Electric vehicle power battery SOC prediction method based on deep learning

Abstract

The invention discloses an electric vehicle power battery SOC prediction method based on deep learning. The method comprises the following steps: firstly, carrying out data cleaning on real vehicle data, selecting a plurality of characteristics such as voltages and temperatures with high correlation with the SOC to estimate the SOC, then carrying out fully adaptive noise set empirical mode decomposition on the estimated SOC, decomposing the SOC into a plurality of intrinsic mode functions (IMF) with different frequencies and amplitudes, and finally, carrying out adaptive noise set empirical mode decomposition on the plurality of IMFs. Meanwhile, key structure parameters in a time convolutional neural network (TCN) are optimized by using a white whale optimization algorithm (BWO) to obtain an optimal network structure, and finally, the IMF decomposed by fully adaptive noise ensemble empirical mode decomposition (CEEMDAN) and the screened optimal characteristic parameters are used as input characteristics of the TCN to carry out battery SOC prediction.

Description

Technical Field

[0001] The present invention relates to the field of electric vehicle power batteries, and in particular to a method for predicting the SOC of a power battery. Background Art

[0002] The battery state of charge (SOC) is the core of the battery state parameters, which quantifies the remaining available energy of the battery. It can be expressed as the percentage of the ratio of the remaining battery capacity Qt of the battery at the current moment to the rated total capacity Q0 of the battery. However, it cannot be obtained directly through measurement and can only be estimated indirectly by measuring parameters such as current and voltage.

[0003] In addition, battery SOC is an instantaneous indicator, which is continuously variable and has nonlinear dynamic changes. It is difficult to accurately estimate the battery SOC online. At the same time, other factors during the operation of electric vehicles will also make the online estimation of battery SOC more difficult. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the SOC of a power battery to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides an electric vehicle power battery SOC prediction method based on the CEEMDAN-BWO-TCN model, and the method steps are as follows.

[0006] First, relevant data is collected through the Internet of Vehicles platform, and the data is processed and cleaned. Then, several features with high correlation with battery SOC are selected as input through correlation analysis, and normalized. Specifically, the following are performed: Relevant data was collected through the Internet of Vehicles platform, and battery-related rechargeable energy storage device data, vehicle data, alarm data and extreme value data were selected.

[0007] Missing values ​​are handled by deleting the missing values ​​method, and outliers are handled by directly eliminating them.

[0008] In order to facilitate analysis, all data were resampled, the original data sampling frequency was compressed to once per hour, and the average SOC per hour was taken as the sampling point data.

[0009] A correlation analysis was performed on the battery data collected by the Internet of Vehicles platform, and the Pearson correlation coefficient, Kendall correlation coefficient, and Spearman correlation coefficient were calculated respectively. These three battery voltage characteristics (maximum voltage, minimum voltage, and total voltage) and two battery temperature characteristics (minimum temperature and maximum temperature) were selected as input features of this deep learning algorithm to estimate the battery SOC.

[0010] Normalize the battery characteristics. Here, the min-max normalization method is used to normalize the battery characteristics. The calculation formula is as follows: In the formula and are the minimum and maximum values ​​of each type of battery characteristic sequence data, is the battery characteristic value at the current time point, is the battery characteristic value after min-max normalization, is the battery SOC sequence output by the deep learning algorithm, are the battery characteristics after denormalization.

[0011] Furthermore, the main contents of the White Whale Optimization Algorithm (BWO) include: BWO sets a balance factor To switch between exploration and development modes, When , the algorithm is in the exploration stage, and when When , the algorithm switches to the development stage, and its calculation formula is as follows: In the formula is a random seed, and its value range is (0,1); is the maximum number of iterations of the algorithm. increase, The value range of will be reduced to (0,0.05).

[0012] Further, through The position update function of the BWO algorithm in the exploration phase is as follows: In the formula Expressed as The white whale in The position update value in the dimension; ( j =1,2,3,... d) For the total dimension The number of random dimensions selected; Indicates that the the location of the white whale; and is a random seed with a value range of (0,1); and They respectively represent the swimming status of the beluga whale.

[0013] Furthermore, in the development stage of the BWO algorithm, a flight strategy is introduced to enhance the convergence effect of the BWO algorithm, and its expression is as follows: In the formula is the random jump intensity; and is a random seed with a value range of (0,1); is the best position found during the iteration.

[0014] Furthermore, the flight function The calculation of is as follows: In the formula are two random numbers that are normally distributed; is a constant with a value of 1.5.

[0015] Furthermore, a whale fall phase is set, and its probability is defined as , which is calculated as follows: In the formula is the maximum number of iterations.

[0016] Furthermore, in the whale fall phase of the BWO algorithm, the white whale fall step length and position are used to update the position, and its mathematical model is as follows: In the formula , , is a random seed with a value range of (0,1); is the step size of the white whale position update; is the step size factor for position update; and They are the upper and lower bounds of the BWO algorithm when searching for the best solution.

[0017] Furthermore, the CEEMDAN method is used to decompose the preprocessed battery SOC sequence into 12 intrinsic mode function subsequences with different frequencies and amplitudes. The subsequence frequencies are distributed from high to low, showing a decreasing trend.

[0018] Furthermore, a CEEMDAN-BWO-TCN combined prediction model was constructed, which specifically includes: First, the battery data is processed and feature screening is performed, and finally it is input into the TCN model.

[0019] The preprocessed battery SOC sequence is decomposed into multiple intrinsic mode functions with different frequencies and amplitudes using the CEEMDAN method and finally input into the TCN model.

[0020] The BWO algorithm is used to optimize the key structural parameters of the neural network that are finally input into the TCN model to obtain the optimal network structure, and the optimal parameters are input into the TCN model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a model framework diagram for predicting battery SOC according to the present invention.

[0022] Figure 2 This is a flow chart of the BWO algorithm in the present invention.

[0023] Figure 3 This is the flow chart of battery SOC prediction by BWO-TCN of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present invention.

[0025] Reference Figure 1 The present invention provides a SOC prediction method of a power battery based on deep learning, comprising the following steps: First, relevant data was collected through a car company’s Internet of Vehicles monitoring platform, and battery-related rechargeable energy storage device data, vehicle data, alarm data, and extreme value data were selected.

[0026] The real vehicle data collected here does not have many missing values, so the data is directly processed by removing null values. Then the Z-score principle is used to detect outliers, mainly through the average value. and standard deviation The calculation formula is as follows: The detection shows that the outliers in the data are almost all zero values. Since there are few outliers and they almost never appear continuously, the outliers are directly eliminated.

[0027] The Internet of Vehicles platform, the source of data, has been connected to the National Monitoring and Management Platform for New Energy Vehicles. According to the regulations of the national testing platform, the sampling time interval for all data is 20s. However, the data collection frequency of once every 20s is too small, which is not conducive to subsequent analysis. Therefore, the sampling frequency needs to be reset and all data needs to be resampled. The original data sampling frequency is compressed to once every 1h, and the average SOC per hour is taken as the sampling point data.

[0028] The characteristic parameters such as total battery current, total voltage, maximum single cell voltage, minimum single cell voltage, battery temperature, mileage and insulation resistance are collected in the Internet of Vehicles monitoring platform. In the deep learning algorithm, in order to estimate the SOC of electric vehicle batteries, it is necessary to find battery features that have a greater impact on the battery SOC. Here, the collected battery data is analyzed for correlation, and the Pearson correlation coefficient, Kendall correlation coefficient and Spearman correlation coefficient are calculated respectively. These three battery voltage features (maximum voltage, minimum voltage and total voltage) and two battery temperature features (minimum temperature and maximum temperature) are selected as the input features of this deep learning algorithm to estimate the battery SOC.

[0029] In the constructed deep learning algorithm, five battery features such as total voltage, single cell voltage and temperature are required as input to estimate the battery SOC. Feature inputs of different dimensions will directly affect the accuracy of the output results of the deep learning algorithm and increase the training time. Here, the min-max normalization method is used to normalize the battery features. The calculation formula is as follows: In the formula and are the minimum and maximum values ​​of each type of battery characteristic sequence data, is the battery characteristic value at the current time point, is the battery characteristic value after min-max normalization, is the battery SOC sequence output by the deep learning algorithm, are the battery characteristics after denormalization.

[0030] The normalized characteristic parameters are input into the SOC prediction model.

[0031] The SOC collected in

[0037] is modally decomposed using the CEEMDAN method, and the SOC sequence is decomposed into 12 intrinsic mode function subsequences of different frequencies and amplitudes, which are input into the SOC prediction model.

[0032] While the above steps are being carried out, the relevant parameters of TCN are optimized through the BWO optimization algorithm, and the optimized parameters are input into the SOC prediction model to achieve the final SOC prediction.

[0033] Figure 2 The following is the BWO optimization algorithm flow: First, the BWO algorithm will initialize parameters, randomly select the location of the beluga whale, and calculate the fitness value based on the objective function.

[0034] BWO will calculate the value of the balance factor to determine whether the algorithm enters the exploration phase or the development phase. The calculation formula is as follows: In the formula is a random seed, and its value range is (0,1); is the maximum number of iterations of the algorithm. increase, The value range of will be reduced to (0,0.05).

[0035] When the balance factor is greater than 0.5, it will enter the exploration phase and update the position. The update function is as follows: In the formula Expressed as The white whale in Position update value in dimension; ( j =1,2,3,... d) For the total dimension The number of random dimensions selected; represents the position of the beluga whale in a random dimension; and is a random seed with a value range of (0,1); and They respectively represent the swimming status of the beluga whale.

[0036] When the balance factor is less than 0.5, it will enter the development stage. In the development stage, in order to enhance the convergence effect of the BWO algorithm, the flight strategy is introduced. Its expression is as follows: In the formula is the random jump intensity; and is a random seed with a value range of (0,1); is the best position found in the iterative process, where the flight function The calculation of is as follows: In the formula are two random numbers that are normally distributed; is a constant with a value of 1.5.

[0037] Then the fitness values ​​of the above stages are calculated to find the optimal fitness value.

[0038] Finally, the balance factor will be compared and whale fall probability To decide to enter the whale fall phase to update the position and determine whether the algorithm termination conditions are met. The calculation formula is: In the formula is the maximum number of iterations.

[0039] If the whale falls, the position is updated using the white whale's falling step length and position. The mathematical model is as follows: In the formula , , is a random seed with a value range of (0,1); is the step size of the white whale position update; is the step size factor for position update; and They are the upper and lower bounds of the BWO algorithm when searching for the best solution.

[0040] In the TCN algorithm, the selection of its hyperparameters will greatly affect the prediction accuracy of the algorithm. The hyperparameters of the TCN algorithm include the convolution kernel size, the number of convolution layers, and the dilation factor. The prediction accuracy is greatly affected by the parameters. By using the BWO algorithm to automatically optimize the above hyperparameters, the prediction accuracy of TCN can be greatly improved. Figure 3 The algorithm flow chart for optimizing TCN for BWO, the specific steps are as follows: Firstly, the electric vehicle data collected by the Internet of Vehicles platform is processed, and the battery features with a greater correlation with the battery SOC are selected as the input features of the algorithm through correlation coefficient calculation.

[0041] Set the algorithm optimization termination conditions and determine the maximum number of iterations and the maximum population size.

[0042] Set the constraints for optimizing the TCN neural network and determine the upper and lower boundaries of the TCN hyperparameters.

[0043] The objective function used to optimize the TCN neural network is set to predict the root mean square error (RMSE) of the battery SOC as the objective function of the White Whale algorithm for optimization. The optimal hyperparameters of TCN are determined by minimizing the RMSE. The calculation formula is as follows: In the formula Optimize the objective function of TCN for the BWO algorithm; is the root mean square error when predicting the battery SOC; , and They are the output sequence length, predicted SOC sequence and actual SOC sequence during TCN prediction, respectively.

[0044] Calculate the fitness value of the white whale, find the optimal fitness and optimal position at the current number of iterations, and save them at the same time.

[0045] Through the balance factor To determine the position update strategy of Beluga and calculate the fitness value at the current position.

[0046] Through the balance factor and whale fall probability To calculate the jump factor And update the position of the white whale and calculate the fitness value of the current position.

[0047] Determine whether the algorithm termination condition is met. If so, output the optimal fitness value and proceed to the next stage. If not, return to step

[0044] to continue iterating and optimizing until the termination condition is met, and then use the optimized parameters in the SOC prediction model.

[0048] Finally, it should be noted that the embodiments described above are only descriptions of the scheme of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical scheme of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for predicting SOC of electric vehicle power battery based on deep learning, characterized in that: It includes the following steps: S1. Collect relevant data through the Internet of Vehicles platform, process and clean the data, and then select several features with high correlation with battery SOC as input through correlation analysis, and normalize them; S2. Based on the TCN neural network optimized by the Beluga algorithm (BWO), the SOC of the real vehicle power battery is estimated, and the optimal value of the TCN parameters is found with the minimization of RMSE as the objective function; S3. CEEMDAN is used to decompose the SOC sequence. The SOC sequence is decomposed into 12 subsequences to maximize the extraction of the data characteristics of the battery SOC sequence and reduce the impact of noise data; S4. Construct a CEEMDAN-BWO-TCN combined prediction model, input the screened features, the intrinsic mode function IMF decomposed by CEEMDAN, and the key structural parameters in the TCN neural network optimized by the BWO algorithm into the prediction module for SOC prediction.

2. The electric vehicle power battery SOC prediction method based on deep learning according to claim 1 is characterized in that: The S1 comprises the following steps: S2-1. Collect relevant data through the Internet of Vehicles platform, and select battery-related rechargeable energy storage device data, vehicle data, alarm data and extreme value data; S2-2, missing values ​​are processed by deleting missing values, and outliers are processed by directly eliminating outliers; S2-3. For the convenience of analysis, all data are resampled, the original data sampling frequency is compressed to once every hour, and the average SOC per hour is taken as the sampling point data; S2-4. Perform correlation analysis on the battery data collected by the Internet of Vehicles platform, calculate the Pearson correlation coefficient, Kendall correlation coefficient and Spearman correlation coefficient respectively, and select these three battery voltage characteristics (maximum voltage, minimum voltage and total voltage) and two battery temperature characteristics (minimum temperature and maximum temperature) as the input characteristics of this deep learning algorithm to estimate the battery SOC; S2-5. Normalize the battery characteristics. Here, the min-max normalization method is used to normalize the battery characteristics.

3. The electric vehicle power battery SOC prediction method based on deep learning according to claim 1 is characterized in that: The S2 comprises the following steps: The White Whale Optimizer (BWO) sets a balancing factor To switch between exploration and development mode, When , the algorithm is in the exploration stage, and when When , the algorithm switches to the development stage, and its calculation formula is as follows: In the formula is a random seed, and its value range is (0,1); is the maximum number of iterations of the algorithm. increase, The value range of will be reduced to (0,0.05); S3-2, the position update function of the BWO algorithm exploration phase is as follows: In the formula Expressed as The white whale in the The position update value in the dimension; ( j =1,2,3,... d) For the total dimension The number of random dimensions selected; Indicates that the the location of the white whale; and is a random seed with a value range of (0,1); and They respectively represent the swimming status of the beluga whale; S3-3, in the development stage of the BWO algorithm, the flight strategy is introduced to enhance the convergence effect of the BWO algorithm. Its expression is as follows: In the formula is the random jump intensity; and is a random seed with a value range of (0,1); is the best position found during the iteration. Flight function The calculation of is as follows: In the formula are two random numbers that are normally distributed; is a constant with a value of 1.5; S3-4, the whale fall stage of the BWO algorithm uses the white whale fall step length and position to update the position. Its mathematical model is as follows: In the formula, , , is a random seed with a value range of (0,1), is the step size of the white whale position update; is the step size factor for position update; and are the upper and lower bounds of the BWO algorithm respectively. The probability of a white whale falling is defined as , which is calculated as follows: In the formula, is the maximum number of iterations.

4. The electric vehicle power battery SOC prediction method based on deep learning according to claim 1 is characterized in that: The S3 comprises the following steps: S4-1. Use the CEEMDAN method to decompose the preprocessed battery SOC sequence into 12 intrinsic mode function subsequences with different frequencies and amplitudes; the subsequence frequencies are distributed from high to low, showing a decreasing trend.

5. The electric vehicle power battery SOC prediction method based on deep learning according to claim 1 is characterized in that: The S4 comprises the following steps: S5-1, process battery data and perform feature screening, and finally input it into the TCN model; S5-2, using the CEEMDAN method to decompose the preprocessed battery SOC sequence into multiple intrinsic mode functions of different frequencies and amplitudes, and finally input them into the TCN model; S5-3. Use the BWO algorithm to optimize the key structural parameters of the neural network that are finally input into the TCN model to obtain the optimal network structure, maximize the prediction accuracy of the TCN model, and finally input it into the TCN model.

Images