 as the operating system.

[0159] This invention utilizes three publicly available datasets to conduct SOC prediction tests. A brief description of the datasets is as follows:

[0160] Dataset 1: This dataset, provided by the National Key Laboratory of Intelligent Vehicle Safety Technology, consists entirely of lithium battery data from real-world vehicle operation. It can be used for lithium battery health status estimation, remaining life prediction, and fault diagnosis. The data comprises 300 operational vehicles (now discontinued), with mileage ranging from 0 to 500,000 kilometers and operating cycles from 0.5 to 4 years. It covers all driving conditions, including charging, discharging, and stationary periods.

[0161] Dataset 2: The vehicle is a BAIC EU500 equipped with CATL NCM batteries. The nominal capacity is 145Ah, and the battery pack contains 90 cells in series and 32 temperature sensors. This dataset comes from the public dataset provided in reference

[28] , which contains charging data of 20 commercial electric vehicle battery packs. The data spans more than two years, approximately 29 months. It includes 9 features such as voltage, current, temperature, energy, and SOC. The running data of vehicle No. 7 and vehicle No. 20 were selected for simulation testing.

[0162] Dataset 3: 2023 Digital Car Competition—New Energy Vehicle Power Battery Health Status and Remaining Life Prediction Dataset. The data used consists of recent years' data from 13 new energy vehicles of a certain brand, equipped with different types of on-board energy storage devices: lithium iron phosphate batteries and lithium manganese oxide batteries. Data collection fields are based on the national standard GB / T32960. Vehicle models cover two typical regions, Northeast and South China, and vehicle characteristics cover typical user usage characteristics such as long-term storage followed by driving, continuous driving, and driving followed by storage.

[0163] To verify the feasibility of the optimization scheme, each selected dataset contains 30,000 data points. For dataset one, temperature was used as the variable, and two vehicle operation data points were randomly selected to test the SOC prediction performance of electric vehicle batteries under different temperature conditions. For dataset two, vehicle operation data points numbered 7 and 14 were selected for simulation testing to verify whether the optimization scheme's generalization ability for SOC prediction meets the standard when used under normal vehicle operating conditions. For dataset three, data from two different energy storage devices were selected to verify the optimization capability of the method across different on-board energy storage batteries. This method of setting and switching variables can intuitively demonstrate the feasibility of the prediction scheme.

[0164] To express prediction accuracy intuitively as a percentage of error, let the percentage of error be T. Then, the following formula can be used for calculation:

[0165] (twenty three)

[0166] In the formula, Indicates the optimized index, Indicates the state before optimization The value T obtained after calculating the index using this formula is the prediction error.

[0167] Since the raw data may include fields such as voltage, current, and temperature, and their dimensions and numerical ranges vary greatly, data standardization will be used for result analysis in the following result evaluation:

[0168] 1. Let the different features be A, B, and C. Calculate the mean of each feature. and standard deviation ;

[0169] 2. Apply the formula to each feature point.

[0170] (twenty four)

[0171] (25)

[0172] (26)

[0173] 3. Standardize Use it as input to the model and analyze the results.

[0174] In the presented prediction charts, V1 indicates the prediction result without optimization; V2 indicates the prediction result optimized only for the encoder; V3 indicates the prediction result optimized only for the decoder; V4 indicates the prediction result optimized for both the encoder and decoder; and "true" represents the actual observed value. The horizontal axis of the chart represents the data sequence number, and the vertical axis represents the predicted value after standardization. Corresponding to each prediction chart is a table recording various evaluation metrics related to that prediction result.

[0175] Figures 2 to 5 To show the prediction results using dataset 1:

[0176] Table 1. SOC prediction results at battery temperatures of -10~10℃

[0177] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0728 0.2459 0.8751 24.54 Encoder optimization 0.0654 0.2196 0.9026 11.13 Decoder optimization 0.0535 0.1655 0.9228 24.06 En-De optimization 0.0284 0.1230 0.9651 11.44

[0178] Table 2. SOC prediction results at battery temperatures of 10~20℃

[0179] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.2116 0.2934 0.8838 28.87 Encoder optimization 0.1346 0.2013 0.9305 13.18 Decoder optimization 0.1332 0.1921 0.9388 29.05 En-De optimization 0.0655 0.1508 0.9722 13.73

[0180] Table 3. SOC prediction results at battery temperatures of 20~30℃

[0181] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0167 0.0806 0.9526 31.93 Encoder optimization 0.0264 0.1425 0.9372 14.71 Decoder optimization 0.0102 0.0701 0.9756 31.81 En-De optimization 0.0075 0.0647 0.9814 15.02

[0182] Table 4. SOC prediction results at battery temperatures of 30~45℃

[0183] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0504 0.2088 0.9001 25.97 Encoder optimization 0.0160 0.1048 0.9659 11.98 Decoder optimization 0.0107 0.0701 0.9798 26.03 En-De optimization 0.0088 0.0707 0.9826 12.20

[0184] Depend on Figures 2 to 5As can be seen, after dynamic feature recalibration optimization, the predicted SOC values ​​of the battery at different temperatures show a significant improvement in the Informer model's prediction fit. Tables 1 to 4 show that both MSE and MAE indices decrease significantly. The indicators improved by 3%-12%, the prediction error decreased by 60.8%-76.1%, and the prediction time was improved by 53.1%.

[0185] Figure 6 and Figure 7 To demonstrate the prediction results using dataset 2:

[0186] Table 5. SOC Prediction Results for Vehicle 7

[0187] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0870 0.2008 0.8686 46.03 Encoder optimization 0.041 0.1090 0.9556 21.96 Decoder optimization 0.0433 0.1112 0.9562 45.61 En-De optimization 0.0302 0.0849 0.9692 21.51

[0188] Table 6. SOC Prediction Results for Vehicle 20

[0189] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0438 0.1349 0.9673 39.64 Encoder optimization 0.0330 0.1085 0.9721 18.96 Decoder optimization 0.0246 0.0761 0.9735 39.13 En-De optimization 0.0241 0.0661 0.9781 18.69

[0190] Depend on Figures 6 to 7 As can be seen, after dynamic feature recalibration optimization, the predicted SOC values ​​of the vehicle battery under different operating conditions show a significant improvement in the prediction fit of the Informer model. As shown in Tables 5 and 6, the MSE and MAE indices both decrease significantly. The indicators improved by 1.12%-11.58%, the prediction error decreased by 33%-76.6%, and the prediction time was improved by 52.85%.

[0191] Figure 8 and Figure 9 To illustrate the prediction results using dataset 3:

[0192] Table 7. SOC Prediction Results of Electric Vehicles Equipped with Lithium Iron Phosphate Batteries

[0193] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0410 0.1297 0.9613 43.39 Encoder optimization 0.0393 0.1128 0.9670 19.11 Decoder optimization 0.0308 0.0744 0.9696 40.29 En-De optimization 0.2849 0.0615 0.9746 18.34

[0194] Table 8. SOC Prediction Results of Electric Vehicles Equipped with Lithium Manganese Oxide Batteries

[0195] MSE MAE <![CDATA[R 2 ]]> Time / s Unoptimized 0.0144 0.0667 0.9702 43.25 Encoder optimization 0.0116 0.0617 0.9721 19.69 Decoder optimization 0.0129 0.0411 0.9801 43.69 En-De optimization 0.0130 0.0388 0.9891 19.92

[0196] Depend on Figure 8 and Figure 9 As can be seen, after dynamic feature recalibration optimization, the predicted SOC values ​​of electric vehicles equipped with different on-board energy storage devices show a significant improvement in the Informer model's prediction fit. Tables 7 and 8 show that both MSE and MAE indices have decreased significantly. The indicator improved by 2%, the prediction error decreased by 63.4%, and the prediction time was improved by 53.94%.

[0197] This Informer battery SOC prediction model design based on encoder-decoder bidirectional calibration can effectively enhance the prediction performance and generalization ability of the Informer model.

[0198] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0199] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An Informer battery SOC prediction method based on bidirectional calibration, characterized in that: Includes the following steps: S1. Data preprocessing: Collect key battery parameter data during the operation of electric vehicles, clean and select features to generate a dynamic feature matrix, and realize adaptive calibration of the input feature weights. S2, Encoder Optimization: Input the dynamic feature matrix generated in step S1 into the Informer encoder, optimize the input feature weight allocation through a learnable feature space projection operator; adopt a dual-path residual structure to fuse the feature outputs of the attention residual path and the convolution enhancement path; introduce a gradient decoupling mechanism to block the gradient backpropagation of the dynamic feature matrix in stages and suppress parameter oscillations in the early stage of training. S3. Decoder optimization: The dynamic feature matrix is ​​embedded in the Informer model decoder to achieve feature recalibration; the relationship between encoder output and decoder input is established through cross-attention mechanism; a hierarchical parameterized residual learning strategy is adopted; hierarchical learnable health factors and channel-level factors are introduced to optimize feature fusion and improve model stability. S4. Model Training and Prediction: The optimized encoder output is fed into the decoder. After feature fusion and mapping, the predicted battery SOC value is output. The prediction performance is evaluated by the coefficient of determination R², mean square error MSE, and mean absolute error MAE.

2. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The generation of the dynamic feature matrix in step S1 includes: S11. Calculate the Pearson correlation coefficient between the characteristics and SOC. The formula is: ； in, These are the observed values ​​of features and SOC, respectively. , respectively, are the average values ​​of the features and SOC, where n is the number of samples; S12. Use the K-means clustering algorithm to partition dimensions. The objective function is: ； Where K is the number of clusters, It is the center of the k-th cluster. These are data points; S13. Calculate feature weights using the normalized correlation coefficient. The formula is: ； in, is the correlation coefficient between the j-th feature and SOC, and m is the total number of features; S14. Calculate the correlation matrix between features. Generate dynamic feature matrix The formula is: ； ； in, It is the correlation coefficient between the i-th feature and the j-th feature. It is the feature index of the j-th feature under the i-th working condition.

3. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The dual-path residual structure of the encoder in step S2 satisfies: Attention residual path: Stable updates are achieved through layer normalization and random deactivation, as shown in the formula: ； in, , D is the feature dimension. Let h be a small value with a denominator of zero, h be the single-layer output vector, and m be the random mask vector. This represents the probability of inactivation. For the first The output of each attention head, It is a dynamic feature matrix; Convolution Enhancement Path: Dilated convolution and a non-linear activation function are used to capture local patterns. The formula is as follows: ； Where Conv represents the convolution operation, and the convolution kernel is... Step length ,filling α is the activation function parameter; Dual-output weighted fusion: ； in These are adaptive weight parameters.

4. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: In step S2, the gradient decoupling mechanism is implemented through alternating optimization, decomposing the objective function into two sub-problems: ； ； in To fix the dynamic feature matrix parameters, To optimize attention / convolution parameters, To fix the attention / convolution parameters, To optimize the dynamic feature matrix parameters, To determine the number of training iterations, This is the loss function.

5. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The feature recalibration operation of the decoder in step S3 is as follows: ； in For decoder input, For batch size, The length of the decoded sequence. For feature dimensions; It is a trainable parameter matrix; This is the bias vector.

6. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The hierarchical parameterized residual learning strategy in step S3 satisfies: Hierarchical feature update: ； Output feature calibration: ； in, Here, ⊙ represents the sigmoid function, ⊙ denotes channel-by-channel product, and N is the total number of decoder layers. For hierarchical learnable health factors, As a channel-level factor, Here, d represents the transformation function of the current layer of the decoder, and d represents the feature dimension. when As the network approaches zero, it degenerates into an identity mapping, requiring optimization of gradient flow. when When the value approaches 1, the transformation of the current layer is strengthened.

7. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The calculation method for the cross-attention mechanism in step S3 is as follows: ； ； in, For decoder input, Encoder output, The length of the encoded sequence. Let be the projection matrix. For feature dimension, is the head dimension, and Softmax is the normalization function.

8. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The key battery parameters in step S1 include at least one of the following: total voltage, total current, ambient temperature, vehicle speed, and motor temperature.

9. The Informer battery SOC prediction method based on bidirectional calibration according to claim 1, characterized in that: The formulas for calculating the coefficient of determination R², mean square error MSE, and mean absolute error MAE in step S4 are as follows: ； ； ； in, Represents the actual observed value. This represents the model's predicted value. It is the average of the actual values.

Images