A method for estimating state of charge of lithium-ion batteries based on ensemble learning

Abstract

The invention relates to the technical field of lithium-ion battery management, and discloses a lithium-ion battery state of charge estimation method based on ensemble learning. By fusing a CNN-Transformer network with a Kalman filter, various problems existing in the prior art are effectively overcome. The introduction of the Kalman filter can update and correct the SOC estimation value in real time, significantly reducing the influence of the accumulated error of the ampere-hour integration method in long-term operation, and its recursive property allows the model to be dynamically adjusted, reducing the dependence on the precise initial SOC value, thereby improving the accuracy of long-term estimation. Through the CNN-Transformer architecture, the effect of processing real-time data streams is achieved, and the SOC estimation can be performed without the battery being stationary, so that it is not affected by the battery hysteresis effect. Because it relies on real-time data rather than static open-circuit voltage measurement, more accurate SOC estimation is provided in real-time applications.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery management technology, and in particular to a lithium-ion battery charge state estimation method based on ensemble learning. Background Art

[0002] As the global climate change problem becomes increasingly serious, the public's awareness of reducing carbon footprint and protecting the environment has significantly increased. In this context, electric vehicles have received widespread attention and have gradually become an important trend in the field of transportation. Lithium-ion batteries have been recognized as the main energy storage device for electric vehicles due to their high energy density, high charge and discharge efficiency, low self-discharge rate, long life and high cycle stability. Therefore, in order to ensure the safe and stable operation of lithium-ion batteries, it is very important to establish a reliable battery management system (BMS) to monitor the operating status of the battery; however, the open circuit voltage under different SOCs is nonlinear and non-monotonic, and it is difficult to directly judge the SOC by voltage. In addition, due to the complexity of the actual operating conditions of the vehicle, the key signals of the power battery system will be interfered by noise, which will affect the accuracy of SOC estimation; at the same time, the battery will gradually age during the cycle, which will lead to an increase in the error of SOC estimation.

[0003] Currently, there are several main methods used in the field of battery SOC estimation, namely coulomb counting method, open circuit voltage method, model-based method and data-driven method.

[0004] The coulomb counting method estimates the SOC of the battery by measuring the charge and discharge current of the battery within a certain period of time and performing time integration. The advantages of this method are that it is relatively simple to implement and low in cost. However, the coulomb counting method also has obvious disadvantages. The main disadvantage is that with the increase of usage time, the accuracy of SOC estimation will gradually decrease due to factors such as the accumulation of current measurement errors and battery aging. The open circuit voltage method estimates the SOC of the battery by measuring the voltage (Open Circuit Voltage, OCV) of the battery in a static state. It is based on the certain corresponding relationship between the OCV and SOC of the battery, and determines the current SOC value by looking up a pre-established OCV-SOC mapping table or curve. The advantages of this method include simple operation, low cost, and the ability to provide higher estimation accuracy after the battery has been fully stationary. However, the disadvantage of the open circuit voltage method is that it relies on the battery to be able to reach a true open circuit state, which is difficult to achieve in practical applications.

[0005] The model-based method estimates the battery SOC by constructing a combination of one or more circuit elements (such as resistors, capacitors) to simulate the electrochemical behavior of the battery, and estimates the SOC through an observer after systematic identification of the circuit model. The advantage of this method is that it can better reflect the dynamic response of the battery under different working conditions. However, the disadvantage of this method is that the accuracy of the model depends on the quality of parameter identification and the adequacy of the data, and complex models may lead to large computational complexity.

[0006] The data-driven method for estimating battery SOC mainly relies on a large amount of battery operation data. By learning the battery charging and discharging behavior and SOC change law from historical data, a nonlinear relationship between battery operation data and SOC is established. The advantages of this method include the ability to process high-dimensional data, capture complex nonlinear relationships, and not being restricted by the battery's physical and chemical models. Neural network models are usually combined with other methods to improve the accuracy and robustness of SOC estimation. However, the disadvantage of the data-driven method is that a large amount of labeled data is required for training, and the data quality directly affects the model performance; the generalization ability of the model depends on the diversity and representativeness of the training data; the model may perform poorly under battery aging or unseen working conditions; in addition, the complexity of the algorithm may lead to high computational costs, and the model's interpretability is poor, making it difficult to intuitively understand its decision-making process. Summary of the invention

[0007] The purpose of the present invention is to provide a lithium-ion battery state of charge estimation method based on ensemble learning to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A lithium-ion battery state of charge estimation method based on ensemble learning, comprising:

[0010] Preprocessing of time series: acquiring the original data set at a preset acquisition frequency through the charging and discharging position and normalizing it, and constructing the time series of the normalized data set to input into the neural network;

[0011] Data-driven model construction, using convolutional neural networks to extract features from the input data set, and training and learning local features and complex abstract representations in the data set based on a hierarchical structure to serve as effective feature input for the SOC prediction task;

[0012] The Kalman filter is constructed to obtain a smooth output effect by fusing the estimated SOC of the convolutional neural network with the SOC calculated by the ampere-hour integration method;

[0013] Model integration and fusion: Through integrated learning theory, the SOC estimation model is trained at fixed temperature nodes in multiple temperature scenarios for weighted fusion to achieve the SOC model in any temperature scenario;

[0014] The evaluation index setting and result verification are as follows: the model performance is evaluated by setting the evaluation index, and the evaluation index includes the error value and the root mean square error value.

[0015] As a further solution of the present invention: the preprocessing step of the time series specifically includes:

[0016] Obtain the original data collected and sent back by the data acquisition sensor on the host side of the charging and discharging equipment, and record the voltage, current, temperature and capacity data of the SOC at different capacity steps under different temperature discharge conditions to establish the original data set;

[0017] By using the maximum and minimum normalization method, the data in the original data set is mapped based on the maximum and minimum values ​​to compress it into the interval [0,1]. The normalization equation is characterized as follows: , where x * is the normalized value, x min is the minimum value in the sequence, x max is the maximum value in the sequence, x is the original data, and takes the voltage, current and temperature values ​​respectively;

[0018] The normalized data set is batched and windowed through the time series building module to obtain several data batches, which are input into the neural network one by one.

[0019] As a further solution of the present invention: it is characterized in that in the step of extracting features from the input data set through a convolutional neural network, the convolution operation formula is characterized as follows:

[0020] ,

[0021] Among them, X is used to represent the input data set array, Y is the output of the convolution layer, ReLU represents the activation function used in the convolution process, W represents the convolution kernel weight, b represents the bias of the convolution kernel, and W and b are calculated and updated through the back propagation process of the neural network.

[0022] As a further solution of the present invention: the step of constructing the data-driven model also includes a single attention calculation process, including:

[0023] The input sequence X is converted into an embedding vector E through the embedding layer, represented as: , where W e is the embedding matrix, b e is the embedding bias;

[0024] Adding the positional encoding to the embedding vector yields E', which can be represented as: , where P is the position encoding vector;

[0025] Calculate the self-attention output SelfAttention(Q,K,V), represented as:

[0026] ,

[0027] Among them, softmax is the activation function, Wq, Wk, Wv are learnable weight matrices;

[0028] The output of the self-attention layer SelfAttention(Q,K,V) is residually connected to the input E' and layer normalization is applied, which is represented as:

[0029] ;

[0030] The output after residual connection is passed through the feedforward neural network and represented as:

[0031] ,

[0032] Among them, W1, W2 are the weight matrices of the feedforward neural network, b1 and b2 are biases, which are updated and calculated through the back propagation process of the neural network;

[0033] The output of the feedforward neural network is again residually connected, and layer normalization is applied to obtain the final encoder layer output, which is represented as:

[0034] .

[0035] As a further solution of the present invention: in the step of constructing the Kalman filter, the step of calculating the SOC by the ampere-hour integration method includes:

[0036] A first-order equivalent model is established for the battery. This model can reflect the dynamic response of the battery. Kirchhoff's law shows that:

[0037] ,

[0038] Among them, Uoc is the open circuit voltage, R0 is the ohmic internal resistance, Rp is the polarization internal resistance, Cp is the polarization capacitance, Up is the voltage across the polarization capacitance, and U is the operating voltage of the battery;

[0039] The change in SOC calculated by the ampere-hour integration method is:

[0040] ,

[0041] Where C0 is the initial capacity of the battery, and k is the loss coefficient;

[0042] Assume that the input u of the model is the working current of the power battery, the output y is the working voltage of the power battery, and the two state variables are selected as SOC and polarization voltage Up. The standard state space expression is:

[0043] ,

[0044] ,

[0045] Among them, ω k and k represent process noise and observation noise respectively. The coefficient matrices in the state equation and observation equation are obtained by parameter identification of the battery.

[0046] As a further solution of the present invention: the step of training the SOC estimation model at a fixed temperature node under multiple temperature scenarios for weighted fusion to realize the SOC model under any temperature scenario includes:

[0047] Assume that the basic models are trained at four temperature nodes respectively, and the weights b1, b2, b3 and b4 of the four basic models are calculated as follows:

[0048] ,

[0049] Among them, a1, a2, a3, and a4 are the temperature differences between the current temperature node and the fixed temperature node respectively;

[0050] Normalize the calculated weight values:

[0051] .

[0052] As a further solution of the present invention: the error value and the root mean square error value are respectively characterized as follows:

[0053] ,

[0054] Among them, SOC k is the true value of SOC at time k, SOC * k is the predicted SOC value at time k.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows: by integrating the CNN-Transformer network with the Kalman filter, various problems existing in the prior art are effectively overcome; the introduction of the Kalman filter, real-time updating and correction of the SOC estimation value, significantly reduces the impact of the cumulative error of the ampere-hour integration method in long-term operation, and its recursive nature allows the model to be dynamically adjusted, reducing the dependence on the precise initial SOC value, thereby improving the accuracy of long-term estimation; through the CNN-Transformer architecture, the effect of processing real-time data streams is achieved, and SOC estimation can be performed without the battery being stationary, so that it is not affected by the battery hysteresis effect, and because it relies on real-time data rather than static open-circuit voltage measurement, it provides more accurate SOC estimation in real-time applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The figure is a flowchart of a lithium-ion battery state of charge estimation method based on ensemble learning.

[0057] Figure 2 Schematic diagram of the CNN-Transformer network structure in a lithium-ion battery state of charge estimation method based on ensemble learning.

[0058] Figure 3 It is the first-order equivalent current diagram of the battery in a lithium-ion battery state of charge estimation method based on ensemble learning.

[0059] Figure 4 Schematic diagram of Kalman filter construction in a lithium-ion battery state of charge estimation method based on ensemble learning.

[0060] Figure 5 Schematic diagram of model integration and fusion in a lithium-ion battery state of charge estimation method based on ensemble learning.

[0061] Figure 6 This is a graph showing the training results of the CNN-Transformer base model at -7°C and 0°C in a lithium-ion battery state of charge estimation method based on ensemble learning.

[0062] Figure 7 This is a graph of the integrated learning results at -7℃ and 0℃ in a lithium-ion battery state of charge estimation method based on integrated learning.

[0063] Figure 8 This is a graph of the root mean square error results of the estimation at -7°C and 0°C in a lithium-ion battery state of charge estimation method based on ensemble learning. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] The purpose of the present invention is to propose a new SOC estimation method, which solves the above-mentioned shortcomings of the prior art by combining CNN-Transformer with Kalman filter; the method uses convolutional neural network to process raw data, captures local features, and uses self-attention mechanism to capture long-distance dependencies in data through Transformer encoder network. Specifically, three measurable variables V, I, and T in the actual operation of the vehicle are selected as neural network inputs. First, a convolutional neural network is constructed to process raw data, and local features in the data are learned through multiple convolutional layers and pooling layer operations are used to reduce the spatial dimension of the data. Then, the feature data processed by the convolutional neural network is input to the Transformer encoder network, and its self-attention mechanism is used to capture long-distance dependencies in the input data; the Kalman filter is combined to further reduce the estimation error, and the SOC estimation values ​​at different temperatures are linearly weighted through integrated learning theory to obtain the SOC estimation value at any temperature node; this method aims to improve the accuracy of SOC estimation, reduce the dependence on initial values ​​and sensor accuracy, and at the same time reduce the computational complexity and enhance the generalization ability and real-time application ability of the model.

[0066] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0067] like Figure 1 The present invention provides a method for estimating the state of charge of a lithium-ion battery based on ensemble learning, which comprises the following steps:

[0068] S10, preprocessing of the time series, obtaining the original data set at a preset acquisition frequency through the charging and discharging position and performing normalization processing, and constructing the time series of the normalized data set to input into the neural network;

[0069] S20, data-driven model construction, extracting features from the input data set through a convolutional neural network, training and learning local features and complex abstract representations in the data set based on a hierarchical structure to serve as effective feature input for the SOC prediction task;

[0070] S30, construction of Kalman filter, by fusing the estimated SOC of convolutional neural network with the SOC calculated by ampere-hour integration method to obtain a smooth output effect;

[0071] S40, model integration and fusion, through integrated learning theory, the SOC estimation model is trained at fixed temperature nodes in multiple temperature scenarios for weighted fusion to achieve the SOC model in any temperature scenario;

[0072] S50, setting of evaluation indicators and result verification, evaluating the model performance by setting evaluation indicators, wherein the evaluation indicators include error value and root mean square error value.

[0073] In this embodiment, a lithium-ion battery state of charge estimation method based on ensemble learning is provided. By integrating the CNN-Transformer network and the Kalman filter, various problems existing in the prior art are effectively overcome. The introduction of the Kalman filter, real-time updating and correction of the SOC estimation value, significantly reduces the impact of the cumulative error of the ampere-hour integration method in long-term operation, and its recursive nature allows the model to be dynamically adjusted, reducing the dependence on the precise initial SOC value, thereby improving the accuracy of long-term estimation; through the CNN-Transformer architecture, the effect of processing real-time data streams is achieved, and the SOC estimation can be performed without the battery being stationary, so that it is not affected by the battery hysteresis effect, and because it relies on real-time data rather than static open-circuit voltage measurement, it provides a more accurate SOC estimation in real-time applications; Based on these advantages, the shortcomings of the relevant prior art are summarized here, including: Cumulative error problem: The ampere-hour integration method is prone to cumulative errors in practical applications, and has high requirements on the initial SOC value and sensor accuracy. These errors will increase with the accumulation of time, thereby affecting the accuracy of the estimation; Difficulty in real-time application: The open-circuit voltage method requires electricity The open circuit voltage is measured when the battery is stationary for a long time, which is difficult in real-time applications. In addition, the battery has a hysteresis effect. The SOC corresponding to the same voltage during charging and discharging may be different, resulting in large errors in SOC calculation; model dependence: Although the model-based methods used to estimate battery SOC, such as Kalman filters and particle filters, can better resist noise interference and have low dependence on initial values, they are highly dependent on the accuracy of the battery model. If the model parameters are inaccurate, large errors will be caused; high computational complexity: data-driven methods such as neural networks, although they do not rely on high-precision battery models, require the collection of a large amount of data, which is not conducive to online real-time estimation. The training results are too dependent on sample quality and are prone to overfitting problems; it is difficult to cover multi-temperature scenarios: neural networks can usually only train models at a few fixed temperature nodes, which limits the generalization ability and accuracy of the model under different temperature conditions; when the battery works at different temperatures, its chemical properties and physical responses will change, resulting in increased complexity in SOC estimation; therefore, when the battery works at a temperature not included in the training data, existing methods may not be able to provide accurate SOC estimation.

[0074] At the same time, the present invention reduces the dependence on accurate battery models; the CNN part is responsible for extracting local features from the original data, while the Transformer encoder network captures long-distance dependencies in the time series data. This data-driven method reduces the dependence on the accuracy of model parameters and improves the flexibility and accuracy of estimation; through the carefully designed CNN-Transformer architecture, the present invention optimizes computational efficiency when extracting features; although neural networks usually require a large amount of data, the present invention linearly weights the models at different temperature nodes through ensemble learning theory, reduces the dependence on large-scale data sets, and avoids the risk of overfitting; the present invention weights the prediction results of the models trained at different temperature nodes by applying ensemble learning theory, so that the SOC estimation value can adapt to any temperature scenario; this strategy makes the model not only effective at the temperature considered during training, but also generalizes to unseen temperature conditions, thereby improving the applicability and accuracy of the model.

[0075] As another preferred embodiment of the present invention, the time series preprocessing step specifically includes:

[0076] Obtain the original data collected and sent back by the data acquisition sensor on the host side of the charging and discharging equipment, and record the voltage, current, temperature and capacity data of the SOC at different capacity steps under different temperature discharge conditions to establish the original data set;

[0077] By using the maximum and minimum normalization method, the data in the original data set is mapped based on the maximum and minimum values ​​to compress it into the interval [0,1]. The normalization equation is characterized as follows: , where x * is the normalized value, x min is the minimum value in the sequence, x max is the maximum value in the sequence, x is the original data, and takes the voltage, current and temperature values ​​respectively;

[0078] The normalized data set is batched and windowed through the time series building module to obtain several data batches, which are input into the neural network one by one.

[0079] In this embodiment, it specifically includes a data acquisition module, a normalization module and a time series construction module. The data acquisition module is responsible for collecting raw data from the upper computer of the charging and discharging equipment, and retains the voltage, current, temperature and capacity data of SOC from 100% to 0% under the discharge conditions at different temperatures to provide the original data set for other modules. The time series construction module is responsible for batching and windowing the normalized data set and processing it into a form that can be input to the neural network. Specifically, the raw data is divided into 1672 batches, and the neural network processes one batch of data at a time. The length of the sliding window within each batch is set to 360, and the 361st feature value is predicted by the input features of the first 360 time points, and the prediction of the entire sequence is realized as the window slides.

[0080] As another preferred embodiment of the present invention, in the step of extracting features from the input data set through a convolutional neural network, the convolution operation formula is represented as:

[0081] ,

[0082] Among them, X is used to represent the input data set array, Y is the output of the convolution layer, ReLU represents the activation function used in the convolution process, W represents the convolution kernel weight, b represents the bias of the convolution kernel, and W and b are calculated and updated through the back propagation process of the neural network.

[0083] Furthermore, the step of constructing the data-driven model also includes a single attention calculation process, including:

[0084] The input sequence X is converted into an embedding vector E through the embedding layer, represented as: , where W e is the embedding matrix, b e is the embedding bias;

[0085] Adding the positional encoding to the embedding vector yields E', which can be represented as: , where P is the position encoding vector;

[0086] Calculate the self-attention output SelfAttention(Q,K,V), represented as:

[0087] ,

[0088] Among them, softmax is the activation function, Wq, Wk, Wv are learnable weight matrices;

[0089] The output of the self-attention layer SelfAttention(Q,K,V) is residually connected to the input E' and layer normalization is applied, which is represented as:

[0090] ;

[0091] The output after residual connection is passed through the feedforward neural network and represented as:

[0092] ,

[0093] Among them, W1, W2 are the weight matrices of the feedforward neural network, b1 and b2 are biases, which are updated and calculated through the back propagation process of the neural network;

[0094] The output of the feedforward neural network is again residually connected, and layer normalization is applied to obtain the final encoder layer output, which is represented as:

[0095] .

[0096] In this embodiment, combined with the attached Figure 2 , the digital drive model construction specifically includes a convolutional neural network module and a Transformer network module. The convolutional neural network module is responsible for feature extraction of the input data, and automatically learns the local features and complex abstract representations in the data through its hierarchical structure, thereby providing effective feature input for subsequent SOC prediction tasks; the Transformer network module is responsible for further deep feature extraction and global context relationship modeling of the data after convolution processing, and captures the long-distance dependencies in the data through the self-attention mechanism, thereby enhancing the model's ability to understand the overall structure of the input data, thereby improving the prediction accuracy and generalization ability of the final task. The present invention only needs to use the encoder part of the Transformer network. In view of the fact that the Transformer network adopts a multi-head attention mechanism, each attention head will process the data in parallel.

[0097] like Figure 3 , Figure 4 As shown, as another preferred embodiment of the present invention, in the step of constructing the Kalman filter, the step of calculating the SOC by the ampere-hour integration method includes:

[0098] A first-order equivalent model is established for the battery. This model can reflect the dynamic response of the battery. Kirchhoff's law shows that:

[0099] ,

[0100] Among them, Uoc is the open circuit voltage, R0 is the ohmic internal resistance, Rp is the polarization internal resistance, Cp is the polarization capacitance, Up is the voltage across the polarization capacitance, and U is the operating voltage of the battery;

[0101] The change in SOC calculated by the ampere-hour integration method is:

[0102] ,

[0103] Where C0 is the initial capacity of the battery, and k is the loss coefficient;

[0104] Assume that the input u of the model is the working current of the power battery, the output y is the working voltage of the power battery, and the two state variables are selected as SOC and polarization voltage Up. The standard state space expression is:

[0105] ,

[0106] ,

[0107] Among them, ω k and k represent process noise and observation noise respectively. The coefficient matrices in the state equation and observation equation are obtained by parameter identification of the battery.

[0108] In this embodiment, after the equivalent circuit model is established and the SOC is calculated using the ampere-hour integration method, the Kalman filtering process is performed; it is divided into two steps: time update and measurement update:

[0109] Time update:

[0110] ,

[0111] Measurement Update:

[0112] ,

[0113] Among them, x k / k-1 is the prior state estimate at time k, which is a prediction of the state at time k based on the optimal estimate at time k-1, without considering the observed data at time k. k-1 / k-1 is the posterior estimate at time point k-1, which is the optimal estimate after considering the observed data at time point k-1. k / k is the posterior estimate at time point k, which is the optimal estimate after considering the observed data at time point k, x represents the state estimate, P represents the covariance of the estimation error, A is the state transfer matrix, B is the control input matrix, u is the control variable, Q is the process noise covariance, C is the observation matrix, y is the observation vector, R is the observation noise covariance, and K is the Kalman gain.

[0114] For the SOC estimation problem, the state vector x = SOC, and the measurement vector is the SOC output of the CNN-Transformer network, so the state space expression established is as follows:

[0115] ,

[0116] Among them, I k-1 is the current at time k-1, Cn is the standard capacity of the battery, ω is the process noise, and υ is the measurement noise.

[0117] like Figure 5 As shown, as another preferred embodiment of the present invention, the step of training the SOC estimation model at a fixed temperature node under multiple temperature scenarios for weighted fusion to realize the SOC model under any temperature scenario includes:

[0118] Assume that the basic models are trained at four temperature nodes respectively, and the weights b1, b2, b3 and b4 of the four basic models are calculated as follows:

[0119] ,

[0120] Among them, a1, a2, a3, and a4 are the temperature differences between the current temperature node and the fixed temperature node respectively;

[0121] Normalize the calculated weight values:

[0122] .

[0123] In this embodiment, combined with the attached Figure 5 The fourth step of the proposed SOC estimation method is model integration and fusion. In this step, the integrated learning theory is applied to realize the SOC estimation in any temperature scenario by weighted fusion of the SOC estimation models trained at fixed temperature nodes in multiple temperature scenarios. The specific implementation is divided into three steps:

[0124] The first step of the model integration and fusion module is to construct a base model, with the lowest root mean square error of the Transformer at a fixed temperature node being used as the standard for retaining the base model.

[0125] The second step of the model integration and fusion module is weight determination, which determines the distance between the current temperature and other fixed temperature nodes, and then determines the weight value and normalizes it. Assuming that the temperature of the battery to be estimated is x°C, the basic model is trained at four temperature nodes respectively.

[0126] Step 3 of the model integration and fusion module is to weight the calculation results of the fixed temperature nodes to obtain the SOC estimation value at the current temperature. Assume that the models trained under the fixed temperature nodes are SOC x1℃ , SOC x2℃ , SOC x3℃ , SOC x4℃ , then the estimated SOC value of the current temperature node is:

[0127] .

[0128] As another preferred embodiment of the present invention, the error value and the root mean square error value are respectively characterized as follows:

[0129] ,

[0130] Among them, SOC k is the true value of SOC at time k, SOC * k is the predicted SOC value at time k.

[0131] In this embodiment, for the error value and the root mean square error value (RMSE), the return value of RMSE is a numerical value, and the smaller its value is, the better the prediction effect of the entire sequence is. At the same time, in order to verify the lithium-ion battery state of charge estimation method based on ensemble learning proposed in the present invention, verification was carried out on battery data sets at three temperatures of -20°C, -7°C and 0°C; as described in the fourth step of model integration and fusion of the present invention, base models were trained at these three temperatures, and the base model was retained with the lowest root mean square error as the standard; Figure 6 As shown in the figure, they are the base model result diagrams at -20℃, -7℃ and 0℃ respectively. It can be seen from the result diagram of the base model that the constructed neural network can extract the hidden features of the input data well, and the fitting result of SOC is good; Figure 7 As shown in the figure, the input data of three separate temperature nodes, -20℃, -7℃ and 0℃, are estimated by ensemble learning at two other temperature nodes. It can be seen from the image that even if the model at the other two temperature nodes is used for estimation, a good estimation effect is still achieved. The estimated root mean square error is shown in Figure 8 As shown, the RMSE of the base model is stable at 0.034~0.042, and the RMSE of the ensemble learning is stable at 0.031~0.044; the purpose of the present invention is to realize SOC estimation in multi-temperature scenarios, and the data under the fixed temperature node are used for verification, so the estimation result of the ensemble learning is lower than the RMSE of the base model with the best training effect at the current temperature; at the same time, after the introduction of Kalman filtering, the estimation accuracy of SOC is greatly reduced at the RMSE level compared with no filter, which verifies the effectiveness of the method proposed in the present invention.

[0132] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0133] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the disclosure in the specification and examples. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0134] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A lithium-ion battery state of charge estimation method based on ensemble learning, characterized in that: Include: Preprocessing of time series: acquiring the original data set at a preset acquisition frequency through the charging and discharging position and normalizing it, and constructing the time series of the normalized data set to input into the neural network; Data-driven model construction, using convolutional neural networks to extract features from the input data set, and training and learning local features and complex abstract representations in the data set based on a hierarchical structure to serve as effective feature input for the SOC prediction task; The Kalman filter is constructed to obtain a smooth output effect by fusing the estimated SOC of the convolutional neural network with the SOC calculated by the ampere-hour integration method; Model integration and fusion: Through integrated learning theory, the SOC estimation model is trained at fixed temperature nodes in multiple temperature scenarios for weighted fusion to achieve the SOC model in any temperature scenario; The evaluation index setting and result verification are as follows: the model performance is evaluated by setting the evaluation index, and the evaluation index includes the error value and the root mean square error value.

2. A lithium-ion battery state of charge estimation method based on ensemble learning according to claim 1, characterized in that: The time series preprocessing steps specifically include: Obtain the original data collected and sent back by the data acquisition sensor on the host side of the charging and discharging equipment, and record the voltage, current, temperature and capacity data of the SOC at different capacity steps under different temperature discharge conditions to establish the original data set; By using the maximum and minimum normalization method, the data in the original data set is mapped based on the maximum and minimum values ​​to compress it into the interval [0,1]. The normalization equation is characterized as follows: Among them, the x * is the normalized value, x min is the minimum value in the sequence, x max is the maximum value in the sequence, x is the original data, and takes the voltage, current and temperature values ​​respectively; The normalized data set is batched and windowed through the time series building module to obtain several data batches, which are input into the neural network one by one.

3. A lithium-ion battery state of charge estimation method based on ensemble learning according to claim 2, characterized in that: In the step of extracting features from the input data set by using a convolutional neural network, the convolution operation formula is represented as: ， Among them, X is used to represent the input data set array, Y is the output of the convolution layer, ReLU represents the activation function used in the convolution process, W represents the convolution kernel weight, b represents the bias of the convolution kernel, and W and b are calculated and updated through the back propagation process of the neural network.

4. The method for estimating the state of charge of a lithium-ion battery based on ensemble learning according to claim 3, characterized in that: The steps of constructing the data-driven model also include a single attention calculation process, including: The input sequence X is converted into an embedding vector E through the embedding layer, represented as: Among them, W e is the embedding matrix, b e is the embedding bias; Adding the positional encoding to the embedding vector yields E', which can be represented as: , where P is the position encoding vector; Calculate the self-attention output SelfAttention(Q,K,V), represented as: ， Among them, softmax is the activation function, Wq, Wk, Wv are learnable weight matrices; The output of the self-attention layer SelfAttention(Q,K,V) is residually connected to the input E' and layer normalization is applied, which is represented as: ； The output after residual connection is passed through the feedforward neural network and represented as: ， Among them, W1, W2 are the weight matrices of the feedforward neural network, b1 and b2 are biases, which are updated and calculated through the back propagation process of the neural network; The output of the feedforward neural network is again residually connected, and layer normalization is applied to obtain the final encoder layer output, which is represented as: 。 5. The method for estimating the state of charge of a lithium-ion battery based on ensemble learning according to claim 1, characterized in that: In the step of constructing the Kalman filter, the step of calculating the SOC by the ampere-hour integration method includes: A first-order equivalent model is established for the battery. This model can reflect the dynamic response of the battery. Kirchhoff's law shows that: ， Among them, Uoc is the open circuit voltage, R0 is the ohmic internal resistance, Rp is the polarization internal resistance, Cp is the polarization capacitance, Up is the voltage across the polarization capacitance, and U is the operating voltage of the battery; The change in SOC calculated by the ampere-hour integration method is: ， Where C0 is the initial capacity of the battery, and k is the loss coefficient; Assume that the input u of the model is the working current of the power battery, the output y is the working voltage of the power battery, and the two state variables are selected as SOC and polarization voltage Up. The standard state space expression is: ， ， Among them, ω k and k represent process noise and observation noise respectively. The coefficient matrices in the state equation and observation equation are obtained by parameter identification of the battery.

6. The method for estimating the state of charge of a lithium-ion battery based on ensemble learning according to claim 1, characterized in that: The step of training the SOC estimation model at a fixed temperature node under multiple temperature scenarios for weighted fusion to realize the SOC model under any temperature scenario includes: Assume that the basic models are trained at four temperature nodes respectively, and the weights b1, b2, b3 and b4 of the four basic models are calculated as follows: ， Among them, a1, a2, a3, and a4 are the temperature differences between the current temperature node and the fixed temperature node respectively; Normalize the calculated weight values: 。 7. The method for estimating the state of charge of a lithium-ion battery based on ensemble learning according to claim 1, characterized in that: The error value and the root mean square error value are respectively characterized as: ， Among them, SOC k is the true value of SOC at time k, SOC * k is the predicted SOC value at time k.

Images