Battery charge state estimation method, device and equipment based on end-cloud collaboration

Abstract

The embodiment of the invention provides a battery state-of-charge estimation method, device and equipment based on end-cloud collaboration, and relates to the technical field related to batteries. The method comprises the following steps: acquiring a first state-of-charge estimation value of a single battery in an electrochemical energy storage battery cabin from a cloud end; the first state-of-charge estimation value is obtained by the cloud through prediction based on real-time operation data of a single battery based on a CNN-LSTM hybrid model fused with negative resistance physical characteristics; obtaining a second state of charge estimation value of the single battery from the terminal; the second state-of-charge estimation value is obtained by the terminal through state estimation based on real-time operation data through a target equivalent circuit model corresponding to the current working condition of the single battery and a preset state estimation algorithm; and through an adaptive Kalman filtering algorithm, carrying out fusion processing on the first state-of-charge estimation value and the second state-of-charge estimation value to obtain a target state-of-charge estimation value of the single battery. The method is used for improving the estimation precision and efficiency of the state of charge.

Description

Technical Field

[0001] This application relates to the field of battery-related technologies, and in particular to a method, apparatus and equipment for estimating the state of charge of a battery based on edge-cloud collaboration. Background Technology

[0002] With the development of new energy technologies, energy storage systems are playing an increasingly important role in energy management. In energy storage systems, the state of charge (SOC) of a battery is a key parameter for measuring its remaining energy, directly affecting the battery's charge and discharge control, safety protection, and lifespan management.

[0003] However, the state of charge (SOC) of a battery has a nonlinear relationship with external parameters such as voltage and current, and is affected by complex factors such as temperature fluctuations, battery aging, and changes in charge and discharge rates, making it difficult for traditional estimation methods to balance accuracy and real-time performance. Summary of the Invention

[0004] This application provides a battery state of charge estimation method, apparatus, and device based on edge-cloud collaboration, which aims to improve the accuracy and efficiency of state of charge estimation and achieve a balance between accuracy, real-time performance, and robustness in SOC estimation.

[0005] In a first aspect, embodiments of this application provide a battery state-of-charge estimation method based on edge-cloud collaboration, the method comprising:

[0006] The first state of charge (SOC) estimate of a single cell in the electrochemical energy storage battery compartment is obtained from the cloud. The first SOC estimate is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM) that incorporates the physical characteristics of negative resistance.

[0007] The second state of charge (SOC) estimate of the individual battery cell is obtained from the terminal. The second SOC estimate is obtained by the terminal through state estimation based on the real-time operating data using a target equivalent circuit model corresponding to the current operating condition of the individual battery cell and a preset state estimation algorithm. The target equivalent circuit model can be any one of a zero-order equivalent circuit model, a first-order equivalent circuit model, or a second-order equivalent circuit model.

[0008] The first state of charge estimate and the second state of charge estimate are fused using an adaptive Kalman filter algorithm to obtain the target state of charge estimate of the single cell.

[0009] In one possible implementation, the step of fusing the first state-of-charge estimate and the second state-of-charge estimate using an adaptive Kalman filter algorithm to obtain the target state-of-charge estimate of the single cell includes:

[0010] The first state of charge estimate is used as the observation prediction value, and the second state of charge estimate is used as the state observation value.

[0011] Based on the confidence interval attached to the first state of charge estimate, the observation noise covariance is determined;

[0012] Based on the innovative sequence of the state prediction values, the process noise covariance is adaptively adjusted;

[0013] Based on the observed noise covariance and the adjusted process noise covariance, the Kalman gain is calculated to perform weighted fusion of the observed predicted value and the state observed value to obtain the target state of charge estimate.

[0014] In one possible implementation, the CNN-LSTM hybrid model is trained using a staged training strategy:

[0015] First, freeze the LSTM layer and train the CNN layer separately. Use the training data and attach a temporary output layer for pre-training until the error of the state of charge predicted by the CNN layer alone is lower than the first threshold.

[0016] Unfreeze the LSTM layer, input the feature vector extracted by the CNN layer into the LSTM layer for end-to-end joint training until the preset termination condition is met, and then end the training to obtain the CNN-LSTM hybrid model.

[0017] In the joint training, an attention mechanism layer is employed, which is configured to dynamically adjust the attention weights on the negative resistance physical features based on the operating condition features in the input data.

[0018] In one possible implementation, the terminal calculates the second state of charge estimate in the following manner:

[0019] Based on the real-time operating data of the individual battery, the current operating condition of the individual battery is determined, and based on the current operating condition, the target equivalent circuit model is determined.

[0020] Based on the real-time operating data, the model parameters of the target equivalent circuit model are identified online using the recursive least squares method.

[0021] The identified model parameters are substituted into the state equation constructed based on the ampere-hour integration method and the observation equation constructed based on the target equivalent circuit model. The state is estimated by the dual Kalman filter algorithm, and the estimated state of charge of the single cell is output.

[0022] In one possible implementation, determining the current operating condition of the individual battery cell based on its real-time operating data includes:

[0023] Based on the real-time operating data of the individual battery cells, the current standard deviation within a preset time window is calculated, and the voltage-current dynamic correlation coefficient within the same time window is calculated.

[0024] If the current standard deviation is less than the first preset threshold, then the current operating condition of the single battery cell is determined to be the standby operating condition.

[0025] If it is determined that the current standard deviation is greater than the second preset threshold and the voltage-current dynamic correlation coefficient indicates a negative correlation, then the current operating condition of the single cell is determined to be the inertia support operating condition; wherein, the second preset threshold is greater than the first preset threshold;

[0026] Otherwise, the current operating condition of the single battery cell is determined to be the charging condition.

[0027] In one possible implementation, determining the target equivalent circuit model based on the current operating conditions includes:

[0028] If the current operating condition is a standby operating condition, then the zero-order equivalent circuit model is selected as the target equivalent circuit model;

[0029] If the current operating condition is a charging condition, then the first-order equivalent circuit model is selected as the target equivalent circuit model.

[0030] If the current operating condition is an inertia-supported operating condition, then the second-order equivalent circuit model is selected as the target equivalent circuit model.

[0031] In one possible implementation, before obtaining the first state-of-charge estimate of a single cell in the electrochemical energy storage battery compartment from the cloud, the method further includes:

[0032] Query the prior value of negative resistance corresponding to the current state of charge, temperature and current operating condition of the single cell from the three-dimensional feature library of negative resistance;

[0033] Based on the real-time voltage and current obtained from the terminal, the prior value of the negative resistance is fine-tuned online using the recursive least squares method to obtain the corrected negative resistance value.

[0034] The corrected negative resistance value is provided to the cloud and / or the terminal for calculating the estimated state of charge.

[0035] In one possible implementation, the electrochemical energy storage battery compartment includes multiple battery clusters connected in parallel, each battery cluster including multiple individual cells connected in series; the method further includes:

[0036] When it is determined that the electrochemical energy storage battery compartment is in a charging state, the maximum value among the target state of charge estimates of each individual battery cell is determined as the state of charge estimate of the battery cluster.

[0037] When it is determined that the electrochemical energy storage battery compartment is in a discharged state, the minimum value among the target state of charge estimates of each individual battery cell is determined as the state of charge estimate of the battery cluster.

[0038] Secondly, embodiments of this application provide a battery state-of-charge estimation device based on edge-cloud collaboration, the device comprising:

[0039] The first processing unit is used to obtain the first state of charge estimate of a single cell in the electrochemical energy storage battery compartment from the cloud; wherein, the first state of charge estimate is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of convolutional neural network CNN-long short-term memory network LSTM that incorporates the physical characteristics of negative resistance.

[0040] The second processing unit is used to obtain a second state of charge estimation value of the single battery cell from the terminal; wherein the second state of charge estimation value is obtained by the terminal through state estimation based on the real-time operating data using a target equivalent circuit model corresponding to the current operating condition of the single battery cell and a preset state estimation algorithm;

[0041] The third processing unit is used to fuse the first state of charge estimate and the second state of charge estimate using a dual adaptive Kalman filter algorithm to obtain the target state of charge estimate of the single cell.

[0042] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0043] The memory stores computer-executed instructions;

[0044] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0045] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0046] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0047] The battery state-of-charge (SOC) estimation method, apparatus, and device based on edge-cloud collaboration provided in this application combine a high-precision deep learning model in the cloud with a lightweight equivalent circuit model on the terminal through an edge-cloud collaborative architecture, and fuse them using adaptive Kalman filtering, achieving a balance between accuracy, real-time performance, and robustness in SOC estimation. This method is particularly suitable for complex application scenarios such as electrochemical energy storage battery compartments, providing technical support for battery health management and safe operation. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 A flowchart illustrating a battery state-of-charge estimation method based on edge-cloud collaboration, provided for an embodiment of this application;

[0050] Figure 2 A framework diagram of a battery state-of-charge estimation method based on an adaptive extended Kalman filter algorithm provided in this application embodiment;

[0051] Figure 3 A comparative schematic diagram of fused SOC and ampere-hour integral SOC provided for embodiments of this application;

[0052] Figure 4 A schematic diagram of a battery state of charge estimation device based on edge-cloud collaboration provided in an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0054] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0055] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0056] With the advancement of new energy technologies, energy storage systems are playing an increasingly crucial role in energy management. Lithium-ion batteries, with their high energy density, long cycle life, and environmentally friendly characteristics, have become the preferred choice for energy storage systems. Accurately estimating their state of charge (SOC)—the ratio of remaining usable capacity to maximum usable capacity—is essential for avoiding overcharging and over-discharging, ensuring battery safety, and improving efficiency. Battery management systems (BMS) rely on precise SOC values ​​to achieve reasonable control over the battery.

[0057] Currently, methods for estimating the state of charge (SOC) of batteries are mainly divided into two categories: physical model-based methods and data-driven methods. Physical model-based methods simulate the electrochemical behavior of the battery by establishing an equivalent circuit model (such as an RC equivalent circuit) and then combining it with algorithms such as the Extended Kalman Filter (EKF) for SOC estimation. These methods rely on the accurate identification of battery parameters (such as ohmic internal resistance and polarization resistance), but in practical applications, battery parameters dynamically drift with aging and temperature changes, leading to model inaccuracies. Furthermore, traditional Kalman filtering is prone to error accumulation in strongly nonlinear systems and is difficult to adapt to complex operating conditions.

[0058] Data-driven methods, on the other hand, use neural networks (such as backpropagation neural networks or backpropagation neural networks optimized by genetic algorithms) to learn nonlinear mapping relationships in historical data to predict the state of charge (SOC) value. However, these methods are limited by the computing resources on the terminal side, making it difficult to process high-dimensional time-series data in real time. Furthermore, their generalization ability is insufficient in extreme conditions not covered by the samples (such as low temperatures or rapid charging and discharging). In addition, training cloud-based models relies on a large amount of high-precision experimental data; if the data distribution deviates from the actual operating conditions, it can easily lead to model overfitting.

[0059] It is evident that current methods for estimating the state of charge (SOC) of batteries suffer from several problems: insufficient environmental adaptability, limited robustness under dynamic operating conditions, reliance on high-precision data acquisition, and error accumulation. These issues prevent a balance between computational accuracy and real-time performance, severely impacting the reliability and practicality of SOC estimation in complex scenarios. Therefore, there is an urgent need for a collaborative architecture that integrates cloud-based deep learning models with real-time algorithms at the terminal end to achieve high-precision, low-latency, and robust online SOC estimation, providing reliable data support for the safe operation and intelligent management of energy storage systems.

[0060] Therefore, this application aims to address the aforementioned technical pain points through an edge-cloud collaborative architecture and a multi-method fusion strategy: First, it overcomes the accuracy bottleneck of single methods by leveraging the cloud-based CNN-LSTM deep learning model's ability to mine massive amounts of historical data, combined with the real-time fusion characteristics of edge-side dual Kalman filtering, to achieve high-precision SOC estimation under complex operating conditions, especially compensating for the estimation bias of traditional methods in electrochemical polarization and concentration polarization coupling scenarios; Second, it overcomes the defects of initial value and accumulated error by using edge-side ampere-hour integration and dynamic correction of the cloud model, combined with the error suppression mechanism of dual Kalman filtering, to eliminate errors accumulated during long-term operation. The model employs several techniques: First, it accumulates differences to address estimation drift caused by unknown or misjudged initial SOC. Second, it balances real-time performance with generalization capability by utilizing lightweight edge-side algorithms (ampere-hour integration and dual Kalman filtering) to meet millisecond-level response requirements, while the cloud-based deep model continuously trains to adapt to new scenarios such as battery aging and environmental changes, ensuring stable performance of the estimation model under different temperatures and charge / discharge rates. Third, it enhances model interpretability and engineering practicality by relying on a second-order RC equivalent circuit model to intuitively depict electrochemical polarization and concentration polarization, providing physical meaning support for the output of the data-driven model and avoiding the uncontrollable risks of "black box models" under abnormal operating conditions.

[0061] Based on this, this application provides an edge-cloud collaborative method for estimating the state of charge (SOC) of a battery. Through an edge-cloud collaborative architecture and a multi-method fusion strategy, it achieves high-precision, real-time, and robust SOC estimation. Specifically, the cloud is responsible for high-precision model training and parameter optimization, while the terminal executes a lightweight algorithm to ensure real-time response. Through dynamic correction and bidirectional feedback, a triple-guarantee system of "data-driven feature extraction + physical model constraints + real-time error correction" is formed. This application overcomes the limitations of a single model or data-driven method by integrating the generalization capability of the cloud-based deep learning model with the real-time performance of the terminal's physical model. It combines a dual Kalman filter mechanism to dynamically suppress error accumulation and embeds an electrochemical equivalent circuit model with a physical feature library, thus solving the challenge of estimating the SOC under complex operating conditions.

[0062] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0063] It should be noted that the execution subject of the battery state of charge estimation method based on edge-cloud collaboration provided in this application embodiment can be a battery state of charge estimation device based on edge-cloud collaboration. This device can be deployed on electronic devices such as edge devices, and this application embodiment does not impose any restrictions.

[0064] Figure 1This is a flowchart illustrating a battery state-of-charge estimation method based on edge-cloud collaboration, provided as an embodiment of this application. This method can be executed by edge devices, achieving more accurate and reliable battery state-of-charge estimation by combining a high-precision model in the cloud with localized estimation on the terminal. Figure 1 As shown, the method may include:

[0065] S101. Obtain the estimated first state of charge of a single cell in the electrochemical energy storage battery compartment from the cloud; wherein, the estimated first state of charge is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of convolutional neural network CNN-long short-term memory network LSTM that incorporates the physical characteristics of negative resistance.

[0066] For example, in this step, the edge device can obtain the estimated first state of charge (SOC) of each individual cell in the electrochemical energy storage battery compartment from a cloud server via a communication module (such as a wireless network or IoT interface). The electrochemical energy storage battery compartment typically includes multiple battery clusters connected in parallel. Each battery cluster contains multiple individual cells connected in series. For instance, in a 280Ah battery compartment, one compartment contains eight battery clusters connected in parallel. Each cluster contains 416 individual cells connected in series. Therefore, the current in a single battery cluster is the same, while the individual cell voltage and temperature data will vary depending on the operation of the battery compartment.

[0067] Among them, the negative resistance physical characteristic refers to the change in internal resistance caused by polarization during the charging and discharging process of a battery. It manifests as a negative resistance phenomenon in the relationship between battery voltage and current under specific operating conditions. This characteristic can reflect the dynamic characteristics of the battery and help improve the model's ability to capture the nonlinear behavior of the battery.

[0068] The CNN-LSTM hybrid model is a deep learning model that typically includes an input layer, CNN layers, pooling layers, LSTM layers, fully connected layers, and an output layer. It combines the feature extraction capabilities of convolutional neural networks (CNNs) with the time series modeling capabilities of long short-term memory networks (LSTMs). CNNs are used to extract spatial features from battery operating data (such as voltage, current, and temperature), while LSTMs are used to learn temporal dependencies, thereby achieving accurate prediction of SOC (State of Charge).

[0069] Understandably, Convolutional Neural Networks (CNNs) mainly consist of convolutional layers and pooling layers. They extract latent features from sample data through convolutional kernels and add activation functions after the convolutional layers to introduce non-linear features. CNNs feature weight sharing and local connectivity, which can significantly reduce the number of model parameters, accelerate the training process, and improve generalization performance. Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs). They add input gates, output gates, and forget gates to the hidden layers and introduce units for storing information, enabling them to handle long-term dependencies in time series. Sequence feature estimation is a typical time series analysis problem, but its application scenarios often involve highly complex dynamic environments and require real-time processing capabilities, which poses a challenge to feature extraction from data. Convolutional Neural Networks (CNNs) can mine latent features from sample data, while Long Short-Term Memory (LSTM) networks excel at capturing long-term dependencies in time series, thus having a significant advantage in processing time series data. Therefore, this application's embodiments adopt a hybrid model—using LSTM as the backend architecture of CNNs, named the CNN-LSTM hybrid model. This model extracts features from the data using CNN, and then concatenates them with LSTM to learn the hidden data relationships in the time series, thereby enabling the prediction of SOC values.

[0070] Real-time operating data of individual cells includes, but is not limited to, data such as individual cell voltage, individual cell temperature, and cell cluster current. This data can be collected by sensors deployed on the terminal and forwarded to the cloud in real time through edge devices.

[0071] Since the one-dimensional CNN-LSTM hybrid model requires a three-dimensional tensor as input, the raw real-time running data can first be converted into a time series format. Specifically, a time series tensor is constructed based on the time series length N of the model input as input, with both input and output structures having a "many-to-one" relationship. Considering that transmission latency can affect real-time performance in practical applications, the CNN-LSTM hybrid model focuses on SOC estimation for future time steps, i.e., predicting the battery's state of charge through extrapolation. After the cloud collects the real-time running data of a single battery cell, it is preprocessed (e.g., data cleaning, normalization, data transformation, etc.). Then, the processed data is input into the pre-trained CNN-LSTM hybrid model. This model, by incorporating the physical characteristics of negative resistance, can more accurately simulate the electrochemical behavior of the battery and output the first estimated state of charge value. Finally, the cloud sends this value to the edge device.

[0072] By using a CNN-LSTM hybrid model that integrates the physical characteristics of negative resistance, the cloud can leverage its powerful computing resources to perform complex modeling, overcoming the shortcomings of traditional methods in modeling the nonlinear characteristics of batteries, and improving the accuracy and robustness of SOC estimation in dynamically changing operating conditions.

[0073] Alternatively, in one possible embodiment, the CNN-LSTM hybrid model can be trained using a staged training strategy:

[0074] S1. First, freeze the LSTM layer and train the CNN layer separately. Use the training data and attach a temporary output layer for pre-training until the error of the charge state predicted by the CNN layer alone is lower than the first threshold.

[0075] S2. Unfreeze the LSTM layer and input the feature vector extracted by the CNN layer into the LSTM layer for end-to-end joint training until the preset termination condition is met, and then end the training to obtain the CNN-LSTM hybrid model. In the joint training, an attention mechanism layer is used, which is configured to dynamically adjust the attention weights on the physical features of negative resistance according to the working condition features in the input data.

[0076] For example, in order to address the problems of gradient instability, slow convergence speed, and susceptibility to local optima that are prone to occur in deep hybrid models due to their numerous parameters and complex structures, CNN-LSTM hybrid models can be trained using a staged training strategy.

[0077] Before training, training data can be acquired and divided into a 70% training set, 15% validation set, and 15% test set to ensure the fairness of model evaluation. During training, the parameters of the LSTM layer are first frozen (i.e., their weights remain unchanged), and the CNN layer is trained separately to enhance the CNN's ability to extract local features from the input data.

[0078] For example, the input tensor shape during CNN training is (N, 60, 5), where N is the batch size, 60 represents the time step (corresponding to 1 hour of data, each time step can be considered as 1 minute), and 5 is the feature dimension, including voltage (U), current (I), temperature (T), current standard deviation (σ_I), and voltage-current correlation coefficient (ρ). The data is pre-processed with min-max normalization to normalize each feature to the [0,1] interval to ensure training stability and convergence speed. The CNN layer receives temporal data (e.g., a 30-second subsequence can be extracted as local input), extracts spatial features through convolution operations, and outputs a 128-dimensional feature vector. To complete pre-training, a temporary output layer (a fully connected layer consisting of 1 neuron) is added after the CNN layer to directly predict the SOC value. The CNN layer is trained under supervised learning using the training set data for 5-10 epochs, focusing on optimizing the convolution kernel parameters to make it highly responsive to local features such as current abrupt changes and voltage fluctuations. The goal of pre-training is to make the mean absolute error (MAE) of the CNN predicting the SOC alone below a first threshold (e.g., 3%), so as to lay the foundation for subsequent LSTM layer fusion and ensure that the CNN has learned meaningful battery features.

[0079] Through pre-training, CNNs can independently identify key local patterns (such as the moment of a sudden change in current), avoiding insufficient feature extraction due to interference from LSTM during joint training, thereby improving the overall stability of the model.

[0080] After the CNN layers are pre-trained, the LSTM layers are unfrozen, and the 128-dimensional feature vectors output by the CNN layers are used as inputs to the LSTM layers for end-to-end joint training. This stage aims to combine the local feature extraction capabilities of CNNs with the long-term dependency modeling capabilities of LSTMs, and dynamically weight key features through an attention mechanism.

[0081] For example, the feature vectors extracted by the CNN layer are fed into an LSTM layer, which captures long-term dependencies in the time series. Subsequently, an attention mechanism layer is connected, configured to dynamically adjust the attention weights on the physical feature of negative resistance (Rneg) based on operating conditions in the input data, such as high current fluctuations or low temperatures. For instance, under inertial support conditions (σ_I > 20A and ρ < 0), the attention weights can be increased to over 0.7 to emphasize the impact of the negative resistance Rneg physical feature on SOC estimation. During training, the Adam optimizer can be used with an initial learning rate of 0.001 and a learning rate decay strategy (decreasing by 10% every 5 epochs), a batch size of 300, and 50 training epochs. Regularization measures are applied during training, including adding a BatchNorm layer after the convolutional layers to accelerate convergence and reduce overfitting, and using Dropout (rate=0.2) to randomly deactivate neurons in the fully connected layers. After each training round, the SOC estimation error (such as MAE and RMSE) is calculated on the validation set. An early stopping strategy is adopted: if the MAE on the validation set does not improve for 5 consecutive rounds, training is stopped and the current CNN layer weights are saved to prevent overfitting.

[0082] Joint training enables the model to utilize both local features and temporal context simultaneously, while the attention mechanism ensures that the model adaptively focuses on key physical features (such as Rneg) under different operating conditions, thereby improving the accuracy and generalization ability of SOC estimation. Finally, the model outputs the SOC prediction value through a fully connected layer and applies a sigmoid activation function to constrain it within a reasonable range.

[0083] Optionally, after training, the feature regions of interest to the CNN layers can be visualized using Grad-CAM to verify their consistency with physical meaning (such as the activation level at the moment of a sudden change in current). If a mismatch is found in the feature response, the convolution kernel size can be adjusted or specific working condition samples can be added to retrain the model.

[0084] In practical applications, the cloud can complete data training, parameter output, and feature library updates every 24 hours, providing high-precision parameter support for edge inference and ensuring that the model continuously adapts to changes in battery status. For example, from 0:00 to 0:30 every day, it automatically reads the new data (approximately 50,000 records) added in the previous 24 hours; from 0:30 to 3:30, it automatically executes the above training process to generate a new model; from 3:30 to 4:00, it compares the model with the historical best model; if the MAE decreases by more than 0.1%, it is pushed to the edge device; from 4:00 to 4:30, the edge device loads the new model, completing a seamless switch, and uses the new model for SOC estimation.

[0085] The phased training strategy detailed in this optional embodiment, combined with the attention mechanism, constitutes an efficient and reliable CNN-LSTM hybrid model construction scheme. It not only effectively solves the technical difficulties of deep model training in practice, but also improves the model's ability to capture and utilize the complex nonlinear behavior of batteries, especially key physical features (negative resistance), from a mechanistic perspective. This provides a solid model foundation for achieving high-precision first SOC estimation in the cloud.

[0086] Optionally, to improve the accuracy of the estimation results, the SOC value can also be displayed in the cloud. Based on this, a random noise digit v between 0 and 1 is added, and the final cloud-based estimate is recorded as... See the following formula (1):

[0087] (1)

[0088] S102. Obtain the second state of charge estimate of a single cell from the terminal; wherein, the second state of charge estimate is obtained by the terminal through a target equivalent circuit model corresponding to the current operating condition of the single cell and a preset state estimation algorithm, based on real-time operating data; the target equivalent circuit model is any one of the zero-order equivalent circuit model, the first-order equivalent circuit model, and the second-order equivalent circuit model.

[0089] For example, in this step, the edge device obtains a second state-of-charge estimate of the individual battery cell from a terminal (such as a battery management system, BMS, or local sensor). The target equivalent circuit model refers to a circuit model selected based on the battery's dynamic characteristics, including any one of a zero-order equivalent circuit model, a first-order equivalent circuit model, or a second-order equivalent circuit model. These models simulate the battery's internal impedance and polarization effects using circuit elements (such as resistors and capacitors).

[0090] The zero-order equivalent circuit model is the simplest, containing only a voltage source and internal resistance, suitable for static or low-precision scenarios. The first-order equivalent circuit model adds an RC parallel network to the zero-order model to simulate battery polarization effects, suitable for general dynamic operating conditions. The second-order equivalent circuit model adds two RC parallel networks to the first-order model, providing a more refined description of the battery's dynamic response, suitable for high-precision scenarios. The current operating condition refers to the battery's real-time operating state, such as charge / discharge rate, temperature, and load changes. In this embodiment, the terminal adaptively selects the most suitable equivalent circuit model based on the operating condition to balance computational complexity and estimation accuracy.

[0091] Preset state estimation algorithms refer to algorithms used for state estimation, such as Kalman filtering (KF), extended Kalman filtering (EKF), or unscented Kalman filtering (UKF). These algorithms are based on equivalent circuit models and recursively estimate the SOC value using real-time running data.

[0092] The terminal can monitor the real-time operating data of a single battery cell and select a target equivalent circuit model based on the current operating conditions (such as current change rate or temperature threshold). Then, it processes the real-time operating data using a preset state estimation algorithm (such as EKF), and outputs a second SOC estimate through model equations and observation updates. Finally, the terminal sends this value to the edge device so that the edge device can obtain the second state of charge estimate of the single battery cell.

[0093] By using an equivalent circuit model and state estimation algorithm, the terminal can achieve low-latency and low-computational-cost SOC estimation locally. The equivalent circuit model has a simple structure and is easy to implement. Furthermore, through adaptive selection of operating conditions, the applicability of the model in different scenarios is ensured. This step compensates for the lack of real-time performance of cloud models that may be caused by communication delays or data loss.

[0094] Optionally, in one possible embodiment, the terminal calculates the second state of charge estimate in the following manner:

[0095] S10. Based on the real-time operating data of a single battery cell, determine the current operating condition of the single battery cell, and determine the target equivalent circuit model according to the current operating condition.

[0096] S20. Based on real-time operating data, the model parameters of the target equivalent circuit model are identified online using the recursive least squares method.

[0097] S30. Substitute the identified model parameters into the state equation constructed based on the ampere-hour integration method and the observation equation constructed based on the target equivalent circuit model, and perform state estimation through the dual Kalman filter algorithm to output the estimated state of charge of a single cell.

[0098] Optionally, determining the current operating condition of a single battery cell based on its real-time operating data may include:

[0099] S11. Based on the real-time operating data of a single battery cell, calculate the current standard deviation within a preset time window, and calculate the voltage-current dynamic correlation coefficient within the same time window.

[0100] S12. If the current standard deviation is less than the first preset threshold, then the current operating condition of the single cell is determined to be the standby operating condition.

[0101] S13. If it is determined that the current standard deviation is greater than the second preset threshold and the voltage-current dynamic correlation coefficient indicates a negative correlation, then the current operating condition of the single cell is determined to be the inertia support operating condition; wherein, the second preset threshold is greater than the first preset threshold.

[0102] S14. Otherwise, determine the current operating condition of the individual battery as the charging condition.

[0103] For example, the terminal can continuously collect real-time operating data of individual battery cells, including terminal voltage U and load current I (positive for charging, negative for discharging). To quantify and analyze the dynamic behavior of the battery, the terminal can set a preset time window Tw (e.g., 5 seconds) and calculate the current standard deviation σ_I and the voltage-current dynamic correlation coefficient ρ within this sliding window. The current standard deviation measures the degree of current fluctuation within the window; a larger value indicates more drastic current changes and a higher demand for battery dynamic response. The voltage-current dynamic correlation coefficient measures the correlation between the directions of voltage and current changes within the same time window.

[0104] When the calculated current standard deviation is less than the set first preset threshold (this threshold can be set according to system noise and typical current fluctuation range in standby mode, for example, 0.2A), it indicates that the battery is in a near-static or extremely low power exchange state. In this case, the current operating condition of a single battery cell can be determined as standby mode.

[0105] When the calculated current standard deviation exceeds the second preset threshold (this threshold is used to identify high dynamic processes, calibrated for typical operating conditions of a 280Ah lithium iron phosphate battery, and can be set to 20A, but can also be adjusted according to the application scenario), and the voltage-current dynamic correlation coefficient indicates a negative correlation (i.e., ρ < 0), then the current operating condition is determined to be an inertia support condition. In this case, it usually corresponds to the battery's drastic and rapid charging and discharging switching to provide rapid frequency support when the grid frequency fluctuates. At this time, the voltage and current are negatively correlated, which is a typical manifestation of "negative resistance" characteristics, and the battery exhibits the strongest dynamic characteristics.

[0106] When the current operating condition is neither a standby condition nor an inertia support condition, it can be considered a charging condition. Understandably, the "charging condition" here is a broad concept, covering a wide range of operating states from stable charging to general dynamic discharging (the degree of dynamics is between standby and inertia support).

[0107] In addition, the current operating condition of a single battery cell can be determined by other methods, and this application embodiment does not impose any limitations. For example, when the current |I| < 14A and the current standard deviation σ_I < 0.2A, it is determined to be a standby operating condition; when the current I > 28A and the forward ratio is 100%, it is determined to be a charging operating condition; when |I| ≥ 84A and σ_I > 20A, it is determined to be an inertia support operating condition.

[0108] Optionally, determining the target equivalent circuit model based on the current operating conditions may include:

[0109] S15. If the current operating condition is standby, then select the zero-order equivalent circuit model as the target equivalent circuit model.

[0110] S16. If the current operating condition is charging, then select the first-order equivalent circuit model as the target equivalent circuit model.

[0111] S17. If the current operating condition is an inertia-supported operating condition, then select the second-order equivalent circuit model as the target equivalent circuit model.

[0112] For example, in standby mode, current fluctuations are minimal, and battery polarization effects and dynamic processes can be ignored. The zero-order RC model contains only an ideal voltage source (open-circuit voltage OCV) and an ohmic internal resistance R0. The model parameters only include R0, resulting in the simplest structure, minimal computational load, and an EKF state vector simplified to [SOC]. The computation time is ≤1s. In such static or quasi-static scenarios, using the zero-order RC model is sufficient to meet the estimation accuracy requirements while maximizing the conservation of terminal computing resources.

[0113] During charging, batteries exhibit significant polarization effects. A first-order RC model, based on the zero-order RC model, adds a parallel RC network (model parameters also include polarization resistance and polarization capacitance, i.e., R0 + Rp1 / Cp1) to describe the battery's relaxation process. The state vector is [SOC, Up1], and the computation time is ≤5s. It effectively characterizes the main dynamic characteristics under these conditions, achieving a good balance between accuracy and computational complexity, making it an ideal choice for everyday applications with varying operating conditions.

[0114] Under highly dynamic operating conditions such as inertia support, the battery polarization process becomes exceptionally complex, with the superposition of fast and slow polarization effects. The second-order RC model, comprising two parallel RC networks (model parameters include R0 + Rp1 / C1 + Rp2 / Cp2), can more precisely describe the battery's voltage response under drastically changing current excitation, particularly capturing the dynamic behavior related to the "negative resistance" characteristic more accurately. Its state vector is [SOC, Up1, Up2], and the computation time is ≤10s. Although it has the highest computational cost, the second-order RC model provides the highest model fidelity, ensuring the accuracy and stability of SOC estimation under the most demanding operating conditions, which is crucial for ensuring the safe and reliable operation of grid ancillary services.

[0115] Based on the essential differences in the dynamic characteristics of the battery under different operating conditions, the embodiments of this application adaptively select the zero-order, first-order, or second-order equivalent circuit model that matches it. On the one hand, this avoids the waste of computing resources caused by using complex models under simple operating conditions, and on the other hand, it ensures that there is sufficient model accuracy to maintain estimation performance under complex operating conditions. Thus, it achieves the optimal balance between computing efficiency and estimation accuracy at the system level, and significantly improves the practicality and adaptability of the terminal SOC estimation method.

[0116] After selecting the target equivalent circuit model, the terminal can use the latest collected real-time operating data to identify the model parameters of the target equivalent circuit model online using the recursive least squares method. The recursive least squares method is a highly efficient online parameter estimation algorithm. Its core lies in the fact that it does not require storing historical data blocks; it only needs to use new sampled data, combined with the parameter estimates and covariance matrix from the previous time step, to update the parameter estimates for the current time step through a recursive formula. In this embodiment, the model parameters of the target equivalent circuit model can first be retrieved from a pre-built RC model parameter mapping library, and then processed to obtain the final model parameters.

[0117] Among them, the pre-built RC model parameter mapping library has calibrated the basic parameters for different operating conditions for zero-order / first-order / second-order RC models (such as the zero-order model retaining only the ohmic resistance R0, the first-order model adding the polarization resistance R1 / C1, and the second-order model supplementing R2 / C2), forming a mapping table of "operating condition-model order-parameter range" (such as standby-zero-order-R0=0.003~0.005Ω).

[0118] Online identification can reflect the changes in model parameters with SOC, temperature, and aging degree in real time, overcoming the problem of accumulated estimation errors caused by changes in battery characteristics in fixed parameter models. Moreover, the recursive least squares method has low computational load and low storage requirements, making it very suitable for real-time operation in embedded systems of terminals, ensuring the timeliness and accuracy of parameter identification.

[0119] Furthermore, the model parameters identified above are substituted into the state-space model, and then the dual Kalman filter algorithm is used for state estimation, finally outputting a high-precision second SOC estimate. Taking the second-order equivalent circuit equation as an example, the results of parameter identification R0, R1, C1, R2, and C2 are substituted into the improved RC equivalent circuit equation, which can include the following formulas (2) to (4):

[0120] (2)

[0121] (3)

[0122] (4)

[0123] in, , These are the voltages of the first RC circuit at times K and K+1, respectively; , These are the voltages of the second RC circuit at times K and K+1, respectively; Let be the open-circuit voltage, which is a function of the state of charge (SOC). Let K be the input current at time K (positive for charging, negative for discharging). Let τ be the terminal voltage at time K (which can be measured); τ1 = R1C1, which represents the time constant of the first RC loop; τ2 = R2C2, which represents the time constant of the second RC loop.

[0124] In this embodiment, the state vector x = SOC(K+1) can be constructed using the ampere-hour integration method and deployed on the terminal side, and can take the form of the following formula (5):

[0125] (5)

[0126] in, yes The state of charge of the battery at any given time. yes The state of charge of the battery at any given time. From The integral of the battery current over time from time k+T. This is the battery's rated capacity.

[0127] The observation equation is constructed based on the target equivalent circuit model, and establishes the mathematical relationship between battery terminal voltage and SOC (through the OCV-SOC relationship curve), current and model parameters. In this embodiment, the observation equation can incorporate the negative resistance Rneg correction value issued by the edge terminal, and the specific formula can be the following formula (6):

[0128] (6)

[0129] Among them, open circuit voltage This refers to the potential difference between the two terminals of a battery when it is unloaded (i.e., no current is flowing through it). The fact that the open-circuit voltage is a function of the state of charge (SOC) means that the battery's open-circuit voltage will change as the remaining charge of the battery changes.

[0130] The observation equation for the terminal voltage can be given by formula (7), the state equation for the terminal voltage can be given by formula (8), and the measurement variable for the terminal voltage can be given by formula (9):

[0131] (7)

[0132] = [SOC (8)

[0133] (9)

[0134] Combining the above formulas (2) to (9), the dual Kalman filter algorithm is used for processing, and a high-precision second SOC estimate can be output. The dual Kalman filter algorithm consists of two parallel Kalman filters: a state filter (main filter) for estimating the system's state variables (i.e., SOC), and a parameter filter (auxiliary filter) for estimating the system's model parameters. The two filters alternately perform state filtering and parameter filtering in each sampling period. State filtering estimates the SOC by using standard Kalman filtering steps (prediction and update) while keeping the current parameter estimates fixed. Parameter filtering updates the parameter estimates by treating the model parameters as state variables while keeping the current state estimate (SOC) fixed, using another Kalman filter.

[0135] This application utilizes a dual Kalman filter algorithm to simultaneously and optimally estimate the battery's state of charge (SOC) and model parameters, forming a virtuous cycle of "parameter identification supporting state estimation, and state estimation assisting parameter identification," significantly improving the overall accuracy of SOC estimation. Simultaneously, by using terminal voltage feedback in the observation equation, it effectively corrects the accumulated errors caused by initial errors, current measurement noise, and capacity decay in the ampere-hour integration method. Furthermore, this algorithm can handle uncertainties in state and parameter estimation, effectively suppressing measurement noise and model mismatch, making it suitable for complex real-world operating environments.

[0136] Optionally, in one possible embodiment, before obtaining the estimated first state of charge of a single cell in the electrochemical energy storage battery compartment from the cloud, the method provided in this application embodiment may further include:

[0137] S01. Query the prior value of negative resistance corresponding to the current state of charge, temperature and current operating condition of a single cell from the three-dimensional feature library of negative resistance.

[0138] S02. Based on the real-time voltage and current obtained from the terminal, the prior value of the negative resistance is fine-tuned online using the recursive least squares method to obtain the corrected negative resistance value; wherein, the corrected negative resistance value is provided to the cloud and / or the terminal for calculating the estimated state of charge value.

[0139] For example, the three-dimensional feature library of negative resistance can be stored locally or in the cloud. It is a pre-generated database or mapping table that stores the prior negative resistance value Rneg of a single cell under different states of charge, temperatures, and operating conditions. For instance, experiments can be conducted under combined conditions of temperature 0℃~55℃ (step 10℃), SOC 5%~95% (step 5%), and current rate 0.1C~1.5C (step 0.1C). The experiment is run for 30 seconds under each (temperature, SOC, current rate) combination at a sampling frequency of 1 time / second. The Rneg value is calculated by fitting the voltage-current curve using the least squares method. Each combination is repeated 3 times and the average value is taken, resulting in a total of (temperature 12) × (SOC 19) × (current rate 15) = 3420 sets of data. Finally, the data is stored in a cloud MySQL database in the form of key-value pairs (e.g., "25℃-50%-0.2C":"-0.0025Ω") with a confidence interval of ±0.0003Ω.

[0140] Edge devices can use the estimated or measured current state of charge (SOC), current temperature, and identified current operating condition of a single cell as joint input keys to perform a matching query in a three-dimensional negative resistance feature library. The query algorithm can be nearest neighbor matching or three-dimensional linear interpolation to output the best-matching prior negative resistance value. After obtaining the prior negative resistance value, the edge device then uses the latest real-time operating data to fine-tune the prior negative resistance value online using the recursive least squares method to obtain the corrected negative resistance value. For example, it requests the prior Rneg value corresponding to the current (SOC, temperature, operating condition) from the cloud (e.g., SOC=50%, 25℃, inertia support → Rneg=-0.0025Ω); combined with the 1Hz U / I data on the edge side, it performs "real-time fine-tuning using the least squares method": Rneg'=Rneg×(1+α). Where α is the correction coefficient (α=0.05~0.1, dynamically adjusted according to the deviation between the measured U value and the model prediction value).

[0141] The corrected negative resistance value Rneg' can be provided to the cloud and / or the terminal for calculating the estimated state of charge (SOC). When provided to the cloud, the corrected negative resistance value is used as an important input feature or constraint when running a CNN-LSTM hybrid model for the first SOC estimation. This significantly improves the accuracy and physical consistency of the cloud-based deep learning model in capturing the complex nonlinear dynamics of the battery, particularly the negative resistance effect. When provided to the terminal, the corrected negative resistance value can be incorporated into the observation equations of its target equivalent circuit model. This allows the equivalent circuit model to more accurately describe the battery terminal voltage behavior, especially under high-dynamic inertia support conditions, thereby indirectly improving the accuracy of the second SOC estimation based on the equivalent circuit model.

[0142] The negative resistance prior value query and online fine-tuning mechanism introduced in this optional embodiment serves as a pre-enhancement step in the edge-cloud collaborative SOC estimation process, combining prior physical knowledge of the battery with real-time data-driven correction. It not only injects crucial physical information into the cloud model, enhancing its extrapolation capability and interpretability, but also provides dynamically accurate parameters for the terminal's equivalent circuit model, thereby improving the reliability of the first and second SOC estimates from the outset.

[0143] S103. The first state of charge estimate and the second state of charge estimate are fused using an adaptive Kalman filter algorithm to obtain the target state of charge estimate of a single cell.

[0144] For example, in this step, the edge device uses an adaptive Kalman filter algorithm to fuse the first SOC estimate obtained from the cloud and the second SOC estimate obtained from the terminal to obtain the final target SOC estimate. The adaptive Kalman filter algorithm is an improved Kalman filter algorithm that can automatically adjust the statistical characteristics of process noise and observation noise to cope with uncertainties or changes in the system model. This algorithm improves the stability and accuracy of filtering by estimating noise parameters in real time. The fusion processing refers to weighted combination of the estimates from the cloud and the terminal to leverage their respective advantages and obtain a more reliable result.

[0145] By fusing the data using an adaptive Kalman filter algorithm, the high accuracy of the cloud model and the real-time performance of the terminal model can be effectively combined, overcoming the limitations of a single model. The adaptive nature of the algorithm ensures its robustness even under uncertain environments such as battery aging and temperature variations, significantly improving the accuracy and reliability of SOC estimation. Furthermore, performing the fusion process at the edge reduces the computational burden on the cloud and minimizes communication overhead.

[0146] Optionally, in one possible embodiment, the first state-of-charge estimate and the second state-of-charge estimate are fused using an adaptive Kalman filter algorithm to obtain the target state-of-charge estimate of the single cell, which may include:

[0147] S100. The estimated value of the first state of charge is used as the observation prediction value, and the estimated value of the second state of charge is used as the state observation value.

[0148] S200. Determine the observation noise covariance based on the confidence interval attached to the first state of charge estimate;

[0149] S300, an innovative sequence based on state prediction values, adaptively adjusts the process noise covariance;

[0150] S400. Calculate the Kalman gain based on the observation noise covariance and the adjusted process noise covariance to perform weighted fusion of the observation predictions and state observations, thereby obtaining the target state of charge estimate.

[0151] For example, when constructing a Kalman filter state-space model, the first SOC estimate provided by the cloud can be used as the observed prediction value, and the second SOC estimate provided by the terminal can be used as the state observation value. Furthermore, when the cloud distributes the first SOC estimate, it also provides a confidence interval for that estimate. This confidence interval reflects the cloud model's self-assessment of the reliability of the current estimation result. The edge device can dynamically determine the observation noise covariance matrix R based on the width of this confidence interval. Specifically, a wider confidence interval indicates higher uncertainty in the cloud model regarding the current estimation result, thus requiring a larger observation noise covariance; conversely, a narrower confidence interval requires a smaller observation noise covariance.

[0152] Furthermore, the innovation sequence refers to the sequence of differences between the observed state values ​​(the second SOC estimate) and the predicted state values ​​(the predicted values ​​obtained by extrapolating the state estimate from the previous time step using the state equation). This sequence reflects the degree of mismatch between the system model and the actual process. Edge devices monitor and analyze the statistical characteristics of the innovation sequence (such as its autocorrelation function or covariance) in real time. If the innovation sequence indicates unmodeled dynamics or increased model error, the process noise covariance matrix Q is adaptively increased to make the filter more confident in the observed values; conversely, if the model matches well, the process noise covariance is decreased to make the filter more confident in the predicted values.

[0153] Then, based on the observation noise covariance R determined in step S200 and the process noise covariance Q adaptively adjusted in step S300, the Kalman gain H can be calculated according to the standard Kalman filter formula. The Kalman gain is essentially the weight assigned to the state observation. Finally, using the calculated Kalman gain H, the observed prediction value (first SOC estimate) and the state observation value (second SOC estimate) are optimally weighted and fused to obtain the target state of charge estimate of a single cell. Among them, the final formula of the combined extended dual Kalman filter and second-order RC fusion algorithm can be the following formula (10):

[0154] (10)

[0155] in, These are preset coefficients. It is the estimated value of the first state of charge, that is = .

[0156] This optional embodiment optimizes role allocation by assigning the high-precision cloud model and the real-time-performance terminal model to the roles of observation prediction and state observation, respectively, laying the foundation for effective fusion. By adaptively determining observation noise using confidence intervals provided by the cloud, dynamic uncertainty management is achieved, enabling the fusion process to intelligently respond to changes in the reliability of cloud results. By adaptively adjusting process noise based on innovative sequences, the inaccuracies of the state equation can be automatically compensated, exhibiting stronger adaptability and robustness to model mismatch caused by battery aging and sudden changes in operating conditions. Finally, optimal weighted fusion is performed through dynamically calculated Kalman gain, achieving optimal utilization of the two information sources in an uncertain environment, thereby obtaining a more accurate and reliable SOC estimate than either single source.

[0157] For example, Figure 2 This is a framework diagram of a battery state-of-charge estimation method based on an adaptive extended Kalman filter algorithm, provided as an embodiment of this application. Figure 2 As shown, by selecting state variables and the error covariance matrix, combining the CNN-LSTM weighted SOC input and RC model parameters, and utilizing the five core formulas of the Extended Kalman Filter (EPF) for state estimation, and then adaptively adjusting the noise factor to dynamically optimize filtering performance, a more accurate and stable SOC estimation result can be output. The overall process integrates filtering methods and deep learning techniques, improving the adaptability and accuracy of SOC estimation.

[0158] For example, Figure 3 This is a comparative diagram of a fused SOC and an ampere-hour integral SOC provided as an embodiment of this application. For example... Figure 3 As shown in the figure, the vertical axis represents SOC (%), and the horizontal axis represents time. The "Modify AH Integral SOC" curve represents the SOC value estimated using the modified ampere-hour integration method, while the "EMS system SOC" curve represents the SOC value estimated by the battery energy management system. Throughout the time range, the two curves and values ​​are similar for most of the period, but a significant deviation begins to appear in the mid-to-late stages (approximately after 4 × 10^5 on the horizontal axis). This indicates that the two estimation methods may produce different results in the later stages of battery use due to differences in algorithms or parameters.

[0159] According to experiments, the battery state-of-charge (SOC) estimation method based on edge-cloud collaboration provided in this application demonstrated excellent performance in continuous daily operation data for more than three months. The fused SOC algorithm of this application converged to within 3% of the ampere-hour integral error. Specifically, the terminal voltage MAE was 53.8088 mV, the SOC error MAE was 0.01208%, the maximum SOC error was 2.0009, the minimum SOC error was -0.068886, the mean SOC error was 0.006647%, the median SOC error was -4.1337e-06%, and the SOC error RMSE was 0.026744%. The battery SOC estimation method based on edge-cloud collaboration provided in this application can be used in actual energy storage projects to perform online SOC estimation for battery clusters and individual battery cells in energy storage containers.

[0160] Furthermore, in this embodiment, the electrochemical energy storage battery compartment may include multiple battery clusters connected in parallel, each battery cluster comprising multiple individual cells connected in series. Therefore, when estimating the state of charge (SOC) of the battery, all individual cells in the battery compartment can be selected concurrently, and the SOC of each individual cell can be estimated to determine the target SOC value for each individual cell. Furthermore, when there is a need to determine the SOC value of the battery cluster, if the electrochemical energy storage battery compartment is determined to be in a charging state, the SOC value of the battery cluster can be determined by selecting the largest target SOC value among the individual cells; if the electrochemical energy storage battery compartment is determined to be in a discharging state, the SOC value of the battery cluster can be determined by selecting the smallest target SOC value among the individual cells.

[0161] The battery state-of-charge (SOC) estimation method based on edge-cloud collaboration provided in this application achieves triple protection—data-driven feature extraction, physical model constraints, and real-time error correction—through deep integration of an edge-cloud collaborative architecture, a hybrid model, and an adaptive filtering mechanism. This improves SOC estimation accuracy by more than 40% compared to traditional methods, maintaining an accuracy within ±2% even under extreme conditions. From a practical application perspective, this solution significantly reduces the computing power requirements of terminal hardware while ensuring high accuracy, allowing stable operation without high-performance chips and significantly reducing industrialization costs. Furthermore, the high-precision SOC output can directly support battery equalization control and lifespan prediction, thereby extending battery cycle life by more than 15%, effectively reducing battery replacement frequency and carbon emissions. It provides a complete solution for large-capacity energy storage power stations that combines high accuracy, high reliability, low cost, and environmental benefits, achieving a balance between accuracy, real-time performance, and robustness in SOC estimation. It is particularly suitable for complex application scenarios such as electrochemical energy storage battery compartments, providing technical support for battery health management and safe operation.

[0162] Figure 4 A schematic diagram of a battery state-of-charge estimation device based on edge-cloud collaboration is provided for an embodiment of this application, as shown below. Figure 4 As shown, the battery state of charge estimation device 40 based on edge-cloud collaboration provided in this embodiment includes: a first processing unit 401, a second processing unit 402 and a third processing unit 403.

[0163] The first processing unit 401 is used to obtain the first state of charge estimate of a single cell in the electrochemical energy storage battery compartment from the cloud; wherein, the first state of charge estimate is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of convolutional neural network CNN-long short-term memory network LSTM that incorporates the physical characteristics of negative resistance.

[0164] The second processing unit 402 is used to obtain the second state of charge estimation value of a single battery cell from the terminal; wherein, the second state of charge estimation value is obtained by the terminal through a target equivalent circuit model corresponding to the current operating condition of the single battery cell and a preset state estimation algorithm, based on real-time operating data for state estimation;

[0165] The third processing unit 403 is used to fuse the first state of charge estimate and the second state of charge estimate using a dual adaptive Kalman filter algorithm to obtain the target state of charge estimate of a single cell.

[0166] In one possible implementation, the third processing unit 403 is specifically used for:

[0167] The first state of charge estimate is used as the observation prediction value, and the second state of charge estimate is used as the state observation value.

[0168] The observation noise covariance is determined based on the confidence interval attached to the first state of charge estimate.

[0169] Innovative sequences based on state prediction values ​​adaptively adjust process noise covariance;

[0170] The Kalman gain is calculated based on the observation noise covariance and the adjusted process noise covariance to perform weighted fusion of the observation predictions and state observations, thereby obtaining the target state of charge estimate.

[0171] In one possible implementation, the CNN-LSTM hybrid model is trained using a staged training strategy:

[0172] First, freeze the LSTM layer and train the CNN layer separately. Use the training data and attach a temporary output layer for pre-training until the error of the charge state predicted by the CNN layer alone is lower than the first threshold.

[0173] Unfreeze the LSTM layer, input the feature vectors extracted by the CNN layer into the LSTM layer for end-to-end joint training until the preset termination condition is met, and then end the training to obtain the CNN-LSTM hybrid model.

[0174] In the joint training, an attention mechanism layer is used, which is configured to dynamically adjust the attention weights on the physical features of negative resistance based on the operating conditions in the input data.

[0175] In one possible implementation, the terminal calculates the second state of charge estimate in the following manner:

[0176] Based on the real-time operating data of individual cells, the current operating condition of the individual cells is determined, and the target equivalent circuit model is determined according to the current operating condition.

[0177] Based on real-time operational data, the model parameters of the target equivalent circuit model are identified online using the recursive least squares method.

[0178] The identified model parameters are substituted into the state equations constructed based on the ampere-hour integration method and the observation equations constructed based on the target equivalent circuit model. The state is estimated by the dual Kalman filter algorithm, and the estimated state of charge of a single cell is output.

[0179] In one possible implementation, the current operating condition of a single battery cell is determined based on its real-time operating data, including:

[0180] Based on the real-time operating data of individual cells, the current standard deviation within a preset time window is calculated, and the voltage-current dynamic correlation coefficient within the same time window is also calculated.

[0181] If the current standard deviation is less than the first preset threshold, the current operating condition of the single cell is determined to be the standby operating condition.

[0182] If the current standard deviation is determined to be greater than the second preset threshold and the voltage-current dynamic correlation coefficient indicates a negative correlation, then the current operating condition of the single cell is determined to be the inertia support operating condition; wherein the second preset threshold is greater than the first preset threshold.

[0183] Otherwise, the current operating condition of the individual battery cell is determined to be the charging condition.

[0184] In one possible implementation, the target equivalent circuit model is determined based on the current operating conditions, including:

[0185] If the current operating condition is standby, then the zero-order equivalent circuit model is selected as the target equivalent circuit model.

[0186] If the current operating condition is charging, then the first-order equivalent circuit model is selected as the target equivalent circuit model.

[0187] If the current operating condition is an inertia-supported operating condition, then the second-order equivalent circuit model is selected as the target equivalent circuit model.

[0188] In one possible implementation, the first processing unit 401, before obtaining the estimated first state of charge of a single cell in the electrochemical energy storage battery compartment from the cloud, is further configured to:

[0189] Query the prior value of negative resistance corresponding to the current state of charge, temperature and current operating conditions of a single cell from the three-dimensional feature library of negative resistance;

[0190] Based on the real-time voltage and current obtained from the terminal, the prior value of the negative resistance is fine-tuned online using the recursive least squares method to obtain the corrected negative resistance value.

[0191] The corrected negative resistance value is provided to the cloud and / or terminal for calculating the estimated state of charge.

[0192] In one possible implementation, the electrochemical energy storage battery compartment includes multiple battery clusters connected in parallel, each battery cluster including multiple individual cells connected in series; the third processing unit 403 is further used for:

[0193] When it is determined that the electrochemical energy storage battery compartment is in a charging state, the maximum value among the target state of charge estimates of each individual cell is determined as the state of charge estimate of the battery cluster.

[0194] When it is determined that the electrochemical energy storage battery compartment is in a discharged state, the minimum value among the target state of charge estimates of each individual cell is determined as the state of charge estimate of the battery cluster.

[0195] The apparatus provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0196] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. These modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented in software via processing element calls, while others are implemented in hardware. Furthermore, they can be stored as program code in the device's memory, and the data processing modules can be called and executed by a specific processing element. The implementation of other modules is similar. These modules can be fully or partially integrated together, or implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0197] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0198] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0199] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0200] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0201] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0202] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0203] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0204] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0205] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0206] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0207] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0208] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0209] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0210] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0211] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0212] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A battery state-of-charge estimation method based on edge-cloud collaboration, characterized in that, The method includes: The first state of charge (SOC) estimate of a single cell in the electrochemical energy storage battery compartment is obtained from the cloud. The first SOC estimate is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM) that incorporates the physical characteristics of negative resistance. The second state of charge (SOC) estimate of the individual battery cell is obtained from the terminal. The second SOC estimate is obtained by the terminal through state estimation based on the real-time operating data using a target equivalent circuit model corresponding to the current operating condition of the individual battery cell and a preset state estimation algorithm. The target equivalent circuit model can be any one of a zero-order equivalent circuit model, a first-order equivalent circuit model, or a second-order equivalent circuit model. The first state of charge estimate and the second state of charge estimate are fused using an adaptive Kalman filter algorithm to obtain the target state of charge estimate of the single cell.

2. The method according to claim 1, characterized in that, The step of fusing the first state-of-charge (POC) estimate and the second POC estimate using an adaptive Kalman filter algorithm to obtain the target POC estimate for the individual battery cell includes: The first state of charge estimate is used as the observation prediction value, and the second state of charge estimate is used as the state observation value. Based on the confidence interval attached to the first state of charge estimate, the observation noise covariance is determined; Based on the innovative sequence of the state prediction values, the process noise covariance is adaptively adjusted; Based on the observed noise covariance and the adjusted process noise covariance, the Kalman gain is calculated to perform weighted fusion of the observed predicted value and the state observed value to obtain the target state of charge estimate.

3. The method according to claim 1, characterized in that, The CNN-LSTM hybrid model is trained using a staged training strategy. First, freeze the LSTM layer and train the CNN layer separately. Use the training data and attach a temporary output layer for pre-training until the error of the state of charge predicted by the CNN layer alone is lower than the first threshold. Unfreeze the LSTM layer, input the feature vector extracted by the CNN layer into the LSTM layer for end-to-end joint training until the preset termination condition is met, and then end the training to obtain the CNN-LSTM hybrid model. In the joint training, an attention mechanism layer is employed, which is configured to dynamically adjust the attention weights on the negative resistance physical features based on the operating condition features in the input data.

4. The method according to claim 1, characterized in that, The terminal calculates the second estimated state of charge value in the following manner: Based on the real-time operating data of the individual battery, the current operating condition of the individual battery is determined, and based on the current operating condition, the target equivalent circuit model is determined. Based on the real-time operating data, the model parameters of the target equivalent circuit model are identified online using the recursive least squares method. The identified model parameters are substituted into the state equation constructed based on the ampere-hour integration method and the observation equation constructed based on the target equivalent circuit model. The state is estimated by the dual Kalman filter algorithm, and the estimated state of charge of the single cell is output.

5. The method according to claim 4, characterized in that, Determining the current operating condition of a single battery cell based on its real-time operating data includes: Based on the real-time operating data of the individual battery cells, the current standard deviation within a preset time window is calculated, and the voltage-current dynamic correlation coefficient within the same time window is calculated. If the current standard deviation is less than the first preset threshold, then the current operating condition of the single battery cell is determined to be the standby operating condition. If it is determined that the current standard deviation is greater than the second preset threshold and the voltage-current dynamic correlation coefficient indicates a negative correlation, then the current operating condition of the single cell is determined to be the inertia support operating condition; wherein, the second preset threshold is greater than the first preset threshold; Otherwise, the current operating condition of the single battery cell is determined to be the charging condition.

6. The method according to claim 5, characterized in that, Determining the target equivalent circuit model based on the current operating conditions includes: If the current operating condition is a standby operating condition, then the zero-order equivalent circuit model is selected as the target equivalent circuit model; If the current operating condition is a charging condition, then the first-order equivalent circuit model is selected as the target equivalent circuit model. If the current operating condition is an inertia-supported operating condition, then the second-order equivalent circuit model is selected as the target equivalent circuit model.

7. The method according to any one of claims 1-6, characterized in that, Before obtaining the estimated first state of charge of a single cell in the electrochemical energy storage battery compartment from the cloud, the method further includes: Query the prior value of negative resistance corresponding to the current state of charge, temperature and current operating condition of the single cell from the three-dimensional feature library of negative resistance; Based on the real-time voltage and current obtained from the terminal, the prior value of the negative resistance is fine-tuned online using the recursive least squares method to obtain the corrected negative resistance value. The corrected negative resistance value is provided to the cloud and / or the terminal for calculating the estimated state of charge.

8. The method according to any one of claims 1-6, characterized in that, The electrochemical energy storage battery compartment includes multiple battery clusters connected in parallel, each battery cluster including multiple individual cells connected in series; the method further includes: When it is determined that the electrochemical energy storage battery compartment is in a charging state, the maximum value among the target state of charge estimates of each individual battery cell is determined as the state of charge estimate of the battery cluster. When it is determined that the electrochemical energy storage battery compartment is in a discharged state, the minimum value among the target state of charge estimates of each individual battery cell is determined as the state of charge estimate of the battery cluster.

9. A battery state-of-charge estimation device based on edge-cloud collaboration, characterized in that, The device includes: The first processing unit is used to obtain the first state of charge estimate of a single cell in the electrochemical energy storage battery compartment from the cloud; wherein, the first state of charge estimate is obtained by the cloud based on the real-time operating data of the single cell using a hybrid model of convolutional neural network CNN-long short-term memory network LSTM that incorporates the physical characteristics of negative resistance. The second processing unit is used to obtain a second state of charge estimation value of the single battery cell from the terminal; wherein the second state of charge estimation value is obtained by the terminal through state estimation based on the real-time operating data using a target equivalent circuit model corresponding to the current operating condition of the single battery cell and a preset state estimation algorithm; The third processing unit is used to fuse the first state of charge estimate and the second state of charge estimate using a dual adaptive Kalman filter algorithm to obtain the target state of charge estimate of the single cell.

10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

Images