Cloud edge collaborative intelligent management method and device for energy storage battery

Abstract

The invention provides an energy storage battery management system and a control method, belongs to the field of energy storage battery management, and is used for solving the problems of low battery state estimation precision, insufficient fault diagnosis sensitivity and poor full life cycle adaptability in related technologies. By combining a neural network-Kalman filtering cascaded state estimation algorithm and a twin structure feature coding fault diagnosis algorithm and matching with an edge-cloud collaborative model self-evolution mechanism, accurate estimation of a battery state and timely diagnosis and positioning of a fault are realized, and the full-working-condition adaptability and operation reliability of a system are improved.

Description

Technical Field

[0001] This application relates to the field of energy storage battery management, and more particularly to a cloud-edge collaborative intelligent management method and device for energy storage batteries. Background Technology

[0002] With the rapid development of the new energy industry, energy storage batteries, as core equipment for energy storage and dispatch, are widely used in power systems, new energy vehicles, and other fields. The operating status of energy storage batteries directly affects the safety, stability, and economy of the entire system; therefore, efficient and precise management of energy storage batteries has become a key research focus in this field.

[0003] In existing technologies, energy storage battery management mostly adopts a single edge-side local management or cloud-based centralized management mode. The single edge-side management mode is limited by local computing resources and is difficult to achieve global optimization across multiple sites; the single cloud-based management mode suffers from high data transmission latency, which cannot meet the needs of real-time battery status monitoring and rapid fault response. At the same time, the state estimation algorithms in existing management methods mostly use a single model, which is difficult to adapt to the characteristics of batteries under different temperatures and aging levels, and the fault diagnosis algorithms have weak ability to identify micro-faults, which are prone to missed or false diagnoses.

[0004] Furthermore, existing technologies lack efficient algorithms for edge-cloud collaboration. The data preprocessing stage is too generic and cannot match the characteristics of energy storage battery operating data, resulting in low-quality input data for core algorithms. Simultaneously, model parameter updates rely heavily on data from single sites, making it difficult to adapt to the varying operating states of batteries across multiple sites, further impacting the overall effectiveness and applicability of management methods. These shortcomings make it difficult for existing energy storage battery management methods to simultaneously meet the demands of real-time performance, accuracy, and global optimization, thus failing to meet the technical requirements for the large-scale development of the energy storage industry. Summary of the Invention

[0005] This application provides a cloud-edge collaborative intelligent management method and device for energy storage batteries. It can solve the shortcomings of existing energy storage battery management methods by combining a cloud-edge collaborative architecture with multi-algorithm fusion design, taking into account the requirements of real-time performance, accuracy and global optimization.

[0006] Firstly, this application provides a cloud-edge collaborative intelligent management method for energy storage batteries. It includes an edge-side data acquisition step, an edge-side core computing step, a data upload step, a cloud-based global optimization step, a strategy distribution and fusion step, and an execution step. The edge-side data acquisition step involves collecting the operating parameters of the energy storage battery through multi-dimensional sensing devices. The edge-side core computing step involves an edge intelligent terminal calling a first algorithm to process the operating parameters, generating real-time battery state estimation results and / or fault diagnosis results, and generating a first control command based on the results. The first algorithm is a combined algorithm, at least including a battery state estimation algorithm using a cascaded neural network and Kalman filter, and a battery fault diagnosis algorithm using a twin structure feature encoding model. The diagnostic algorithm involves the following steps: First, the edge intelligent terminal encrypts the operating parameters, state estimation results, and / or fault diagnosis results and uploads them to the cloud server platform via an industrial communication network. Second, the cloud global optimization step involves the cloud server platform calling a second algorithm to process the uploaded data, performing global optimization calculations or digital twin simulation verification, and generating a second optimization strategy. Third, the strategy distribution and fusion step involves the cloud distributing the second optimization strategy to the edge side, and the edge intelligent terminal fusing the first control command and the second optimization strategy to generate a final control signal. Fourth, the execution step involves the actuator receiving the final control signal to complete the charging / discharging regulation or fault protection action of the energy storage battery.

[0007] By adopting the above technical solutions, a collaborative architecture combining real-time processing at the edge and global optimization in the cloud is constructed. At the edge, the combined algorithm enables real-time and accurate estimation of battery status and fault diagnosis. In the cloud, optimization strategies are generated through global optimization and simulation verification. The control commands after the fusion of the two take into account both real-time response and global optimization, effectively improving the overall performance of energy storage battery management and solving the problem that it is difficult to balance real-time performance and optimization under a single management mode.

[0008] Furthermore, the execution process of the battery state estimation algorithm in the edge-side core computing step includes acquiring historical and real-time operating parameters of the energy storage battery to construct time series input data, inputting the time series input data into a pre-trained neural network model to output a priori estimate of the battery state, constructing state equations and observation equations based on the battery equivalent circuit model or electrochemical mechanism model, substituting the priori estimate output by the neural network as the initial state into the filtering algorithm, correcting the error through two sub-steps of state prediction and observation update, and outputting the final battery state estimate. After each charge-discharge cycle, the neural network model is fine-tuned using new operating data.

[0009] By adopting the above technical solution, which combines the advantages of data-driven models and physical models, the neural network model can fully explore the patterns in historical and real-time data, while the filtering algorithm corrects the prior estimates based on physical mechanisms. At the same time, by cyclically fine-tuning and optimizing the model parameters, the accuracy and stability of battery state estimation are improved.

[0010] Furthermore, the execution process of the battery state estimation algorithm also includes a temperature adaptive correction step. The temperature adaptive correction step involves calculating the battery internal resistance correction value at different temperatures in real time based on the correlation between temperature and battery internal resistance during the construction of the state equation and observation equation. The internal resistance correction value is then substituted into the state equation or observation equation of the filtering algorithm to replace the fixed internal resistance parameter.

[0011] By adopting the above technical solution, dynamic compensation for the impact of temperature on battery state estimation is achieved, enabling the state estimation algorithm to adapt to different temperature environments. This avoids the problem of decreased estimation accuracy under low or high temperature environments caused by fixed parameters, and improves the environmental adaptability of the algorithm.

[0012] Furthermore, the execution process of the battery state estimation algorithm also includes an aging adaptive correction step, which involves calculating the battery health state (SOH), embedding the battery health state (SOH) as a correction factor into the state equation of the filtering algorithm, recalculating the battery health state (SOH) after each set number of charge-discharge cycles, and synchronously updating the correction factor in the state equation.

[0013] By adopting the above technical solution, the battery aging state is incorporated into the state estimation model. The estimation results are optimized by dynamically updated health state correction factors, enabling the algorithm to adapt to the aging characteristics of the battery throughout its entire life cycle and ensuring the accuracy of state estimation at different aging stages.

[0014] Furthermore, the execution process of the battery fault diagnosis algorithm in the edge-side core computing step includes collecting battery operating data, extracting multi-dimensional fault features and constructing feature vectors, constructing a twin-structured feature encoding model, the model containing two encoding branches with the same structure, the first branch being trained with normal battery data to reconstruct input features, and the second branch being used to process the feature vector of the battery to be diagnosed, with residual connections introduced in the encoding process to improve the ability to retain micro-fault features, the feature vector being simultaneously input into the two encoding branches to output normal feature encoding and feature encoding to be diagnosed respectively, calculating the residual between the two feature encodings, judging whether a fault has occurred and its severity based on the comparison result of the residual with a preset threshold, and locating the specific fault type by combining the residual contribution analysis.

[0015] By adopting the above technical solution, the twin structure model is used to achieve accurate comparison between normal and diagnostic features. Residual connection enhances the ability to retain micro-fault features. Through residual analysis and contribution positioning, the sensitivity and accuracy of fault diagnosis are improved, and micro-faults can be effectively identified and the fault type can be clearly defined.

[0016] Furthermore, the execution process of the battery fault diagnosis algorithm also includes a feature dynamic weighting step. The feature dynamic weighting step involves assigning initial weights to multi-dimensional fault features based on the battery fault mechanism, adjusting the feature weights in real time according to the battery health status (SOH), constructing a new feature vector using a weighted summation method, and immediately updating the feature weights when the battery charge / discharge rate or temperature fluctuation exceeds a set threshold.

[0017] By adopting the above technical solution, the fault feature weights are dynamically adjusted according to the battery status, so that the core fault features can be effectively identified under different operating conditions. This improves the adaptability of the fault diagnosis algorithm to batteries under different operating conditions and health states, and further ensures the accuracy of diagnosis.

[0018] Furthermore, it also includes an adaptive data preprocessing step, which is located after the edge-side data acquisition step and before the edge-side core computing step, and includes at least one of the following processes: filtering the original acquired data using a low-pass filter that matches the main frequency of battery voltage fluctuations; dynamically adjusting the sampling frequency according to the complexity of the operating conditions; adopting differentiated compression strategies for data features of different importance; and using outlier removal criteria to remove outliers from the filtered data.

[0019] By adopting the above technical solution, adaptive preprocessing is performed on the characteristics of energy storage battery operation data, which effectively reduces noise interference, improves the targeting and efficiency of data acquisition, ensures the quality of input data for core algorithms, and provides data support for the accuracy of subsequent state estimation and fault diagnosis.

[0020] Furthermore, it also includes a model self-evolution step, in which each edge node initializes the first algorithm based on local historical data, updates the local algorithm parameters using a parameter update algorithm with a dynamic adjustment factor, the dynamic adjustment factor being dynamically adjusted according to the battery health state (SOH), the edge node extracts the gradient data of the algorithm update, encrypts it, and uploads it to the cloud, the cloud performs weighted aggregation of the gradients of each node according to the set aggregation weight to generate global algorithm parameters, the cloud distributes the global algorithm parameters to each edge node to update the local algorithm, and completes local iteration and global aggregation according to the set period.

[0021] By adopting the above technical solution, the algorithm model achieves multi-site collaborative evolution. It not only utilizes local data to adapt to the battery characteristics of a single site, but also aggregates and integrates the experience of multiple sites through the cloud. Dynamically adjusting factors ensures the rationality of parameter updates under different health states, thereby improving the model's versatility and adaptability.

[0022] Furthermore, it also includes a model distillation co-optimization step, in which a complex optimization model is trained in the cloud based on aggregated data from multiple sites. The complex optimization model is used as the teacher model and the lightweight model on the edge side is used as the student model. A distillation loss function with battery life loss constraints is constructed to transfer the knowledge of the teacher model to the student model. The distilled student model is then sent to the edge intelligent terminal as a component of the first algorithm. The control commands generated by the student model on the edge side are uploaded to the cloud for simulation verification through a digital twin model. If the life loss exceeds a threshold, the cloud is triggered to re-distill.

[0023] By adopting the above technical solutions, the migration of complex model knowledge to lightweight models on the edge side is realized, which takes into account both the real-time computing needs of the edge side and the optimization performance of the model. Lifetime loss constraints and digital twin verification ensure the security of the control strategy and further improve the reliability of cloud-edge collaborative management.

[0024] Secondly, this application provides a cloud-edge collaborative intelligent management device for energy storage batteries. The device is used to execute the cloud-edge collaborative intelligent management method for energy storage batteries as described in any of the first aspects above, including a multi-dimensional sensor network, an edge intelligent terminal, a cloud server platform, an industrial communication link, and an actuator. The multi-dimensional sensor network is used to collect operating parameters of the energy storage battery. The edge intelligent terminal adopts a heterogeneous architecture and is configured to execute the edge-side core computing steps, strategy fusion steps, and related algorithm steps of the method. The cloud server platform includes a computing cluster, a digital twin module, and a data security module, and is configured to execute the cloud-based global optimization steps and related aggregation and model distillation steps of the method. The industrial communication link is used to realize data interaction between the edge intelligent terminal and the cloud server platform. The actuator is used to execute the charge / discharge regulation or fault protection actions of the method.

[0025] By adopting the above technical solutions, a dedicated hardware carrier is provided for the cloud-edge collaborative intelligent management method of energy storage batteries. The functions and steps of each module are precisely adapted, the heterogeneous architecture of the edge intelligent terminal ensures real-time computing performance, the cloud module achieves global optimization and security, and the industrial communication link ensures the stability of data interaction, thus ensuring that the entire management method can be implemented efficiently and reliably.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A cloud-edge collaborative intelligent management solution for energy storage batteries is provided, which balances real-time performance and global optimization, thereby improving the accuracy and reliability of management.

[0028] The fusion of multiple algorithms and dynamic correction design improves the adaptability of state estimation and fault diagnosis to environmental and aging conditions.

[0029] The model's self-evolution and distillation optimization enhance the versatility and feasibility of the solution, ensuring the safety and economy of energy storage battery operation.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of a cloud-edge collaborative intelligent management system for energy storage batteries is shown in an embodiment of this application.

[0033] Figure 2 A flowchart of a cloud-edge collaborative intelligent management method for energy storage batteries is shown in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a cloud-edge collaborative intelligent management method and device for energy storage batteries. Through the integration of cloud-edge collaborative architecture and multiple algorithms, it can accurately realize battery state estimation and fault diagnosis, take into account real-time response and global optimization, improve management reliability and adaptability, and ensure the safe and economical operation of energy storage batteries.

[0037] In a first aspect, embodiments of this application disclose a cloud-edge collaborative intelligent management system for energy storage batteries.

[0038] Figure 1 A schematic diagram of a cloud-edge collaborative intelligent management system for energy storage batteries is shown in an embodiment of this application.

[0039] Reference Figure 1 The system includes a multi-dimensional sensor network, edge intelligent terminals, a cloud server platform, industrial communication links, and actuators. It is a distributed hardware environment system that supports the implementation of cloud-edge collaborative intelligent management methods for energy storage batteries.

[0040] A multi-dimensional sensor network is used to collect operating parameters of the energy storage battery, specifically covering multi-dimensional physical characteristic data such as voltage, current, temperature, internal resistance, and deformation, providing comprehensive and accurate raw data input for subsequent state estimation and fault diagnosis. The edge intelligent terminal adopts a heterogeneous architecture and is configured to execute the core edge-side computation steps, policy fusion steps, and related algorithm steps of the method described above. It undertakes the core tasks of data preprocessing, local lightweight model inference, and control command generation, and is a key carrier for achieving real-time edge-side response. The cloud server platform includes a computing cluster, a digital twin module, and a data security module, configured to execute the cloud-based global optimization steps and related aggregation and model distillation steps of the method described above. Leveraging large-scale computing power, it supports complex model training, multi-site data aggregation, global policy optimization, and digital twin simulation verification. The industrial communication link is used to realize data interaction between the edge intelligent terminal and the cloud server platform. A low-latency, highly reliable bidirectional communication channel is built using protocols such as 5G and MQTT to ensure efficient synchronization of command issuance and data upload between the edge and the cloud. The actuator is used to perform the charging and discharging regulation or fault protection actions of the method, receive the control signals output by the edge intelligent terminal, and complete specific operations such as equal energy transfer, emergency circuit cut-off, and temperature regulation.

[0041] The components form a closed-loop hardware system through clear connections. Data collected by the multi-dimensional sensor network is transmitted to the edge intelligent terminal. The edge intelligent terminal interacts with the cloud server platform via an industrial communication link to realize data and command interaction. The edge intelligent terminal directly connects to the actuator and outputs control commands. The overall architecture takes into account both the real-time performance of the edge side and the global optimization capabilities of the cloud, providing a solid hardware environment support for the implementation of cloud-edge collaborative intelligent management methods for energy storage batteries.

[0042] Secondly, embodiments of this application disclose a cloud-edge collaborative intelligent management method for energy storage batteries.

[0043] Figure 2 A flowchart of a cloud-edge collaborative intelligent management method for energy storage batteries is shown in an embodiment of this application.

[0044] Reference Figure 2 The method specifically includes the following steps:

[0045] S1: The edge-side data acquisition step involves collecting the operating parameters of the energy storage battery through multi-dimensional sensing devices.

[0046] The operating parameters mentioned here are all acquired by the multi-dimensional sensor network in the system hardware, specifically including the individual unit voltage. ( This corresponds to 16 individual battery cells, and the data is collected by a voltage acquisition module (unit: V) and loop current. (Data acquired by a current Hall sensor, unit: A), distributed temperature ( , The number of temperature acquisition points is measured by a 16-channel fiber optic temperature sensor, in °C. (High-frequency internal resistance is also mentioned.) (Data collected by a high-frequency internal resistance tester, test frequency 1kHz, unit μΩ), shell deformation (Data acquired by MEMS strain gauges, unit με). The acquisition frequency of these parameters is set to 100Hz, and the acquisition period is synchronized with the period of subsequent data preprocessing and core calculation steps, providing comprehensive raw data support for subsequent algorithms.

[0047] The method in this step also includes an adaptive data preprocessing step, which is located after the edge-side data acquisition step and before the edge-side core calculation step. The adaptive data preprocessing step includes at least one of the following processes: filtering the original acquired data using a low-pass filter that matches the main frequency of battery voltage fluctuations; dynamically adjusting the sampling frequency according to the complexity of the operating conditions; adopting differentiated compression strategies for data features of different importance; and using outlier removal criteria to remove outliers from the filtered data.

[0048] A low-pass filter matching the dominant frequency of battery voltage fluctuations is used to filter the raw acquired data. The core purpose is to filter out high-frequency noise interference and retain effective signals related to the battery's operating state. The dominant frequency of battery voltage fluctuations is typically in the range of 0.1~10Hz; therefore, the cutoff frequency of the low-pass filter... Adaptively matching the main frequency range, the main frequency is adjusted in real time by detecting current voltage fluctuations through a preset frequency recognition algorithm. This low-pass filter employs a first-order RC low-pass filtering algorithm, and its transfer function is: in, For complex frequency variables, For filtering resistors (pre-set based on hardware circuit parameters, unit Ω). This is the filter capacitor (pre-set based on hardware circuit parameters, unit: F). The filtered output signal. The calculation formula is:

[0049]

[0050] In the formula, The raw data before filtering (such as real-time values ​​of parameters like voltage and temperature, sourced from the raw output of a multi-dimensional sensor network). For integration variables, The current time (in seconds). The filtered output signal. As input data for subsequent core calculation steps, it can effectively improve the computational accuracy of the core algorithm. The sampling frequency is dynamically adjusted based on the complexity of the operating conditions, which is determined by comprehensively evaluating the rate of change of the currently collected data. An evaluation index for operating condition complexity is defined. The calculation formula is as follows: in, To evaluate the number of sampling points within the window (pre-set to 10 sampling points). The first in the current window The collected values ​​at each sampling point (sourced from the original collected data or the preliminary data after filtering). For the first The collected values ​​at each sampling point. When ( When the high complexity threshold is preset to 0.05, it is considered a high complexity condition, and the sampling frequency is increased to 200Hz; when... ( When the low complexity threshold is preset to 0.01, it is considered a low complexity condition, and the sampling frequency is reduced to 50Hz; when... At the same time, the basic sampling frequency is maintained at 100Hz. The dynamic adjustment of the sampling frequency is achieved through the clock control module of the edge intelligent terminal, ensuring timely capture of critical data when operating conditions change, while reducing hardware resource consumption under non-critical operating conditions.

[0051] Differentiated compression strategies are adopted for data features of different importance. First, the importance weights of each of the 28 features are determined offline using the Analytic Hierarchy Process (AHP). ( The weight range is 0~1, and The weighted data is pre-stored in the SSD of the edge intelligent terminal. Based on the weight, the features are divided into three categories: core features (…). ), General characteristics ( Secondary features () The core features include individual cell voltage. High-frequency internal resistance These characteristics directly affect the accuracy of battery state estimation and require a lossless compression strategy; common characteristics include loop current. Distributed temperature It employs lossy compression with a moderate compression ratio; secondary characteristics include shell deformation. Electromagnetic interference intensity The algorithm employs lossy compression with a high compression ratio. Lossless compression of core features uses the LZ77 algorithm, which essentially uses a sliding window to find duplicate data sequences and replaces them with "distance-length" pairs. Specifically, the sliding window is divided into a history buffer (storing processed data) and a look-ahead buffer (storing data to be processed). Let the length of the history buffer be... bytes, lookahead buffer length is Bytes (all pre-set based on the storage capacity of the edge smart terminal), for the data sequence in the lookahead buffer, search for the longest matching sequence in the history buffer, if the matching length is... Then output (distance) ,length Otherwise, a single character is output. The core feature data compressed by LZ77 is distortion-free, ensuring the accuracy of key parameters in subsequent state estimation. The compressed data is labeled "core-compressed" and temporarily stored on SSD.

[0052] Both general and secondary features are compressed using wavelet transform lossily. The core principle is to transform the data to the wavelet domain through multi-scale decomposition, threshold quantize the wavelet coefficients, and then reconstruct them to achieve data compression. Here, the db4 wavelet basis is selected (pre-set to balance compression effect and computational complexity), and the decomposition scale is... The compression ratio determines the decomposition scale (a higher compression ratio corresponds to a larger decomposition scale). For general features, a compression ratio of 5:1 is set, corresponding to a decomposition scale of [missing information]. For secondary features, a compression ratio of 15:1 is set, corresponding to a decomposition scale. The compression ratio parameter is pre-calibrated through offline testing and stored in the policy configuration file of the edge intelligent terminal. The wavelet decomposition formula is:

[0053]

[0054] In the formula, For the first Low-frequency approximation coefficients (preserving key data characteristics). For the first High-frequency detail coefficients of the layer (corresponding to data noise or minor features); , These are the low-pass and high-pass filter coefficients of the db4 wavelet (pre-embedded in the FPGA of the edge intelligent terminal, utilizing the parallel computing capabilities of the FPGA to improve decomposition efficiency). For the first Wavelet coefficients of the initial layer ( The feature data to be compressed originates from the filtered output data. ); , This is a discrete-time index. The quantization process uses a hard threshold function, and the threshold... ,in The standard deviation of the high-frequency coefficients is calculated from the current data sequence to be compressed. ( (mean of high-frequency coefficients) This is the current data sequence length (dynamically changing with the sampling frequency, derived from the number of sampling points adjusted for operational complexity). When... At that time, Set to 0; when At that time, retain Data compression is achieved by discarding small-amplitude, high-frequency coefficients. The compressed coefficients retain only... and some The data is reconstructed using inverse wavelet transform. The inverse transform formula is: The final reconstructed data is the compressed output. Outlier removal criteria are used to remove outliers from the filtered data; here, the 3σ criterion (Laida criterion) is used. This criterion is based on the normal distribution of the data and is suitable for removing outliers caused by random noise. First, the mean of the data is calculated within a preset statistical window. and standard deviation Statistical window length (Pre-set, balancing statistical reliability and real-time performance), the calculation formula is:

[0055]

[0056] In the formula, This is the number of sampling points within the statistics window (preset to 20 sampling points). The filtered first Each sampling point data (source: low-pass filtered output data). When a certain sampling point data... satisfy When this occurs, the data is determined to be an outlier, and linear interpolation is used to replace the outlier. The interpolation formula is: in, , These are the normal sampling points adjacent to the outlier, respectively. , , These represent the acquisition times (in seconds, provided by the BeiDou timing module of the edge intelligent terminal, with a synchronization accuracy of <1ms) for the corresponding sampling points. The data stream after outlier removal. As the final preprocessed result, it is transmitted to the core computing module of the edge intelligent terminal to provide high-quality input data for subsequent core computing steps on the edge side. It should be noted that if the data needs to be uploaded to the cloud, the various feature data after differential compression will have data identifier headers (including feature type, compression method, data length, etc.) added by the intelligent communication adaptation module, and then uploaded through the 5G communication link according to the strategy of "core feature priority transmission". The identifier header information is generated by the ARM core of the edge intelligent terminal to ensure that the cloud can accurately parse and match the corresponding processing algorithm after receiving it.

[0057] S2: The core computing step on the edge side is that the edge intelligent terminal calls the first algorithm to process the running parameters, generate real-time battery state estimation results and / or fault diagnosis results, and generate a first control command based on the results. The first algorithm is a combined algorithm, which includes at least a battery state estimation algorithm using a cascaded neural network and Kalman filter and a battery fault diagnosis algorithm using a twin structure feature encoding model.

[0058] The operating parameters here are the output data after adaptive data preprocessing in step S1, including 28-dimensional feature data (such as individual unit voltage) after filtering, outlier removal, and differential compression and restoration. Loop current Distributed temperature The edge intelligent terminal adopts a heterogeneous architecture of ARM Cortex-A72 core and FPGA (Xilinx Artix-7). The FPGA is responsible for high-frequency data operations in the first algorithm (such as Kalman filter matrix operations and feature encoding convolution operations), while the ARM core is responsible for algorithm scheduling, result fusion, and control instruction generation. The two interact through the AXI bus to ensure that the computation latency is <3ms. The generation logic of the first control instruction is as follows: if only the state estimation result is output, a charge / discharge regulation instruction is generated based on the deviation between the state estimation value and the preset threshold; if the fault diagnosis result is output, a fault protection instruction (such as cutting off the circuit and starting temperature control) is generated first. The instruction format is pre-fixed in the control logic of the edge intelligent terminal and can directly drive the actuator.

[0059] In this step, the execution process of the battery state estimation algorithm includes acquiring historical and real-time operating parameters of the energy storage battery to construct time series input data, inputting the time series input data into a pre-trained neural network model to output a prior estimate of the battery state, constructing state equations and observation equations based on the battery equivalent circuit model or electrochemical mechanism model, substituting the prior estimate output by the neural network as the initial state into the filtering algorithm, correcting the error through two sub-steps of state prediction and observation update, and outputting the final battery state estimate. After each charge-discharge cycle, the neural network model is fine-tuned using new operating data.

[0060] Specifically, historical operating parameters are derived from preprocessed data of nearly 30 charge-discharge cycles stored in the 128GB SSD of the edge intelligent terminal, while real-time operating parameters are the output data of the current cycle in step S1. The time-series input data is constructed using sliding window sampling, with a window length of... (Pre-set, corresponding to 60 sampling periods, balancing temporal correlation and real-time performance), meaning each time series sample is... ,in Let be the 28-dimensional eigenvector at time t. This represents the transpose operation. The pre-trained neural network model uses a Long Short-Term Memory (LSTM) network, with a model structure of "input layer - 2 LSTM hidden layers - fully connected output layer". The input layer dimension is... (Corresponding to window length and feature dimension), the first LSTM hidden layer has 128 neurons, the second layer has 64, and the output layer dimension is 2 (corresponding to the battery state estimation results: prior estimates of remaining charge (SOC) and state of health (SOH)). The model parameters (weights, biases) are generated offline through a cloud server cluster. The training data is preprocessed data aggregated from multiple sites, and after training, it is distributed to the edge intelligent terminal for permanent storage. The core calculation formula of the LSTM model is:

[0061]

[0062] In the formula, , , Let be the activation values ​​of the input gate, forget gate, and output gate at time t, respectively. It is the sigmoid activation function; Let t represent the candidate cell state at time t. It is the hyperbolic tangent activation function; , The cell states at time t and time t-1 are shown respectively. , These are the hidden layer outputs at time t and time t-1, respectively. This is the model weight matrix (sourced from offline training in the cloud). This is the model bias vector (sourced as above); This represents element-wise multiplication. The output of the LSTM model is the prior estimate of the battery state. ,in Let be the prior estimate of SOC at time t. Let SOH be the prior estimate at time t.

[0063] The filtering algorithm selected is the Extended Kalman Filter (EKF), adapted to the nonlinear characteristics of the battery. Its foundation is the state equations and observation equations constructed from a second-order RC equivalent circuit model. The state vector of the second-order RC equivalent circuit model is... ,in , These are the polarization voltages (in V) of the two parallel RC branches; the observation vector is... ,in This refers to the battery terminal voltage (i.e., the single-cell voltage collected in step S1). (Unit: V). The state equation and the observation equation are as follows:

[0064]

[0065] In the formula, The state prediction value at time k is based on the state at time k-1. This is the state estimate at time k-1; The loop current at time k (sourced from the data after preprocessing in step S1, with positive for charging and negative for discharging, in A). This is the state transition function. For observation functions; The process noise (with a mean of 0 and a variance of ) Gaussian distribution, The process noise covariance matrix is ​​pre-calibrated as a diagonal matrix through offline testing. ); The observed noise (following a mean of 0 and a variance of ) Gaussian distribution, To observe the noise variance, the voltage sensor accuracy is set to... State transition function The specific expression is:

[0066]

[0067] In the formula, The coulomb efficiency is preset to 0.98~1.0 and is adaptively adjusted during charging and discharging. The sampling period (originating from the sampling frequency dynamically adjusted in step S1, i.e.) (unit: s) The rated capacity of the battery (sourced from the battery's factory parameters, pre-stored in the edge smart terminal, unit Ah). These are the resistances (in Ω) of the two RC branches. , These are the time constants (in seconds) of the two RC branches. The capacitance (in F) of the two RC branches is as follows. All parameters are from the battery equivalent circuit model and were pre-calibrated through offline pulse testing. Observation function The specific expression is:

[0068]

[0069] In the formula, Let be the open-circuit voltage (in V) at time k. The mapping relationship between and SOC is constructed using a lookup table built through offline testing and stored in the edge intelligent terminal. Interpolation is then used to obtain the corresponding value for any SOC. ; This represents the battery's internal resistance in ohms (in Ω), which will be dynamically updated through a temperature adaptive correction step. The EKF state prediction step is based on the state estimate at time k-1. Calculate the state prediction value at time k With the prediction error covariance matrix The formula is:

[0070]

[0071] In the formula, State transition function exist Jacobian matrix at the location, It is its transpose matrix; Let be the estimated error covariance matrix at time k-1. The observation update step is to calculate the Kalman gain. State estimates With the estimated error covariance matrix The formula is:

[0072]

[0073] In the formula, For observation function exist Jacobian matrix at the location, Let it be its transpose matrix. This represents the matrix inversion operation; The observed value at time k (i.e., the single-cell voltage after preprocessing in step S1) ); It is the identity matrix. The initial state of EKF. Prior estimates output by the LSTM model Initial error covariance matrix Pre-set as a diagonal matrix The final output battery state estimate is... The core result is (Remaining battery power) is used for generating subsequent control commands.

[0074] Each complete charge-discharge cycle (by detecting the current) The LSTM model is fine-tuned using the preprocessed data (input) and the accurate state estimate (label) output by the EKF, determined by a change in SOC exceeding 80% (either from positive to negative or from negative to positive). The fine-tuning employs a stochastic gradient descent (SGD) optimizer with a learning rate of [missing information]. (Pre-defined), the loss function is the mean squared error (MSE):

[0075]

[0076] In the formula, The number of samples within the loop. This is the exact EKF estimate for the i-th sample. The LSTM prediction value for the i-th sample is used to update the LSTM model parameters by minimizing the loss function, thereby improving the accuracy of subsequent prior estimates.

[0077] The execution process of the battery state estimation algorithm also includes a temperature adaptive correction step. This step involves calculating the battery internal resistance correction value in real time at different temperatures based on the correlation between temperature and battery internal resistance during the construction of the state equation and observation equation. This correction value is then substituted into the state equation or observation equation of the filtering algorithm to replace the fixed internal resistance parameter. Here, the temperature refers to the distributed temperature collected in step S1. average ( (Number of temperature sampling points), the relationship between temperature and internal resistance was constructed through offline testing, specifically using a polynomial fitting model:

[0078]

[0079] In the formula, For temperature The corrected value for the internal resistance of the ohm (in Ω); The fitting coefficients are obtained in advance through offline testing, with a test temperature range of -20℃ to 60℃, and stored in the edge intelligent terminal. The average current temperature (Unit: °C). [The following is a list of parameters:] [To be filled in] Replace observation function Fixed internal resistance That is, the corrected observation function is: This correction adapts the observation equations to the current temperature environment, avoiding observation errors caused by temperature changes and improving the accuracy of state estimation. Simultaneously, the RC branch resistance... The same form of temperature polynomial model is also used for correction. The corrected parameters are substituted into the state transition function to ensure the temperature adaptability of the state equation.

[0080] The execution process of the battery state estimation algorithm also includes an aging adaptive correction step, which involves calculating the battery health state (SOH), embedding the battery health state (SOH) as a correction factor into the state equation of the filtering algorithm, and recalculating the battery health state (SOH) and updating the correction factor in the state equation synchronously after each set number of charge-discharge cycles.

[0081] The State of Health (SOH) of a battery is defined as the ratio of its current usable capacity to its rated capacity, and is calculated using the following formula:

[0082]

[0083] In the formula, The current usable capacity of the battery (in Ah) is calculated through charge-discharge cycle testing: after constant current charging to the upper voltage limit (pre-set to the battery's rated voltage, such as 3.65V), followed by 1 hour of rest and constant current discharging to the lower voltage limit (pre-set to 2.5V), the integral of the charge during the discharge process is the current usable capacity. ,in Discharge time (in seconds). Discharge current (unit: A); The rated capacity (source: battery manufacturer parameters). With SOH embedded as a correction factor into the capacity term of the state transition function, the corrected SOC calculation formula is: The State of Health (SOH) is recalculated every 5 charge-discharge cycles (the number of cycles is counted by the edge intelligent terminal), and the state equation is updated synchronously. This feature adapts state estimation to battery aging characteristics, ensuring estimation accuracy throughout the entire lifespan.

[0084] The execution process of the battery fault diagnosis algorithm includes collecting battery operating data, extracting multi-dimensional fault features and constructing feature vectors, constructing a twin structure feature encoding model, the model containing two encoding branches with the same structure, the first branch being trained with normal battery data to reconstruct input features, and the second branch being used to process the feature vector of the battery to be diagnosed, with residual connections introduced in the encoding process to improve the ability to retain micro-fault features, the feature vector being simultaneously input into the two encoding branches to output normal feature encoding and feature encoding to be diagnosed respectively, calculating the residual between the two feature encodings, judging whether a fault has occurred and its severity based on the comparison result of the residual with a preset threshold, and locating the specific fault type by combining the residual contribution analysis;

[0085] The battery operation data here is the 28-dimensional feature data after preprocessing in step S1. The multi-dimensional fault features include time-domain and frequency-domain features: time-domain features include the mean, variance, peak value, kurtosis, and waveform factor of voltage, current, and temperature (5 categories, 28 dimensions per category, totaling 140 dimensions); frequency-domain features include the peak value, centroid, and harmonic amplitude of the voltage and current signals obtained through Fast Fourier Transform (FFT) (3 categories, 28 dimensions per category, totaling 84 dimensions). The final extracted fault feature dimension is 224. The feature vector is constructed as follows: ,in It is a 140-dimensional time-domain feature vector. The 84-dimensional frequency domain feature vector is constructed and then normalized to the [0,1] interval using the min-max normalization method. The normalization formula is as follows:

[0086]

[0087] In the formula, The normalized feature vectors, , These are the minimum and maximum values ​​for each feature dimension (obtained in advance through statistical analysis of normal battery data and stored in the edge intelligent terminal).

[0088] The Siamese structural feature encoding model contains two encoding branches with identical structures (branch A and branch B). Each encoding branch consists of 4 layers of convolutional neural networks (CNN) and 2 layers of fully connected layers. The core structure is "Conv1d(64, 3) -BatchNorm - ReLU - MaxPool1d - Conv1d(128, 3) - BatchNorm - ReLU - MaxPool1d- Conv1d(256, 3) - BatchNorm - ReLU - MaxPool1d - Conv1d(512, 3) - BatchNorm- ReLU - Flatten - Linear(1024) - Linear(256)", where Conv1d is a one-dimensional convolutional layer (the number of output channels is before the kernel size), BatchNorm is a batch normalization layer, ReLU is an activation function, MaxPool1d is a one-dimensional max pooling layer, Flatten is a flattening layer, and Linear is a fully connected layer. Branch A uses the normalized feature vector of a normal battery. Offline training, the training objective is to reconstruct the input features, and the loss function is the reconstruction error:

[0089]

[0090] In the formula, This is the normal number of training samples. For the reconstructed feature vector of branch A, It is an L2 norm. The network parameters of branch B are shared with those of branch A. Residual connections are introduced during the encoding process. The residual connection formula is:

[0091]

[0092] In the formula, The input features are those of the l-th layer. The output of the convolution operation at layer l is used to directly pass low-level features through residual connections, improving the ability to preserve micro-fault features. This is achieved by normalizing the feature vector to be diagnosed. Two encoding branches are input simultaneously; branch A outputs the normal feature code. Branch B outputs the feature code to be diagnosed. The residuals between the two are calculated as the Euclidean distance:

[0093]

[0094] In the formula, , These are the values ​​of the d-th dimension of the two feature codes, respectively. Preset fault determination threshold. (Pre-calibrated through offline testing with normal and faulty samples, and stored in the edge intelligent terminal), when When, it is determined to be fault-free; when When a fault is detected, and the severity of the fault is related to... Positive correlation ( The larger the value, the more severe the fault. Residual contribution analysis calculates the contribution of each feature dimension to the residual. The contribution value is realized, and the contribution value is calculated using the following formula:

[0095]

[0096] In the formula, For the contribution of the d-th feature encoding, the top 5 feature dimensions with the highest contribution are selected and back-mapped to the original 28-dimensional operating parameters (e.g., if mapped to voltage features, it is determined to be a voltage-related fault; if mapped to temperature features, it is determined to be a temperature-related fault), thus realizing fault type localization.

[0097] The execution process of the battery fault diagnosis algorithm also includes a feature dynamic weighting step. The feature dynamic weighting step is to assign initial weights to multi-dimensional fault features based on the battery fault mechanism, adjust the feature weights in real time according to the battery health status (SOH), construct a new feature vector using a weighted summation method, and immediately update the feature weights when the battery charge / discharge rate or temperature fluctuation exceeds a set threshold.

[0098] Based on battery fault mechanisms (such as the strong correlation between overcharge faults and voltage characteristics, and the strong correlation between overheat faults and temperature characteristics), initial weights are assigned to the 224-dimensional fault features. ( The weight range is 0~1, and The weights are pre-determined through expert experience and offline testing and stored in the edge intelligent terminal. The weights are adjusted in real-time based on the State of Health (SOH), using the following formula:

[0099]

[0100] In the formula, The feature weights are dynamically adjusted. The weighting adjustment coefficient is set to 0.5, and SOH represents the current health status calculated by the aging adaptive correction step. When SOH decreases (accelerated battery aging), the weights of aging-related features (such as internal resistance and capacity decay-related features) are increased to improve the diagnostic sensitivity of aging-related faults. The formula for constructing a new feature vector by weighted summation is:

[0101]

[0102] In the formula, The weighted feature vectors, The d-th dimension is the normalized feature. Charge / discharge rate. (Unit: C), Temperature fluctuation (Unit: °C), Preset magnification threshold Temperature fluctuation threshold (All preset), when or Immediately recalculate the dynamic weights. And update the feature vector to ensure that fault diagnosis is adapted to extreme working conditions.

[0103] It also includes a model self-evolution step, in which each edge node initializes the first algorithm based on local historical data, updates the local algorithm parameters using a parameter update algorithm with a dynamic adjustment factor, the dynamic adjustment factor being dynamically adjusted according to the battery health state (SOH), the edge node extracts the gradient data of the algorithm update, encrypts it, and uploads it to the cloud, the cloud performs weighted aggregation of the gradients of each node according to the set aggregation weight to generate global algorithm parameters, the cloud distributes the global algorithm parameters to each edge node to update the local algorithm, and completes local iteration and global aggregation according to the set period.

[0104] Each edge node's local historical data consists of preprocessed data from nearly 10 charge-discharge cycles stored within itself, along with corresponding state estimates and fault diagnosis results. The initialization parameters for the first algorithm are the initial model parameters (consistent with the parameters of the pre-trained LSTM and Siamese coding models) distributed from the cloud. The parameter update algorithm with dynamic adjustment factors uses the stochastic gradient descent (SGD) algorithm with momentum, and the parameter update formula is:

[0105]

[0106] In the formula, These are the local algorithm parameters (including the weights and biases of the LSTM model and the Siamese coding model) at time t+1 and time t, respectively. The momentum coefficient (preset to 0.9); The learning rate (dynamic adjustment factor) is dynamically adjusted according to SOH, and the adjustment formula is as follows: ,in The base learning rate is preset, and the state of health (SOH) is the current health status. The learning rate is reduced when the battery ages (SOH decreases) to avoid parameter oscillation. The parameter gradient at time t is obtained by differentiating the loss function of the local training data (MSE loss of LSTM, reconstruction loss of Siamese model) with respect to the parameters. Gradient data is extracted from edge nodes. Then, the data is encrypted using the AES-256 encryption algorithm (the key is pre-distributed to each edge node by the cloud) and uploaded to the cloud server platform via the 5G communication link (SA mode). The cloud then sets the aggregation weight based on the number of batteries and runtime of each edge node. ( The more batteries and the longer the runtime, the greater the weight (with pre-defined weight allocation rules), and the weighted aggregation generates the global gradient. The formula is:

[0107]

[0108] In the formula, The total number of edge nodes participating in self-evolution. The encrypted gradient uploaded to the i-th edge node (obtained after decryption in the cloud). Global algorithm parameters. It is obtained by subtracting the global gradient update from the initial global parameters:

[0109]

[0110] In the formula, These are the initial global parameters. The global learning rate is preset. The self-evolution cycle is set to 7 days (timed by a cloud clock). Within each cycle, a closed-loop process of "local parameter update - gradient upload - cloud aggregation - parameter distribution - local update" is completed once, realizing the continuous self-evolution of the first algorithm and improving its adaptability to different scenarios and batteries with different aging levels.

[0111] S3: The data upload step involves the edge intelligent terminal encrypting the operating parameters, state estimation results, and / or fault diagnosis results and then uploading them to the cloud server platform via the industrial communication network.

[0112] The operating parameters here are the final output data after adaptive data preprocessing in step S1, specifically including 28-dimensional feature data after filtering, outlier removal, and differential compression (if uploaded), covering individual unit voltages. ( ), loop current Distributed average temperature High-frequency internal resistance Core and general feature data are collected, while secondary feature data is selectively uploaded only when global analysis is required in the cloud. The state estimation result is the final estimate output by the battery state estimation algorithm in step S2, including the remaining power. Polarization voltage and health status ,in and The core data is uploaded, with polarization voltage uploaded synchronously as auxiliary data. The fault diagnosis result is the output of the battery fault diagnosis algorithm in step S2, including the fault determination result (0 indicates no fault, 1 indicates fault) and the quantification value of the fault severity. (i.e., feature coding residuals) and fault type location results (such as the coding values ​​corresponding to text labels like voltage anomaly, temperature anomaly, etc.). If the fault is determined to be absent, only the fault determination result needs to be uploaded; if the fault is determined to be present, all fault diagnosis-related data must be uploaded completely. All data to be uploaded should be organized in a structured data format to construct data frames. The Header is the data frame header, which includes the edge node ID (pre-assigned by the cloud to uniquely identify each edge terminal) and the data collection timestamp. (Provided by the BeiDou timing module of the edge intelligent terminal, with millisecond accuracy), data type identifier (01 represents operating parameters, 10 represents state estimation results, and 11 represents fault diagnosis results) and data length. (Unit: bytes) The core data body stores the specific data content to be uploaded, arranged in little-endian byte order; the tail is the end of the data frame, storing the data checksum, used for integrity verification after the cloud receives the data.

[0113] Data encryption employs the SM4 block cipher algorithm (a symmetric encryption algorithm), a national standard. This algorithm is suitable for lightweight encryption needs in industrial scenarios and can be efficiently implemented on the ARM core of edge intelligent terminals. The encryption key is pre-distributed to each edge node by the cloud server platform through the initial security handshake process. The key length is 128 bits and it is stored in the security encryption chip of the edge intelligent terminal (such as the encryption engine of STM32L476) to prevent key leakage. The core of the SM4 algorithm is block encryption, which encrypts the data to be encrypted in 128-bit (16-byte) blocks. If the length of the data to be encrypted is less than 128 bits, it is padded using PKCS#7. The padding rule is as follows: Let the number of padding bytes be... ( ), then fill The number is The formula for filling the bytes is: in The original data segment to be encrypted. This indicates byte concatenation operations. ( (Indicates the data length in bytes), padding ensures that each encrypted block is 128 bits long.

[0114] The round function of the SM4 encryption algorithm is the core computational unit, and its mathematical expression is:

[0115]

[0116] In the formula, The four 32-bit input data for the current encryption round (obtained by splitting 128-bit block data); For the first wheel( The master key consists of 32 round keys, which are generated from the 128-bit master key using a key expansion algorithm. It is a nonlinear transformation function, consisting of two parts: S-box substitution and linear transformation. This represents the XOR operation. The nonlinear transformation function... The specific expression is:

[0117]

[0118] In the formula, 32-bit input data; For the S-box replacement operation, the 32-bit data is split into four 8-bit bytes. Each byte is replaced using a pre-defined S-box table, which is a 16×16 byte mapping table pre-embedded in the encryption chip of the edge smart terminal. The replacement formula is as follows: ,in It is an 8-bit binary number, with the first 4 bits being the row index and the last 4 bits being the column index;

[0119] For linear transformation operations, the mathematical expression is:

[0120]

[0121] In the formula, The 32-bit data after replacing the S-box; This indicates a left shift operation. After the left shift, any bits exceeding 32 bits are discarded, and the least significant bits are padded with 0s. The SM4 algorithm performs 32 rounds of function operations, and the output of the final round... Perform a reverse transformation to obtain 128-bit encrypted ciphertext. This completes the encryption of single-group data. For multi-group data to be encrypted, the ECB (Electronic Codebook) mode is used for group encryption to ensure the parallelism of the encryption process and adapt to the real-time processing needs of edge intelligent terminals.

[0122] Round keys are generated using a key expansion algorithm, based on a 128-bit master key. (each) (32-bit), first with system parameters The initial key is obtained by performing an XOR operation. The formula is: In the formula, These are preset system parameters, with specific values... , , , (All numbers are hexadecimal and pre-embedded in the encryption algorithm program). Then, based on the initial key... Iteratively generate 32 round keys The iterative formula is:

[0123]

[0124] In the formula, ; These are 32 preset wheel constants, each a 32-bit hexadecimal number (e.g., ...). , (etc.), pre-stored in an encrypted chip; The nonlinear transformation function for key expansion, and the encryption round function. The structure remains the same, except that the left shift bits of the linear transformation part are adjusted to 13 or 23 bits. The linear transformation formula is as follows: .

[0125] The 32 round keys generated through the above key expansion process correspond one-to-one with the master key, ensuring consistency between the encryption and decryption processes. The same round key is used for decryption as for encryption, with only the order of the round function operations reversed.

[0126] The industrial communication network adopts a dual-mode redundancy architecture with "5G SA (Standalone) as the primary mode and Industrial Ethernet as the secondary mode." Edge intelligent terminals achieve dual-mode communication adaptation through built-in 5G modules (supporting NR bands) and the Industrial Ethernet interface. The selection of the communication link is dynamically switched based on real-time requirements. When data upload latency requirements are <50ms (e.g., fault diagnosis results, SOC mutation data), the 5G SA link is prioritized, with a downlink rate ≥100Mbps and an uplink rate ≥50Mbps, meeting the requirements for high-frequency, small-batch real-time data uploads. When uploading batches of historical data (e.g., complete operating parameters for a charge-discharge cycle), the Industrial Ethernet link can be selected, transmitting data via TCP / IP protocol to ensure data transmission stability and high-capacity adaptability. The communication parameters of the 5G SA link are dynamically configured through the communication management module of the edge intelligent terminal, with the modulation and demodulation method using QPSK or 16QAM, the specific selection based on channel quality parameters. (Channel quality indicator, value range 0~15) is determined when When using 16QAM modulation (higher spectral efficiency), QPSK modulation is used (with stronger anti-interference capability), and channel quality parameters are... The 5G module detects and feeds back the data to the communication management module in real time.

[0127] To ensure the integrity and reliability of uploaded data, an integrity check is required after data encryption and before uploading. The CRC32 cyclic redundancy check algorithm is used to generate a checksum, which is embedded in the Tail field of the data frame. The core of the CRC32 algorithm is to treat the data to be checked as a binary polynomial. , and the preset generator polynomial The remainder obtained by performing modulo-2 division is the check digit, and its mathematical expression is: In the formula, The data to be verified (encrypted ciphertext data) (The polynomial corresponding to the binary sequence concatenated with the data frame header). Indicates will Shift left by 32 bits; The quotient polynomial of modulo 2 division; The polynomial for generating the CRC32 standard is as follows: (Pre-embedded in the verification algorithm program); It is a remainder polynomial, corresponding to a 32-bit checksum. After calculating the checksum, the edge intelligent terminal uploads it to the cloud along with the encrypted data frame. Upon receiving the data, the cloud uses the same CRC32 algorithm to verify the received encrypted data and header. If the calculated checksum matches the checksum in the Tail field, the data is considered complete and proceeds with subsequent decryption and processing. If they do not match, a retransmission request is sent to the edge intelligent terminal. Upon receiving the retransmission request, the edge intelligent terminal re-encrypts the corresponding data frame and uploads it. The preset number of retransmissions is 3. If the verification still fails after 3 retransmissions, a fault log is recorded and reported to the cloud operation and maintenance module.

[0128] Furthermore, data uploading employs a combined strategy of "on-demand uploading + periodic uploading." State estimation results are uploaded synchronously according to the sampling period of step S1 (i.e., periodic uploading under dynamic sampling frequency), while operating parameters are uploaded at a fixed 1-minute interval (if the sampling frequency is less than 1 minute / time, they are uploaded synchronously with the sampling period). Fault diagnosis results are immediately triggered for uploading upon completion of the diagnosis (on-demand uploading), ensuring that the cloud can obtain critical data in real time while avoiding excessive occupancy of the communication link. Upload data flow control uses a token bucket algorithm with a preset token generation rate. bytes / second, token bucket capacity bytes, when the amount of data to be uploaded At that time, you can directly obtain the token and upload it; when At this time, it is necessary to wait for tokens to be generated until enough tokens are accumulated to avoid communication congestion caused by a large amount of data being uploaded in a short period of time. The flow control formula of the token bucket algorithm is: In the formula, The number of available tokens in the current token bucket (in bytes); The time interval between two token generation (default is 0.1 seconds); the number of available tokens is dynamically updated through this formula to achieve smooth control of upload traffic.

[0129] S4: The cloud-based global optimization step involves the cloud server platform calling the second algorithm to process the uploaded data, performing global optimization calculations or digital twin simulation verification, and generating a second optimization strategy.

[0130] The uploaded data here is the complete data encrypted and uploaded in step S3, and then decrypted and verified by the cloud. It includes the operating parameters, state estimation results, and fault diagnosis results of the energy storage batteries at each edge node. The cloud server platform adopts a distributed cluster architecture (composed of data storage nodes, computing nodes, and simulation nodes). The data governance module cleans and merges the multi-source uploaded data to form a global dataset. Data cleaning mainly removes abnormal data caused by communication interference during the upload process. It uses the statistically based 3σ criterion (consistent with the outlier removal logic in step S1) to calculate the mean for each type of data (such as the SOC sequence and temperature sequence of each edge node) within a preset time window (the window length is preset to 1 hour, and the corresponding data volume dynamically changes with the sampling frequency of the edge nodes). and standard deviation The formula is:

[0131]

[0132] In the formula, This represents the number of global data samples within the time window. For the i-th global data sample (sourced from similar data uploaded by each edge node); when If data is deemed abnormal, it is replaced with the average of data from adjacent edge nodes during the same period to ensure the reliability of the global dataset. Data fusion employs a weighted average fusion strategy, fusing similar state estimation results (such as SOC) uploaded by different edge nodes at the same time. The fusion weights are... Dynamic allocation of historical estimation accuracy based on edge nodes; historical estimation accuracy Defined as the reciprocal of the average deviation between the historical estimated value of the edge node and the backtracked verification value in the cloud, the fusion formula is:

[0133]

[0134] In the formula, Let SOC be the globally fused value at time t. Let be the SOC estimate uploaded by the i-th edge node at time t. ( That is, calculated based on the backtracking verification results of the last 10 time points. The cloud-based backtracking verification value at time tk (obtained from digital twin model simulation) improves the accuracy and completeness of global data through data fusion.

[0135] The second algorithm comprises a global optimization algorithm and a digital twin simulation algorithm. The global optimization algorithm employs an improved particle swarm optimization (PSO) algorithm, which addresses the collaborative operation requirements of multi-edge node energy storage battery clusters. It constructs a multi-objective optimization model with the dual optimization objectives of "minimizing global energy consumption and minimizing battery life loss." The optimization variables are the charging and discharging power of each edge node energy storage battery. ( (t is the scheduling time, in hours). The constraints include battery charging and discharging power constraints, SOC safety constraints, and grid interaction power constraints. The specific constraint equations are as follows:

[0136]

[0137] In the formula, These are the minimum and maximum charge and discharge power of a single edge node battery, respectively (sourced from battery factory parameters and pre-stored in the cloud). The safe operating range of battery SOC (pre-set based on battery life characteristics); These represent the minimum and maximum power requirements for the interaction between the energy storage cluster and the power grid (pre-distributed to the cloud by the power grid dispatch center). The objective function of the dual optimization objectives is transformed into a single objective function using a weighted summation method, as shown in the formula:

[0138]

[0139] In the formula, Let be the global energy consumption cost at time t. Let t be the global battery lifetime loss cost. The target weight (dynamically adjusted based on grid peak and off-peak electricity prices and battery replacement costs, peak time) Valley Time Energy consumption costs , The grid electricity price at time t (distributed in real-time from the grid dispatch center to the cloud); lifetime loss cost. The battery cycle life loss is quantified based on the rainflow counting method, and the formula is as follows: , Let be the number of cycle life losses of the battery at the i-th edge node at time t. This is the total cycle life of the battery (source: battery manufacturing parameters). Replacement cost for a single battery (preset).

[0140] The improved PSO algorithm enhances optimization performance by introducing adaptive inertia weights. The formula is dynamically adjusted according to the number of iterations:

[0141]

[0142] In the formula, The maximum and minimum values ​​of the inertia weight (preset).

[0143]

[0144] Let t be the maximum number of iterations (default 100), and t be the current iteration number. The formulas for updating the particle's velocity and position are:

[0145]

[0146] In the formula, Let be the velocity of the i-th particle in the j-th dimension (corresponding to the charging and discharging power of the j-th edge node) in the t-th iteration. For the corresponding position; Learning factors (pre-set). A random number within the interval [0,1]; This represents the historical best position of the i-th particle. The global optimal position for all particles is represented. The optimal charging and discharging power for each edge node is obtained using an improved PSO algorithm. This constitutes the core content of the global optimization strategy.

[0147] Digital twin simulation verification is based on a digital twin model of an energy storage battery cluster. This model employs a hybrid modeling approach combining mechanism and data-driven methods, integrating battery electrochemical mechanism models with deep learning models to accurately reproduce the operating states of batteries at each edge node in the physical world. The mechanism part is based on a pseudo-two-dimensional (P2D) electrochemical model, describing the ion transport and charge transfer processes inside the battery. The core equations include the solid-phase diffusion equation, the electrolyte diffusion equation, and the Butler-Volmer charge transfer equation. The solid-phase diffusion equation is as follows:

[0148]

[0149] In the formula, Let be the solid-phase lithium-ion concentration at the battery particle radius r at time t. Here, r is the solid-phase lithium-ion diffusion coefficient (derived from battery material parameters, pre-stored in the cloud), and r is the particle radius (range 0~). , (The radius of the battery active material particles is preset). The data-driven part uses a CNN-LSTM hybrid model to compensate for the output error of the P2D model. The input is the operating parameters (current, temperature) uploaded by the edge nodes, and the output is the error compensation value. The final digital twin model output is obtained through data fusion. ,in This is the state output (SOC, terminal voltage, etc.) of the P2D model. This is the error compensation value for the data-driven model.

[0150] The core of digital twin simulation verification is to compare the state estimation results uploaded by edge nodes with the simulation results of the digital twin model and calculate the deviation. ,like ( If the preset deviation thresholds are set (SOC deviation threshold is 0.02V, voltage deviation threshold is 0.05V), then the edge node state estimation result is considered reliable, and the global optimization strategy is effective; if If this happens, the global optimization algorithm will be triggered to iterate and solve the problem again. The final generated second optimization strategy includes the optimal charging and discharging power curves of each edge node, the SOC target trajectory, and temperature control threshold adjustment suggestions, which are sent to each edge smart terminal through the 5G communication link as the constraint basis for the first algorithm on the edge side to generate control commands.

[0151] The method in this step also includes a model distillation co-optimization step. The model distillation co-optimization step involves training a complex optimization model in the cloud based on aggregated data from multiple sites, using the complex optimization model as the teacher model and the lightweight model on the edge as the student model, constructing a distillation loss function with battery life loss constraints to transfer knowledge from the teacher model to the student model, and sending the distilled student model to the edge intelligent terminal as a component of the first algorithm. The edge side uploads the control commands generated by the student model to the cloud for simulation verification through a digital twin model. If the life loss exceeds a threshold, the cloud is triggered to re-distill.

[0152] The multi-site aggregated data consists of historical operating parameters and corresponding global optimization results uploaded by each edge node (obtained by the improved PSO algorithm). The data volume needs to cover at least 100 charge-discharge cycles to ensure sufficient training of the teacher model. The complex optimization model (teacher model) adopts a deep neural network (DNN) with a structure of "input layer - 4 fully connected hidden layers - output layer". The input layer has a dimension of 12 (including 12 key features such as current SOC, current, temperature, and grid electricity price). The number of neurons in each hidden layer is 256, 128, 64, and 32, respectively. The output layer has a dimension of 1 (corresponding to the optimal charge-discharge power command). The training optimizer uses Adam with a preset learning rate of 1e-3 and the loss function is mean squared error. ,in The number of training samples. Let be the globally optimal charge / discharge power for the i-th sample. This represents the predicted power of the teacher model.

[0153] The lightweight edge-side model (student model) employs a shallow neural network with an input layer, two fully connected hidden layers, and an output layer. The input layer dimension is the same as the teacher model (12-dimensional), the hidden layer neurons are 64 and 32 respectively, and the output layer dimension is 1, ensuring efficient operation on the ARM core of edge smart terminals (inference latency <2ms). A distillation loss function incorporates battery life loss constraints, balancing prediction accuracy and lifespan protection. The formula is:

[0154]

[0155] In the formula, The soft loss is the cross-entropy between the softened probability output by the teacher model and the output probability by the student model. The hard loss is the mean square error between the student model's predicted value and the actual optimal power. For lifetime loss constraints, These are the weighting coefficients (pre-calibrated through offline testing). The calculation of soft loss requires temperature softening of the teacher model output; the formula is:

[0156]

[0157] In the formula, The softening temperature is preset (the higher the temperature, the stronger the softening). These are the softening probabilities output by the teacher model and the student model, respectively. The number of output categories (here, 1, representing continuous power, approximated using continuous probability). Lifetime loss constraint. The formula is calculated based on the battery cycle life loss model. ,in Predict the number of battery cycle life losses corresponding to the power for the student model (calculated by rainflow counting method). This refers to the total cycle life of the battery.

[0158] The model distillation training process is completed in a cloud GPU cluster (such as NVIDIA Tesla V100). After training, the student model is quantized and compressed (using INT8 quantization to convert the model weights from 32-bit floating-point numbers to 8-bit integers) to reduce model storage capacity and inference power consumption. The quantization formula is as follows:

[0159]

[0160] In the formula, The weights are quantized 8-bit integers. The weights are the original 32-bit floating-point numbers. The minimum and maximum values ​​of the weights are obtained from the statistical analysis of the trained model weights, and `round` is the rounding operation. The quantized student model is sent to each edge intelligent terminal through an encrypted communication link, replacing the power prediction module in the original first algorithm, and participating in the generation of control commands as a component of the first algorithm.

[0161] The edge side will generate control commands (charge and discharge power commands) from the student model. The operating parameters are uploaded synchronously to the cloud. The cloud uses a digital twin model to simulate the battery's operating state under the control command and calculate the corresponding lifespan loss. (Obtained through simulation) and the preset lifetime loss threshold (The comparison is based on 0.1% of the total battery cycle life per day, as preset; the judgment condition is...) If the decision criteria are met, the student model can continue to be used; otherwise... This triggers cloud-based re-distillation, retraining the teacher model and optimizing the weight coefficients of the distillation loss function. A new student model is generated and sent to the edge side to ensure that the edge side control commands meet the requirements of real-time performance while also taking into account the battery life protection requirements.

[0162] S5: The strategy distribution and fusion steps are as follows: the cloud distributes the second optimization strategy to the edge side, and the edge intelligent terminal integrates the first control command and the second optimization strategy to generate the final control signal.

[0163] The second optimization strategy here is the final strategy after global optimization calculation or digital twin simulation verification in step S4, which includes the optimal charging and discharging power trajectory corresponding to each edge node. SOC target range Temperature control threshold adjustment value The policy data, including core content such as fault collaborative handling suggestions, is encapsulated in structured JSON format and further enhanced with a cloud-based digital signature to ensure integrity and immutability. The digital signature employs the SM2 asymmetric encryption algorithm, with the cloud acting as the signatory using its private key. (Generated and securely stored by the cloud key management center) The hash value of the policy data is signed. The hash value is calculated using the SM3 hash algorithm, and the formula is:

[0164]

[0165] In the formula, The original data for the second optimization strategy. This is a 256-bit hash value of the strategy data. The digital signature is generated in the cloud; the policy data and the digital signature are sent together to the corresponding edge smart terminal via a 5G SA encrypted communication link (consistent with the upload link in step S3), and the cloud public key is pre-stored on the edge side. (Delivered during the initial configuration phase from the cloud), upon receipt, the signature's validity is verified using the public key. The verification formula is:

[0166]

[0167] In the formula, The hash value recalculated by the edge side for the received policy data, if If the verification passes, the policy data is received and parsed; if the verification fails, the policy is rejected and a policy retransmission request is sent to the cloud. The retransmission mechanism is consistent with the data retransmission logic of step S3 (maximum of 3 retransmissions) to ensure the security and reliability of policy distribution.

[0168] The first control command is a local control command generated by the edge intelligent terminal through the first algorithm in step S2. Its core content includes real-time charging and discharging power commands. The generation logic for fault protection commands (if a fault exists) and basic temperature control commands has been defined in step S2, and here they serve as the local decision-making basis for the fusion input. The core of the edge intelligent terminal's fusion of the first control command and the second optimization strategy is the dynamic weighted fusion of the charging and discharging power commands. The fusion weight is dynamically allocated based on "local real-time priority + cloud-based global optimal priority," and the fusion weight coefficient is defined. (Local instruction weights) and (Cloud-based strategy weights) satisfy The weighting formula is:

[0169]

[0170] In the formula, This represents the deviation between the current battery SOC and the median SOC of the cloud strategy. (Source: SOC target range in the second optimization strategy). Set a preset SOC deviation threshold (pre-set based on battery operation safety). This is the weight adjustment coefficient (pre-calibrated through offline testing to adjust the sensitivity of the weights to changes in bias). When At that time, local commands and cloud policies have equal weight; when At that time, increase the weight of local instructions to prioritize real-time battery safety (such as preventing the SOC from exceeding the safe range); when At the same time, increase the weight of cloud-based strategies to prioritize achieving global optimization goals.

[0171] The formula for calculating the fusion of charge and discharge power commands is as follows:

[0172]

[0173] In the formula, This is the final charge / discharge power command after fusion. The local charging and discharging power in the first control command. This represents the optimal charge / discharge power for the corresponding edge node in the second optimization strategy. After fusion, the power command needs to be constrained and verified to ensure it complies with battery hardware safety limits. The verification constraints are consistent with the charge / discharge power constraints in step S4.

[0174]

[0175] In the formula, The minimum and maximum charge and discharge power of a single edge node battery (sourced from battery factory parameters, consistent with step S4) are used to ensure that the power command after fusion is within the hardware safety range through constraint verification.

[0176] For the fusion of temperature control commands and fault handling commands, the principle of "fault priority and hierarchical coverage" is adopted: if the first control command includes a fault protection command (such as a serious fault diagnosed in step S2), then the fault protection command (such as cutting off the charging and discharging circuit or starting emergency cooling) is directly used. At this time, the temperature control threshold adjustment value in the cloud-based second optimization strategy is adjusted accordingly. This serves only as an auxiliary adjustment basis after fault recovery; if there is no fault, the basic temperature control command and the cloud-based temperature control threshold adjustment value are combined to generate the final temperature control command. The formula is:

[0177]

[0178] In the formula, The basic temperature control threshold (preset) in the first control command. The final temperature control threshold after fusion. This is the current average battery temperature (sourced from the average distributed temperature after preprocessing in step S1). The temperature control lag threshold is preset to avoid frequent start-stop of the temperature control equipment; finally, the temperature control command and the charge / discharge power command after constraint verification together constitute the core control signal.

[0179] The final control signal needs to be converted into a standardized electrical signal that the actuator can recognize. A PWM (Pulse Width Modulation) signal is used to drive the charge / discharge controller, and a digital level signal is used to control the temperature control equipment. The signal conversion formula is as follows:

[0180]

[0181] In the formula, This refers to the duty cycle of the PWM signal, ranging from 0 to 100%, corresponding to a charging / discharging power of... arrive The linear adjustment; after receiving the PWM signal, the charge and discharge controller realizes the corresponding charge and discharge power output through the power conversion circuit, and the temperature control device executes the corresponding heat dissipation or heating operation after receiving the digital level signal.

[0182] After generating the final control signal, the edge intelligent terminal needs to record the fusion process data (including fusion weight, input command values, and final output signal values) and upload it to the cloud synchronously according to the data upload cycle of step S3. This serves as the basis for the cloud to evaluate the effectiveness of the fusion strategy and optimize the second algorithm, forming a closed-loop control logic of "cloud strategy issuance - edge fusion execution - data feedback evaluation" to ensure the continuous optimization and iteration of the entire technical solution.

[0183] S6: The execution step is for the actuator to receive the final control signal and complete the charging and discharging regulation or fault protection action of the energy storage battery.

[0184] The final control signal here is a standardized electrical signal generated by the edge intelligent terminal in step S5 after fusing the first control command and the second optimization strategy, including the PWM signal (duty cycle) driving the charge and discharge controller. The actuator adopts an integrated architecture of "charge and discharge control unit + temperature control execution unit + fault protection unit" and digital level signals for controlling temperature control equipment. Each unit interacts with the edge intelligent terminal through the CAN bus to ensure real-time response to control commands. The response delay must meet <1ms (guaranteed by the hardware performance parameters of the actuator and verified in advance through offline testing).

[0185] The charging and discharging regulation is executed by the charging and discharging control unit, whose core component is a bidirectional DC / DC converter. Its function is to convert the PWM duty cycle command in the final control signal into actual charging and discharging current and voltage regulation, enabling bidirectional power flow between the energy storage battery and the grid / load. The output voltage of the bidirectional DC / DC converter has a linear relationship with the PWM duty cycle. Based on the continuous current mode (CCM) principle of inductor, its voltage gain formula is:

[0186]

[0187] In the formula, For the voltage gain of the DC / DC converter, The DC bus voltage of the converter (the target value is preset to 750V to match the grid voltage level). The terminal voltage of the energy storage battery (source: average single-cell voltage after preprocessing in step S1) (where n is the number of individual cells). This formula enables precise matching of the PWM duty cycle and the bus voltage, thereby regulating the charging and discharging power. The closed-loop control of the charging and discharging current employs a PI regulation algorithm, using the fused charging and discharging power command. The derived target current To control the target, the formula for calculating the target current is:

[0188]

[0189] In the formula, The target charge / discharge current (positive during charging and negative during discharging). The output of the PI regulator is the real-time battery terminal voltage; the output of the PI regulator is the correction amount for the PWM duty cycle. The final drive duty cycle is The mathematical expression for the PI control algorithm is:

[0190]

[0191] In the formula, The deviation between the target current and the actual current. The real-time output current of the charge / discharge controller (acquired by a current sensor and fed back to the edge intelligent terminal). This is the proportionality coefficient. The integral coefficient (both are pre-calibrated offline to ensure the stability and speed of current regulation); through PI closed-loop control, the actual charging and discharging current accurately tracks the target current, achieving... The corresponding power adjustment effect.

[0192] The fault protection action is executed by the fault protection unit. The trigger condition is that the final control signal contains the fault protection command generated in step S2 (such as overvoltage, overtemperature, or overcurrent faults). Core actions include disconnecting the main charging / discharging circuit, activating the emergency cooling / fire extinguishing device, and issuing a fault alarm signal. Main circuit disconnection is achieved by controlling a high-speed DC contactor. The contactor's response time is ≤20ms, meeting the requirement for rapid fault isolation. The execution logic of the fault protection action must follow the "graded response" principle, quantifying the fault severity. (Source: S2 step fault diagnosis results) Three levels of fault classification: Level 1 fault ( Minor malfunction (Preset to 0.3) Only current / voltage limiting regulation is performed, without disconnecting the main circuit; Level 2 fault ( Moderate fault (Preset to 0.6) Cut off the main charging and discharging circuit and start normal heat dissipation; Level 3 fault ( (Severe fault) The main circuit will be cut off, the emergency cooling / fire extinguishing device will be activated, and an audible and visual alarm will be issued. The formula for determining the quantification value of fault severity is:

[0193]

[0194] In the formula, A preset fault level threshold (determined by the battery manufacturer's fault safety standards and pre-stored in the edge intelligent terminal) ensures the accuracy and rationality of fault protection actions. Temperature control actions are executed by the temperature control execution unit, based on the temperature control instructions in the final control signal. The core function of the temperature control execution unit is to stabilize the battery temperature at the final temperature control threshold after fusion, allowing users to choose whether to activate the heat dissipation device (air cooling / liquid cooling), the heating device, or maintain the current state. Nearby. When cooling is activated, if the average battery temperature... ( (Preset heat dissipation level 1 threshold), start the air cooling device; if ( (Preset secondary heat dissipation threshold), simultaneously activating both air cooling and liquid cooling systems. The heat dissipation power adjustment of the liquid cooling system is related to the PWM duty cycle, as shown in the formula:

[0195]

[0196] In the formula, This refers to the real-time heat dissipation power of the liquid cooling device. This is the maximum heat dissipation power of the liquid cooling device (hardware parameter, preset). Maximum permissible overheat value (preset); heating power when heating is started. ,in This represents the maximum heating power of the heating device. To ensure the maximum permissible low temperature, the heating process is kept stable to avoid damage to the battery from sudden temperature rises.

[0197] After the actuator completes the charge / discharge regulation or fault protection action, it needs to collect real-time operating data after the action through the sensor array (voltage, current, temperature sensors) and feed it back to the edge intelligent terminal to form a closed-loop verification of the action effect.

[0198] Verification indicators include charge / discharge power deviation. Battery temperature deviation The fault clearance status (whether the fault diagnosis result returns to fault-free after the fault protection action) has a preset verification threshold of [value missing]. , ;like and If the verification conditions are met, the action is deemed valid; if the verification conditions are not met, the edge intelligent terminal regenerates the control signal to drive the actuator to make secondary adjustments. The number of secondary adjustments is preset to 2. If the standard is still not met, a cloud alarm is triggered, and an emergency optimization strategy is generated by the cloud and sent for execution.

[0199] In addition, the operating status data of the actuator (including the number of actions executed, cumulative running time, fault trigger records, and energy consumption data) needs to be uploaded to the edge intelligent terminal in real time and synchronized to the cloud according to the data upload cycle of step S3. This serves as an important basis for the cloud to evaluate the reliability of the actuator and optimize the control strategy. It forms a complete connection with the closed-loop logic of "cloud strategy distribution - edge fusion execution - data feedback evaluation" mentioned above, ensuring the stable and efficient operation of the entire energy storage battery management system.

[0200] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0201] The technical effectiveness of this energy storage battery management system relies on a complete technical chain: "precise edge sensing - adaptive preprocessing - heterogeneous collaborative core computing - secure and efficient data interaction - global model self-evolution." Each stage of the technical approach works in synergy and progresses progressively, ultimately achieving the core objectives of accurate battery status estimation, timely fault diagnosis, full lifecycle adaptation, and stable and reliable system operation. The specific logical derivation is as follows:

[0202] First, an edge-side data acquisition system built through a multi-dimensional sensor network achieves full-domain coverage acquisition of battery operating parameters. This technology integrates multi-dimensional sensors for voltage, current, temperature, internal resistance, and casing deformation, combined with a 100Hz base sampling frequency, providing comprehensive and real-time raw data input for subsequent algorithms. Building upon this, adaptive data preprocessing techniques utilize low-pass filtering to match the battery voltage fluctuation frequency, dynamic sampling frequency adjustment driven by operational complexity, differential compression based on the analytic hierarchy process (AHP), and outlier removal using the 3σ criterion. These techniques achieve noise filtering, redundancy reduction, and quality improvement of the raw data. Low-pass filtering ensures the retention of effective signals; dynamic sampling balances data timeliness and hardware consumption; differential compression reduces transmission and storage pressure while ensuring the preservation of core features; and outlier removal avoids interference from invalid data in subsequent calculations. Together, these technologies output high-quality, highly reliable input data for the core algorithm, laying the foundation for accurate computation.

[0203] Secondly, the edge intelligent terminal adopts a heterogeneous architecture of ARM and FPGA, providing hardware support for the efficient operation of the core computing algorithm. The FPGA is responsible for high-frequency matrix operations and convolution operations, while the ARM is responsible for algorithm scheduling and result fusion. Low-latency data interaction is achieved through the AXI bus, ensuring that the computational latency of the core algorithm is controlled within a preset range. The core algorithm adopts a cascaded architecture of neural network and Kalman filter. The LSTM network is used to mine the temporal correlation of operating parameters and output the prior estimate of battery state. Then, the state equation and observation equation are constructed by combining the second-order RC equivalent circuit model. The EKF algorithm is used to correct the error of the prior estimate, thereby improving the accuracy of state estimation. Temperature adaptive correction and aging adaptive correction technologies are superimposed. The algorithm parameters are dynamically adjusted through the temperature-internal resistance correlation model and SOH correction factor, so that the state estimation can adapt to different temperature environments and battery aging stages, further ensuring the estimation accuracy throughout the entire operating conditions and the entire life cycle.

[0204] Meanwhile, the fault diagnosis algorithm constructed by the twin structure feature coding model enhances the ability to extract and retain micro-fault features through dual-branch shared parameters and residual connection design; combined with the construction of multi-dimensional fault features (time domain + frequency domain) and dynamic weighting strategy, it allocates initial weights based on fault mechanism and dynamically adjusts them with SOH, charge / discharge rate, and temperature fluctuations to improve the identification and diagnostic sensitivity of fault features; through feature coding residual calculation and contribution analysis, it achieves accurate fault determination, severity quantification, and type positioning, providing a reliable basis for the generation of fault protection instructions and reducing the risk of battery fault expansion.

[0205] The data interaction process employs the national standard SM4 encryption algorithm and structured data frame design to ensure the security and integrity of operating parameters, state estimation, and fault diagnosis results during transmission over industrial communication networks, preventing data leakage and transmission errors. The combination of a core feature priority transmission strategy and a 5G communication link ensures real-time uploading of critical data, providing timely data support for cloud-based global management. The model self-evolution technology, through a collaborative mode of local parameter iteration at edge nodes and global aggregation in the cloud, combined with a learning rate strategy dynamically adjusted according to the State of Health (SOH), continuously optimizes algorithm parameters. Simultaneously, by utilizing the aggregated weight allocation of each edge node, the global model adapts to the battery operating characteristics under different scenarios, enhancing the system's adaptability to diverse application scenarios.

[0206] Finally, the first control command generated by the edge side based on the core calculation results, through the logical design of "state deviation adjustment" and "fault priority protection", directly drives the actuator to realize charge and discharge regulation or fault handling, forming an edge-side closed loop of "acquisition-processing-computation-control". The cloud, by receiving the data uploaded by each edge node, realizes the monitoring and management of the battery's operating status across the entire domain. Combined with the distribution of global algorithm parameters, it coordinates and optimizes the operating performance of the entire system. The synergistic cooperation of technical means in each link ultimately achieves the technical effects of precise control of the energy storage battery's operating status, timely fault warning and handling, efficient and stable system operation, and full life cycle adaptation.

[0207] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A cloud-edge collaborative intelligent management method for energy storage batteries, characterized in that, This includes edge-side data acquisition steps, edge-side core computing steps, data upload steps, cloud-based global optimization steps, policy distribution and integration steps, and execution steps. The edge-side data acquisition process involves collecting the operating parameters of the energy storage battery through multi-dimensional sensing devices. The core computing steps on the edge side involve the edge intelligent terminal calling a first algorithm to process the operating parameters, generating real-time battery state estimation results and / or fault diagnosis results, and generating a first control command based on the results. The first algorithm is a combined algorithm, which at least includes a battery state estimation algorithm using a cascaded neural network and Kalman filter, and a battery fault diagnosis algorithm using a twin structure feature encoding model. The data upload step involves the edge intelligent terminal encrypting the operating parameters, state estimation results, and / or fault diagnosis results, and then uploading them to the cloud server platform via the industrial communication network. The cloud-based global optimization process involves the cloud server platform calling a second algorithm to process the uploaded data, performing global optimization calculations or digital twin simulation verification, and generating a second optimization strategy. The strategy distribution and integration process involves the cloud distributing the second optimization strategy to the edge, and the edge intelligent terminal integrating the first control command and the second optimization strategy to generate the final control signal. The execution steps involve the actuator receiving the final control signal to complete the charging and discharging regulation or fault protection actions of the energy storage battery.

2. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 1, characterized in that, The execution process of the battery state estimation algorithm in the edge-side core computing step includes: Obtain historical and real-time operating parameters of the energy storage battery to construct time-series input data. The time-series input data is fed into a pre-trained neural network model to output a prior estimate of the battery state. State equations and observation equations are constructed based on battery equivalent circuit models or electrochemical mechanism models. The prior estimate output by the neural network is used as the initial state and substituted into the filtering algorithm. The error is corrected through two sub-steps: state prediction and observation update, and the final battery state estimate is output. After each charge-discharge cycle, the neural network model is fine-tuned using new operational data.

3. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 2, characterized in that, The execution process of the battery state estimation algorithm also includes a temperature adaptive correction step, which is as follows: during the process of constructing the state equation and the observation equation, the battery internal resistance correction value at different temperatures is calculated in real time based on the correlation between temperature and battery internal resistance, and the internal resistance correction value is substituted into the state equation or observation equation of the filtering algorithm to replace the fixed internal resistance parameter.

4. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 2 or 3, characterized in that, The execution process of the battery state estimation algorithm also includes an aging adaptive correction step, which is as follows: calculate the battery health state SOH, embed the battery health state SOH as a correction factor into the state equation of the filtering algorithm, and recalculate the battery health state SOH and update the correction factor in the state equation synchronously after each set number of charge-discharge cycles.

5. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 1, characterized in that, The execution process of the battery fault diagnosis algorithm in the edge-side core computing step includes: Collect battery operation data to extract multi-dimensional fault features and construct feature vectors. A twin-structured feature encoding model is constructed, comprising two structurally identical encoding branches. The first branch is trained using normal battery data to reconstruct input features, while the second branch processes the feature vector of the battery to be diagnosed. Residual connections are introduced during the encoding process to improve the ability to retain micro-fault features. The feature vector is simultaneously input into two encoding branches, which output normal feature codes and features to be diagnosed, respectively. Calculate the residual between the two feature codes, and determine whether a fault has occurred and its severity based on the comparison result of the residual with a preset threshold. The specific fault type can be identified by combining residual contribution analysis.

6. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 5, characterized in that, The execution process of the battery fault diagnosis algorithm also includes a feature dynamic weighting step, which is as follows: assign initial weights to multi-dimensional fault features based on the battery fault mechanism, adjust the feature weights in real time according to the battery health status (SOH), construct a new feature vector using a weighted summation method, and update the feature weights immediately when the battery charge / discharge rate or temperature fluctuation exceeds a set threshold.

7. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 1, characterized in that, It also includes an adaptive data preprocessing step, which is located after the edge-side data acquisition step and before the edge-side core computing step, and includes at least one of the following processes: The raw data was filtered using a low-pass filter that matched the main frequency of battery voltage fluctuations. The sampling frequency is dynamically adjusted according to the complexity of the operating conditions; Different compression strategies are adopted for data features of different importance; Outlier removal criteria are used to remove outliers from the filtered data.

8. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 1, characterized in that, It also includes a model self-evolution step, which is as follows: Each edge node initializes the first algorithm based on local historical data, and updates the local algorithm parameters using a parameter update algorithm with a dynamic adjustment factor. The dynamic adjustment factor is dynamically adjusted according to the battery health state (SOH). The gradient data updated by the edge node extraction algorithm is encrypted and uploaded to the cloud. The cloud-based algorithm generates global algorithm parameters by weighting and aggregating the gradients of each node according to the set aggregation weights. The cloud distributes global algorithm parameters to each edge node to update the local algorithm, and completes local iteration and global aggregation according to the set period.

9. The cloud-edge collaborative intelligent management method for energy storage batteries according to claim 1, characterized in that, It also includes a model distillation co-optimization step, which is as follows: A complex optimization model is trained in the cloud based on aggregated data from multiple sites. This complex optimization model serves as the teacher model, while the lightweight model on the edge serves as the student model. A distillation loss function with battery life loss constraints is constructed to transfer knowledge from the teacher model to the student model. The distilled student model is sent to the edge intelligent terminal as part of the first algorithm. The control commands generated by the student model are uploaded to the cloud for simulation verification using a digital twin model. If the lifespan loss exceeds the threshold, the cloud will be triggered to re-distill.

10. A cloud-edge collaborative intelligent management device for energy storage batteries, characterized in that, The device is used to execute the cloud-edge collaborative intelligent management method for energy storage batteries according to any one of claims 1-9, and includes a multi-dimensional sensor network, an edge intelligent terminal, a cloud server platform, an industrial communication link, and an actuator. Multi-dimensional sensor networks are used to collect operating parameters of energy storage batteries. The edge intelligent terminal adopts a heterogeneous architecture and is configured to execute the edge-side core computing steps, policy fusion steps, and related algorithm steps of the method described above. The cloud server platform includes a computing cluster, a digital twin module, and a data security module, configured to execute the cloud-based global optimization step and related aggregation and model distillation steps of the method described above. Industrial communication links are used to enable data interaction between edge intelligent terminals and cloud server platforms. The actuator is used to perform charge / discharge regulation or fault protection actions of the method.

Images