Fusion of physical mechanism and cost-sensitive battery heat generation prediction method
By integrating physical mechanisms and cost-sensitive learning, a LightGBM model was constructed, which solved the problems of robustness in battery temperature prediction and accuracy in prediction under extreme conditions, and achieved high-precision, low-cost battery thermal safety monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI UNIV OF ARTS & SCI
- Filing Date
- 2026-04-01
- Publication Date
- 2026-05-29
AI Technical Summary
Existing battery temperature prediction technologies have poor robustness throughout the battery's entire life cycle. Deep learning models lack generalization ability under extreme conditions. The long-tailed distribution of real-vehicle operating data leads to insufficient reliability of prediction results for high-risk operating conditions. Traditional mechanism models rely on easily drifting parameters and cannot provide reliable thermal runaway warnings.
By integrating physical mechanisms and cost-sensitive learning strategies, a physically guided feature is constructed through the Bernardi heat generation equation. Combined with temporal statistical features and a dynamic weighted loss function, an improved LightGBM model is trained, giving high weight to extreme working conditions and optimizing the model's prediction accuracy under high-risk conditions.
It improves the accuracy and interpretability of battery temperature prediction, overcomes the peak shaving phenomenon of traditional models under high temperature conditions, enhances the sensitivity and accuracy of thermal safety early warning, reduces engineering implementation costs, and is suitable for deployment in vehicle battery management systems.
Smart Images
Figure CN122109850A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle battery technology, and more specifically, to a method for predicting battery heat generation that integrates physical mechanisms and cost sensitivity. Background Technology
[0002] In recent years, with the popularization of electric vehicles (EVs), battery thermal safety has become a core bottleneck. Battery thermal runaway often begins with the accumulation of abnormal heat. Battery management systems (BMS) need to sense battery temperature accurately in real time, but existing sensor solutions are costly and can only measure surface temperature, failing to reach the core heat source.
[0003] To compensate for the shortcomings of direct temperature measurement solutions, academia and industry have gradually shifted their research towards battery temperature prediction. However, existing technical solutions generally have multiple technical pain points and are difficult to adapt to the safety management requirements of power batteries throughout their entire life cycle and under all operating conditions.
[0004] First, traditional mechanism prediction models are built on the theory of electrochemical thermal equilibrium. For example, the classic thermal equilibrium model proposed by Lü Qiao et al. is highly dependent on key physical property parameters such as battery heat transfer coefficient, specific heat capacity, and thermal resistance. These parameters will dynamically drift with battery cycle aging and health state decay, resulting in poor robustness of the model throughout the battery's entire life cycle.
[0005] Secondly, to avoid the parameter dependence problem mentioned above in the mechanism model, mainstream technologies mostly use deep learning algorithms such as Long Short-Term Memory Network (LSTM) and Backpropagation Neural Network (BP) to build prediction models. However, such models are typical "black box" models, lacking the underlying theoretical support of electrochemical thermodynamics and battery reaction kinetics. This results in extremely poor generalization ability under unseen extreme and boundary conditions, and insufficient reliability of prediction results.
[0006] More importantly, battery operating data in real-world vehicle scenarios exhibits a typical long-tail distribution, with approximately 95% of the data representing low-load steady-state conditions. These conditions show gradual temperature changes and are relatively easy to predict. Only 5% of the data represents high-risk transient conditions such as fast charging and rapid acceleration, which are difficult to predict. Existing data-driven models generally use the mean squared error (MSE) loss function for model training. However, the model optimization process is dominated by massive amounts of steady-state data, leading to a significant "peak clipping" phenomenon in critical high-temperature and rapid temperature rise scenarios. This means that the predicted temperature is significantly lower than the actual internal temperature of the battery cell, resulting in missed thermal safety warnings. Consequently, the BMS cannot provide reliable early warning information for thermal runaway, making it difficult to ensure the thermal safety of electric vehicles under extreme operating conditions. Summary of the Invention
[0007] To address the aforementioned technical problems, this application provides a method for predicting battery heat generation that integrates physical mechanisms and cost sensitivity.
[0008] The technical solution provided in this application is as follows: A method for predicting battery heat generation that integrates physical mechanisms and cost-sensitive factors includes: Acquire prediction samples, including full-condition operating data of the battery pack collected at preset time intervals; Based on the physical heat generation mechanism, the predicted samples are reconstructed to build physical guidance features and time-series statistical features. The physical guidance features are used to characterize the physical correlation between battery heat generation and current and temperature. The time-series statistical features are used to characterize the battery's transient fluctuations, heat accumulation, and state switching. The predicted samples are divided into multiple operating condition categories based on the full operating condition data, and weights are set for the corresponding predicted samples according to the number of samples in each operating condition category, thereby constructing a dynamic weighted loss function based on the inverse frequency domain transformation of the operating condition. Using the physical guidance features, time-series statistical features, and predicted sample weights as inputs, and the measured battery temperature as a label, an improved LightGBM prediction model is trained based on the cost-sensitive loss function to obtain a battery heat generation prediction model. The real-time collected battery monitoring data is input into the battery heat generation prediction model to obtain the battery heat generation prediction result.
[0009] In one possible implementation, the full-condition operating data includes data acquisition timestamps, state of charge (SOC), vehicle speed, total odometer reading, charge / discharge status, total voltage, total current, minimum voltage of individual cells, maximum voltage of individual cells, minimum temperature value, maximum temperature value, battery voltage, and probe temperature.
[0010] One possible implementation involves reconstructing features of the predicted samples based on physical heat generation mechanisms, including: The entropy coefficient was determined using the characteristic mapping method based on the Bernardi heat production mechanism equation. The physical correlation between battery heat generation and current and temperature is determined based on the entropy-thermal coefficient, and the physical guidance features are constructed. The physical guidance features include Joule thermal features, reversible thermal features, and power features. The Joule thermal features are used to capture irreversible heat during high-current fast charging, the reversible thermal features are used to capture the entropy-thermal absorption / release effect under low current, and the power features are used to characterize the energy change rate of the battery system.
[0011] In one possible implementation, the time-series statistical features include transient fluctuation features and thermal accumulation features, constructed based on a multi-scale sliding window operator, and the method includes: Set time windows of different lengths and calculate the mean and standard deviation of the current within each window. The short window is used to calculate the standard deviation of the current to capture transient fluctuations and reflect the severity of the operating conditions. The long window is used to calculate the mean of the current to capture heat accumulation and simulate the thermal hysteresis caused by heat capacity.
[0012] In one possible implementation, the timing statistical features include state switching features and individual unit voltage standard deviation features; The state switching characteristics are obtained by performing first-order difference calculations on current and voltage, and are used to identify the switching points of the battery's operating state, including the battery switching from rest to charging and from constant current to constant voltage. The standard deviation of the individual cell voltage is determined based on the total number of cells connected in series in the battery pack, the individual cell voltage, and the average individual cell voltage, and is used to characterize the internal impedance differences of the battery pack.
[0013] One possible implementation involves setting weights for the predicted samples, including: The total number of samples in the predicted samples is counted. ; Based on the full-condition operation data, the prediction samples are divided into K There are 10 working condition categories, and the number of samples in each working condition category is counted. The operating condition categories include low-load steady-state operating condition, fast-charging operating condition, and rapid acceleration operating condition. Based on the total number of samples The number of samples to predict the working condition category of the sample and number of working condition categories K Calculate the weight coefficients of each predicted sample. Among them, the number of samples The smaller the operating condition category, the higher its corresponding weight coefficient. w i The larger.
[0014] One possible implementation involves constructing a dynamically weighted loss function based on the inverse frequency domain transform of the operating conditions, including: The weight coefficients corresponding to each predicted sample are introduced into the benchmark loss function as weighting factors. The dynamic weighted loss function is constructed by weighted summation of the sample prediction errors. The benchmark loss function is a regression loss function used to measure the deviation between the predicted temperature and the actual temperature.
[0015] In one possible implementation, the physical guidance features, temporal statistical features, and predicted sample weights are used as inputs, and the measured battery temperature is used as a label. An improved LightGBM prediction model is trained based on the cost-sensitive loss function to obtain a battery heat generation prediction model, including: The feature matrix constructed from the physical guidance features and temporal statistical features, along with the predicted sample weights, are input into the initial LightGBM model. Calculate the sample gradient of the initial LightGBM model, and sort the predicted samples according to the absolute value of the gradient; A predetermined proportion of the predicted samples are selected as the large gradient sample set, and a predetermined proportion of the predicted samples are randomly sampled from the remaining samples as the small gradient sample set. The gradient of the samples in the small gradient sample set is then corrected by multiplying the gradient by an amplification factor. Using a leaf-wise growth strategy, a decision tree is constructed based on the set of large gradient samples and the corrected set of small gradient samples. During the splitting process, the gradient calculation is corrected using the weights of the predicted samples. Iterative training is performed with the goal of minimizing the dynamic weighted loss function to obtain the battery heat generation prediction model.
[0016] In one possible implementation, after obtaining the battery heat generation prediction model, the method further includes: The contribution of each feature to the prediction result is calculated using the SHAP method; The contribution rate is used to verify whether the Joule thermal characteristics and the prediction results are linearly positively correlated, and to determine whether the reversible thermal characteristics and the prediction results are non-monotonic, in order to confirm that the battery heat generation prediction model has learned the electrochemical physical laws.
[0017] In one possible implementation, the battery monitoring data input to the battery heat generation prediction model includes sampling timestamps, total current I, total voltage U, and a list of individual cell voltages. Probe temperature T, State of charge (SOC).
[0018] Compared with the prior art, the technical solution of this application has the following beneficial effects: This application integrates the physical mechanism of power battery heat generation with a cost-sensitive learning strategy to construct the LightGBM battery heat generation prediction model, effectively solving the pain points of existing battery temperature prediction technologies, such as parameter drift in the mechanism model, lack of physical data in purely data-driven models, and peak shaving in high-risk operating condition predictions caused by long-tail data from real vehicles. Through physical-guided feature reconstruction based on the Bernardi heat generation equation, the model achieves both high accuracy and physical interpretability, avoiding the black-box defects of pure data fitting and eliminating reliance on easily drifting physical property parameters. By leveraging a dynamically weighted cost-sensitive loss function based on inverse frequency domain transformation of operating conditions, high weights are assigned to rare high-risk operating condition samples such as fast charging and rapid acceleration, amplifying their gradient optimization signals. This completely overcomes the problem of traditional loss functions being dominated by low-load steady-state data, eliminates peak shaving in high-temperature operating condition predictions, and significantly improves the sensitivity and accuracy of thermal safety warnings. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating a battery heat generation prediction method that integrates physical mechanisms and cost sensitivity, as provided in Embodiment 1 of this application.
[0020] Figure 2 This is a flowchart of the method for training a battery heat generation prediction model provided in Embodiment 1 of this application. Detailed Implementation
[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0022] Example 1 See Figure 1 This is a flowchart illustrating a battery heat generation prediction method that integrates physical mechanisms and cost sensitivity, as provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include: Step 101: Obtain prediction samples, including full-condition operation data of the battery pack collected at preset time intervals.
[0023] The aforementioned prediction samples originate from the electric vehicle BMS cloud platform and are vehicle-to-everything (V2X) data. In this embodiment, the real-vehicle monitoring data obtained from the BMS includes, but is not limited to, terminal time (i.e., data collection timestamp), state of charge (SOC), vehicle speed, total odometer, charge / discharge status, total voltage, total current, minimum voltage (MinVAT), maximum voltage (MaxVAT), minimum temperature (MinT), maximum temperature (MaxT), battery voltage list, and probe temperatures list.
[0024] The preset sampling period is set to 10 seconds / time (10Hz). Of course, the above sampling period can also be set according to the situation. This application embodiment does not make specific limitations, but the value must ensure that transient current impacts such as rapid acceleration can be captured.
[0025] Step 102: Based on the physical heat generation mechanism, the above-mentioned predicted samples are reconstructed to build physical guidance features and time-series statistical features.
[0026] Among them, the aforementioned physical guidance features are used to characterize the physical correlation between battery heat generation and current and temperature, including Joule thermal features, reversible thermal features and power features.
[0027] In this embodiment of the application, in order to solve the entropy-heat coefficient To address the difficulty of obtaining real-time data, instead of directly using raw current and voltage data, an approximate substitution is made based on the Bernardi heat generation mechanism equation using the characteristic mapping method, expressed as:
[0028] In the formula, q Indicates the total heat production rate. Indicates the operating current (charge is negative and discharge is positive). Indicates battery volume, Indicates open-circuit voltage. Indicates terminal voltage. Indicates the battery's thermodynamic temperature. It represents the entropy heat coefficient.
[0029] Based on the above formula, the relationship between heat generation and the square of current is derived. and the current-temperature product There is a strong physical correlation between them, expressed as:
[0030] In the formula, This indicates the total internal resistance of the battery, including ohmic internal resistance, polarization internal resistance, etc. This indicates the amount of heat generated.
[0031] Based on this, the aforementioned physical guidance features are constructed. Among them, the aforementioned Joule heating feature is characterized as... This is used to capture irreversible heat during high-current fast charging, dominating high-temperature operating conditions. The aforementioned reversible thermal characteristics are characterized as follows: This is used to capture the entropic heat absorption / excitation effect under low current. The aforementioned power characteristics are characterized as... , representing the rate of energy change of the battery system, reflects the overall energy exchange intensity. The aforementioned thermal inertia characteristics are characterized by the mean and standard deviation of the current, calculated through a multi-scale sliding window, and are used to reflect the hysteresis of the battery's thermal response.
[0032] The aforementioned time-series statistical features are used to characterize the battery's transient fluctuations, thermal accumulation, and state transitions. Specifically, the transient fluctuation and thermal accumulation features are constructed based on a multi-scale sliding window operator. By setting time windows of different lengths, the mean and standard deviation of the current within each window are calculated. A short window is used to calculate the current standard deviation to capture transient fluctuations and reflect the severity of operating conditions. A long window is used to calculate the current mean to capture thermal accumulation characteristics and simulate thermal hysteresis caused by heat capacity.
[0033] In this embodiment, the average current is calculated using a multi-scale sliding window (Window=5s, 20s). and standard deviation Thermal inertia characteristics are constructed to reflect the hysteresis of the battery's thermal response.
[0034] Define window size as This indicates the number of time steps contained within the window. At time , for the input sequence Sliding window operator Will Time mapping to a dimensional vector
[0035]
[0036] The vector Includes from Time's up Complete historical information of the moment.
[0037] Sliding window parameters, short window This corresponds to 5 seconds of historical data, capturing transient changes (such as rapid acceleration and sudden braking). Long window ( It corresponds to 20 seconds of historical data to capture the heat accumulation effect.
[0038] By calculating the standard deviation of the short-window current, a transient fluctuation characteristic (Std) is constructed to reflect the severity of the operating condition, expressed as:
[0039] In the formula, This represents the standard deviation of the current with a window length of 5. k Represents historical observation values. This represents the moving average.
[0040] By calculating the average current over a long window, a heat accumulation characteristic (Mean) is constructed to simulate the heat accumulation caused by heat capacity, expressed as:
[0041] In the formula, This represents the average power value with a window length of 20.
[0042] The battery state switching characteristics are obtained by performing first-order difference calculations on current and voltage, and are used to identify the switching points of the battery's operating state, including the switching points from the battery being idle to charging and from constant current to constant voltage.
[0043] In addition, this application designs an analytical algorithm to extract statistics, and specifically introduces the standard deviation of individual unit voltages ( The feature is represented as:
[0044] In the formula, This refers to the total number of battery cells connected in series in the battery pack, such as 96 or 108. For the parsed first i Individual cell voltage, This represents the average voltage of individual cells. This indicator directly reflects the inconsistencies within the battery pack. The voltage standard deviation reflects the imbalance within the battery pack; inconsistencies can lead to localized overheating, also known as the "weakest link" effect. The above characteristics of the individual cell voltage standard deviation... The larger the value, the greater the difference in internal impedance, and the higher the risk of thermal runaway.
[0045] Step 103: Based on the above full-condition operation data, divide the above prediction samples into multiple condition categories, and set weights for the corresponding prediction samples according to the number of samples in each condition category, and construct a dynamic weighted loss function based on the inverse frequency domain transformation of the condition.
[0046] Among them, this application uses statistical prediction of the total number of samples. Based on the full-condition operating data of the prediction samples, the prediction samples are divided into... K There are 10 working condition categories, and the number of samples in each working condition category is counted. The aforementioned operating conditions include, but are not limited to, low-load steady-state operating conditions and fast-charging operating conditions.
[0047] Based on the total number of samples The number of samples to predict the working condition category of the sample and number of working condition categories K Calculate the weight coefficients of each predicted sample. Among them, the number of samples The smaller the operating condition category, the higher its corresponding weight coefficient. w i The larger.
[0048] Specifically, the above sample weights w i Represented as:
[0049] It is calculated using a dynamic inverse transformation based on the frequency of occurrence of different operating conditions. For fast-charging operating conditions with extremely small sample sizes, its... Minimal, thus making the calculated weights Extremely high. In the experiments of this application, the weight of rare fast charging conditions was automatically increased to 18,000 times the baseline value.
[0050] To address the imbalanced categories in real-world vehicle data, this application redesigns the objective function of LightGBM. Traditional regression models typically use mean squared error (MSE) as the objective function, aiming to minimize the global average error. However, in cases of extreme data imbalance (long-tailed distribution), this function becomes dominated by a massive amount of steady-state samples. The regression loss function used to measure the deviation between predicted and actual temperatures is expressed as:
[0051] As an improvement, this application designs a dynamic weighted loss function based on the inverse frequency domain transform of the operating conditions to reshape the optimization objective, expressed as:
[0052] In the formula, This represents the weighted loss function. Indicates the dynamic weighting coefficient. N Represents the total number of samples. Indicates the first i The actual observed value at each moment. Indicates the first i The predicted output value of the model at each time point.
[0053] This application uses the aforementioned regression loss function as the benchmark loss function, and constructs the aforementioned dynamic weighted loss function by introducing the weight coefficients corresponding to each predicted sample as weighting factors into the benchmark loss function and by weighting and summing the sample prediction errors.
[0054] Step 104: Using the above-mentioned physical guidance features, time-series statistical features, and predicted sample weights as inputs, and the measured battery temperature as a label, train the improved LightGBM prediction model based on the above-mentioned cost-sensitive loss function to obtain the battery heat generation prediction model.
[0055] In this application, the LightGBM algorithm is selected as the model, utilizing its one-sided gradient sampling GOSS technique. GOSS retains large gradient samples (i.e., weighted fast-charging samples) and only randomly samples small gradient samples (steady-state data), perfectly matching the cost-sensitive strategy of this application. This dynamic weighting strategy plays a crucial corrective role in the gradient descent process. When calculating the residual gradient, LightGBM multiplies it by this large weighting coefficient. .
[0056] First-order gradient:
[0057] Second gradient:
[0058] This means that even a small prediction bias in the model under extreme high-temperature conditions such as fast charging will significantly amplify the gradient signal of its backpropagation, forcing the tree model to prioritize finding split points that can reduce the residuals of these high-risk samples during the splitting and growth process. This overcomes the "peak clipping" effect caused by the long-tail distribution globally and achieves accurate capture of extreme temperatures.
[0059] As a specific implementation, this application uses the LightGBM algorithm for training and leverages LightGBM's leaf-wise strategy to grow a decision tree, quickly fitting nonlinear thermal characteristics. The split gain formula based on unilateral gradient sampling GOSS maintains high efficiency when processing weighted massive amounts of data. For example... Figure 2 As shown, the specific implementation steps of the above method include: Step 1041: Input the feature matrix constructed from the above physical guidance features and temporal statistical features, along with the predicted sample weights, into the initial LightGBM model.
[0060] Step 1042: Calculate the sample gradient of the initial LightGBM model and sort the predicted samples according to the absolute value of the gradient.
[0061] Step 1043: Select a predetermined proportion of predicted samples as the large gradient sample set, randomly sample a predetermined proportion of predicted samples from the remaining samples as the small gradient sample set, and multiply the sample gradients in the small gradient sample set by an amplification factor for correction.
[0062] In this embodiment of the application, based on the absolute value of the gradient Sort the predicted samples. (Previous selection) Large gradient samples (set) From the remaining Random sampling in small gradient samples (gather For sets The sample gradient multiplied by the magnification factor To maintain consistency in data distribution.
[0063] Step 1044: Using the leaf-wise growth strategy, construct a decision tree based on the above large gradient sample set and the corrected small gradient sample set. During the splitting process, use the above predicted sample weights to correct the gradient calculation. Iteratively train the model to minimize the above dynamic weighted loss function and obtain the battery heat generation prediction model.
[0064] The core idea of Gradient Boosting Decision Tree (GBDT) is to iteratively train weak learners (decision trees) to fit the residuals of the previous model. Suppose we want to train the... tree The goal of the entire ensemble model is to minimize the regularized loss function, expressed as:
[0065] In the formula, It is a loss function (such as MSE). It's a real label. It was before The prediction results of the wheel, It is a regularization term. n Represents the total number of samples. Represents the t-th tree for sample x i The predicted value, Indicates the first Each model function.
[0066] To support various custom loss functions and improve convergence speed, LightGBM (and XGBoost) utilizes a second-order Taylor expansion to approximate the objective function. Expanding on this, we get:
[0067] In the formula, and These are the first-order gradient and the second-order gradient (Hessian) of the loss function, respectively, expressed as:
[0068] Structural Score and Optimal Leaf Weight Decision Tree Divide the input space into Each leaf node is a non-overlapping region. Corresponding to a predicted value .set up To fall into the first The sample set of leaf nodes. The regularization term is usually defined as... After removing the constant term, the objective function can be rewritten as an accumulation of leaf nodes:
[0069] In the formula, Represents the L2 regularization coefficient. This represents the complexity cost / node splitting threshold.
[0070] To minimize this objective function, this application assigns weights to each leaf node. By taking the derivative and setting it to 0, we obtain the optimal weights. , is represented as:
[0071] Substituting the optimal weights back into the objective function yields the minimum loss value for the tree structure, i.e., the structure score, which is expressed as:
[0072] This forms the mathematical basis for evaluating split quality in all GBDT variants (including LightGBM and XGBoost).
[0073] Furthermore, this application no longer uses the model as a black box, but instead employs the SHAP (Shapley Additive ex Planations) method to verify whether the model has learned the electrochemical mechanism.
[0074] Specifically, the SHAP value calculation principle is based on cooperative game theory, and features j For the sample x Contribution of predicted values Defined as:
[0075] Where F represents the set of all features, and S represents a subset of features. This represents the expected prediction of the model when only the subset S features are retained. Indicates that it does not contain features j The complete collection.
[0076] The implementation steps of TreeExplainer are as follows: Since directly calculating the above formula has a complexity of... This invention employs the TreeExplainer algorithm path integral for tree models, and uses the hierarchical characteristics of the LightGBM tree structure to recursively calculate the deviation contribution of feature j to the predicted mean at each node of the tree.
[0077] If the mechanistic verification standard is met, then the feature Dependency graph curves are presented That is, a linear positive correlation. Because Since it is a quadratic term, and the SHAP value reflects the contribution, it proves that the model has learned it. Joule's law. If the characteristic... If the dependency graph shows a "U-shaped" or non-monotonic change, it proves that the model has captured the endothermic / exothermic characteristics of reversible heat.
[0078] Step 105: Input the real-time collected battery monitoring data into the battery heat generation prediction model above to obtain the battery heat generation prediction result.
[0079] The battery monitoring data input into the battery heat generation prediction model includes, but is not limited to, sampling timestamps, total current I, total voltage U, and a list of individual cell voltages. Probe temperature T, state of charge (SOC), etc.
[0080] Compared with the prior art, the technical solution provided in Embodiment 1 of this application has the following beneficial effects: (1) This application solves the prediction failure problem caused by the "long-tail distribution" of real vehicle data. It innovatively introduces a cost-sensitive learning strategy and designs a dynamic weighted loss function based on the inverse frequency domain transformation of operating conditions. By automatically assigning extremely high penalty weights to scarce fast charging / high-temperature samples (up to 18,000 times in experiments), the model is forced to focus on extreme operating conditions. This effectively overcomes the "mean regression" trap of traditional models, significantly improves the prediction accuracy for high-risk operating conditions such as fast charging, eliminates the "peak-shaving" phenomenon, and ensures the sensitivity of safety monitoring.
[0081] (2) Integrating physical mechanisms, achieving both high precision and interpretability. This application does not rely on pure data fitting, but instead constructs a physical feature map (Joule heat factor) based on the Bernardi heat production equation. Reversible heat factor This gives the model "gray box" characteristics. Firstly, it boasts high accuracy; as shown in Table 1, experimental verification shows that the RMSE of this application is only 0.344℃, outperforming BP neural networks (typically >0.5℃) and traditional ensemble learning models. Secondly, it is interpretable; SHAP analysis proves that the model successfully learned physical laws such as "the square of the current dominates Joule heating" and "temperature difference dominates heat dissipation," rather than simple numerical correlations.
[0082] Table 1: Comparison data of different models in experiments
[0083] (3) Extremely high computational efficiency, suitable for cloud / edge deployment: The improved LightGBM algorithm (combining single-sided gradient sampling GOSS and mutually exclusive feature binding EFB) is adopted. When processing millions of real vehicle data, the training time is only about 20 seconds (a significant improvement compared to CatBoost's 358 seconds). Its millisecond-level inference speed fully meets the real-time requirements of the vehicle-cloud collaborative BMS platform.
[0084] (4) Low cost and no additional hardware required. This application can achieve high-precision prediction using only the existing voltage, current and temperature sensor data of the BMS, without the need to install expensive EIS equipment or embedded sensors, which greatly reduces the cost of engineering implementation.
[0085] (5) Thermal inertia and battery pack consistency were considered. The thermal accumulation effect (thermal inertia) of the battery was captured by multi-scale sliding window feature extraction; the "single cell voltage standard deviation" feature was introduced to effectively identify abnormal temperature rise caused by inconsistency (short board effect) within the battery pack, thereby improving the thermal safety early warning capability.
[0086] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for predicting battery heat generation that integrates physical mechanisms and cost sensitivity, characterized in that, include: Acquire prediction samples, including full-condition operating data of the battery pack collected at preset time intervals; Based on the physical heat generation mechanism, the predicted samples are reconstructed to build physical guidance features and time-series statistical features; The physical guidance feature is used to characterize the physical correlation between battery heat generation and current and temperature; The time-series statistical features are used to characterize the battery's transient fluctuations, thermal accumulation, and state switching; The predicted samples are divided into multiple operating condition categories based on the full operating condition data, and weights are set for the corresponding predicted samples according to the number of samples in each operating condition category, thereby constructing a dynamic weighted loss function based on the inverse frequency domain transformation of the operating condition. Using the physical guidance features, time-series statistical features, and predicted sample weights as inputs, and the measured battery temperature as a label, an improved LightGBM prediction model is trained based on the cost-sensitive loss function to obtain a battery heat generation prediction model. The real-time collected battery monitoring data is input into the battery heat generation prediction model to obtain the battery heat generation prediction result.
2. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, The data includes the timestamp of the collected data, state of charge (SOC), vehicle speed, total odometer reading, charge / discharge status, total voltage, total current, minimum voltage of a single cell, maximum voltage of a single cell, minimum temperature value, maximum temperature value, battery voltage, and probe temperature.
3. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, Based on the physical heat generation mechanism, the predicted samples are reconstructed using features, including: The entropy coefficient was determined using the characteristic mapping method based on the Bernardi heat production mechanism equation. The physical correlation between battery heat generation and current and temperature is determined based on the entropy-thermal coefficient, and the physical guidance features are constructed. The physical guidance features include Joule thermal features, reversible thermal features, and power features. The Joule thermal features are used to capture irreversible heat during high-current fast charging, the reversible thermal features are used to capture the entropy-thermal absorption / release effect under low current, and the power features are used to characterize the energy change rate of the battery system.
4. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, The time-series statistical features include transient fluctuation features and thermal accumulation features, which are constructed based on a multi-scale sliding window operator. The method includes: Set time windows of different lengths and calculate the mean and standard deviation of the current within each window. The short window is used to calculate the standard deviation of the current to capture transient fluctuations and reflect the severity of the operating conditions. The long window is used to calculate the mean of the current to capture heat accumulation and simulate the thermal hysteresis caused by heat capacity.
5. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, The time-series statistical features include state switching features and individual unit voltage standard deviation features; The state switching characteristics are obtained by performing first-order difference calculations on current and voltage, and are used to identify the switching points of the battery's operating state, including the battery switching from rest to charging and from constant current to constant voltage. The standard deviation of the individual cell voltage is determined based on the total number of cells connected in series in the battery pack, the individual cell voltage, and the average individual cell voltage, and is used to characterize the internal impedance differences of the battery pack.
6. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, Setting weights for the predicted samples includes: The total number of predicted samples is counted. ; Based on the full-condition operation data, the predicted samples are divided into K There are 10 working condition categories, and the number of samples in each working condition category is counted. The operating condition categories include low-load steady-state operating condition, fast-charging operating condition, and rapid acceleration operating condition. Based on the total number of samples The number of samples to predict the working condition category of the sample and number of working condition categories K Calculate the weight coefficients of each predicted sample. Among them, the number of samples The smaller the working condition category, the higher the corresponding weight coefficient. w i The larger.
7. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, Construct a dynamic weighted loss function based on the inverse frequency domain transform of operating conditions, including: The weight coefficients corresponding to each predicted sample are introduced into the benchmark loss function as weighting factors. The dynamic weighted loss function is constructed by weighted summation of the sample prediction errors. The benchmark loss function is a regression loss function used to measure the deviation between the predicted temperature and the actual temperature.
8. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, Using the physical guidance features, temporal statistical features, and predicted sample weights as inputs, and the measured battery temperature as the label, an improved LightGBM prediction model is trained based on the cost-sensitive loss function to obtain a battery heat generation prediction model, including: The feature matrix constructed from the physical guidance features and temporal statistical features, along with the predicted sample weights, are input into the initial LightGBM model. Calculate the sample gradient of the initial LightGBM model, and sort the predicted samples according to the absolute value of the gradient; A predetermined proportion of the predicted samples are selected as the large gradient sample set, and a predetermined proportion of the predicted samples are randomly sampled from the remaining samples as the small gradient sample set. The gradient of the samples in the small gradient sample set is then corrected by multiplying the gradient by an amplification factor. Using a leaf-wise growth strategy, a decision tree is constructed based on the set of large gradient samples and the corrected set of small gradient samples. During the splitting process, the gradient calculation is corrected using the weights of the predicted samples. Iterative training is performed with the goal of minimizing the dynamic weighted loss function to obtain the battery heat generation prediction model.
9. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, After obtaining the battery heat generation prediction model, the method further includes: The contribution of each feature to the prediction result is calculated using the SHAP method; The contribution rate is used to verify whether the Joule thermal characteristics and the prediction results are linearly positively correlated, and to determine whether the reversible thermal characteristics and the prediction results are non-monotonic, in order to confirm that the battery heat generation prediction model has learned the electrochemical physical laws.
10. The method for predicting battery heat generation by integrating physical mechanisms and cost sensitivity as described in claim 1, characterized in that, The battery monitoring data input to the battery heat generation prediction model includes sampling timestamps, total current I, total voltage U, and a list of individual cell voltages. Probe temperature T, State of charge (SOC).