Retired battery remaining life prediction method and system, and battery management system

Abstract

The application provides a kind of decommissioned battery remaining life prediction method, system and battery management system, the decommissioned battery remaining life prediction method includes: end side equipment carries out feature extraction and correlation screening to the operation data of decommissioned battery, and battery characteristic data after screening is uploaded to side node;Based on the battery characteristic data after screening, the inference is carried out using battery life prediction model, and the remaining life prediction value is obtained;Wherein, the battery life prediction model uploads the parameters of the battery life prediction model by the side node, the cloud server carries out aggregation optimization to the parameters uploaded by each side node and issues the updated parameters, and the side node receives the updated parameters and updates.The method provided by the application can improve the accuracy and real-time performance of the remaining life prediction of decommissioned battery.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and specifically to a method, system, and battery management system for predicting the remaining lifespan of retired batteries. Background Technology

[0002] Retired batteries refer to power batteries that still have remaining capacity after being retired from equipment such as new energy vehicles, and are typically reused in energy storage systems. Accurately predicting the remaining usable life (RUL) of retired batteries during this reuse process is crucial for ensuring the safe and stable operation of energy storage systems and reducing maintenance costs.

[0003] In related technologies, existing methods for predicting the remaining life of retired batteries primarily rely on cloud servers for processing. This approach has the following drawbacks: First, battery management systems are limited by computing resources, making it difficult to deploy complex prediction models on the edge, resulting in underutilization of edge computing resources. Second, the battery types and operating conditions vary significantly across energy storage stations, making models trained at a single station difficult to apply to other stations. Third, reliance on cloud servers leads to significant network transmission latency, making it difficult to meet the real-time requirements of energy storage systems. These factors reduce the accuracy and real-time performance of remaining life prediction. Summary of the Invention

[0004] In view of this, it is necessary to provide a method, system and battery management system for predicting the remaining life of retired batteries, so as to solve the technical problems of low accuracy and real-time performance in predicting the remaining life of retired batteries in the prior art.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for predicting the remaining lifespan of retired batteries, applied to a retired battery energy storage system, comprising end-side equipment deployed in a battery management system, edge-side nodes deployed in each energy storage station, and a cloud server, the method comprising: The edge device performs feature extraction and correlation filtering on the operating data of retired batteries, and uploads the filtered battery feature data to the edge node. The side node uses the filtered battery feature data and a battery life prediction model to infer the remaining life prediction value of the retired battery. Specifically, the battery life prediction model uploads its parameters through the side nodes, the cloud server aggregates and optimizes the parameters uploaded by each side node and distributes updated parameters, and the side nodes receive the updated parameters and update them.

[0006] In one possible implementation, the operating data includes voltage data, state of charge, temperature data, and internal resistance data, and the battery characteristic data includes voltage characteristics, temperature characteristics, and internal resistance characteristics; the end-side device performs feature extraction on the operating data of the retired battery, including: The voltage characteristic is determined by differentiating the voltage data with respect to the state of charge. The standard deviation of the temperature data is defined as the temperature characteristic. The ratio of the change in internal resistance data to the change in the number of cycles is defined as the internal resistance characteristic.

[0007] In one possible implementation, the relevance screening step includes: Based on the joint probability distribution of the battery feature data and the remaining life tag of the retired battery, as well as their respective marginal probability distributions, the mutual information value between each of the battery feature data and the remaining life tag is calculated. The battery feature data is filtered for correlation based on the mutual information value.

[0008] In one possible implementation, before the side node uploads the parameters of the battery life prediction model, the method further includes: The parameters are encrypted using a homomorphic encryption algorithm.

[0009] In one possible implementation, the cloud server aggregates and optimizes the parameters uploaded by each side node, including: A weighted average algorithm is used to aggregate the parameters uploaded by each edge node to generate aggregated parameters; The aggregated parameters are iteratively optimized using a multimodal pre-trained model to generate updated parameters.

[0010] In one possible implementation, a weighted average algorithm is used to aggregate the parameters uploaded by each edge node, generating aggregated parameters, including: The weight of each side node is determined based on the proportion of historical data samples of retired batteries stored in each side node to the total sample size. The aggregated parameters are obtained by summing the products of the parameters uploaded by each side node and their respective weights.

[0011] In one possible implementation, before the cloud server sends the updated parameters, the following is also included: The updated parameters are compressed using low-rank decomposition.

[0012] In one possible implementation, the battery life prediction model is an LSTM network, and the network structure of the LSTM network includes an input layer, an LSTM layer, and a fully connected layer. The number of nodes in the input layer is equal to the dimension of the battery feature data; The LSTM layer is used to extract the time dependency of the filtered battery feature data; The fully connected layer is used to output the remaining lifetime prediction value.

[0013] Secondly, the present invention also provides a system for predicting the remaining life of retired batteries, applied to retired battery energy storage systems, comprising: The edge module, deployed in the battery management system, is used to extract features and filter correlations from the operating data of retired batteries, and upload the filtered battery feature data to the edge module. The edge module, deployed at each energy storage station, is used to infer the remaining life prediction value of the retired battery based on the battery characteristic data using the battery life prediction model; the edge module is also used to upload the parameters of the battery life prediction model, and to receive and update the updated parameters. The cloud module, deployed on a cloud server, is used to aggregate and optimize the parameters uploaded by each side module and then distribute the updated parameters.

[0014] Thirdly, the present invention also provides a battery management system for use in retired battery energy storage systems, including edge nodes and a cloud server, wherein the battery management system includes: The data acquisition module is used to collect operational data of retired batteries in real time. The feature processing module is used to extract features and filter correlations from the running data, and upload the filtered battery feature data to the side node. The communication module is used to upload the filtered battery feature data to the side node, so that the side node can use the battery life prediction model to infer the remaining life prediction value of the retired battery. Specifically, the battery life prediction model uploads its parameters through the side nodes, the cloud server aggregates and optimizes the parameters uploaded by each side node and distributes updated parameters, and the side nodes receive the updated parameters and update them.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The method for predicting the remaining lifespan of decommissioned batteries provided by this invention reduces the amount of data transmission between the end-side device and the edge-side device by performing feature extraction and correlation filtering on the operational data before uploading it. Operational data typically has a high sampling frequency and large volume, and direct uploading would consume significant network bandwidth. The end-side device converts the operational data into feature vectors and removes redundant features through correlation filtering, significantly reducing the amount of data uploaded and alleviating the communication burden. Edge-side nodes perform inference based on locally deployed battery lifespan prediction models, obtaining the remaining lifespan prediction value without waiting for a response from the cloud server. This local inference method avoids the impact of network transmission delays on real-time performance, improving the real-time accuracy of remaining lifespan prediction. The battery lifespan prediction model is continuously updated through collaborative operation between edge-side nodes and the cloud server. Edge-side nodes use local data for training and upload parameters, while the cloud server aggregates and optimizes the parameters from each site before issuing updates. This mechanism allows models from various energy storage stations to share global knowledge, while the original data does not need to leave the local machine, improving the model's generalization ability and prediction accuracy while ensuring data security. As the number of connected energy storage stations increases, the global model is further optimized, improving the generalization ability and prediction accuracy of the battery life prediction model, thereby enhancing the accuracy and real-time performance of predicting the remaining life of retired batteries. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic flowchart of an embodiment of the method for predicting the remaining life of retired batteries provided by the present invention; Figure 2 A schematic diagram of an embodiment of the retired battery remaining life prediction system provided by the present invention; Figure 3 This is a schematic diagram of an embodiment of the battery management system provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0020] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] This invention provides a method, system, and battery management system for predicting the remaining lifespan of retired batteries, which will be described below.

[0023] Figure 1 This is a flowchart illustrating an embodiment of the method for predicting the remaining life of decommissioned batteries provided by the present invention. The implementing entity of this embodiment is a decommissioned battery energy storage system, including end-side devices deployed in the battery management system, edge-side nodes deployed in each energy storage station, and a cloud server. The method for predicting the remaining life of decommissioned batteries includes: S101. The end-side device performs feature extraction and correlation filtering on the operating data of the retired battery, and uploads the filtered battery feature data to the side-side node. S102. The side node, based on the filtered battery feature data, uses a battery life prediction model to infer the remaining life prediction value of the retired battery; wherein, the battery life prediction model is updated by the parameters uploaded by the side node, the cloud server aggregates and optimizes the parameters uploaded by each side node and sends out updated parameters, and the side node receives the updated parameters and updates them.

[0024] The retired battery energy storage system includes N (N>1) energy storage stations, each energy storage station is equipped with a side node, and the battery clusters in each energy storage station are connected to the corresponding BMS.

[0025] The edge devices are deployed within the Battery Management System (BMS). The edge devices and edge nodes are connected via the internal network of the energy storage station, while the edge nodes and the cloud server are connected via a wide area network. The edge devices upload feature data, and the edge nodes upload model parameters and receive updated parameters from the cloud server, forming a closed data flow loop.

[0026] The end-side equipment is deployed within the BMS, which is connected to the retired battery to collect real-time operational data during the charging and discharging process. This operational data includes parameters such as voltage, current, temperature, and state of charge.

[0027] A battery life prediction model is a pre-trained neural network model that can estimate the remaining lifespan of a battery based on input feature data. The parameters of a battery life prediction model refer to the weights and biases that the model learns through training. These parameters determine how the model maps the input feature data to the output remaining lifespan prediction value, and thus determine the accuracy of the model's prediction.

[0028] Specifically, the edge device extracts key features characterizing the battery's health status from operational data, namely battery feature data, such as voltage change rate, temperature gradient, and internal resistance change rate. To reduce the amount of battery feature data transmitted and improve transmission efficiency, the edge device performs correlation screening on the battery feature data, retaining features highly correlated with the battery's remaining life and removing redundant or low-correlation features. The filtered battery feature data is uploaded to the edge node. In this way, the edge device only uploads highly correlated feature data to the edge node, thereby reducing the amount of data transmitted and improving the real-time performance of remaining life prediction.

[0029] Edge nodes are deployed at each energy storage station, and each station can be configured with one or more edge nodes. After receiving feature data uploaded by end-side devices within its station, the edge node inputs it into a locally stored battery life prediction model. This model, pre-trained, can calculate the remaining lifespan of the battery based on the input feature data. The edge node uses the model output as the predicted remaining lifespan of the retired battery.

[0030] Each edge node at an energy storage station trains its local model using historical operational data from retired batteries stored on the node, obtaining updated parameters. These edge nodes then upload these parameters to a cloud server. The cloud server receives the parameters from each edge node, aggregates and optimizes them, generating a globally optimized set of parameters. The cloud server then distributes the updated parameters to each edge node. Each edge node receives the parameters from the cloud server and replaces the original parameters in its local model with the updated parameters, thus completing the model update. Because the models at different energy storage stations can share global knowledge, the generalization ability and prediction accuracy of the battery life prediction model are improved.

[0031] In summary, the remaining life prediction method for decommissioned batteries provided in this invention reduces the amount of data transmission between the end-side device and the edge-side device by extracting features and filtering correlations from the operational data before uploading it. Operational data typically has a high sampling frequency and large volume, and direct uploading would consume significant network bandwidth. The end-side device converts the operational data into feature vectors and removes redundant features through correlation filtering, significantly reducing the amount of data uploaded and alleviating the communication burden. Edge-side nodes perform inference based on locally deployed battery life prediction models, obtaining the remaining life prediction value without waiting for a response from the cloud server. This local inference method avoids the impact of network transmission delays on real-time performance, improving the real-time accuracy of remaining life prediction. The battery life prediction model is continuously updated through collaborative operations between edge-side nodes and the cloud server. Edge-side nodes use local data for training and upload parameters, while the cloud server aggregates and optimizes the parameters from each site before issuing updates. This mechanism allows models at various energy storage stations to share global knowledge, while the original data does not need to leave the local storage location, improving the model's generalization ability and prediction accuracy while ensuring data security. As the number of connected energy storage stations increases, the global model is continuously optimized, improving the generalization ability and prediction accuracy of the battery life prediction model, thereby enhancing the accuracy and real-time performance of battery remaining life prediction.

[0032] In some embodiments of the present invention, the operating data includes voltage data, state of charge, temperature data, and internal resistance data, and the battery characteristic data includes voltage characteristics, temperature characteristics, and internal resistance characteristics; step S101 includes: determining the voltage characteristic by differentiating the voltage data with respect to the state of charge; determining the temperature characteristic by determining the standard deviation of the temperature data; and determining the internal resistance characteristic by determining the ratio of the change in internal resistance data to the change in the number of cycles.

[0033] In some embodiments of the present invention, in step S101, the end-side device extracts features from the operating data of the retired battery. The operating data includes voltage data, state of charge, temperature data, and internal resistance data. The extracted battery feature data includes voltage features, temperature features, and internal resistance features.

[0034] Among them, voltage characteristic refers to the rate of change of battery terminal voltage with respect to state of charge, denoted as The voltage characteristic is calculated by numerically differentiating the voltage data V with respect to the state of charge (SOC), i.e. This characteristic reflects the degree of battery polarization and the degradation of usable capacity as the battery ages. The peak of the curve will shift.

[0035] Temperature characteristics refer to the standard deviation of temperature at each sampling point within the battery pack, denoted as . The temperature characteristics are calculated as follows: ,in For the first i Each temperature sample value, The average temperature is denoted by N, and N is the number of sampling points. This feature reflects the consistency of the internal temperature distribution of the battery pack; abnormal temperature rise indicates increased internal resistance or localized aging.

[0036] Internal resistance characteristic refers to the rate of change of the battery's internal resistance with the number of cycles, denoted as . The internal resistance characteristic is calculated as follows: Where ΔR is the change in internal resistance, This represents the change in the number of battery cycles. This characteristic is directly related to the wear and tear of electrode materials and the degree of electrolyte decomposition, and is a key indicator for assessing the health of the battery.

[0037] Specifically, the end-side equipment collects voltage data (V), state of charge (SOC), and temperature data at each sampling point from retired batteries during the charging and discharging process. And the internal resistance data R. Within the preset calculation period, the terminal device calculates the voltage characteristics. The numerical derivative value. For temperature characteristics, the end-side device calculates the standard deviation of each temperature sampling point. For internal resistance characteristics, the end-side device records the change in internal resistance with the number of cycles and calculates... The ratio. The end-side device will calculate the ratio. , , This is used as the extracted battery feature data.

[0038] Understandably, this embodiment significantly reduces the amount of data transmitted from the terminal side to the edge side by extracting three lightweight features: voltage features, temperature features, and internal resistance features. Compared to directly uploading raw operational data, the amount of feature data transmitted can be reduced by more than an order of magnitude. Simultaneously, these three types of features describe the battery's aging state from three dimensions: electrochemical polarization, thermal characteristics, and internal resistance decay, respectively. These features complement each other, providing a more comprehensive reflection of the health status of retired batteries and offering effective information input for subsequent remaining lifespan prediction.

[0039] In some embodiments of the present invention, the correlation screening step includes: calculating the mutual information value between each of the battery feature data and the remaining life tag according to the joint probability distribution of the battery feature data and the remaining life tag and their respective marginal probability distributions; and performing correlation screening on the battery feature data based on the mutual information value.

[0040] Mutual information is a measure of the degree of interdependence between two random variables; the larger the value, the stronger the correlation between the two variables.

[0041] Battery characteristic data refers to multimodal feature vectors such as voltage characteristics, temperature characteristics, and internal resistance characteristics extracted by end-side devices from raw operating data.

[0042] The remaining life tag refers to the number of cycles or usage time a battery has remaining from the current moment until the end of its lifespan. It can be derived from the battery's full life cycle record in historical data.

[0043] Specifically, the edge device calculates the mutual information value between each battery feature data and the remaining life tag based on the joint probability distribution of the battery feature data and the remaining life tag, as well as their respective marginal probability distributions. The formula for calculating the mutual information value is as follows: ; Where X represents a type of battery characteristic data, such as voltage or temperature characteristics, and RUL represents the remaining life tag. This represents the summation of all possible values ​​of the battery characteristic data X (voltage, temperature, internal resistance characteristics, etc.); This represents summing over all possible values ​​of RUL; The feature data X takes the value of x And RUL takes the value r The joint probability density, p ( x The battery characteristic data X takes the value of ) x The marginal probability density of X itself, p ( r ) is the value of RUL r The edge probability density of the RUL itself. The end-side device double-sums all possible values ​​of the battery feature data and the remaining life tag to obtain the mutual information value between the battery feature data and the remaining life tag.

[0044] The edge device performs correlation filtering on battery feature data based on the calculated mutual information value, retaining features with larger mutual information values ​​and removing features with smaller mutual information values. For example, features with mutual information values ​​greater than a preset threshold can be retained, while features with mutual information values ​​less than or equal to the preset threshold can be removed. The preset threshold is 1, meaning features with mutual information values ​​greater than 1 are retained, i.e., features positively correlated with remaining lifespan are retained. The filtering formula is as follows: ; Understandably, this embodiment evaluates the importance of each feature by calculating the mutual information value between feature data and remaining lifetime labels, and selects highly correlated features accordingly. This allows the edge device to upload only feature data that contributes significantly to remaining lifetime prediction, reducing the amount of data transmission between the edge and the side, reducing communication bandwidth usage, and avoiding interference from low-correlation features on the model inference accuracy, thereby improving the accuracy of remaining lifetime prediction by subsequent edge nodes.

[0045] In some embodiments of the present invention, before the side node uploads the parameters of the battery life prediction model, the method further includes: encrypting the parameters using a homomorphic encryption algorithm.

[0046] Homomorphic encryption is a type of encryption that allows direct mathematical operations on encrypted data. Homomorphic encryption algorithms include, but are not limited to: Paillier semi-homomorphic encryption, which supports addition operations on encrypted data; and fully homomorphic encryption, which supports both addition and multiplication operations on encrypted data. This embodiment preferably uses the Paillier semi-homomorphic encryption algorithm to achieve a balance between security and computational efficiency.

[0047] Specifically, after completing model training locally, each edge node obtains updated parameters. The edge node then encrypts the updated parameters using the Paillier semi-homomorphic encryption algorithm, generating ciphertext. Instead of directly uploading plaintext parameters, the edge node uploads the encrypted ciphertext to the cloud server. Upon receiving the ciphertext from each edge node, the cloud server performs aggregation operations on the ciphertext without decryption, obtaining aggregated ciphertext. The cloud server then distributes the aggregated ciphertext to each edge node. Upon receiving the aggregated ciphertext, each edge node decrypts it using its locally stored private key to obtain the aggregated parameters and uses them to update its local model.

[0048] Understandably, this embodiment employs a homomorphic encryption algorithm to encrypt the model parameters, ensuring that the model parameters exist in ciphertext form during transmission, thereby enhancing data transmission security. Simultaneously, the cloud server can directly perform aggregation calculations on the encrypted parameters without decryption, preventing the cloud server from accessing the parameter information of individual edge nodes and protecting the data privacy of each energy storage station. Furthermore, an end-to-end encrypted transmission link is formed between the edge nodes and the cloud server, ensuring that the original model parameters always flow in ciphertext form and are only decrypted locally, complying with data security regulations for industrial data protection.

[0049] In some embodiments of the present invention, the cloud server performs aggregation optimization on the parameters uploaded by each side node, including: aggregating the parameters uploaded by each side node using a weighted average algorithm to generate aggregated parameters; and iteratively optimizing the aggregated parameters using a multimodal pre-trained model to generate updated parameters.

[0050] In some embodiments of the present invention, a weighted average algorithm is used to aggregate the parameters uploaded by each side node to generate aggregated parameters, including: determining the weight of each side node based on the proportion of the historical data sample size of retired batteries stored by each side node to the total sample size; and summing the products of the parameters uploaded by each side node and their respective weights to obtain the aggregated parameters.

[0051] Among them, aggregation optimization refers to the process by which the cloud server merges and further optimizes the parameters uploaded by each side node.

[0052] The weighted average algorithm is an implementation of the aggregation operation, which performs weighted summation by assigning different weights to the parameters of different edge nodes.

[0053] A multimodal pre-trained model refers to a deep learning model that has been pre-trained on large-scale, multi-source data. For example, a multimodal pre-trained model uses the Transformer architecture. Compared to recurrent neural networks, it can better capture long-range dependencies in sequential data. This model can handle various types of input data, such as voltage, temperature, and internal resistance.

[0054] Specifically, the cloud server aggregates and optimizes the parameters uploaded by each side node. It can first use a weighted average algorithm to aggregate the parameters uploaded by each side node to generate aggregated parameters. Then, it uses a multimodal pre-trained model to iteratively optimize the aggregated parameters to generate updated parameters.

[0055] Specifically, the cloud server first obtains the sample size of historical retired battery data stored in each edge node, and determines the weight of each node based on the proportion of its sample size to the total sample size. Then, the parameters uploaded by each edge node are multiplied by their respective weights, and all products are summed to obtain the aggregated parameters. This aggregation method ensures that edge nodes with larger data volumes have a higher contribution to the global parameters.

[0056] Furthermore, the cloud server inputs the aggregated parameters into the multimodal pre-trained model for iterative optimization. This involves using global data to perform several rounds of iterative training on the pre-trained model, adjusting the model parameters through the backpropagation algorithm to gradually reduce the model's prediction error on the global data. After iterative optimization, updated parameters are generated and distributed to each edge node.

[0057] Understandably, this embodiment achieves effective fusion of global model parameters and performance improvement by combining weighted average aggregation with multimodal pre-training optimization. Weighted average aggregation is simple and efficient, and the larger the sample size of a site, the greater its impact on global parameters, consistent with the objective laws of data-driven approaches. The multimodal pre-trained model can fully utilize the diversity of global data in the cloud, improving the model's generalization ability to different operating conditions and battery types. Compared to models trained independently on each side node, the global model after aggregation optimization exhibits lower prediction errors and stronger robustness.

[0058] In some embodiments of the present invention, before the updated parameters are sent out by the cloud server, the updated parameters are further compressed using low-rank decomposition.

[0059] Low-rank decomposition refers to a mathematical method that decomposes a high-dimensional matrix into a product of multiple low-dimensional matrices. Low-rank decomposition includes, but is not limited to, singular value decomposition, nonnegative matrix decomposition, and principal component analysis.

[0060] Specifically, after the cloud server aggregates and optimizes the parameters uploaded by each edge node, it generates updated parameters to be distributed. This parameter matrix is ​​usually high-dimensional, and direct distribution would consume a lot of network bandwidth. The cloud server uses low-rank decomposition technology to compress the updated parameters, decomposing the original high-dimensional parameter matrix into several low-dimensional matrices, and only distributing the low-dimensional matrices instead of the complete parameter matrix.

[0061] As a preferred approach, the cloud server uses singular value decomposition to compress the update amount of model parameters. Let the global model parameters after aggregation in round t be... The global model parameters after aggregation in round t+1 are: Then the update amount of the model parameters is The cloud server performs singular value decomposition on the update, obtaining... U and Let Σ be an orthogonal matrix, and let Σ be a diagonal matrix. The elements on the diagonal are singular values. The magnitude of the singular values ​​reflects the importance of the corresponding decomposition matrix; the larger the singular value, the greater the contribution of the corresponding component to the original matrix.

[0062] The cloud server retains only the top few decomposition matrices with the largest singular values, discarding those with smaller singular values, thus compressing the parameter update size. For example, retaining only the decomposition matrices corresponding to the top 10% of singular values ​​with the highest contribution percentage can compress the parameter update size to less than 10% of the original size. The compressed update package typically does not exceed 10MB. The cloud server distributes this compressed update package to each edge node. Upon receiving the compressed package, the edge nodes reconstruct an approximate parameter update size through inverse operations and add it to their local model parameters to complete the model update.

[0063] Understandably, this embodiment compresses the model parameters through low-rank decomposition before distribution, reducing the amount of data transmitted over the network between the cloud server and edge nodes, lowering communication bandwidth usage, and enabling high-frequency iterative updates of the model. Simultaneously, the compressed parameters retain the key information of the original parameters, ensuring that the prediction accuracy after model updates is not significantly affected.

[0064] In some embodiments of the present invention, the battery life prediction model is an LSTM network, the network structure of which includes an input layer, an LSTM layer, and a fully connected layer; the number of nodes in the input layer is the dimension of the battery feature data; the LSTM layer is used to extract the time dependency of the filtered battery feature data; and the fully connected layer is used to output the remaining life prediction value.

[0065] The number of nodes in the input layer is set to the dimension of the battery feature data. Since the battery feature data in this embodiment includes three dimensions: voltage feature, temperature feature, and internal resistance feature, the number of nodes in the input layer is set to 3.

[0066] The LSTM layer is used to extract the time-dimensional dependencies of the filtered battery feature data. In this embodiment, the number of nodes in the LSTM layer is set to 16.

[0067] The fully connected layer is used to map the features extracted by the LSTM layer to the final remaining lifetime prediction, and its number of nodes is set to 1.

[0068] The edge nodes input the battery feature data received from the end side into the LSTM network in chronological order. After forward computation, the network outputs a value, which is the predicted remaining life of the retired battery.

[0069] Understandably, this embodiment fully leverages the modeling capabilities of LSTM networks for time-series data by employing an LSTM network as the prediction model for edge nodes. Battery degradation is essentially a time-series process, with the current battery state closely related to its historical states. Through its gating mechanism, the LSTM network can effectively capture long-term dependencies, avoiding gradient vanishing or exploding problems, thereby improving the accuracy of remaining lifetime prediction. Simultaneously, the LSTM network has a moderate number of parameters, allowing deployment on edge nodes with limited computing power, ensuring real-time inference.

[0070] In one specific implementation, the edge nodes employ a federated transfer learning mechanism, as follows: Local model deployment: Deploy a lightweight LSTM model on the edge nodes. Network structure: Input layer (3 nodes) → LSTM layer (16 nodes) → Fully connected layer (1 node). The federal training process is as follows: (1) Local training: Each edge node is trained for 10 epochs based on local data with a learning rate of 0.001; (2) Parameter encryption: Paillier semi-homomorphic encryption is used to encrypt the model gradient; (3) Secure aggregation: The cloud server performs a weighted average of all edge parameters and updates the global model; (4) Model distribution: The updated global model is distributed from the cloud to each edge node for local fine-tuning.

[0071] Federated learning aggregation algorithm (FedAvg) calculation formula: ; In the formula, K represents the number of edge nodes. For the first Round global model parameters; Let be the sample size of the k-th node. This represents the total sample size. For the parameters updated locally at the k-th node, For learning rate, Let be the local loss function for the k-th node.

[0072] The cloud-based multimodal pre-training and incremental updates are as follows: (1) Global pre-trained model: The Transformer pre-trained model is deployed in the cloud, and the input is a multimodal feature sequence; the model training uses public datasets or self-test datasets, and the training sample size is ≥1000 batteries' full life cycle data.

[0073] (2) Incremental model update: Low-rank decomposition (SVD) is used to compress model parameters. Only the decomposition matrix corresponding to the top 10% of singular values ​​is distributed, and the update data volume is ≤10MB; after the edge node loads the update package, the model update can be completed by fine-tuning 5 epochs based on local data.

[0074] Low-rank decomposition is used to implement incremental model updates, only distributing parameter changes: ; In the formula, ,Σ, As a low-rank decomposition matrix, the parameter update amount can be compressed to less than 10% of the original size, and only edge nodes need to be loaded. The model update can then be completed.

[0075] To better implement the method for predicting the remaining life of decommissioned batteries in this embodiment of the invention, this embodiment also provides a system for predicting the remaining life of decommissioned batteries, applied to decommissioned battery energy storage systems, such as... Figure 2 The diagram shown is a structural schematic of a retired battery remaining life prediction system, including: The edge module 100 is deployed in the battery management system and is used to extract features and filter correlations from the operating data of retired batteries. The filtered battery feature data is then uploaded to the edge module 200. Side-side module 200, deployed in each energy storage station, is used to infer the remaining life prediction value of the retired battery based on the battery characteristic data using the battery life prediction model; the side-side module is also used to upload the parameters of the battery life prediction model, and to receive updated parameters and update them. The cloud module 300, deployed on a cloud server, is used to aggregate and optimize the parameters uploaded by each side module 200 and then distribute the updated parameters.

[0076] Specifically, the retired battery energy storage system adopts a three-level architecture of end-edge-cloud, including end-side module 100, edge-side module 200 and cloud module 300.

[0077] The edge module 100 is deployed within the battery management system. The edge module 100 is used to extract features and filter correlations from the operational data of retired batteries, and then uploads the filtered battery feature data to the edge module 200. Each battery management system can be configured with one edge module 100, and each edge module 100 is responsible for managing a group of retired batteries connected to it.

[0078] Side-side modules 200 are deployed in various energy storage stations. Each energy storage station can be configured with one or more side-side modules 200. The side-side modules 200 are used to infer the remaining lifespan of retired batteries based on received battery characteristic data using a battery lifespan prediction model. The side-side modules 200 are also used to upload the parameters of the battery lifespan prediction model to the cloud module 300, and to receive updated parameters from the cloud module 300 and update the local model accordingly.

[0079] The cloud module 300 is deployed on a cloud server. The cloud module 300 is used to aggregate and optimize the parameters uploaded by each side module 200, generate updated parameters, and distribute the updated parameters to each side module 200.

[0080] In this embodiment, the decommissioned battery energy storage system utilizes an end-side module deployed within the battery management system to perform feature extraction and correlation filtering. The raw operational data is converted into compact feature vectors at the end-side, with only the filtered feature data needing to be uploaded to the edge-side module. This significantly reduces the amount of data transmission between the end-side and edge-side modules, alleviating the communication burden on the internal network of the energy storage station. Furthermore, the end-side module has a relatively low computational load, with processing latency typically within 10 milliseconds.

[0081] By deploying edge modules at each energy storage station and using local models for inference, remaining lifetime prediction can be completed at the edge without waiting for a response from the cloud module. This avoids the impact of network transmission latency on the real-time performance of the prediction. The inference latency of the edge modules is typically within 100 milliseconds, which can meet the real-time control requirements of the energy storage system.

[0082] By configuring the edge module to upload model parameters and the cloud module to aggregate, optimize, and update parameters for each site, the models of each energy storage station can share global knowledge without sharing raw data. This solves the data silo problem, improves the model's generalization ability and prediction accuracy across different sites, and ensures the security and compliance of the raw data. As the number of connected energy storage stations increases, the data aggregated by the cloud module becomes richer, and the model prediction error can be gradually reduced from the initial 6% to less than 3%.

[0083] The present invention also provides a battery management system for use in retired battery energy storage systems, including edge nodes and a cloud server, such as... Figure 3 The diagram shown is a structural schematic of a battery management system, which includes: The data acquisition module 310 is used to collect the operating data of retired batteries in real time; Feature processing module 320 is used to extract features and filter correlations from the running data, and upload the filtered battery feature data to the side node. The communication module 330 is used to upload the filtered battery feature data to the side node, so that the side node can use the battery life prediction model to infer the remaining life prediction value of the retired battery; wherein, the battery life prediction model uploads the parameters of the battery life prediction model through the side node, the cloud server aggregates and optimizes the parameters uploaded by each side node and sends out updated parameters, and the side node receives the updated parameters and updates itself.

[0084] Specifically, the data acquisition module 310 is used to collect the operating data of the retired battery in real time. The data acquisition module 310 is directly connected to the retired battery and acquires the original operating parameters of the battery such as voltage, current, temperature, and state of charge according to the preset sampling frequency.

[0085] The feature processing module 320 is connected to the data acquisition module 310 and is used to extract features and filter correlations from the acquired running data, and output the filtered battery feature data to the communication module 330.

[0086] The communication module 330 is connected to the feature processing module 320 and is used to upload the filtered battery feature data to the side node. After receiving the feature data, the side node uses the battery life prediction model to perform inference and obtain the predicted remaining life value of the retired battery.

[0087] It should be noted that the battery life prediction model involves the side nodes uploading model parameters, the cloud server aggregating and optimizing the parameters uploaded by each side node and distributing the updated parameters, and the side nodes receiving the updated parameters and updating them.

[0088] By incorporating a feature processing module within the battery management system, raw operational data is transformed into compact feature vectors and subjected to correlation filtering at the edge. Only the filtered feature data needs to be uploaded to the edge nodes, effectively reducing the data transmission volume between the battery management system and the edge nodes and alleviating the communication burden on the internal network of the energy storage station. Furthermore, the computational load for feature extraction and correlation filtering is relatively small, enabling efficient operation on the limited computing resources of the battery management system.

[0089] By configuring a communication module to upload the filtered feature data to the edge nodes, the edge nodes can then use the battery life prediction model for inference, eliminating the need for the battery management system to perform the computational tasks of model inference. This approach allows the battery management system to maintain a lightweight design, eliminating the need to deploy complex deep learning models and reducing hardware costs and power consumption requirements.

[0090] The battery life prediction model is continuously updated through collaborative operation between edge nodes and the cloud server, while the battery management system (BMS) is only responsible for providing filtered feature data and does not need to participate in the model training and update process. This approach simplifies and standardizes the BMS function, making it easy to replicate and deploy across different energy storage stations and reducing system maintenance complexity.

[0091] Accordingly, this application also provides a computer-readable storage medium for storing a computer-readable program or instruction. When the program or instruction is executed by a processor, it can implement the steps or functions of the method for predicting the remaining life of retired batteries provided in the above-described method embodiments.

[0092] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0093] The foregoing has provided a detailed description of the method, system, and battery management system for predicting the remaining life of retired batteries provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for predicting the remaining life of a retired battery, characterized in that, The method, applicable to retired battery energy storage systems, includes end-side devices deployed in the battery management system, edge-side nodes deployed in each energy storage station, and a cloud server. The method comprises: The edge device performs feature extraction and correlation filtering on the operating data of retired batteries, and uploads the filtered battery feature data to the edge node. The side node uses the filtered battery feature data and a battery life prediction model to infer the remaining life prediction value of the retired battery. Specifically, the battery life prediction model uploads its parameters through the side nodes, the cloud server aggregates and optimizes the parameters uploaded by each side node and distributes updated parameters, and the side nodes receive the updated parameters and update them.

2. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, The operating data includes voltage data, state of charge data, temperature data, and internal resistance data, and the battery characteristic data includes voltage characteristics, temperature characteristics, and internal resistance characteristics. The end-side device performs feature extraction on the operational data of the retired batteries, including: The voltage characteristic is determined by differentiating the voltage data with respect to the state of charge. The standard deviation of the temperature data is defined as the temperature characteristic. The ratio of the change in internal resistance data to the change in the number of cycles is defined as the internal resistance characteristic.

3. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, The relevance screening steps include: Based on the joint probability distribution of the battery feature data and the remaining life tag of the retired battery, as well as their respective marginal probability distributions, the mutual information value between each of the battery feature data and the remaining life tag is calculated. The battery feature data is filtered for correlation based on the mutual information value.

4. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, Before the side node uploads the parameters of the battery life prediction model, the following is also included: The parameters are encrypted using a homomorphic encryption algorithm.

5. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, The cloud server aggregates and optimizes the parameters uploaded by each side node, including: A weighted average algorithm is used to aggregate the parameters uploaded by each edge node to generate aggregated parameters; The aggregated parameters are iteratively optimized using a multimodal pre-trained model to generate updated parameters.

6. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, A weighted average algorithm is used to aggregate the parameters uploaded by each edge node, generating aggregated parameters, including: The weight of each side node is determined based on the proportion of historical data samples of retired batteries stored in each side node to the total sample size. The aggregated parameters are obtained by summing the products of the parameters uploaded by each side node and their respective weights.

7. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, Before the updated parameters are sent out by the cloud server, the following is also included: The updated parameters are compressed using low-rank decomposition.

8. The method for predicting the remaining lifespan of retired batteries according to claim 1, characterized in that, The battery life prediction model is an LSTM network, and the network structure of the LSTM network includes an input layer, an LSTM layer, and a fully connected layer. The number of nodes in the input layer is equal to the dimension of the battery feature data; The LSTM layer is used to extract the time dependency of the filtered battery feature data; The fully connected layer is used to output the remaining lifetime prediction value.

9. A system for predicting the remaining lifespan of retired batteries, characterized in that, Applications in retired battery energy storage systems include: The edge module, deployed in the battery management system, is used to extract features and filter correlations from the operating data of retired batteries, and upload the filtered battery feature data to the edge module. The edge module, deployed at each energy storage station, is used to infer the remaining life prediction value of the retired battery based on the battery characteristic data using the battery life prediction model; the edge module is also used to upload the parameters of the battery life prediction model, and to receive and update the updated parameters. The cloud module, deployed on a cloud server, is used to aggregate and optimize the parameters uploaded by each side module and then distribute the updated parameters.

10. A battery management system, characterized in that, The battery management system, applicable to retired battery energy storage systems, includes edge nodes and a cloud server, and comprises: The data acquisition module is used to collect operational data of retired batteries in real time. The feature processing module is used to extract features and filter correlations from the running data, and upload the filtered battery feature data to the side node. The communication module is used to upload the filtered battery feature data to the side node, so that the side node can use the battery life prediction model to infer the remaining life prediction value of the retired battery. Specifically, the battery life prediction model uploads its parameters through the side nodes, the cloud server aggregates and optimizes the parameters uploaded by each side node and distributes updated parameters, and the side nodes receive the updated parameters and update them.

Images