Energy storage power station thermal runaway early warning method based on intra-cluster consistency and residual fusion

Abstract

The present application relates to the technical field of energy storage equipment and control method, and particularly relates to a kind of energy storage station thermal runaway early warning method based on cluster consistency and residual fusion, the normal change caused by stripping current fluctuation and ambient temperature is calculated using improved electro-thermal coupling model theory state;For the energy storage dense stacking characteristics, introduce position correction factor.Construct space-time dual residual, namely longitudinal use physical model residual (self-comparison), transverse use cluster consistency residual (horizontal comparison), solve LFP voltage insensitive problem.Isolation forest algorithm is used to score the residual vector, and the model parameters are self-calibrated and updated using the specific standing / float period of energy storage to adapt to battery aging, and then a closed-loop feedback data processing process is formed.

Description

Technical Field

[0001] This invention relates to the technical field of energy storage equipment and control methods, specifically to an early warning method for thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals. Background Technology

[0002] Lithium iron phosphate (LiFePO4, LFP) batteries have become the mainstream cell type for large-scale electrochemical energy storage systems (such as containerized energy storage power stations) due to their high safety, long cycle life, and low cost. However, as the scale of a single energy storage system continues to expand (typically containing tens of thousands to hundreds of thousands of cell cells), the prevention and control of thermal runaway risks faces severe challenges. Current mainstream thermal runaway early warning technologies mainly rely on fixed threshold alarm mechanisms in the Battery Management System (BMS) or artificial intelligence models based on pure data-driven approaches, but these have significant limitations in practical applications and cannot meet the requirements for "early, accurate, and reliable" early warning.

[0003] First, the inherent electrochemical characteristics of LFP batteries result in insufficient voltage sensitivity. Within the 15%–85% state of charge (SOC) range, the open-circuit voltage curve of LFP batteries is extremely flat. The voltage drop caused by early faults such as minor internal short circuits is minimal (often below millivolts), and is easily masked by measurement noise, contact resistance fluctuations, or current ripple. This causes traditional threshold alarm mechanisms based on voltage anomalies to fail, making it impossible to detect precursors to thermal runaway.

[0004] Secondly, the complexity of thermal management brought about by large-scale integration exacerbates the problem of sensor response lag. Energy storage battery clusters typically adopt a high-density stacking structure, with limited heat dissipation channels between battery cells and long and uneven heat conduction paths. When local overheating or the initial reaction of thermal runaway occurs inside a battery cell, its core temperature may rise rapidly. However, due to the delay in heat conduction, temperature sensors arranged on the surface of the module often take several minutes or even tens of minutes to detect a significant temperature rise, severely weakening the timeliness of early warning.

[0005] Furthermore, existing intelligent early warning methods face a dual dilemma of "false alarm fatigue" and insufficient generalization ability in practical applications. On the one hand, if a simple combination of multivariate thresholds (such as temperature > 60℃ and voltage < 2.5V) is used, although it is simple to implement, it is essentially a post-event response. It usually triggers an alarm only when thermal runaway has entered an irreversible stage, thus losing the early intervention window. On the other hand, although the pure data-driven AI models (such as LSTM, CNN and other deep learning methods) that have emerged in recent years have certain pattern recognition capabilities, they have three major defects: (1) The model has high computational complexity and is difficult to deploy in real time on resource-constrained edge BMS devices; (2) It lacks robustness to normal operating condition interference (such as the temperature rise caused by high current charging and discharging during peak shaving and frequency modulation), and is prone to misjudging dynamic operating conditions as faults; (3) Thermal runaway events themselves are extremely low-probability negative samples, and real fault data is extremely scarce, resulting in insufficient model training, poor generalization performance, and a high false alarm rate on site.

[0006] Therefore, based on the above problems, it is of great significance to study a new early warning method for thermal runaway, which breaks through the limitations of traditional threshold methods and AI models, and achieves high sensitivity and low false alarm thermal runaway early warning. Summary of the Invention

[0007] The purpose of this invention is to provide an early warning method for thermal runaway in energy storage stations based on intra-cluster consistency and residual fusion, so as to solve the existing technical problems in the background art.

[0008] To address the aforementioned technical problems, the present invention provides the following technical solution: a method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals, comprising the following steps: Step 1: Data Acquisition and Processing; Collect raw measured data from the sensors, and after timestamp alignment and filtering, output a time-synchronized and noise-reduced state vector. ; For the time when the data is available, For the first Individual battery cells; Step 2: Construction of a position-sensitive electrothermal coupling model; based on the filtered current in the state vector obtained in Step 1. and alignment temperature The theoretically predicted temperature is calculated based on the electrothermal coupling model formula. ; Step 3: Construction of spatiotemporal fusion residual feature vector; based on the aligned temperature obtained in Step 1. Compared with the theoretically predicted temperature obtained in step two Generate discrete sampling times Longitudinal residuals and lateral residuals Then, the longitudinal and lateral residuals are combined to form a multidimensional spatiotemporal fusion residual feature vector. ;in For the residual integral term; Step four: Anomaly scoring is performed using the isolated forest model; based on the multi-dimensional spatiotemporal fusion residual feature vector obtained in step three. After inputting into the pre-trained isolation forest model, anomaly scores are output after scoring calculation. and alarm signals; Step 5, Parameter Library Update and Feedback: Based on the static data collected in Step 1, reverse-engineer the physical parameters and update the parameters. Write back to the system parameter library; the updated parameters will be used in step two in subsequent operating cycles, forming a closed-loop feedback.

[0009] Based on the above technical solution, the timestamp alignment in step one includes the following process: Based on the arrival time of current data Based on this, linear interpolation is performed to align the voltage and temperature data; assuming the voltage data is at... and The value at time and ,and ,but: Aligned voltage The calculation formula is: ; Aligned temperature The calculation formula is: .

[0010] Based on the above technical solution, the anti-aliasing filtering process in step one is as follows: The current data is processed using a low-pass filter with a preset cutoff frequency, and the filtered current is calculated using discretized difference equations. The discretized difference equation is as follows: ; in, This represents the current discrete sampling time. for Raw current data collected at all times, coefficients This is a filtering operator.

[0011] Based on the above technical solution, the theoretically predicted temperature is calculated using the electrothermal coupling model formula in step two. This includes the following processes: Calculate the total heat production rate with nonlinear entropy heat compensation The calculation formula is as follows: ; in For the first The internal resistance of a single battery cell at its current state of charge and temperature. The entropy-thermal coefficient varies with the state of charge; The heat balance differential equation is based on the lumped parameter method, and a position correction factor is introduced. The calculation formula is as follows:

[0012] in, This refers to the specific heat capacity of a single battery cell. The mass of a single battery cell; For the first The thermal resistance of each battery cell is calculated using the following formula: , The natural convection heat transfer coefficient is... The heat dissipation area of ​​a single battery cell; position correction factor. The initial value is obtained from the calibration, and the updated value is obtained from step five.

[0013] Based on the above technical solution, the longitudinal residual in step three The calculation formula is: ； in, and The first The battery cell in the first Measured temperature and theoretically predicted temperature at each discrete sampling time; To extract the characteristics of the gradually varying temperature rise caused by a small internal short circuit, the longitudinal residual is amplified by the residual integral term, and the calculation formula is as follows:

[0014] in, The number of sampling points included in the set time window. The sampling time interval, This is the sequence of sampling times within the time window.

[0015] Based on the above technical solution, the third step... Each battery cell is in Lateral residuals at each discrete sampling time The difference between the measured temperature data and the median measured temperature of the battery cluster it belongs to is calculated using the following formula: .

[0016] in, This represents the total number of individual battery cells within the battery cluster.

[0017] Based on the above technical solution, the pre-training of the isolated forest in step four includes the following process: Collect historical data of the entire battery cluster within a preset time period during the healthy operation period of the energy storage station; Based on the above historical data, construct a system containing a preset number of... A model of an isolated forest with isolated trees; Each isolation tree is recursively generated by randomly selecting feature dimensions and random cut points until a leaf node contains only one sample or reaches the maximum depth. Complete the pre-training process in the isolated forest.

[0018] Based on the above technical solution, the scoring calculation in step four includes the following process: Real-time multidimensional spatiotemporal fusion residual feature vector Given a tree, calculate the path length from the root node to a leaf node. ; Calculate the average path length of all isolated trees ; Anomaly scores are calculated based on average path length, using the following formula: ,in Given a sample size, the expected average path length is determined; and an alarm signal is issued based on the anomaly scoring criteria.

[0019] Based on the above technical solution, the parameter library update and feedback in step five are calculated using the recursive least squares method, including the following process: Define the observation equation: ;in, For the first Measured temperature difference observations at discrete sampling times; For the first A regression vector that contains current and temperature data at all times; The vector of physical parameters to be identified; e This is the observation error; Gain calculation: ; in, For the first Gain vector at time step; P ( -1) is the first The covariance matrix at time -1; Forgetting factor; Parameter update: ;in, and They are parameter vectors In the Time and the The estimated value at time -1; Covariance matrix update: ;in It is the identity matrix; Closed-loop feedback application: After recursive calibration, from the finally converged parameter estimation vector Extract the updated internal resistance value and position correction factor Update the model parameter library used in step two to complete the adaptive calibration.

[0020] Based on the above technical solution, the triggering conditions for updating and providing feedback to the parameter library are as follows: Current And duration Meanwhile, within the preset sliding window, the mean rate of change of battery temperature is within the preset range and its variance is less than the preset stability threshold.

[0021] The beneficial effects of the technical solution provided by this invention are as follows: Extremely high sensitivity and lead time: Compared with the traditional threshold method, the present invention can identify tiny internal short circuits through residual accumulation and issue an early warning more than 60 minutes before the thermal runaway of LFP battery occurs, while the traditional threshold method often only has 0-5 minutes.

[0022] Low false alarm rate: The normal temperature rise caused by the operating condition (high current) is decoupled through the physical model, and the ambient temperature change is decoupled through lateral consistency, so that the false alarm rate can be maintained even under severe operating conditions such as frequency modulation and peak modulation.

[0023] Adapting to LFP characteristics: Overcoming the industry challenge of monitoring the voltage plateau period of lithium iron phosphate batteries, and filling the blind spot of voltage monitoring by utilizing thermodynamic characteristics.

[0024] Low computational cost: Compared with deep neural networks, the lumped parameter model and isolated forest algorithm of this solution are simple to operate, reducing computational cost by 90%, and can be directly deployed in the local controller (EMU / BCU) of the energy storage power station without the need for expensive GPU servers. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a schematic diagram of the position-sensitive electrothermal coupling model in this invention; Figure 3 This is a graph showing the entropy-heat variation of lithium iron phosphate battery with SOC in this invention. Figure 4This is a comparison chart of the application effects of spatiotemporal fusion residuals and physical model residuals alone in this invention; Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and embodiments: In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] In the description of this invention, it should be understood that the terms "left", "right", "front", "rear", "top", "bottom", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] In an embodiment of the present invention, Or written as Represents physical time or timestamp. Represents discrete sampling moments; the two are linked by the sampling period. Association, that is, satisfying This definition, used throughout the text, distinguishes between the continuous computation of physical models and the discrete processing of digital systems, ensuring that physical modeling, i.e., continuous timestamps, is the basis for this distinction. Algorithm analysis, i.e., discrete sampling time The logical chain is complete.

[0028] like Figures 1 to 4 As shown, a thermal runaway early warning method for energy storage stations based on the fusion of intra-cluster consistency and residuals includes the following steps: Step 1: Data Acquisition and Processing; Collect raw measured data from the sensors, and after timestamp alignment and filtering, output a time-synchronized and noise-reduced state vector. ;in For filter current, To align the voltage, To align the temperature, To align with ambient temperature, For the time when the data is available, For the first Individual battery cells; Step 2: Construction of a position-sensitive electrothermal coupling model; based on the filtered current in the state vector obtained in Step 1. and alignment temperature The theoretically predicted temperature is calculated based on the electrothermal coupling model formula. ; Step 3: Construction of spatiotemporal fusion residual feature vector; based on the aligned temperature obtained in Step 1. Compared with the theoretically predicted temperature obtained in step two Generate longitudinal residuals and lateral residuals Then, the longitudinal and lateral residuals are combined to form a multidimensional spatiotemporal fusion residual feature vector. ; Step four: Anomaly scoring is performed using the isolated forest model; based on the multi-dimensional spatiotemporal fusion residual feature vector obtained in step three. After inputting into the pre-trained isolation forest model, anomaly scores are output after scoring calculation. and alarm signals; Step 5, Parameter Library Update and Feedback: Based on the static data collected in Step 1, reverse-engineer the physical parameters and update the parameters. Write back to the system parameter library; the updated parameters will be used in step two in subsequent operating cycles, forming a closed-loop feedback.

[0029] This invention provides a thermal runaway early warning method for energy storage stations based on the fusion of intra-cluster consistency and residuals. It utilizes an improved electrothermal coupling model to calculate the theoretical state, eliminating normal variations caused by current fluctuations and ambient temperature. For the dense stacking characteristics of energy storage, a position correction factor is introduced. A spatiotemporal dual residual is constructed: vertically, it utilizes the physical model residual (self-comparison); horizontally, it utilizes the intra-cluster consistency residual (horizontal comparison), addressing the LFP voltage insensitivity problem. An isolated forest algorithm is used to score anomalies in the residual vector, and the unique rest / float charging cycle of energy storage is used for model parameter self-calibration and updates to adapt to battery aging, thus forming a closed-loop feedback data processing flow.

[0030] Based on the above technical solution, the timestamp alignment in step one includes the following process: Based on the arrival time of current data Based on this, linear interpolation is performed to align the voltage and temperature data; assuming the voltage data is at... and The value at time and ,and ,but: Aligned voltage The calculation formula is: ; Aligned temperature The calculation formula is: .

[0031] in, , To and Two adjacent original sampling times.

[0032] The original measured data includes current. Battery cell voltage Battery cell temperature and ambient temperature .

[0033] Based on the above technical solution, the anti-aliasing filtering process in step one is as follows: The current data is processed using a low-pass filter with a preset cutoff frequency, and the filtered current is calculated using discretized difference equations. The discretized difference equation is as follows: ; in, This represents the current discrete sampling time. for Raw current data collected at all times, coefficients This is a filtering operator.

[0034] A second-order Butterworth low-pass filter is used to process the current data, with its cutoff frequency preset based on the noise frequency of the switching power supply. Discretized difference equations are designed to calculate the filtered current. Formula, where coefficients The filtering operator is determined by the type, order, and ratio of the sampling frequency to the cutoff frequency of the low-pass filter.

[0035] In a preferred embodiment, when the sampling frequency is 10Hz and the cutoff frequency is 0.5Hz, the coefficient value is: .

[0036] To eliminate high-frequency noise from switching power supplies To mitigate computational interference, a second-order Butterworth low-pass filter is used to process the current data, and the discretized difference equations are designed as shown above.

[0037] Based on the above technical solution, the theoretically predicted temperature is calculated using the electrothermal coupling model formula in step two. This includes the following processes: Calculate the total heat production rate with nonlinear entropy heat compensation The calculation formula is as follows: ; in For the first The internal resistance of a single battery cell at its current state of charge and temperature. The entropy-thermal coefficient varies with the state of charge; for The filtered current constantly flowing through the battery cluster; For the first Alignment temperature of individual battery cells; The heat balance differential equation is based on the lumped parameter method, and a position correction factor is introduced. The calculation formula is as follows:

[0038] in, This refers to the specific heat capacity of a single battery cell. The mass of a single battery cell; For the first The thermal resistance of each battery cell is calculated using the following formula: , The natural convection heat transfer coefficient is... The heat dissipation area of ​​a single battery cell; position correction factor. The initial value is obtained from the calibration, and the updated value is obtained from step five.

[0039] The entropy-thermal coefficient is based on the lithium iron phosphate battery. The mapping table is obtained by looking up the table. Preferably, the entropy-heat coefficient The acquisition process includes: pre-constructing lithium iron phosphate batteries -SOC mapping table, which is obtained through adiabatic calorimeter (ARC) experiments, in SOC Every interval within the interval Set a feature point; specifically, in SOC When the value is negative (e.g.) (exothermic); at SOC of When the value is close to ; in SOC When the value suddenly turns positive (e.g.) (Heat absorption). During model runtime, real-time coefficients are obtained through table lookup and linear interpolation, thereby eliminating spurious temperature rise predictions at the charging end.

[0040] Preferably, the initial value of the position correction factor is obtained from the factory calibration, and the real-time value during operation is dynamically updated by the self-calibration algorithm in step five and written into the parameter library to achieve closed-loop feedback.

[0041] This step includes an entropy-heat calculation term that varies nonlinearly with SOC in the electrothermal coupling model, which is particularly suitable for the characteristics of lithium iron phosphate batteries and eliminates the thermal prediction error during the battery's plateau period. At the same time, a position correction factor is introduced into the electrothermal coupling model, which can correct the heat dissipation efficiency through position differences. This is particularly suitable for high-density stacked energy storage battery cluster systems, and the thermal runaway warning is more timely.

[0042] Based on the above technical solution, the longitudinal residual in step three The calculation formula is: ； in, and The first The battery cell in the first Measured temperature and theoretically predicted temperature at each discrete sampling time; To extract the characteristics of the gradually varying temperature rise caused by a small internal short circuit, the longitudinal residual is amplified by the residual integral term, and the calculation formula is as follows:

[0043] in, The number of sampling points included in the set time window. The sampling time interval, This is the sequence of sampling times within the time window.

[0044] The longitudinal residual reflects whether the individual battery cell itself violates physical laws. When this occurs, it indicates the presence of an additional abnormal heat source, meaning an internal short circuit has occurred.

[0045] Preferably, the window Set as ,against The micro-internal short circuit has an extremely small instantaneous temperature rise (such as...). (Signal-to-noise ratio) is easily overwhelmed by noise. By setting the integral accumulation within a window, the signal-to-noise ratio can be improved. More than twice. The principle behind the improved signal-to-noise ratio is that sensor measurement noise, such as thermocouple ripple and ADC sampling noise, is usually manifested as Gaussian white noise with a mean of zero. Although its instantaneous value in the time domain may mask the slowly varying temperature rise, by performing integration calculation within a sliding window of length W, the noise terms will cancel each other out due to their randomness, while the trend residual term reflecting the heat generation of the internal short circuit will accumulate linearly over time.

[0046] According to statistical principles, the signal component after integration grows much faster than the random fluctuation of the noise component. Experimental data shows that, at a sampling frequency of 10Hz and a window... With this configuration, for amplitudes that are only The slowly varying signal, after integration and accumulation, can significantly deviate from the background noise baseline, thereby increasing the effective signal-to-noise ratio (SNR) of feature identification from less than 1dB in the original data to more than 20dB, thus enabling reliable extraction of early signs of thermal runaway.

[0047] Based on the above technical solution, the third step... Each battery cell is in Lateral residuals at each discrete sampling time The difference between the measured temperature data and the median measured temperature of the battery cluster it belongs to is calculated using the following formula: .

[0048] in, This represents the total number of individual battery cells within the battery cluster.

[0049] The measured temperature set of the entire battery cluster output in step one is used to statistically calculate the lateral residuals. The median, rather than the mean, is used as the group benchmark. This is because when individual cells in the battery cluster experience severe failures, such as extremely high temperatures, the mean will be pulled up, causing the residuals of other normal cells to become negative and resulting in misjudgments. The median, on the other hand, is robust and is not affected by extreme outliers, and can more accurately isolate faulty cells, ensuring the robustness of the thermal runaway early warning method.

[0050] Specifically, the longitudinal residual reflects the degree to which a single cell deviates from the physical model. In the initial stage of a small internal short circuit, such as 1000-5000Ω, the instantaneous temperature rise is extremely small, about 0.01 degrees Celsius / min, and is easily drowned out by measurement noise. By setting the integral accumulation of the window W, the signal-to-noise ratio can be improved by more than 10 times, extracting the extremely slow heat accumulation signal. The lateral residual is obtained by subtracting the median within the battery cluster to remove common-mode interference from the environment and operating conditions, such as the large current temperature rise caused by frequency modulation and peak modulation or changes in ambient temperature. After the environmental interference is canceled out, the lateral characteristics of the faulty cell exhibit relatively abrupt changes, such as... Figure 4 As shown, the residuals of the original physical model, represented by the light-colored dashed line, exhibit significant fluctuations under environmental interference. In contrast, the spatiotemporal fusion residuals processed in step three of this invention, represented by the dark-colored solid line, maintain extremely high stability before the fault occurs. This background suppression capability stems from the noise filtering and exponential amplification of the feature signal by the longitudinal residual integral term. The faulty individual can be effectively isolated simply by taking the instantaneous value. Therefore, the asymmetric processing of the longitudinal and lateral residuals in feature extraction is the core design of this algorithm to reduce the false alarm rate.

[0051] Based on the above technical solution, the pre-training of the isolated forest in step four includes the following process: Collect historical data of the entire battery cluster within a preset time period during the healthy operation period of the energy storage station; Based on the above historical data, construct a system containing a preset number of... A model of an isolated forest with isolated trees; Each isolation tree is recursively generated by randomly selecting feature dimensions and random cut points until a leaf node contains only one sample or reaches the maximum depth. Complete the pre-training process in the isolated forest.

[0052] Based on the above technical solution, the scoring calculation in step four includes the following process: Real-time multidimensional spatiotemporal fusion residual feature vector Given a tree, calculate the path length from the root node to a leaf node. ; Calculate the average path length of all isolated trees ; Anomaly scores are calculated based on average path length, using the following formula: ,in Given a sample size, the expected average path length is determined; and an alarm signal is issued based on the anomaly scoring criteria.

[0053] The anomaly scoring criteria are as follows: If And duration This triggered a thermal runaway warning.

[0054] Based on the above technical solution, the parameter library update and feedback in step five are calculated using the recursive least squares method, including the following process: Define the observation equation: ;in, For the first Measured temperature difference observations at discrete sampling times; For the first A regression vector that contains current and temperature data at all times; The vector of physical parameters to be identified; e This is the observation error; The vector Includes the internal resistance value to be identified and position correction factor Equal parameter set Gain calculation: ; in, For the first Gain vector at time step; For the first The covariance matrix at time -1; Forgetting factor; The forgetting factor is set as follows This is used to assign higher weights to new data and forget about old, outdated data.

[0055] Parameter update: ;in, and They are parameter vectors In the Time and the The estimated value at time -1; This step enables the parameter set to be updated simultaneously at the current moment; Covariance matrix update: ;in It is the identity matrix; Closed-loop feedback application: After recursive calibration, from the finally converged parameter estimation vector Extract the updated internal resistance value and position correction factor Update the model parameter library used in step two to complete the adaptive calibration.

[0056] Based on the above technical solution, the triggering conditions for updating and providing feedback to the parameter library are as follows: Current And duration Meanwhile, within the preset sliding window, the mean of the first derivative of the battery temperature change rate is within the preset range and its variance is less than the preset stability threshold.

[0057] The battery clusters are in a static or float charging state, meeting the current requirements. And duration .

[0058] The temperature stability determination logic is met: within a preset sliding window, the statistical characteristics of the temperature change rate (first derivative) of each battery cell are calculated. The mean of the change rate must be within a preset quasi-steady-state range, and the variance of the change rate must be below a preset stability threshold to ensure the signal-to-noise ratio of the input data for the recursive least squares method. Data quality is assessed using statistical methods to ensure that the temperature sequence entering the algorithm is smooth and conforms to physical cooling laws, thereby guaranteeing the convergence and accuracy of parameter identification.

[0059] The logic behind this trigger condition is that the parameter library update and feedback process utilizes data from the resting period for reverse identification of physical parameters. Since the recursive least squares (FF-RLS) method is highly sensitive to the stability of the input signal, drastic temperature changes or high-frequency disturbances can cause the identification results to diverge. By introducing the mean range of the rate of change and variance constraints, this embodiment can determine whether the system has entered a "quasi-steady state," thereby eliminating non-stationary noise such as sudden environmental changes and ensuring that the extracted internal resistance value R and position correction factor λ accurately reflect the battery aging state, achieving adaptive calibration.

[0060] This method employs recursive least squares with a forgetting factor (FF-RLS), which introduces a forgetting factor. ), the value is in In this process, newly acquired data can be given higher weight, while older historical data can be gradually forgotten. This is especially important in battery parameter updates, where the internal resistance and thermal parameters of individual battery cells are not constant but drift slowly over their entire lifespan (SOH decay). By setting a forgetting factor... It can track this slow time-varying characteristic in real time and achieve adaptive updates of model parameters, thereby avoiding false alarms caused by outdated parameters.

[0061] To fully demonstrate that this invention can achieve the technical effects of providing a 60-minute early warning and reducing computing power by 90%, the following data support based on theoretical calculations and simulation experiments is provided: 1. Computing power requirements have been significantly reduced. The algorithm in the early warning method of this application was compared with the computational cost (FLOPs) of the mainstream Bidirectional Long Short-Term Memory (BiLSTM) network on the same embedded hardware platform (such as STM32F4 series MCU) in a single inference: (1) Comparison object (BiLSTM deep learning model): Assume the model structure is: input layer (10-dimensional) -> hidden layer (64-dimensional) -> output layer (1-dimensional).

[0062] A single forward inference operation involves a large number of matrix multiplications. The estimation formula is: .

[0063] The computational workload is approximately: This is a floating-point operation.

[0064] (2) The proposed scheme (physical model + isolated forest): The physical model section only involves simple algebraic addition, subtraction, multiplication, and division (such as...). The calculation involves simple difference equations and approximately 50 floating-point operations.

[0065] Isolation Forest part: Essentially, it is a path search of a binary tree, which only involves the logical judgment of if (value < threshold) and hardly consumes floating-point computing power. Suppose there are 100 trees with an average tree depth of 10, only about 1000 simple comparison operations are required. The total computational amount (equivalent FLOPs) is much lower than 500.

[0066] Conclusion: . That is, the computing power requirement of this invention is only about 5% of that of the deep learning model, and the reduction in computing power exceeds 90%. This data proves the feasibility of this solution for deployment at low-cost edge devices (BCU).

[0067] 2. Simulation experiment verification of early warning lead Since it is difficult to obtain a large amount of real thermal runaway data, we constructed a thermal-electrical coupling simulation model of internal short circuit in the battery cell based on COMSOL Multiphysics. For the verification of intra-cluster consistency (horizontal), in order to avoid the waste of computing power in establishing thousands of highly complex models, we adopted the method of constructing virtual battery clusters, which is as follows: (1) Construction of simulation model (virtual battery cluster) Fault benchmark model: Establish 1 high-precision finite element model containing the physical field of internal short circuit (set as the battery cell ), and introduce a 10Ω internal short circuit resistance at moment.

[0068] Normal benchmark model: Establish 1 completely identical normal battery cell model (without internal short circuit).

[0069] Generation of virtual population: Use the output data of the normal benchmark model to copy and generate copies of data (simulating the standard battery module of 52 strings in a battery cluster). In order to simulate the sensor noise in the real world and the slight differences between individual battery cells, Gaussian white noise with different distributions is superimposed on these 51 copies of data .

[0070] Combination: Combine (fault data) with (virtual normal data) to form a complete virtual battery cluster data set.

[0071] (2) Verification process (calculation of horizontal residual) At minutes, the core temperature of the faulty battery cell slightly increases, and the surface temperature increases by about 0.8℃ compared with the normal value.

[0072] Calculation of horizontal residual: The system calculates the horizontal residual of Since the other 51 cells did not experience a temperature rise, the cluster average remained almost unchanged, resulting in... It deviates significantly from 0.

[0073] Environmental interference elimination: Assuming the overall ambient temperature rises by 5°C, all battery cells (including faulty and normal ones) will be affected. Both will rise by 5°C simultaneously. At this time, An increase of 5.8℃ and a cluster mean increase of 5.0℃, subtracting the two, yields the lateral residual. It remained at 0.8℃. This proves that the lateral model can effectively counteract environmental common-mode interference.

[0074] (3) Comparison results: Traditional threshold method (setting) ):until The surface temperature only reached 60°C after a few minutes, triggering the alarm.

[0075] The thermal runaway early warning method applied for here: At minute, longitudinal residual integral term Cumulative exceedance, and horizontally consistent residuals The cell was confirmed to be significantly deviating from the population. The anomaly score output by the isolated forest exceeded 0.85, triggering a level 2 alert.

[0076] (4) Conclusion: The alarm time of this invention ( min) compared to traditional thresholding methods ( min), 70 minutes ahead of schedule.

[0077] It should be noted that in the simulation verification phase described above, we avoided building 52 complex finite element models by reusing and adding noise. In the actual engineering deployment phase, this invention runs a lightweight lumped parameter model (i.e., algebraic formula) rather than a complex finite element model. A single microprocessor (MCU) can easily compute thousands of such lightweight formulas in parallel, thus eliminating the computational bottleneck.

[0078] The foregoing has shown and described the basic principles and main features of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments. Therefore, the embodiments should be considered as exemplary and not restrictive. The scope of the present invention is defined by the appended claims rather than the foregoing description. Therefore, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the present invention.

[0079] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A thermal runaway early warning method for energy storage stations based on intra-cluster consistency and residual fusion, characterized in that, Includes the following steps: Step 1, Data Acquisition and Processing: Collect raw measured data from the sensor, and after timestamp alignment and filtering, output a time-synchronized and noise-reduced state vector; For the time when the data is available, For the first Individual battery cells; Step 2: Construction of a position-sensitive electrothermal coupling model; based on the filtered current in the state vector obtained in Step 1. and alignment temperature The theoretically predicted temperature is calculated based on the electrothermal coupling model formula. ; Step 3: Construction of spatiotemporal fusion residual feature vector; based on the aligned temperature obtained in Step 1. Compared with the theoretically predicted temperature obtained in step two Generate discrete sampling times Longitudinal residuals and lateral residuals Then, the longitudinal and lateral residuals are combined to form a multidimensional spatiotemporal fusion residual feature vector. ;in For the residual integral term; Step 4: Use the isolated forest model to determine anomalies; Based on the multidimensional spatiotemporal fusion residual feature vector obtained in step three After inputting into the pre-trained isolation forest model, anomaly scores are output after scoring calculation. and alarm signals; Step 5, Parameter Library Update and Feedback: Based on the static data collected in Step 1, reverse-engineer the physical parameters and update the parameters. Write back to the system parameter library; the updated parameters will be used in step two in subsequent operating cycles, forming a closed-loop feedback.

2. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, Timestamp alignment in step one Includes the following processes: Based on the arrival time of current data Based on this, linear interpolation is performed to align the voltage and temperature data; assuming the voltage data is at... and The value at time and ,and ,but: Aligned voltage The calculation formula is: ； Aligned temperature The calculation formula is: 。 3. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, The anti-aliasing filtering process in step one is as follows: The current data is processed using a low-pass filter with a preset cutoff frequency, and the filtered current is calculated using discretized difference equations. The discretized difference equation is as follows: ； in, This represents the current discrete sampling time. for Raw current data collected at all times, coefficients This is a filtering operator.

4. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, In step two, the theoretical predicted temperature is calculated using the electrothermal coupling model formula. This includes the following processes: Calculate the total heat production rate with nonlinear entropy heat compensation The calculation formula is as follows: ; in For the first The internal resistance of a single battery cell at its current state of charge and temperature. The entropy-thermal coefficient varies with the state of charge; The heat balance differential equation is based on the lumped parameter method, and a position correction factor is introduced. The calculation formula is as follows: in, This refers to the specific heat capacity of a single battery cell. The mass of a single battery cell; For the first The thermal resistance of each battery cell is calculated using the following formula: , The natural convection heat transfer coefficient is... The heat dissipation area of ​​a single battery cell; position correction factor. The initial value is obtained from the calibration, and the updated value is obtained from step five.

5. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, Longitudinal residual in step three The calculation formula is: ； in, and The first The battery cell in the first Aligned temperature and theoretically predicted temperature at each discrete sampling time; To extract the gradually varying temperature rise characteristics caused by minute internal short circuits, the longitudinal residual is amplified by the residual integral term, and the calculation formula is as follows: in, The number of sampling points included in the set time window. The sampling time interval, This is the sequence of sampling times within the time window.

6. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, In step three Each battery cell is in Lateral residuals at each discrete sampling time The difference between the measured temperature data and the median measured temperature of the battery cluster it belongs to is calculated using the following formula: 。 in, This represents the total number of individual battery cells within the battery cluster.

7. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, The pre-training of the isolated forest in step four includes the following process: Collect historical data of the entire battery cluster within a preset time period during the healthy operation period of the energy storage station; Based on the above historical data, construct a system containing a preset number of... A model of an isolated forest with isolated trees; Each isolation tree is recursively generated by randomly selecting feature dimensions and random cut points until a leaf node contains only one sample or reaches the maximum depth. Complete the pre-training process for the isolated forest.

8. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, The scoring calculation in step four includes the following process: Real-time multidimensional spatiotemporal fusion residual feature vector Given a tree, calculate the path length from the root node to a leaf node. ; Calculate the average path length of all isolated trees ; Anomaly scores are calculated based on average path length, using the following formula: ,in Given a sample size, the expected average path length is determined; and an alarm signal is issued based on the anomaly scoring criteria.

9. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, In step five, the parameter database update and feedback are calculated using the recursive least squares method. Includes the following processes: Define the observation equation: ;in, For the first Measured temperature difference observations at discrete sampling times; For the first A regression vector that contains current and temperature data at all times; The vector of physical parameters to be identified; e This is the observation error; Gain calculation: ; in, For the first Gain vector at time step; For the first The covariance matrix at time -1; Forgetting factor; Parameter update: ;in, and They are parameter vectors In the Time and the The estimated value at time -1; Covariance matrix update: ;in It is the identity matrix; Closed-loop feedback application: After recursive calibration, from the finally converged parameter estimation vector Extract the updated internal resistance value and position correction factor Update the model parameter library used in step two to complete the adaptive calibration.

10. The method for early warning of thermal runaway in energy storage stations based on the fusion of intra-cluster consistency and residuals as described in claim 1, characterized in that, The triggering conditions for updating and providing feedback to the parameter library are as follows: Current And duration Meanwhile, within the preset sliding window, the mean rate of change of battery temperature is within the preset range and its variance is less than the preset stability threshold.

Images