A federated learning collaborative optimization method for energy storage battery safety prevention and control

Abstract

The application relates to the field of new energy technology, in particular to a federal learning collaborative optimization method for safety prevention and control of energy storage batteries. First, multi-dimensional data of the batteries is extracted, and a double-cycle federal optimization architecture is constructed. The double-cycle federal optimization architecture comprises an outer cycle module and an inner cycle module. The outer cycle module is used for periodic updating and optimization of a global model, and the inner cycle module is used for iterative optimization of a local model. Then, based on the double-cycle federal optimization architecture, a federal learning model is constructed, inner cycle training is carried out, first updating parameters are obtained and uploaded to a central server. After aggregation of the first updating parameters, outer cycle training is carried out, second updating parameters are obtained. After multiple cycles, the first updating parameters and the second updating parameters are fused to obtain double-time optimization precision parameters. According to the double-time optimization precision parameters, a global loss function, a local loss function and a time precision loss function are calculated, and a target optimization function is obtained by weighted summation. Finally, an optimization algorithm is used to optimize the running state.

Description

Technical Field

[0001] The present invention relates to the field of new energy technologies, and in particular to a federated learning collaborative optimization method for safety control of energy storage batteries. Background Art

[0002] Federated learning, a distributed machine learning technology, demonstrates potential in energy storage battery fault prediction and load optimization by training neural network models locally and aggregating global models while protecting data. Existing technologies employ horizontal federated learning to optimize global neural networks by sharing model parameters, adapting to similar data scenarios across different energy storage systems. Vertical federated learning integrates heterogeneous feature data from battery manufacturers and operators to improve model prediction accuracy. Supporting technologies such as data protection mechanisms and local computing support enhance training efficiency and security.

[0003] However, existing federated learning methods have limitations in energy storage battery safety management. The operating environments and battery types of different energy storage systems result in non-independent data distribution, poor adaptability of neural network models, and insufficient generalization capabilities. Federated learning relies on frequent gradient or parameter synchronization, resulting in high communication and computational overhead, limiting its application on resource-constrained devices. Model parameters may be exposed to security threats, impacting system reliability.

[0004] To address the above problems, the present invention proposes an improved federated learning collaborative optimization method, develops a distributed learning method framework called double-loop federated optimization learning, integrates the characteristics of multi-source data, enhances the accuracy of fault prediction and state estimation, further optimizes the training parameters through the dynamic weight adjustment algorithm of the optimization algorithm, adapts to the data distribution of different battery types, compresses the model, improves the model convergence speed and generalization ability, optimizes the data protection strategy to reduce computing overhead, ensures safe and efficient training, and realizes efficient distributed management of energy storage battery safety prevention and control. Summary of the Invention

[0005] The present invention is specifically a federated learning collaborative optimization method for energy storage battery safety control, which is used for safety control and optimization of energy storage batteries. The specific implementation steps are: first, extract battery multidimensional data and build a dual-loop federated optimization architecture, which includes an outer loop module and an inner loop module. The outer loop module is used for periodic update optimization of the global model, and the inner loop module is used for iterative optimization of the local model; then, based on the dual-loop federated optimization architecture, build a federated learning model, perform inner loop training, obtain the first update parameter and upload it to the central server; secondly, aggregate the first update parameter and perform an outer loop to obtain the second update parameter; after multiple cycles, fuse the first update parameter and the second update parameter to obtain the dual-time optimization precision parameter, calculate the global loss function, local loss function and time precision loss function according to the dual-time optimization precision parameter, and perform weighted summation to obtain the target optimization function; finally, optimize the operating state through the optimization algorithm. The present invention integrates the characteristics of multi-source data to enhance the accuracy of fault prediction and state estimation. The optimized data protection strategy reduces computational overhead and ensures safe and efficient training. The present invention solves the collaborative problem of model accuracy, real-time control, and long-term life in the safety prevention and control of energy storage batteries through the deep coupling of federated learning and multi-objective optimization.

[0006] A federated learning collaborative optimization method for energy storage battery safety control includes: dividing multiple nodes to extract multi-dimensional battery data, constructing a dual-loop federated optimization architecture, the dual-loop federated optimization architecture including an outer loop module and an inner loop module, the outer loop module is used for periodic update optimization of a global model, and the inner loop module is used for iterative optimization of a local model; constructing a federated learning model based on the dual-loop federated optimization architecture, the federated learning model including a local model and a global model; performing inner loop training through the inner loop module to obtain a first update parameter and uploading it to a central server; aggregating the first update parameter, performing outer loop training through the outer loop module to obtain a second update parameter; fusing the first update parameter and the second update parameter to obtain a dual-time optimization precision parameter; calculating a global loss function, a local loss function and a time precision loss function according to the dual-time optimization precision parameter, and then performing a weighted summation thereof to obtain a target optimization function; minimizing the target optimization function through an optimization algorithm and optimizing each node.

[0007] Preferably, the process of building a dual-loop federated optimization architecture includes:

[0008] Each node participating in federated learning corresponds to an energy storage battery. Each node collects multi-dimensional data during the operation of the battery, including voltage, current, temperature, and internal resistance, and preprocesses these data. The preprocessing methods include data cleaning and normalization operations. The inner loop module is mainly used for the iterative optimization of the local model. The initial model or the global model is trained using the local multi-dimensional data of each node to obtain a local model that is more adapted to the characteristics of the local data. It is responsible for the periodic update optimization of the global model. By aggregating the first update parameters uploaded by each node, the global model is updated so that it can adapt to the overall data distribution of all participating nodes. The federated learning model based on deep neural network is initialized on the central server and the initial parameters of the model are set.

[0009] Preferably, the inner loop training, obtaining the first updated parameter and uploading it to the central server comprises:

[0010] According to the global model sent by the central server received by each node, training is performed using local multi-dimensional data based on the global model. During the training process, the deep learning optimization algorithm is used to update the model parameters. After each specified number of local training iterations, the performance indicators of the local model on the validation set are calculated and compared with the local optimization accuracy parameters. If the local optimization accuracy requirements are met, the local model training is stopped and the local model update parameters are obtained; otherwise, the local training iterations are continued, and the local model update parameters are encrypted using homomorphic encryption technology, and the encrypted local model update parameters are uploaded to the central server.

[0011] Preferably, the process of aggregating the first update parameters and performing outer loop training to obtain the second update parameters includes:

[0012] The central server receives the encrypted local model update parameters uploaded by each node. After collecting the specified number of local model update parameters, it prepares to update the global model. The central server temporarily stores and manages the received update parameters. The root central server uses the homomorphic encryption key to decrypt the collected local model update parameters, and aggregates these update parameters using dynamic weights to obtain the second update parameters. The global model update parameters are calculated using the second update parameters. According to the global model update parameters, the global federated learning model is updated. The updated global model is further processed and optimized, and sent to each node through a secure communication channel for the node to conduct the next round of local model training, and check whether the global model meets the overall optimization goals and accuracy requirements. If so, the training process ends.

[0013] Preferably, the process of fusing the first update parameter and the second update parameter to obtain the dual-time optimization precision parameter includes:

[0014] According to each inner and outer loop training, a cluster analysis is performed on multiple first and second update parameters, and similar update parameters are classified into one category. The weighted average of the update parameters in each category is calculated as the representative update parameter of each category. The representative update parameters of different categories are fused to obtain the dual-time optimization precision parameters.

[0015] Preferably, the process of calculating the global loss function, the local loss function and the time precision loss function includes:

[0016] The dual-time optimization precision parameter is used as the weight factor, and the prediction errors of different nodes are weighted summed to obtain the global loss function; the dual-time optimization precision parameter is used as the constraint factor of the local model complexity, and the mean square error is used to analyze the prediction error of local data to obtain the local loss function; the dual-time optimization precision parameter is used as the time weight factor, and the prediction errors at different time points are weighted summed to obtain the time accuracy loss function.

[0017] Preferably, the process of obtaining the target optimization function includes:

[0018] Based on the obtained global loss function, local loss function and time accuracy loss function, a weighted summation is taken to obtain the target optimization function. The weight coefficient is determined according to the actual needs of energy storage battery safety prevention and control and the importance of each loss function in the scenario.

[0019] Preferably, the optimization algorithm minimizes the target optimization function and optimizes each node in the process including:

[0020] By minimizing the target optimization function through the optimization algorithm, the battery temperature, power and frequency are maintained within a reasonable range, so that the optimized operation of the battery is coordinated with the control strategy of the entire energy storage system; after optimization, the battery performance indicators are tested and evaluated, including the battery's capacity retention rate, internal resistance change, and cycle life. The optimized battery performance indicators and operating status are fed back to the central server, and the optimization algorithm and parameter settings are further adjusted based on the feedback information.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The dual-federation optimization architecture proposed in the present invention constructs a dual-loop federation optimization architecture comprising an outer loop module and an inner loop module. The outer loop module is responsible for the periodic update optimization of the global model, while the inner loop module is responsible for the iterative optimization of the local model. This two-layer loop structure enables both the global model and the local model to be fully optimized, improving the overall optimization effect. The dual-federation optimization architecture enables the global model to better reflect the characteristics of different nodes, thereby enhancing the model's adaptability and accuracy.

[0023] 2. The target optimization function proposed in the present invention forms a comprehensive target optimization function by weighted summation of the global loss function, the local loss function and the time precision loss function. The target optimization function is based on the dual-time precision parameters, which are obtained by clustering analysis of multiple parameters. It can more accurately reflect the optimization accuracy of the model at different time and space scales. By adjusting the weight coefficient, it can flexibly highlight the optimization focus of different loss functions and realize all-round and multi-level comprehensive performance optimization of energy storage batteries.

[0024] 3. The optimization algorithm proposed in the present invention optimizes control parameters by minimizing the target optimization function to meet constraints such as temperature and power. It not only focuses on the performance indicators of the energy storage battery itself, but also conducts collaborative optimization with other systems. After optimization, the battery performance indicators are tested and evaluated, including the battery's capacity retention rate, internal resistance change, and cycle life. The optimized battery performance indicators and operating status are fed back to the central server. Based on the feedback information, the optimization algorithm and parameter settings are further adjusted to coordinate the operating status of the energy storage battery with the control strategy of the entire system, thereby achieving efficient collaborative operation of the energy storage battery and the external system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flow chart of a federated learning collaborative optimization method for energy storage battery safety control proposed in an embodiment of the present invention;

[0026] Figure 2 This is a flow chart of the inner loop of the dual-loop federated optimization architecture proposed in an embodiment of the present invention;

[0027] Figure 3 This is a flow chart of the outer loop of the dual-loop federated optimization architecture proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] Existing federated learning methods have limitations in energy storage battery safety management. The varying operating environments and battery types of different energy storage systems lead to non-independent data distribution, poor adaptability of neural network models, and insufficient generalization capabilities. Federated learning relies on frequent gradient or parameter synchronization, resulting in high communication and computational overhead, limiting its application on resource-constrained devices. Furthermore, model parameters may be exposed to security threats, impacting system reliability.

[0030] The present invention relates to the field of new energy technologies, and more specifically to a federated learning collaborative optimization method for energy storage battery safety control. The method is suitable for further optimizing the safety control of energy storage batteries, improving safety control efficiency, and enhancing optimization effects. To illustrate the effectiveness of the method and system of the present invention, a detailed description will be provided in conjunction with the accompanying drawings of this embodiment and the following embodiments.

[0031] Example 1

[0032] The present application embodiment discloses a federated learning collaborative optimization method for energy storage battery safety control, such as Figure 1 As shown, safety control is performed on the lithium-ion battery pack in the type A grid energy storage system. The grid energy storage system includes multiple distributed energy storage sites. Each site manages a group of batteries. The operating status of the batteries needs to be monitored and optimized in real time to ensure safety, power balance and cycle life. First, the multi-dimensional data of the battery is extracted to construct a dual-loop federated optimization architecture. The dual-loop federated optimization architecture includes an outer loop module and an inner loop module. The outer loop module is used for periodic update optimization of the global model, and the inner loop module is used for iterative optimization of the local model. Then, based on the dual-loop federated optimization architecture, a federated learning model is constructed, and inner loop training is performed to obtain the first update parameter and upload it to the central server. Secondly, the first update parameter is aggregated and then an outer loop is performed to obtain the second update parameter. After multiple cycles, the first update parameter and the second update parameter are fused to obtain the dual-time optimization precision parameter. The global loss function, local loss function and time precision loss function are calculated according to the dual-time optimization precision parameter, and the weighted sum of them is obtained to obtain the target optimization function. Finally, the operating status is optimized through the optimization algorithm.

[0033] Furthermore, a dual-loop federated optimization architecture is constructed. Corresponding to the above steps, the specific implementation is as follows:

[0034] Each node participating in federated learning is mapped to a specific energy storage battery. Each node collects multidimensional data from the battery's operation in real time, including but not limited to voltage, current, temperature, and internal resistance. This collected multidimensional data is preprocessed, including data cleaning to remove noise and outliers and normalization to standardize the data. On the central server, a federated learning model is initialized based on a deep neural network. Initial model parameters, including network structure parameters, weight matrix, and bias terms, are set. The communication protocol between the central server and each node is defined to ensure the accuracy and timeliness of data transmission, including but not limited to encrypted communication, packet format, and transmission frequency. Each node is assigned a unique identifier so that the central server can distinguish and manage data and parameters uploaded by different nodes. A data storage module is established on the central server to temporarily store local model update parameters uploaded by each node and to store intermediate results used for aggregate calculations. The central server sets a period for updating the global model and a number of iterations for iterative optimization of the local model to ensure the coordinated operation of the internal and external loops. A verification module is developed on the central server to verify that the global model meets the overall optimization objectives and accuracy requirements, and corresponding verification metrics and thresholds are set.

[0035] This embodiment performs data preprocessing operations, initializes a federated learning model based on a deep neural network on a central server, and sets reasonable initial parameters. This ensures that the model has a good performance foundation in the early stages of training. The facility uses a dual architecture of internal and external loops to continuously optimize the global model and the local model, enabling the model to better adapt to the actual operating status of the energy storage battery, improving optimization accuracy, and achieving safety control and performance improvement of the energy storage battery.

[0036] Furthermore, the inner loop training is performed through the inner loop module to obtain the first updated parameters and upload them to the central server, such as Figure 2 As shown, corresponding to the above steps, the specific implementation includes:

[0037] Specifically, each node receives a global model based on a deep neural network, which serves as the initial model. The local federated learning model is trained using local multidimensional data, including voltage, current, temperature, and internal resistance. During training, the Adam optimization algorithm, a deep learning optimization algorithm, is used to update model parameters. After every 10 local training iterations, the mean squared error (MSE) of the performance metric on the local model validation set is calculated and compared to a local optimization accuracy threshold of 0.01. If the MSE is less than the local optimization accuracy threshold, local training is terminated and the first updated parameter a is obtained. Otherwise, iterations continue until the maximum number of iterations is reached or the accuracy requirement is met. The first updated parameter a is encrypted using homomorphic encryption and uploaded to the central server via a secure communication channel. Timestamps are recorded during the upload process to support temporal precision loss calculations. During the encrypted upload to the central server, the convergence speed, computing resource usage, and communication latency of each node's local training are simultaneously recorded. An optimization feedback log is generated and fed back to the central server for dynamic adjustment of the number of inner loop iterations and the optimization algorithm to improve training efficiency and model performance.

[0038] This implementation uses Adam as an optimization algorithm for training on multidimensional data, updates model parameters, completes local training iterations a certain number of times, calculates performance metrics and compares them against a local optimization accuracy threshold, and uploads the first updated parameters encrypted after the iterations are complete, reducing the risk of data leakage. Through high-precision local model training and privacy protection, accurate battery status predictions for large-scale energy storage sites are ensured, improving the training efficiency and adaptability of the overall system.

[0039] Furthermore, the first update parameters are aggregated and the outer loop training is performed through the outer loop module to obtain the second update parameters, such as Figure 2 As shown, corresponding to the above steps, the implementation process includes:

[0040] The central server receives the encrypted local model update parameters uploaded by each node during the outer loop iteration, where represents the first update parameter. The server stores these parameters in a secure database and records the timestamp and node identifier for each received parameter to ensure traceability and support the calculation of time precision loss. The central server waits until it receives the encrypted parameters uploaded by 80% of the total number of nodes, or when the maximum waiting time of 1 second is reached, it starts the global model update to ensure sufficient node participation while maintaining system efficiency. The central server uses the homomorphic encryption key to decrypt the collected encrypted parameters and obtain the decrypted first update parameter a. The decryption process is performed in a secure computing environment to prevent unauthorized access. Dynamic weights are calculated for each node with weight coefficients of 0.4, 0.4, and 0.2 Using dynamic weights The decrypted first update parameter a is summed for each participating node to obtain the second update parameter b, and the performance indicator mean square error MSE is calculated. If the global model satisfies MSE < 0.005, the training process is terminated; otherwise, further optimization is performed. The optimized global model is encrypted using a secure communication protocol and sent to each node through a secure channel for the next round of local model training.

[0041] This embodiment decrypts the update parameters, aggregates them with the first update parameters using dynamic weights, obtains the second update parameters, and calculates the global model update parameters using the second update parameters. Based on the global model update parameters, the global federated learning model is updated and further optimized, thereby reducing the impact of noisy data or unreliable nodes, accelerating convergence, and saving computing resources.

[0042] Furthermore, the fusing of the first update parameter and the second update parameter to obtain the dual-time optimization precision parameter corresponds to the above steps and is specifically implemented as follows:

[0043] In each inner and outer loop training iteration, the first update parameter a generated by each node during the iteration and the second update parameter b generated by the central server are collected. In order to reduce or block the impact of low-quality data and enhance the robustness and anti-interference ability of the system, the collected first and second update parameters are clustered and analyzed, and similar update parameters are grouped using a clustering algorithm. For each identified cluster, the product of the cluster parameter and the inverse of the local MSE is summed to calculate the representative update parameter within the cluster, which is used as the weighted average of the cluster parameters. The dual-time optimization accuracy parameters are obtained by summing the fusion weights and the weighted average of the cluster parameters. ,The fusion process is adaptive, and the fusion weights are adjusted according to the real-time feedback of the target optimization function.

[0044] This embodiment performs cluster analysis on multiple update parameters, classifies similar update parameters into one category, calculates the average value of the update parameters in each category, and uses it as the representative update parameter of each category. The representative update parameters of different categories are fused to obtain dual-time optimization precision parameters, ensuring that the clustering reflects the diversity of battery operating conditions, improving the accuracy of battery status prediction, and reducing the impact of noise or unbalanced data on certain nodes.

[0045] Furthermore, the calculation of the global loss function, the local loss function, and the time precision loss function corresponds to the above steps, and the specific implementation includes:

[0046] The dual-time optimization accuracy parameter is used as the weight to perform weighted summation of the prediction error MSE of different nodes to obtain the global loss function ; Taking the dual-time optimization accuracy parameter as the constraint factor of the local model complexity, the mean square error is used to analyze the prediction error of the local data to obtain the local loss function ; Using the dual-time optimization accuracy parameter as the time weight factor, the prediction errors at different time points are weighted and summed to obtain the time accuracy loss function .

[0047] This embodiment calculates the global loss function, local loss function and time precision loss function through dual-time optimization precision parameters, ensuring that the loss function captures diverse error characteristics, enhancing the reliability of battery state prediction under different operating conditions and the precise control of battery parameters under dynamic conditions, while retaining implementation flexibility.

[0048] Furthermore, the target optimization function is obtained, corresponding to the above steps, and is specifically implemented as follows:

[0049] by , , Weight coefficient, satisfying =1, the initial value is =0.4, =0.4, =0.2, the global loss function , local loss function And time accuracy loss function Perform weighted summation to obtain the target optimization function To ensure the needs of different energy storage scenarios and maintain consistent performance, the results are fed back to the central server, and the weight coefficient and scenario adjustment coefficient are iteratively optimized through the reinforcement learning algorithm. When the temperature rises significantly, the , increases when voltage and current rise The weight of The weight of .

[0050] This embodiment obtains the target optimization function by taking weighted summation based on the obtained global loss function, local loss function and time precision loss function, and records the performance indicators of the target optimization function, including MSE, convergence speed and weight adjustment effect, to enhance dynamic adaptability and facilitate real-time monitoring and optimization of the battery's operating status.

[0051] Furthermore, the minimization of the target optimization function and optimization of each node by the optimization algorithm corresponds to the above steps, and the specific implementation includes:

[0052] By minimizing the objective optimization function using the deep learning optimization algorithm Adam , adjust the battery operating status corresponding to each node to ensure that the battery temperature remains between 20°C and 40°C, the power fluctuation rate is less than 10%, and the frequency deviation is within ±0.1Hz. The optimized model is used to predict power demand and adjust the node's charging and discharging strategy to coordinate with the overall control strategy of the energy storage system. After each feedback, the optimization effect is verified by evaluating the value of the performance indicator MSE of the global model. If the performance does not reach the expected 0.005, the reinforcement learning algorithm is used to further optimize the hyperparameters and control parameters. The verification results are recorded and fed back to the central server to iteratively improve the optimization process. If some nodes fail to feedback parameter results and parameters in a timely manner due to battery maintenance or communication delays, the central server adopts an outdated tolerance mechanism and uses the weighted average of historical parameters to update the global model to avoid training stagnation, reduce the impact of communication delays, and balance training efficiency and resource consumption.

[0053] This embodiment uses the Adam deep learning optimization algorithm to minimize the target optimization function, reduce the MSE value, ensure the stability of battery status prediction, and use an adaptive optimization strategy to adjust the battery status in real time, maintain the battery temperature, and feed the optimized battery status back to the central server in real time, reducing the convergence time, ensuring the stability of the optimization effect in dynamic scenarios, and improving the prediction accuracy of the global model.

[0054] This embodiment implements a federated learning collaborative optimization method for energy storage battery safety control. The specific process mainly includes the following steps: constructing a dual-loop federated optimization architecture, which includes an outer loop for periodic update optimization of the global model and an inner loop for iterative optimization of the local model; using the multidimensional data to perform inner loop training on the local federated learning model to obtain first update parameters, which are uploaded; aggregating the first update parameters to obtain a global model and then sending it to each node, performing an outer loop to obtain second update parameters; after multiple cycles, fusing the first and second update parameters to obtain dual-time optimization precision parameters, and calculating the global loss function, local loss function, and time precision loss function based on the dual-time optimization precision parameters; obtaining a target optimization function by weighted summation of the obtained loss functions, minimizing the target optimization function through an optimization algorithm, and optimizing the corresponding operating state of each node. The present invention proposes a dual-federated optimization architecture, target optimization function, and optimization algorithm for the above process. Starting from multidimensional data, the comprehensive performance indicators of the battery are tested and evaluated, and optimization and prediction are made, thereby achieving efficient and safe control of the energy storage battery.

[0055] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0056] Example 2

[0057] The embodiment of the present application discloses a federated learning collaborative optimization method for energy storage battery safety control, which is applied to a 4-node battery management system in an electric vehicle. Each node corresponds to a battery module in the battery pack. The data of the battery module includes voltage, resistance, temperature and internal resistance. It is necessary to monitor and optimize the battery status in real time, and the time accuracy requirement is high. First, the multi-dimensional data of the battery is extracted, and a dual-loop federated optimization architecture is constructed. The dual-loop federated optimization architecture includes an outer loop module and an inner loop module. The outer loop module is used for periodic update optimization of the global model, and the inner loop module is used for iterative optimization of the local model. Then, based on the dual-loop federated optimization architecture, a federated learning model is constructed, and inner loop training is performed to obtain the first update parameter and upload it to the central server. Secondly, the first update parameter is aggregated and then an outer loop is performed to obtain the second update parameter. After multiple cycles, the first update parameter and the second update parameter are fused to obtain a dual-time optimization precision parameter. According to the dual-time optimization precision parameter, the global loss function, the local loss function and the time accuracy loss function are calculated, and the weighted sum is obtained to obtain the target optimization function. Finally, the operating state is optimized by the optimization algorithm.

[0058] First, multi-dimensional data of the battery is collected and a dual-federated optimization architecture is constructed. The system is divided into four nodes, each equipped with a sensor to collect real-time data from the battery module, including voltage, resistance, temperature, and internal resistance, and perform preprocessing operations, including noise and outlier removal, normalization, and data standardization; the central server initializes a federated learning model based on a deep neural network, which consists of a three-layer fully connected network with randomly set initial parameters.

[0059] Each node receives the global model from the central server and performs an inner-loop training using local multidimensional data. The training process is as follows: optimization algorithm, training iterations, performance evaluation, and encrypted upload. The optimization algorithm uses the Adam optimizer. After each iteration, the mean squared error (MSE) of the local model on the validation set is calculated. If the MSE is less than 0.01, training is terminated, and the first updated parameter a is obtained.

[0060] Encrypted using a homomorphic encryption scheme, the data is uploaded to the central server via the CAN bus to ensure data privacy.

[0061] Wait for at least 3 nodes to complete the upload or the time is greater than 2s, start the global model update, use the key to decrypt the parameters, and obtain the first updated parameter a after decryption. , , The weight coefficients, 0.3, 0.3, and 0.4, are aggregated with the node reliability score and MSE to obtain the dynamic weight of each node. Then use the dynamic weight to aggregate the first update parameter a of the decryption, calculate the second update parameter, update the global model parameters through the second update parameter b, calculate the mean square error MSE, and terminate the training if the global MSE < 0.005. Collect the first update parameter a and the second update parameter b, perform cluster analysis on them, and use the clustering algorithm to calculate as the weighted average of the parameters within the cluster By integrating the representative update parameters of different clusters, the dual-time optimization accuracy parameters are obtained. The obtained dual-time optimization precision parameters are combined with MSE, sample actual value and predicted value to obtain 、 and , and generate the optimization objective function according to the weights of 0.3, 0.3 and 0.4 .

[0062] By minimizing the optimization objective function through an optimization algorithm, operating status adjustments were made, and the coolant flow rate was dynamically adjusted to maintain the temperature between 20°C and 40°C. The charge and discharge rates were optimized, with a power fluctuation rate of <8%. The capacity retention rate remained at 94% after 800 cycles. The internal resistance growth rate was <4%. The cycle life was expected to be 1,200 times. The optimization time was reduced from 0.5 seconds to 0.42 seconds.

[0063] This example demonstrates the efficient application of this method in optimizing the safety and control of electric vehicle battery systems. By collaboratively optimizing battery status through federated learning, it significantly improves safety, efficiency, and lifespan, making it suitable for dynamic driving and fast charging scenarios.

[0064] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A federated learning collaborative optimization method for energy storage battery safety control, characterized in that: The following steps are involved: Divide multiple nodes, extract multi-dimensional battery data, and build a dual-loop federated optimization architecture. The dual-loop federated optimization architecture includes an outer loop module and an inner loop module. The outer loop module is used for periodic update optimization of the global model, and the inner loop module is used for iterative optimization of the local model. Building a federated learning model based on the dual-loop federated optimization architecture, the federated learning model including a local model and a global model; performing inner-loop training through an inner-loop module to obtain first updated parameters and uploading them to a central server; Aggregating the first update parameters, performing outer loop training through the outer loop module, and obtaining the second update parameters; fusing the first update parameters and the second update parameters to obtain the dual-time optimization precision parameters; The process of fusing the first update parameter and the second update parameter to obtain the dual-time optimization precision parameter includes: Based on each inner and outer loop training, a cluster analysis is performed on multiple first and second update parameters, similar update parameters are grouped into one category, a weighted average of the update parameters in each category is calculated as the representative update parameter of each category, and the representative update parameters of different categories are fused to obtain the dual-time optimization accuracy parameter; Calculating a global loss function, a local loss function, and a time precision loss function according to the dual-time optimization precision parameters, and then performing a weighted summation thereof to obtain a target optimization function; the process of calculating the global loss function, the local loss function, and the time precision loss function includes: The dual-time optimization precision parameter is used as a weight factor to perform weighted summation of the prediction errors of different nodes to obtain the global loss function. The dual-time optimization precision parameter is used as a constraint factor for the complexity of the local model. The prediction error of local data is analyzed using the mean square error to obtain the local loss function. The dual-time optimization precision parameter is used as a time weight factor to perform weighted summation of the prediction errors at different time points to obtain the time accuracy loss function. The target optimization function is minimized and each node is optimized through the optimization algorithm.

2. The method for federated learning collaborative optimization of energy storage battery safety control according to claim 1, characterized in that: The process of building a dual-loop federated optimization architecture includes: Each node participating in federated learning corresponds to an energy storage battery. Each node collects multi-dimensional data during the battery operation process, including voltage, current, temperature, and internal resistance, and preprocesses these data. The preprocessing methods include data cleaning and normalization operations; an inner loop module is constructed for iterative optimization of the local model, and the initial model or global model is trained using the local multi-dimensional data of each node to obtain a local model that is more adapted to the local data characteristics; it is responsible for the periodic update optimization of the global model, and updates the global model by aggregating the first update parameters uploaded by each node so that it can adapt to the overall data distribution of all participating nodes; and initializes the federated learning model based on deep neural network on the central server and sets the initial parameters of the model.

3. The method for federated learning collaborative optimization of energy storage battery safety control according to claim 1, characterized in that: The inner loop training is performed by the inner loop module to obtain the first update parameter and upload it to the central server. The implementation process includes: According to the global model sent by the central server received by each node, training is performed using local multi-dimensional data based on the global model. During the training process, the deep learning optimization algorithm is used to update the model parameters. After each specified number of local training iterations, the performance indicators of the local model on the validation set are calculated and compared with the local optimization accuracy parameters. If the local optimization accuracy requirements are met, the local model training is stopped and the local model update parameters are obtained; otherwise, the local training iterations are continued, and the local model update parameters are encrypted using homomorphic encryption technology, and the encrypted local model update parameters are uploaded to the central server.

4. The method for federated learning collaborative optimization of energy storage battery safety control according to claim 1, characterized in that: The process of aggregating the first update parameters and performing outer loop training through the outer loop module to obtain the second update parameters includes: The central server receives the encrypted local model update parameters uploaded by each node. After collecting the specified number of local model update parameters, it prepares to update the global model. The central server temporarily stores and manages the received update parameters. The central server uses the homomorphic encryption key to decrypt the collected local model update parameters, and aggregates these update parameters using dynamic weights to obtain the second update parameters. The global model update parameters are calculated using the second update parameters. According to the global model update parameters, the global federated learning model is updated. The updated global model is further processed and optimized, and sent to each node through a secure communication channel for the node to conduct the next round of local model training, and check whether the global model meets the overall optimization goals and accuracy requirements. If so, the training process ends.

5. The method for federated learning collaborative optimization of energy storage battery safety control according to claim 1, characterized in that: The process of obtaining the target optimization function includes: Based on the obtained global loss function, local loss function and time accuracy loss function, a weighted summation is taken to obtain the target optimization function. The weight coefficient is determined according to the actual needs of energy storage battery safety prevention and control and the importance of each loss function in the scenario.

6. A federated learning collaborative optimization method for energy storage battery safety control according to claim 1, characterized in that The minimizing the target optimization function and optimizing each node by the optimization algorithm includes: By minimizing the target optimization function through the optimization algorithm, the battery temperature, power and frequency are maintained within a reasonable range, so that the optimized operation of the battery is coordinated with the control strategy of the entire energy storage system; after optimization, the battery performance indicators are tested and evaluated, including the battery's capacity retention rate, internal resistance change, and cycle life. The optimized battery performance indicators and operating status are fed back to the central server, and the optimization algorithm and parameter settings are further adjusted based on the feedback information.

Images