A neural network fused battery energy storage system soc evaluation method

Abstract

The application provides a battery energy storage system SOC evaluation method based on neural network fusion. The SOC data are decomposed in multiple time scales by using an empirical mode decomposition algorithm, and then complexity analysis is performed by using sample entropy. The components with similar complexity are classified and aggregated by taking the average sample entropy as an index. Furthermore, six input factors, namely, total voltage, current, insulation resistance, average voltage, average temperature and cumulative charging capacity, are used to perform SOC sub-band prediction by using a multi-time scale neural network fusion algorithm, so that the evaluation accuracy of the battery energy storage system SOC is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance of battery energy storage systems, and specifically relates to a neural network fusion-based SOC evaluation method for battery energy storage systems. Background Technology

[0002] In response to the dual-carbon strategy and to promote the clean energy transition, electrochemical energy storage, as a flexible energy storage resource, has developed rapidly in recent years. Among them, lithium batteries have gradually become the main force due to their advantages such as high energy density and long cycle life. Accurate State of Charge (SOC) assessment is indispensable for improving the application performance of battery energy storage systems. SOC describes the remaining usable capacity of a battery in its current state, and for energy storage systems, it represents the duration for which the battery can be continuously charged or discharged under specific conditions. Therefore, SOC estimation can update the remaining capacity of the battery in a timely manner, prevent overcharging and over-discharging, improve the charging and discharging efficiency of the battery, and extend its lifespan. Although many studies on SOC assessment have been conducted, the main focus is on individual battery cells, with relatively little research on battery energy storage systems. Battery energy storage systems are more complex than individual battery cells in terms of cell number, topology, element coupling degree, and operating characteristics, making analysis more difficult. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, this invention proposes a neural network fusion-based method for evaluating the State of Charge (SOC) of battery energy storage systems. The method utilizes an empirical mode decomposition algorithm to decompose SOC data across multiple time scales, then performs complexity analysis using sample entropy. Using the average sample entropy as an indicator, components with similar complexity are classified and aggregated. Finally, using six input factors—total voltage, current, insulation resistance, average voltage, average temperature, and cumulative charging capacity—a multi-neural network fusion algorithm is employed to predict SOC in different frequency bands, thereby improving the accuracy of SOC evaluation for battery energy storage systems.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A neural network fusion-based method for evaluating the State of Charge (SOC) of a battery energy storage system is proposed. This method utilizes an empirical mode decomposition algorithm to decompose SOC data into multiple time scales, then performs complexity analysis using sample entropy. The average sample entropy is used as an indicator to classify and aggregate components with similar complexity. Finally, using six input factors—total voltage, current, insulation resistance, average voltage, average temperature, and cumulative charging capacity—a multi-time scale neural network fusion algorithm is employed to predict SOC in different frequency bands, thereby improving the accuracy of SOC evaluation for battery energy storage systems.

[0006] Furthermore, deviation standardization is used to linearly transform the input and output data so that their values ​​are between [0,1].

[0007]

[0008] In the formula, data is a certain input sample; max(data) is the maximum value of the input sample; min(data) is the minimum value of the input sample; and d is the normalized data value.

[0009] Furthermore, to evaluate the prediction accuracy of the SOC of the battery energy storage system, the root mean square error (RMSE) is used, and the calculation formula is shown below:

[0010]

[0011] In the formula, y i For actual data; y i ′ represents the predicted data; n represents the amount of data.

[0012] Furthermore, the empirical mode decomposition algorithm includes:

[0013] The empirical mode decomposition algorithm decomposes the signal sequence into a series of intrinsic mode functions c that are symmetric about the time axis. i (i = 1, 2, ..., n) and residual trend component r n The sum; the intrinsic mode function satisfies the following conditions: 1) the number of extreme points and zero-crossing points are equal or differ by no more than 1 throughout the entire data sequence; 2) at any time point, the local mean is zero; the decomposition terminates when one of the following requirements is met: 1) c i or r n Less than the predetermined value; 2)r n It becomes a monotonic function, and the intrinsic mode function can no longer be obtained;

[0014] The expression for signal x(t) after decomposition using the empirical mode decomposition algorithm is:

[0015]

[0016] The SOC curve was decomposed using the empirical mode decomposition algorithm to obtain 12 IMF components and 1 residual component. The frequencies of the 12 IMF components decreased sequentially, while the residual component monotonically decreased.

[0017] Furthermore, the sample entropy is calculated as follows:

[0018] Assume a time series {x} i Let x(1), x(2), ..., x(N), where N is the number of data points:

[0019] 1) Transform the sequence {x} i The vectors are arranged in order to form an m-dimensional vector, namely X(i)=[x(i),x(i+1),…,x(i+m-1)], where i=1,2,…,N-m+1;

[0020] 2) Define the distance d between X(i) and X(j) m (X(i), X(j)) is the maximum value of the difference between the corresponding elements of the two:

[0021]

[0022] For each i value, calculate d between X(i) and the remaining vectors X(j) (j = 1, 2, …, N - m + 1 & j ≠ i) m (X(i), X(j));

[0023] 3) Given a similarity tolerance r (r > 0), count the number of d m (X(i), X(j)) < r, and then calculate its ratio to the total number of distances N - m, denoted as

[0024]

[0025] where j = 1, 2, …, N - m + 1 & j ≠ i; num is the number of d m (X(i), X(j)) < r;

[0026] This process is defined as the X(i) template matching process, representing the matching probability of any X(j) with the template;

[0027] 4) Obtain The average value of is:

[0028] I

[0029] 5) Increase the dimension to m + 1, and repeat steps 1) to 3), then the average value of B m+1 (r) is:

[0030]

[0031] Thus, the sample entropy definition is obtained:

[0032]

[0033] When N takes a finite value, the sample entropy estimated value is:

[0034]

[0035] Calculate the value of each IMF component using the sample entropy, thus dividing the SOC curve into a high-frequency segment and a low-frequency segment. Among them, the high-frequency segment consists of the five components with the largest sample entropy among the 12 IMF components, and the low-frequency segment consists of the remaining seven components plus the residual component among the 12 IMF components.

[0036] Furthermore, the multi-timescale neural network SOC prediction is achieved through a feedforward neural network and a long short-term memory neural network;

[0037] The high-frequency bands are predicted using the feedforward neural network and the long short-term memory neural network, respectively.

[0038] Beneficial effects:

[0039] Existing conventional analyses primarily focus on individual battery cells, with limited SOC analysis of battery energy storage systems. Most neural network prediction methods employ single models, making it difficult to maintain consistent performance across different research subjects. This invention utilizes a fusion of multiple neural networks for frequency-band SOC prediction of battery energy storage systems, improving prediction accuracy. While conventional analyses primarily use voltage, current, and temperature data from individual battery cells as model inputs, this invention employs six data types: total voltage, current, insulation resistance, average voltage, average temperature, and cumulative charging capacity, providing a more multi-dimensional and richer data type. Empirical mode decomposition is used to divide the data into different frequency bands, but there is no fixed method for frequency band reconstruction. This invention proposes using the average sample entropy as an evaluation index to achieve classification and aggregation of high-frequency and low-frequency bands. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the State of Charge (SOC) of a battery energy storage system.

[0041] Figure 2 This refers to the total voltage of the battery energy storage system.

[0042] Figure 3 This refers to the current in the battery energy storage system.

[0043] Figure 4 Insulation resistance of battery energy storage system;

[0044] Figure 5 This represents the average voltage of the battery energy storage system.

[0045] Figure 6 The average temperature of the battery energy storage system;

[0046] Figure 7 Accumulated charging capacity for the battery energy storage system;

[0047] Figure 8a , Figure 8b , Figure 8c For IMF and residual components; where, Figure 8a These are the IMF1-IMF4 components; Figure 8b These are IMF5-IMF8 components; Figure 8c For IMF9-IMF12 components and residual components;

[0048] Figure 9 The sample entropy of each IMF component;

[0049] Figure 10 For SOC high-frequency components;

[0050] Figure 11 This refers to the low-frequency components of the SOC.

[0051] Figure 12 The results are for high-frequency component prediction based on backpropagation (BP).

[0052] Figure 13 The results are for low-frequency component prediction based on LSTM.

[0053] Figure 14 The results are predictions from a combination of BP and LSTM. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0055] The neural network fusion-based SOC assessment method for battery energy storage systems of this invention selects six parameters—total voltage, current, insulation resistance, average voltage, average temperature, and cumulative charging capacity—to predict the SOC (State of Charge). The sampling interval is 1 minute, and the duration is 5760 minutes. Figures 1 to 7 As shown, the State of Charge (SOC) of a battery energy storage system is predicted using a BP neural network and a long short-term memory neural network, with the first 80% used as training data and the last 20% used as test data.

[0056] To better utilize data-driven algorithms for SOC evaluation, sample data preprocessing is necessary to prevent large gradient updates. This invention employs deviation normalization, linearly transforming the input and output data to ensure their values ​​fall within the range [0,1].

[0057]

[0058] In the formula, data represents a specific input sample; max(data) is the maximum value of the input sample; min(data) is the minimum value of the input sample; and d is the normalized data value. To evaluate the prediction accuracy of the SOC of a battery energy storage system, the root mean square error (RMSE) is used, and the calculation formula is shown below:

[0059]

[0060] In the formula, y i For actual data; y i ′ represents the predicted data; n represents the amount of data.

[0061] This invention utilizes an empirical mode decomposition algorithm to decompose State of Charge (SOC) data across multiple time scales. Then, it employs sample entropy for complexity analysis, using the average sample entropy as an indicator to classify and aggregate components with similar complexity. Finally, it uses six input factors—total voltage, current, insulation resistance, average voltage, average temperature, and cumulative charging capacity—and a multi-time-scale neural network fusion algorithm to predict SOC in different frequency bands, thereby improving the accuracy of SOC assessment for battery energy storage systems. The empirical mode decomposition algorithm is as follows:

[0062] Empirical Mode Decomposition (EMD) does not require predefined basis functions. It can obtain localized features simply by decomposing the signal based on its own time-scale characteristics, making it very suitable for processing nonlinear and non-stationary random signals.

[0063] EMD decomposes a signal sequence into a series of intrinsic mode functions (IMFs) that are symmetric about the time axis. i (i = 1, 2, ..., n) and residual trend component r n The sum of the values. The IMF (Inter-Functional Method) must satisfy the following conditions: 1) The number of extreme points and zero-crossing points is equal or differs by no more than 1 throughout the entire data sequence; 2) At any given time point, the local mean is zero. The decomposition terminates if one of the following conditions is met: 1) c i or r n Less than the predetermined value; 2)r n It becomes a monotonic function, and its intrinsic mode function can no longer be obtained. In summary, the expression for signal x(t) after EMD decomposition is:

[0064]

[0065] Using the EMD algorithm Figure 1 The SOC curve in the image is decomposed, and the decomposed curve is shown in Figure 8, consisting of 12 IMF components and 1 residual component. From IMF1 to IMF12, the frequency decreases sequentially, and the residual component monotonically decreases.

[0066] The sample entropy includes:

[0067] The more complex the data sequence, the greater the sample entropy, and vice versa. The principle is as follows:

[0068] Assume a time series {x} i Let x(1), x(2), ..., x(N) be the number of data points.

[0069] 1) Transform the sequence {x}i} They are sequentially composed into an m-dimensional vector, that is, X(i) = [x(i), x(i+1), …, x(i+m-1)], where i = 1, 2, …, N-m+1.

[0070] 2) Define the distance d between X(i) and X(j) m (X(i), X(j)) as the maximum value of the differences between the corresponding elements of the two:

[0071]

[0072] For each i value, calculate d between X(i) and the remaining vectors X(j) (j = 1, 2, …, N-m+1 & j ≠ i) m (X(i), X(j)).

[0073] 3) Given a similarity tolerance r (r > 0), count the number of d m (X(i), X(j)) < r, and then calculate its ratio to the total number of distances N-m, denoted as B i m (r):

[0074]

[0075] In the formula, j = 1, 2, …, N-m+1 & j ≠ i; num is the number of d m (X(i), X(j)) < r. This process is defined as the template matching process of X(i), and B i m (r) represents the matching probability of any X(j) with the template.

[0076] 4) Further obtain the average value of B i m (r) as:

[0077]

[0078] 5) Increase the dimension to m+1, and repeat steps 1) to 3), then the average value of B m+1 (r) is:

[0079]

[0080] Thus, the definition of sample entropy can be obtained:

[0081]

[0082] When N takes a finite value, the estimated value of sample entropy is:

[0083]

[0084] The value of sample entropy is related to m and r, but their consistency is good. The trend of entropy value change is not affected by m and r. In this invention, m = 2 and r = 0.2SD (r is generally 0.1 to 0.25SD, where SD is the standard deviation of the time series).

[0085] The value of each IMF component is calculated using sample entropy, such as... Figure 9 As shown, the sample entropy values ​​generally show a gradual decreasing trend as the frequency of the IMF component bands decreases. This means that as the frequency decreases, data fluctuations decrease, complexity decreases, and self-similarity improves. IMF1-IMF5 have relatively large sample entropy values, with a significant decrease between IMF5 and IMF6. IMF6-IMF12 have relatively small sample entropy values, decreasing roughly linearly, with all values ​​less than 0.1. Furthermore, the average sample entropy is calculated to be 0.3028. The sample entropy values ​​of IMF1-IMF5 are above the average, while those of IMF6-IMF12 are all below the average. Therefore, the SOC curve is divided into two frequency bands: a high-frequency band composed of the five components of IMF1-IMF5, and a low-frequency band composed of the seven components of IMF6-IMF12 plus a residual component. The high-frequency band is shown below. Figure 10 As shown, the low-frequency band is as follows Figure 11 As shown.

[0086] The multi-timescale neural network fusion algorithm for SOC prediction includes:

[0087] A feedforward (BP) neural network consists of an input layer i, one or more hidden layers j, and an output layer O. It has efficient nonlinear mapping ability, generalization ability, and good fault tolerance.

[0088] Long Short-Term Memory (LSTM) neural networks can efficiently learn and acquire the nonlinear characteristics of time-series data, and can effectively solve the problems of gradient explosion or vanishing gradients.

[0089] Two strategies, BP and LSTM, were used to predict high and low frequency bands respectively, and the prediction results are shown in Table 1. BP significantly improved the accuracy of high-frequency band prediction, with an RMSE of only 0.0463; for low-frequency band prediction, the RMSE was 0.1529. Reconstruction of the predicted high-frequency and low-frequency band components resulted in a reconstructed RMSE of 0.01442 due to the complementarity of the data. This demonstrates that segmented prediction is more accurate than direct prediction. With LSTM, the prediction errors for both high and low frequency bands were similar to those of direct prediction, but the overall prediction accuracy improved after reconstruction. Therefore, after SOC frequency segmentation, the regularity of each band is enhanced, and the accuracy is improved to some extent when using the same neural network algorithm for prediction. Furthermore, the data complementarity after reconstruction further enhances the accuracy of the reconstructed data.

[0090] Table 1. Prediction results of high and low frequency bands using different algorithms

[0091]

[0092] To further verify the advantages of frequency band prediction, since BP calculation is relatively fast, prediction of high-frequency components is performed, such as... Figure 12 ; LSTM is used to predict low-frequency components, such as Figure 13 After prediction, data reconstruction is performed, such as... Figure 14 As shown in Table 2, combining different neural networks can also improve the accuracy of SOC prediction. Therefore, using frequency-band prediction can improve the accuracy of SOC prediction compared to direct prediction. Furthermore, in terms of computation time, BP takes 0.1672s for high-frequency prediction, and LSTM takes 235.18s for low-frequency prediction, for a total of approximately 235.35s. This is 37.03% faster than LSTM's 373.75s without frequency-band prediction.

[0093] Table 2 Prediction Results of the Combined Algorithm

[0094]

[0095] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the State of Charge (SOC) of a battery energy storage system using neural network fusion, characterized in that, The SOC data is decomposed into multiple time scales using the empirical mode decomposition algorithm. Then, the complexity is analyzed using sample entropy. The average sample entropy is used as an indicator to classify and aggregate components with similar complexity. Finally, using six input factors, namely total voltage, current, insulation resistance, average voltage, average temperature and cumulative charging capacity, the SOC is predicted in different frequency bands through a multi-time scale neural network fusion algorithm, thereby improving the SOC evaluation accuracy of the battery energy storage system. The sample entropy is calculated as follows: Assuming time series for ,in For data volume: 1) Change the sequence Composed in order 3D vector, i.e. ,in ; 2) Definition and Distance between The maximum difference between corresponding elements of the two: (4) For each Value Calculation With the remaining vectors Between ; 3) Given similarity tolerance Statistical analysis of each Value The number of [items] is then calculated, and its distance is compared with the total number of [distances]. The ratio of is denoted as : (5) In the formula, ; for quantity; This process is defined as Template matching process, Indicate any one The probability of matching the template; 4) Obtain The average value is: (6) 5) Increase the dimension to Repeat steps 1) to 3), then The average value is: (7) Thus, the definition of sample entropy is obtained: (8) when When taking finite values, the estimated sample entropy is: (9) The value of each IMF component is calculated using sample entropy, thereby dividing the SOC curve into a high-frequency segment and a low-frequency segment. The high-frequency segment consists of the five components with the largest sample entropy among the 12 IMF components, and the low-frequency segment consists of the remaining seven components and residual components among the 12 IMF components. The multi-timescale neural network SOC prediction is achieved through a feedforward neural network and a long short-term memory neural network. The feedforward neural network and the long short-term memory neural network are used to predict the high and low frequency bands, respectively.

2. The method for evaluating the State of Charge (SOC) of a battery energy storage system based on neural network fusion according to claim 1, characterized in that, By using deviation standardization, the input and output data are linearly transformed so that their values ​​are between [0,1]. (1) In the formula, data is a certain input sample; The maximum value of the input sample; The minimum value of the input sample; These are the normalized data values.

3. The method for evaluating the State of Charge (SOC) of a battery energy storage system based on neural network fusion according to claim 1, characterized in that, To evaluate the prediction accuracy of the State of Charge (SOC) of a battery energy storage system, the Root Mean Square Error (RMSE) is used, and the calculation formula is shown below: (2) In the formula, This is actual data; For the predicted data; n is the amount of data.

4. The method for evaluating the State of Charge (SOC) of a battery energy storage system based on neural network fusion according to claim 1, characterized in that, The empirical mode decomposition algorithm includes: The empirical mode decomposition algorithm decomposes the signal sequence into a series of intrinsic mode functions that are symmetric about the time axis. With residual trend components The sum; the intrinsic mode function satisfies the following conditions: 1) the number of extreme points and zero-crossing points is equal or differs by no more than 1 throughout the entire data sequence; 2) at any time point, the local mean is zero; the decomposition terminates when one of the following requirements is met: 1) or Less than the predetermined value; 2) It becomes a monotonic function, and the intrinsic mode function can no longer be obtained; Signal The expression after decomposition by the empirical mode decomposition algorithm is: (3) The SOC curve was decomposed using the empirical mode decomposition algorithm to obtain 12 IMF components and 1 residual component. The frequencies of the 12 IMF components decreased sequentially, while the residual component monotonically decreased.

Images