A battery SOC prediction method based on long short-term memory network

Abstract

The present invention proposes a battery SOC prediction method based on a long short-term memory network, which can capture the battery SOC prediction sequence and make the prediction result more accurate. The present invention constructs a three-layer radial basis neural network according to the structure of the modified long short-term memory network; then downloads NASA's battery data set, and organizes the data format from data to SOC according to information such as voltage, current, temperature, time, etc.; then randomly initializes the parameters of the improved long short-term memory network; secondly, inputs the data into the constructed long short-term memory network to obtain the output; again, according to the time series gradient descent method, updates the relevant parameters; again, determines the model parameters, and the model training process ends; finally, inputs the test data into the trained model to perform battery SOC prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of battery performance prediction, and in particular to a battery SOC prediction method based on a long short-term memory network. Background Art

[0002] As a new type of green energy, lithium-ion batteries have been widely used as energy supply devices in devices such as mobile phones and cars. However, lithium-ion battery energy storage itself is a complex chemical reaction process. Unlike traditional oil vehicles, it is impossible to directly determine the current available energy by a fixed amount of oil. In addition, the continuous charging and discharging during the battery cycle is a nonlinear change process, and it is impossible to directly use a simple linear model to estimate the state of charge (SOC). If gasoline is left unused, the oil will not be lost, but for lithium batteries, even if they are left unused, the available power of the battery will change over time.

[0003] The estimation methods of battery SOC can be divided into two categories, namely empirical model-based methods and data-driven methods. The empirical model-based methods require accurate modeling of the battery's operating conditions, but modeling the actual working conditions is a complex nonlinear problem, and there is no way to accurately solve it through empirical modeling. The data-driven method uses relevant data parameters of the battery, such as voltage, internal resistance, etc., and builds a nonlinear model to fit this complex nonlinear relationship by building a relationship between them and the battery SOC. This method avoids single modeling of the battery's operating conditions and makes it adaptable to a variety of operating conditions based on real data. However, the existing data-driven methods have poor fitting capabilities and require the use of nonlinear modeling methods with stronger capabilities.

[0004] At present, a lithium-ion battery SOC prediction method based on a recurrent neural network has been proposed (patent application publication number: CN 108519556 A). This method predicts SOC through an improved recurrent neural network-long short-term memory network, realizing a high-precision SOC estimation method. However, due to the direct use of the long short-term memory network structure, the dynamic timing characteristics inside the battery cannot be fully captured. Summary of the invention

[0005] In view of this, the present invention proposes a battery SOC prediction method based on a long short-term memory network, which can capture the battery SOC prediction time series and make the prediction result more accurate.

[0006] To achieve the above objectives, the present invention provides a lithium-ion battery SOC prediction method based on a long short-term memory network.

[0007] The present invention discloses a battery SOC prediction method based on a long short-term memory network. According to the structure of the modified long short-term memory network, a three-layer radial basis neural network is constructed; then, the battery data set of NASA is downloaded, and the data format from data to SOC is organized according to voltage, current, temperature and time; then, the parameters of the improved long short-term memory network are randomly initialized; secondly, the data is input into the constructed long short-term memory network to obtain output; thirdly, the relevant parameters are updated according to the temporal gradient descent method; thirdly, the model parameters are determined, and the model training process ends; finally, the test data is input into the trained model to predict the battery SOC.

[0008] Among them, in the structure of the modified long short-term memory network, after the model is built, the back propagation algorithm is used to update the weight W in the formula * ; The formula is as follows:

[0009] f t =σ(W ft ·x t +W fh ·h t-1 +W fc ·C t-1 +b f )

[0010] i t =σ(W it ·x t +W ih ·h t-1 +W ic ·C t-1 +b i )

[0011] o t =σ(W ot ·x t +W oh ·h t-1 +W oc ·C t-1 +b o )

[0012] Among them, W ** represents the weight of the model, b * Represents the bias weight, W ** The black dots in the figure represent the multiplication of the matrix and the vector, and the * represents the product of the corresponding elements of the vector. t represents the input of the current step, h t-1 represents the implicit state of the previous step, C t-1 Indicates the cell state of the previous step, i t represents the input gate, o t represents the output gate, f trepresents the forget gate;

[0013] Through the weight updating process, the nonlinear relationship between the actual output and input data of the SOC prediction is learned. After the weight is determined, the test data is input to obtain the SOC prediction result of the test battery.

[0014] The specific steps for constructing a three-layer radial basis neural network are as follows: first, construct the correspondence between the external characteristic parameters of the battery and the SOC results, then normalize the data, and finally determine the time length in the long short-term memory network by determining the width of the sliding window.

[0015] The output of the LSTM network refers to the final output SOC prediction result, and the input layer represents the data parameters related to the battery life prediction. In the process of organizing data input, the SOC prediction results of multiple moments are spliced ​​together as a sample input of the LSTM network; the battery external parameters form a vector x with a dimension of n. Considering the time window K, the number of samples input to a model is K×n. The calculation process is that K x are used separately, and each input in the time window is x. t express.

[0016] When predicting the SOC at the current moment, the battery external characteristic parameters at previous moments are comprehensively considered, and when predicting the SOC at the current time T, the battery external characteristic parameters from time T-(K-1) to T are considered.

[0017] Beneficial Effects

[0018] 1. The present invention constructs a three-layer radial basis neural network according to the structure of the modified long short-term memory network; then downloads NASA's battery data set, and organizes the data format from data to SOC according to information such as voltage, current, temperature, time, etc.; then randomly initializes the parameters of the improved long short-term memory network; secondly, inputs the data into the constructed long short-term memory network to obtain output; thirdly, updates the relevant parameters according to the temporal gradient descent method; thirdly, determines the model parameters, and the model training process ends; finally, inputs the test data into the trained model to predict the battery SOC.

[0019] 2. The present invention transforms the structure of the long short-term memory network and learns the nonlinear relationship between the real output and input data of the SOC prediction through the weight updating process. After the weight is determined, the test data is input to obtain the SOC prediction result of the test battery.

[0020] 3. The present invention uses a new type of long short-term memory network to predict SOC, which has the characteristics of capturing the timing of battery SOC prediction, so that the prediction result is more accurate; in the long short-term memory network, the input layer represents the data parameters related to the battery life prediction, and these battery external parameters form a vector x with a dimension of n, that is, the number of neurons in the input layer is n. In addition, considering the time window K, the number of samples input into a model is K×n, but the calculation process is K x separately, and each input in the time window is x. t By modifying the original long short-term memory network, a new network structure was designed, which is more accurate than the existing technology in battery performance modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The schematic diagram for comparing the classic long short-term memory network and the improved long short-term memory network provided by the present invention is shown in FIG. The dotted line in the figure represents the connection structure proposed by the present invention for capturing the internal charge and discharge timing characteristics of the battery. DETAILED DESCRIPTION

[0022] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0023] Long short-term memory network is a model that can capture the sequence information between different time steps of data. The specific structure of the classic long short-term memory network (Long Short-Time Memory, LSTM) is as follows Figure 1 As shown, this is a schematic diagram of the calculation process of a time step. The input of the structure of each time step includes the input x of the current step. t , the hidden state h of the previous step t-1 and cell state C t-1 , the output of the current time step is the cell state C of the current time step t and the hidden state h t The input x of the current step t and the hidden state h of the previous step t-1 Can determine the input gate i t , output gate o t and forget gate f t The function of each gate is to control the amount of information. The value of the gate ranges from 0 to 1, where 0 represents complete isolation of information and 1 represents complete reception of information. The intermediate value represents allowing a certain amount of information to pass through a gate. The input x of the current step t and the hidden state h of the previous step t-1 First, a nonlinear mapping is performed to output the intermediate cells of the current step. and input gate i t Multiply to get the input of the current step. At the same time, the cell state C of the previous stept-1 and forget gate f t Multiply them to get the residual after forgetting in the previous step, and add them together to get the cell state C output at the current time step. t ,C t After nonlinear mapping, it is multiplied by the output gate to obtain the implicit state h of the current step output t The specific calculation steps are as follows:

[0024] f t =σ(W ft ·x t +W fh ·h t-1 +b f ) (1)

[0025] i t =σ(W it ·x t +W ih ·h t-1 +b i ) (2)

[0026] o t =σ(W ot ·x t +W oh ·h t-1 +b o ) (3)

[0027]

[0028]

[0029] h t =o t *tanh(C t ) (6)

[0030] Among them, W ** represents the weight of the model, b * Represents the bias weight, W ** The black dots in the figure represent the multiplication of matrices and vectors, * represents the product of the corresponding elements of the vectors, and tanh represents a nonlinear mapping:

[0031]

[0032] Through the reasoning of time steps, the relevant information is finally stored in h t Among them, h t Then, by mapping through the fully connected layer, the estimated SOC output value can be obtained.

[0033] In the present invention, the structure of the above long short-term memory network is modified. Specifically, formulas (1), (2) and (3) are selectively modified to obtain:

[0034] f t =σ(W ft ·x t +W fh ·h t-1 +W fc ·C t-1 +b f ) (7)

[0035] i t =σ(W it ·x t +W ih ·h t-1 +W ic ·C t-1 +b i ) (8)

[0036] o t =σ(W ot ·x t +W oh ·h t-1 +W oc ·C t-1 +b o ) (9)

[0037] After the model is built, the back propagation algorithm can be used to update the weight W in the above formula * Through the above weight update process, the nonlinear relationship between the real output and input data of SOC prediction can be learned. After the weight is determined, the test data can be input to obtain the SOC prediction result of the test battery.

[0038] In the long short-term memory network of the present invention, the input layer represents the data parameters related to the battery life prediction. These battery external parameters form a vector x with a dimension of n, that is, the number of neurons in the input layer is n. In addition, considering the time window K, the number of samples input into a model is K×n, but the calculation process is K x separately, and each input in the time window is x. t express.

[0039] According to the structure of the modified LSTM network, a three-layer radial basis neural network is constructed; then NASA's battery data set is downloaded, and the data format from data to SOC is organized according to information such as voltage, current, temperature, and time; then the parameters of the improved LSTM network are randomly initialized; secondly, the data is input into the constructed LSTM network to obtain the output; again, the relevant parameters are updated according to the temporal gradient descent method; again, the model parameters are determined, and the model training process is completed; finally, the test data is input into the trained model to predict the battery SOC.

[0040] The present invention constructs a three-layer radial basis neural network according to the structure of the modified long short-term memory network. The specific steps are as follows:

[0041] First, construct a correspondence between the external characteristic parameters of the battery and the SOC result, then normalize the data, and finally determine the time length in the long short-term memory network by determining the width of the sliding window, that is, when predicting the SOC at the current moment, comprehensively consider the external characteristic parameters of the battery at previous moments. Assuming that the width of the time window is K, it means that when predicting the SOC at the current time T, the external characteristic parameters of the battery from time T-(K-1) to T must be considered. Specifically, for example, if the time window K is 2, it means that the external parameters of the battery at time T-1 and T must be considered to predict the battery SOC. Specifically, the battery SOC prediction method based on the long short-term memory network of the present invention uses an improved long short-term memory network connection structure such as Figure 1 The dotted line shows the connections in the improved long short-term memory network and the weights obtained through training, and also includes the use of the weights to predict the battery SOC. The output of the long short-term memory network refers to the final output SOC prediction result. In the process of organizing data input, the SOC prediction results of multiple moments are spliced ​​together as a sample input of the long short-term memory network.

[0042] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A battery SOC prediction method based on long short-term memory network, characterized in that: According to the structure of the modified LSTM network, a three-layer radial basis neural network was constructed. Then, NASA's battery data set was downloaded and the data format from data to SOC was organized according to voltage, current, temperature, and time. Then, the parameters of the improved long short-term memory network are randomly initialized; secondly, the data is input into the constructed long short-term memory network to obtain the output; Again, update the relevant parameters according to the temporal gradient descent method; Again, the model parameters are determined and the model training process is completed; finally, the test data is input into the trained model to predict the battery SOC; In the structure of the modified long short-term memory network, after the model is built, the back propagation algorithm is used to update the weight W in the formula * ; The formula is as follows: f t =σ(W ft ·x t +W fh ·h t-1 +W fc ·C t-1 +b f ) i t =σ(W it ·x t +W ih ·h t-1 +W ic ·C t-1 +b i ) o t =σ(W ot ·x t +W oh ·h t-1 +W oc ·C t-1 +b o ) Among them, W ** represents the weight of the model, b * Represents the bias weight, W ** The black dots in the figure represent the multiplication of the matrix and the vector, and the * represents the product of the corresponding elements of the vector. t represents the input of the current step, h t-1 represents the implicit state of the previous step, C t-1 Indicates the cell state of the previous step, i t represents the input gate, o t represents the output gate, f t represents the forget gate; Through the weight update process, the nonlinear relationship between the real output and input data of the SOC prediction is learned. After the weight is determined, the test data is input to obtain the SOC prediction result of the test battery; The specific steps of constructing a three-layer radial basis neural network are as follows: first, construct the correspondence between the external characteristic parameters of the battery and the SOC result, then normalize the data, and finally determine the time length in the long short-term memory network by determining the width of the sliding window; The output of the LSTM network refers to the final output SOC prediction result. The input layer represents the data parameters related to the battery life prediction. In the process of organizing data input, the SOC prediction results of multiple moments are spliced ​​together as a sample input of the LSTM network. The battery external parameters form a vector x with a dimension of n. Considering the time window K, the number of samples input to a model is K×n. The calculation process is to use K x separately, and each input in the time window is x. t express.

2. The method according to claim 1, characterized in that When predicting the SOC at the current moment, the battery external characteristic parameters at previous moments are comprehensively considered. When predicting the SOC at the current time T, the battery external characteristic parameters from time T-(K-1) to T are considered.

Images