Method and system for monitoring temperature of energy storage battery in high altitude area

Abstract

The invention discloses a method and a system for monitoring the temperature of an energy storage battery in a high-altitude area. Through environment adaptive feature reconstruction, thermodynamic model correction and multi-frequency EIS fusion feature extraction are utilized to generate a feature vector adaptive to the plateau environment. The thermodynamic model correction adjusts the convective heat transfer coefficient by introducing an air pressure correction coefficient; the multi-frequency EIS fusion feature selects a specific frequency point impedance value to construct a vector, and the weight is dynamically adjusted through a random forest algorithm. The lightweight hybrid machine learning model integrates the advantages of LightGBM, 1D-CNN and a physical constraint particle filter model, and realizes accurate prediction for different working conditions. The plateau exclusive training mechanism covers data enhancement, transfer learning and online calibration, and the generalization ability and adaptability of the model are improved. According to the method, the problems of low temperature monitoring precision and poor model adaptability of the energy storage battery in the high-altitude area are solved, the monitoring precision and the system reliability are remarkably improved, and a guarantee is provided for safe and stable operation of a high-altitude energy storage system.

Description

Technical Field

[0001] This invention relates to the field of temperature monitoring of energy storage batteries in high-altitude areas, and specifically to a method and system for monitoring the temperature of energy storage batteries in high-altitude areas. Background Technology

[0002] With the advancement of the global energy transition, energy storage systems are being used more and more widely in high-altitude areas. However, high-altitude areas have low air pressure and large temperature differences, and traditional methods for monitoring the temperature of energy storage batteries have many shortcomings.

[0003] 1. Existing technologies do not fully consider the impact of air pressure at high altitudes on battery heat dissipation, leading to reduced accuracy of thermodynamic models and an inability to accurately predict battery temperature. Traditional heat conduction models do not accurately describe the changes in convective heat transfer coefficients under low air pressure environments, resulting in significant discrepancies between model predictions and actual heat dissipation.

[0004] 2. Traditional temperature monitoring methods are insufficient in capturing changes in the internal electrochemical characteristics of batteries under low-temperature conditions, making it difficult to effectively integrate multi-frequency EIS features, resulting in insufficient feature extraction. The kinetic characteristics of internal electrochemical reactions change significantly under low-temperature conditions, but existing methods cannot comprehensively capture these changes, thus affecting the accuracy of temperature monitoring.

[0005] 3. Existing machine learning models have weak generalization ability in the complex environment of high altitudes, consume a lot of computational resources, and waste measurement resources significantly, failing to meet the high accuracy, real-time performance, and reliability requirements of temperature monitoring for energy storage systems in high-altitude areas. For example, traditional LSTM models are prone to overfitting when dealing with complex operating conditions in high-altitude areas and have high computational resource requirements, making them difficult to apply effectively in resource-constrained energy storage systems.

[0006] To address the aforementioned issues, this invention proposes a method and system for monitoring the temperature of energy storage batteries in high-altitude and cold regions based on environmental adaptive feature reconstruction and lightweight hybrid learning. The aim is to improve the accuracy and reliability of temperature monitoring for energy storage batteries in high-altitude areas while reducing the consumption of computing and measurement resources. Summary of the Invention

[0007] This invention aims to solve the problems of low temperature monitoring accuracy and poor model adaptability of energy storage batteries in high-altitude areas in the prior art, and provides a temperature monitoring system and method for energy storage batteries in high-altitude areas to achieve accurate and real-time monitoring of battery temperature and ensure the safe and stable operation of energy storage systems in high-altitude areas.

[0008] Specifically, the objectives of this invention are as follows:

[0009] 1. A thermodynamic model correction method that can adapt to environmental changes in high-altitude areas is proposed. By introducing an air pressure correction coefficient, the heat dissipation characteristics of the battery under different air pressures are accurately reflected, thereby improving the applicability and accuracy of the thermodynamic model in high-altitude areas.

[0010] 2. Develop a multi-frequency EIS fusion feature extraction method, which uses a random forest feature selection algorithm to dynamically adjust the weights of each frequency band, fully explore the internal electrochemical characteristics of the battery, and enhance the characterization capability of temperature monitoring features.

[0011] 3. Construct a lightweight hybrid machine learning model that combines the advantages of LightGBM, 1D-CNN and physically constrained particle filter models to predict temperature under steady-state, transient and data-missing conditions, thereby improving the model's generalization ability and prediction accuracy under complex conditions.

[0012] 4. Design a plateau-specific training mechanism, including data augmentation, transfer learning, and online calibration, to make full use of limited plateau data resources, improve the training efficiency and adaptability of the model, and ensure that the model maintains high accuracy in long-term operation.

[0013] The specific plan is as follows:

[0014] A method for monitoring the temperature of energy storage batteries in high-altitude areas includes:

[0015] S1. The thermodynamic model is corrected using the environmental adaptive feature reconstruction module, and multi-frequency EIS fusion features are extracted. The thermodynamic model is corrected by introducing a pressure correction coefficient to adjust the convective heat transfer coefficient. The multi-frequency EIS fusion features select impedance values ​​at specific frequencies to construct vectors, and the weights are dynamically adjusted using the random forest algorithm.

[0016] S2. Construct a lightweight hybrid machine learning model, including LightGBM, 1D-CNN and physically constrained particle filter model, to predict the core temperature of the battery.

[0017] S3. The model is trained and optimized using a plateau-specific training mechanism, including data augmentation, transfer learning, and online calibration.

[0018] Furthermore, step S1 specifically includes: environmental adaptive feature reconstruction.

[0019] (1) Thermodynamic model correction

[0020] At high altitudes, the drop in air pressure significantly impacts the heat dissipation efficiency of batteries, a process that traditional thermal conduction models cannot accurately describe. Therefore, this invention introduces an air pressure correction coefficient to adjust the convective heat transfer coefficient in the cylindrical battery thermal conduction equation, as shown in the following formula:

[0021]

[0022] in:

[0023] h high-alt The corrected convective heat transfer coefficient reflects the actual heat dissipation efficiency in high-altitude areas.

[0024] h: Convective heat transfer coefficient under standard atmospheric pressure, determined based on experimental data of the battery under standard atmospheric pressure (typically 101.3 kPa).

[0025] P: Local actual air pressure, which is measured in real time by the air pressure sensor installed in the energy storage system.

[0026] P0: Standard atmospheric pressure, usually taken as 101.3 kPa, used as a reference pressure for comparison.

[0027] The revised thermodynamic model dynamically adjusts the convective heat transfer coefficient based on the actual air pressure, making it more consistent with the actual heat dissipation conditions at high altitudes. This provides an accurate theoretical basis for subsequent temperature monitoring and ensures that the system outputs reliable results stably at different altitudes.

[0028] (2) Multi-frequency EIS fusion features

[0029] Under low-temperature conditions, the internal electrochemical characteristics of the battery change significantly, and the impedance characteristics at different frequencies differ markedly. This invention selects 10Hz (Z) as the reference. 10 ), 100Hz (Z) 100 ) and 1kHz (Z 1000 Three characteristic frequency points are used to construct a three-dimensional impedance vector [Z]. 10 Z 100 Z 1000 The weights of each frequency band are dynamically adjusted using a random forest feature selection algorithm, and the fusion formula is as follows:

[0030]

[0031] in:

[0032] The fused average temperature is the final temperature estimate that integrates impedance information from three frequency points. (Average temperature extracted from multi-frequency EIS fusion features) It doesn't directly refer to the battery's core temperature, but rather a characteristic temperature calculated based on the fusion of multi-frequency impedance data. This characteristic temperature comprehensively reflects the relationship between the battery's electrochemical characteristics and temperature at different frequencies. It is mainly used to capture the temperature change trends and characteristic information inside the battery, rather than directly representing the battery's core temperature.

[0033] w1, w2, w3: Weight coefficients for the three frequency bands, dynamically adjusted by the random forest feature selection algorithm based on ambient temperature, reflecting the importance of impedance characteristics of each frequency band at different temperatures.

[0034] f1, f2, f3: These are the mapping functions between impedance and temperature for the corresponding frequency bands. They are obtained by fitting a large amount of experimental data and describe the relationship between impedance changes and temperature changes in this frequency band.

[0035] Z 10 Z 100 Z 1000 : Impedance values ​​at frequencies of 10Hz, 100Hz and 1kHz, respectively, were obtained by electrochemical impedance spectroscopy.

[0036] Based on the temperature field information provided by the modified thermodynamic model, the multi-frequency EIS fusion features utilize the random forest algorithm to dynamically adjust the weights according to the ambient temperature, so that the impedance features of each frequency band are reasonably allocated to different temperatures, providing high-quality feature inputs for subsequent machine learning model training and prediction, and enhancing the accuracy of temperature monitoring.

[0037] Furthermore, step S2 specifically includes: using a lightweight hybrid machine learning model to predict battery temperature.

[0038] Battery temperature prediction is crucial for battery management systems, effectively improving battery performance, extending lifespan, and enhancing safety. This invention presents a lightweight hybrid machine learning model that integrates the advantages of multiple machine learning algorithms and physical models to accurately and efficiently predict core battery temperatures under various operating conditions.

[0039] (1) Model Architecture

[0040] This lightweight hybrid machine learning model consists of three main models: the LightGBM model for steady-state conditions, the 1D-CNN model for transient conditions, and the physically constrained particle filter model to address scenarios with missing data at extremely low temperatures. Through mechanisms such as model selection and division of labor, and model fusion logic, the model achieves complementary advantages, meeting the core temperature prediction needs of batteries under different operating conditions and complex environments.

[0041] (2) Model selection and division of labor

[0042] ①LightGBM model (steady-state condition)

[0043] Input features:

[0044] X a (Battery voltage) reflects the potential level of the battery during operation and is closely related to the battery's charge and discharge state and internal electrochemical reactions, providing electrochemical state information for the model.

[0045] Xb (Battery current) reflects the rate of charge flow during charging and discharging. The magnitude and direction of the current affect the battery's heat generation and dissipation.

[0046] X c (Battery self-discharge rate) measures the rate at which a battery loses its charge naturally when no external circuit is connected, and is related to the heat generated by side reactions inside the battery.

[0047] X d (Battery health status) comprehensively evaluates battery performance indicators such as cycle life and capacity decay, which affect battery thermal characteristics.

[0048] T o (Ambient temperature) provides the temperature conditions of the environment in which the battery operates, and is a key parameter for heat exchange between the battery and the environment.

[0049] (Average temperature after fusion) is obtained by combining impedance information from three frequency points. Multi-frequency impedance reflects the electrochemical and thermal characteristics of the battery at different time scales. The average temperature after fusion can more accurately characterize the internal temperature state.

[0050] h high-alt (The corrected convective heat transfer coefficient) characterizes the strength of convective heat transfer between the battery surface and the surrounding environment. It is affected by factors such as air velocity and battery surface structure, and has a direct impact on the battery's heat dissipation effect.

[0051] Output core temperature estimate:

[0052] That is, the battery core temperature value predicted by the model.

[0053] Prediction process:

[0054] LightGBM is a gradient boosting decision tree algorithm that constructs multiple decision trees. The construction process for each tree is as follows: starting from the root node, the data is recursively split based on the input features. The optimal splitting features and thresholds are selected to make the temperature values ​​of samples within a node as close as possible, ultimately forming a leaf node. The output of each decision tree is the average temperature of the corresponding leaf node. The formula is:

[0055]

[0056] Where K is the number of decision trees, f k (x i ) represents the input feature x of the k-th decision tree. i The model integrates the predictions from multiple decision trees to obtain the final core temperature estimate. Each input feature interacts during the decision tree splitting process, influencing the tree structure and leaf node temperature values. For example, the battery voltage X... a and current Xb The battery's heat generation capacity is determined by the ambient temperature T. o Heat transfer coefficient h high-alt Synergistic effect on battery heat dissipation, health status X d Changing the battery's thermal properties, self-discharge rate X c This causes additional heat generation, fusion temperature It provides prior information on the internal thermal state. These information are interconnected and, through model calculations, enable accurate prediction of the battery core temperature under steady-state conditions.

[0057] ② 1D-CNN model (transient condition)

[0058] Input features:

[0059] Ts (time-series temperature data) records the sequence of changes in battery surface temperature over time, reflecting the historical temperature dynamics of the battery and providing temperature change trend information for the model.

[0060] Z 100 (Impedance value change sequence at 100Hz frequency) obtains the impedance time sequence change of the battery at 100Hz frequency, reflecting the rapid electrochemical reaction kinetics and thermal characteristics of the electrode-electrolyte interface inside the battery, which is related to the heat generation and transfer of the battery in the transient process.

[0061] h high-alt (The corrected convective heat transfer coefficient), the same as the heat transfer coefficient in the LightGBM model, provides the model with information on heat dissipation conditions.

[0062] Output core temperature change rate estimate:

[0063] Predict the rate of change of battery core temperature over time.

[0064] Prediction process:

[0065] 1D-CNN extracts local features by sliding the convolutional kernels of convolutional layers across the input temporal data. The convolution operation formula is:

[0066]

[0067] Among them, w i These are the convolution kernel weight parameters, x j+i Let be the value of the input time-series data at position j+i, b be the bias term, and σ be the activation function (commonly the ReLU function). The convolution kernel is applied to the time-series temperature data Ts and the impedance value change sequence Z. 100 Slide the slider upwards to extract features within a local time window, such as trends in temperature rise or fall, and abrupt changes in impedance. Simultaneously, the heat transfer coefficient h... high-altAs an additional input channel, it is fused with the aforementioned features to jointly influence the output of the convolutional layer. After multiple convolutional and pooling operations, the model maps the extracted features to the output space of the core temperature change rate, obtaining an estimate of the core temperature change rate. The time-series temperature data Ts provides real-time temperature feedback, and the impedance value change sequence Z... 100 The heat transfer coefficient h reflects the rapid dynamic changes of the heat source in an electrochemical reaction. high-alt The three factors—the rate of heat loss, the speed of heat dissipation, and the rate of heat loss—work together to enable the model to accurately capture the rapid changes in core temperature under transient conditions.

[0068] ③ Physically constrained particle filter model (for scenarios with missing ultra-low temperature data)

[0069] Input features:

[0070] Ambient temperature, especially in ultra-low temperature environments, is the main driving factor for battery heat loss, determining the temperature difference between the battery and the environment and the direction of heat transfer.

[0071] Air pressure affects air density and thermophysical properties, which in turn alters the battery's convective heat transfer and heat dissipation.

[0072] h high-alt (The corrected convective heat transfer coefficient) is consistent with the heat transfer coefficient in the LightGBM and 1D-CNN models, and is used in this model to correct the heat dissipation calculation at ultra-low temperatures.

[0073] Output core temperature estimate:

[0074] Predicted battery core temperature value.

[0075] Prediction process:

[0076] This model describes the dynamic changes in the battery core temperature based on thermodynamic equations:

[0077]

[0078] in, Let be the core temperature of the battery at time k, and α be the thermal diffusivity (reflecting the rate of heat diffusion within the battery), ▽ 2T represents the temperature gradient change (reflecting the non-uniform heat distribution inside the battery), Δt is the time step, β is the pressure-related heat dissipation coefficient, P0 is the standard atmospheric pressure, and P is the local actual atmospheric pressure. This equation shows that the core temperature is determined by the previous temperature, internal heat diffusion, and heat dissipation terms. Particle filtering estimates the temperature by approximating the posterior probability density function of the particle set. During initialization, a set of particles is randomly generated within the possible core temperature range, with each particle representing a possible core temperature state. In the prediction phase, based on the thermodynamic equation and the input ambient temperature, pressure, heat transfer coefficient, etc., the temperature value of each particle at the current moment is predicted. In the update phase, particle weights are calculated using a small amount of available observation data (such as battery surface temperature).

[0079]

[0080] in, z is the weight of the i-th particle at time k. (k) For the observed values, This is an observational model (assuming a Gaussian distribution). After weight updates, particles are resampled, retaining high-weight particles and discarding low-weight particles. Finally, the particle mean or weighted mean is used as the core temperature estimate. In this model, ambient temperature determines the direction and rate of heat loss, air pressure affects the heat dissipation coefficient and thus changes the heat dissipation effect, and the heat transfer coefficient h... high-alt The heat dissipation terms were revised to better reflect actual heat dissipation conditions at ultra-low temperatures. These parameters collectively drive the thermodynamic equations, which, through particle filtering algorithms, enable a reasonable estimate of the battery core temperature in scenarios where ultra-low temperature data is scarce.

[0081] ④ Model fusion logic

[0082] The hybrid model allocates tasks in real time based on battery operating conditions and environmental conditions. After data preprocessing, a weighted average method is used to fuse the prediction results of each model to generate the final core temperature estimate.

[0083]

[0084] Among them, w i The weights are those of the i-th model, pre-set based on the prediction performance of each model under specific operating conditions. Under steady-state conditions, the LightGBM model has a larger weight because it can fully utilize multi-dimensional features to accurately predict temperature; under transient conditions, the 1D-CNN model has a higher weight due to its ability to capture temporal features; in scenarios where ultra-low temperature data is scarce, the physically constrained particle filter model has a dominant weight, relying on physical principles for reasonable estimation.

[0085] (3) Model selection and usage methods

[0086] ① Operating condition judgment

[0087] The sensor collects data such as battery SOC, current, and ambient temperature in real time, and determines the operating condition according to preset rules:

[0088] When SOC>50% and current change rate is less than 0.5A / min, it is considered a steady-state condition. At this time, the battery charging and discharging process is relatively stable, and the internal heat generation and dissipation are in dynamic equilibrium, which is suitable for temperature prediction using the LightGBM model.

[0089] When SOC≤50% or current change rate≥0.5A / min, it is judged as transient condition, the battery is in charge-discharge transition or high current charge-discharge stage, internal heat generation and temperature distribution change rapidly, and is predicted by 1D-CNN model.

[0090] When the ambient temperature is below -40℃ or the sensor malfunction causes data loss, the scenario of ultra-low temperature data loss is entered. The battery heat dissipation conditions are extreme and data acquisition is difficult. At this time, the physical constraint particle filter model is used for temperature prediction.

[0091] ② Data preprocessing

[0092] The collected battery data is cleaned to remove outliers and noisy data; normalization is then performed to unify features with different dimensions and value ranges into the [0,1] interval, using the following formula:

[0093]

[0094] Where x' is the normalized value, x is the original data, and x' is the normalized value. min and x max These represent the minimum and maximum values ​​of the data, respectively. Simultaneously, a sliding window is applied to the time-series data to create a time-series window suitable for input to the 1D-CNN model; feature fusion calculations are performed on the multi-frequency EIS data to obtain the fused average temperature. As one of the input features of the LightGBM model.

[0095] ③ Model Input Preparation

[0096] Based on the different model requirements, corresponding input feature vectors are prepared. For the LightGBM model, the battery voltage, current, self-discharge rate, health status, ambient temperature, fusion temperature, and heat transfer coefficient are combined into a feature vector; for the 1D-CNN model, an input tensor containing time-series temperature data, a 100Hz impedance value change sequence, and heat transfer coefficient is constructed; for the physics-constrained particle filter model, the data structure of ambient temperature, air pressure, and heat transfer coefficient is organized to provide input for model prediction.

[0097] ④ Model prediction

[0098] Input the prepared input features into the corresponding model, run the model to calculate, and obtain the output results of each model:

[0099] LightGBM model outputs core temperature estimates.

[0100] The 1D-CNN model outputs an estimate of the core temperature change rate.

[0101] The physical constraint particle filter model outputs the core temperature estimate.

[0102] Furthermore, step S3 specifically includes: implementation of a high-altitude-specific training mechanism.

[0103] (1) Data Augmentation Implementation

[0104] ① Constructing a Generative Adversarial Network (GAN): Building a GAN architecture that includes a generator G and a discriminator D. The generator is used to generate virtual data, and the discriminator is used to determine whether the data is real data.

[0105] ②Prepare real data and random noise: Collect a small amount of real data of the battery at low pressure of -40℃ in the high-altitude environment as real data x, which follows the real data distribution pdata; at the same time, generate random noise z, which follows the random noise distribution pz.

[0106] ③ Adversarial training process: Random noise z is input into the generator G to generate virtual data G(z). Real data x is input into the discriminator D, and the discriminator output logD(x) is calculated; the generated virtual data G(z) is input into the discriminator D, and the discriminator output log(1-D(G(z))) is calculated. By minimizing the Jensen-Shannon divergence between the generated data and the real data, the parameters of the generator and the discriminator are updated. The objective function is:

[0107]

[0108] During training, the generator continuously learns the distribution characteristics of plateau environmental data and generates virtual samples with statistical characteristics similar to real data in order to increase the amount of measured data on the plateau and alleviate the problem of data scarcity.

[0109] (2) Implementation of transfer learning

[0110] ① Plain Data Pre-training: A lightweight hybrid machine learning model is pre-trained on a rich battery dataset from a plain region. The plain data is used to learn the general temperature variation patterns and feature representations of batteries, optimizing the model parameters θ to achieve better temperature prediction performance in a plain environment.

[0111] ② Model Transfer and Parameter Fine-tuning: The pre-trained model is transferred to a high-altitude region. A small-sample dataset from the high-altitude region is collected, containing characteristic data of the battery under high-altitude conditions (such as voltage, current, impedance, etc.) and corresponding core temperature data. The model parameters are fine-tuned using the high-altitude small-sample dataset, focusing on adjusting the feature weights and bias terms related to the high-altitude environment. By minimizing the distribution difference between the source domain (plains) and the flat domain (high-altitude region), a domain adaptation method is adopted, with the objective function as follows:

[0112] m θ inL task (θ)+λL domain (θ)

[0113] Among them, L task L is the task loss function, which measures the model's error in the temperature prediction task; domain λ is the domain adaptation loss function, which measures the difference in data distribution between the source and flat domains; λ is the balancing parameter, used to control the weights of the task loss and the domain adaptation loss; θ is the model parameter. This approach allows the model to better adapt to the plateau environment, improving the accuracy of temperature predictions in plateau regions.

[0114] (3) Implementation of online calibration

[0115] ① Set calibration trigger conditions: Set the daily off-peak electricity period as the online calibration trigger time. During off-peak electricity periods, the energy storage system is in a relatively low load or idle state, at which time a pulsed EIS (electrochemical impedance spectroscopy) scan is triggered to reduce interference with the normal operation of the battery.

[0116] ② Pulsed EIS Scan Operation: After trigger calibration, a short-duration pulse excitation signal is applied to the battery. Using an electrochemical impedance spectroscopy (EIS) instrument, the impedance data of the battery is measured under pulse excitation to obtain the impedance frequency response characteristics of the battery in its current state, thereby obtaining the impedance data of the battery in its current state.

[0117] ③ Least squares method to update fitting coefficients: The acquired current state impedance data of the battery and the real-time ambient temperature are used as input features X. i (Including battery impedance data and ambient temperature, etc.), based on the actual temperature measurement value y of the battery. i As observed values. Using the least squares method, by minimizing the sum of squared residuals:

[0118]

[0119] New impedance-temperature mapping models f1, f2, and f3 were calculated. f1, f2, and f3 describe the relationship between input characteristics and battery temperature. The updated mapping models f1, f2, and f3 will be used in subsequent temperature monitoring models to more accurately reflect the impedance-temperature relationship of the battery in the current high-altitude environment, thereby improving the real-time performance and accuracy of temperature monitoring.

[0120] A temperature monitoring system for energy storage batteries in high-altitude areas includes:

[0121] The environmental adaptive feature reconstruction module is used to correct the thermodynamic model and extract multi-frequency EIS fusion features. The thermodynamic model is corrected by introducing a pressure correction coefficient to adjust the convective heat transfer coefficient. The multi-frequency EIS fusion features select impedance values ​​at specific frequencies to construct vectors and dynamically adjust the weights using a random forest algorithm.

[0122] A lightweight hybrid machine learning model module, including LightGBM, 1D-CNN, and physically constrained particle filter models, is used to predict battery core temperature.

[0123] A plateau-specific training mechanism module is used for training and optimizing the model.

[0124] The beneficial effects of this invention are as follows:

[0125] (1) A pressure-adaptive thermodynamic model correction method was proposed, which solved the problem that the traditional thermodynamic model could not accurately describe the convective heat transfer coefficient in the low-pressure environment of the plateau, and significantly improved the applicability of the model in high-altitude areas and the accuracy of temperature prediction.

[0126] (2) The multi-frequency EIS fusion feature extraction method was improved. Through the dynamic weight allocation mechanism, the correlation information between impedance of different frequency bands and temperature change at low temperature was fully explored, which solved the problem of insufficient feature extraction in low temperature environment by traditional methods and enhanced the feature representation ability.

[0127] (3) A lightweight hybrid machine learning model was constructed, which integrates the advantages of LightGBM (steady-state), 1D-CNN (transient) and physical constraint particle filter (data missing) models. Combined with the intelligent working condition judgment mechanism, it realizes high-precision temperature prediction under complex working conditions on plateau (steady-state, transient, and ultra-low temperature data missing), while significantly reducing the computational resource consumption of the model.

[0128] (4) A plateau-specific training mechanism was designed, which comprehensively utilizes generative adversarial network (GAN) data augmentation, transfer learning and online calibration techniques to effectively overcome the limitation of scarce real data in plateau areas and improve the model's generalization ability, adaptability and long-term stability in the special environment of plateau (such as reducing model drift).

[0129] The system systematically solves the core problems in temperature monitoring of energy storage batteries in high-altitude areas, such as low monitoring accuracy (the model is not adapted to low air pressure and insufficient low temperature feature extraction), poor model adaptability (weak generalization ability under complex plateau conditions and scarce data), and high resource consumption (heavy computational burden of traditional models). Attached Figure Description

[0130] Figure 1 The system architecture diagram illustrates the entire system process, from environment-adaptive feature reconstruction to hybrid machine learning models, and then to the plateau-specific training mechanism.

[0131] Figure 2 The flowchart for thermodynamic model correction shows how to correct the convective heat transfer coefficient in the thermodynamic model based on the actual gas pressure.

[0132] Figure 3 The flowchart for multi-frequency EIS fusion features demonstrates how impedance data from different frequencies are fused to calculate the characteristic temperature. Figure 4 This diagram illustrates the lightweight hybrid machine learning model architecture, showcasing the input, processing, and output processes of three models (LightGBM, 1D-CNN, and physically constrained particle filter model), as well as the model fusion logic.

[0133] Figure 5 A flowchart of the plateau-specific training mechanism is provided, illustrating the three key training components: data augmentation, transfer learning, and online calibration. Detailed Implementation

[0134] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0135] As shown in the figure, this invention provides a method for monitoring the temperature of energy storage batteries in high-altitude areas, comprising the following steps:

[0136] 1. Environmental Adaptive Feature Reconstruction

[0137] (1) Thermodynamic model correction

[0138] ① Barometric Pressure Data Acquisition: Install a high-precision barometric pressure sensor in the energy storage system, ensuring that the sensor's measurement range covers the expected high-altitude barometric pressure range (typically 50 kPa to 101.3 kPa) and the measurement accuracy is better than 1 kPa. Connect the barometric pressure sensor to the data acquisition system, set the acquisition frequency to once per minute, collect the local barometric pressure value P in real time, and store it in the system database.

[0139] ② Determine the standard convective heat transfer coefficient: Under standard laboratory conditions (temperature 25℃, atmospheric pressure 101.3 kPa), the heat dissipation of the energy storage battery was tested. Different heat dissipation conditions were simulated by adjusting the wind speed, and the difference between the battery surface temperature and the ambient temperature, as well as the corresponding wind speed, were measured. The convective heat transfer coefficient under different wind speeds was calculated using heat transfer formulas. The most frequently occurring convective heat transfer coefficient value was selected as the convective heat transfer coefficient h under standard atmospheric pressure. Typically, the standard convective heat transfer coefficient for cylindrical batteries is approximately 10 W / (m²). 2 ·K).

[0140] ③ Dynamically calculate the corrected convective heat transfer coefficient: Read the local air pressure value P in real time from the system database and substitute it into the formula. Where P0 = 101.3 kPa. For example, when the local air pressure is monitored to be 70 kPa, the corrected convective heat transfer coefficient is calculated to be h. high-alt = 10 × (70 / 101.3) 0.7 ≈8.1W / (m 2 ·K). The calculated h high-alt The data is updated in real time to the thermodynamic model, providing accurate heat dissipation parameters for subsequent temperature monitoring.

[0141] (2) Multi-frequency EIS fusion feature extraction

[0142] ① EIS Data Acquisition System Setup: An electrochemical impedance spectroscopy (EIS) instrument is integrated into the energy storage system, with its measurement frequency range set to 10 Hz to 1 kHz, and three frequency points corresponding to 10 Hz (Z). 10 ), 100Hz (Z) 100 ) and 1kHz (Z 1000 Connect the impedance spectrometer to the battery testing system to ensure stable acquisition of battery impedance data under different operating conditions.

[0143] ② Impedance data acquisition process: During battery charge / discharge cycles, an EIS measurement is triggered every 30 minutes. Before each measurement, ensure the battery is in a static state for 5 minutes to reduce measurement errors. After acquiring the impedance value at each frequency point, record the current battery temperature, SOC, and other relevant parameters, and store them in a data file.

[0144] ③ Construct a three-dimensional impedance vector: Extract the impedance values ​​at 10Hz, 100Hz, and 1kHz frequencies from the acquired EIS data, denoted as Zi. 10 Z 100 and Z 1000 Construct a three-dimensional impedance vector [Z] 10 Z 100 Z 1000 For example, in a certain measurement, Z was obtained. 10=5+3jΩ, Z 100 =2+1jΩ and Z 1000 =1+0.5jΩ, then the impedance vector is [5+3j,2+1j,1+0.5j].

[0145] ④ Method for determining the mapping function: A battery testing platform was built in the laboratory to conduct EIS measurements and temperature monitoring experiments on the battery at multiple temperature points (-40℃, -20℃, 0℃, 25℃, 50℃). At each temperature point, a large amount of sample data was collected to establish the correspondence between impedance values ​​and temperature. Using a multiple linear regression method, the mapping function between impedance and temperature for each frequency band was fitted and obtained.

[0146] For example, for a 10Hz frequency point, the mapping function is f1(Z). 10 )=0.2×Re(Z 10 )+0.1×Im(Z 10 +10; For a frequency of 100Hz, the mapping function is f2(Z) + 10; 100 )=0.3×Re(Z 100 )+0.15×Im(Z 100 +8; For a frequency of 1kHz, the mapping function is f3(Z) + 8; 1k )=0.4×Re(Z 1k )+0.2×Im(Z 1k )+5. Where Re(Z) and Im(Z) represent the real and imaginary parts of the impedance, respectively.

[0147] ⑤ Application of Random Forest Feature Selection Algorithm: A large number of battery data samples covering different temperatures, SOCs, and operating conditions were collected and divided into training and testing sets. The random forest model was trained using the training set, with 100 decision trees in the forest and a maximum depth of 10 for each tree. By analyzing the classification results of each decision tree on the samples, the importance score of impedance features in each frequency band was calculated. Based on the importance score, the weight coefficients w1, w2, and w3 of the three frequency bands were dynamically adjusted. For example, at an ambient temperature of -20℃, the algorithm calculated w1 = 0.4, w2 = 0.3, and w3 = 0.3. These weight coefficients were then applied to the multi-frequency EIS fusion feature calculation.

[0148] ⑥ Fusion calculation process: During real-time monitoring, the impedance values ​​of each frequency band are substituted into the corresponding mapping function to obtain the temperature estimate of each frequency band.

[0149] For example, when Z 10 When =5+3jΩ, f1(Z) 10 )=0.2×5+0.1×3+10=11.3℃; when Z 100 When =2+1jΩ, f2(Z) 100)=0.3×2+0.15×1+8=9.15℃; when Z 1000 When =1+0.5jΩ, f3(Z) 1k The temperature is calculated as follows: 0.4 × 1 + 0.2 × 0.5 + 5 = 5.7℃. Then, a weighted average temperature is obtained by weighting the values ​​according to the weighting coefficients. The average temperature after fusion As the output of the environmental adaptive feature reconstruction, it is provided to the subsequent machine learning model for temperature monitoring.

[0150] 2. Lightweight Hybrid Machine Learning Model

[0151] (1) Model selection and division of labor

[0152] ①LightGBM model

[0153] Data preparation: Extract historical data from the energy storage system's database, including battery voltage X. a Current X b Self-discharge rate X c Health status X d Ambient temperature T o and the impedance vector Z generated by multi-frequency EIS fusion features vec The corresponding core temperature measurements were collected as labels. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0154] Model initialization: Install the LightGBM library in the Python environment and import the necessary modules. Set the basic parameters of the model, such as tree depth of 5, learning rate of 0.1, number of iterations of 100, and loss function of mean squared error (MSE).

[0155] Model Training: Input the training set data into the LightGBM model to start the training process. During training, the model automatically builds multiple decision trees, and the outputs of each tree are summed to obtain the final predicted value. By calculating the error between the predicted and actual values, the model parameters are continuously adjusted to optimize the model's performance. Validation is performed on the validation set, and metrics such as RMSE and MAE are monitored to prevent overfitting.

[0156] Model Evaluation and Optimization: The trained LightGBM model is evaluated using a test set, and the RMSE, MAE, and R-value are calculated. 2 The model's predictive performance is evaluated using metrics such as tree depth and learning rate. Based on the evaluation results, the model parameters are fine-tuned to further improve the model's accuracy and generalization ability.

[0157] ② 1D-CNN model

[0158] Data preparation: Extract time-series temperature data Ts and Z from the energy storage system database. 100 h high-alt Window data. The data undergoes preprocessing, including data cleaning and normalization. The processed dataset is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0159] Model Construction: Build a 1D-CNN model using TensorFlow or PyTorch frameworks. The model structure includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer. For example, the input layer size is [sequence_length, 1], the convolutional layers use 32 convolutional kernels of size 3, the pooling layers use max pooling of size 2, and the fully connected layers have 64 neurons.

[0160] Model Training: The training set data is input into the 1D-CNN model, with a learning rate of 0.001, and the Adam optimization algorithm is used for model training. During training, the model's weights and biases are adjusted using the backpropagation algorithm to minimize the error between predicted and actual values. Validation is performed on a validation set, monitoring metrics such as the model's loss function and accuracy to prevent overfitting.

[0161] Model Evaluation and Optimization: The trained 1D-CNN model is evaluated using a test set, and various performance metrics are calculated. Based on the evaluation results, the model structure is adjusted, such as increasing or decreasing the number of convolutional layers and adjusting the size of the convolutional kernels, to improve the model's prediction accuracy for the core temperature change rate under transient conditions.

[0162] ③ Physically constrained particle filter model

[0163] Data preparation: Collect limited observational data of the battery under ultra-low temperature scenarios where data is scarce, including observed values ​​of the battery core temperature and corresponding ambient temperature, air pressure, and other information. Simultaneously, prepare the battery's thermodynamic parameters, such as the thermal diffusivity α and the pressure-dependent heat dissipation coefficient β.

[0164] Model initialization: In MATLAB or Python, use the particle filter toolbox or write your own particle filter algorithm. Set the number of particles to 1000 and define the particle state vectors, including battery core temperature, ambient temperature, etc. Initialize the particle set, typically using a Gaussian distribution.

[0165] Model training and updating: The observation data is sequentially input into the particle filter model, and the model is updated according to the thermodynamic equations. Perform state prediction and use the weight update formula Update the particle weights. Through resampling, remove particles with lower weights and duplicate particles with higher weights, maintaining the diversity and representativeness of the particle set.

[0166] Model Evaluation and Optimization: The prediction results of the particle filter model are evaluated under a simulated cryogenic data-scarce scenario. Error metrics such as RMSE and MAE between predicted and actual values ​​are calculated. Based on the evaluation results, model parameters, such as the number of particles, process noise covariance, and observation noise covariance, are adjusted to improve the model's temperature estimation performance in data-scarce scenarios.

[0167] (2) Model fusion logic

[0168] Data Acquisition and Preprocessing System Integration: The energy storage system integrates multiple sensors, including voltage sensors, current sensors, SOC sensors, ambient temperature sensors, air pressure sensors, and electrochemical impedance spectroscopy, to ensure real-time and accurate acquisition of relevant battery data. Data from all sensors is uniformly connected to the data acquisition system for initial data cleaning and processing, removing outliers and noise. Then, data is synchronously acquired at preset time intervals (e.g., once per minute), and the acquired data is normalized to a value range between [0,1] to facilitate subsequent model processing and analysis.

[0169] Operating condition judgment rules are set: Based on the battery's charge and discharge characteristics, SOC change patterns, and relevant standards and specifications, detailed operating condition judgment rules are formulated. For example, when SOC > 50% and the current change rate is less than 0.5A / min, it is judged as a steady-state operating condition; when SOC ≤ 50% or the current change rate is greater than or equal to 0.5A / min, it is judged as a transient operating condition. Simultaneously, a processing mode for ultra-low temperature data loss scenarios is activated when the ambient temperature is below -40℃ or when there is sensor failure.

[0170] Model Fusion Module Development: This section describes the development of a model fusion module that integrates the outputs of the LightGBM, 1D-CNN, and Physically Constrained Particle Filter (PCP) models. The module defines weight coefficients for each model, which can be dynamically adjusted based on performance under different conditions and data reliability. For example, under steady-state conditions, the LightGBM model receives higher weights; under transient conditions, the 1D-CNN model receives more weights; and in scenarios with limited cryogenic data, the PCP model's output is primarily considered. The final core temperature estimate is calculated using a weighted average method. This data is then displayed in real time on the monitoring interface as the output of the entire monitoring system, providing decision support for the operation and management of the energy storage system.

[0171] 3. High-altitude-specific training mechanism

[0172] (1) Data Augmentation

[0173] ① Data Collection and Preprocessing: In high-altitude areas, collect a small amount of real battery operation data, including information such as battery voltage, current, temperature, and impedance. Preprocess the collected data, including data cleaning, missing value imputation, and outlier handling, to ensure data quality and integrity.

[0174] ② GAN Model Construction: A Generative Adversarial Network (GAN) is constructed using a deep learning framework (such as TensorFlow or PyTorch), consisting of two sub-networks: a generator and a discriminator. The generator takes a random noise vector as input and outputs a virtual data sample with the same format as the real data. The discriminator takes a data sample as input (which can be real data or virtual data generated by the generator) and outputs a discrimination result (representing the probability value of whether the data is real or virtual).

[0175] ③ Model Training Process: Preprocessed real-world plateau data is used as training data for the discriminator, while virtual data generated by the generator is also input into the discriminator. By alternately training the generator and discriminator, the generator continuously learns the distribution characteristics of the plateau environment data, generating virtual data that becomes increasingly similar to the real data. During training, the Jensen-Shannon divergence between the generated data and the real data is monitored. Training is stopped when the divergence gradually decreases and the quality of the generated data reaches a high level.

[0176] ④ Virtual Data Fusion and Application: The generated virtual data is fused with real data at a certain ratio (e.g., 1:1 or 2:1) to expand the training dataset. The fused dataset is then used to train a lightweight hybrid machine learning model, improving the model's generalization ability and adaptability in high-altitude environments, especially enhancing the model's robustness and reliability under extreme conditions where data is scarce.

[0177] (2) Transfer learning

[0178] ① Model Pre-training on Plain Data: In plain areas, a large amount of battery operating data was collected, including input and output data under different operating conditions, temperatures, and SOCs. This data was used to pre-train lightweight hybrid machine learning models (including LightGBM, 1D-CNN, and physically constrained particle filter models) to enable the model to learn the temperature variation patterns and feature representations of batteries under general environmental conditions. During pre-training, reasonable training parameters were set, such as learning rate, batch size, and number of iterations, to ensure the model achieves good fitting performance and high prediction accuracy on plain data.

[0179] ② Model Transfer and Parameter Fine-tuning: The pre-trained model is transferred to the plateau region, and fine-tuned using a small sample dataset collected from the plateau. During fine-tuning, the focus is on adjusting feature weights and biases relevant to the plateau environment, such as the corrected convective heat transfer coefficient and the weight coefficients of multi-frequency EIS fusion features. Simultaneously, a domain adaptation loss function is added. This allows the model to better adapt to the data distribution in the high-altitude environment. During fine-tuning, the model's performance metrics, such as RMSE and MAE, are monitored. When the model's predictive performance on high-altitude data reaches its optimal level, the fine-tuning process is stopped, and the fine-tuned model parameters are saved.

[0180] ③ Domain Adaptation Optimization Strategy: In the transfer learning process, to better balance task loss and domain adaptation loss, a reasonable balancing parameter λ is set. In the early stages of training, a larger weight can be given to the domain adaptation loss, allowing the model to adapt more quickly to the data distribution in the high-altitude environment. As training progresses, the value of λ is gradually decreased, allowing the model to focus more on the predictive performance of the task itself. This approach achieves a smooth transition and performance optimization of the model in the high-altitude environment.

[0181] (3) Online calibration

[0182] ① Pulsed EIS Scan Trigger Mechanism: In the control strategy of the energy storage system, a pulsed EIS scan is automatically triggered during off-peak hours (such as 12:00 to 6:00 AM). During this period, the battery is in a relatively static or low-load operating state. A short pulse excitation signal (such as a square wave signal with an amplitude of 1V and a width of 10ms) is applied to the battery, and impedance data of the battery at different frequencies is collected.

[0183] ② Data Acquisition and Preprocessing Process: During each pulsed EIS scan, impedance data of the battery's current state is acquired in real time, including impedance values ​​at 10Hz, 100Hz, and 1kHz frequencies, as well as real-time ambient temperature data. The acquired impedance data is preprocessed, including noise removal and data smoothing, to ensure data accuracy and reliability.

[0184] ③ Fitting coefficient update method: Using preprocessed impedance data and ambient temperature as input features, and corresponding battery temperature measurements as observations, an impedance-temperature fitting model is established using the least squares method. New fitting coefficients are obtained by minimizing the sum of squared residuals. For example, in a certain online calibration, the new fitting coefficients are [0.22, 0.35, 0.43], which will be used for subsequent multi-frequency EIS fusion feature calculations.

[0185] ④ Model Parameter Update and Feedback: The updated impedance-temperature fitting coefficients are fed back into the lightweight hybrid machine learning model to update relevant parameters. In the LightGBM model, the weight coefficients of the multi-frequency EIS fusion features are updated; in the 1D-CNN model, the convolution kernel weights related to impedance features are adjusted; and in the physically constrained particle filter model, parameters related to heat dissipation are corrected. In this way, the model can reflect the battery's temperature characteristics in real time under its current state, ensuring that the model's prediction accuracy and reliability remain at a high level over time.

[0186] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several adjustments and improvements without departing from the core concept of the present invention, and these adjustments and improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring the temperature of energy storage batteries in high-altitude areas, characterized in that, include: S1. The thermodynamic model is corrected using the environmental adaptive feature reconstruction module, and multi-frequency EIS fusion features are extracted. The thermodynamic model is corrected by introducing a pressure correction coefficient to adjust the convective heat transfer coefficient. Multi-frequency EIS fusion features select impedance values ​​at specific frequencies to construct vectors, and dynamically adjust the weights using a random forest algorithm. S2. Construct a lightweight hybrid machine learning model, including LightGBM, 1D-CNN and physically constrained particle filter model, to predict the core temperature of the battery. S3. The model is trained and optimized using a plateau-specific training mechanism, including data augmentation, transfer learning, and online calibration.

2. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: In step S1, the environmental adaptive feature reconstruction module corrects the convective heat transfer coefficient in the cylindrical battery thermal conduction equation by introducing a pressure correction coefficient, as shown in the following formula: Among them, h high-alt represents the corrected convective heat transfer coefficient; h represents the convective heat transfer coefficient under standard atmospheric pressure; P represents the local actual air pressure; P0 represents standard atmospheric pressure.

3. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: In step S1, the environment adaptive feature reconstruction module selects three feature frequencies: 10Hz imaginary part, 100Hz real part, and 1kHz real part, to construct a three-dimensional impedance vector [Z]. 10 Z 100 Z 1000 The weights of each frequency band are dynamically adjusted using a random forest feature selection algorithm, and the fusion formula is as follows: in, Z represents the average temperature after fusion; w1, w2, and w3 represent the weighting coefficients of the three frequency bands, respectively; f1, f2, and f3 are the mapping functions of impedance and temperature for the corresponding frequency bands, respectively. 10 Z 100 Z 1000 These represent the impedance values ​​at frequencies of 10Hz, 100Hz, and 1kHz, respectively.

4. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: In the lightweight hybrid machine learning model of step S2, the LightGBM model is suitable for steady-state conditions, and the input feature is (X a ,X b ,X c ,X d ,T o Z vec ,h high-alt ), where X a X is the battery voltage. b X is the battery current. c X represents the battery's self-discharge rate. d For battery health status, T o For ambient temperature, h represents the average temperature after fusion. high-alt The corrected convective heat transfer coefficient; output core temperature estimate. LightGBM is a gradient boosting decision tree algorithm that constructs multiple decision trees. The output of each decision tree is the average temperature of the corresponding leaf node, as shown in the formula: Where K is the number of decision trees, f k (x i ) represents the input feature x of the k-th decision tree. i The output; The 1D-CNN model is designed for transient conditions, with input time series data T. s Z 100 h high-alt Output core temperature change rate estimate 1D-CNN extracts local features by sliding the convolutional kernels of convolutional layers across the input temporal data. The convolution operation formula is as follows: Among them, w i These are the convolution kernel weight parameters, x j + i Let b be the value of the input time series data at position j+i, b be the bias term, and σ be the activation function; Physically constrained particle filter models are used to address scenarios with missing ultra-low temperature data, with input h. high-alt Output the predicted battery core temperature value Core temperature estimation based on thermodynamic equations and observational data: in, Let be the core temperature of the battery at time k, and α be the thermal diffusivity. It represents the temperature gradient change, Δt is the time step, β is the heat dissipation coefficient related to air pressure, P0 is the standard atmospheric pressure, and P is the local actual air pressure.

5. The high-altitude energy storage battery temperature monitoring system according to claim 4, characterized in that: The selection method for the LightGBM, 1D-CNN, and physically constrained particle filter models is based on real-time data acquisition of battery SOC, current, ambient temperature, etc., specifically as follows: (1) When SOC>50% and current change rate is less than 0.5A / min, it is determined to be a steady-state condition, and the LightGBM model is selected for temperature prediction. (2) When SOC≤50% or current change rate is greater than or equal to 0.5A / min, it is determined to be a transient condition, and a 1D-CNN model is selected for prediction; (3) When the ambient temperature is below -40℃ or the sensor malfunction causes data loss, the ultra-low temperature data loss scenario is entered, and the physical constraint particle filter model is selected for temperature prediction.

6. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: In the data augmentation process of step S3, a generative adversarial network is used to generate virtual data in a plateau environment to expand the training dataset.

7. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: In the transfer learning process of step S3, the model is first pre-trained on plain data, then the pre-trained model is transferred to the plateau region, and the model parameters are fine-tuned using a small sample dataset from the plateau.

8. The high-altitude energy storage battery temperature monitoring system according to claim 1, characterized in that: During the online calibration process in step S3, a pulsed EIS scan is automatically triggered during off-peak hours each day to update the impedance-temperature fitting coefficients and feed them back into the lightweight hybrid machine learning model to update the model parameters.

9. A temperature monitoring system for energy storage batteries in high-altitude areas, characterized in that, To implement the method of any one of claims 1-8, comprising: The environmental adaptive feature reconstruction module is used to correct the thermodynamic model and extract multi-frequency EIS fusion features. The thermodynamic model is corrected by introducing a pressure correction coefficient to adjust the convective heat transfer coefficient. The multi-frequency EIS fusion features select impedance values ​​at specific frequencies to construct vectors and dynamically adjust the weights using a random forest algorithm. A lightweight hybrid machine learning model module, including LightGBM, 1D-CNN, and physically constrained particle filter models, is used to predict battery core temperature. A plateau-specific training mechanism module is used for training and optimizing the model.

Images