Lithium battery SoH prediction method based on MD-LSTM

Abstract

The invention discloses a lithium battery SoH prediction method based on MD-LSTM. The method comprises the following steps: determining rated capacity, discharge cut-off voltage and full charge voltage of a lithium battery; carrying out a discontinuous charging and discharging experiment on the lithium battery, and recording data information in the experiment process; obtaining an input variable and an output variable; preprocessing data, and establishing a quaternary time sequence of each cycle; constructing a battery aging data set D, D = {delta U, delta T, T ', SoH}; splitting the battery aging data set D into a training set and a test set; training the battery SoH prediction model to obtain an MD-LSTM model; and carrying out validity verification on the MD-LSTM model by using the test set, and outputting a prediction result. According to the method, the lithium battery SoH is predicted by using the MD-LSTM model, so that the accuracy of lithium battery SoH prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery state prediction, and specifically relates to a lithium battery SoH prediction method based on MD-LSTM. Background Art

[0002] Lithium-ion batteries have been widely used in many fields such as electric bicycles, electric vehicles, military equipment, aerospace, etc. due to their high energy density, long service life and environmental protection. However, during repeated charging and discharging, the components such as electrodes and diaphragms inside the battery will degrade to varying degrees, resulting in the attenuation of battery capacity and power performance, and even threatening the personal safety of users. In order to ensure the safe and reliable use of lithium-ion batteries, it is particularly important to accurately evaluate the battery's state of health (SoH).

[0003] The main methods currently used to estimate the SoH of lithium-ion batteries can be summarized as model-based methods and data-driven methods. The model-based method estimates the health status of the battery by establishing an equivalent circuit model or a chemical reaction kinetic model of the battery, combined with measured parameters such as voltage, current, and temperature. The advantage is that it can reflect the internal degradation mechanism of the battery and has a high prediction accuracy. The disadvantage is that it requires a large number of electrochemical parameter measurements, the modeling is complex, and it is difficult to promote in practical applications. The data-driven method uses machine learning algorithms, such as support vector machines and neural networks, to learn the battery degradation law from a large amount of measured data and then predict SoH. The advantage is that there is no need for complex electrochemical models and it can be easily applied to actual battery management systems. Make full use of various operating parameters of the battery, such as voltage, current, temperature, etc., to more comprehensively characterize the battery state. The disadvantage is that it requires a lot of domain knowledge and trial and error process, and the prediction accuracy is limited by the selection of input features. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a lithium battery SoH prediction method based on MD-LSTM to solve the problem of poor prediction accuracy of the existing scheme, and based on the voltage, temperature and other characteristics of the lithium battery during discharge and the MD-LSTM model, to enhance the stability of the model prediction and improve the prediction accuracy.

[0005] According to the technical solution of the present invention, the present invention provides a lithium battery SoH prediction method based on MD-LSTM, comprising the following steps: Step S1, determining the rated capacity, discharge cut-off voltage and full charge voltage of the lithium battery; Step S2, performing a discontinuous charge and discharge experiment on the lithium battery, wherein the charge cut-off condition is that the battery rises to the full charge voltage, and the discharge cut-off condition is that the voltage drops to the discharge cut-off voltage, recording the data information during the experiment, and calculating the discharge capacity in each cycle according to the integral of the discharge current and the discharge time in the data information, calculating the ratio of the discharge capacity of each cycle of the lithium battery to the rated capacity, and obtaining the battery health state SoH of the lithium battery; the experiment termination condition is that the SoH drops to the end-of-life threshold; Step S3, obtaining input variables and output variables; The input variables are: {( ∆U i , ∆U i-1 , ... , ∆U i-4 ) , ( ∆T i , ∆T i-1 , ... , ∆T i-4 ) , ( T' i , T' i-1 , ... , T' i-4 ) , ( SoH i , SoH i-1 , ... , SoH i-4 )}; The output variables are: { SoH i+1 , SoH i+2 , ... , SoH i+5}; in, ∆U i It represents the hysteresis voltage data that changes per unit time during the i-th cycle discharge process. ∆T i It represents the temperature difference during the discharge process of the ith cycle. T' i represents the ambient temperature in the i-th cycle, SoH i Indicates the battery health status in the i-th cycle; Step S4, data preprocessing; during the preprocessing process, data cleaning is performed for each input variable and output variable, including identifying and processing errors, missing values ​​and outliers in the data, detecting and correcting erroneous data, and establishing a quaternary time series for each cycle; Step S5, constructing a battery aging data set D according to the quaternary time series of each cycle of step S4, D={ ∆U , ∆T , T' , SoH}; Step S6, splitting the battery aging data set D into a training set and a test set; Step S7, training the MD-LSTM model; the training process includes: sampling the quaternary time series in the battery aging data set D, generating actual values ​​and model input values ​​for calculating the model prediction error, inputting the model input values ​​into the MD-LSTM model, outputting the prediction values, and iterating the model multiple times using the training set to obtain the MD-LSTM model; Step S8: Use the test set to verify the effectiveness of the MD-LSTM model.

[0006] Furthermore, in step S1, the rated capacity calibrated by the manufacturer when the lithium battery leaves the factory is obtained as the rated capacity, the OCV-SOC curve of the lithium battery is obtained, and the voltage when the SOC is 5% is determined as the discharge cut-off voltage of each experiment, and the voltage when the SOC is 90% is determined as the full charge voltage of each experiment.

[0007] Furthermore, in step S2, the battery health state SoH formula of the lithium battery is expressed as: ; in, t 1 is the initial discharge time, t 2 is the discharge end time, I is the current during the discharge process, Q rated is the rated capacity of the lithium battery.

[0008] Further, in step S2, the life end threshold is 70%.

[0009] Furthermore, in step S3, the hysteresis voltage changes ∆U It refers to the change in hysteresis voltage within 1 minute after the lithium battery stops discharging.

[0010] Further, in step S7, the network input of the MD-LSTM model is the first 5 ∆ U , ∆T , T' , SoH , as the charge and discharge cycle characteristics; the output is the charge and discharge cycle discharge process after the last 5 charge and discharge cycles SoH .

[0011] Further, in step S7, the MD-LSTM model is a long short-term memory network hybrid fully connected network structure, comprising a total of 6 hidden layers, in which 3 LSTM layers and 3 fully connected layers are arranged alternately, the first layer is an LSTM layer with 128 neurons, the second layer is a fully connected layer with 64 neurons, the third layer is an LSTM layer with 64 neurons, the fourth layer is a fully connected layer with 32 neurons, the fifth layer is an LSTM layer with 32 neurons, and the sixth layer is a fully connected layer with 5 neurons; The LSTM layer is configured with at least a forget gate, an input gate, and an output gate. The forget gate, the input gate, and the output gate are used to control the retention and forgetting of data. The structural unit expression of the MD-LSTM model is: ; in, x t is the input of the current time step, h t-1 is the hidden state at the previous time step, C t-1 is the cell state at the previous time step, i t 、f t 、o t They are input gate, forget gate and output gate. is the candidate cell state at the current time step, C t is the cell state at the current time step, h t is the hidden state at the current time step, W f is the forget gate weight matrix, W i is the input gate weight matrix, W c is the current cell state weight matrix, W o is the output gate weight matrix, b f , b i , b c , b o They are the forget gate, input gate, current cell state and output gate bias. W d is the fully connected layer weight matrix, b d is the fully connected layer bias, f is the activation function of the fully connected layer.

[0012] Furthermore, in step S7, the loss function used in the model training process is the mean square error MSE: ; in, f(x) is the model prediction value, y is the true value, n is the total number of samples.

[0013] Furthermore, in step S8, if the MAPE on the test set is less than or equal to 3%, the test passes and the model is valid; if the MAPE on the test set is greater than 3%, the amount of data is increased and the above experimental steps are repeated until the MAPE on the test set is less than or equal to 3%, and the experiment ends.

[0014] Furthermore, in step S8, the process of performing validity verification is to perform prediction error evaluation analysis based on the predicted SoH data and the actual SoH data, including the following steps: Calculate the mean absolute percentage error MAPE: ; in, represents the predicted value, 𝑦 𝑖 Indicates the actual value, j Represents the number of predicted values.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are as follows: The present invention proposes the characteristics of the changing hysteresis voltage from the voltage curve of the lithium battery discharge process, that is, the voltage value that rises within one minute after the battery reaches the cut-off voltage, and the characteristics of the temperature that changes during the discharge process, and uses these characteristics as neural network inputs, using LSTM networks and DNN networks to establish a lithium battery SoH prediction model based on MD-LSTM; and uses certain prior knowledge to improve the correlation between network input and output, helping the network to better understand the relationship between them, and improving the accuracy of the lithium battery SoH prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 4 is a structural diagram of an MD-LSTM model according to an embodiment of the present invention.

[0017] Figure 2 Schematic diagram of a lithium battery SoH prediction method based on MD-LSTM according to an embodiment of the present invention.

[0018] Figure 3 This is a diagram showing prediction results of a lithium battery SoH prediction method based on MD-LSTM according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present invention provides a lithium battery SoH prediction method based on MD-LSTM, and relates to the technical field of lithium battery state prediction. The method comprises: obtaining the rated capacity and OCV-SOC curve calibrated by the manufacturer when the lithium battery leaves the factory, determining the voltage when the SOC is 5% as the discharge cut-off voltage of each experiment, and the voltage when the SOC is 90% as the full charge voltage of each experiment; placing the lithium battery under three conditions of low temperature, room temperature and high temperature for discontinuous charge and discharge experiments; the experimental termination condition is that the SoH drops to 70%; recording data information during the experiment; data preprocessing, detecting and correcting error data, and establishing a multidimensional sequence of each cycle, wherein each sequence includes a changing hysteresis voltage ∆U, a cell temperature change ∆T during discharge, an ambient temperature T' during discharge, and a battery SoH during battery discharge; constructing a battery aging data set D according to the four-element time series of each cycle; splitting the data set D into a training set and a test set; using the training set to train the MD-LSTM model to obtain a battery SoH prediction model based on MD-LSTM; using the test set to test the battery SoH prediction model, and outputting the prediction result. The present invention mainly aims to solve the problem of poor prediction accuracy of existing solutions. Based on the voltage, temperature and other characteristics of lithium batteries during discharge and the MD-LSTM model, the stability of model prediction is enhanced and the prediction accuracy is improved.

[0020] See also Figure 1 , Figure 2 , a lithium battery SoH prediction method based on MD-LSTM according to an embodiment of the present invention comprises the following steps: Step S1, determine the rated capacity, discharge cut-off voltage U1 and full charge voltage U2 of the lithium battery. For example, the rated capacity calibrated by the manufacturer when the lithium battery leaves the factory is obtained as the rated capacity, the OCV-SOC curve of the lithium battery is obtained, and the voltage when the SOC is 5% is determined as the discharge cut-off voltage of each experiment. U 1. The voltage when SOC is 90% is the full charge voltage of each experiment U 2.

[0021] Step S2, placing the lithium battery under three conditions of low temperature (such as -20℃~5℃), room temperature (such as 23±2℃) and high temperature (such as 30℃~50℃) for discontinuous charge and discharge experiments; the charging cut-off condition is that the battery rises to the full charge voltage U2, the battery charging method is constant current charging, the discharge cut-off condition is that the voltage drops to the discharge cut-off voltage U1, and the data information during the experiment is recorded (the data information at least includes the corresponding discharge current and discharge time), and the discharge capacity in each cycle is calculated according to the integral of the discharge current and discharge time in the data information, and the ratio of the discharge capacity of the lithium battery in each cycle to the rated capacity is calculated to obtain the battery health state SoH of the lithium battery; the experiment termination condition is that the SoH drops to the end-of-life threshold, and the end-of-life threshold is, for example, 70%.

[0022] Furthermore, in step S2, the battery health state SoH formula of the lithium battery (in the i-th cycle) is expressed as: ; in, t 1 is the initial discharge time, t 2 is the discharge end time, I is the current during the discharge process, Q rated is the rated capacity of the lithium battery.

[0023] Step S3, obtaining input variables and output variables, and calculating the hysteresis voltage according to the data information in the experimental process described in step S2 ∆U , Battery cell temperature changes during discharge ∆T , Ambient temperature during discharge T' , during battery discharge SoH ; Specifically, the input variables are, for example: {( ∆U i , ∆U i-1 , ... , ∆U i-4 ) , ( ∆T i , ∆T i-1 , ... , ∆T i-4 ) , ( T' i , T' i-1 , ... , T' i-4 ) , ( SoH i , SoH i-1 , ... , SoH i-4 )}; The output variables are: { SoH i+1 , SoH i+2 , ... , SoH i+5}; in, ∆U i It represents the hysteresis voltage data that changes per unit time during the i-th cycle discharge process. ∆T i It represents the temperature difference during the discharge process of the ith cycle. T'i represents the ambient temperature in the i-th cycle, SoH i Indicates the battery health status in the i-th cycle.

[0024] After the lithium battery stops discharging, its terminal voltage will rise temporarily to a certain extent. This rising voltage is the hysteresis voltage. Specifically, the hysteresis voltage refers to the voltage that continues to rise on the basis of the discharge cut-off voltage after the lithium battery stops discharging. In step S3, the hysteresis voltage changes ∆U It refers to the change in hysteresis voltage within a period of time (for example, 1 minute) after the lithium battery stops discharging.

[0025] Step S4, data preprocessing; in the process of preprocessing, data cleaning is performed for each input variable and output variable, including identifying and processing errors, missing values ​​and outliers in the data, detecting and correcting erroneous data, and establishing a quaternary time series for each cycle. The quaternary time series refers to a time series containing four variables (respectively ∆U , ∆T , T' , SoH ) time series.

[0026] Step S5, constructing a data set; constructing a battery aging data set D according to the quaternary time series of each cycle of step S4, D={ ∆U,∆T,T',SoH}.

[0027] Step S6, dividing the data set, splitting the battery aging data set D into a training set and a test set.

[0028] Step S7, training the MD-LSTM model; the training process includes: performing sliding window sampling on the quaternary time series in the battery aging data set D, generating actual values ​​and model input values ​​for calculating model prediction errors, inputting the model input values ​​into the MD-LSTM model, outputting prediction values, and iterating the model multiple times using the training set to obtain the optimal MD-LSTM model.

[0029] Furthermore, step S7 specifically includes the following contents: The network input of the MD-LSTM model is the first 5 ∆U , ∆T , T' , SoH , as the charge and discharge cycle characteristics; the output is the charge and discharge cycle discharge process after the last 5 charge and discharge cycles SoH ; The model is an MD-LSTM model, which is a long short-term memory network (LSTM) hybrid fully connected network structure. The network structure diagram of the model is as follows: Figure 1As shown, it contains 6 hidden layers in total, among which 3 LSTM layers and 3 fully connected layers are arranged alternately, the first layer is the LSTM layer, the number of neurons is 128, the second layer is the fully connected layer, the number of neurons is 64, the third layer is the LSTM layer, the number of neurons is 64, the fourth layer is the fully connected layer, the number of neurons is 32, the fifth layer is the LSTM layer, the number of neurons is 32, and the sixth layer is the fully connected layer, the number of neurons is 5; the LSTM layer is at least configured with a forget gate, an input gate and an output gate, and the retention and forgetting of data are controlled by the forget gate, the input gate and the output gate; the structural unit expression of the model is: ; in, x t is the input of the current time step, h t-1 is the hidden state at the previous time step, C t-1 is the cell state at the previous time step, i t 、f t 、o t They are input gate, forget gate and output gate. is the candidate cell state at the current time step, C t is the cell state at the current time step, h t is the hidden state at the current time step, W f is the forget gate weight matrix, W i is the input gate weight matrix, W c is the current cell state weight matrix, W o is the output gate weight matrix, b f , b i , b c , b o They are the forget gate, input gate, current cell state and output gate bias. W d is the fully connected layer weight matrix, b d is the fully connected layer bias, f is the activation function of the fully connected layer (ReLU, Sigmoid, Tanh, etc.).

[0030] Furthermore, the loss function used in the model training process is the mean square error (MSE): ; in, f(x) is the model prediction value, y is the true value, n is the total number of samples.

[0031] Step S8, use the test set to verify the effectiveness of the MD-LSTM model; if the MAPE (mean absolute percentage error) on the test set is less than or equal to 3%, the test passes and the model is valid; if the MAPE on the test set is greater than 3%, change the experimental temperature conditions, test the battery under different temperature conditions, increase the amount of experimental data, and continue to repeat the above experimental steps until the MAPE on the test set is less than or equal to 3%, then end the experiment.

[0032] Furthermore, step S8 specifically includes the following contents: The process of validity verification is to conduct prediction error evaluation analysis based on the predicted SoH data and the actual SoH data, including the following steps: Calculate the mean absolute percentage error: ; in, represents the predicted value, 𝑦 𝑖 Indicates the actual value, j Represents the number of predicted values.

[0033] In order to more intuitively illustrate the solution of the present invention, the above solution is described in detail below in conjunction with specific implementation cases.

[0034] Lithium battery SoH prediction method based on multivariate time series LSTM, the flow chart is as follows Figure 2 As shown, the solution includes: In step S1, in this case, four 18650 lithium batteries are selected for experiment to obtain the rated capacity (2Ah) and OCV-SOC curve of the four lithium batteries; In step S2, the four lithium batteries are numbered as batteries 1, 2, 3, and 4, and the OCV-SOC curve of each battery is consulted to determine the charge and discharge cut-off voltage. The termination condition of the experiment is that the rated capacity of the battery drops to 1.4Ah. Among them, batteries 1, 2, and 3 are training set battery groups, and battery 4 is a test set battery group. The first three groups are tested at low temperature of 4°C, room temperature of 24°C, and high temperature of 44°C, respectively. The fourth group is tested at 30°C. The data collected from batteries 1, 2, and 3 are experimental training sets, and the data collected from battery 4 are experimental test sets (in order to verify the generalization ability of the model, the test set selects temperature experimental conditions different from those of the training set); In step S3, the current data, voltage data, temperature data, and time information during the discharge process of the experiment are recorded; the input variables in this case include: The hysteresis voltage per unit time after the battery is discharged ∆U , the unit time is one minute; Battery cell temperature changes during discharge ∆T ; Ambient temperature T' ; Battery health status at each discharge cycle SoH ; In step S4, data is preprocessed to establish a quaternary time series for each cycle of four lithium batteries; In step S5, a data set is constructed, and a battery aging data set D is constructed according to the quaternary time series of each cycle of step S4, where D={ ∆U,∆T,T',SoH}; In step S6, the data set D is divided into a training set (experimental data of batteries No. 1, 2, and 3) and a test set (experimental data of battery No. 4); In step S7, the MD-LSTM model is trained using the training set; In step S8, the MD-LSTM model is validated using the test set; if the MAPE on the test set is less than or equal to 3%, the test is passed and the model is valid; if the MAPE on the test set is greater than 3%, the experimental temperature conditions are changed, the battery is tested under different temperature conditions, the amount of experimental data is increased, and the above experimental steps are repeated until the MAPE on the test set is less than or equal to 3%, and the experiment is terminated; This case uses the above-mentioned No. 4 battery to verify the SoH model and output the prediction results; Figure 3 As shown in the figure, it can be clearly seen that the predicted value is close to the true value. The MAPE on the test set is 2.10%, and the overall prediction effect of the model is good.

[0035] In summary, the present invention proposes the characteristics of the changing hysteresis voltage from the voltage curve of the lithium battery discharge process, that is, the voltage value that rises within one minute after the battery reaches the cut-off voltage, and the characteristics of the temperature that changes during the discharge process, and uses these characteristics as neural network input, using LSTM network and DNN network to establish a lithium battery SoH prediction model based on MD-LSTM; and uses certain prior knowledge to improve the correlation between network input and output, helping the network to better understand the relationship between them, and improving the accuracy of the lithium battery SoH prediction model.

Claims

1. A lithium battery SoH prediction method based on MD-LSTM, characterized in that: The following steps are involved: Step S1, determining the rated capacity, discharge cut-off voltage and full charge voltage of the lithium battery; Step S2, performing a discontinuous charge and discharge experiment on the lithium battery, wherein the charge cut-off condition is that the battery rises to the full charge voltage, and the discharge cut-off condition is that the voltage drops to the discharge cut-off voltage, recording the data information during the experiment, and calculating the discharge capacity in each cycle according to the integral of the discharge current and the discharge time in the data information, calculating the ratio of the discharge capacity of each cycle of the lithium battery to the rated capacity, and obtaining the battery health state SoH of the lithium battery; the experiment termination condition is that the SoH drops to the end-of-life threshold; Step S3, obtaining input variables and output variables; The input variables are: {( ∆U i , ∆U i-1 , ... , ∆U i-4 ) , ( ∆T i , ∆T i-1 , ... , ∆T i-4 ) , ( T' i , T' i-1 , ... , T' i-4 ) , ( SoH i , SoH i-1 , ... , SoH i-4 )}； The output variables are: { SoH i+1 , SoH i+2 , ... , SoH i+5}； in, ∆U i It represents the hysteresis voltage data that changes per unit time during the i-th cycle discharge process. ∆T i It represents the temperature difference during the discharge process of the ith cycle. T' i represents the ambient temperature in the i-th cycle, SoH i Indicates the battery health status in the i-th cycle; Step S4, data preprocessing; during the preprocessing process, data cleaning is performed for each input variable and output variable, including identifying and processing errors, missing values ​​and outliers in the data, detecting and correcting erroneous data, and establishing a quaternary time series for each cycle; Step S5, constructing a battery aging data set D according to the quaternary time series of each cycle of step S4, D={ ∆U , ∆T , T' , SoH }; Step S6, splitting the battery aging data set D into a training set and a test set; Step S7, training the MD-LSTM model; the training process includes: sampling the quaternary time series in the battery aging data set D, generating actual values ​​and model input values ​​for calculating the model prediction error, inputting the model input values ​​into the MD-LSTM model, outputting the prediction values, and iterating the model multiple times using the training set to obtain the MD-LSTM model; Step S8: Use the test set to verify the effectiveness of the MD-LSTM model.

2. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S1, the rated capacity calibrated by the manufacturer when the lithium battery leaves the factory is obtained as the rated capacity, the OCV-SOC curve of the lithium battery is obtained, and the voltage when the SOC is 5% is determined as the discharge cut-off voltage of each experiment, and the voltage when the SOC is 90% is determined as the full charge voltage of each experiment.

3. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S2, the battery health state SoH formula of the lithium battery is expressed as: ； in, t 1 is the initial discharge time, t 2 is the discharge end time, I is the current during the discharge process, Q rated is the rated capacity of the lithium battery.

4. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S2, the life end threshold is 70%.

5. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S3, the hysteresis voltage changes ∆U It refers to the change in hysteresis voltage within 1 minute after the lithium battery stops discharging.

6. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S7, the network input of the MD-LSTM model is the first 5 ∆U , ∆T , T' , SoH , as a charge-discharge cycle characteristic; The output is the last 5 charge and discharge cycles during the charge and discharge cycle. SoH .

7. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S7, the MD-LSTM model is a long short-term memory network hybrid fully connected network structure, which includes 6 hidden layers, in which 3 LSTM layers and 3 fully connected layers are arranged alternately, the first layer is an LSTM layer with 128 neurons, the second layer is a fully connected layer with 64 neurons, the third layer is an LSTM layer with 64 neurons, the fourth layer is a fully connected layer with 32 neurons, the fifth layer is an LSTM layer with 32 neurons, and the sixth layer is a fully connected layer with 5 neurons; The LSTM layer is configured with at least a forget gate, an input gate, and an output gate. The forget gate, the input gate, and the output gate are used to control the retention and forgetting of data. The structural unit expression of the MD-LSTM model is: ； in, x t is the input of the current time step, h t-1 is the hidden state at the previous time step, C t-1 is the cell state at the previous time step, i t 、f t 、o t They are input gate, forget gate and output gate. is the candidate cell state at the current time step, C t is the cell state at the current time step, h t is the hidden state at the current time step, W f is the forget gate weight matrix, W i is the input gate weight matrix, W c is the current cell state weight matrix, W o is the output gate weight matrix, b f , b i , b c , b o They are the forget gate, input gate, current cell state and output gate bias. W d is the fully connected layer weight matrix, b d is the fully connected layer bias, f is the activation function of the fully connected layer.

8. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S7, the loss function used in the model training process is the mean square error MSE: ； in, f(x) is the model prediction value, y is the true value, n is the total number of samples.

9. The lithium battery SoH prediction method based on MD-LSTM according to claim 1, characterized in that: In step S8, if the MAPE on the test set is less than or equal to 3%, the test passes and the model is valid; if the MAPE on the test set is greater than 3%, the amount of data is increased and the above experimental steps are repeated until the MAPE on the test set is less than or equal to 3%, and the experiment ends.

10. The lithium battery SoH prediction method based on MD-LSTM according to claim 9, characterized in that: In step S8, the process of performing validity verification is to perform prediction error evaluation analysis based on the predicted SoH data and the actual SoH data, including the following steps: Calculate the mean absolute percentage error MAPE: ； in, represents the predicted value, 𝑦 𝑖 Indicates the actual value, j Represents the number of predicted values.

Images