A method, device, equipment, medium and product for evaluating a state of an energy storage battery

Abstract

The application discloses a kind of energy storage battery state evaluation method, device, equipment, medium and product, involve battery state evaluation field, the method includes actual energy storage battery data is preprocessed and feature extraction, obtain multiple actual aging state estimation input characteristics and actual dynamic state estimation input characteristics;According to actual aging state estimation input characteristics, utilize aging state estimation model to estimate, obtain aging state estimation result;According to actual dynamic state estimation input characteristics, utilize dynamic state estimation model to estimate, obtain multiple preliminary state estimation results;Multiple preliminary state estimation results are fused using fusion method, obtain dynamic state estimation result;The fusion method is weighted fusion method or fusion model fusion method;The weight of the weighted fusion method is determined using the estimation result of aging state estimation model in training process.This application can improve the accuracy of energy storage battery state evaluation.

Description

Technical Field

[0001] This application relates to the field of battery state assessment, and in particular to a method, apparatus, equipment, medium, and product for assessing the state of energy storage batteries. Background Technology

[0002] To ensure the safety, reliability, and efficiency of energy storage battery systems, a reliable and advanced battery management system must be established to accurately and timely monitor battery status. Energy storage batteries exhibit various states, including State of Health (SOH), State of Charge (SOC), State of Energy (SOE), and State of Power (SOP). Accurate state estimation is crucial for ensuring the battery operates within a safe and efficient range and for determining its driving range. The SOH state is a long-term aging scale measured in cycles, while the other states are short-term dynamic performance scales measured in seconds. The internal state of an energy storage battery cannot be directly measured and must be estimated using measurable macroscopic signals through specific methods. However, due to the complex electrochemical kinetics and multi-physics coupling within the battery, it is a highly complex, dynamic, time-varying, nonlinear electrochemical system. Furthermore, there is a complex coupling relationship between the battery's SOH and dynamic performance states; therefore, achieving accurate joint estimation of multiple states remains a challenging task. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, device, medium, and product for assessing the state of energy storage batteries, which can improve the accuracy of energy storage battery state assessment.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for assessing the state of an energy storage battery, including: Obtain actual energy storage battery data; The actual energy storage battery data is preprocessed and feature extracted to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features; the aging state estimation input features include the peak feature and peak position feature of the IC curve obtained based on the ICA method; the dynamic state estimation input features include preprocessed current signal, voltage signal and temperature signal; The aging state estimation result is obtained by estimating the aging state based on the actual aging state estimation input features using an aging state estimation model; the aging state estimation model is constructed based on the Transformer model. Based on the actual dynamic state estimation input features, the dynamic state estimation model is used to estimate the state, resulting in multiple preliminary state estimation results. Each dynamic state estimation model is built based on a CNN-Transformer model, and the multiple dynamic state estimation models share a CNN network. Multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

[0005] In one embodiment, the training process of the dynamic state estimation model specifically includes: Acquire historical energy storage battery data and actual dynamic status; the historical energy storage battery data is historical dynamic operating condition data under a reference aging state. Preprocessing and feature extraction are performed on the historical dynamic operating condition data under the reference aging state to obtain the historical dynamic state estimation input features; By using the historical dynamic state estimation input features, the real dynamic state is used as a label to train the base model, thus obtaining the dynamic state estimation model.

[0006] In one embodiment, the training process of the fusion model in the fusion model fusion method specifically includes: For dynamic operating condition data under multiple aging states, multiple dynamic state estimation models are used sequentially to determine the historical preliminary state estimation results. The historical preliminary state estimation results determined by the multiple dynamic state estimation models are then spliced ​​together to obtain the spliced ​​result. Using the splicing result as input and the corresponding real dynamic state as label, the meta-model is trained to obtain the fusion model.

[0007] In one embodiment, during the training process, the aging state estimation model and the dynamic state estimation model utilize a Bayesian optimization algorithm for hyperparameter optimization, and the gradient descent algorithm used is the Adam algorithm.

[0008] In one embodiment, multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results, specifically including: The current aging state is estimated using an aging state estimation model based on actual energy storage battery data. The weights are determined based on the distance between the reference aging state and the current aging state for each dynamic state estimation model. Multiple preliminary state estimation results are fused using the weights.

[0009] In one embodiment, multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results, specifically including: Multiple preliminary state estimation models are input into a fusion model for fusion to obtain dynamic state estimation results.

[0010] Secondly, this application provides an energy storage battery state assessment device, comprising: The acquisition module is used to obtain actual energy storage battery data; The preprocessing and feature extraction module is used to preprocess and extract features from the actual energy storage battery data to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features. The aging state estimation input features include the peak features and peak position features of the IC curve obtained based on the ICA method. The dynamic state estimation input features include preprocessed current signals, voltage signals, and temperature signals. The aging state estimation module is used to estimate the aging state based on the actual aging state estimation input features using the aging state estimation model, and obtain the aging state estimation result; the aging state estimation model is built based on the Transformer model; The dynamic state estimation module is used to estimate the actual dynamic state estimation input features using a dynamic state estimation model to obtain multiple preliminary state estimation results; each of the dynamic state estimation models is built based on a CNN-Transformer model; the multiple dynamic state estimation models share a CNN network; The fusion module is used to fuse multiple preliminary state estimation results using a fusion method to obtain a dynamic state estimation result; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

[0011] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the energy storage battery state assessment method.

[0012] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned energy storage battery state assessment method.

[0013] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned energy storage battery state assessment method.

[0014] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, apparatus, device, medium, and product for assessing the state of energy storage batteries. It constructs an aging state estimation model using a Transformer model. Due to the attention mechanism of the Transformer model, it can model dependencies without considering distances in the input or output sequences, directly calculating the dependencies between any two positions, thus better capturing global information. Compared to traditional convolution and recursion operations for time series processing, it offers higher parallelism and computational efficiency, solving the accuracy and efficiency problems in time series modeling. Furthermore, it uses a CNN-Transformer model for dynamic state estimation. By adding a CNN before the Transformer network, its powerful feature extraction capabilities are utilized to automatically extract features, solving the problem of poor estimation accuracy caused by the difficulty in feature extraction due to complex and changing dynamic conditions and high real-time requirements in dynamic state estimation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is an application environment diagram of an energy storage battery state assessment method according to an embodiment of this application; Figure 2 A schematic flowchart illustrating a method for assessing the state of an energy storage battery according to an embodiment of this application; Figure 3 This is a schematic diagram of a method for assessing the state of energy storage batteries. Figure 4 A functional module diagram of an energy storage battery state assessment device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Currently, state estimation for energy storage batteries can be categorized into direct measurement methods, model-based methods, and data-driven methods. Direct measurement methods estimate the battery state directly through ampere-hour integration or by measuring characteristic parameters such as battery resistance. Model-based methods primarily involve establishing battery models (equivalent circuit models, fractional-order models, and electrochemical models) and identifying parameters. Finally, they use observation algorithms such as Kalman filters to estimate the battery state. Model-based methods are closed-loop systems and can reduce uncertainty interference through self-calibration. Data-driven methods establish machine learning or deep learning models, inputting measurable battery signals into the model for training, and then using the trained model to estimate the battery state.

[0019] Direct measurement methods have stringent requirements for experimental conditions and sensor accuracy, making them suitable only for specific environments such as laboratories. Furthermore, the ampere-hour integration method, as an open-loop approach, is severely affected by initial value bias and cumulative errors. Model-based methods suffer from the drawbacks of complex battery model building processes and the need for timely parameter updates and adjustments under different temperatures and operating conditions to maintain estimation accuracy. Data-driven methods are highly dependent on data quality and quantity for estimation accuracy and lack generalization across different operating conditions and temperatures. Additionally, the accuracy of dynamic state estimation gradually deteriorates with battery aging. As batteries age, various parameters within the battery system change, significantly impacting the estimation accuracy of model-based methods. Further optimization and adjustments complicate the model, leading to reduced efficiency. The distribution of battery charge and discharge data also varies with different aging processes and degrees, posing a challenge for data-driven methods. It is necessary to find high-quality features that effectively reflect the aging degree of batteries under different conditions.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The energy storage battery state assessment method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server. Terminal 101 can send actual energy storage battery data to server 102. Server 102 receives the actual energy storage battery data and performs preprocessing and feature extraction on the actual energy storage battery data to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features. The aging state estimation input features include peak features and peak position features of the IC curve obtained based on the ICA method. The dynamic state estimation input features include preprocessed current signals, voltage signals, and temperature signals. Based on the actual aging state estimation input features, an aging state estimation model is used to estimate the aging state to obtain an aging state estimation result. The aging state estimation model is built based on the Transformer model. Based on the actual dynamic state estimation input features, a dynamic state estimation model is used to estimate the dynamic state to obtain multiple preliminary state estimation results. Each dynamic state estimation model is built based on the CNN-Transformer model. Multiple dynamic state estimation models share a CNN network. The multiple preliminary state estimation results are fused using a fusion method to obtain a dynamic state estimation result. The fusion method is a weighted fusion method or a fusion model fusion method. The weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process. Server 102 can feed back the obtained aging state estimation results and dynamic state estimation results to terminal 101. In addition, in some embodiments, the energy storage battery state assessment can also be performed by server 102 or terminal 101 alone. For example, terminal 101 can directly perform energy storage battery state assessment based on the actual energy storage battery data to be processed, or server 102 can perform energy storage battery state assessment from the data storage system.

[0022] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 102 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0023] In one exemplary embodiment, such as Figure 2 As shown, a method for assessing the state of an energy storage battery is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the following steps are included.

[0024] Step 201: Obtain actual energy storage battery data.

[0025] Step 202: Preprocess and extract features from the actual energy storage battery data to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features; the aging state estimation input features include the peak features and peak position features of the IC curve obtained based on the ICA method; the dynamic state estimation input features include the preprocessed current signal, voltage signal and temperature signal.

[0026] Step 203: Estimate the aging state using the aging state estimation model based on the actual aging state estimation input features to obtain the aging state estimation result; the aging state estimation model is constructed based on the Transformer model.

[0027] Step 204: Based on the actual dynamic state estimation input features, estimate using a dynamic state estimation model to obtain multiple preliminary state estimation results; each dynamic state estimation model is constructed based on a CNN-Transformer model; multiple dynamic state estimation models share a CNN network.

[0028] Step 205: The multiple preliminary state estimation results are fused using a fusion method to obtain the dynamic state estimation result; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

[0029] An aging state estimation model is constructed using the Transformer model. Due to the Transformer's attention mechanism, it can model dependencies without considering distances in the input or output sequences, directly calculating the dependencies between any two positions, thus better capturing global information. Compared to traditional convolutional and recursive operations for time series processing, it offers higher parallelism and computational efficiency, solving the accuracy and efficiency problems in time series modeling. Dynamic state estimation is performed based on the CNN-Transformer model. By adding a CNN before the Transformer network, its powerful feature extraction capabilities are utilized to automatically extract features, addressing the problem of poor estimation accuracy caused by the difficulty in feature extraction in complex and ever-changing dynamic conditions and high real-time requirements.

[0030] In an exemplary embodiment, the actual energy storage battery data is preprocessed and feature extracted to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features. Data preprocessing refers to using methods such as filtering out abnormal data and fixing sampling intervals to preprocess the current, voltage, and temperature data in the energy storage battery data, thereby removing abnormal data and noise introduced by sampling.

[0031] The aging state is reflected by the State of Health (SOH) of the energy storage battery, which is a state quantity on a long time scale in units of cycles. The dynamic state includes various state quantities such as State of Charge (SOC), State of Operation (SOP), and State of Energy (SOE). Although these state quantities reflect different dynamic characteristics of the energy storage battery, the techniques for estimating them have common characteristics. They are all dynamic performance scales on a short time scale in units of seconds. When estimating them, it is necessary to deal with dynamic and complex operating conditions and extreme operating conditions, and at the same time, the real-time requirements are extremely high. Moreover, the estimation of these states is affected by the aging of the energy storage battery.

[0032] The data-driven aging state estimation model employs a deep learning model, with the Transformer model at its core. The Transformer model is a deep learning model suitable for solving sequence problems. The core mechanism of the Transformer model is self-attention, which can model dependencies without considering distances in the input or output sequences, directly calculating the dependencies between any two positions, better capturing global information, and exhibiting higher parallelism and computational efficiency.

[0033] The aging state estimation feature extraction is based on incremental capacity analysis (ICA). Incremental capacity (IC) data and curves are generated based on the incremental capacity of continuous voltage steps in the energy storage battery data. In the IC curve, the voltage plateau will be transformed into a clearly visible peak. Its peak value and peak position are strongly correlated with the micro phase transition during the battery aging process. The IC curve is often used as an auxiliary means for battery aging characterization.

[0034] The data-driven dynamic state estimation model adopts a deep learning model, which is built around the CNN-Transformer model with a convolutional neural network (CNN) as the core. CNN has a powerful feature extraction capability. When facing the complex and ever-changing dynamic conditions and high real-time requirements that are often encountered in dynamic state estimation, it can automatically extract features and use the Transformer model to capture temporal relationships.

[0035] The input format required to build a model refers to the fact that in deep learning models that process time series data, the input needs to be constructed into a time series format. First, the feature data is constructed into a two-dimensional matrix according to a certain time series length. A two-dimensional matrix is ​​regarded as a time series. Then, multiple time series are constructed into several three-dimensional tensors. A three-dimensional tensor is regarded as a batch. During training, training is performed on a batch basis, thereby effectively improving the training efficiency of the network.

[0036] In an exemplary embodiment, during the training process, the aging state estimation model and the dynamic state estimation model utilize a Bayesian optimization algorithm for hyperparameter optimization, and the gradient descent algorithm used is the Adam algorithm.

[0037] Hyperparameter optimization techniques employ the Bayesian optimization algorithm. This algorithm uses a Gaussian process to model the objective function and finds the global optimum by iteratively updating the model. It is suitable for computationally intensive tasks such as automatic hyperparameter optimization in deep learning. The gradient descent algorithm uses the Adam algorithm. Training deep learning models requires gradient descent, which is an improvement on other gradient descent algorithms. During model training, Adam dynamically adjusts the learning rate of each parameter through first- and second-moment estimation of the gradient, while introducing bias correction. This ensures that the learning rate has a defined range for each iteration, thus stabilizing the parameters and significantly improving convergence speed and accuracy.

[0038] The division of training and test sets refers to using cross-validation to train and validate the model. Data from different batteries are used as the test set, while data from the remaining batteries are used as the training set. During model training, the data from the batteries used as the training set and the dynamic state labels are used to form the training set. During model testing, the data from the batteries used as the test set are input into the trained model for state estimation. The output estimation results are compared with the actual aging state and dynamic state labels of the batteries used as the test set for error analysis to evaluate the model's performance.

[0039] In an exemplary embodiment, the training process of the dynamic state estimation model specifically includes: acquiring historical energy storage battery data and real dynamic state; the historical energy storage battery data is historical dynamic operating condition data under a reference aging state; preprocessing and feature extraction of the historical dynamic operating condition data under the reference aging state to obtain historical dynamic state estimation input features; using the historical dynamic state estimation input features, training the base model with the real dynamic state as a label to obtain the dynamic state estimation model.

[0040] In an exemplary embodiment, the training process of the fusion model in the fusion model fusion method specifically includes: using multiple dynamic state estimation models to determine the historical preliminary state estimation results for dynamic operating condition data under multiple aging states in sequence; concatenating the historical preliminary state estimation results determined by the multiple dynamic state estimation models to obtain the concatenated result; using the concatenated result as input and the corresponding real dynamic state as label, training the meta-model to obtain the fusion model.

[0041] In an exemplary instance, multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results. Specifically, this includes: estimating the current aging state using an aging state estimation model based on actual energy storage battery data; determining weights based on the distance between the reference aging state and the current aging state corresponding to each dynamic state estimation model; and fusing the multiple preliminary state estimation results using the weights.

[0042] In another exemplary embodiment, multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results. Specifically, this includes inputting multiple preliminary state estimation models into a fusion model for fusion to obtain dynamic state estimation results.

[0043] The ensemble learning method first selects multiple reference aging states. Under different reference aging states, dynamic state estimation base models are trained using corresponding dynamic operating condition data. All base models have the same model structure but different model parameters. All base models share a CNN network as the underlying feature extractor to reduce the number of parameters and further improve model efficiency. The Transformer network retains independent task-specific layers to fully learn the dynamic characteristics of the battery under different aging states. During estimation, multiple dynamic state estimation base models are estimated simultaneously. Considering the differences in practical application scenarios, two ensemble schemes for dynamic state estimation base models are proposed: Scheme 1 uses a dynamic weighting method based on aging states, determining dynamic weights based on the current aging state estimation result and the corresponding reference aging states of different base models, and determining the final dynamic state estimation result through weighted summation; Scheme 2 uses a stacking ensemble method, first constructing a meta-model based on multiple fully connected layers, then integrating the estimation results of multiple base models into meta-features, and inputting the meta-features into the meta-model for training and validation to obtain the final dynamic state estimation result.

[0044] The purpose of this application is to address the problem that changes in battery parameters and data distribution due to aging in practical applications lead to difficulties in guaranteeing estimation accuracy when using a single method for dynamic performance state estimation. The main objective is to provide a data-driven method for energy storage battery state assessment in energy storage power stations, characterized by strong practicality, high estimation accuracy, and robustness. This application can be used for multi-state joint estimation of energy storage batteries, showing diverse application prospects in energy storage power station scenarios. This combined method can achieve high-precision estimation of the aging state of energy storage batteries, and simultaneously achieve high-precision dynamic state estimation under different aging states. Firstly, energy storage battery data is acquired through various channels, including energy storage power station battery management systems, cloud-based battery monitoring platforms, manufacturer-provided data, and publicly available datasets. This data includes current, voltage, temperature, power tags, capacity tags, and energy tags. Then, the collected current, voltage, and temperature data undergo data preprocessing. Then, a state estimation model is constructed based on a data-driven approach, specifically an aging state estimation model and a dynamic state estimation model. First, the data-driven aging state estimation model is constructed by extracting features from preprocessed energy storage battery data, using these features as input. Next, a data-driven dynamic state estimation model is constructed, using current, voltage, and temperature signals collected from the energy storage battery data as input. For training and validation of the aging state and dynamic state estimation models, the preprocessed and feature-extracted data are first divided into training and test sets, respectively. The data is then constructed into the required input format for the model and input into the model for training. During training, hyperparameter optimization techniques are used to optimize the training process, and gradient descent is used to train the network. The trained model is then used to estimate the aging and dynamic states of the energy storage battery based on the test set. The preprocessing, feature extraction, and input format construction for the signals collected in the test set are the same as for the training set. Finally, an ensemble learning method is used to correct the dynamic state estimation results under different aging states based on the aging state estimation results of the energy storage battery, thereby achieving joint estimation of the aging and dynamic states.

[0045] This embodiment refers to the invention, and the overall process is as follows: Figure 3 As shown, the process mainly consists of the following four steps: data preprocessing and feature extraction; deep learning model construction; model training and validation; and joint estimation based on ensemble learning.

[0046] Step 1: Data preprocessing and feature extraction.

[0047] Taking the cyclic aging data of ternary material system energy storage batteries as an example, eight 0.74Ah pouch batteries were used and placed in a MK53 hot chamber at 40℃. Constant current charge-discharge cyclic aging tests were conducted using an 8-channel bioMPG 205 battery tester. The collected current, voltage, and temperature data were preprocessed by removing outliers and fixing the sampling interval.

[0048] When estimating aging status, the first step is to extract features using the ICA method based on the processed data. The ICA method generates IC data and curves based on the incremental capacity of continuous voltage steps in the energy storage battery data. In the IC curve, the voltage plateau will be transformed into a clearly visible peak. Its peak value and peak position are strongly correlated with the microscopic phase transitions during the battery aging process. The IC curve is often used as an auxiliary means of characterizing battery aging. The calculation of IC is as follows: .

[0049] in, Represents capacity increment, Represents the voltage step size. Represents current. Represents the time step.

[0050] The obtained IC curve is further filtered using a Savitzky Golay (SG) filter. The SG filter is a time-domain filtering method based on a local least squares polynomial approximation algorithm, which performs well in processing curves with obvious peaks and valleys. The specific principle is as follows: .

[0051] in, This represents the filtered signal. Indicates the index of the data currently being processed. Represents a neighborhood index. Represents the original signal. Let be the size of the sliding window, which is equal to 2m0 + 1. These are the coefficients of the SG filter. This is the size of the filtering window.

[0052] Further feature extraction is performed based on the processed curve, including the peak value and peak position of the IC curve: .

[0053] in, The index representing the extracted peak value. Represents voltage. Representative index The corresponding voltage data, This represents the IC value corresponding to the voltage. The voltage coordinates representing the extracted peak value. and The table index represents the adjacent voltages corresponding to the voltage. This is the derivative of the IC data corresponding to the voltage data point.

[0054] For dynamic state estimation, due to the complex and ever-changing dynamic operating conditions and the extremely high requirements for real-time performance in practical applications, model development and verification are directly based on the preprocessed raw data.

[0055] Step 2: Deep learning model construction.

[0056] The aging state estimation model employs a deep learning model, built around the Transformer model, a deep learning model suitable for solving sequence problems. The Transformer model consists of an encoder and a decoder, internally including pointwise, stacked self-attention mechanisms and fully connected layers. The core mechanism of the Transformer model is self-attention, which can model dependencies without considering distances in the input or output sequences, directly calculating dependencies between any two positions, better capturing global information, and exhibiting higher parallelism and computational efficiency. Since state estimation is a many-to-one regression problem, based on feature sequences rather than sequence-to-sequence transformations or generation to estimate battery states, only the encoder part of the Transformer is used. The encoder part of the Transformer model can effectively model time series data and capture long-term dependencies in the time series. The Transformer encoder model includes an input layer, a self-attention mechanism layer, a residual connection and normalization layer, a feedforward neural network layer, and an output layer.

[0057] First, in the input layer, the input battery feature data is positionally encoded. Positional encoding uses trigonometric functions to inject positional information into the time series, generating a unique code for each vector at each location. This code is then added to the input time series data. The trigonometric function positional encoding can be described as follows: .

[0058] .

[0059] in, Indicates location, Representing feature dimension, Encoding the position of trigonometric functions, This indicates that the position code is for an even position. This indicates that the position code is for an odd-numbered position.

[0060] Then the input timing data and position code are added together: .

[0061] .

[0062] in, Given the input sequence, the input to the m-th encoder layer is... The output is , The length of the time series. The final output of the input layer, superscript This indicates that the variable is an input sequence. Encoding the position of trigonometric functions, For input feature time series data, subscript This indicates that the variable is the time series data of the input features of the input layer.

[0063] Then, the self-attention mechanism layer is entered. For a single head, the input sequence is transformed linearly to generate a query (Q), a key (K), and a value (V): .

[0064] in, , and These are the linear transformation weight matrices, and are the learnable parameters, with superscripts... , and These indicate that the linear transformation weight matrix corresponds to the linear transformation weight matrix for the query, key, and value, respectively. This is the input to the m-th encoder layer.

[0065] Then, the scaling dot product attention is calculated, which involves calculating the dot product of each dot with all others, dividing each result by a normalization factor, and then processing it using the softmax activation function. .

[0066] in, , , These are query, key, and value matrices, respectively. This represents the softmax activation function, used to normalize values ​​to a range of weights between 0 and 1, where the sum of the weights equals 1. As the key dimension, Scaling factor This is the result of scaling the dot product attention calculation.

[0067] The multi-head attention mechanism is built upon the scaled dot product attention mechanism described above. It concatenates the attention outputs of each head and performs a linear transformation by multiplying them by the projection matrix, ultimately yielding the output of the multi-head attention mechanism. .

[0068] .

[0069] in, This represents a splicing operation. Represents the projection matrix. To ultimately obtain the output of the multi-head attention mechanism, The result of splicing the attention outputs of each head.

[0070] Then, we proceed to the residual connection and normalization layer: .

[0071] The layer normalization formula is as follows: .

[0072] In the above formula, And β are learnable parameters. The mean, To prevent small constants with a denominator of 0, The variance of the input. For the output of the residual connection and normalization layer, The input to the representation layer is normalized. The formula for calculating layer normalization is given.

[0073] Then, it enters the feedforward neural network for nonlinear transformation: .

[0074] in, This represents the activation function. This represents the output of the feedforward neural network layer. and These are the weights and biases of the first layer of the feedforward neural network. and These are the weights and biases of the second layer of the feedforward neural network, respectively.

[0075] The feedforward neural network is followed by another residual connection and a normalization layer, and the final output is set to... .

[0076] Finally, the input is fed into the output layer, which consists of fully connected layers. The output layer outputs the state estimation results, and a sigmoid function is added as the activation function in the fully connected layers. .

[0077] in, For the aging state estimation results, the subscript is... This indicates that the variable represents the result of an aging state estimate. It is the Sigmoid activation function. and These are the weights and biases of the output layer in a Transformer, with subscripts... This indicates that the variable is a variable of the output layer in the Transformer.

[0078] The dynamic state estimation model is built around a CNN-Transformer model with a CNN pre-built. CNNs have powerful feature extraction capabilities, automatically extracting features to address the complex and ever-changing dynamic conditions and high real-time requirements often encountered in dynamic state estimation, and using the Transformer model to capture temporal relationships. A CNN consists of three layers: an input layer, hidden layers, and an output layer. The input layer transmits raw data to the network; the hidden layers include convolutional layers and max-pooling layers; the output layer is generally composed of fully connected layers, responsible for generating the output for the next layer. The convolutional layer is the core of the CNN. It extracts local features from higher-level inputs and passes all information to lower layers to obtain more complex and abstract features. Each convolutional layer and pooling layer includes a ReLU function as an activation function to accelerate model convergence. Max-pooling layers are added after the convolutional layers for data dimensionality reduction, further reducing the computational burden on the model. CNNs, with their local connectivity and weight sharing, can significantly reduce the number of model parameters, accelerate training, and improve generalization performance. One-dimensional CNNs are mainly used for time series problems. The input data is fed into the input layer and then passed through the convolutional layer for convolution calculations. .

[0079] in, The original input sequence, index This indicates that the variable is the original input sequence. Represents the activation function. and These represent the weights and biases of the convolutional layer, respectively. The size of the convolution kernel. Represents the output of the convolutional layer, subscript This indicates that the variable represents the output of the convolutional layer.

[0080] Then, the data enters the pooling layer for further dimensionality reduction: .

[0081] in, This is the data after dimensionality reduction.

[0082] Finally, the data is output through the output layer, which is typically composed of fully connected layers: .

[0083] in, It is the Sigmoid activation function. This is the final output of the CNN. and These represent the weights and biases of the output layer in a CNN, with subscripts... This indicates that the variable is a variable of the output layer in the CNN.

[0084] After feature extraction via CNN, the output is fed into the Transformer model to capture and model temporal relationships, ultimately outputting dynamic state estimation results. The detailed formula for the Transformer has been given above and will not be repeated here; a simplified formula is as follows: .

[0085] in, For dynamic state estimation results, subscript This indicates that the variable is a result of dynamic state estimation.

[0086] Step 3: Model training and validation.

[0087] The model training and validation employed cross-validation. Data from different batteries were used as the test set, while data from the remaining batteries served as the training set. For aging state estimation, during model training, data preprocessing and feature extraction were performed on the training set data, and then the extracted feature data and capacity labels were used to construct the training set. During model testing, the preprocessing and feature extraction of the measurable signals in the test set were the same as for the training set. After model training, the feature data from the batteries used as the test set were input into the trained model for aging state estimation. For dynamic state estimation, raw current, voltage, and temperature data were directly used as input. The model was trained using data from the training set and dynamic state labels, and the data from the test set was input into the trained model for dynamic state estimation. The output estimation results were compared with the actual aging state and dynamic state labels of the batteries used as the test set for error analysis to evaluate the model's performance. The model training utilized Bayesian optimization techniques for automatic hyperparameter optimization. Bayesian optimization is a sequential optimization method based on a probabilistic model, particularly suitable for computationally expensive black-box function optimization, which continuously seeks the optimal value of the objective function through a Gaussian process. .

[0088] in, For hyperparameters, For the optimal value of the hyperparameter, ( ) represents minimizing the cost function. This represents the range of hyperparameter optimization.

[0089] The hyperparameters to be optimized include the learning rate, the size and number of convolutional kernels, and the number of neurons in each fully connected layer. The model is trained using the Adam gradient descent algorithm. Adam is an adaptive learning rate optimization algorithm, which, compared to traditional gradient descent, optimizes training speed and performance by updating the learning rate during training. During training, the Adam algorithm dynamically adjusts the learning rate of each parameter through first-moment and second-moment predictions of the gradient, while introducing bias correction to ensure that the learning rate has a defined range for each iteration, thus stabilizing the parameters and significantly improving convergence speed and accuracy. Its principle is as follows: Updated in time: .

[0090] in, This represents the number of iterations.

[0091] Gradient calculation: .

[0092] in, The parameters from the previous iteration, For loss function, For gradient calculation, The gradient of the parameters on the objective function.

[0093] Prediction of the first moment of the gradient: .

[0094] in, This is the first-order moment exponential decay rate, used to control the smoothness of momentum, and is typically taken as 0.9. The gradient is the exponential moving average, i.e., the first moment, where the initial value is... , It is the first moment of the previous iteration.

[0095] Prediction of the second moment of the gradient: .

[0096] in, It is the exponential moving average of the squared gradient, i.e., the second moment, where the initial value is... , The first moment of the previous iteration, The second-order moment exponential decay rate is used to control the smoothness of the adaptive learning rate, and is usually set to 0.99.

[0097] Deviation correction: .

[0098] in, and These are the corrected first and second moments, respectively.

[0099] Parameter update: .

[0100] in, The parameters for the current iteration step. For learning rate, To avoid small constants with a denominator of 0.

[0101] Step 4: Joint estimation based on ensemble learning.

[0102] The joint estimation strategy adopted in this application corrects the dynamic state estimation results based on the aging state estimation results, thereby reducing the impact of battery aging on dynamic state estimation. The collaborative strategy is based on ensemble learning, a machine learning method that combines the computational results of different weak learners to obtain the final decision. First, multiple reference aging states are selected: 80%, 85%, 90%, 95%, and 100%. For each reference aging state, corresponding dynamic state estimation base models are trained using dynamic operating condition data. All base models have the same model structure but different model parameters. All base models share a CNN network as the underlying feature extractor, which not only captures the common patterns contained in the original battery signal across tasks but also effectively reduces the number of parameters, further improving model efficiency. The Transformer network retains independent task-specific layers to fully learn the dynamic characteristics of the battery under different aging states. During estimation, multiple dynamic state estimation base models are performed simultaneously. Considering the differences in practical application scenarios, two ensemble schemes for dynamic state estimation base models are proposed: Option 1 employs a dynamic weighting method based on aging states. Dynamic weights are determined based on the current aging state estimation result and the reference aging states corresponding to different base models. The final dynamic state estimation result is determined by weighted summation; that is, the closer the current battery aging state is to the corresponding aging state during the training of a certain base model, the higher its weight. The weighting function is designed based on the softmax function. .

[0103] in This indicates the current aging state of the battery. For the first Each base model corresponds to an aging state. For the corresponding number Each weight, When calculating the sum of all distances, the index is... The base model corresponds to the aging state.

[0104] The final dynamic state estimation result is as follows: .

[0105] in, The corresponding aging state is The dynamic state estimation results of the base model, This represents the final dynamic state estimation result.

[0106] In the final verification, a random aging state within the range of 100%-80% was selected, and its dynamic operating conditions were estimated using the above method. The estimation results were then compared with the actual values ​​to evaluate the estimation effectiveness. This scheme has clear physical interpretability and low computational cost, making it suitable for application in automotive environments and scenarios with high safety requirements.

[0107] Option 2 employs the Stacking ensemble method. Stacking is an advanced ensemble learning approach that combines the estimation results of multiple base models by training a meta-model. A meta-model is constructed based on multiple fully connected layers. The outputs of the multiple base models are then integrated into meta-features, which are input into the meta-model for training and validation to obtain the final dynamic state estimation result. The meta-model is built upon a stack of multiple fully connected layers. The computation of each fully connected layer is as follows: .

[0108] in, and These are the weights and biases of the metamodel, respectively. This is the Sigmoid activation function.

[0109] During the training and validation of the meta-model, multiple battery aging intervals are first divided based on the reference aging states corresponding to the dynamic state estimation base model: [100%, 95%], [95%, 90%], [90%, 85%], and [85%, 80%]. Then, three aging states are randomly selected from each of the different aging intervals, resulting in a total of 12 randomly selected aging states. The aging state estimation model is used to estimate the aging state of each selected data point. Based on the obtained aging state estimation results, the corresponding dynamic operating condition data is selected, and all dynamic state estimation base models are used to estimate the dynamic state of the aforementioned dynamic operating condition data. The dynamic state estimation results of all base models are concatenated as meta-features, and the corresponding real dynamic state values ​​are used as labels to construct the training set. For final validation, an aging state within the 100%-80% range is randomly selected, and its dynamic operating condition is estimated using the above method. The estimation results are then compared with the real values ​​to evaluate the estimation effect. This data-driven approach is suitable for cloud-based big data platforms and the development of novel batteries with unknown aging mechanisms.

[0110] This application constructs an aging state estimation model based on the Transformer model. Based on the self-attention mechanism, it can model dependencies without considering the distance in the input or output sequences, directly calculate the dependencies between any two positions, better capture global information, and has higher parallelism and computational efficiency compared with traditional convolution and recursion operations for processing time series, thus solving the problems of accuracy and efficiency in time series modeling.

[0111] This application presents a dynamic state estimation model based on the CNN-Transformer model. By adding a CNN before the Transformer network, the powerful feature extraction capability is used to automatically extract features, which solves the problem of poor estimation accuracy caused by the difficulty in feature extraction due to the complex and ever-changing dynamic conditions and high real-time requirements in dynamic state estimation.

[0112] This application employs ensemble learning technology for joint estimation of aging state and dynamic state. Multiple dynamic state estimation base models are constructed based on different reference aging states of the battery. Then, the base models are integrated according to the difference between the current actual aging state and the reference aging state corresponding to the base model, thereby achieving high-precision dynamic state estimation under different aging states and solving the problem of difficulty in achieving high-precision dynamic state estimation under different aging states of the battery.

[0113] In this application, all dynamic state estimation base models share a CNN network as the underlying feature extractor, while the Transformer network retains independent task-specific layers to fully learn the dynamic characteristics of the battery under different aging states. While ensuring that the model can fully capture the dynamic characteristics of the energy storage battery under different aging states, the number of parameters is reduced and the model efficiency is improved. This solves the problem that in cross-task problems, it is usually difficult to fully utilize the common patterns in the data in order to ensure the learning of different specific tasks, which leads to redundant model parameters and low efficiency.

[0114] This application proposes two parallel approaches for integrating the base model: a dynamic weighting method based on aging state and a stacking integration method. The two methods are suitable for different scenarios. The dynamic weighting method based on aging state has clear physical interpretability and low computational cost, making it suitable for automotive environments and scenarios with high safety requirements. The stacking integration method, on the other hand, is based on data-driven construction and is suitable for cloud big data platforms and the research and development of novel batteries with unknown aging mechanisms. This solves the problem that specific methods are difficult to fully adapt to specific scenarios due to the complexity and variability of application scenarios in practical applications.

[0115] Based on the same inventive concept, this application also provides an energy storage battery state assessment device for implementing the above-described energy storage battery state assessment method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the energy storage battery state assessment device provided below can be found in the limitations of the energy storage battery state assessment method described above, and will not be repeated here.

[0116] In one exemplary embodiment, such as Figure 4 As shown, a battery state assessment device is provided, comprising: The acquisition module is used to obtain actual energy storage battery data.

[0117] The preprocessing and feature extraction module is used to preprocess and extract features from the actual energy storage battery data to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features. The aging state estimation input features include the peak features and peak position features of the IC curve obtained based on the ICA method. The dynamic state estimation input features include the preprocessed current signal, voltage signal, and temperature signal.

[0118] The aging state estimation module is used to estimate the aging state based on the actual aging state estimation input features using the aging state estimation model to obtain the aging state estimation result; the aging state estimation model is built based on the Transformer model.

[0119] The dynamic state estimation module is used to estimate the actual dynamic state estimation input features using a dynamic state estimation model to obtain multiple preliminary state estimation results; each of the dynamic state estimation models is built based on a CNN-Transformer model; the multiple dynamic state estimation models share a CNN network.

[0120] The fusion module is used to fuse multiple preliminary state estimation results using a fusion method to obtain a dynamic state estimation result; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

[0121] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores energy storage battery state assessment data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an energy storage battery state assessment method.

[0122] Those skilled in the art will understand that Figure 5 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method embodiments.

[0123] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0124] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method embodiments.

[0125] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0126] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0127] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0128] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assessing the state of an energy storage battery, characterized in that, The energy storage battery status assessment method includes: Obtain actual energy storage battery data; The actual energy storage battery data is preprocessed and feature extracted to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features; the aging state estimation input features include the peak feature and peak position feature of the IC curve obtained based on the ICA method; the dynamic state estimation input features include preprocessed current signal, voltage signal and temperature signal; The aging state estimation result is obtained by estimating the aging state based on the actual aging state estimation input features using an aging state estimation model; the aging state estimation model is constructed based on the Transformer model. Based on the actual dynamic state estimation input features, the dynamic state estimation model is used to estimate the state, resulting in multiple preliminary state estimation results. Each dynamic state estimation model is built based on a CNN-Transformer model, and the multiple dynamic state estimation models share a CNN network. Multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

2. The energy storage battery state assessment method according to claim 1, characterized in that, The training process of the dynamic state estimation model specifically includes: Acquire historical energy storage battery data and actual dynamic status; the historical energy storage battery data is historical dynamic operating condition data under a reference aging state. Preprocessing and feature extraction are performed on the historical dynamic operating condition data under the reference aging state to obtain the historical dynamic state estimation input features; By using the historical dynamic state estimation input features, the real dynamic state is used as a label to train the base model, thus obtaining the dynamic state estimation model.

3. The energy storage battery state assessment method according to claim 1, characterized in that, The training process of the fusion model in the fusion model fusion method specifically includes: For dynamic operating condition data under multiple aging states, multiple dynamic state estimation models are used sequentially to determine the historical preliminary state estimation results. The historical preliminary state estimation results determined by the multiple dynamic state estimation models are then spliced ​​together to obtain the spliced ​​result. Using the splicing result as input and the corresponding real dynamic state as label, the meta-model is trained to obtain the fusion model.

4. The energy storage battery state assessment method according to claim 1, characterized in that, During the training process, the aging state estimation model and the dynamic state estimation model use Bayesian optimization algorithm to optimize hyperparameters, and the gradient descent algorithm used is the Adam algorithm.

5. The energy storage battery state assessment method according to claim 1, characterized in that, Multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results, specifically including: The current aging state is estimated using an aging state estimation model based on actual energy storage battery data. The weights are determined based on the distance between the reference aging state and the current aging state for each dynamic state estimation model. Multiple preliminary state estimation results are fused using the weights.

6. The energy storage battery state assessment method according to claim 1, characterized in that, Multiple preliminary state estimation results are fused using a fusion method to obtain dynamic state estimation results, specifically including: Multiple preliminary state estimation models are input into a fusion model for fusion to obtain dynamic state estimation results.

7. A device for assessing the state of an energy storage battery, characterized in that, The energy storage battery status assessment device includes: The acquisition module is used to obtain actual energy storage battery data; The preprocessing and feature extraction module is used to preprocess and extract features from the actual energy storage battery data to obtain multiple actual aging state estimation input features and actual dynamic state estimation input features. The aging state estimation input features include the peak features and peak position features of the IC curve obtained based on the ICA method. The dynamic state estimation input features include preprocessed current signals, voltage signals, and temperature signals. The aging state estimation module is used to estimate the aging state based on the actual aging state estimation input features using the aging state estimation model, and obtain the aging state estimation result; the aging state estimation model is built based on the Transformer model; The dynamic state estimation module is used to estimate the actual dynamic state estimation input features using a dynamic state estimation model to obtain multiple preliminary state estimation results; each of the dynamic state estimation models is built based on a CNN-Transformer model; the multiple dynamic state estimation models share a CNN network; The fusion module is used to fuse multiple preliminary state estimation results using a fusion method to obtain a dynamic state estimation result; the fusion method is a weighted fusion method or a fusion model fusion method; the weights of the weighted fusion method are determined using the estimation results of the aging state estimation model during the training process.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the energy storage battery state assessment method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the energy storage battery state assessment method according to any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the energy storage battery state assessment method according to any one of claims 1-6.

Images