Battery charge state prediction method and system based on convolutional neural network

Abstract

The invention discloses a battery charge state prediction method and system based on a convolutional neural network. The method comprises the following steps: acquiring power consumption data of a battery; preprocessing the power consumption data; obtaining a first predicted SOC of the battery at each sampling time according to a double-Kalman filtering algorithm and the power utilization data; forming a data set by using the power consumption data and the first predicted SOC; dividing the data set into a training set, a verification set and a test set; constructing a convolutional neural network model CNN; training a convolutional neural network model by using the training set to obtain a trained convolutional neural network model; verifying the trained convolutional neural network model by using the verification set to obtain a convolutional neural network model passing verification; and testing the verified convolutional neural network model by using the test set to obtain a second predicted SOC of the battery at the next moment. The method can effectively improve the quality of a data set, reduces the introduction of errors, and improves the precision of a prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management, and in particular to a battery state of charge prediction method and system based on convolutional neural network. Background Art

[0002] At present, the battery SOC prediction methods are mainly divided into two categories: model-driven methods and data-driven methods. In the model-driven method, theoretical calculation methods based on battery modeling are adopted, including open circuit voltage model, fractional order model and electrochemical model. These methods rely on accurate battery modeling and a strong theoretical basis to calculate the SOC value. The model-driven method has high requirements on the rationality of battery modeling and the accuracy of the theoretical basis. The setting or calculation of model parameters has a significant impact on the accuracy of SOC, so high requirements are placed on the accuracy of the model. The data-driven method uses machine learning or deep learning networks, such as SVN, FNN, GRU, LSTM, etc., to predict SOC. Therefore, the data-driven method needs to use a large number of data sets containing battery features and SOC labels to train models such as neural networks, and achieve accurate prediction of SOC by learning and fitting the mapping relationship between relevant feature parameters and SOC.

[0003] In contrast, although the data-driven method does not rely on the accuracy of the theoretical model, it may introduce a certain degree of error during the model training and fitting process due to the quality of the data set. Summary of the invention

[0004] In order to solve the technical problems existing in the background technology, the present invention proposes a battery state of charge prediction method and system based on convolutional neural network.

[0005] The present invention proposes a battery state of charge prediction method based on a convolutional neural network, comprising:

[0006] Obtaining battery power consumption data; wherein the power consumption data includes sampling time, voltage, current and temperature;

[0007] Preprocessing of electricity consumption data;

[0008] According to the dual Kalman filter algorithm and power consumption data, the first predicted SOC of the battery at each sampling time is obtained;

[0009] The power consumption data and the first predicted SOC are combined into a data set;

[0010] Divide the dataset into training, validation and test sets;

[0011] Construct a convolutional neural network model CNN;

[0012] The convolutional neural network model is trained using the training set to obtain a trained convolutional neural network model;

[0013] Use the validation set to validate the trained convolutional neural network model and obtain a validated convolutional neural network model;

[0014] The verified convolutional neural network model is tested using the test set to obtain the second predicted SOC of the battery at the next moment.

[0015] Preferably, after the power consumption data and the predicted SOC are combined into a data set, the method further includes:

[0016] Form a feature matrix atlas based on the data set;

[0017] The feature matrix atlas is divided into training set, validation set and test set.

[0018] Preferably, forming a feature matrix atlas according to the data set specifically includes:

[0019] Sort the sample data in the data set according to the sampling time;

[0020] Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence;

[0021] A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step length, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram; wherein each row or column in the feature matrix diagram represents a subsequence;

[0022] All feature matrix graphs are combined to form a feature matrix graph set.

[0023] Preferably, after all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram, the method further includes:

[0024] Standardize the feature matrix.

[0025] Preferably, the convolution formula of the convolutional neural network is

[0026] (Z k ) i,j =(X*W k ) i,j +b k ;

[0027] In the formula, (Z k ) i,j represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; bk represents the bias of the convolution kernel k; * represents the convolution operation.

[0028] The present invention also proposes a battery state of charge prediction system based on a convolutional neural network, comprising:

[0029] An acquisition module is used to acquire the power consumption data of the battery; wherein the power consumption data includes sampling time, voltage, current and temperature;

[0030] A preprocessing module, used for preprocessing electricity consumption data;

[0031] A dual Kalman filter module, used to obtain a first predicted SOC of the battery at each sampling time according to a dual Kalman filter algorithm and power consumption data;

[0032] A data set building module, used for forming a data set from the power consumption data and the first predicted SOC;

[0033] The partitioning module is used to divide the data set into training set, validation set and test set;

[0034] Model building module, used to build convolutional neural network model CNN;

[0035] A training module is used to train the convolutional neural network model using a training set to obtain a trained convolutional neural network model;

[0036] The verification module is used to verify the trained convolutional neural network model using the verification set to obtain a verified convolutional neural network model;

[0037] The prediction module is used to test the verified convolutional neural network model using the test set to obtain the second predicted SOC of the battery at the next moment.

[0038] Preferably, it also includes:

[0039] A data set processing module, used for forming a feature matrix atlas according to the data set;

[0040] The partition module is also used to partition the feature matrix atlas into training set, validation set and test set.

[0041] Preferably, the data set processing module is used to form a feature matrix atlas according to the data set, specifically including:

[0042] The data set processing module is used to sort the sample data in the data set according to the sampling time;

[0043] Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence;

[0044] A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step length, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram; wherein each row or column in the feature matrix diagram represents a subsequence;

[0045] All feature matrix graphs are combined to form a feature matrix graph set.

[0046] Preferably, the data set processing module is also used to perform standardization processing on the feature matrix diagram.

[0047] Preferably, the convolution formula of the convolutional neural network is

[0048] (Z k ) i,j =(X*W k ) i,j +b k ;

[0049] In the formula, (Z k ) i,j represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; b k represents the bias of the convolution kernel k; * represents the convolution operation.

[0050] In the present invention, the proposed battery state of charge prediction method and system based on convolutional neural network aims at the technical problem that the energy storage battery data is not fully feature extracted, resulting in incomplete and inefficient data features, easy loss of battery information, and affecting the accuracy and reliability of battery health state prediction. The acquired power consumption data is first preprocessed to remove outliers and noise data, and then the dual Kalman filter algorithm is used to extract data features on long time scales and short time scales to obtain the first predicted SOC of the battery at each sampling time, and the power consumption data and the first predicted SOC are combined into a data set to facilitate the division of training set, verification set and test set, so as to train, verify and test the constructed convolutional neural network model to obtain the second predicted SOC of the battery at the next moment.

[0051] The present invention can effectively improve the quality of the data set, reduce the introduction of errors, and improve the accuracy of the prediction results. In addition, CNN has a good tolerance for noisy data, can extract more complex features from large-scale data, process nonlinear relationships more efficiently, and has higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the flow of a battery state of charge prediction method based on a convolutional neural network in one embodiment of the present invention.

[0053] Figure 2 The figure is a flow chart of a dual Kalman filter algorithm in one embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] Reference Figure 1 The present invention proposes a method and system for predicting battery state of charge based on a convolutional neural network, comprising:

[0056] Obtaining battery power consumption data; wherein the power consumption data includes sampling time, voltage, current and temperature;

[0057] Preprocess the electricity consumption data to remove outliers and noise data;

[0058] According to the dual Kalman filter algorithm and power consumption data, the first predicted SOC of the battery at each sampling time is obtained;

[0059] The power consumption data and the first predicted SOC are combined into a data set;

[0060] Divide the dataset into training, validation and test sets;

[0061] Construct a convolutional neural network model CNN;

[0062] The convolutional neural network model is trained using the training set to obtain a trained convolutional neural network model;

[0063] Use the validation set to validate the trained convolutional neural network model and obtain a validated convolutional neural network model;

[0064] The verified convolutional neural network model is tested using the test set to obtain the second predicted SOC of the battery at the next moment.

[0065] In order to solve the technical problem that energy storage battery data is not fully feature extracted, resulting in incomplete and inefficient data features, easy loss of battery information, and affecting the accuracy and reliability of battery health status prediction, the present invention first pre-processes the acquired power consumption data to remove outliers and noise data, and then uses a dual Kalman filter algorithm to extract data features on long time scales and short time scales to obtain the first predicted SOC of the battery at each sampling time, and the power consumption data and the first predicted SOC are combined into a data set to facilitate division into training set, verification set and test set, so as to train, verify and test the constructed convolutional neural network model to obtain the second predicted SOC of the battery at the next moment.

[0066] The present invention can effectively improve the quality of the data set, reduce the introduction of errors, and improve the accuracy of the prediction results. In addition, CNN has a good tolerance for noisy data, can extract more complex features from large-scale data, process nonlinear relationships more efficiently, and has higher prediction accuracy.

[0067] In order to make the samples in the data set more suitable for the convolutional neural network model CNN, thereby further improving the quality of the data set, in this embodiment, after the power consumption data and the predicted SOC are combined into a data set, the following is also included:

[0068] Form a feature matrix atlas based on the data set;

[0069] The feature matrix atlas is divided into training set, validation set and test set.

[0070] Specifically, a feature matrix atlas is formed according to the data set, including:

[0071] Sort the sample data in the data set according to the sampling time;

[0072] Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence;

[0073] A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step size, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram;

[0074] All feature matrix graphs are combined to form a feature matrix graph set.

[0075] In which, each row or column in the feature matrix diagram represents a subsequence.

[0076] In a further embodiment, after all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram, the method further includes:

[0077] The feature matrix graph is standardized so that the data values ​​in the feature matrix graph are normalized to between 0 and 1. Each matrix element corresponds to the value of a pixel, forming a pseudo image that can be input into CNN.

[0078] like Figure 2 As shown, in this embodiment, according to the dual Kalman filter algorithm and the power consumption data, the first predicted SOC of the battery at each sampling time is obtained, which specifically includes:

[0079] Obtain the initial SOC of the battery based on the power consumption data;

[0080] Construct a second-order RC battery equivalent circuit model and obtain the battery state equation and output equation;

[0081] The initial SOC of the battery is used as the input state quantity, and the voltage equation corresponding to the preset second-order RC battery equivalent circuit model is used as the output equation, and the extended Kalman filter algorithm is used to perform a primary estimation of the battery SOC;

[0082] The battery SOC obtained by the primary estimation is used as the input state quantity, and the ampere-hour integration method is used as the output equation. The battery SOC is estimated secondary using the Kalman filter algorithm, and the battery SOC obtained by the secondary estimation is used as the first predicted SOC.

[0083] Among them, the Kalman filter algorithm adopts the extended Kalman filter, and the state and observation equations are approximately linearized using the first-order Taylor expansion:

[0084]

[0085] In the formula, ff(x k ,u k ) represents the state transition equation, which describes how the system state is transferred from time step k to time step k+1; x k represents the state of the system at time step k; u k represents the external effect on the system, and the parameter transfer equation g(x k ,u k ) describes the change of parameters over time; ω k-1 The random noise introduced in the state transfer process is the observation vector at time k; y k is used to represent the innovation sequence, i.e., the difference between the observed value and the predicted observed value; v k is the observation noise, which is usually assumed to have a mean of 0; A k-1 is the state transfer matrix; C k Represents the observation matrix, which is used to map the state vector to the observation space.

[0086] In this embodiment, the parameter identification part is divided into a long time scale and a short time scale, which can help deal with the problem of parameter changes at different time scales.

[0087] In the long time scale, the double Kalman filter is used to identify the parameters with slow changes, such as ohmic internal resistance, polarization internal resistance, polarization capacitance, charge transfer internal resistance, charge transfer capacitance, battery capacity, etc. The double Kalman filter has:

[0088]

[0089] In the formula, represents the covariance matrix of the predicted state, i.e. the estimated covariance before the observation update; ∑ r The covariance matrix representing the process noise represents the uncertainty in the evolution of the system state; represents the Kalman gain matrix, which balances the error between the predicted state and the observed state and is used to adjust the state estimation; represents the observation matrix, which describes how the observations are associated with the states; e The covariance matrix represents the measurement noise, which represents the uncertainty in the observations; represents the updated state estimate, that is, the estimate of the current state after combining the observed information; The observation equation of the system, which shows how the observations are derived from the states and control inputs; represents the updated state covariance matrix, that is, the correction of the covariance after combining the observation information.

[0090] Among them, the short-time scale state Kalman filter is:

[0091]

[0092] In the formula, represents the updated state estimate at the current time k, which is the state estimate obtained by combining the new observation information; Represents the state transfer equation, describing the system state from time k-1 according to the input u k-1 and parameters Change to the state at time k; A represents the covariance matrix of the predicted state, that is, the covariance matrix before the observation update at time k; k-1 represents the state transfer matrix, which describes the transition process of the system state from time n-1 to time n; ∑ W The covariance matrix of the process noise represents the uncertainty in the state transition process; L k represents the Kalman gain matrix, which balances the error between prediction and observation and determines how to update the state estimate; represents the updated covariance matrix after combining the observation information; C k represents the observation matrix, which represents the linear relationship between state and observation; ∑ V represents the observation noise covariance matrix, which represents the uncertainty in the observation; Represents the predicted state estimate, which is a prediction based on the state estimate at the previous moment; represents the updated state covariance matrix, that is, the covariance after combining the observation update; I represents the identity matrix.

[0093] It should be noted that the convolutional neural network CNN in this embodiment is mainly used to process data with a grid-like topological structure. The convolution layer can effectively extract spatial features through the local receptive field and weight sharing mechanism, and has high computational efficiency.

[0094] Among them, the convolution formula of the convolutional neural network is

[0095] (Z k ) i,j =(X*W k ) i,j +b k ;

[0096] (Z k ) i,j represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; b k represents the bias of the convolution kernel k; * represents the convolution operation.

[0097] The present invention also proposes a battery state of charge prediction system based on a convolutional neural network, comprising:

[0098] An acquisition module is used to acquire the power consumption data of the battery; wherein the power consumption data includes sampling time, voltage, current and temperature;

[0099] A preprocessing module, used for preprocessing the electricity consumption data;

[0100] A dual Kalman filter module, used to obtain a first predicted SOC of the battery at each sampling time according to a dual Kalman filter algorithm and power consumption data;

[0101] A data set building module, used for forming a data set from the power consumption data and the first predicted SOC;

[0102] The partitioning module is used to divide the data set into training set, validation set and test set;

[0103] Model building module, used to build convolutional neural network model CNN;

[0104] A training module is used to train the convolutional neural network model using a training set to obtain a trained convolutional neural network model;

[0105] The verification module is used to verify the trained convolutional neural network model using the verification set to obtain a verified convolutional neural network model;

[0106] The prediction module is used to test the verified convolutional neural network model using the test set to obtain the second predicted SOC of the battery at the next moment.

[0107] In this embodiment, it also includes:

[0108] A data set processing module, used for forming a feature matrix atlas according to the data set;

[0109] The partition module is also used to partition the feature matrix atlas into training set, validation set and test set.

[0110] The data set processing module is used to form a feature matrix atlas according to the data set, specifically including:

[0111] The data set processing module is used to sort the sample data in the data set according to the sampling time;

[0112] Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence;

[0113] A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step length, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram; wherein each row or column in the feature matrix diagram represents a subsequence;

[0114] All feature matrix graphs are combined to form a feature matrix graph set.

[0115] In a further embodiment, the data set processing module is further used to perform standardization processing on the feature matrix graph.

[0116] Among them, the convolution formula of the convolutional neural network is

[0117] (Z k ) i,j =(X*W k ) i,j +b k ;

[0118] In the formula, (Z k ) i,j represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; b k represents the bias of the convolution kernel k; * represents the convolution operation.

[0119] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A battery state of charge prediction method based on convolutional neural network, characterized in that: include: Obtaining battery power consumption data; wherein the power consumption data includes sampling time, voltage, current and temperature; Preprocessing of electricity consumption data; According to the dual Kalman filter algorithm and power consumption data, the first predicted SOC of the battery at each sampling time is obtained; The power consumption data and the first predicted SOC are combined into a data set; Divide the dataset into training, validation and test sets; Construct a convolutional neural network model CNN; The convolutional neural network model is trained using the training set to obtain a trained convolutional neural network model; Use the validation set to validate the trained convolutional neural network model and obtain a validated convolutional neural network model; The verified convolutional neural network model is tested using the test set to obtain the second predicted SOC of the battery at the next moment.

2. The method for predicting battery state of charge based on a convolutional neural network according to claim 1, characterized in that: After the power consumption data and predicted SOC are combined into a data set, it also includes: Form a feature matrix atlas based on the data set; The feature matrix atlas is divided into training set, validation set and test set.

3. The battery state of charge prediction method based on convolutional neural network according to claim 2, characterized in that: A feature matrix atlas is formed based on the data set, including: Sort the sample data in the data set according to the sampling time; Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence; A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step length, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram; wherein each row or column in the feature matrix diagram represents a subsequence; All feature matrix graphs are combined to form a feature matrix graph set.

4. The method for predicting battery state of charge based on convolutional neural network according to claim 3, characterized in that: After all the subsequences captured by each sliding window are stacked in order to form a feature matrix diagram, it also includes: Standardize the feature matrix.

5. The method for predicting battery state of charge based on convolutional neural network according to claim 1, characterized in that: The convolution formula of the convolutional neural network is (Z k ) i,j =(X*W k ) i,j +b k ; In the formula, (Z k ) i,j represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; b k represents the bias of the convolution kernel k; * represents the convolution operation.

6. A battery state of charge prediction system based on convolutional neural network, characterized in that: include: An acquisition module is used to acquire the power consumption data of the battery; wherein the power consumption data includes sampling time, voltage, current and temperature; A preprocessing module, used for preprocessing electricity consumption data; A dual Kalman filter module, used to obtain a first predicted SOC of the battery at each sampling time according to a dual Kalman filter algorithm and power consumption data; A data set building module, used for forming a data set from the power consumption data and the first predicted SOC; The partitioning module is used to divide the data set into training set, validation set and test set; Model building module, used to build convolutional neural network model CNN; A training module is used to train the convolutional neural network model using a training set to obtain a trained convolutional neural network model; The verification module is used to verify the trained convolutional neural network model using the verification set to obtain a verified convolutional neural network model; The prediction module is used to test the verified convolutional neural network model using the test set to obtain the second predicted SOC of the battery at the next moment.

7. The battery state of charge prediction system based on convolutional neural network according to claim 6, characterized in that: Also includes: A data set processing module, used for forming a feature matrix atlas according to the data set; The partition module is also used to partition the feature matrix atlas into training set, validation set and test set.

8. The method for predicting battery state of charge based on convolutional neural network according to claim 7, characterized in that: The data set processing module is used to form a feature matrix atlas according to the data set, which specifically includes: The data set processing module is used to sort the sample data in the data set according to the sampling time; Each sample data is concatenated into a vector according to the first predicted SOC and its power consumption data to form a subsequence; A sliding window algorithm is used on the time axis to capture subsequences according to a predetermined step length, and all subsequences captured by each sliding window are stacked in order to form a feature matrix diagram; wherein each row or column in the feature matrix diagram represents a subsequence; All feature matrix graphs are combined to form a feature matrix graph set.

9. The method for predicting battery state of charge based on convolutional neural network according to claim 8, characterized in that: The data set processing module is also used to standardize the feature matrix graph.

10. The method for predicting battery state of charge based on convolutional neural network according to claim 6, characterized in that: The convolution formula of the convolutional neural network is (Z k ) i,j =(X*W k ) i,j +b k ; In the formula, (Z k ) i,j Represents the value of the convolution output feature k at position (i, j); X represents the input feature map; W k represents the weight matrix of the convolution kernel (or filter) k; b k Represents the bias of the convolution kernel k; * represents the convolution operation.

Images