Battery temperature prediction method based on multi-model fusion

Abstract

The invention belongs to the field of data science and machine learning, and constructs a prediction system fusing a convolutional neural network (CNN), a long-short-term memory network (LSTM), a seasonal autoregressive integral moving average model (SARIMAX) and a data-driven modeling method in order to solve the problem of battery temperature prediction. The system focuses on analyzing transient temperature change during fast charging of the battery and a temperature evolution rule caused by aging, an exclusive data feature extraction and fusion strategy is formulated for different working conditions, and model parameters are optimized. Key performance indexes such as prediction accuracy, model robustness and working condition adaptability are remarkably improved, and the method has a wide application prospect and an important practical value in thermal management of the battery of the electric vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management, and in particular to a battery temperature prediction method based on multi-model fusion. Background Art

[0002] Time series prediction is of great significance in many fields. In the field of battery management, accurate battery temperature prediction is crucial for ensuring battery safety, improving performance, and extending service life. As battery application scenarios become increasingly complex, the requirements for temperature prediction accuracy are also increasing.

[0003] Traditional statistical models have numerous limitations when processing battery temperature data. Models such as ARMA and ARIMA are only applicable to stationary or difference-stationary time series. However, battery temperature data is large-scale, high-dimensional, and nonlinear. Using such models not only requires tedious stationarity testing and parameter selection, but also limits forecasting accuracy.

[0004] Although machine learning and deep learning models have been introduced to the field of battery temperature prediction, they also face challenges. Long short-term memory networks (LSTMs) can capture long-term dependencies, but they are slow when trained on large-scale battery data and are sensitive to hyperparameters. Convolutional neural networks (CNNs) can extract local features but struggle to capture overall trends and long-term patterns in battery temperature. While temporal convolutional networks have improved upon CNNs, a single model still cannot meet the complex requirements of battery temperature prediction. Faced with the significant seasonal variations in battery data, the prediction accuracy of existing single models is significantly reduced.

[0005] Hybrid models are gradually emerging in the field of battery temperature prediction. Combining the advantages of LSTM and CNN can improve prediction performance, and weighted prediction strategies are also widely adopted. How to efficiently build and optimize hybrid models, as well as reasonably select weighting strategies, remain the focus and difficulty of current research. Ordinary ambient temperature prediction methods do not consider the impact of special conditions such as battery fast charging and aging on temperature changes, which is essentially different from battery temperature prediction. At present, related research still has a lot of room for exploration in data processing, model fusion and prediction accuracy improvement. The present invention aims to fill these gaps and provide a more effective battery temperature prediction solution. Summary of the Invention

[0006] Aiming at the problem of low battery temperature prediction accuracy in the prior art, the present invention discloses a battery temperature prediction method based on multi-model fusion, which is used to predict the temperature of the power battery of an electric vehicle.

[0007] The present invention is achieved through the following technical solutions:

[0008] The present invention provides a battery temperature prediction method based on multi-model fusion, including

[0009] Step 1: Obtain relevant data for predicting battery temperature, perform preprocessing, and construct a data set;

[0010] Step 2: Build a CNN-LSTM model, a SARIMAX model, and a data-driven model, and train the models using the battery temperature dataset.

[0011] Step 3: Combine the test results of the CNN-LSTM model, SARIMAX model, and data-driven model to build a temperature prediction model, and use the data set built in step 1 to process the temperature prediction model;

[0012] Step 4: Optimize the weights of step 3 to optimize the temperature prediction model.

[0013] As a further solution, the step 1 specifically includes:

[0014] Obtain temperature data of the power battery of an open-source electric vehicle under multiple different driving cycles;

[0015] Collect battery status data, environmental data, and vehicle driving data from multiple monitoring points within the battery pack;

[0016] Calculate the battery internal resistance at different SOC and SOH, and record the heat generation rate at different charge and discharge rates under different internal resistance;

[0017] Perform missing value processing, smoothing and denoising on the above data;

[0018] The processed data is divided into training set, validation set and test set, and the battery fast charging data and battery aging data are marked.

[0019] As a further solution, the CNN-LSTM model contains two convolutional layers, a pooling layer, an LSTM layer and a fully connected layer, wherein:

[0020] The first convolution layer uses 64 filters of size 3, and the activation function adopts ReLU. A batch normalization layer and a one-dimensional maximum pooling layer are added. The second convolution layer uses 32 filters of size 3, and the activation function adopts ReLU. A batch normalization layer and a one-dimensional maximum pooling layer are added. The LSTM layer uses 100 units to capture long-term dependencies and adds a batch normalization layer.

[0021] As a further solution, the LSTM layer uses the SOH attenuation gradient as the gate input and introduces the cell state into the gate connection. The formula is

[0022] i t =σ(W xi x t+W ci c t-1 +W hi h t-1 +b i )

[0023] f t =σ(W xf x t +W ci c t-1 +W hf h t-1 +b f )

[0024] o t =σ(W xo x t +W ci c t-1 +W ho h t-1 +b o )

[0025] σ is the sigmoid activation function, W is the weight matrix, b is the bias vector, h t-1 is the hidden state of the previous time step t-1, C t-1 is the cell state at the previous time step t-1.

[0026] As a further solution, the Adam optimizer and mean square error loss function are used to adjust the CNN-LSTM model training parameters.

[0027] As a further solution, the formula of the SARIMAX model is:

[0028] φ(B)(1-B S )Y t =θ(B)(1-θ(B s ))ε t +βX t ,

[0029] The formula of the SARIMAX model is:

[0030] φ(B)(1-B S )Y t =θB)(1-θ(B s ))ε t +βX t ,

[0031] Where φ(B) and θ(B) are the autoregressive and moving average polynomials, respectively, S is the seasonal cycle, and X t is an exogenous variable; 1-Θ(B s ) indicates that the seasonal trend is eliminated by differentiating the time series, Yt represents the observed value of the time series at time t, B is the lag factor, ε t is the white noise error at time t, β is the exogenous variable X t The coefficient of

[0032] The SARIMAX model is trained using the training set data, and the parameters are determined by Bayesian optimization, including:

[0033] Seasonal autoregressive order (P): represents the order of φ(B);

[0034] Seasonal difference order (D): indicates the seasonal difference of time series data to eliminate seasonal trends;

[0035] Seasonal moving average order (Q): determines 1-Θ(B s ) order;

[0036] and seasonal cycle (S).

[0037] As a further solution, the data-driven model calculates the battery internal resistance at different SOCs based on the VI method, determines the heat generation rate according to the current internal resistance and operating conditions, and obtains the heating temperature.

[0038] As a further solution, in step 4, optimizing the weights includes:

[0039] The dynamic weight coefficient formula is defined as

[0040]

[0041] Among them, i represents the model number, w i Represents the weight corresponding to the model, C rate is the charge and discharge rate, SOH is the battery health status, is the temperature change rate;

[0042] The grid search method is used to traverse different weight combinations and determine the optimal weight with the goal of minimizing the prediction error.

[0043] As a further solution, in step 4, optimizing the weights further includes:

[0044] Under the condition of C-rate>2C, within the normal rate range of the battery, the weights of the CNN-LSTM model, SARIMAX model, and data-driven model are set to 0.65, 0.25, and 0.1 respectively;

[0045] Under the condition of SOH < 80%, the weights of the CNN-LSTM model, SARIMAX model, and data-driven model are set to 0.6, 0.2, and 0.2, respectively.

[0046] As a further solution, the temperature prediction model is evaluated by calculating the mean square error, mean absolute error and root mean square error.

[0047] The characteristics and beneficial effects of the present invention are:

[0048] (1) In terms of accuracy, the present invention effectively captures the characteristics of battery temperature data through a unique hybrid model architecture and a precise weighted prediction strategy. CNN is used to accurately extract the local temperature change characteristics of the battery at different time scales during the charging and discharging process, such as the subtle fluctuations in the temperature of each monitoring point inside the battery module. LSTM, with its gating mechanism, effectively captures the dependency of temperature changes during long-term battery use, fully considering the long-term impact of battery aging on temperature changes. The SARIMAX model models the seasonal impact of ambient temperature on battery temperature changes and the battery charge and discharge cycle characteristics, comprehensively combining the battery's own characteristics to significantly improve the prediction performance. The weighted prediction strategy comprehensively considers multiple factors of the battery and dynamically adjusts the weights. Experiments show that the prediction error is significantly reduced under different working conditions. Compared with the mean square error of 2.68 for CNN-LSTM, the mean square error of the hybrid model is reduced to 0.16, the mean absolute error of CNN-LSTM is 1.28, and the mean absolute error of the hybrid model is reduced to 0.32, providing an accurate decision-making basis for battery thermal management, ensuring stable battery performance and extending battery service life.

[0049] (2) In terms of robustness, the multi-model fusion method of the present invention provides a high tolerance for data noise and outliers. In the face of data loss, abnormal fluctuations, and complex operating conditions, the various models work together to ensure reliable predictions. When 15% of the simulated sensor failure data is missing, the prediction mean square error drops from 9.3 of the traditional model to 4.9, a 47% reduction in mean square error. This significantly reduces the safety risks caused by data anomalies and ensures safe and stable battery operation.

[0050] (3) In terms of flexibility, users can flexibly adjust the weights based on the battery usage scenario and aging degree. For example, when an electric vehicle is fast-charging at high speed, the weight of the hybrid CNN-LSTM model can be increased. When the battery is aging and charging and discharging for a long time, the seasonal autoregressive integrated moving average model (SARIMAX) and LSTM weight can be increased. Data simulation shows that after adjusting the weights, the prediction error changes little, and the mean square error decreases from 5.3 to 4.6, a 13% reduction in the mean square error. This allows the model to quickly adapt to new operating conditions and meet different battery management requirements, thereby improving the practicality and adaptability of the model.

[0051] (4) The present invention also focuses on special situations such as fast charging and aging of batteries, and applies them to the battery thermal management system. By accurately predicting temperature changes, starting heat dissipation or heating measures in advance, and optimizing charging and discharging strategies, the battery management level is improved and some problems existing in battery management are solved. It has unique advantages in the field of battery temperature prediction and has effectively promoted the development of the battery industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the process of the method in the embodiment of the present invention;

[0053] Figure 2 Schematic diagram showing the comparison between the predicted temperature and the actual temperature in an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of MSE comparison of different models in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0056] In one embodiment, Figure 1 As shown, the present invention provides a battery temperature prediction method based on multi-model fusion, including

[0057] Step 1: Obtain relevant data for predicting battery temperature, preprocess the data, and construct a data set;

[0058] In an embodiment of the present invention, the data related to predicting battery temperature mainly includes battery status data. The battery status data mainly includes SOC (capacity), SOH (health, SOH = actual battery capacity / factory capacity), SOF (functional state). The core parameters that directly affect the battery status data include battery voltage, battery current, battery temperature, battery internal resistance, number of charge and discharge cycles, cumulative charge amount / discharge amount, cumulative operating time / mileage, number of deep discharges / depth, remaining cruising range, full charge cruising range, charging state, and energy flow state. The data of the battery temperature itself includes the temperature value itself, including the temperature of the single cell, the module temperature, and the overall system temperature. Then there is the temperature change rate, which is an indicator that can reflect the severity of the battery's working state. It also includes temperature extremes, especially the highest temperature, because overheating may cause thermal runaway, and battery temperature difference data. Excessive temperature difference will lead to uneven performance of the battery pack.

[0059] In an embodiment of the present invention, the data related to predicting the battery temperature also includes exogenous variable data such as ambient temperature, humidity, and vehicle driving status (such as vehicle speed).

[0060] Step 2: Build a CNN-LSTM model, a SARIMAX model, and a data-driven model, and train the models using the battery temperature dataset.

[0061] Step 3: Combine the test results of the CNN-LSTM model, SARIMAX model, and data-driven model to build a temperature prediction model, and use the data set built in step 1 to process the temperature prediction model;

[0062] Step 4: Optimize the weights of step 3 to optimize the temperature prediction model.

[0063] In the above embodiment, constructing the data set includes reading the battery temperature data file, identifying outliers, and using interpolation to cover the outliers to ensure that the data is complete and smooth. The feature and target variables are scaled using a standardization method, and the formula is: Where x is the original data, μ is the mean, and σ is the standard deviation. Generate a window of time series data according to the battery charge and discharge cycle and temperature change law, and construct a data set suitable for deep learning model training. For the fast charging data of the battery, the data window is divided separately, and the window step size is appropriately reduced. In the present invention, the fast charging data only considers the impact of temperature on fast charging; for the battery aging data, including battery capacity attenuation, internal resistance change, number of charge and discharge cycles, etc., the data set is divided according to the degree of aging, and different time series windows are generated respectively to provide more accurate data support for subsequent model training. At the same time, the internal resistance, heat generation rate and other data obtained from the test in the data-driven model are preprocessed and feature integrated together with other battery data. In the data preprocessing process, the data is also smoothed to remove noise interference and improve the stability of the data.

[0064] In the above embodiment, the CNN-LSTM model includes multiple layers of convolutional layers, pooling layers, LSTM layers and fully connected layers. Based on the battery module structure and the distribution of temperature monitoring points, the first convolutional layer uses 64 filters of size 3, and the activation function adopts ReLU. A batch normalization layer and a one-dimensional maximum pooling layer are added to effectively extract the local temperature characteristics of the battery. In the present invention, the local temperature characteristics of the battery are defined as the change law of the local temperature data of the battery. The second convolutional layer uses 32 filters, and a batch normalization layer and a one-dimensional maximum pooling layer are also added. Considering the length and complexity of the historical temperature data of the battery, the LSTM layer uses 100 units to capture long-term dependencies, and a batch normalization layer is added. The data is flattened by the Flatten layer, and after passing through a fully connected layer of 64 neurons and a random inactivation layer, the output layer uses a single neuron to predict the battery temperature.

[0065] During the fast charging phase, CNN uses dilated convolution technology to expand the temporal receptive field, covering 15 seconds of historical data. The convolution operation formula in one dimension is: (x is the input sequence, w is the convolution kernel weight, y is the output, d is the step size parameter, which is used to determine the interval for selecting elements from the input sequence and is set to 100. j is the loop index in the summation operation, which is used to traverse the weight vector w and determine the position of the elements involved in the calculation in the input x. By inserting intervals between convolution kernel elements to increase the receptive field, combined with the spatial attention mechanism to focus on temperature mutation modules such as the center of the battery cell, and collecting test data from the difficult heat dissipation area in the middle of the battery cell, the error of local temperature feature extraction is reduced by 32%. This allows the model to keenly capture the local rapid changes in battery temperature during the fast charging stage, such as temperature change trends, providing strong support for prediction.

[0066] For aging batteries, the LSTM layer uses the SOH attenuation gradient as the gate input and introduces the cell state into the gate connection. The improved formula is:

[0067] i t =σ(W xi x t +W ci c t-1 +W hi h t-1 +b i )

[0068] f t =σ(W xf x t +W ci c t-1 +W hf h t-1 +b f )

[0069] o t =σ(W xo x t +W ci c t-1 +W ho h t-1 +b o )

[0070] Among them, σ is the sigmoid activation function, W is the weight matrix, b is the bias vector, h t-1 It is the hidden state of the previous time step t-1, responsible for short-term information transmission and output calculation. t-1The cell state at the previous time step t-1 is used to preserve and transmit long-term information. Simulation results show that this design can suppress the long-term dependency loss problem caused by aging, reducing the prediction mean square error from 8.25 to 5.8, a decrease of 29%. This helps the model better capture the long-term trend of temperature changes in aging batteries and improves the accuracy of temperature prediction for aging batteries. In the present invention, the long-term trend of battery temperature changes is defined as the long-term variation pattern of battery temperature data.

[0071] During the training process, the training parameters are adjusted according to the fast charging and aging conditions of the battery. For training sets with a relatively large proportion of fast charging data, the learning rate is appropriately increased to speed up the model's learning of the fast charging temperature change characteristics. For aging battery data, the number of training rounds is increased so that the model can better learn the aging characteristics. The Adam optimizer and the mean square error loss function are used to adjust the model training parameters. In addition, the present invention also introduces battery health status and the like as auxiliary training data to further improve the accuracy of the model's prediction of battery temperature. During the training process, the early stopping method can be used to prevent the model from overfitting. By monitoring the loss value on the validation set, the training is stopped when the loss value no longer decreases within a certain number of rounds. Through targeted structural design and training parameter adjustment, the CNN-LSTM model can effectively learn the local and long-term characteristics of battery temperature data and improve prediction accuracy.

[0072] In the above embodiment, a SARIMAX model is trained using the battery's historical temperature data and its related features, such as exogenous variables such as ambient temperature, humidity, and vehicle driving status. The seasonal cycle is determined based on the temperature variation cycle of the battery's operating environment, such as the temperature variation within a day and the temperature difference between different seasons. The SARIMAX model formula is: φ(B)(1-B S )Y t =θ(B)(1-θ(B s ))δ t +βX t , where φ(B) and θ(B) are autoregressive and moving average polynomials, respectively. φ(B) describes the linear relationship between the current value and the past value of the sequence. 1-Θ(B s ) indicates that the seasonal trend is eliminated by differentiating the time series, Y t represents the observed value of the time series at time t, B is the shift factor, that is, the time series is shifted back k time steps, here represents the observed value of the battery temperature at time t, s is the seasonal cycle, X t is an exogenous variable, ε t is the white noise error at time t, representing unexplained random fluctuations, and β is the exogenous variable X t The coefficient of the exogenous variable X t For the dependent variable Y tAt the same time, the relevant information in the data-driven model (such as the trend of internal resistance change under different aging degrees) is used as a supplement to the exogenous variables to optimize the model training. When determining the parameters of the SARIMAX model, the cross-validation method can be used to select the optimal parameter combination to improve the generalization ability of the model. By rationally utilizing exogenous variables and targeted adjustment of parameters, the SARIMAX model can effectively capture the seasonal and long-term trend characteristics of battery temperature. In the present invention, the seasonal characteristics of battery temperature are defined as the data change patterns of battery temperature in different seasons, and the long-term trend characteristics of battery temperature are defined as the long-term change patterns of battery temperature in different seasons.

[0073] In the above embodiment, the data-driven model tests the internal resistance of the battery pack at different SOC intervals of 10%, wherein the SOH test is conducted from 100% to 70% with an interval of 10%, and the temperature rise rate at different internal resistances and different charge / discharge rates is tested. The internal resistance data is sorted out as the battery pack BOL data. During the use of the battery, according to the voltage and current under different working conditions, the VI method (U=E-Ir, record the voltage and current under different working conditions, use the least squares method to fit a straight line, and regard the slope as the internal resistance at the current SOC) is used to calculate the internal resistance of the battery pack at different SOCs. According to the current internal resistance and charge / discharge rate, the current temperature rise rate is queried, and the temperature rise is calculated based on the duration. The internal resistance update of the current battery pack at different SOCs at different aging degrees is monitored in real time. Each time the low SOC internal resistance is updated, the internal resistance at all SOCs is traversed. If the low SOC internal resistance is smaller than the high SOC internal resistance, it means that the high SOC internal resistance has not been updated for a long time. The long-term unupdated internal resistance value is overwritten according to the following method:

[0074]

[0075] Among them, DCR highsocupdate Indicates that the DCR update value under high SOC has not been updated for a long time. nowsoc Indicates that DCR is updated under the current SOC. BoLnowsoc Indicates the last time BOL updated SOC DCR, DCR highBoLsoc Represents the DCR at high SOC (BOL). The heat generation rate is determined based on the current internal resistance and operating conditions, and the temperature rise is calculated. The calculated temperature rise is used as an independent prediction result in subsequent weighted predictions. Initially, internal resistance and heat generation rate can be tested, and the average value can be obtained from multiple experiments to reduce experimental error and improve the accuracy of the data-driven model. By calculating internal resistance and heat generation rate, the data-driven model provides a physical property-based prediction basis for the hybrid model.

[0076] In the above embodiment, in order to achieve accurate prediction of battery temperature, the present invention also sets different weights for the CNN-LSTM model, SARIMAX model, and data-driven model, and performs weighted averaging on the prediction results to give full play to the advantages of each model and adapt to the temperature change characteristics under different battery operating conditions and states.

[0077] The dynamic weight coefficient formula is defined as:

[0078]

[0079] Among them, i represents the model number, w i Represents the weight corresponding to the model, C rate is the charge and discharge rate, SOH is the battery health status, is the temperature change rate. This formula uses Bayesian optimization to determine the parameter combination. Since the charge and discharge rate directly affects the chemical reaction rate within the battery, which in turn determines the rate of temperature change, higher charge and discharge rates are typically accompanied by more dramatic temperature changes, making them a key consideration in weight allocation. In the formula, k1 indicates that the charge and discharge rate has a significant impact on the weight coefficient, reflecting its important role in battery temperature changes.

[0080] Battery health reflects the changes in the physical and chemical structure of the battery. Batteries with different degrees of aging have different temperature change patterns. For batteries in poor health, the internal resistance may increase and the heat generation characteristics may change, which needs to be reflected in the weight calculation. In the formula, k2 = 0.3, which means that the battery health status also has an important impact on the weight coefficient. The temperature change rate intuitively reflects the dynamic changes in battery temperature. The greater the rate of change, the more drastic the temperature change, and this change trend needs to be captured more accurately during prediction. In the formula, k3 = 0.1. Although its weight is smaller than the first two, it is still an indispensable part of the weight calculation. The denominator ∑exp(·) is the normalization of the sum of all related exponentials to ensure that the weight coefficient ω of each model i The sum is 1, thus achieving reasonable weight distribution.

[0081] By traversing different weight combinations through a grid search method, the optimal weight is determined with the goal of minimizing the prediction error. The weight range and search step size are appropriately set based on the actual needs and cost-effectiveness of battery prediction. For example, considering the potential safety risks of battery overheating, when predicting high temperature conditions, the weight search range is appropriately adjusted to focus on improving prediction accuracy. A grid search parameter space is defined, and for each set of weight combinations, the mean square error of the weighted prediction results is calculated using the formula:

[0082]

[0083] Among them, yi is the true temperature value, is the weighted predicted temperature value, and n is the number of test samples. When calculating the prediction error, in addition to the mean square error, indicators related to battery safety and performance are introduced, such as the proximity of the predicted temperature to the battery thermal runaway temperature threshold, the impact of the predicted temperature on the battery aging rate, etc., to make the weight optimization more in line with the actual application scenarios of battery prediction. At the same time, real-time monitoring of information such as the fast charging and aging status of the battery and the change of internal resistance in the data-driven model is carried out, and the weight search range and step size are dynamically adjusted to ensure that the model can always obtain the optimal weight. When performing grid search, parallel computing can be used to improve search efficiency and quickly find the best weight combination. This weighted prediction strategy based on multi-index optimization and dynamic adjustment can give full play to the advantages of each model, significantly improve the accuracy and reliability of the prediction, and meet the strict requirements of the battery management system for temperature prediction under different working conditions.

[0084] In the present invention, if the fast charge rate exceeds 2C, or the battery health is lower than 80%, the fixed weights in Table 1 are used.

[0085] Table 1

[0086]

[0087] Under fast charging conditions (C rate >2°C), the battery's internal chemical reactions proceed rapidly, resulting in rapid and complex temperature changes. The CNN-LSTM model, with its ability to quickly capture local features and learn time series, is able to effectively capture the short-term, rapid changes in battery temperature during the fast-charging phase. Therefore, it is given a higher weight, set to 0.65. The SARIMAX model excels at capturing long-term trends and seasonality, but these characteristics are relatively less important during the fast-charging phase, so its weight is adjusted to 0.25. The data-driven model calculates key data such as internal resistance based on the battery's physical properties, providing a physical basis for prediction. However, its role during the fast-charging phase is relatively limited, so its weight is determined according to a formula and is set to 0.10 here.

[0088] Under slow charging conditions, battery temperature changes are relatively gradual, and the seasonal impact and long-term trends of ambient temperature have a more pronounced effect on battery temperature. The SARIMAX model is well-suited to modeling these seasonal and long-term trends, so its weight is increased; the weight of the CNN-LSTM model is adjusted accordingly; and the weight of the data-driven model is calculated using the same formula. When the battery is at a shallow depth of discharge, its temperature variation characteristics are similar to those under slow charging conditions, with ambient temperature and long-term trends having a greater impact on battery temperature. Therefore, the weight of the SARIMAX model is appropriately increased, the weight of the CNN-LSTM model is adjusted accordingly, and the weight of the data-driven model is determined according to the formula.

[0089] As batteries age, their internal physical and chemical structures change, leading to different temperature trends. For batteries with advanced age (SOH < 80%), the weight of LSTM in the CNN-LSTM model is further increased. LSTM can better capture the temperature dependence of batteries during long-term use and adapt to the long-term temperature trends of aging batteries. At the same time, the weight of the data-driven model in prediction is gradually increased. Changes in physical properties such as the internal resistance of aging batteries have a more critical impact on temperature prediction, making the predictions provided by the data-driven model based on these physical properties more valuable.

[0090] The present invention deeply integrates CNN, LSTM, SARIMAX and data-driven models. The hybrid model analyzes and processes battery temperature data from different angles, comprehensively captures various characteristics of battery temperature changes, and provides powerful model support for accurate prediction of battery temperature.

[0091] In practical applications, the present invention deploys the trained hybrid model into the battery management system of an electric vehicle. While the vehicle is in motion, the BMS collects real-time data such as battery temperature, SOC, charge / discharge rate, and ambient temperature and inputs it into the model. If, while the vehicle is traveling at high speed, the model predicts that the battery temperature will exceed the safety threshold within 10 minutes, the BMS activates the active cooling system, increasing the fan speed and reducing the charge / discharge rate. Practical applications have verified that the hybrid model can accurately predict battery temperature changes in advance, ensuring safe battery operation and extending battery life.

[0092] As a specific example of the invention, the following detailed case is provided:

[0093] Data preparation stage:

[0094] The temperature data of the power battery of an open source electric vehicle under multiple different driving cycles are selected as experimental data.

[0095] Through the on-board battery management system, information such as temperature, battery state of charge, charge and discharge rate, ambient temperature, humidity, and vehicle speed at multiple monitoring points in the battery module is collected at a frequency of once per second. A total of 50 consecutive complete driving cycles were collected.

[0096] Identify and process missing values: For a small number of consecutive missing values, linear interpolation is used to fill them in. For time periods with a large number of missing values, the missing values ​​are estimated and filled in by combining data from other similar operating conditions within the same driving cycle and the physical properties of the battery. In the present invention, the physical properties of the battery refer to the characteristics of the battery itself. For example, the battery capacity will decrease with the number of charge and discharge cycles, and the battery internal resistance will increase with the number of charge and discharge cycles.

[0097] Abnormal values ​​are judged based on the normal operating temperature range of the battery and the reasonable relationship between the charge and discharge rate and temperature. For abnormal temperature values ​​outside the normal range, normal data at adjacent time points are used for mean filtering.

[0098] The temperature data units are uniformly converted to Celsius, and all feature data are standardized so that their mean is 0 and their variance is 1.

[0099] The preprocessed data was divided into a 70% training set, a 15% validation set, and a 15% test set in chronological order. When dividing the training set, the fast-charging data and the aging battery data were clearly labeled, and their respective proportions were calculated to provide a basis for subsequent model training and parameter adjustment.

[0100] Following the requirements of the data-driven model, we tested the internal resistance of the battery at different SOC and SOH levels, recorded data such as heat generation rate, and integrated this data into the training, validation, and test sets. To ensure data quality, we conducted multiple cross-validations on the collected data to check its consistency and reliability.

[0101] Model building and training phase:

[0102] CNN-LSTM model: The constructed CNN-LSTM model includes: the first convolutional layer uses 64 filters of size 3, the activation function uses ReLU, a batch normalization layer and a one-dimensional maximum pooling layer; the second convolutional layer has 32 filters and the same operation; the LSTM layer has 100 units and a batch normalization layer; it is flattened by a Flatten layer, passed through a fully connected layer with 64 neurons and a random inactivation layer, and the output layer has 1 neuron to predict the battery temperature.

[0103] The Adam optimizer was used during training. When the proportion of fast-charging data in the training set exceeded 30%, the learning rate was increased from 0.001 to 0.0015. When the proportion of aged battery data exceeded 20%, the number of training epochs was increased from 100 to 120. The validation set was used for evaluation every 10 epochs. Training was terminated early if the validation set loss did not decrease for five consecutive epochs. During training, intermediate model results were regularly saved to facilitate subsequent analysis of the model's training progress and performance changes. Data augmentation techniques were also used, such as random flipping and translation of the training data, to increase data diversity and improve the model's generalization capabilities.

[0104] SARIMAX Model: A SARIMAX model was trained using the training set data and relevant features. Parameters were determined using Bayesian optimization: seasonal autoregressive order (P) = 1, seasonal differencing order (D) = 1, seasonal moving average order (Q) = 1, and seasonal period S = 1440. A seasonal differencing order D = 1 indicates first-order seasonal differencing of the time series data to eliminate seasonal trends, and a seasonal moving average order Q = 1 indicates a linear relationship between the errors of the current and previous seasons. For training sets containing aged battery data exceeding 30%, the autoregressive order (P) was adjusted to 3 and the seasonal autoregressive order (P) to 2 to better fit the temperature variation of aged batteries. Furthermore, some data from the data-driven model (such as the internal resistance variation trend at different aging levels) was incorporated as exogenous variables into model training. Parallel computing technology was used to accelerate the SARIMAX model training process and reduce training time. Furthermore, performance evaluation was performed on the models under different parameter combinations, and performance curves were plotted to visually demonstrate the model's performance under different parameter settings, providing a reference for parameter adjustment. By optimizing parameters and rationally utilizing exogenous variables, the SARIMAX model can more accurately capture the seasonal and long-term trend characteristics of battery temperature. Parallel computing technology and performance evaluation methods improve the efficiency of model training and the scientific nature of parameter adjustment.

[0105] Data-driven model: In the early stages of the experiment, the internal resistance of the battery pack was tested at 10% SOC intervals and 10% SOH intervals. The VI method was used to calculate the internal resistance at different SOCs, and the heat generation rate at different charge and discharge rates under different internal resistances was recorded. In actual applications, the internal resistance update is monitored in real time, and the internal resistance at different SOCs is updated according to established rules. The heat generation rate is determined based on the current internal resistance and operating conditions, and the heating temperature is calculated. To improve the accuracy of the data-driven model, the calculation of the heat generation rate and heat dissipation effect is calibrated multiple times, and the parameters in the calculation model are optimized based on the physical properties of the actual battery and experimental data. At the same time, a dynamic update mechanism for the internal resistance and heat generation rate is established. According to changes in the battery's operating conditions, the internal resistance data at different power levels is updated in a timely manner to ensure the real-time and accuracy of the model. The data-driven model provides reliable physical property data support for the hybrid model, ensuring the accuracy and real-time nature of the prediction results.

[0106] Model optimization phase:

[0107] A grid search method was used to traverse weight combinations. Under fast-charging conditions, the weight range was [0.6, 0.8] with a step size of 0.05; under slow-charging conditions, the weight range was [0.4, 0.6] with a step size of 0.05; and under shallow depth of discharge conditions, the weight range was [0.4, 0.6] with a step size of 0.05. The mean squared error (MSE) of the weighted prediction results for each weight combination was calculated on the validation set, and the combination with the lowest MSE was selected. For battery data with a degree of aging exceeding 50%, the prediction error was calculated by increasing the weight of indicators related to the predicted temperature and battery aging rate. For example, the deviation between the predicted temperature and the temperature threshold for accelerated battery aging was included in the error calculation, and the weight was increased from 10% to 20%. Ultimately, weights that better reflect the characteristics of aged batteries were determined. For example, in this experiment, for batteries with a high degree of aging, the hybrid CNN-LSTM model had a weight of 0.45, the SARIMAX model had a weight of 0.35, and the data-driven model had a weight of 0.2. During the grid search, a distributed computing platform was utilized to improve search efficiency and quickly identify the optimal weight combination. Furthermore, an adaptive weight adjustment strategy was employed to dynamically adjust weights based on the battery's real-time status and prediction error, further improving the model's prediction accuracy. This maximized the hybrid model's prediction accuracy, meeting the high-precision requirements for battery temperature prediction under various operating conditions.

[0108] Model evaluation phase:

[0109] Evaluation indicators: The hybrid model is evaluated using a test set, and the mean square error, mean absolute error, and root mean square error are calculated. At the same time, evaluation indicators related to battery performance and safety are added, such as calculating the deviation between the predicted temperature and the battery's optimal operating temperature range to evaluate the impact on the battery aging rate; analyzing the impact of the predicted temperature on the energy consumption of the battery thermal management system, adjusting the heat dissipation or heating strategy based on the predicted temperature, and calculating the corresponding energy consumption changes. In addition, a battery thermal runaway risk assessment indicator is introduced to evaluate the possibility of thermal runaway in the battery by comparing the predicted temperature with the thermal runaway temperature threshold. A comprehensive evaluation indicator system can evaluate model performance from multiple dimensions, more accurately reflect the effectiveness of the model in actual applications, and provide a strong basis for model optimization and improvement.

[0110] Result analysis: Figure 2 As shown in the figure, the prediction results of the hybrid model are closer to the actual temperature and have higher prediction accuracy than those using the CNN-LSTM model or the SARIMAX model to predict temperature. It can be seen from the figure that the predicted temperature curve basically coincides with the actual temperature curve, and also shows a good tracking trend during the stage of drastic battery temperature changes. Compared with the single CNN-LSTM model and SARIMAX model, the hybrid model has significantly reduced various error indicators. Figure 3As shown, the MSE of the hybrid model on the test set is 0.85, which is lower than the MSE of the CNN-LSTM and SARIMAX models, and the prediction is more accurate. Further analysis of the performance of the model under different working conditions shows that the prediction accuracy of the hybrid model under complex working conditions such as fast charging and aging is particularly obvious. For example, under fast charging conditions, the MSE of the hybrid model is 30% lower than that of the single CNN-LSTM model and 40% lower than that of the single SARIMAX model. At the same time, by conducting comparative experiments with other advanced battery temperature prediction methods, the superiority of the hybrid model of the present invention is verified, and a more reliable temperature prediction solution is provided for the battery management system. The experimental results fully demonstrate the excellent performance of the hybrid model of the present invention in battery temperature prediction. It has obvious advantages over other models and can provide more accurate and reliable temperature prediction services for the battery management system to meet the needs of actual applications.

[0111] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A battery temperature prediction method based on multi-model fusion, characterized by: include Step 1: Obtain relevant data for predicting battery temperature, preprocess the data, and construct a data set; Step 2: Build a CNN-LSTM model, a SARIMAX model, and a data-driven model, and train the models using the battery temperature dataset. Step 3: Combine the test results of the CNN-LSTM model, SARIMAX model, and data-driven model to build a temperature prediction model, and use the data set built in step 1 to process the temperature prediction model; Step 4: Optimize the weights of step 3 to optimize the temperature prediction model.

2. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: The step 1 specifically includes: Obtain temperature data of the power battery of an open-source electric vehicle under multiple different driving cycles; Collect battery status data, environmental data, and vehicle driving data from multiple monitoring points within the battery pack; Calculate the battery internal resistance at different SOC (State of Charge) and SOH (State of Health), and record the heat generation rate at different charge and discharge rates under different internal resistances; Perform missing value processing, smoothing and denoising on the above data; The processed data is divided into training set, validation set and test set, and the battery fast charging data and battery aging data are marked.

3. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: The CNN-LSTM model contains two convolutional layers, a pooling layer, an LSTM layer and a fully connected layer, where: The first convolution layer uses 64 filters of size 3, and the activation function adopts ReLU. A batch normalization layer and a one-dimensional maximum pooling layer are added. The second convolution layer uses 32 filters of size 3, and the activation function adopts ReLU. A batch normalization layer and a one-dimensional maximum pooling layer are added. The LSTM layer uses 100 units to capture long-term dependencies and adds a batch normalization layer.

4. The battery temperature prediction method based on multi-model fusion according to claim 3, characterized in that: The LSTM layer uses the SOH attenuation gradient as the gate input and introduces the cell state into the gate connection. The formula is i t =σ(W xi x t +W ci c t-1 +W hi h t-1 +b i ) f t =σ(W xf x t +W ci c t-1 +W hf h t-1 +b f ) o t =σ(W xo x t +W ci c t-1 +W ho h t-1 +b o ) σ is the sigmoid activation function, W is the weight matrix, b is the bias vector, h t-1 is the hidden state of the previous time step t-1, C t-1 is the cell state at the previous time step t-1.

5. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: It also includes tuning the CNN-LSTM model training parameters using the Adam optimizer and the mean squared error loss function.

6. The battery temperature prediction method based on multi-model fusion according to claim 2, characterized in that: The formula of the SARIMAX model is: φ(B)(1-B S )Y t =θ(B)(1-θ(B s ))e t +βX t , Where φ(B) and θ(B) are the autoregressive and moving average polynomials, respectively, S is the seasonal cycle, and X t is an exogenous variable; 1-Θ(B s ) indicates that the seasonal trend is eliminated by differentiating the time series, Y t represents the observed value of the time series at time t, B is the lag factor, ε t is the white noise error at time t, β is the exogenous variable X t The coefficient of The SARIMAX model is trained using the training set data, and the parameters are determined by Bayesian optimization, including: Seasonal autoregressive order (P): represents the order of φ(B); Seasonal difference order (D): indicates the seasonal difference of time series data to eliminate seasonal trends; Seasonal moving average order (Q): determines 1-Θ(B s ) order; and seasonal cycle (S).

7. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: The data-driven model calculates the battery internal resistance at different SOCs based on the VI method, determines the heat generation rate based on the current internal resistance and operating conditions, and obtains the heating temperature.

8. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: In the step 4, optimizing the weights includes: The dynamic weight coefficient formula is defined as Among them, i represents the model number, w i Represents the weight corresponding to the model, C rate is the charge and discharge rate, SOH is the battery health, is the temperature change rate; The grid search method is used to traverse different weight combinations and determine the optimal weight with the goal of minimizing the prediction error.

9. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: In the step 4, optimizing the weight further includes: In C rate Under the working condition of >2C, within the normal rate range of the battery, the weights of the CNN-LSTM model, SARIMAX model, and data-driven model are set to 0.65, 0.25, and 0.1 respectively; Under the condition of SOH < 80%, the weights of the CNN-LSTM model, SARIMAX model, and data-driven model are set to 0.6, 0.2, and 0.2, respectively.

10. The battery temperature prediction method based on multi-model fusion according to claim 1, characterized in that: It also includes the evaluation of the temperature prediction model by calculating the mean square error, mean absolute error, and root mean square error.

Images