Construction method of lithium ion battery pulse power prediction model
By constructing a dynamic confidence-weighted fusion model that combines linear and nonlinear models, the accuracy and robustness issues of lithium-ion battery pulse power prediction are solved. This enables accurate prediction and battery aging adaptation across different SOC ranges, making it suitable for new energy vehicle BMS battery management systems.
Patent Information
- Application Number
- CN202511637470.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-10
- Publication Date
- 2026-01-06
AI Technical Summary
Existing methods for predicting the pulse power of lithium-ion batteries suffer from low accuracy, poor robustness, difficulty in adapting to battery aging and changes in operating conditions, high data requirements, and inability to achieve accurate predictions across different SOC ranges.
A dynamic confidence-weighted fusion model is constructed, combining linear and nonlinear models. Prediction is performed using multiple linear regression and an adaptive ACO-BP network. Virtual samples are generated using a generative adversarial network for data preprocessing and real-time monitoring feedback, thereby optimizing model training and adjustment.
It improves the accuracy and robustness of lithium-ion battery pulse power prediction, adapts to different operating conditions and battery aging, reduces interference from data anomalies and changes in operating conditions, and supports the deployment of real-time control systems.
Smart Images
Figure CN121276348A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lithium-ion battery performance testing and prediction technology, and more specifically, to a method for constructing a lithium-ion battery pulse power prediction model. Background Technology
[0002] The pulse power of a lithium-ion battery is a key indicator of its instantaneous discharge capability, and accurate prediction of this parameter is of great significance for battery management system (BMS) design and electric vehicle range optimization. Existing technologies for pulse power analysis and prediction mainly fall into five categories: experimental measurement methods, equivalent circuit model methods, electrochemical model methods, data-driven model methods, and hybrid model methods. Data results show that experimental measurement methods are time-consuming, equivalent circuit simulation methods have low accuracy, and data-driven and hybrid models have high data requirements and poor interpretability. Furthermore, battery pulse power varies significantly across different state of charge (SOC) ranges and exhibits both linear and nonlinear characteristics. Achieving functional prediction in the application environment of lithium-ion batteries requires complex or difficult-to-implement technical aspects in parameter setting and model selection. Long-term use of lithium-ion batteries leads to aging and performance degradation, causing drift in model prediction accuracy.
[0003] Therefore, there is a need for a lithium-ion battery pulse power prediction scheme that takes into account the characteristics of battery pulse power, the application environment of lithium-ion batteries, and their own characteristics, and can balance robustness, flexibility, and accuracy. Summary of the Invention
[0004] To achieve the above objectives, this application provides a method for constructing a lithium-ion battery pulse power prediction model, comprising the following steps: Generating a lithium-ion battery pulse power prediction sample set includes: collecting feature parameters and target parameters, performing outlier processing, two-stage normalization and small sample enhancement on the feature parameters, and generating a lithium-ion battery pulse power prediction sample set. The pulse power prediction model is defined as a dynamic confidence-weighted fusion model, which includes a linear model and a nonlinear model. The linear model is used to obtain the first predicted pulse power P. * And with a first confidence level C1, the nonlinear model is used to obtain the second predicted pulse power P. ** The second confidence level C2; the first predicted pulse power P * Second predicted pulse power P ** The pulses are fused to generate the predicted pulse power P′. The dynamic confidence-weighted fusion model is trained in parallel based on the lithium-ion battery pulse power prediction sample set, and the weights of the dynamic confidence-weighted fusion model are optimized. The real-time monitoring and feedback process includes: performing N predictions using a dynamic confidence-weighted fusion model, comparing the predicted N pulse powers P′ with the actual pulse power P, and determining whether the deviation analysis results meet the prediction criteria; if the prediction criteria are not met, adjusting the anomaly detection sensitivity of the dynamic confidence-weighted fusion model and performing parallel training again; if the prediction criteria are met, the pulse power prediction model is complete; furthermore, the adjustment of the anomaly detection sensitivity includes: lowering the isolated forest threshold to more rigorously identify outliers in new data, and increasing the feature update frequency to collect more feature points for data updates.
[0005] The linear model is achieved through multiple linear regression, and is expressed as follows: P * =a×SOC+b×I+c×t+d×R+e, where P * For predicting pulse power, I is the pulse current, t is the pulse time, R is the internal resistance, and a to e are linear coefficients.
[0006] The nonlinear model uses an adaptive ACO-BP network or a shallow neural network to obtain the second predicted pulse power P. ** ; When the number of samples > the standard sample value, an adaptive ACO-BP network is used; the input layer has 8 nodes, the hidden layer uses a dynamic number of nodes, the activation function is Leaky-ReLU, and the output layer has 1 node; the ACO parameters are dynamically adjusted according to the feature variance; when the number of samples ≤ the standard sample value, a shallow neural network is used.
[0007] Furthermore, before calculating the confidence level, the coefficient of determination R² and mean absolute percentage error (MAPE) of the validation set are calculated. The first confidence level C2 and the second confidence level C2 are calculated as follows: C = k1 × R 2 +k2×(1−MAPE), where R is the resistance, MAPE is the mean absolute percentage error, k1 and k2 are confidence adjustment coefficients, and k1+k2=1; The confidence level C1 is calculated based on the coefficient of determination and mean absolute percentage error of the linear model. The confidence level calculated based on the coefficient of determination and mean absolute percentage error of the nonlinear model is the second confidence level C2; If |C1-C2|>Sc, adjust the confidence adjustment coefficient and perform confidence correction.
[0008] Furthermore, when generating the predicted pulse power P′, the weights of the dynamic confidence-weighted fusion model are used to adjust the first predicted pulse power P′. * Second predicted pulse power P **To merge; The first confidence level C1 and the second confidence level C2 correspond to the first weight W1 and the second weight W2, and are expressed as: W1=α×C1, W2=α×C2, where α is the working condition adaptation coefficient, and W1+W2=1; The first predicted pulse power P * Second predicted pulse power P ** The method for achieving fusion is as follows: P′=W1×P * +W2×P ** , where W1+W2=1.
[0009] Furthermore, before parallel training, the lithium-ion battery pulse power prediction sample set is stratified according to SOC intervals and stratified sampling is performed to ensure that the SOC interval distribution in each set is consistent after the sample set is divided into training set, validation set and test set.
[0010] The parallel training process includes stages such as dynamic feature selection, online incremental learning, cross-type adaptation, and health-related adaptation. In the feature dynamic screening stage, the cycle period is determined, and feature screening is carried out by improving the SHAP value analysis method. Features whose weights are less than a specified value for multiple consecutive times are placed in the dormant pool. During the online incremental learning phase, the cycle period is determined, and the regularization coefficient is dynamically adjusted based on the rules of similarity between the distributions of new and old data to achieve adaptive adjustment of data stability. Transfer learning is achieved by transferring model parameters during the cross-type adaptation phase, which can be accomplished by simply fine-tuning the weights of the output layer. During the health adaptation phase, if the SOH decreases by more than 10%, the internal resistance-power relationship coefficient is corrected.
[0011] The characteristic parameters include: charge state (SOC), pulse current (I), pulse time (t), and internal resistance (R), while the target parameter is the pulse power (P). Outlier handling includes overall distribution anomaly identification, local sudden anomaly identification, and outlier filling. Two-stage normalization refers to standardizing outliers using Z-Score and then scaling them to the [0,1] interval using the Min-Max method, while saving the two-stage mapping parameters. Small sample augmentation refers to using generative adversarial networks to generate virtual samples when the number of samples is ≤50, while simultaneously performing cross-combination operations on existing features to enrich the feature dimensions.
[0012] Furthermore, the cross-combination operation includes: multiplying the SOC by the pulse current to reflect the combined effect of "remaining charge - instantaneous current" on power, and dividing the internal resistance by the pulse time to reflect the coupling relationship of "internal resistance magnitude - pulse duration".
[0013] According to the method for constructing a lithium-ion battery pulse power prediction model provided by this invention, data preprocessing optimization, parallel model training, and real-time feedback adjustment are performed to improve prediction accuracy and engineering applicability. Specifically, this is reflected in the following aspects: 1) Small sample adaptability: Virtual samples are generated through GAN and the linear model is constrained by physical knowledge to solve the problem of insufficient data in the early stage of new battery testing; 2) Improved high-precision prediction: Dynamic confidence weighted fusion of linear and nonlinear models takes into account the linear / nonlinear characteristics under different operating conditions and reduces the bias of a single model; 3) Improved robustness: Layered outlier processing, real-time error monitoring, and adaptive adjustment effectively resist the interference of data anomalies and changes in operating conditions; 4) Improved engineering practicality: Parallel computing, parameter storage, and reverse conversion are supported, making it easy to deploy to real-time control systems (such as the battery management system BMS of new energy vehicles). Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating the construction steps of a lithium-ion battery pulse power prediction model according to an embodiment of the present invention. Figure 2 This is a comparison chart of test data tested using a lithium-ion battery pulse power prediction model. Figure 3 This is a diagram showing the error between simulated fitting data and actual values obtained from testing using a lithium-ion battery pulse power prediction model. Detailed Implementation
[0015] This invention provides a method for constructing a lithium-ion battery pulse power prediction model, which improves prediction accuracy and engineering applicability by optimizing data preprocessing, parallel model training, and real-time feedback adjustment.
[0016] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0017] like Figure 1 As shown, the method for constructing a lithium-ion battery pulse power prediction model provided by this invention includes the following steps: Step S100: Generate a lithium-ion battery pulse power prediction sample set; In this step, feature parameters and target parameters are collected, and the feature parameters are subjected to outlier processing, two-stage normalization and small sample enhancement to generate a lithium-ion battery pulse power prediction sample set. 1) When collecting characteristic parameters and target parameters, the characteristic parameters collected include: charge SOC, pulse current I, pulse time t, and internal resistance R. The target parameter collected is the pulse power P.
[0018] 2) The outlier handling in this step adopts a layered approach. The first layer uses the IQR method to identify overall distribution anomalies; Specifically, the calculation formula is Q1 = 25th quantile, Q3 = 75th quantile, IQR = Q3 - Q1, and outliers are determined as <(Q1 - 1.5 × IQR) or >(Q3 + 1.5 × IQR). The IQR method can quickly identify extreme anomalies in the overall distribution (such as current jumps caused by sensor failures) and improve the isolated forest by capturing local sudden anomalies (such as sudden changes in internal resistance caused by instantaneous polarization of the battery) through feature importance weights (obtained by random forest pre-training).
[0019] The second layer uses an improved isolated forest algorithm for local burst anomaly identification: the anomaly score threshold is adjusted by introducing feature importance weights (dynamic range 0.6-0.8); the anomaly score calculation formula of the improved isolated forest algorithm is: Score=∑(w_i×s_i), where w_i is the importance weight of the i-th feature (obtained through random forest pre-training), and s_i is the original anomaly score.
[0020] The above two layers of processing, through dual identification, significantly reduce the interference of outliers on the model. The third layer employs a three-level weighted strategy for outlier imputation: when the data volatility is <5%, the mean weight is 0.8 and the median weight is 0.2; when 5% ≤ volatility < 15%, the mean and median weights are each 0.5; when volatility ≥ 15%, the mean weight is 0.2 and the median weight is 0.8.
[0021] The three-level weighted strategy adjusts the weights of the mean and median according to volatility, avoiding the bias that occurs with fixed-filling methods when data fluctuations are large.
[0022] 3) Two-stage normalization refers to standardizing outliers using Z-Score and then scaling them to the [0,1] interval using the Min-Max method to save the two-stage mapping parameters; Z-Score standardization is based on the transformation of the data mean (μ) and standard deviation (σ), with the formula x′=σx−μ, which eliminates the influence of outliers on the data distribution; When using the Min-Max method, the feature parameter values are scaled to the [0,1] interval: the formula is x′′=max(x′)−min(x′)x′−min(x′), which unifies the units of each feature (e.g., SOC is a percentage, current is amperes) to avoid the model being biased towards features with large values; Two-stage normalization can preserve the two-stage mapping parameters (μ, σ, min, max) for inverse transformation during subsequent prediction (restoring the true power value from the normalization result).
[0023] 4) Small sample augmentation is a sample expansion strategy for situations with a small sample size: When the number of samples is ≤50, a generative adversarial network is used to generate virtual samples, and existing features are cross-combined to enrich the feature dimensions without the need to collect additional real data, which is highly practical in engineering.
[0024] The specific requirements for the cross-combination of features include multiplying the SOC by the pulse current to reflect the combined effect of "remaining charge - instantaneous current" on power, and dividing the internal resistance by the pulse time to reflect the coupling relationship of "internal resistance magnitude - pulse duration".
[0025] In actual battery operation, data is easily affected by sensor errors and operating condition fluctuations, resulting in problems such as many outliers and insufficient sample size. The stratified processing in this step specifically addresses these pain points.
[0026] Step S110: Construct a dynamic confidence-weighted fusion model; First, the pulse power prediction model is defined as a dynamic confidence-weighted fusion model, which includes a linear model and a nonlinear model. The linear model is used to obtain the first predicted pulse power P. * And with a first confidence level C1, the nonlinear model is used to obtain the second predicted pulse power P. ** The second confidence level C2; the first predicted pulse power P * Second predicted pulse power P ** By fusing the signals, the predicted pulse power P′ can be generated.
[0027] 1) The linear model is achieved through multiple linear regression, expressed as: P * =a×SOC+b×I+c×t+d×R+e, where P * For predicting pulse power, I is the pulse current, t is the pulse time, R is the internal resistance, and a to e are linear coefficients; When the sample size is ≤50, lithium-ion battery physics knowledge and empirical formulas are incorporated into the linear model as constraints, such as instantaneous power P=U×I, Ohm's law, electrochemical reaction kinetic formulas, etc.
[0028] 2) The nonlinear model uses an adaptive ACO-BP network or a shallow neural network to obtain the second predicted pulse power P. ** ; When the number of samples exceeds the standard sample value (e.g., the standard sample value is set to 50), an adaptive ACO-BP network is used, with 8 nodes in the input layer, a dynamic number of nodes in the hidden layer (6-10 nodes, adaptively adjusted according to the feature dimension), and the activation function is Leaky-ReLU (to solve the dead neuron problem). The output layer has 1 node. The ACO parameters are dynamically adjusted according to the feature variance (40-50 ants when the variance is >0.2, otherwise 20-30 ants). This avoids computational redundancy caused by too many ants when the variance is small, while ensuring search accuracy when the variance is large. The ACO-BP model supports parallel computation and parameter storage. The training strategy is as follows: the initial learning rate is 0.01, the Adam optimizer is used (momentum coefficient 0.9, second moment coefficient 0.999), the MAPE is verified every 5 rounds, and if it does not decrease for 2 consecutive rounds, the learning rate is multiplied by 0.4. Training stops when MAPE < 4% or the learning rate < 1e-6. Through this training strategy, both training efficiency and prediction accuracy can be balanced, which meets the requirements of "fast convergence + controllable error" in engineering.
[0029] When the number of samples is less than or equal to the standard sample value (e.g., the standard sample value is set to 50), a shallow neural network is used.
[0030] Since battery power characteristics combine linear (such as low-pulse scenarios dominated by Ohm's law) and nonlinear (such as high-pulse scenarios dominated by electrochemical reaction kinetics), this invention adopts parallel training of "linear + nonlinear" dual models. It not only covers the basic linear relationship by improving multiple linear regression, but also captures complex nonlinear features through an adaptive ACO-BP network. The technical path is highly consistent with the physical characteristics of batteries.
[0031] Before calculating the confidence score, calculate the coefficient of determination R² (reflecting the goodness of fit of the model, the closer to 1 the better) and the mean absolute percentage error (MAPE) of the validation set (reflecting the prediction accuracy, the smaller the better). Then, calculate the confidence score using the formula, expressed as: C=k1×R 2 +k2×(1−MAPE), where MAPE is the mean absolute percentage error, k1 and k2 are confidence adjustment coefficients, and k1+k2=1; The confidence level calculated based on the mean absolute percentage error of the linear model is the first confidence level C1; The confidence level calculated based on the mean absolute percentage error of the nonlinear model is the second confidence level C2; Determine the standard value Sc for the confidence error (e.g., set Sc=0.3) to judge whether the confidence level setting is reasonable. If |C1-C2|>Sc, then adjust the confidence level adjustment coefficient to correct the confidence level.
[0032] Then, the dynamic confidence weighted fusion model is trained in parallel based on the lithium-ion battery pulse power prediction sample set, and the weights of the dynamic confidence weighted fusion model are optimized. Before parallel training, the lithium-ion battery pulse power prediction sample set is stratified according to SOC intervals (20%–50%, 50%–80%, 80%–100%) and stratified sampling is performed to ensure that the SOC interval distribution in each set is consistent after the sample set is divided into training set, validation set and test set.
[0033] In practical applications, battery pulse power is affected by multiple parameters such as SOC, current, and internal resistance, and the characteristics of different SOC ranges (20%–50%, 50%–80%, 80%–100%) differ significantly—power in the low SOC range is limited by the risk of lithium plating, while power in the high SOC range is limited by polarization effects. In this step, data is divided by stratified sampling (stratified by SOC range) to ensure the consistency of the training set distribution with the actual operating conditions and to avoid overfitting the model in a single SOC range.
[0034] When generating the predicted pulse power P′, the first predicted pulse power P′ is calculated using the weights of the dynamic confidence-weighted fusion model. * Second predicted pulse power P ** The fusion is performed; the first confidence level C1 and the second confidence level C2 correspond to the first weight W1 and the second weight W2, which are expressed as: W1=α×C1, W2=α×C2, where α is the working condition adaptation coefficient, and W1+W2=1; The first predicted pulse power P * Second predicted pulse power P ** The method for achieving fusion is as follows: P′=W1×P * +W2×P ** , where W1+W2=1.
[0035] The parallel training process includes stages such as dynamic feature selection, online incremental learning, cross-type adaptation, and health adaptation. In the dynamic feature selection stage, a cycle period (e.g., 30 cycles) is determined, and features are selected using an improved SHAP value analysis method, placing features with weights consistently below a specified value into a dormant pool. In the online incremental learning stage, a cycle period (e.g., 50 cycles) is determined, and the regularization coefficient is dynamically adjusted based on the similarity rules of new / old data distributions to achieve adaptive adjustment of data stability. In the cross-type adaptation stage (e.g., applied to ternary lithium / lithium iron phosphate), model parameter transfer is performed to achieve transfer learning; at this point, only fine-tuning the weights of the output layer is required. In the health adaptation stage, if the state of harmonics (SOH) decreases by more than 10%, the internal resistance-power relationship coefficient is corrected.
[0036] The parallel training process also optimizes the training strategy for small samples, including: increasing the L1 or L2 regularization coefficient, adjusting the training parameters to reduce the learning rate or increase the number of training rounds, and training multiple simple models with different initialization parameters, averaging or voting on the prediction results to achieve ensemble learning.
[0037] After the above training process, the dynamic confidence weighted fusion model has been initially constructed.
[0038] Step S120: Perform real-time monitoring feedback; First, error monitoring is performed, which involves making N predictions using a dynamic confidence weighted fusion model, comparing the predicted pulse power P′ with the actual pulse power P, and determining whether the deviation analysis results meet the prediction standard. If the prediction criteria are not met, the anomaly detection sensitivity of the dynamic confidence weighted fusion model is adjusted, and parallel training is performed again; if the prediction criteria are met, the pulse power prediction model is completed.
[0039] For example, calculate the deviation between the actual power and the predicted value 10 times, and trigger an alarm when the deviation is >8% for 3 consecutive times or >15% for a single time; adjust the anomaly detection sensitivity of the dynamic confidence weighted fusion model.
[0040] Adjustments to anomaly detection sensitivity include: lowering the isolated forest threshold and increasing the feature update frequency. For example, lowering the isolated forest threshold by 0.1 (from 0.7 to 0.6) can more rigorously identify outliers in new data. At the same time, increasing the feature update frequency allows for more feature points to be collected to update the data.
[0041] Because batteries experience aging and performance degradation over long-term use, model prediction accuracy can drift. Real-time monitoring and feedback effectively address this issue. Error monitoring triggers an alert when "three consecutive deviations > 8% or a single deviation > 15%" to promptly detect model drift or sudden anomalies. The strategy for adjusting anomaly detection sensitivity after an alert improves anomaly detection sensitivity and updates feature data without manual intervention, ensuring the model maintains high accuracy even when battery performance changes.
[0042] This closed-loop design of "monitoring-early warning-adjustment" avoids the accuracy decline problem caused by the traditional model's "one-time training and long-term use", ensuring the reliability of the solution throughout the battery's entire life cycle.
[0043] This invention provides specific embodiments, taking LCO / / HC pouch batteries as an example. Through testing, their state of charge (SOC), pulse current I, pulse time t, internal resistance R, and pulse power P can be determined. A total of 48 sets of training data were obtained, such as... Figure 2As shown, each vertical line represents a set of sample data. Due to different pulse times, the line width varies for each set of data. The data was organized into a two-dimensional series of charge SOC (%), pulse current I (A), pulse time t (ms), internal resistance R (mΩ), and pulse power P (W) for training. By comparing with the test set data, the mean absolute percentage error (MAPE) of the model was calculated to be only 1.79% (e.g., ...). Figure 3 As shown in the figure, this indicates that the prediction model can predict the pulse performance of the battery relatively accurately.
[0044] The method for constructing a lithium-ion battery pulse power prediction model provided by this invention can improve prediction accuracy and engineering applicability by optimizing data preprocessing, parallel model training, and real-time feedback adjustment. Specifically, it is reflected in the following aspects: 1) Small sample adaptability: By generating virtual samples through GAN and constraining linear models with physical knowledge, the problem of insufficient data in the early stage of new battery testing is solved; 2) High-precision prediction: Dynamic confidence weighted fusion of linear and nonlinear models takes into account the linear / nonlinear characteristics under different operating conditions and reduces the bias of a single model; 3) Strong robustness: Layered outlier processing, real-time error monitoring, and adaptive adjustment effectively resist the interference of data anomalies and changes in operating conditions; 4) Engineering applicability: Supports parallel computing, parameter storage, and reverse conversion, which is convenient for deployment in real-time control systems (such as the battery management system of new energy vehicles).
[0045] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A method for constructing a lithium-ion battery pulse power prediction model, characterized in that, The method comprises the following steps: The method comprises the following steps: The structure of the pulse power prediction model is defined as a dynamic confidence weighted fusion model, the dynamic confidence weighted fusion model includes a linear model and a nonlinear model, wherein the linear model is used to obtain a first predicted pulse power P * and a first confidence C1, the nonlinear model is used to obtain a second predicted pulse power P ** and a second confidence C2; the first predicted pulse power P * and the second predicted pulse power P ** are fused to generate a predicted pulse power P′; According to the lithium ion battery pulse power prediction sample set, the dynamic confidence weighted fusion model is trained in parallel, and the weight of the dynamic confidence weighted fusion model is optimized; Real-time monitoring feedback is performed, including: N times of prediction by the dynamic confidence weighted fusion model, deviation comparison between the predicted N predicted pulse powers P' and the actual pulse power P, judgment whether the deviation analysis result meets the prediction standard; if the prediction standard is not met, the abnormal detection sensitivity of the dynamic confidence weighted fusion model is adjusted, and parallel training is performed again; if the prediction standard is met, the pulse power prediction model is constructed.
2. The construction method of claim 1, wherein, The linear model is realized by multiple linear regression, and is expressed as: P * = a x SOC + b x I + c x t + d x R + e, where P * is the predicted pulse power, I is the pulse current, t is the pulse time, R is the internal resistance, and a through e are linear coefficients.
3. The construction method of claim 1, wherein, The nonlinear model adopts an adaptive ACO-BP network or a shallow neural network to obtain a second predicted pulse power P ** ; When the number of samples is greater than the sample value standard, an adaptive ACO-BP network is used; wherein the input layer is set to 8 nodes, the hidden layer uses dynamic node number, the activation function is Leaky-ReLU, and the output layer uses 1 node; ACO parameters are dynamically adjusted according to feature variance; when the number of samples is less than or equal to the sample value standard, a shallow neural network is used.
4. The construction method of claim 1, wherein, Before calculating the confidence, the determination coefficient R2 and the mean absolute percentage error MAPE of the validation set are calculated; The calculation method of the first confidence C2 and the second confidence C2 is: C2=k1×R 2 +k2×(1−MAPE), wherein R is resistance, MAPE is mean absolute percentage error, k1 and k2 are confidence adjustment coefficients, and k1+k2=1. The confidence calculated according to the determination coefficient and the mean absolute percentage error of the linear model is the first confidence C1; The confidence calculated according to the determination coefficient and the mean absolute percentage error of the nonlinear model is the second confidence C2; If |C1-C2|>Sc, adjust the confidence adjustment coefficient and correct the confidence.
5. The construction method of claim 1, wherein, The first predicted pulse power P * and the second predicted pulse power P ** are fused by dynamically confidence-weighted fusing the weights of the models. The first confidence C1 and the second confidence C2 correspond to the first weight W1 and the second weight W2, and are expressed as: W1=α×C1, W2=α×C2, wherein α is a working condition adaptation coefficient, and W1+W2=1; The first predicted pulse power P * and the second predicted pulse power P ** The implementation method of fusion is: P' = W1 x P * + W2 x P ** where W1 + W2 = 1.
6. The construction method of claim 1, wherein, Before the parallel training, the lithium ion battery pulse power prediction sample set is divided into layers according to the SOC interval and is sampled in layers, so that the SOC interval distribution in each set is consistent after the sample set is divided into a training set, a validation set and a test set.
7. The construction method of claim 1, wherein, The parallel training process comprises the following stages: feature dynamic screening, online incremental learning, cross-type adaptation, health degree adaptation and the like. In the feature dynamic screening stage, the cycle period is determined, feature screening is performed by improving the SHAP value analysis, and features with continuous multiple weights less than a specified value are put into a dormant pool; In the online incremental learning stage, the cycle period is determined, the regularization coefficient is dynamically adjusted based on the similarity of new / old data distribution to realize adaptive adjustment of data stability; In the cross-type adaptation stage, model parameter migration is performed to realize transfer learning, at this time, only the weight of the output layer is fine-tuned to realize it. In the health degree adaptation stage, if the SOH decreases by more than 10%, the internal resistance and power relationship coefficient is corrected.
8. The construction method of claim 1, wherein, The feature parameters include state of charge SOC, pulse current I, pulse time t and internal resistance R, and the target parameter is pulse power P. The abnormal value processing includes overall distribution anomaly identification, local burst anomaly and abnormal value filling. The two-stage normalization refers to Z-Score standardization processing of outlier influence, and then Min-Max method is used to scale to the interval [0, 1], and the two-stage mapping parameters are saved. The small sample enhancement refers to that when the sample quantity is less than or equal to 50, a generative adversarial network is used to generate virtual samples, and existing features are subjected to cross combination operation to enrich the feature dimension.
9. The construction method of claim 8, wherein, The cross combination operation includes: The SOC is multiplied by the pulse current to reflect the influence of "residual power-instantaneous current" on power, and the internal resistance is divided by the pulse time to reflect the coupling relationship of "internal resistance size-pulse duration".
10. The construction method of claim 1, wherein, The adjustment of the abnormality detection sensitivity includes reducing the isolation forest threshold and increasing the feature update frequency setting.