Method for predicting SOC, SOE and BST of battery based on hybrid deep learning network

Abstract

The invention discloses a method for predicting SOC, SOE and BST of a battery based on a hybrid deep learning network, and the method comprises the steps: inputting a data sequence of current, voltage, surface temperature and charge-discharge capacity of the battery obtained under different test conditions into a constructed hybrid deep learning network for training; estimating the state of the battery by using the trained network model, and estimating the SOC, SOE and BST of the battery at the current t moment and the future moment; the deep learning network sequentially comprises an Informer network, a GRU network and a KAN network from the input side to the output side. The method provided by the invention not only can improve the prediction precision of the model, but also can predict the SOC, SOE and BST of the battery at the future moment in a multi-scale output form, and provides reliable reference information for controlling the temperature of the battery.

Description

Technical Field

[0001] The present invention relates to the technical field of battery state prediction, and in particular to a method for predicting battery SOC, SOE and BST based on a hybrid deep learning network. Background Art

[0002] As the market share of new energy vehicles continues to increase, the market's requirements for the safety performance of new energy vehicles are also getting higher and higher. As the core component of electric vehicles, the battery management system mainly provides various battery status information to provide safety protection for the operation of electric vehicle systems. The battery status mainly includes key indicators such as battery state of charge (SOC), battery state of energy (SOE), and battery state of health (SOH).

[0003] At present, the main methods for estimating battery SOC and SOE are ampere-hour integration method, Kalman filter algorithm, machine learning and deep learning. The ampere-hour integration method is greatly affected by the initial SOC value and is prone to cumulative errors. The Kalman filter algorithm is heavily dependent on the accuracy of the battery model, and the model parameters are also affected by different conditions. The model accuracy under different conditions also varies greatly. In addition, the parameter selection in the Kalman filter algorithm will also have a great impact on the estimation results of battery SOC and SOE. These complex coupling relationships also make battery state estimation very challenging. Some machine learning technologies have failed to fully learn the mapping relationship between current, voltage and temperature and battery SOC and SOE due to their simple network structure and simple machine learning algorithms.

[0004] With the development of artificial intelligence technology and computing hardware, deep learning technology has gradually become a research hotspot in the field of battery state estimation, such as Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU) neural network and Transformer network. For example, Chinese patent document CN112540298B proposes a lithium-ion battery SOC estimation method based on deep-transfer learning, which completes the battery SOC estimation between different battery data sets. However, this method only estimates a single battery state. Chinese patent document CN118444159A proposes a battery pack SOC and SOE estimation method driven by Mamba architecture to estimate the battery pack SOC and SOE. Chinese patent document CN117214755A proposes a lithium-ion battery SOC, SOE and SOH joint estimation method based on hybrid neural network to jointly estimate battery SOC, SOE and SOH.

[0005] The deep learning network inputs of the above methods are current, voltage and temperature, and no other effective features are added. Battery surface temperature (BST) also has a significant impact on battery performance, life and safety. Predicting the battery BST at future times can prevent battery thermal runaway and provide safety protection for the battery. However, none of the above methods can effectively predict the battery surface temperature. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings and defects of the prior art and to provide a method for predicting battery SOC, SOE and BST based on a hybrid deep learning network in a multi-input and multi-output form. The three states of SOC, SOE and BST can be predicted in a multi-output form, which can not only estimate the battery SOC and SOE at the same time, but also predict the battery surface temperature BST at a certain moment in the future.

[0007] A method for predicting battery SOC, SOE and BST based on a hybrid deep learning network, comprising the steps of:

[0008] Perform operating tests on the battery at different temperatures to obtain test data, extract information on current I, voltage U, and battery surface temperature BST from the test data, and calculate the battery SOC and SOE;

[0009] The ampere-hour integration method is used to integrate the current based on the final charge and discharge capacity of the previous cycle to obtain the battery's charge and discharge capacity C. n2 ;

[0010] The current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 Normalization processing forms a data set, which is divided into a training set and a test set;

[0011] Construct a deep learning network for estimating the battery SOC, SOE, and battery surface temperature BST state, which includes the Informer network, GRU network, and KAN network from the input side to the output side;

[0012] Based on current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 As input, battery SOC, SOE and BST are outputs. The deep learning network is trained using the training set and verified using the validation set to obtain a deep learning network prediction model.

[0013] The current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 As model input, the deep learning network prediction model is used to estimate the battery SOC, SOE and battery surface temperature BST at the current time t and future times.

[0014] Among them, the SOC at the current time t is calculated by the following formula:

[0015]

[0016] In the formula, SOC(t0) is the initial SOC value, SOC(t) is the SOC value at the current time t, η is the Coulomb rate, η=1, I t is the current flowing through the battery at time t, C N is the actual capacity of each working condition.

[0017] Among them, the SOE at the current time t is calculated by the following formula:

[0018]

[0019] In the formula, SOE(t0) is the initial SOE value, SOE(t) is the SOE value at the current time t, η is the Coulomb multiplier, η=1, U t is the battery voltage at time t, I t is the current flowing through the battery at time t, E N is the total available energy for each operating condition.

[0020] Among them, the charge and discharge capacity C at the current time t n2 Calculated by the following formula:

[0021]

[0022] In the formula, C n2 (t0) is the capacity value at the initial moment, which is 0 during discharge and charge, η is the Coulomb rate, η=1, I t is the current flowing through the battery at time t, C n2 (t) is the charge and discharge capacity at the current moment.

[0023] The training set includes mixed working conditions and CC-CV working conditions at various temperatures, and the test set includes standard working conditions at each temperature.

[0024] Wherein, when the deep learning network prediction model predicts,

[0025] The model input is X input =[I(t);U(t);BST(t);C n2 (t)],

[0026] Model output Y output =[SOC(t);SOE(t);BST(t+1)];

[0027] Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time tn2 , predict the SOC and SOE at the current time t, and the battery surface temperature BST at time t+1.

[0028] Wherein, when the deep learning network prediction model predicts,

[0029] Model input X input =[I(t);U(t);BST(t);C n2 (t)],

[0030] Model output Y output =[SOC(t+1);SOE(t+1);BST(t+2)],

[0031] Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 , predict the SOC and SOE at time t+1, and the battery surface temperature BST at time t+2.

[0032] Wherein, when the deep learning network prediction model predicts,

[0033] Model input X input =[I(t);U(t);BST(t);C n2 (t)],

[0034] Model output Y output =[SOC(t); SOE(t); BST(t+1); SOC(t+1); SOE(t+1); BST(t+2)];

[0035] Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 , predict the SOC and SOE at the current time t and time t+1, and the battery surface temperature BST at time t+1 and time t+2.

[0036] Among them, the current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 Normalization is performed by the following formula:

[0037] x normal =(x signal -x min ) / (x max -x min );

[0038] In the formula, x min and x max is the minimum and maximum value of the feature, x signal is the actual measured data in the data, x normalThe data is normalized.

[0039] Among them, the mean absolute error MAE and root mean square error RMSE are used to evaluate the prediction performance of the deep learning network prediction model, including:

[0040]

[0041]

[0042] In the formula, y t is the actual SOC, SOE and BST value of the battery, y * t are the predicted SOC, SOE and BST values, and M is the total number of cycles in the test set.

[0043] The present invention utilizes data at multiple temperatures for training. The trained model can realize battery state estimation at different temperatures, and outputs battery SOC, SOE and battery surface temperature BST in a multi-scale manner, which can not only realize the estimation of the current state, but also realize the state estimation at the next moment at one time.

[0044] The present invention incorporates the battery surface temperature into the output of the deep learning network model, which can effectively estimate the battery surface temperature and provide valuable reference information for battery thermal runaway.

[0045] The present invention uses the cumulative capacity obtained by the ampere-hour integration method as the characteristic of the model. The characteristic is not affected by the initial capacity value and can obtain accurate battery SOC, SOE and BST prediction results.

[0046] The hybrid deep learning network model proposed in the present invention is a hybrid network, including three neural networks: an informer network, a GRU network (Gated Recurrent Unit, GRU) network, and a KAN (Kolmogorov-Arnold Networks, KAN). The hybrid network can perform parallel calculations, which not only improves the computational efficiency and prediction performance, but also can handle the temporal relationship in the sequence data well. Using the KAN network as the network output part, the flexibility and learnability of its internal activation function can reflect the complexity of multivariate functions, making the model more interpretable.

[0047] Compared with the prior art, the method provided by the present invention can not only improve the prediction accuracy of the model, but also predict the battery SOC, SOE and battery surface temperature BST at future moments in the form of multi-scale output, providing reliable reference information for controlling the battery temperature. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1It is a flow chart of a method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to an embodiment of the present invention.

[0049] Figure 2 4 is a structural diagram of a deep learning network according to an embodiment of the present invention.

[0050] Figure 3 This is a structural diagram of the Informer network module of the deep learning network of an embodiment of the present invention.

[0051] Figure 4 This is a structural diagram of the GRU network module of the deep learning network of an embodiment of the present invention.

[0052] Figure 5 4 is a structural diagram of a KAN network module of a hybrid deep learning network according to an embodiment of the present invention.

[0053] Figure 6 The battery SOC, SOE and BST prediction results in prediction mode 1 of the embodiment of the present invention.

[0054] Figure 7 The battery SOC, SOE and BST prediction results in the prediction mode 2 of the embodiment of the present invention.

[0055] Figure 8 , Fig. 9 The battery SOC, SOE and BST prediction results in prediction mode 3 of the embodiment of the present invention are shown. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0057] The embodiment of the present invention first inputs the data sequence of the battery's current, voltage, battery surface temperature and charge and discharge capacity under different test conditions into the constructed hybrid deep learning network for training, and then uses the trained network model to estimate the battery state, that is, the battery SOC, SOE and battery surface temperature BST state.

[0058] Specifically, a battery test data set is first obtained, then the battery test data set is normalized, and the normalized data is input into the constructed deep learning network model for training. Finally, the trained deep learning network model is used to predict the three states of battery SOC, SOE and BST, thereby achieving accurate battery status at multiple moments in the future.

[0059] Figure 1Flow chart of a method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to an embodiment of the present invention. Figure 1 As shown, the specific implementation steps of the method for predicting battery SOC, SOE and BST based on a hybrid deep learning network in an embodiment of the present invention are as follows:

[0060] Step 1: Perform operating condition tests on the battery at different temperatures, extract test information such as current I, voltage U, BST, etc. from the test data, and calculate the battery SOC and SOE based on the test information;

[0061] Step 2: Using the ampere-hour integration method, the current is integrated based on the final charge and discharge capacity of the previous cycle to obtain the charge and discharge capacity C of the battery. n2 ;

[0062] Step 3: Current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 The normalization process forms a data set, which is divided into a training set and a test set.

[0063] Step 4: construct a deep learning network for estimating the battery SOC, SOE and battery surface temperature BST state, which includes an Informer network, a GRU network (Gated Recurrent Unit, GRU), and a KAN network (Kolmogorov-Arnold Networks, KAN) from the input side to the output side, such as Figure 2 As shown;

[0064] Step 5: Current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 As input, battery SOC, SOE and BST are outputs. The deep learning network is trained using the training set and verified using the validation set to obtain a deep learning network prediction model.

[0065] Step 6: The current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t are n2 As model input, the deep learning network prediction model is used to estimate the battery SOC, SOE and battery surface temperature BST at the current time t and future times.

[0066] In the above step 1, the operating condition test includes urban test conditions (Urban Dynamometer Driving Schedule, UDDS), highway fuel economy driving conditions (Highway Fuel Economy Driving Schedule, HWFET), LA92 test driving conditions (LA92 Dynamometer Driving Schedule, LA92), supplemental federal test driving conditions (Supplemental Federal Test Procedure Driving Schedule, US06), constant current - constant voltage (constant current and constant voltage, CC-CV) charging conditions.

[0067] In the specific implementation, the open source battery data set of the Institute of Automotive Research and Technology at McMaster University can be selected as the research target. The test includes the above operating condition information, as well as mixed information of various operating conditions. The test temperature includes five temperatures: -10℃, 0℃, 10℃, 25℃ and 40℃.

[0068] In the embodiment of the present application, the battery SOC is defined as the ratio of the remaining available capacity to the rated capacity, which is mainly obtained by integrating the current over time. The calculation process is as follows:

[0069]

[0070] In the formula, SOC(t0) is the initial SOC value, SOC(t) is the SOC value at the current time t, η is the Coulomb rate, η=1, I t is the current flowing through the battery at time t, C N is the actual capacity of each working condition.

[0071] In the embodiment of the present application, the battery SOE is defined as the ratio of the available energy stored in the battery to the total energy, and its expression is as follows:

[0072]

[0073] In the formula, SOE(t0) is the initial SOE value, SOE(t) is the SOE value at the current time t, η is the Coulomb multiplier, η=1, U t is the battery voltage at time t, I t is the current flowing through the battery at time t, E N is the total available energy for each operating condition.

[0074] In the above step 2, the present embodiment will add the capacity information obtained by the ampere-hour integration method as the input of the model, denoted as C n2, which can improve the prediction accuracy of the model. It is worth noting that C n2 It is not affected by the initial SOC. When the battery starts testing, it is 0. When switching working conditions, it is necessary to record the final charge and discharge capacity of the previous cycle, and then perform current integration based on this. The calculation formula is as follows:

[0075]

[0076] In the formula, C n2 (t0) is the capacity value at the initial moment, which is 0 during discharge and charge. n2 (t) is the charge and discharge capacity at the current moment.

[0077] In step 3 above, the reason for normalizing the data is that the battery test conditions are different, resulting in different values ​​of I, U, BST and C. n2 The distribution differences are large, so in order to improve the training efficiency and estimation accuracy of the model, the input features need to be normalized. In the embodiment of the present application, the normalization is performed in the following manner.

[0078] I, U, BST and C n2 Normalized to [0,1], the expression is as follows:

[0079] x normal =(x signal -x min ) / (x max -x min )(4)

[0080] In the formula, x min and x max is the minimum and maximum value in the feature, x signal is the actual measured data in the data, x normal The data is normalized.

[0081] In step 4 of the embodiment of the present application, the constructed hybrid deep learning network can not only perform parallel calculations to improve computing efficiency and prediction performance, but also can handle the temporal relationship in sequence data well. Using the KAN network as the output part of the network, the flexibility and learnability of its internal activation function can reflect the complexity of multivariate functions, making the model more interpretable.

[0082] Among them, the Informer network can improve the self-attention mechanism, reduce memory usage, and reduce computational complexity; there are mainly three improvements: (a) a probabilistic sparse self-attention mechanism is proposed, which effectively improves the time complexity and computational complexity; (b) the model uses self-attention distillation technology to halve the sequence length of each layer through convolution pooling, thereby extracting the main attention features; (c) the Informer network uses a generative decoding method, which only requires one forward step to obtain the output of a long sequence, avoiding the cumulative error propagation in the prediction stage. The Informer network is mainly composed of an encoder and a decoder.

[0083] like Figure 3 As shown, in an embodiment of the present application, the Informer network includes, from the input side to the output side, an encoder consisting of a multi-head sparse self-attention mechanism, distillation, a multi-head sparse self-attention mechanism, and a distillation connection, and a decoder including a masked multi-head probabilistic sparse self-attention mechanism and a multi-head attention mechanism; wherein, after the encoder processes the input Xen, a cascaded feature map is obtained to input the multi-head attention mechanism of the decoder, and at the same time, Xen is input to the decoder, and after processing by the decoder, the decoder outputs the corresponding feature map to connect to the fully connected layer for processing, and finally the fully connected layer outputs the Informer network processing result.

[0084] In the Informer network, the input part uses sine and cosine functions to generate time encoding to enhance the global position context and local time context of the time series input, and then maps the input features to the embedding space through a linear layer. The elements PE(pos,2i) and PE(pos,2i+1) in the sequence are calculated by sine and cosine functions respectively, and the calculation method is as follows:

[0085]

[0086] Where pos is the time step, i is the index of the feature dimension, and d is the embedding dimension.

[0087] The multi-head sparse self-attention mechanism reduces the complexity of attention calculation at each layer, thereby improving the efficiency of model prediction; the multi-head sparse self-attention mechanism can select the query matrix that has the main contribution to the attention value and perform dot product operations. The calculation formula of the attention value is simplified as follows:

[0088]

[0089] Where Q, K, and V are query, key, and value matrices, respectively; softmax is the normalized activation function; d k is the dimension of the matrix; To select only a small number of important query matrices.

[0090] In the Informer network, the encoder is a structure designed to extract long-distance dependencies of long sequence inputs. After the input data passes through the probabilistic sparse self-attention mechanism, the feature map will still remain in the sink. The Informer network uses distillation operations to reduce network parameters and continuously extract key features as the number of stacked layers increases. Distillation uses one-dimensional convolution and maximum pooling operations to reduce the dimension and compress the input data, thereby reducing memory usage. Specifically, the distillation operation can compress the data into a smaller size without losing important information, so that the data can be processed and stored more efficiently. Among them, the operation of distilling the jth layer to the j+1th layer is as follows:

[0091]

[0092] In the formula, It represents the attention value calculated by the head-probability sparse self-attention mechanism, σ is the ELU linear unit activation function, Conv1d is a one-dimensional convolution operation in the time dimension, and MaxPoo is a maximum pooling operation, which can reduce memory usage and reduce the consumption of computing resources.

[0093] In the present invention, the GRU network is an improved network of the recursive neural network, which can effectively solve the gradient vanishing problem of the network in the prediction of longer time series. The GRU network has the advantages of fewer parameters, simple structure, higher computational efficiency, etc., and is more suitable for building larger networks.

[0094] like Figure 4 As shown, the GRU network consists of two gate control units, input X t , output H t Introducing a reset gate helps capture short-term dependencies in the time series, while introducing an update gate helps capture long-term dependencies with larger step sizes in the time series.

[0095] The working process of the GRU network is described as follows:

[0096] R t =σ GRU (X t W xr +H t-1 W hr +b r )(8)

[0097] Z t =σ GRU (X t W xz +H t-1 W hz +b z)(9)

[0098]

[0099] Where W xr and W hr and and is the network weight parameter; b r 、b z and b h is the network bias parameter; σ GRU The sigmoid activation function converts the value range of each element in the reset gate and the update gate to between 0 and 1. The tanh activation function converts the value range of all elements to between [-1, 1]. is a candidate hidden state; H t is the hidden layer state at the current moment; H t-1 is the hidden state at the previous moment; Represents the product of each element in a vector.

[0100] In the embodiment of the present application, the KAN network is a neural network architecture based on mathematical theory, inspired by the Kolmogorov-Arnold representation theorem. The theorem provides a method to represent any multivariable continuous function with a finite number of single-variable function combinations. KANs are designed to improve the expressiveness and interpretability of the model, especially when dealing with complex function fitting and high-dimensional data.

[0101] The KAN network structure diagram is as follows Figure 5 As shown in the figure, in the KAN network, the input and output of each layer are transformed by the combination and summation of a series of one-dimensional functions, where a KAN layer transforms from the input vector X (l) To the output vector X (l+1) The transformation can be expressed as:

[0102]

[0103] In the formula, x l+1,j is the output of the jth node in the l+1th layer, x l,i is the input of the ith node in the lth layer, φ l,j,i is a unary activation function connecting two nodes. In matrix form, it is expressed as:

[0104]

[0105] Where φ is the function matrix corresponding to the lth KAN layer.

[0106] The general KAN network is a combination of L layers, with input x(L-1) Through the activation function matrix φ L-1 Calculate y:

[0107]

[0108] Among them, the parameterized activation function φ(x) of the KAN network is defined as follows:

[0109] φ(x)=w(b(x)+spline(x)) (15)

[0110] b(x)=silu(x)=x / (1+e -x ) (16)

[0111] In the embodiment of the present application, the data set is divided into a training set and a test set. The training set is a mixed working condition and a CC-CV working condition at various temperatures, and the test set is a standard working condition at each temperature.

[0112] In the embodiment of the present application, during model training, according to I, U, BST and C n2 The result of feature normalization is input to the model X input =[I(t);U(t);BST(t);C n2 (t)], so that the model outputs Y output =[SOC(t); SOE(t); BST(t+1)], train the constructed hybrid deep learning network to obtain the deep learning network prediction model, and finally use the trained deep learning network prediction model to estimate the battery SOC, SOE and battery surface temperature BST in real time.

[0113] In the embodiment of the present application, mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the prediction performance of the hybrid neural network model.

[0114]

[0115]

[0116] In the formula, y t is the actual SOC, SOE and BST value of the battery, y * t are the predicted SOC, SOE and BST values, and M is the total number of cycles in the test set.

[0117] In the embodiment of the present application, the model can provide three prediction methods.

[0118] The first prediction mode: the model input Xinput =[I(t);U(t);BST(t);C n2 (t)], the output of the model Y output =[SOC(t); SOE(t); BST(t+1)], the function of the model is to directly predict the SOC and SOE values ​​at the current time t, and the temperature value at the future time t+1.

[0119] Second prediction mode: model input X input =[I(t);U(t);BST(t);C n2 (t)], the output of the model Y output =[SOC(t+1); SOE(t+1); BST(t+2)], the function of the model is to directly predict the SOC and SOE values ​​at the future time t+1, and the temperature value at the future time t+2.

[0120] The third prediction mode: the model input X input =[I(t);U(t);BST(t);C n2 (t)], the output of the model Y output =[SOC(t); SOE(t); BST(t+1); SOC(t+1); SOE(t+1); BST(t+2)], the function of the model is to directly predict the SOC and SOE values ​​at the current time t and t+1, as well as the temperature values ​​at the future time t+1 and t+2.

[0121] Among the three prediction modes mentioned above, the third prediction mode can output the battery status at multiple moments in the future at one time, providing the battery with accurate battery status reference information.

[0122] In order to highlight the prediction effect of the present invention, the Informer network is selected for comparison. input =[I(t); U(t); BST(t)], and the outputs are the same as those of the present invention.

[0123] It is worth noting that the present invention combines the training sets at all temperatures for training, and then uses the trained deep learning model to realize the battery SOC, SOE and BST prediction at various temperatures.

[0124] In the case of the first prediction mode, the prediction results of 40℃ are as follows: Figure 6 As shown in Table 1, the statistical results of each error at all temperatures are

[0125] Table 1 Prediction results at all temperatures under the first prediction mode (%)

[0126]

[0127] Table 2 Prediction results at all temperatures under the second prediction mode (%)

[0128]

[0129] Table 3 Prediction results at all temperatures under the second prediction mode (%)

[0130]

[0131] from Figure 6 , 7 As can be seen from Figures 8 and 9, the battery SOC, SOE and BST prediction results of the present invention can be well consistent with the actual values ​​without large fluctuations.

[0132] It can be seen from the statistical results that under the above three prediction modes, the battery SOC, SOE and BST prediction errors estimated by the deep learning network model of the present invention are all smaller than those of the Informer network.

[0133] The above prediction results show that under the three prediction mode methods, the present invention can obtain higher battery SOC, SOE and BST prediction accuracy, which shows that the method proposed in the present invention can effectively estimate various battery states and provide reliable reference information for cloud online deployment.

[0134] The basic principles and main features of the present invention and the advantages of the present invention are shown and described above. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0135] Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the claims are included in the present invention.

[0136] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for predicting battery SOC, SOE and BST based on a hybrid deep learning network, characterized in that: Includes steps: Perform operating tests on the battery at different temperatures to obtain test data, extract information on current I, voltage U, and battery surface temperature BST from the test data, and calculate the battery SOC and SOE; The ampere-hour integration method is used to integrate the current based on the final charge and discharge capacity of the previous cycle to obtain the charge and discharge capacity C of the battery. n2 ; The current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 Normalization processing forms a data set, which is divided into a training set and a test set; Construct a deep learning network for estimating the battery SOC, SOE, and battery surface temperature BST state, which includes the Informer network, GRU network, and KAN network from the input side to the output side; Based on current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 As input, battery SOC, SOE and BST are outputs. The deep learning network is trained using the training set and verified using the validation set to obtain a deep learning network prediction model. The current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 As model input, the deep learning network prediction model is used to estimate the battery SOC, SOE and battery surface temperature BST at the current time t and future times.

2. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The SOC at the current time t is calculated by the following formula: In the formula, SOC(t0) is the initial SOC value, SOC(t) is the SOC value at the current time t, η is the Coulomb rate, η=1, I t is the current flowing through the battery at time t, C N is the actual capacity of each working condition.

3. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The SOE at the current time t is calculated by the following formula: In the formula, SOE(t0) is the initial SOE value, SOE(t) is the SOE value at the current time t, η is the Coulomb multiple, η=1, U t is the battery voltage at time t, I t is the current flowing through the battery at time t, E N is the total available energy for each operating condition.

4. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The charge and discharge capacity C at the current time t n2 Calculated by the following formula: In the formula, C n2 (t0) is the capacity value at the initial moment, which is 0 during discharge and charge, η is the Coulomb rate, η=1, I t is the current flowing through the battery at time t, C n2 (t) is the charge and discharge capacity at the current moment.

5. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The training set is the mixed working condition and CC-CV working condition at each temperature, and the test set is the standard working condition at each temperature.

6. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: When the deep learning network prediction model predicts, The model input is X input =[I(t);U(t);BST(t);C n2 (t)], Model output Y output =[SOC(t);SOE(t);BST(t+1)]; Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 , predict the SOC and SOE at the current time t, and the battery surface temperature BST at time t+1.

7. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: When the deep learning network prediction model predicts, Model input X input =[I(t);U(t);BST(t);C n2 (t)], Model output Y output =[SOC(t+1);SOE(t+1);BST(t+2)], Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 , predict the SOC and SOE at time t+1, and the battery surface temperature BST at time t+2.

8. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: When the deep learning network prediction model predicts, Model input X input =[I(t);U(t);BST(t);C n2 (t)], Model output Y output =[SOC(t); SOE(t); BST(t+1); SOC(t+1); SOE(t+1); BST(t+2)]; Used to input the current I, voltage U, battery surface temperature BST, charge and discharge capacity C at the current time t n2 , predict the SOC and SOE at the current time t and time t+1, and the battery surface temperature BST at time t+1 and time t+2.

9. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The current I, voltage U, battery surface temperature BST, charge and discharge capacity C n2 Normalization is performed by the following formula: x normal =(x signal -x min ) / (x max -x min ); In the formula, x min and x max is the minimum and maximum value of the feature, x signal is the actual measured data in the data, x normal The data is normalized.

10. The method for predicting battery SOC, SOE and BST based on a hybrid deep learning network according to claim 1, characterized in that: The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the prediction performance of the deep learning network prediction model, including: In the formula, y t is the actual SOC, SOE and BST value of the battery, y * t are the predicted SOC, SOE and BST values, and M is the total number of cycles in the test set.

Images