Cross-domain perception self-evolution energy storage battery management system and protection method thereof

Abstract

The application discloses a cross-domain perception self-evolution energy storage battery management system and a protection method thereof, and belongs to the technical field of energy storage battery management. The system comprises a perception unit, a data fusion unit, an algorithm core unit and a collaborative control unit. The perception unit collects cell microstate, battery body operation and external scene parameters; the data fusion unit pre-processes and outputs a feature vector; the algorithm core unit comprises a double-cycle AI algorithm module of online incremental learning and offline iterative updating, and a digital twin engine module for constructing a special twin for each cell; and the collaborative control unit allocates and schedules tasks according to the cell health state, and reversely requests load adjustment in an abnormal situation. The application realizes accurate cell state perception, model self-evolution updating and health constraint collaborative control, and improves the safety, economy and cycle life of the energy storage system.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery management technology, specifically to a cross-domain sensing self-evolving energy storage battery management system and its protection method that integrates cell microscopic sensing, cross-domain data fusion, self-evolving artificial intelligence algorithms, digital twins, and bidirectional collaborative control of source-grid-load-storage-cell.

[0002] More specifically, this invention belongs to the field of collaborative technology between Battery Management System (BMS) and Energy Internet, and can be widely applied to various scenarios such as large-scale energy storage on the grid side, industrial and commercial energy storage, residential energy storage, and power battery management for new energy vehicles. Background Technology

[0003] A Battery Management System (BMS) is a core component ensuring the safe, efficient, and long-life operation of an energy storage system. Its main functions include monitoring battery status, estimating state parameters, controlling charge and discharge, managing battery balance, and providing safety protection. With the expansion of the energy storage industry and the diversification of application scenarios, especially in collaborative scenarios involving virtual power plants (VPPs), grid frequency regulation, and demand-side response, traditional BMS technology is increasingly unable to meet the demands for high safety, high economic efficiency, and long lifespan. Currently, energy storage BMS technology is at a critical juncture in its transformation from "basic monitoring and control" to "intelligent collaborative management," and related research mainly focuses on the following directions: In terms of sensing system construction, existing BMS generally adopts a macroscopic parameter acquisition scheme with battery pack surface temperature, total voltage, and total current as the core. Some advanced BMS solutions have begun to attempt to integrate cross-domain data such as grid frequency and ambient temperature and humidity, but most of them remain at a shallow fusion level of 3-4 dimensions. In recent years, some researchers have proposed smart battery technology that embeds sensors inside the battery. For example, related research from Beijing Institute of Technology has disclosed an embedded multi-source sensor group for detecting parameters such as internal battery pressure and core temperature. However, this type of solution still has the following limitations: First, it does not involve key microscopic parameters that can reflect the evolution of the internal state of the cell, such as separator impedance and tab temperature gradient; second, the integration method of sensors and cells has not yet formed a standardized process, and the impact on the energy density and reliability of the cells needs to be verified; third, it lacks a systematic fusion mechanism with grid-side, load-side, and environmental-side data, making it difficult to support the needs of source-grid-load-storage coordinated control. This perception mode, characterized by "macro-centricity, lack of micro-level information, and fragmented cross-domain approaches," results in insufficient accuracy of existing BMS in identifying internal cell degradation and early-stage faults. The SOC (State of Charge) estimation often has an error rate exceeding 5% under complex operating conditions, making it difficult to provide reliable data support for refined control.

[0004] In terms of state estimation and algorithmic models, artificial intelligence technology has been widely applied to the estimation of core parameters such as SOC (State of Charge) and SOH (State of Health). Current mainstream solutions mostly employ offline-trained neural network models, such as BP neural networks and LSTM (Long Short-Term Memory) networks, fitting the mapping relationship between input features and state parameters based on historical data. Related research has also attempted to introduce digital twin technology to construct virtual images of battery systems, enabling visualized monitoring of their state. However, these solutions generally suffer from the limitation of "fixed models, difficult to evolve": once deployed, the accuracy of offline-trained models significantly decreases with cell degradation and changes in operating conditions, requiring retraining based on newly collected data; while some solutions mention online learning, they lack effective mechanisms to suppress "catastrophic forgetting," causing the model to lose historical experience while learning new knowledge; data sharing and joint modeling among multiple nodes face challenges related to data privacy and transmission security, limiting the improvement of model generalization capabilities. The current applications of digital twin technology are mostly concentrated on battery manufacturing process optimization or system-level state monitoring. For example, a patent discloses a digital twin modeling method for the forming process of curved lithium batteries, but there are no reports of building a dedicated twin for a single cell that integrates production process parameters and full life cycle operation data, and supports control strategy simulation optimization.

[0005] Regarding balancing and control strategies, existing BMSs mostly employ passive or active balancing strategies based on voltage consistency, with relatively simple control logic. At the source-grid-load-storage coordination level, the interaction between the BMS and virtual power plants (VPPs) or energy management systems is primarily a one-way command response mode: after receiving dispatch commands, the BMS executes charging and discharging operations according to preset logic and reports the results. This mode presents two problems: first, it fails to consider the real-time health status of the cells as a core constraint for dispatch decisions, leading to a contradiction between the power demand of the grid and the lifespan guarantee of the cells—the potential of healthy cells is not fully realized, while degraded cells are accelerated to failure due to overuse; second, it lacks a reverse feedback mechanism under abnormal conditions. When cells experience emergencies such as overheating, the BMS can only perform local protection actions and cannot actively coordinate with the grid to adjust the load, affecting the overall safety and availability of the system. Some research has attempted to introduce SOH-based collaborative control, such as achieving SOH balancing between cells through a dynamically reconfigurable battery network, but this solution relies on complex hardware topology reconfiguration, posing challenges to system cost and reliability.

[0006] In terms of thermal runaway safety protection, traditional BMS mainly relies on threshold alarm mechanisms based on a single parameter, such as triggering a buzzer alarm or cutting off the circuit when the temperature exceeds a set value. In recent years, researchers have begun to focus on early warning methods based on the collaborative sensing of multiple physical quantities, such as simultaneously monitoring the changing trends of parameters such as voltage, current, and pressure, and combining Kalman filtering or mathematical modeling for fault diagnosis. Graph Neural Networks (GNNs), as a deep learning model that can effectively process graph-structured data, have shown excellent performance in fields such as infrared thermal imaging diagnostics of industrial equipment. However, a complete technical solution that applies GNNs to the prediction of thermal runaway in battery cells, constructs a feature system with microscopic parameters such as tab temperature gradient and diaphragm impedance mutation rate as the core, and constructs an adjacency matrix based on the correlation between parameters, and finally links cooling, isolation, inert gas injection and other execution units to achieve graded intervention, has not yet been publicly reported. Existing protection solutions generally suffer from the problems of "lagging early warning and single intervention," making it difficult to achieve accurate prediction before thermal runaway occurs, and even more difficult to match differentiated intervention strategies according to risk levels, resulting in limited thermal diffusion suppression effects.

[0007] In summary, existing energy storage BMS technologies still have significant technological gaps and room for improvement in terms of sensing dimensions, algorithm evolution capabilities, control coordination mechanisms, and safety protection levels. How to construct an intelligent BMS capable of deeply integrating the microscopic state of battery cells with cross-domain macroscopic data, possessing continuous self-evolution capabilities, and achieving bidirectional coordinated control of source, grid, load, and storage with battery cell health as the core constraint, as well as full-chain prediction and graded intervention of thermal runaway based on multi-parameter correlation mining, has become a key technical problem urgently needing to be solved in this field.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0010] This disclosure provides a cross-domain sensing self-evolving energy storage battery management system and its protection method to solve the technical problems of existing energy storage BMS: single sensing dimension, lack of fusion of cell micro parameters and cross-domain data; rigid algorithm model, unable to adapt to cell aging and operating condition changes; extensive management, ignoring individual cell differences; and unidirectional coordination, failing to take cell health as the core constraint of scheduling.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: On one hand, the present invention provides a cross-domain sensing self-evolving energy storage battery management system, comprising: The sensing unit is used to collect microscopic state parameters of the battery cell, battery operating parameters, and external scene parameters. The external scene parameters include at least one of grid-side parameters, environmental-side parameters, and load-side parameters. The sensing unit serves as the system's data acquisition layer, providing multi-dimensional and multi-layered foundational data for subsequent processing.

[0012] A data fusion unit, connected to the sensing unit, is used to preprocess the received multidimensional data. The preprocessing includes data cleaning, time synchronization, and feature extraction, ultimately outputting a fused feature vector. This unit addresses the inconsistencies in time reference, noise level, and dimensionality differences among multi-source data.

[0013] The algorithm core unit, connected to the data fusion unit, is the intelligent decision-making core of the system. This unit specifically includes: A self-evolving dual-loop AI algorithm module is used to estimate the cell state parameters based on the feature vector. This module has a dual-loop self-evolution mechanism of online incremental learning and offline iterative updates, enabling the model to continuously adapt to cell aging and changes in operating conditions.

[0014] The cell digital twin engine module is used to build a unique digital twin for each cell. The twin integrates the cell's manufacturing process parameters and factory inspection data for initialization, and maps the cell's current state in real time based on the feature vector, enabling fine-grained management at the cell level.

[0015] The collaborative control unit is communicatively connected to both the algorithm core unit and the external scheduling center. This unit receives external scheduling commands and, based on the cell health status output by the cell digital twin engine module, allocates scheduling tasks to each cell in a differentiated manner. Simultaneously, when an abnormal cell status is detected, this unit sends a load adjustment request to the external scheduling center, forming a two-way collaborative control closed loop.

[0016] As a preferred technical solution, the sensing unit further includes: The cell micro-sensing subunit is embedded in the cell or on its surface and is used to collect at least one micro-state parameter among tab temperature, cell expansion and diaphragm impedance. The battery body sensing subunit is used to collect at least one of the following operating parameters of the battery cell: voltage, current and temperature. The cross-domain sensing subunit is used to collect at least one external scene parameter from the following: power grid frequency, voltage fluctuation, ambient temperature and humidity, light intensity, wind speed, and load power.

[0017] As a preferred technical solution, the self-evolving dual-loop AI algorithm module further includes: The online fine-tuning loop submodule deploys an incremental learning model, which is used to fine-tune the current model based on newly collected data under preset trigger conditions. A historical knowledge retention mechanism is introduced during the fine-tuning process to prevent the model from forgetting the historical experience it has learned. The offline iterative loop submodule deploys a federated learning mechanism to aggregate and update model parameters across multiple local nodes, and encrypts and protects the model parameters during the aggregation process to ensure data privacy and security.

[0018] More preferably, the incremental learning model in the online fine-tuning recurrent submodule is a long short-term memory network model, and the historical knowledge retention mechanism is a distillation loss function based on a learning-without-forgetting strategy.

[0019] More preferably, when aggregating model parameters, the offline iterative loop submodule calculates dynamic weights based on the data volume and working condition similarity of each local node, and performs a weighted average of the model parameters uploaded by the local nodes based on the dynamic weights.

[0020] As a preferred technical solution, the cell digital twin engine module is also used to simulate the impact of candidate control strategies on the health status of the cell, and to select the optimal control strategy based on the simulation results, thereby realizing the active optimization of the control strategy.

[0021] As a preferred technical solution, the collaborative control unit further includes: The task allocation submodule is used to allocate the power tasks corresponding to external scheduling commands to cells with different health levels based on the cell health status level. The status monitoring and feedback submodule is used to monitor the cell status parameters in real time. When the cell status parameters exceed the preset threshold, it sends a load adjustment request to the external dispatch center and starts the corresponding thermal management execution unit.

[0022] As a preferred technical solution, the system further includes: The thermal runaway prediction and intervention unit, connected to the sensing unit and the algorithm core unit, is used to predict the thermal runaway risk level based on continuous multi-frame monitoring data including seven-dimensional monitoring data such as voltage, current, temperature, micro-vibration, gas concentration, electrode temperature gradient and diaphragm impedance mutation rate, through a pre-trained graph neural network model, and to execute a graded intervention strategy according to the risk level.

[0023] More preferably, the thermal runaway prediction and intervention unit includes: The risk prediction submodule takes as input seven-dimensional monitoring data, including voltage, current, temperature, micro-vibration, gas concentration, electrode temperature gradient and diaphragm impedance mutation rate, and its adjacency matrix is ​​constructed based on the correlation matrix of the seven-dimensional monitoring data. The graded intervention submodule is used to match and execute one or more intervention actions from air cooling, liquid cooling, charge / discharge circuit disconnection, faulty cell isolation, and nitrogen injection based on the predicted risk level.

[0024] On the other hand, the present invention provides a protection method for a cross-domain sensing self-evolving energy storage battery management system based on any of the above technical solutions, comprising the following steps: Step S1: Collect cell micro-state parameters, battery operating parameters, and external scene parameters through the sensing unit; Step S2: The collected multidimensional data is preprocessed by the data fusion unit, and the fused feature vector is output. Step S3: Estimate the cell state parameters through the self-evolving dual-loop AI algorithm module of the core algorithm unit, and map the current state of the cell in real time through the cell digital twin engine module; Step S4: Receive external scheduling instructions through the collaborative control unit, allocate scheduling tasks to each cell in a differentiated manner according to the cell health status, and send a load adjustment request to the external scheduling center when an abnormal cell status is detected; Step S5: The thermal runaway prediction and intervention unit performs risk prediction on continuous multi-frame monitoring data and executes graded intervention strategies according to the risk level.

[0025] The cross-domain sensing self-evolving energy storage battery management system and its protection method provided in this disclosure can achieve the following technical effects: The sensing unit first collects microscopic state parameters of the battery cell, battery operating parameters, and external scenario parameters such as the power grid, environment, and load, forming multi-dimensional raw data. The data fusion unit cleans, synchronizes, and extracts features from this data, outputting a unified feature vector. The self-evolving dual-loop AI algorithm module continuously estimates the state parameters of the battery cell, such as SOC and SOH, based on this feature vector, and achieves adaptive evolution of the model through online incremental learning and offline federated learning mechanisms. The battery cell digital twin engine module constructs a unique digital twin for each battery cell that integrates production process parameters, mapping the current state of the battery cell in real time. The collaborative control unit distributes external scheduling instructions to each battery cell in a differentiated manner according to the health status of the battery cell, and actively sends load adjustment requests to the power grid when the status is abnormal.

[0026] Compared to existing technologies, the advantages of this technical solution are as follows: At the perception level, by integrating cell micro-parameters and cross-domain data, the problem of traditional BMS having a single perception dimension and being "invisible and unclear" about the internal state of the cell is solved, providing a reliable data foundation for precise control. At the algorithm level, through a dual-cycle self-evolution mechanism, the problem of fixed models being unable to adapt to cell aging and changes in operating conditions is solved, ensuring that the accuracy of SOC estimation and SOH prediction remains stable over the long term. At the management granularity level, through a cell-level digital twin engine, a leap from "one-size-fits-all" management of the battery pack to "tailor-made" management of individual cells is achieved, effectively reducing the impact of individual cell differences on the overall lifespan. At the coordination mechanism level, through bidirectional collaborative control with cell health as the core constraint, the one-way mode of passively responding to grid commands by BMS is changed, maximizing the protection of cell lifespan while meeting grid dispatch requirements, and achieving a balance between safety and economy.

[0027] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0028] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of the overall system architecture provided in an embodiment of this disclosure; Figure 2 This is a flowchart of a self-evolving dual-loop AI algorithm framework provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram illustrating the working principle of a battery cell digital twin engine provided in an embodiment of this disclosure; Figure 4 This is a flowchart of a bidirectional coordinated control system for source, grid, load, energy storage, and battery cells provided in an embodiment of this disclosure. Detailed Implementation

[0029] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0030] The terms "first," "second," etc., used in the embodiments of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0031] Unless otherwise stated, the term "multiple" means two or more.

[0032] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0033] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0034] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0035] This invention centers on "deep integration of battery cell and BMS," constructing a closed-loop system encompassing "sensing, fusion, algorithm, control, collaboration, and feedback." The core logic is as follows: The sensing layer synchronously collects multi-dimensional data from the battery cell's micro-state, the battery itself, the grid side, the environment side, and the load side. After data fusion, cleaning, time synchronization, and feature extraction are performed, the data is input to the algorithm core layer. The algorithm layer, through a self-evolving dual-loop AI engine and a battery cell digital twin engine, achieves accurate state prediction and control strategy optimization. The control layer, based on the optimized strategy, drives execution units such as intelligent balancing and gradient cooling, while simultaneously achieving cross-domain linkage with virtual power plants, cloud platforms, and local terminals through a collaborative interaction layer. Finally, the battery cell's operating status is fed back to the sensing layer, forming a continuously iterative closed-loop control system that ensures the safety, long lifespan, and high economic efficiency of the energy storage system.

[0036] Combination Figures 1 to 4 This invention provides a cross-domain sensing self-evolving energy storage battery management system, comprising: The sensing unit is used to collect microscopic state parameters of the battery cell, battery operating parameters, and external scene parameters. The external scene parameters include at least one of grid-side parameters, environmental-side parameters, and load-side parameters. The sensing unit serves as the system's data acquisition layer, providing multi-dimensional and multi-layered foundational data for subsequent processing.

[0037] A data fusion unit, connected to the sensing unit, is used to preprocess the received multidimensional data. The preprocessing includes data cleaning, time synchronization, and feature extraction, ultimately outputting a fused feature vector. This unit addresses the inconsistencies in time reference, noise level, and dimensionality differences among multi-source data.

[0038] The algorithm core unit, connected to the data fusion unit, is the intelligent decision-making core of the system. This unit specifically includes: A self-evolving dual-loop AI algorithm module is used to estimate the cell state parameters based on the feature vector. This module has a dual-loop self-evolution mechanism of online incremental learning and offline iterative updates, enabling the model to continuously adapt to cell aging and changes in operating conditions.

[0039] The cell digital twin engine module is used to build a unique digital twin for each cell. The twin is initialized by integrating the cell's production process parameters and factory inspection data, and maps the cell's current state in real time based on the feature vector, achieving fine-grained management at the cell level.

[0040] The collaborative control unit is communicatively connected to both the algorithm core unit and the external scheduling center. This unit receives external scheduling commands and, based on the cell health status output by the cell digital twin engine module, allocates scheduling tasks to each cell in a differentiated manner. Simultaneously, when an abnormal cell status is detected, this unit sends a load adjustment request to the external scheduling center, forming a two-way collaborative control closed loop.

[0041] In one specific implementation, the data fusion unit and the collaborative control unit are implemented by the same main controller. This main controller employs a microcontroller with a clock frequency of at least 480MHz, equipped with at least 2MB of Flash and 1MB of RAM, such as an STM32H743IGT6 chip. The data fusion unit uses this main controller to perform preprocessing of multidimensional data, including data cleaning, time synchronization, and feature extraction. The collaborative control unit uses this main controller to receive external scheduling commands, allocate tasks, and send load adjustment requests under abnormal conditions. The main controller is also responsible for sensor data acquisition, execution unit driving, and CAN bus communication.

[0042] In another specific implementation, the core algorithm unit and the thermal runaway prediction and intervention unit are implemented using an edge AI chip. This chip employs an edge AI chip with a computing power of at least 16 TOPS, supporting INT8 and FP16 precision operations, such as the Huawei Ascend 310B. The self-evolving dual-loop AI algorithm module and the cell digital twin engine module of the core algorithm unit run on this chip, and the graph neural network model of the thermal runaway prediction and intervention unit also performs inference operations on this chip.

[0043] To reduce the computational load on the edge AI chip, the algorithm core unit is also equipped with an FPGA acceleration module. This FPGA has at least 20K logic cells and 120 DSP slices, such as a Xilinx Artix-7, specifically designed to accelerate the real-time mapping computation of the cell twin in the cell digital twin engine module.

[0044] The storage module uses a combination of 64GB eMMC and 4GB DDR4, with a read and write speed of no less than 200MB / s. It supports power-loss data protection and is used to store the battery cell's full life cycle operation data, model parameters, and log files for the algorithm core unit.

[0045] In another specific embodiment, the sensing unit includes a cell micro-sensing subunit, a battery body sensing subunit, and a cross-domain sensing subunit. The cell micro-sensing subunit is used to collect data on the tab temperature, cell expansion, and separator impedance. In one specific embodiment, this subunit uses a PT1000 miniature platinum resistance sensor to measure the tab temperature, with an accuracy of ±0.1℃ and a range of -40℃ to 120℃; it uses a thin-film strain gauge sensor to measure the cell expansion, with an accuracy of ±0.01mm and a range of 0 to 5mm; and it uses a built-in impedance detection electrode to measure the separator impedance, with an accuracy of ±0.01Ω and a range of 0 to 10Ω.

[0046] The battery body sensing subunit is used to collect the voltage, current, and temperature of the battery cells. In one specific embodiment, this subunit uses a TI ADS1256 voltage acquisition chip to collect the individual cell voltage with an accuracy of ±0.005V and a sampling rate of 10Hz; it uses a LEM LAH 300-S Hall current sensor to collect the total current with an accuracy of ±0.1A and a response time of no more than 1ms; and it uses an NTC thermistor to collect the module temperature with an accuracy of ±0.2℃.

[0047] The cross-domain sensing subunit is used to collect parameters from the power grid, environment, and load sides. In one specific implementation, the power grid side uses a Huawei DTU 650 terminal and a WAGO 850-704 voltage sensor, with a power grid frequency accuracy of ±0.1Hz and a voltage fluctuation accuracy of ±1%, supporting the IEC 61850 protocol. The environment side is equipped with a BH1750 light sensor, an RS-FS01 wind speed sensor, and an SHT30 temperature and humidity sensor, with a light accuracy of ±10 lux, a wind speed accuracy of ±0.1 m / s, and a temperature and humidity accuracy of ±0.5℃ and ±2% RH. The load side uses a State Grid DDZY108 smart meter and a customized load identification module, with a power accuracy of ±0.5%, supporting the classification and identification of industrial, residential, and commercial loads.

[0048] The execution components include an intelligent equalization circuit, a gradient cooling system, an independent cell isolation switch, and a directional pressure relief and inert gas device, used to respond to control commands issued by the collaborative control unit and the thermal runaway prediction and intervention unit.

[0049] The intelligent balancing circuit is used to perform balancing tasks issued by the collaborative control unit. In one specific embodiment, the circuit adopts a bidirectional Buck-Boost topology, paired with an IRF3205 MOSFET, and the balancing current can be adaptively adjusted within the range of 0.1A to 5A, with an efficiency of no less than 95%.

[0050] The gradient cooling system is used to execute cooling commands issued by the collaborative control unit and the thermal runaway prediction and intervention unit. In one specific embodiment, the system integrates axial flow fan cooling and microchannel liquid cooling, with the air cooling velocity ranging from 0.5 m / s to 3 m / s and the liquid cooling rate not less than 5 °C / min.

[0051] The cell-independent disconnect switch is used to execute circuit disconnection and fault cell isolation commands issued by the collaborative control unit and the thermal runaway prediction and intervention unit. In one specific embodiment, the switch uses a Hongfa HF115F magnetic latching relay with a rated voltage of 1000V DC, a rated current of 200A, and an operating time of no more than 1s.

[0052] The directional pressure relief and inert gas device is used to execute high-risk emergency protection commands issued by the thermal runaway prediction and intervention unit. Preferably, the device includes a rupture disc with a burst pressure of 1.2 MPa and a nitrogen storage tank with a volume of 5 L, a pressure relief response time of no more than 0.5 s, and a nitrogen injection flow rate of no less than 10 L / min.

[0053] In one specific embodiment, the energy storage battery management system further includes a fusion interface unit. This fusion interface unit integrates a CAN FD bus, a 100Mbps Ethernet port, and an OTA upgrade interface. It supports a wide voltage input range of 3.3V to 12V, has 2A overcurrent protection, and an IP67 protection rating. This fusion interface unit is compatible with various types of battery cells, including lithium-ion, sodium-ion, and vanadium redox flow cells, and supports hot-swapping without data loss during the process. The fusion interface unit connects the sensing unit and the data fusion unit, and supports communication between the collaborative control unit and an external dispatch center.

[0054] In one specific implementation, the data fusion unit is used to preprocess the multidimensional data collected by the sensing unit to ensure the reliability, synchronization and effectiveness of the data.

[0055] Data Cleaning: The data fusion unit first initializes the extended Kalman filter, setting the state covariance matrix to a diagonal matrix [0.01, 0.01, 0.01] and the measurement noise matrix to a diagonal matrix [0.005, 0.1, 0.2] for the three-dimensional core data of voltage, current, and temperature. This measurement noise matrix matches the sensor accuracy. After receiving the raw data, the voltage, current, and temperature values ​​are extracted, and the prediction and update steps of the extended Kalman filter are performed to complete the data denoising. Subsequently, the 3σ criterion is used to identify outliers in the voltage parameters: if the value deviates from the mean by more than 3 times the standard deviation, the outlier is replaced with the sequence mean to ensure data reliability.

[0056] Time synchronization: The data fusion unit uses the timestamp of the cell micro data as a reference and traverses the data of other dimensions. If the time difference exceeds 1ms, the data value is corrected by interpolation and the timestamp is unified so that the time synchronization error of the multi-dimensional data is no more than 1ms.

[0057] Feature Extraction: The data fusion unit extracts time-domain and frequency-domain features from the fused data. Time-domain features include mean voltage, current variance, maximum temperature rise rate, and membrane impedance abrupt change rate; frequency-domain features are extracted by performing a Fast Fourier Transform on the voltage sequence to extract the dominant frequency peak. Combined with the mean cell expansion, tab temperature gradient, and electrolyte decomposition product concentration, the high-dimensional data is compressed into an 8-dimensional effective feature vector, which is then output to the algorithm core unit.

[0058] Building upon this foundation, the core algorithm units include a self-evolving dual-loop AI algorithm module and a cell digital twin engine module. The self-evolving dual-loop AI algorithm module features a dual-loop self-evolution mechanism with online fine-tuning loops and offline iterative loops. The online fine-tuning loop submodule deploys an incremental learning long short-term memory network model. This model takes the 8-dimensional feature vector output from the data fusion unit as input and the cell's state of charge and health state as output. The model contains a long short-term memory layer and a fully connected layer, taking the output result of the last time step of the long short-term memory layer.

[0059] The online fine-tuning loop is triggered when 100 sets of cell cycle data are collected, or when a sudden change occurs in cross-domain data, including grid voltage fluctuations exceeding 5%. During fine-tuning, a mean squared error loss function and the Adam optimizer are used. The weighted sum of the current task loss and distillation loss is used as the total loss, where the distillation loss incorporates predictions from the old model. A learning-without-forgetting strategy is used to prevent the model from forgetting historical knowledge. After training, the state of charge estimation error is verified. If the error exceeds 2%, the learning rate is reduced and fine-tuning is performed again. This submodule ensures that the fine-tuning time does not exceed 5 seconds, the state of charge estimation error is not greater than 2% over a long period, the health state prediction accuracy is not less than 98%, and finally, the updated model parameters are returned.

[0060] The offline iterative loop submodule deploys a federated learning mechanism for aggregating and updating model parameters across multiple local nodes.

[0061] During local node training, each local node trains the model independently, using the Adam optimizer and mean squared error loss function. Only the model parameters and loss values ​​are output, without disclosing the original data to protect privacy.

[0062] The federated aggregation strategy employs a dynamic weighting method, where weights are calculated as the ratio of the product of local data volume and operational similarity to the sum of all node weights. After weighted averaging of the parameters of each local node, the aggregated global parameters are protected using the Paillier homomorphic encryption algorithm to prevent leakage during transmission.

[0063] The main process of federated learning iterates the global model every 90 days. First, training data is collected from each anonymized local node, the global model is cloned to complete local training and obtain parameters; after aggregating and decrypting the local parameters, the global model is updated, and then the updated model is distributed to each local node to improve the model's generalization ability.

[0064] The cell digital twin engine module builds a unique digital twin for each cell, enabling refined management at the cell level.

[0065] Initialization: This module integrates the cell's manufacturing process parameters and factory inspection data to perform twin initialization. Manufacturing process parameters include electrode thickness, electrolyte injection amount, and the cell's unique identifier; factory inspection data includes initial internal resistance and initial capacity. The initialized cell state parameters include state of charge, state of health, internal resistance, and expansion amount, with an initial error not exceeding 3%.

[0066] Real-time mapping: This module receives multi-dimensional fused sensor data output from the data fusion unit, extracts 8-dimensional feature vectors, inputs them into a pre-trained long short-term memory network model, and predicts state changes. Subsequently, it corrects the state parameters based on physical constraints, including limiting the state of charge to the range of 0 to 1, ensuring that the healthy state only decays irreversibly, and that the internal resistance does not exceed 5 times the initial value. Through the above processing, the cell state is dynamically updated, with a mapping delay of no more than 10ms.

[0067] Strategy Simulation Optimization: This module supports simulating the impact of candidate control strategies on the cell's health status. It iterates through candidate control strategies, including combinations of charge / discharge rates and equalization currents, and simulates the health status after 100 cycles using a simplified electrochemical model. The module calculates the single-cycle degradation rate and selects the optimal strategy with the lowest degradation rate. The simulation takes less than 200ms and can achieve a 30% reduction in the cell degradation rate.

[0068] In one specific implementation, the collaborative control unit executes the following collaborative control logic with a task cycle of 100ms.

[0069] The collaborative control unit receives dispatch instructions from the virtual power plant through a dual-backup approach, using the 5G module as the primary communication channel and the CAN bus as the backup. Upon receiving the instruction, it invokes the cell digital twin engine module to update the cell status based on multi-dimensional fused sensor data for each cell.

[0070] The task allocation submodule categorizes battery cells according to their health status: cells with a health status of 90% or higher are considered healthy; cells with a health status between 80% and 90% are considered average; and cells with a health status below 80% are considered degraded. Healthy cells handle 70% of the high-frequency response power, average cells handle 30% of the basic energy storage power, and degraded cells are prohibited from active charging and discharging. The charge / discharge rate and equalization current adjustment are then performed, a process that takes no more than 500ms.

[0071] The status monitoring and feedback submodule monitors the worst-performing cell in real time. If its temperature is not lower than 50°C or its expansion is not lower than 0.5mm, it sends a request to the virtual power plant to reduce the load to 50% through the 5G module and starts the gradient cooling system of the corresponding cell. If the status is normal, it feeds back the actual executed power to the virtual power plant, forming a collaborative control closed loop.

[0072] In one specific implementation, the thermal runaway prediction and intervention unit includes a risk prediction submodule and a graded intervention submodule.

[0073] The risk prediction submodule constructs a graph neural network model for predicting thermal runaway. This model consists of two graph convolutional layers and one fully connected layer. The input consists of a node feature matrix and an adjacency matrix with seven parameters. The output is the probability distribution of three risk levels: low, medium, and high. The adjacency matrix is ​​constructed by calculating the correlation matrix of the seven parameters and retaining strongly correlated edges with an absolute value not less than 0.3 as the adjacency matrix of the graph neural network. The seven parameters include voltage, current, temperature, micro-vibration, gas concentration, tab temperature gradient, and diaphragm impedance abrupt change rate.

[0074] The risk prediction submodule is also used to locate the highest-risk battery cells. A risk score is calculated based on the anomaly degree weighting of seven-dimensional parameters, where the tab temperature gradient has a weight of 0.3, the separator impedance mutation rate has a weight of 0.25, and the remaining parameters each have a weight of 0.09. The highest-risk battery cell is identified based on the risk score.

[0075] The tiered intervention submodule takes 10 consecutive frames of seven-dimensional monitoring data as input and predicts the risk level using a pre-trained graph neural network model. This model is trained with a mixture of real and simulated data, improving its generalization ability by 40%. Based on the predicted risk level, a tiered intervention strategy is executed: for low risk, air cooling is activated and the charge / discharge rate is limited to no more than 0.8C; for medium risk, air cooling and liquid cooling are activated in tandem and the charge / discharge circuit is cut off; for high risk, directional liquid cooling, nitrogen injection, faulty cell isolation, and remote alarm are implemented.

[0076] In one specific implementation, the collaborative control unit achieves bidirectional communication with the external scheduling center through a fusion interface unit.

[0077] The communication process between the collaborative control unit and the external dispatch center is as follows: The system-side protocol parsing module first receives instructions from the virtual power plant, cloud platform, or local operation and maintenance terminal, encrypts the data using the AES-256 encryption algorithm, and then sends it to the network transmission layer via the 5G industrial module or Ethernet transmission module. The receiving end receives the data through the corresponding transmission module, decrypts it, and verifies its integrity using CRC32, before issuing valid instructions to the execution component. After completing the operation, the execution component feeds back the execution result to the interactive object according to the same process, forming a two-way communication closed loop.

[0078] In terms of protocol adaptation, the collaborative control unit communicates with the virtual power plant using the IEC 61850-7-420 energy storage protocol, with a dispatch command transmission latency of no more than 10ms. Communication with the cloud-based operation and maintenance platform uses MQTT and HTTPS protocols, supporting remote data upload, model upgrades, and status monitoring. Communication with the local operation and maintenance terminal uses RS232 and WebSocket protocols, enabling fault query, parameter configuration, and log export functions, with a transmission latency of no more than 10ms.

[0079] Furthermore, the technical solution of the present invention can be adapted and deployed according to different application scenarios, as detailed below: Grid-side energy storage scenario: For large-scale energy storage systems of 100MWh or more, at the hardware level, 2 to 4 Huawei Ascend 310B edge AI chips are added to meet the computational needs of large-scale cell twins through computing power aggregation, while expanding the micro liquid cooling channels to be independently configured for each cell. At the software level, the number of local nodes in federated learning is expanded to 20, the weight of the collaborative strategy is optimized, and the grid frequency regulation response speed is used as the core indicator to ensure that the response time is no more than 2 seconds.

[0080] For industrial and commercial energy storage scenarios: For industrial and commercial energy storage systems ranging from 1MWh to 10MWh, the hardware adopts a single Huawei Ascend 310B edge AI chip, and the cooling system is mainly air-cooled, with local liquid cooling only configured in high-heat areas. At the software level, the peak shaving and valley filling strategies are optimized to adapt to the power fluctuation characteristics of industrial loads within ±30% within 5 minutes. A new long short-term memory network load prediction module has been added, which can predict load change trends one hour in advance.

[0081] For residential energy storage scenarios: For residential energy storage systems ranging from 5kWh to 20kWh, the hardware is simplified to a combination of an STM32H743 main controller and a lightweight AI module. The lightweight AI module has a computing power of 0.5 TOPS, and the cooling system retains only high-efficiency air cooling. On the software side, the focus is on security and extended lifespan, limiting the charge / discharge rate to 0.3C to 0.8C, eliminating the federated learning module, and retaining only the local incremental learning function to reduce hardware costs and energy consumption.

[0082] Sodium-ion battery scenario: When adapting to sodium-ion batteries, the hardware level adjusts the expansion sensor sampling rate to 3Hz to match the low expansion rate characteristics of sodium-ion batteries. The software level optimizes the physical model of the battery cell digital twin engine module to match the electrochemical characteristics of sodium ions, adjusting the charge / discharge cutoff voltage to 2.5V to 4.0V and expanding the equalization threshold to 8mV to ensure precise matching with the characteristics of sodium-ion batteries.

[0083] The present invention has the following advantages over the prior art: At the sensing level, this invention achieves direct acquisition of microscopic state parameters such as tab temperature, cell expansion, and separator impedance by embedding micro sensors inside or on the surface of the battery cell. This overcomes the limitations of traditional BMS, which relies solely on macroscopic parameters to indirectly infer the internal state of the battery cell. Simultaneously, by synchronously acquiring cross-domain data from the grid, environment, and load sides, a multi-dimensional sensing system integrating the cell's microscopic state, battery operating parameters, and external scenario parameters is constructed. This provides a rich and reliable data foundation for precise control, solving the problems of existing technologies having a single sensing dimension and being "invisible and unclear" regarding the internal state of the battery cell.

[0084] At the algorithm level, the self-evolving double-loop AI algorithm framework proposed in this invention enables the model to adapt to the aging of battery cells and the changes in operating conditions in real time through the online fine-tuning loop, and introduces a distillation loss function based on the learning without forgetting strategy, effectively preventing the model from forgetting historical knowledge during the incremental learning process; through the offline iteration loop, multiple nodes can jointly evolve without leaking the original data, and a dynamic weight aggregation method based on the data volume and operating condition similarity is introduced to improve the generalization ability of the global model in specific scenarios. This double-loop mechanism fundamentally solves the technical problems of accuracy decay and inability to continuously optimize after the deployment of traditional fixed models, making the estimation error of the state of charge stable within 2% in the long term, and the prediction accuracy of the state of health not less than 98%.

[0085] At the management granularity level, this invention constructs a dedicated digital twin for each battery cell by integrating production process parameters and factory inspection data, achieving a leap from the "one-size-fits-all" management at the battery pack level to the "tailor-made" management at the cell level. This twin engine can not only map the cell state in real time with high precision, but also support the simulation and optimization of candidate control strategies, screening the optimal charge and discharge strategies for each cell, effectively reducing the impact of individual cell differences on the overall life of the battery pack, and extending the cycle life of the cell by more than 30%.

[0086] At the cooperation mechanism level, this invention breaks through the limitation of the traditional BMS that only responds to grid commands unidirectionally, and establishes a two-way cooperation control mechanism with the real-time health state of the battery cell as the core constraint. By differentially allocating grid tasks according to the health level, it not only meets the scheduling requirements of the grid such as frequency modulation and peak shaving and valley filling, but also avoids overusing the degraded battery cells; at the same time, when the cell state is abnormal, the system can actively request the grid to adjust the load in the reverse direction, achieving a win-win situation for ensuring the life of the battery cell and the safe operation of the grid.

[0087] At the safety protection level, this invention innovatively applies the graph neural network to the pre-judgment of thermal runaway. By mining the implicit correlations among the seven-dimensional parameters of voltage, current, temperature, micro-vibration, gas concentration, temperature gradient of the tab, and mutation rate of the diaphragm impedance, it can identify the risk evolution trend in advance, and the early warning lead time reaches more than 18 minutes. At the same time, according to the three risk levels of low, medium, and high, differential hierarchical intervention strategies such as air cooling, liquid cooling, circuit cutting, cell isolation, and nitrogen injection are matched, achieving a leap from passive alarm to active pre-judgment and from single disposal to precise linkage, significantly improving the safety of the energy storage system.

[0088] In summary, through systematic innovations in the five aspects of perception, algorithm, model, cooperation, and protection, this invention constructs a full-link, closed-loop intelligent energy storage battery management system, achieving significant improvements in key indicators such as perception accuracy, estimation accuracy, cycle life, cooperation efficiency, and safety protection, and having prominent substantive features and significant progress.

[0089] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0090] This embodiment uses a 5MWh industrial and commercial energy storage system in an automotive parts processing plant as an example. The plant employs approximately 1000 people and experiences significant industrial load fluctuations, with power surges reaching ±25% within 5 minutes. The energy storage system has a capacity of 5MWh, an AC output power of 4MW, and is compatible with a 380V industrial power grid. It utilizes lithium iron phosphate square cells with specifications of 3.2V / 280Ah and a cycle life of no less than 3000 cycles. The module configuration consists of 16 cells connected in series to form one module, and 24 modules connected in parallel to form one energy storage cabinet, for a total of 75 energy storage cabinets. The system operates in an indoor energy storage room with a temperature range of -10℃ to 45℃, a humidity range of 30% to 80%, and an IP54 protection rating.

[0091] In terms of hardware deployment, the data fusion unit and the collaborative control unit are implemented by the same main controller. Each energy storage cabinet is equipped with one STM32H743IGT6 main controller, and the 75 cabinets are networked via a CAN bus. The main cabinet MCU acts as the central control node, responsible for global coordination. The algorithm core unit and the thermal runaway prediction and intervention unit are implemented by an edge AI chip. The main cabinet is equipped with one Huawei Ascend 310B edge AI chip with a computing power of 16 TOPS. It is equipped with a self-evolving dual-loop AI algorithm module, a cell digital twin engine module, and a thermal runaway prediction model, supporting parallel computation of the cell status of 75 cabinets. To reduce the load on the AI ​​chip, the algorithm core unit is also equipped with a Xilinx Artix-7 FPGA acceleration module, which is directly connected to the AI ​​chip via a PCIe 2.0 interface. It is specifically responsible for the real-time mapping computation of the digital twins of 28,800 cells, with a single cell mapping latency of no more than 8ms. The storage module uses a combination of 64GB eMMC and 4GB DDR4, with a read and write speed of no less than 200MB / s. It supports power-loss data protection and is used to store the battery cell's full life cycle operation data, model parameters, and log files.

[0092] The sensing unit comprises a cell micro-sensing subunit, a battery body sensing subunit, and a cross-domain sensing subunit. In the cell micro-sensing subunit, each cell embeds a PT1000 miniature platinum resistance sensor to measure the tab temperature with an accuracy of ±0.1℃; a thin-film strain gauge measures the cell expansion with an accuracy of ±0.01mm; and an integrated impedance detection electrode measures the separator impedance with an accuracy of ±0.01Ω. All micro-sensors are connected to the sub-cabinet MCU via a CAN FD bus with a transmission rate of 5Mbps. In the battery body sensing subunit, each module is equipped with a TI ADS1256 voltage acquisition chip to acquire the voltage of 16 cells with an accuracy of ±0.005V; a LEM LAH 300-S Hall current sensor acquires the total current with an accuracy of ±0.1A; and four NTC thermistors monitor the module temperature with an accuracy of ±0.2℃. In the cross-domain sensing subunit, the power grid side is equipped with a Huawei DTU 650 terminal and a WAGO 850-704 voltage sensor to collect power grid frequency and voltage fluctuations in real time; the environmental side is equipped with a BH1750 light sensor, an RS-FS01 wind speed sensor and an SHT30 temperature and humidity sensor; the load side uses a State Grid DDZY108 smart meter to collect load power, and a customized load identification module is used to identify the type of industrial load.

[0093] The execution components include an intelligent balancing circuit, a gradient cooling system, independent cell isolation switches, and directional pressure relief and inert gas devices. Each module is equipped with one bidirectional Buck-Boost balancing circuit, with the balancing current adaptively adjustable within the range of 0.1A to 5A and an efficiency of no less than 95%. Each energy storage cabinet is equipped with two axial flow fans and one set of miniature liquid cooling channels, with an air cooling velocity of 0.5m / s to 3m / s and a liquid cooling rate of no less than 5℃ / min. Each module is equipped with one Hongfa HF115F magnetic latching relay, with a rated voltage of 1000V DC, a rated current of 200A, and an operating time of no more than 1s. Each energy storage cabinet is equipped with one 1.2MPa rupture disc and one 5L nitrogen storage tank on the top, with a pressure relief response time of no more than 0.5s and a nitrogen injection flow rate of no less than 10L / min. Each cabinet is also equipped with a converged interface module with an IP67 protection rating. It integrates a CAN FD bus, a 100Mbps Ethernet and OTA upgrade interface, supports a wide voltage input from 3.3V to 12V, and is compatible with various types of battery cells, including lithium-ion, sodium-ion, and vanadium redox flow batteries.

[0094] In terms of software operation, the data fusion unit collects multi-dimensional data every 100ms. During data cleaning, an extended Kalman filter is used to denoise voltage, current, and temperature, reducing voltage noise from ±0.01V to ±0.003V. Outliers are removed using the 3σ criterion, with the monthly average outlier rate not exceeding 0.05%. Time synchronization is based on the NTP time of the main cabinet MCU, and sub-cabinet data is synchronized via timestamps, with an error not exceeding 1ms. Feature extraction extracts an 8-dimensional feature vector every 1s, including the mean voltage, current variance, maximum temperature rise rate, diaphragm impedance mutation rate, peak value of the voltage sequence FFT main frequency, mean cell expansion, tab temperature gradient, and electrolyte decomposition product concentration, which is input to the algorithm core unit in real time.

[0095] In the self-evolving dual-loop AI algorithm module, the online fine-tuning loop submodule deploys an incremental learning long short-term memory network model, taking an 8-dimensional feature vector as input and outputting the cell's state of charge (SOC) and health status. Fine-tuning is triggered every 100 sets of cell loop data collected, employing incremental learning based on a learning-without-forgetting strategy. The total loss is calculated by weighted summation of the current task loss and distillation loss, with fine-tuning taking no more than 4.5 seconds. The SOC estimation error is consistently within 1.8%, and the health status prediction accuracy is no less than 98.2%. The offline iterative loop submodule, in conjunction with energy storage systems from three similar surrounding factories, forms a federated learning network, iterating the global model once per quarter. After local nodes independently train the model, only the model parameters are uploaded. Paillier homomorphic encryption is used to protect the aggregated parameters. The weights are calculated as the ratio of the product of the local data volume and the operating condition similarity to the sum of the weights of all nodes. After the global model is updated, the SOC estimation error is further reduced to within 1.5%.

[0096] The cell digital twin engine module constructs a unique digital twin for each cell. During initialization, it integrates cell manufacturing process parameters and factory testing data, including electrode thickness of 0.15mm, electrolyte injection volume of 12g, initial internal resistance of 80mΩ, and initial capacity of 280Ah, constructing digital twins for 28,800 cells with an initial error not exceeding 2.8%. During real-time mapping, an 8-dimensional feature vector is received every 100ms as input to a pre-trained long short-term memory network model to predict state changes. Combined with physical constraints, the state parameters are corrected, limiting the state of charge to the range of 0 to 1, ensuring the healthy state only decays irreversibly, and maintaining internal resistance no more than 5 times the initial value. The mapping delay is no more than 8ms, and the state parameter error is no more than 1.9%. During strategy simulation, the charging and discharging strategy for the next day is simulated at 2 AM daily. A simplified electrochemical model is used to simulate the healthy state after 100 cycles, selecting the optimal strategy with the lowest decay rate. This process takes no more than 180ms, reducing the daily decay rate of the cell by 32%.

[0097] The collaborative control unit executes collaborative control with a task cycle of 100ms. It receives virtual power plant scheduling instructions daily at 00:00, with peak shaving and valley filling targets of 3MW output power from 10:00 to 16:00 during the day and 3MW charging power from 00:00 to 06:00 at night. The cell digital twin engine module is invoked to select approximately 21,600 healthy cells (health status not lower than 90%) for frequency regulation tasks, approximately 5,760 average cells (health status between 80% and 90%) for peak shaving and valley filling tasks, and approximately 1,440 degraded cells (health status below 80%) with a charge / discharge rate limited to no more than 0.5C. Task allocation and execution time does not exceed 500ms. Cell status is monitored in real time. When the temperature of a cell in a module is not lower than 50℃ or the expansion is not lower than 0.5mm, a load reduction request is immediately sent to the virtual power plant via the MQTT protocol, requesting a load reduction to 2MW, and simultaneously activating the gradient cooling system of that cabinet. When in normal condition, the actual execution results are fed back to the virtual power plant via the IEC 61850-7-420 protocol, with a response time of no more than 2.8s and a power regulation accuracy of ±0.8kW.

[0098] In the thermal runaway prediction and intervention unit, the risk prediction submodule builds a graph neural network model, containing two graph convolutional layers and one fully connected layer. The input is 10 consecutive frames of seven-dimensional monitoring data, including voltage, current, temperature, micro-vibration, gas concentration, tab temperature gradient, and diaphragm impedance mutation rate. The adjacency matrix is ​​constructed based on the correlation matrix of the seven-dimensional parameters, retaining strongly correlated edges with an absolute value not less than 0.3. The model is trained with 10 sets of real data and 1000 sets of simulation data, achieving a low-risk identification accuracy of not less than 95.3%, a medium-to-high-risk identification accuracy of not less than 98.5%, and a warning lead time of not less than 18 minutes. A risk score is calculated based on the weighted average of the seven-dimensional parameter anomalies to pinpoint the cell with the highest risk. The graded intervention submodule executes intervention strategies according to the predicted risk level: for low risk, air cooling is activated and the charge / discharge rate is limited to no more than 0.8C; for medium risk, air cooling and liquid cooling are activated in conjunction, and the charge / discharge circuit is cut off; for high risk, directional liquid cooling, nitrogen injection, faulty cell isolation, and remote alarm are implemented. In one instance where a short circuit inside the battery cell caused the temperature to rise from 35°C to 60°C within 10 seconds, the system triggered a medium-risk intervention, activated the linkage between air cooling and liquid cooling, and cut off the isolation switch of the module. Within 15 seconds, the temperature dropped to 45°C without thermal diffusion.

[0099] In terms of collaborative interaction, the collaborative control unit communicates with the virtual power plant using the IEC 61850-7-420 protocol, transmitted via a 5G industrial module. The delay in receiving scheduling commands is no more than 8ms, and the delay in feedback of execution results is no more than 9ms. Communication with the cloud platform uses MQTT and HTTPS protocols, uploading cell status, operating data, and fault logs in real time at a frequency of once every 10 seconds, supporting remote model upgrades and parameter configuration. The local maintenance terminal connects to the main cabinet via RS232 and WebSocket protocols, supporting fault query, parameter configuration, and log export. For industrial and commercial energy storage scenarios, this embodiment introduces a new long short-term memory network load prediction module, predicting factory load changes 1 hour in advance with an accuracy rate of no less than 92%. When the predicted load changes by ±25%, the output power is adjusted in advance. The cooling strategy is dynamically adjusted according to industrial load characteristics, activating air cooling and liquid cooling in conjunction during high daytime loads, and only air cooling during low nighttime loads, reducing daily cooling energy consumption by 15%.

[0100] Through actual operation verification, this embodiment improves the health status prediction accuracy by 15% compared with the traditional BMS by embedding micro sensors and cell digital twin engine; adopts a dual-loop strategy of federated learning and incremental learning, and the state of charge estimation error has been kept within 2% for a long time; assigning tasks according to the cell health level extends the cell life by 32%; and adopts graph neural network model and hierarchical intervention strategy to successfully avoid three potential thermal runaway accidents and ensure the safe operation of the system.

[0101] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0102] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for descriptive purposes only and is not intended to limit the scope of protection. As used in the description herein, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used herein means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0103] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0104] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

Claims

1. A cross-domain sensing self-evolving energy storage battery management system, characterized in that, include: The sensing unit is used to collect microscopic state parameters of the battery cell, operating parameters of the battery body, and external scene parameters, wherein the external scene parameters include at least one of grid-side parameters, environmental-side parameters, and load-side parameters. The data fusion unit, connected to the sensing unit, is used to preprocess the received multidimensional data, including data cleaning, time synchronization and feature extraction, and output the fused feature vector. The core algorithm unit, connected to the data fusion unit, includes a self-evolving dual-loop AI algorithm module and a cell digital twin engine module. The self-evolving dual-loop AI algorithm module is used to estimate the cell state parameters based on the feature vector and has a dual-loop self-evolving mechanism of online incremental learning and offline iterative update. The cell digital twin engine module is used to construct a unique digital twin for each cell. The twin integrates the cell's production process parameters and factory inspection data for initialization and maps the cell's current state in real time based on the feature vector. The collaborative control unit is communicatively connected to both the algorithm core unit and the external scheduling center. It receives external scheduling instructions, allocates scheduling tasks to each cell in a differentiated manner based on the cell health status output by the cell digital twin engine module, and sends a load adjustment request to the external scheduling center when an abnormal cell status is detected.

2. The energy storage battery management system according to claim 1, characterized in that, The sensing unit includes: The cell micro-sensing subunit is embedded in the cell or on its surface and is used to collect at least one micro-state parameter among tab temperature, cell expansion and diaphragm impedance. The battery body sensing subunit is used to collect at least one of the following operating parameters of the battery cell: voltage, current and temperature. The cross-domain sensing subunit is used to collect at least one external scene parameter from the following: power grid frequency, voltage fluctuation, ambient temperature and humidity, light intensity, wind speed, and load power.

3. The energy storage battery management system according to claim 1, characterized in that, The self-evolving dual-loop AI algorithm module includes: The online fine-tuning loop submodule deploys an incremental learning model, which is used to fine-tune the current model based on newly collected data under preset trigger conditions, and introduces a historical knowledge retention mechanism during the fine-tuning process to prevent the model from forgetting. The offline iterative loop submodule deploys a federated learning mechanism to aggregate and update model parameters across multiple local nodes, and encrypts and protects the model parameters during the aggregation process.

4. The energy storage battery management system according to claim 3, characterized in that, The incremental learning model in the online fine-tuning recurrent submodule is a long short-term memory network model, and the historical knowledge retention mechanism is a distillation loss function based on a learning-without-forgetting strategy.

5. The energy storage battery management system according to claim 3, characterized in that, When aggregating model parameters, the offline iterative loop submodule calculates dynamic weights based on the data volume and operating condition similarity of each local node, and performs a weighted average of the model parameters uploaded by the local nodes based on the dynamic weights.

6. The energy storage battery management system according to claim 1, characterized in that, The cell digital twin engine module is also used to simulate the impact of candidate control strategies on the cell health status and to select the optimal control strategy based on the simulation results.

7. The energy storage battery management system according to claim 1, characterized in that, The collaborative control unit includes: The task allocation submodule is used to allocate the power tasks corresponding to external scheduling commands to cells with different health levels based on the cell health status level. The status monitoring and feedback submodule is used to monitor the cell status parameters in real time. When the cell status parameters exceed the preset threshold, it sends a load adjustment request to the external dispatch center and starts the corresponding thermal management execution unit.

8. The energy storage battery management system according to any one of claims 1 to 7, characterized in that, Also includes: The thermal runaway prediction and intervention unit, connected to the sensing unit and the algorithm core unit, is used to predict the thermal runaway risk level based on continuous multi-frame monitoring data including seven-dimensional monitoring data such as voltage, current, temperature, micro-vibration, gas concentration, electrode temperature gradient and diaphragm impedance mutation rate, through a pre-trained graph neural network model, and to execute a graded intervention strategy according to the risk level.

9. The energy storage battery management system according to claim 8, characterized in that, The thermal runaway prediction and intervention unit includes: The risk prediction submodule takes as input seven-dimensional monitoring data, including voltage, current, temperature, micro-vibration, gas concentration, electrode temperature gradient and diaphragm impedance mutation rate, and its adjacency matrix is ​​constructed based on the correlation matrix of the seven-dimensional monitoring data. The graded intervention submodule is used to match and execute one or more intervention actions from air cooling, liquid cooling, charge / discharge circuit disconnection, faulty cell isolation, and nitrogen injection based on the predicted risk level.

10. A protection method based on the cross-domain sensing self-evolving energy storage battery management system according to any one of claims 1 to 9, characterized in that, Includes the following steps: Step S1: Collect cell micro-state parameters, battery operating parameters, and external scene parameters through the sensing unit; Step S2: The collected multidimensional data is preprocessed by the data fusion unit, and the fused feature vector is output. Step S3: Estimate the cell state parameters through the self-evolving dual-loop AI algorithm module of the core algorithm unit, and map the current state of the cell in real time through the cell digital twin engine module; Step S4: Receive external scheduling instructions through the collaborative control unit, allocate scheduling tasks to each cell in a differentiated manner according to the cell health status, and send a load adjustment request to the external scheduling center when an abnormal cell status is detected; Step S5: The thermal runaway prediction and intervention unit performs risk prediction on continuous multi-frame monitoring data and executes graded intervention strategies according to the risk level.

Images