An energy storage battery state monitoring method and system for high-cold high-altitude environment

Abstract

The application provides a kind of energy storage battery state monitoring method and system for high-cold high-altitude environment, innovatively constructs multi-scale physical constraint framework: through the hollow convolution network with time sequence perception ability, multi-dimensional features are extracted, and the physical information of the initial state of the battery is fused;Embedding the inherent physical law constraint of the battery in the feature space, including forced charge conservation, health state one-way attenuation and temperature and pressure adaptive stability control;Design a special compensation mechanism for extreme environments, achieve accurate state estimation by dynamically correcting low-temperature diffusion behavior, and establish a multi-condition cooperative triggering lithium analysis safety warning. Adopting a calibration-free training strategy, only a small amount of boundary data is required to complete model optimization, combined with plateau environment data enhancement to improve adaptability. After light processing, it can run on a resource-limited embedded platform. In the high-cold low-pressure working condition, it effectively solves the core industry problems such as insufficient adaptability in extreme environment, inaccurate physical law and limited hardware deployment.

Description

Technical Field

[0001] This invention belongs to the field of battery management system (BMS) technology, specifically relating to a method and system for monitoring the state of energy storage batteries in cold and high-altitude environments. Background Technology

[0002] With the increasing global demand for clean energy, the development of renewable energy sources such as wind and solar power in high-altitude and cold regions is receiving increasing attention. In these areas, energy storage batteries, as key devices for energy storage and stable power supply, directly affect the reliable operation of the entire energy system. However, the unique characteristics of the high-altitude environment present numerous severe challenges to the state monitoring of energy storage batteries. High-altitude and cold regions are typically characterized by extremely low temperatures (below -20°C) and thin air (pressure below 65 kPa), which significantly impact the physical and chemical properties of batteries. Low temperatures increase electrolyte viscosity, reduce the lithium-ion diffusion coefficient, and lead to increased internal impedance, thereby affecting the battery's charge-discharge performance and the accuracy of state estimation. Low air pressure alters the gas diffusion rate at the electrode-electrolyte interface, accelerating battery capacity decay. Furthermore, the large diurnal temperature range and frequent temperature fluctuations in high-altitude regions can increase thermal stress on battery materials, causing SEI film rupture and further exacerbating battery aging and performance degradation. Traditional battery state monitoring technologies have shown significant shortcomings in addressing these challenges of high-altitude environments. While electrochemical models are accurate, their high computational complexity makes them unsuitable for the real-time requirements of battery management systems (BMS). Equivalent circuit models, though reflecting the macroscopic characteristics of the battery to some extent, suffer from a significant drop in estimation accuracy due to neglecting the time-varying nature of parameters at extreme temperatures. Traditional data-driven models, often based on black-box structures, lack physical interpretability, and data acquisition in high-altitude environments is costly, making it difficult to obtain sufficient labeled data for training. These technical limitations hinder the effective application of existing battery state monitoring methods in high-altitude energy storage scenarios, necessitating a dedicated, efficient battery state monitoring solution specifically designed for high-altitude environments. Summary of the Invention

[0003] In high-altitude and frigid regions, state monitoring of energy storage batteries is crucial for ensuring the stable operation and safety of energy systems. Due to extreme climatic conditions and environmental factors, traditional battery state monitoring technologies face numerous challenges, often failing to accurately and promptly reflect the true state of the batteries. This invention aims to provide a state monitoring solution specifically designed for high-altitude and frigid environments to meet the unique battery monitoring needs of these regions.

[0004] 1. Improve monitoring accuracy: To address the impact of extreme environments in high-altitude and cold regions on battery status, a state monitoring method will be developed that can accurately monitor battery SOC, SOH, and lithium plating risk. This will ensure that the accuracy of SOC estimation and voltage prediction is significantly improved under harsh conditions such as low temperature and low air pressure, meeting the requirements for high-precision monitoring.

[0005] 2. Enhanced Real-Time Performance: Considering the stringent real-time requirements of high-altitude energy storage systems, an efficient state estimation architecture is designed to significantly shorten inference time, enabling the battery management system to respond promptly to changes in battery state and ensuring the stable operation of the energy storage system.

[0006] 3. Ensuring safety: By introducing lithium plating risk factors and early warning mechanisms, the system can reliably warn of lithium plating risks in low-temperature environments, effectively preventing battery damage caused by low-temperature lithium plating and improving the safety and lifespan of the energy storage system.

[0007] 4. Enhance adaptability: Construct an environmental adaptive compensation mechanism, including an Arrhenius diffusion kinetics compensation term and a high-altitude stability constraint, to significantly reduce the impact of low temperature and low air pressure on SOC / SOH estimation, enabling the monitoring method to adapt to the environmental differences in different high-altitude and cold regions.

[0008] 5. Reduce maintenance costs: By adopting a label-free training mechanism, the bottleneck of scarce labeled data in plateau scenes is overcome, reducing the dependence on a large amount of labeled data, reducing data collection and labeling costs, while improving the model's generalization ability and adaptability, and reducing the model retraining and maintenance work caused by environmental changes.

[0009] 6. Achieve lightweight deployment: The model is quantized and compressed to significantly reduce its size while ensuring monitoring performance, meeting the lightweight requirements of embedded systems, reducing hardware costs and power consumption, and making it suitable for long-term operation of remote energy storage power stations in high-altitude and cold regions.

[0010] The specific plan is as follows:

[0011] A method for monitoring the state of energy storage batteries in high-altitude and cold environments includes the following steps:

[0012] S1. Hollow Causal Convolutional Encoding: Collect four-dimensional time-series data of current, voltage, temperature and air pressure, extract features using a hollow causal convolutional encoder, construct multi-scale feature representation by exponentially increasing the expansion factor, and fuse the physical information of the battery's initial state.

[0013] S2. Construct a physical constraint potential space: Apply physical constraints, including SOC conservation constraints, SOH monotonicity constraints, and high-altitude stability constraints, to the potential space to ensure that the characteristics conform to the physical laws of batteries.

[0014] S3, Cold-Adaptive Decoding: An Arrhenius diffusion kinetics compensation term and a lithium plating risk factor are embedded in the cold-adaptive decoder. A label-free training mechanism including thermodynamic gradient loss and plateau characteristic loss is adopted to output SOC, SOH, voltage prediction value and lithium plating risk factor.

[0015] Furthermore, a multi-scale physical constraint framework is constructed, including:

[0016] (1) Dilated Causal Convolutional Encoder

[0017] Multi-scale features of four-dimensional time-series data (current, voltage, temperature, and air pressure) are extracted to provide effective feature representations for subsequent state estimation.

[0018] ①Structural Design: A 4-layer dilated causal convolution is used, with dilation factors of 1, 2, 4, and 8, forming an exponentially growing receptive field. The formula for calculating the receptive field of the 4th layer is:

[0019]

[0020] in, The kernel size is [size]. Let be the expansion factor of the i-th layer.

[0021] ② Channel attention mechanism: A channel attention module is added after each convolutional layer. The calculation method is as follows:

[0022]

[0023] By analyzing the input feature map Perform global average pooling ( ), to obtain the weight vector along the channel dimension, The activation function is then compared with the original feature map. Element-wise multiplication enhances the features of important channels and suppresses the features of unimportant channels.

[0024] ③ Physical Information Injection Layer: The initial SOC (SOC0) and initial SOH (SOH0) are injected as prior physical information into the feature map output by the encoder. The calculation method is as follows:

[0025]

[0026] Where W is a learnable parameter, This refers to feature concatenation, which integrates physical information into the feature representation output by the encoder.

[0027] (2) Physically constrained potential space

[0028] The physical constraints of the battery are incorporated into the feature representation to ensure that the state estimation results conform to the actual physical laws of the battery.

[0029] ①SOC Conservation Constraint: By using gradient-truncation backpropagation, the estimated SOC is ensured to satisfy the conservation law, i.e.:

[0030]

[0031] in, Let I be the change in the hidden state estimate of SOC, and I be the current. For time intervals, This represents the battery's nominal capacity. During the backpropagation process, if the error between the estimated value of the SOC change and the theoretical value calculated based on the coulomb counting principle exceeds 0.005, the gradient is truncated to prevent further error propagation, thereby ensuring the accuracy of the SOC estimate.

[0032] ②SOH Monotonicity Constraint: To ensure the physical law that SOH (state of health of the battery) decreases monotonically with time, the following implementation method is adopted:

[0033]

[0034] in, The hidden state estimate of SOH. The attenuation coefficient is... Let T be the activation function and T be the temperature. By introducing a temperature-related decay term during the SOH update process, the estimated SOH value is adjusted to decrease with each update. Simultaneously, the influence of temperature on the battery aging rate is considered to ensure that the SOH estimate is consistent with reality and monotonically decreasing.

[0035] ③ High-altitude stability constraint: To improve the model's stability in high-altitude environments, a norm constraint is applied to the feature representation of the latent space. The specific expression is as follows:

[0036]

[0037] in, The encoder output feature representation is used to integrate the initial state physical information, where T represents temperature and P represents air pressure. By introducing coefficients related to temperature and air pressure, the upper norm of the feature representation is dynamically adjusted, enabling the model to perform state estimation more stably in high-altitude and frigid environments and reducing the interference of environmental factors on the model output.

[0038] (3) Cold-adaptive decoder

[0039] The feature representations in the latent space that incorporate physical constraints are mapped back to specific battery state quantities (SOC, SOH, voltage, etc.), and adaptive adjustments are made for high-altitude and cold environments to output the final battery state estimate.

[0040] ① Diffusion kinetics compensation: To address the issue of reduced lithium-ion diffusion rate in cold environments, an Arrhenius diffusion kinetics compensation term is introduced into the SOC update model. The formula is as follows:

[0041]

[0042] in, This is the estimated SOC value at time k. Let be the current at time k. For time intervals, For efficiency coefficient, This refers to the battery's nominal capacity. For activation energy, The gas constant is Let E be the temperature at time k. Activation energy E a Segmented setup:

[0043]

[0044] By adjusting the activation energy value according to temperature changes, the diffusion behavior of lithium ions at different temperatures can be simulated more accurately, thereby improving the accuracy of SOC estimation.

[0045] ② Lithium plating risk factor: To provide early warning of lithium plating risk, the calculation method for the lithium plating risk factor is defined as follows:

[0046]

[0047] in, The minimum change in voltage. The average temperature. and These are learnable parameters. When the triggering condition is met:

[0048]

[0049] A level-three warning signal is issued to remind the system to take measures to prevent lithium plating and ensure the safe operation of the battery.

[0050] (4) Logical connections between the parts

[0051] A dilated causal convolutional encoder extracts features from four-dimensional time-series data, yielding feature representations containing multi-scale information. These feature representations are then passed to a physically constrained latent space, where the battery's physical constraints are incorporated, making the features more consistent with the battery's actual physical characteristics. The physically constrained feature representations are then input to a high-altitude adaptive decoder. The decoder adapts to the characteristics of the high-altitude environment, ultimately outputting estimated battery state values ​​(SOC, SOH, voltage, etc.). (Note: The input voltage of the dilated causal convolutional encoder is the system's fundamental data source, while the output voltage of the high-altitude adaptive decoder is the model's prediction of the battery's future state. Although both are related to the battery's voltage characteristics, they occupy different positions and play different roles in the system. The input voltage provides the model with direct information about the battery's current state and is one of the bases for state estimation; the output voltage is the model's prediction of the battery state, used to assist in SOC and SOH estimation, fault warning, and system control and optimization. Through this input-output correlation, the entire monitoring system can achieve real-time and accurate monitoring and prediction of battery state, ensuring the safe and stable operation of energy storage batteries in high-altitude and cold environments.)

[0052] Furthermore, a label-free training mechanism is constructed, specifically including:

[0053] By designing a special loss function and training the model with unlabeled data, the model can learn the intrinsic relationship between battery state and four-dimensional time series data, while meeting the physical constraints of the battery and the special requirements of the high-altitude environment.

[0054] (1) Composition of loss function

[0055] ① Voltage prediction loss ( The mean square error (MSE) is used to measure the difference between the predicted voltage and the actual voltage. The calculation formula is as follows:

[0056]

[0057] Where N is the sample size. and These are the predicted voltage and actual voltage of the i-th sample, respectively. This loss term is used to ensure that the model can accurately predict the battery voltage.

[0058] ② Coulomb count dynamic loss The dynamic changes in SOC are constrained by the coulomb counting principle, and the calculation formula is as follows:

[0059]

[0060] in, Let SOC be the change at the i-th time step. Let be the current at the i-th time step. This loss term helps improve the accuracy of the SOC estimation.

[0061] ③ Boundary constraint loss ( The accuracy of SOC estimation under boundary conditions is ensured by minimizing the difference between the predicted SOC and the boundary calibration data (OCV-SOC mapping at full filling / full discharging points). The calculation formula is as follows:

[0062]

[0063] Where M is the number of boundary calibration data points. For the predicted SOC and OCV of the j-th boundary point j Let SOC be the actual open-circuit voltage at the j-th boundary point. This loss term is used to calibrate the accuracy of the SOC estimate under extreme charge and discharge conditions.

[0064] ④ Thermodynamic gradient loss ( ): This is used to constrain the rate of change of voltage with respect to temperature to be consistent with the actual physical process. The calculation formula is:

[0065]

[0066] in, To predict the derivative of voltage with respect to temperature, This represents the rate of change of actual voltage with respect to temperature. This loss term allows the model to better adapt to the impact of temperature changes on battery voltage.

[0067] ⑤Loss of high-altitude characteristics ( ): Due to the special characteristics of the high-altitude environment, the rate of change of the constraint voltage with respect to air pressure follows a specific law, and the calculation formula is as follows:

[0068]

[0069] in, The voltage derivative with respect to air pressure is used to predict the voltage-to-air-pressure derivative, where T is the temperature. This loss term is used to improve the model's performance in high-altitude, low-pressure environments.

[0070] (2) Weighted optimization of loss function

[0071] The weighted sum of the above loss functions yields the final optimization objective:

[0072]

[0073] in, These are weighting coefficients used to balance the influence of various loss functions during the optimization process. By adjusting these weighting coefficients, the model can achieve optimal overall performance under different tasks and constraints.

[0074] (3) Logical relationship with the multi-scale physical constraint framework

[0075] The label-free training mechanism is closely integrated with a multi-scale physical constraint framework. By minimizing the loss function designed above, the model parameters within the framework are optimized. During training, when the model learns the mapping relationship between four-dimensional time-series data and battery state, it is constrained by various physical constraints in the physical constraint latent space, while also adapting to the special requirements of the high-altitude environment. In this way, the model can not only accurately estimate the battery state, but also ensure the physical rationality of the estimation results and their reliability in the high-altitude environment.

[0076] Furthermore, model training and optimization include:

[0077] (1) Data preprocessing

[0078] Preprocessing the raw data enhances its representativeness and adaptability, thereby improving the training effect of the model.

[0079] ① Plateau data augmentation strategy: Use pressure disturbance and low temperature compensation methods to augment plateau data.

[0080] Air pressure disturbance: Noise of varying amplitude is added based on air pressure levels; the lower the air pressure, the greater the added noise. The calculation formula is as follows:

[0081]

[0082] Where V represents the original voltage data and P represents the air pressure. The voltage data after adding noise.

[0083] In the plateau data augmentation process, `np.random.normal(size=V.shape)` is used to generate random noise with the same shape as the original voltage data `V`. `np.random.normal` is a function in the NumPy library used to generate random numbers that follow a normal distribution. It can simulate random fluctuations in natural phenomena, which are common in actual battery voltage measurements. The parameter `size=V.shape` specifies that the shape of the generated random array should be the same as the shape of the original voltage data `V`. For example, if the original voltage data is a one-dimensional array of shape `(100,)` (representing voltage measurements at 100 time points), then the generated random noise will also be a one-dimensional array of shape `(100,)`, where each element is a random number following a normal distribution. These random numbers have a normal distribution characteristic with a mean of 0 and a standard deviation of 1. In the code implementation, the amplitude of the random noise can be adjusted by multiplying by a coefficient to simulate noise interference of different intensities. By adding this random noise to the original voltage data, voltage measurement errors caused by environmental factors (such as electromagnetic interference, sensor accuracy limitations, etc.) in real-world applications can be simulated. This step increases the diversity of data, enabling the model to learn the characteristics of battery voltage changes in the presence of noise, thereby improving the model's robustness and generalization ability in practical applications.

[0084] Low temperature compensation: Temperature compensation is based on changes in air pressure. The calculation formula is as follows:

[0085]

[0086] Where T represents the raw temperature data and P represents the air pressure. The temperature data is after compensation. Simultaneously, the time-series data is downsampled, reducing the original 10Hz data to 1Hz to decrease the data volume and improve the model's real-time performance.

[0087] In the low-temperature compensation formula, It is the hyperbolic tangent function, defined as:

[0088]

[0089] The hyperbolic tangent function has the following properties: the output range of the hyperbolic tangent function is (-1, 1), and as the input x approaches positive infinity, Approaching 1; when the input x approaches negative infinity, Approaching -1; when the input x is 0, The value is 0. The hyperbolic tangent function is a smooth nonlinear function with good mathematical properties, capable of nonlinearly mapping input values.

[0090] In the low-temperature compensation formula, Its function is to make non-linear adjustments to the temperature compensation value based on changes in air pressure P. Specifically:

[0091] When the air pressure P is greater than 65 kPa, It is a positive value; as air pressure increases, The value gradually approaches 1. At this point, the temperature compensation value... Gradually approaching 15 means that under higher air pressure, the compensated temperature... It will be much lower than the actual temperature T.

[0092] When the air pressure P is less than 65 kPa, The value is negative, and as the air pressure decreases, The value gradually approaches -1. At this point, the temperature compensation value... Gradually approaching -15 means that under low air pressure conditions, the compensated temperature It will be much higher than the actual temperature T.

[0093] When the air pressure P equals 65 kPa, The temperature compensation value is 0, and the compensated temperature is... It equals the actual temperature T.

[0094] This nonlinear adjustment better simulates the complex relationship between air pressure and temperature in high-altitude and frigid environments, making the compensated temperature more consistent with the actual operating state of the battery. In this way, temperature can be reasonably compensated under different air pressure conditions, improving the accuracy of battery status monitoring.

[0095] ② Data Normalization: The four-dimensional time-series data (current, voltage, temperature, and air pressure) are normalized to ensure their values ​​fall within the range of [-1, 1]. The normalization formula is:

[0096]

[0097] Where X represents the original data. and These are the minimum and maximum values ​​of the data, respectively. This is the normalized data. Normalization accelerates model convergence and improves training efficiency.

[0098] (2) Boundary marking protocol

[0099] The model is calibrated using OCV-SOC mapping data of the battery under full charge and full discharge conditions to improve the accuracy of SOC estimation.

[0100] ① OCV-SOC mapping data acquisition: OCV-SOC mapping data are collected at the battery's full charge and full discharge points respectively.

[0101] Charge the circuit to 4.2V using a constant current, then let it stand for 2 hours. Record the open circuit voltage (OCV) and the corresponding SOC value (SOC = 100%) at this time.

[0102] Discharge at constant current to 2.8V, then let stand for 2 hours, and record the open circuit voltage (OCV) and the corresponding SOC value (SOC = 0%) at this time.

[0103] ② Data Usage Strategy: The collected OCV-SOC mapping data is used as boundary calibration data and is only used during the training process, accounting for no more than 6-7% of the training cycle. In this way, without relying on a large amount of labeled data, a small amount of boundary data is used to constrain and calibrate the model, thereby improving the accuracy of the model's state estimation under boundary conditions.

[0104] (3) Model training process

[0105] By iteratively updating the model parameters through optimization algorithms, the model can achieve optimal performance on the training data.

[0106] ① Training parameter settings: The Adam optimization algorithm was used for model training, with an initial learning rate of 0.001, a batch size of 128, and 1000 training epochs. During training, the learning rate was decayed every 100 epochs with a decay coefficient of 0.1 to improve the model's convergence performance.

[0107] ② Regularization techniques: To prevent overfitting, L2 regularization and dropout techniques are used during model training. The weight coefficient for L2 regularization is 0.0001, and the probability of dropout is 0.2-0.5 (adjusted according to different layers of the model). Regularization techniques improve the model's generalization ability, enabling it to better adapt to different working conditions and environmental conditions in practical applications.

[0108] ③ Training Monitoring and Evaluation: During training, the model's loss function value and various performance metrics (such as SOC estimation error, voltage prediction MAE, lithium plating warning F1 score, etc.) are monitored in real time, and the model is evaluated using a validation set. Training is stopped early when the performance metrics on the validation set no longer improve to avoid overfitting. Simultaneously, model parameters and intermediate results are recorded during training to facilitate subsequent model optimization and debugging.

[0109] (4) Logical connection with the preceding sections

[0110] Data preprocessing provides high-quality training data for model training, while the boundary calibration protocol provides crucial boundary information for calibrating the SOC estimate. The model training process utilizes this data and information to iteratively update the model parameters within the multi-scale physical constraint framework by optimizing the loss function in the unlabeled training mechanism. This allows the model to gradually learn the complex mapping relationship between battery state and four-dimensional time-series data, while simultaneously meeting the physical constraints and the specific requirements of the high-altitude environment. The entire training process is a continuous process of optimization and adjustment, ultimately enabling the model to accurately monitor the multi-state of energy storage batteries in high-altitude and frigid environments.

[0111] Furthermore, embedded deployment includes:

[0112] (1) Model quantization and compression

[0113] The trained model is quantized and compressed so that it can run efficiently on resource-constrained embedded devices.

[0114] ① Quantization Method: Post-Training Quantization (PTQ) technology was used to quantize the model from FP32 precision to INT8 precision. By collecting statistical information from the input data, the quantization parameters for each layer (such as quantization range, quantization step size, etc.) were determined, and floating-point weights and activation values ​​were converted into integer representations, thereby reducing the model's storage space and computational load. The quantized model size was compressed from 15MB to 98KB, achieving a compression ratio of approximately 153:1, meeting the lightweight requirements of embedded devices.

[0115] ② Accuracy Calibration: During the quantization process, the Histogram Calibration method is used to calibrate the model's accuracy. By analyzing the distribution of the input data, the selection of quantization parameters is optimized to minimize accuracy loss during quantization. After calibration, the quantized model maintains a small performance gap compared to the original FP32 model across various metrics, ensuring that the model's performance on embedded devices is consistent with its performance in the training environment.

[0116] (2) Deployment platform and optimization

[0117] Choose a suitable embedded platform and optimize the model and system to ensure that the model can run efficiently and stably in practical applications.

[0118] ① Deployment platform selection: The STM32H7 series MCU was selected as the embedded deployment platform. This series of MCUs has a high-performance Cortex-M7 core, abundant storage resources and a variety of peripheral interfaces, which can meet the computational requirements of the model and the communication requirements with the battery management system.

[0119] ② Model Optimization and Porting: Based on the hardware architecture characteristics of the STM32H7, the quantized model was further optimized and ported.

[0120] The Neon instruction set is used to accelerate key operations such as convolution and matrix multiplication in the model, thereby improving the model's inference speed.

[0121] The memory layout of the model is optimized by using data reuse and caching optimization techniques to reduce data read and write operations and memory usage.

[0122] The model is converted into a binary format suitable for running on an MCU, and the corresponding driver and interface code are written to achieve seamless integration between the model and the battery management system.

[0123] ③ Real-time performance guarantee: By optimizing the model structure and algorithm implementation, the inference time of the model on the STM32H7 series MCU is ensured to be no more than 50ms, meeting the stringent real-time requirements of the battery management system. Simultaneously, buffer management is implemented for the model's input and output data to guarantee data continuity and real-time performance.

[0124] (3) Early warning triggering logic and system integration

[0125] Implement a lithium plating risk warning function in an embedded system and integrate the entire state monitoring system with the battery management system to achieve real-time monitoring and control of battery status.

[0126] ① Warning triggering logic: Implement lithium plating risk warning triggering logic in the embedded system:

[0127] When lithium plating risk factors And the current temperature And the rate of temperature change At that time, a Level III warning signal was triggered.

[0128] The warning signal is sent to the upper-level monitoring system through the communication interface of the battery management system, and corresponding measures are taken locally, such as reducing the charging current (by 50%), to prevent lithium plating and ensure the safe operation of the battery.

[0129] ② System Integration and Testing: Integrate the condition monitoring model with the battery management system, completing hardware connections and software interface debugging. Conduct system testing in a real high-altitude and cold-weather energy storage power station environment to verify the model's performance and reliability under real operating conditions. Through comparative analysis with actual battery data, further optimize model parameters and system configuration to improve the overall system performance and stability.

[0130] (4) Logical connection with the preceding sections

[0131] Embedded deployment is a crucial step in applying trained models to real-world monitoring scenarios for energy storage batteries in high-altitude and cold-weather environments. Model quantization and compression ensure the model can run on embedded devices, while deployment platform optimization improves the model's efficiency and stability in embedded environments. Early warning triggering logic and system integration realize the model's practical application, enabling it to provide real-time status monitoring and risk warning functions for battery management systems. The entire embedded deployment process serves as a bridge from model training to practical application, ensuring the model can fulfill its intended role and solve the challenges of energy storage battery status monitoring in high-altitude and cold-weather environments.

[0132] A state monitoring system for energy storage batteries designed for high-altitude and cold environments includes:

[0133] Hollow causal convolutional encoder: By collecting four-dimensional time-series data of current, voltage, temperature and air pressure, features are extracted using a hollow causal convolutional encoder. The expansion factor grows exponentially to construct a multi-scale feature representation and fuses the physical information of the battery's initial state.

[0134] Physical constraint processing module: Apply physical constraints, including SOC conservation constraints, SOH monotonicity constraints, and high-altitude stability constraints, to the feature representation output by the encoder that has fused the initial state physical information, to ensure that the features conform to the physical laws of the battery;

[0135] High-altitude cold-adaptive decoder: An Arrhenius diffusion kinetics compensation term and a lithium plating risk factor are embedded in the high-altitude cold-adaptive decoder, and a label-free training mechanism including thermodynamic gradient loss and plateau characteristic loss is adopted to output SOC, SOH, voltage prediction value and lithium plating risk factor.

[0136] The beneficial effects of this invention are as follows:

[0137] This solution surpasses traditional methods in terms of state estimation accuracy, safety assurance capabilities, and engineering applicability, effectively addressing core industry challenges such as insufficient adaptability to extreme environments, inaccurate physical laws, and limited hardware deployment.

[0138] Specifically,

[0139] 1. By constructing a multi-scale physical constraint framework, the accuracy and reliability of battery state estimation are significantly improved, ensuring that it conforms to physical laws under extreme environments and effectively reducing estimation errors.

[0140] 2. The compensation mechanism and lithium plating risk warning in the high-altitude cold adaptability decoder significantly improve the model's adaptability and safety in low-temperature environments and extend battery life.

[0141] 3. By adopting a label-free training mechanism, the bottleneck of scarce labeled data in plateau scenes is overcome, the data collection cost is reduced, and the generalization ability and adaptability of the model are improved.

[0142] 4. Through model quantization and compression, lightweight deployment of the model was achieved, reducing hardware costs and power consumption, and meeting the operational requirements of embedded platforms in high-altitude and cold regions.

[0143] 5. The early warning triggering logic and system integration enable real-time early warning of lithium plating risk and integration with the battery management system, thereby improving the safety and stability of the energy storage system and ensuring the safe operation of the battery. Attached Figure Description

[0144] Figure 1 This is a flowchart of a multi-state monitoring method for energy storage batteries.

[0145] Figure 2 This is a flowchart of the model training process.

[0146] Figure 3 This is a diagram of an embedded deployment architecture. Detailed Implementation

[0147] 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.

[0148] As shown in the figure, the invention provides a method for monitoring the state of energy storage batteries in high-altitude and cold environments, including the following steps:

[0149] 1. Data Acquisition and Preprocessing

[0150] (1) Data collection

[0151] ① Equipment Selection and Installation: Select high-precision current, voltage, temperature, and air pressure sensors and correctly install them in the corresponding positions on the battery pack. The data acquisition system should be able to synchronously acquire data from these sensors and transmit it to the BMS main control unit via CAN bus or Ethernet.

[0152] Current sensor: A high-precision Hall current sensor is selected and installed in the charging and discharging circuit of the battery pack.

[0153] Voltage sensor: Employs a high-precision voltage acquisition chip, installed at both the positive and negative terminals of the battery pack.

[0154] Temperature sensor: A PT100 platinum resistance temperature sensor is selected and installed in different locations inside the battery pack.

[0155] Barometric pressure sensor: A high-precision barometric pressure sensor is used and installed near the battery pack.

[0156] Data acquisition system: Configured with an embedded data acquisition system, responsible for synchronously acquiring and initially processing the data collected by the sensors.

[0157] ② Data Synchronization and Transmission: A synchronization trigger circuit is designed to ensure time synchronization of sensor data. Data verification mechanisms (such as CRC checksum) are used to ensure the integrity and accuracy of transmitted data.

[0158] (2) Data preprocessing

[0159] ① Plateau data enhancement:

[0160] The formula for air pressure disturbance is Where V is the original voltage data and P is the air pressure. This is the voltage data after noise has been added.

[0161] The formula for low temperature compensation is: T is the raw temperature data, and P is the air pressure. This is the compensated temperature data.

[0162] Temporal downsampling reduces the original 10Hz data to 1Hz, retaining the first point out of every 10 sampling points to reduce the amount of data and improve the model's real-time performance.

[0163] ② Data normalization: Normalize the four-dimensional time series data so that its value range is between [-1, 1]. The formula is as follows: Where X is the original data, and These are the minimum and maximum values ​​of the data, respectively.

[0164] 2. Model Building and Training

[0165] (1) Construction of multi-scale physical constraint framework

[0166] ①Diffuse causal convolutional encoder:

[0167] A four-layer dilated causal convolution is used, with dilation factors of 1, 2, 4, and 8. The formula for calculating the receptive field of the fourth layer is: , It is the kernel size. It is the expansion factor of the i-th layer.

[0168] The channel attention mechanism extracts channel features through global average pooling, and then multiplies them with the original feature map to enhance important features. The formula is as follows: , It is the input feature map. It is an activation function.

[0169] The physical information injection layer injects the initial SOC and SOH as prior information into the feature map output by the encoder. The formula is as follows: W is a learnable parameter. This indicates feature splicing.

[0170] ②Physically constrained potential space:

[0171] The SOC conservation constraint calculates the error between the predicted change in SOC and the theoretical value, which is used as part of the loss function. The formula is as follows: , It is the change in SOC at the i-th time step. is the current at the i-th time step, and N is the number of samples.

[0172] The SOH monotonicity constraint corrects the hidden state estimate of SOH based on temperature and the decay coefficient, ensuring that SOH monotonically decreases. The formula is... , This is the hidden state estimate of SOH. It is the attenuation coefficient. It is the activation function, and T is the temperature.

[0173] The Frobenius norm of the characteristic matrix is ​​calculated under high-altitude stability constraints. The upper limit of the norm is adjusted based on temperature and air pressure, and the characteristic matrix is ​​then scaled. The formula is as follows: , The feature representation of the encoder output is fused with the physical information of the initial state, where T is temperature and P is air pressure.

[0174] ③ Cold-weather adaptability decoder:

[0175] It adopts a multi-task output structure, simultaneously outputting the SOC estimate, SOH estimate, voltage prediction, and lithium plating risk factor.

[0176] SOC decoding introduces an Arrhenius diffusion kinetics compensation term. The formula is as follows: , This is the estimated SOC value at time k. It is the current at time k. It is a time interval. It is the efficiency coefficient. It is the battery's nominal capacity. It is activation energy. It is the gas constant. It is the temperature at time k. Activation energy. Segmentation set as .

[0177] The lithium plating risk factor calculation considers factors such as the minimum voltage variation, average temperature, and atmospheric pressure of the battery. The formula is... , It is the minimum change in voltage. It is the average temperature. and It is a learnable parameter. When the triggering condition is met... At that time, a level-three warning signal was issued.

[0178] (2) Unlabeled training mechanism

[0179] ① Definition of loss function:

[0180] Voltage prediction loss is measured using mean square error (MSE) to determine the difference between the predicted and actual voltage. The formula is as follows: , It is the predicted voltage of the i-th sample. is the actual voltage of the i-th sample, and N is the number of samples.

[0181] , It is the change in SOC at the i-th time step. It is the current at the i-th time step.

[0182] The boundary constraint loss maps the predicted SOC to OCV and then calculates the absolute error by comparing it with the actual OCV data. The formula is as follows: , The SOC is predicted for the j-th boundary point. is the SOC corresponding to the actual open-circuit voltage at the j-th boundary point, and M is the number of boundary calibration data points.

[0183] The absolute error between the predicted voltage derivative with respect to temperature and the actual voltage with respect to the rate of change of temperature is calculated using the thermodynamic gradient loss. The formula is as follows: , It predicts the voltage derivative with respect to temperature. It is the rate of change of actual voltage with respect to temperature.

[0184] The absolute error between the predicted voltage-pressure derivative and the voltage-pressure change rate calculated by the theoretical model is determined by the plateau characteristic loss calculation. The formula is as follows: , It is the predicted voltage derivative with respect to air pressure, where T is temperature.

[0185] Comprehensive Loss Function: The optimization objective is formed by weighted summation of the various loss functions. The formula is as follows: , It is the weighting coefficient.

[0186] ② Training process:

[0187] Initialize the encoder, physical constraint latent space, decoder, and loss function, and use the Adam optimization algorithm for training.

[0188] During the training loop, forward propagation is performed batch by batch, the encoder output is calculated and processed through the physically constrained latent space.

[0189] The processed features are input into the decoder to obtain the battery state prediction value.

[0190] Calculate the comprehensive loss function and update the model parameters through backpropagation.

[0191] After every 100 training rounds, the learning rate is decayed to improve the model's convergence performance.

[0192] 3. Embedded Deployment and System Integration

[0193] (1) Model quantization and compression

[0194] Model quantization: The model is quantized from FP32 precision to INT8 precision using Post-Training Quantization (PTQ) technology. By collecting statistical information from the input data, quantization parameters are determined, and floating-point weights and activation values ​​are converted to integers to reduce storage space and computation.

[0195] Model compression: Pruning the quantized model by removing unimportant weights, and then retraining to restore performance.

[0196] (2) Deployment platform and optimization

[0197] Embedded platform configuration: The STM32H7 series MCU is selected as the embedded deployment platform. It has a high-performance Cortex-M7 core, abundant storage resources and a variety of peripheral interfaces, which can meet the needs of model calculation and communication with BMS.

[0198] Model optimization and porting: The Neon instruction set is used to accelerate critical model operations, and memory layout is optimized to reduce data read / write operations and memory usage. The quantized model is converted to a binary format suitable for MCU operation, and driver and interface code are written to achieve seamless integration between the model and the BMS. Buffer management is implemented for the model's input and output data to ensure data continuity and real-time performance.

[0199] (3) Early warning triggering logic and system integration

[0200] Warning Trigger Logic: In the embedded system, when the lithium plating risk factor is greater than 0.5 and the current temperature is below -15℃ with a temperature change rate greater than 1℃ / min, a level three warning signal is triggered. The warning signal is sent to the upper-level monitoring system through the BMS communication interface, and local measures are taken, such as reducing the charging current to prevent lithium plating.

[0201] System Integration and Testing: Integrate the condition monitoring model with the BMS, completing hardware connections and software interface debugging. Conduct system testing in a real-world high-altitude and cold-weather energy storage power station environment to verify the model's actual performance and reliability. By comparing with actual battery data, further optimize model parameters and system configuration to improve system performance and stability.

[0202] 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 state of energy storage batteries in high-altitude and cold environments, characterized in that, Includes the following steps: S1. Hollow Causal Convolutional Encoding: Collect four-dimensional time-series data of current, voltage, temperature and air pressure, extract features using a hollow causal convolutional encoder, construct multi-scale feature representation by exponentially increasing the expansion factor, and fuse the physical information of the battery's initial state. S2. Construct a physical constraint latent space: In the latent space, apply physical constraints, including SOC conservation constraints, SOH monotonicity constraints, and high-altitude stability constraints, to the feature representation output by the encoder that has fused the initial state physical information, to ensure that the features conform to the physical laws of the battery. S3, Cold-Adaptive Decoding: An Arrhenius diffusion kinetics compensation term and a lithium plating risk factor are embedded in the cold-adaptive decoder, and a label-free training mechanism including thermodynamic gradient loss and plateau characteristic loss is adopted to output SOC, SOH, voltage prediction value and lithium plating risk factor. In step S1, the dilation factors are 1, 2, 4, and 8, respectively. The formula for calculating the receiving field of the 4th layer of the dilated causal convolutional encoder is as follows: in, The kernel size is [size]. Let be the expansion factor of the i-th layer; A channel attention module is added after each convolutional layer, calculated as follows: By analyzing the input feature map Perform global average pooling This yields the weight vector along the channel dimension. The activation function is then compared with the original feature map. Element-wise multiplication enhances the features of important channels and suppresses the features of unimportant channels. The initial SOC and initial SOH are injected as initial state physical information into the feature map output by the encoder, and the calculation method is as follows: Where SOC0 is the initial SOC, SOH0 is the initial SOH, and W is the learnable parameter. This represents feature concatenation, in which the physical information of the initial state is incorporated into the feature representation output by the encoder; In step S2, the SOC conservation constraint is ensured through gradient-truncation backpropagation to guarantee that the estimated SOC satisfies the conservation law. The calculation formula is as follows: in, Let I be the change in the hidden state estimate of SOC, and I be the current. For time intervals, The nominal capacity of the battery is used. During the reverse propagation process, if the error between the estimated value of the SOC change and the theoretical value calculated based on the coulomb counting principle exceeds 0.005, the gradient is truncated to prevent the error from propagating further, thereby ensuring the accuracy of the SOC estimate. In step S2, the SOH monotonicity constraint is used to ensure the physical law that SOH decreases monotonically with time, and its calculation formula is as follows: in, The hidden state estimate of SOH. The attenuation coefficient is... Here, T is the activation function, and T is the temperature. In step S2, the high-altitude stability constraint is used to improve the stability of the model in high-altitude environments. It applies a norm constraint to the feature representation of the latent space, and its calculation formula is as follows: in, The feature representation of the encoder output, which integrates the physical information of the initial state, is T for temperature and P for air pressure; In step S3, an Arrhenius diffusion dynamics compensation term is introduced into the SOC update model to achieve SOC decoding. The calculation formula is as follows: in, This is the estimated SOC value at time k. Let be the current at time k. For time intervals, For efficiency coefficient, This refers to the battery's nominal capacity. For activation energy, The gas constant is Let be the temperature at time k; In step S3, the lithium plating risk factor is used to provide early warning of lithium plating risk, and its calculation formula is as follows: in, This is the minimum voltage change. The average temperature is W, and W and b are learnable parameters. In step S3, the formula for calculating the comprehensive loss function of the unlabeled training mechanism is as follows: in, For voltage prediction loss, For Coulomb count dynamic loss, For boundary constraint loss, For thermodynamic gradient loss, Due to the loss of high-altitude characteristics, These are the weighting coefficients; Thermodynamic gradient loss The formula used to constrain the rate of change of voltage with respect to temperature to be consistent with the actual physical process is as follows: in, To predict the derivative of voltage with respect to temperature, This represents the rate of change of actual voltage with respect to temperature. Loss of high-altitude characteristics Considering the unique characteristics of the high-altitude environment, the rate of change of the constraint voltage with respect to air pressure follows a specific law, and the calculation formula is as follows: in, To predict the derivative of voltage with respect to air pressure, T is the temperature.

2. The method for monitoring the state of energy storage batteries in high-altitude and cold environments according to claim 1, characterized in that, The method is based on an energy storage battery state monitoring system and includes: Hollow causal convolutional encoder: By collecting four-dimensional time-series data of current, voltage, temperature and air pressure, features are extracted using a hollow causal convolutional encoder. The expansion factor grows exponentially to construct a multi-scale feature representation and fuses the physical information of the battery's initial state. Physical constraint processing module: Apply physical constraints, including SOC conservation constraints, SOH monotonicity constraints, and high-altitude stability constraints, to the feature representation output by the encoder that has fused the initial state physical information, to ensure that the features conform to the physical laws of the battery; High-altitude cold-adaptive decoder: An Arrhenius diffusion kinetics compensation term and a lithium plating risk factor are embedded in the high-altitude cold-adaptive decoder, and a label-free training mechanism including thermodynamic gradient loss and plateau characteristic loss is adopted to output SOC, SOH, voltage prediction value and lithium plating risk factor.

Images