Flow battery voltage inflection point SOC verification method based on GBDT machine learning prediction

Abstract

The invention discloses a liquid flow battery voltage inflection point SOC verification method based on GBDT machine learning prediction, and belongs to the technical field of an application foundation combining power electronics and long-time energy storage control. Aiming at the defect that the current ZBFB does not have a formed SOC accurate measurement technology and an evaluation algorithm, the GBDT algorithm in machine learning is applied to SOC real-time calculation and verification of the ZBFB, the work can effectively improve the SOC calculation precision at the voltage inflection point of the ZBFB, and a reference can be provided for developing a universal flow battery SOC algorithm. Compared with the prior art, the method has the following advantages: (1) according to a plurality of groups of representative characteristic data, high-precision evaluation of the SOC at the ZBFB inflection point is realized by adopting a GBDT verification algorithm matched with SOC measurement at the ZBFB discharge voltage inflection point; and (2) the GBDT algorithm verification and timeliness judgment criterion based on limit cycle analysis is applied, the problem that the GBDT algorithm consumes long time is solved, and calculation support is provided for improving the timeliness of the GBDT from the algorithm level.

Description

Technical Field

[0001] The present invention belongs to the application basic technology field combining power electronics with long-term energy storage control, and relates to a high-precision evaluation and verification method of SOC (State of Charge) at a battery discharge voltage inflection point based on GBDT (Gradient Boosting Decision Tree). Background Art

[0002] Accurate measurement of the state of charge (SOC) of zinc-bromine flow battery (ZBFB) is very effective in improving battery reliability, achieving series voltage balancing and suppressing parallel circulating current, and promoting its application in the field of long-term energy storage. However, due to the inherent electrochemical reaction characteristics of the battery itself and the influence of self-discharge, coulomb efficiency, electrolyte temperature, flow rate and discharge rate, an inevitable voltage inflection point will occur in the late stage of discharge. Limited by the sampling frequency of BMS (Battery Management System), the sudden drop in voltage at the inflection point can easily lead to distortion of the sampled voltage and current, thereby affecting the calculation accuracy of SOC. This is difficult to completely solve with the classic SOC calculation method and current measurement technology. Therefore, it is urgent to study a feasible method to provide a solution for SOC identification at the inflection point of ZBFB discharge voltage.

[0003] GBDT is a high-precision prediction algorithm based on time series. Its advantages lie in prediction accuracy, robustness and small sample computing capability. However, due to the real-time adjustment of parameters such as the number and depth of trees and learning rate, the training time is long, which in turn affects the real-time performance of battery SOC monitoring. This is one of the bottlenecks in applying it to SOC identification at the ZBFB voltage inflection point. Previous studies have shown that not all SOC at the ZBFB discharge voltage inflection points are distorted, but only some inflection points have this situation. Since the ZBFB discharge voltage inflection point data has the characteristics of small samples, it is only necessary to perform GBDT identification and correction on the inflection points of SOC distortion, which can greatly reduce the computational cost and running time of GBDT. Summary of the invention

[0004] The purpose of this invention is to solve the bottleneck problem that the classical SOC calculation method cannot achieve high-precision prediction due to the influence of the sampling frequency of the BMS, and to provide a liquid flow battery voltage inflection point SOC identification method based on limit cycle assisted GBDT analysis, which can provide a reference for the development of a general liquid flow battery SOC high-precision calculation algorithm. This method focuses on solving the following problems: ① Developing an efficient GBDT solution algorithm for battery SOC identification based on ZBFB discharge voltage and current; ② Researching a solution method for SOC distortion judgment at the ZBFB discharge voltage inflection point based on limit cycle analysis.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A flow battery voltage inflection point SOC verification method based on GBDT machine learning prediction, the method comprising the following steps:

[0007] Step 1: Establish a comprehensive experimental platform for zinc-bromine flow batteries, including the zinc-bromine flow battery body, BMS, EMS, PCS and load AC motor; among them, BMS is used to monitor the charge and discharge data of the battery body, EMS is used to power the positive and negative electrode circulation pumps of the zinc-bromine flow battery, PCS is used to realize DC / AC power conversion, and the load AC motor is used to simulate the inductive load;

[0008] Step 2: Clarify the measured parameters and identification parameters. The measured parameters include voltage, current, circulation pump speed, charge and discharge time and measured SOC. The identification parameter is the SOC at the voltage inflection point. Use the battery charge and discharge comprehensive tester and BMS to monitor the experimental platform in real time for 24 hours, with a sampling period of 1 minute, to obtain the curves of measured SOC, battery voltage and battery current data changing with time.

[0009] Step 3: Select limited voltage and current sampling data in multiple cycles of ZBFB discharge to identify discharge change characteristics; extract limited feature samples at regular intervals in a whole discharge cycle to avoid sensitive outliers;

[0010] Step 4: Take the extracted limited feature samples as the data set, and give the loss function and learning rate of GBDT in real time. Use GBDT to solve the optimal SOC value of ZBFB and develop a SOC solution paradigm that can realize real-time verification.

[0011] Step 5: Apply limit cycle theory to derive the judgment and verification criteria for using GBDT to calculate the high-precision SOC of ZBFB. By reducing the number of GBDT applications, its calculation efficiency in a complete charge and discharge cycle can be improved to improve the real-time performance of the proposed GBDT algorithm.

[0012] Furthermore, in step 2, the sampling period is 1 minute, and the sampled measured data include the voltage, current, circulation pump speed, charge and discharge time, and measured SOC of the zinc-bromine flow battery experimental platform.

[0013] Furthermore, in step 3, {(t1, SOC1), (t2, SOC2), … (t k ,SOC k ),…,(t N ,SOC N )} A total of N representative small sample feature data.

[0014] Considering that too many trees and too complex depth will lead to a long training time, which will affect the real-time performance of battery SOC monitoring. After feature extraction, the number of small sample feature data is N=6, and the corresponding voltage and current sample data sets are obtained based on this, and sorted from low to high according to the collected voltage values:

[0015] D={(u1,i1),(u2,i2),…,(u k ,i k ),…,(u N ,i N )}.

[0016] Furthermore, in step 4, the SOC of the ZBFB in the time domain after the voltage inflection point is evaluated in real time, and the implementation process of the constructed GBDT model is as follows:

[0017] (1) Input feature sample data D = {(x1, y1), (x2, y2), … (x k ,y k ),...,(x N ,y N )}, input loss function L(y,f(x)); where N is the number of samples, (x k ,y k ) represents the kth feature sample data, 1≤k≤N, χ represents the input space; 1≤k≤N, Y represents the output space; y and f(x) represent the output data and the learner function respectively;

[0018] (2) Initialize weak learner:

[0019]

[0020] And initialize the number of iterations m = 1; where f0(x) is the initialization weak learner function, x is the input feature, and c is the constant taken when the loss function is minimized;

[0021] (3) For any sample data (x k ,y k ), the residual is calculated as:

[0022]

[0023] Among them, r m,k is the pseudo residual of the k-th feature sample data in the m-th iteration, that is, is the negative gradient of the loss function;

[0024] (4) The calculated residual value is used as the true value of the next training, and the fitting data {(xk ,r m,k )} k＝1,2,...,N As the sample data for the next training, a new regression tree T is obtained m , and its corresponding leaf node area is R m,k ,k=1,2,…,J m , where J m is the number of leaf nodes of the mth regression tree;

[0025] (5) Apply regression tree T m Calculate the best fit value for each leaf node of:

[0026]

[0027] Among them, c m,k is the best fitting value of the leaf area, f m-1 (x k ) represents the sample data x given by the learner in the m-1th iteration k The predicted value of

[0028] (6) Refresh the strong learner:

[0029]

[0030] Among them, f m (x) is the learner of the mth iteration, I is an indicator whose value is 1, which is used to fit the classifier in the node area;

[0031] (7) After completing M rounds of iterations, the final boosted regression tree is obtained:

[0032]

[0033] Among them, M is the number of iterations, that is, the number of trees;

[0034] (8) Obtain and analyze the optimal decision tree

[0035] (9) Verification by comparison with basic samples

[0036] (10) If If the predicted value is reasonable, output Otherwise, m=m+1, and the above calculation process is repeated starting from (3).

[0037] According to the algorithm solution process of GBDT, the initialization of the first weak learner should minimize the loss function. Calculate the residual of the measured SOC relative to the first weak learner, and find the optimal partition node of the sample data based on the traversal value of each feature data; this procedure starts at voltage u1 and ends at current i6;

[0038] In order to evaluate the SOC in the time domain after the voltage inflection point, u4 and i4 are selected as sample partitioning nodes for the purpose of optimizing the left and right residuals of the partitioning node, that is, (u4, i4) is selected as the optimal partitioning node; further, based on the minimum variance principle and the traversal value analysis of the left leaf node and the right leaf node, a tree diagram corresponding to the time domain after the voltage inflection point is obtained; based on the discharge current at the SOC to be evaluated, the optimal path for SOC evaluation is selected.

[0039] Furthermore, in step 4, considering that the coulombic efficiency of the battery varies with the discharge current, the learning rate α of GBDT is finally selected to be 0.458; the optimal decision tree of ZBFB at the voltage inflection point can be calculated as shown in the following formula:

[0040]

[0041] Furthermore, in step five, within a complete charge and discharge cycle, the necessity of GBDT for SOC verification is determined by analyzing the discharge voltage limit cycle and the discharge current limit cycle in different discharge time periods, and the calculation efficiency of GBDT within a complete charge and discharge cycle is improved by reducing the number of applications of GBDT.

[0042] In view of the current lack of a mature SOC accurate measurement technology and evaluation algorithm for ZBFB, the present invention applies the GBDT algorithm in machine learning to the real-time calculation and verification of SOC of ZBFB. The work done can effectively improve the SOC calculation accuracy at the voltage inflection point of ZBFB, and can provide a reference for the development of a general liquid flow battery SOC algorithm. Compared with the existing technology, it has the following advantages: (1) According to several groups of representative characteristic data, a GBDT verification algorithm matching the SOC measurement at the ZBFB discharge voltage inflection point can be used to achieve high-precision evaluation of SOC at the ZBFB inflection point; (2) The GBDT algorithm verification and timeliness judgment criteria based on limit cycle analysis are applied to solve the problem of long time consumption of GBDT algorithm, and provide a calculation support for improving the timeliness of GBDT from the algorithm level. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of the method implemented by the present invention;

[0044] Figure 2 A diagram of the comprehensive experimental platform developed for the laboratory, including the ZBFB body, control circuit and AC load;

[0045] Figure 3 It is the measured SOC and the corresponding ZBFB discharge voltage diagram;

[0046] Figure 4 It is the measured SOC and the corresponding ZBFB discharge current diagram;

[0047] Figure 5 The first regression tree diagram of GBDT for evaluating SOC in the [3520,3690] time domain (where the optimal branch is 0.7A≤i(2.6A)<18.10A, indicated by the red line);

[0048] Figure 6 The second regression tree diagram of GBDT for evaluating SOC in the [3520,3690] time domain (where the optimal branch is 0.7A≤i(2.6A)<18.10A, indicated by the red line);

[0049] Figure 7 The third regression tree diagram of GBDT for evaluating SOC in the [3520,3690] time domain (where the optimal branch is 0.7A≤i(2.6A)<18.10A, indicated by the red line);

[0050] Figure 8 The fourth regression tree diagram of GBDT for evaluating SOC in the [3520,3690] time domain (where the optimal branch is 0.7A≤i(2.6A)<18.10A, indicated by the red line);

[0051] Fig. 9 The fifth regression tree diagram of GBDT for evaluating SOC in the [3520,3690] time domain (where the optimal branch is 0.7A≤i(2.6A)<18.10A, indicated by the red line);

[0052] Fig.10 is the ZBFB discharge voltage phase diagram in the time domain [625,2176] (relative to Figure 2 The correct measured SOC is shown, and there is a stable limit cycle);

[0053] Fig.11 It is the ZBFB discharge voltage phase diagram in the time domain [2965,3960] (corresponding to the distorted SOC period, there is no limit cycle);

[0054] Fig.12 is the ZBFB discharge voltage phase diagram in the time domain [3985,4298] (relative to Figure 2 The correct measured SOC is shown, and there is a stable limit cycle);

[0055] Fig.13 is the ZBFB discharge current phase diagram in the time domain [625,2176] (relative to Figure 2 The correct measured SOC is shown, and there is a stable limit cycle);

[0056] Fig.14is the ZBFB discharge current phase diagram in the time domain [2965,3960] (corresponding to the distorted SOC period, there is no limit cycle);

[0057] Fig.15 is the ZBFB discharge current phase diagram in the time domain [3985,4298] (relative to Figure 2 The correct measured SOC is shown, and there is a stable limit cycle). DETAILED DESCRIPTION

[0058] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0059] The present invention is based on a 10kWh zinc-bromine flow battery experimental platform developed in the laboratory. Aiming at the deficiency that the existing methods are still unable to fully realize the accurate measurement of SOC (State of Charge) of zinc-bromine flow batteries, GBDT (Gradient Boosting Decision Tree) is applied to realize high-precision evaluation and verification of SOC at the inflection point of battery discharge voltage, which can provide a reference for the development of a general high-precision calculation algorithm for SOC of flow batteries. The core work of the present invention is to: ① Construct a GBDT modeling, solution and evaluation system for efficient battery SOC evaluation based on finite samples of voltage and current, highlighting the high-precision identification of SOC at the voltage inflection point; ② In order to effectively reduce the computational cost of GBDT and improve its computational efficiency, an SOC verification system based on discharge voltage and current limit cycle analysis is established, and the response time of GBDT is improved by reducing its calculation frequency within a charge and discharge cycle. The following is combined with Figure 1 The specific implementation mode of the present invention is introduced, which specifically comprises the following steps:

[0060] Step 1: Establish a comprehensive experimental platform for zinc-bromine flow batteries (such as Figure 2 , detailed parameter description is shown in Table 1), including zinc-bromine flow battery body, BMS, EMS, PCS and load AC motor. Among them, BMS is used to monitor the charge and discharge data of the battery body, EMS is used to power the positive and negative circulation pumps of the zinc-bromine flow battery, PCS is used to realize DC / AC power conversion, and the load AC motor is used to simulate the inductive load. The model foundation is established for extracting limited data samples and identifying the SOC of the battery;

[0061] Table 1

[0062]

[0063] Step 2: Clearly identify the measured parameters and identification parameters. The measured parameters include voltage, current, circulating pump speed, charge and discharge time, and measured SOC. The identification parameter is the SOC at the voltage inflection point. Use a battery charge and discharge comprehensive tester and BMS to monitor the experimental platform in real time for 24 hours, with a sampling period of 1 minute, and obtain the curves of measured SOC, battery voltage, and battery current data changing with time (such as Figure 3 to Figure 4 );

[0064] Step 3: Select limited voltage and current sampling data in one cycle of the ZBFB discharge process at regular intervals within a whole discharge cycle, extract limited feature samples and avoid sensitive outliers;

[0065] The present invention focuses on solving the problem of poor SOC recognition accuracy of the voltage inflection point of a flow battery. Figure 3 and Figure 4 It shows that when the discharge voltage and current of ZBFB are both 0 (3250min-3690min), the SOC measured by the tester is not 0, so the SOC at this point needs to be corrected. The characteristic samples extracted from the data obtained by the tester are shown in Table 2 below. The first six groups of data are used to subsequently use GBDT to improve the SOC identification accuracy at the voltage inflection point. The seventh group of data is in the interval after the voltage inflection point (3250min-3690min), and the measuring instrument presents an inaccurate SOC value.

[0066] Table 2

[0067] Sampling Data Voltage (V) Current (A) SOC(%) Time (min) <![CDATA[(u1,i1)]]> 51.30 18.10 90.29 1965 <![CDATA[(u2,i2)]]> 51.40 22.30 49.04 3985 <![CDATA[(u3,i3)]]> 53.10 0.70 99.87 2282 <![CDATA[(u4,i4)]]> 53.80 0.40 99.95 625 <![CDATA[(u5,i5)]]> 54.10 -29.60 0.91 4298 <![CDATA[(u6,i6)]]> 54.70 -29.70 81.41 2176 <![CDATA[(u7,i7)]]> 7.60 2.60 13.41 3520

[0068] In order to achieve the goal of improving the SOC identification accuracy at the voltage inflection point, six representative characteristic data are extracted from the displayed charge and discharge cycles: (625, 99.95), (1965, 90.29), (2176, 81.41), (2282, 99.87), (3985, 49.04) and (4298, 0.91). Based on this, the corresponding voltage and current sample data sets are as follows:

[0069] Formula 1:

[0070] Step 4: Take the extracted limited feature samples as the data set, give the loss function and learning rate of GBDT in real time, and develop a SOC solution paradigm that can achieve real-time verification;

[0071] The GBDT algorithm was proposed by Jerome H. Friedman in 1999 and is a regression prediction method in the field of machine learning. The core idea is: ① Perform multiple rounds of iterative training through weak classifier residual basis analysis; ② Fit the negative gradient of the loss function under the current model in each round of iteration; ③ Control the loss function to descend along the gradient direction and converge to the local optimal solution or the global optimal solution as quickly as possible. The implementation process of the GBDT model constructed in step 4 is:

[0072] Step 4.1 Input feature sample data D = {(x1, y1), (x2, y2), … (x k ,y k ),…,(x N ,y N )}, input loss function L(y,f(x)). Where N is the number of samples, (x k ,y k ) represents the kth feature sample data, 1≤k≤N, χ represents the input space. 1≤k≤N, Y represents the output space. y and f(x) represent the output data and the learner function respectively.

[0073] Step 4.2 Initialize the weak learner:

[0074]

[0075] And initialize the number of iterations m = 1. Where f0(x) is the initialization weak learner function, x is the input feature, and c is the constant taken when the loss function is minimized.

[0076] Step 4.3 For any sample data (x k ,y k ), the residual is calculated as:

[0077]

[0078] Among them, r m,k is the pseudo residual of the k-th feature sample data in the m-th iteration, that is, is the negative gradient of the loss function.

[0079] The residual value calculated in step 4.4 is used as the true value for the next training, and the fitting data {(x k ,r m,k )} k＝1,2,...,N As the sample data for the next training, a new regression tree T is obtained m , and its corresponding leaf node area is R m,k ,k=1,2,...,J m , where J mis the number of leaf nodes of the mth regression tree.

[0080] Step 4.5 Apply regression tree T m Calculate the best fit value for each leaf node of:

[0081]

[0082] Among them, c m,k is the best fitting value of the leaf area, f m-1 (x k ) represents the sample data x given by the learner in the m-1th iteration k The predicted value of .

[0083] Step 4.6 Refresh the strong learner:

[0084]

[0085] Among them, f m (x) is the learner of the mth iteration, I is an indicator symbol whose value is 1, which is used to fit the classifier in the node area.

[0086] Step 4.7 After completing M rounds of iterations, the final boosted regression tree is obtained:

[0087]

[0088] Among them, M is the number of iterations, that is, the number of trees.

[0089] Step 4.8 Obtain and analyze the optimal decision tree

[0090] Step 4.9 Verification by comparing with the basic sample

[0091] Step 4.10 If If the predicted value is reasonable, output Otherwise, m=m+1, and repeat the above calculation process starting from step 4.3.

[0092] Given the complexity of the SOC calculation of the ZBFB, when applying GBDT to evaluate the SOC at the inflection point of the discharge voltage, there are still the following related issues to be resolved: (1) It is necessary to select a limited number of voltage and current sampling data within a cycle of the ZBFB discharge process to identify the discharge change characteristics. (2) Given that the coulombic efficiency of the ZBFB varies with the discharge current, the loss function, learning rate and other parameters of the GBDT need to be given in real time. (3) It is necessary to overcome the inherent data sensitivity of the GBDT caused by the sudden voltage drop at the inflection point of the ZBFB discharge voltage. (4) In order to avoid overcompensation, improve the accuracy of SOC prediction and the real-time performance of calculation, and give the optimal SOC value of the ZBFB, it is necessary to develop a SOC solution paradigm that can achieve real-time verification.

[0093] In order to evaluate the SOC of ZBFB in the [3520,3690] time domain in real time, based on Table 2 and the GBDT algorithm solution process, the first weak learner should be initialized to minimize the loss function. The first weak learner can be written as follows:

[0094] Formula 2: f0(x) = (90.29 + 49.04 + 99.87 + 99.95 + 0.91 + 81.41) / 6 = 70.245

[0095] The residual of the measured SOC relative to the first weak learner is shown in Table 3. It should be emphasized that it is necessary to find the optimal partition node of the sample data based on the traversal value of each feature data. This program starts at voltage u1 = 51.30V and ends at current i6 = -29.70A. The core work is to calculate the variance of the two sampled voltage or current partition data for ZBFB.

[0096] Table 3

[0097] Sampling Data Measured SOC <![CDATA[f0(x)]]> Residual <![CDATA[(u1,i1)]]> 90.29 70.245 20.045 <![CDATA[(u2,i2)]]> 49.04 70.245 -21.205 <![CDATA[(u3,i3)]]> 99.87 70.245 29.625 <![CDATA[(u4,i4)]]> 99.95 70.245 29.705 <![CDATA[(u5,i5)]]> 0.91 70.245 -69.335 <![CDATA[(u6,i6)]]> 81.41 70.245 11.165

[0098] Based on the above, u4=54.10V and i4=0.40A are selected as sample partition nodes, that is, (u4, i4) is used to evaluate the optimal partition node of SOC in the [3520, 3690] time domain. Furthermore, based on the minimum variance principle and the traversal value analysis of the left leaf node and the right leaf node, the tree diagram corresponding to the [3520, 3690] time domain can be obtained as follows: Figures 5 to 9 This work is used to evaluate the optimal SOC calculation path. The essence is to improve the SOC calculation accuracy at the ZBFB discharge voltage inflection point based on sample feature data training.

[0099] Compare Figures 5 to 9As shown in Table 3, since the discharge current at the SOC to be evaluated is 2.6A, this value is on the branch 0.7A≤i(2.6A)<18.10A, so 0.7A≤i(2.6A)<18.10A is selected as the best path for SOC evaluation, that is, node 3 is the last strong learner calculation path to be considered.

[0100] In addition, considering that the coulombic efficiency of the battery varies with the discharge current, the learning rate of GBDT is finally selected as α = 0.458. Therefore, the SOC value of ZBFB in the time domain [3520,3690] can be calculated as shown in the following formula:

[0101] Formula 3:

[0102] Comparing Formula 3 with Table 2 shows that, compared with the SOC value of 13.41% obtained from the actual test, the SOC evaluated by GBDT can reach 4.10%, which is very close to the actual SOC value of 0, because the actual values ​​of the discharge voltage and discharge current of ZBFB are both 0. Based on the above analysis, the proposed GBDT can be applied to the evaluation and verification of distorted SOC, and thus can provide a reference basis for the correction of SOC calculation errors caused by sampling distortion at the inflection point of ZBFB discharge voltage.

[0103] Step 5: Apply the limit cycle theory to derive the judgment and verification criteria for using GBDT to calculate the high-precision SOC of ZBFB. By reducing the number of GBDT applications, its calculation efficiency in a complete charge and discharge cycle can be improved to improve the real-time performance of the proposed GBDT algorithm.

[0104] In a complete charge and discharge cycle, the discharge voltage limit loop ( Figures 10-12 )、Discharge current limit loop( Figures 13-15 ) analysis (the corresponding statistical results are shown in Table 4) to determine whether GBDT is needed for SOC verification in different discharge periods. By reducing the number of GBDT applications, its calculation efficiency in a complete charge and discharge cycle can be improved. The difficulty lies in giving the GBDT application criteria based on the limit cycle.

[0105] Table 4

[0106]

[0107] Compare Figures 10-12 , Figures 13-15 Table 4 shows that Fig.10 , Fig.12 and Fig.13 , Fig.15 Correspondingly, the measured SOC at this time is consistent with the ZBFB discharge voltage and discharge current in this time domain, and there is a stable limit cycle; Fig.11 , Fig.14 As shown, in the time domain [2965,3960] range, the measured SOC is 13.41%, while the corresponding discharge voltage and discharge current are 0, that is, the SOC does not match the voltage and current. At this time, the voltage and voltage distortion rate, and the current and current distortion rate cannot form a stable limit cycle.

[0108] Conversely, Figures 10-12 and Figures 13-15 As shown, the following conclusions can be drawn: (1) When the discharge voltage / current conversion rate of the ZBFB can form a stable limit cycle relative to the voltage / current, there is no SOC distortion and GBDT does not need to be used to verify the SOC. (2) When the discharge voltage / current conversion rate of the ZBFB cannot form a stable limit cycle relative to the voltage / current, or the limit cycle is unstable, there may be SOC distortion and GBDT needs to be used to verify the SOC.

[0109] In summary (1) and (2), since the present invention does not need to perform GBDT verification on the SOC at each voltage inflection point of the ZBFB, the calculation cost of the GBDT can be greatly reduced, and the real-time performance of the algorithm in the application of the ZBFB field can be improved.

[0110] In view of the fact that there is no mature SOC accurate measurement technology and evaluation algorithm for ZBFB at present, this paper applies the GBDT algorithm in machine learning to the real-time calculation and verification of SOC of ZBFB, and highlights the following two research results: (1) A GBDT verification algorithm matching the SOC measurement at the ZBFB discharge voltage inflection point is given, which provides an algorithmic basis for the high-precision evaluation of SOC at the ZBFB inflection point; (2) A GBDT algorithm verification and timeliness judgment criterion based on limit cycle analysis is developed, which provides a computational support for improving the timeliness of GBDT from the algorithm level. The analysis of the measured results shows that the work done can effectively improve the SOC calculation accuracy at the ZBFB voltage inflection point, and can provide a reference for the development of a general liquid flow battery SOC algorithm.

Claims

1. A flow battery voltage inflection point SOC verification method based on GBDT machine learning prediction, characterized by: The method comprises the following steps: Step 1: Establish a comprehensive experimental platform for zinc-bromine flow batteries, including the zinc-bromine flow battery body, BMS, EMS, PCS and load AC motor; among them, BMS is used to monitor the charge and discharge data of the battery body, EMS is used to power the positive and negative electrode circulation pumps of the zinc-bromine flow battery, PCS is used to realize DC / AC power conversion, and the load AC motor is used to simulate the inductive load; Step 2: Clarify the measured parameters and identification parameters. The measured parameters include voltage, current, circulation pump speed, charge and discharge time and measured SOC. The identification parameter is the SOC at the voltage inflection point. Use the battery charge and discharge comprehensive tester and BMS to monitor the experimental platform in real time for 24 hours, with a sampling period of 1 minute, to obtain the curves of measured SOC, battery voltage and battery current data changing with time. Step 3: Select limited voltage and current sampling data in multiple cycles of ZBFB discharge to identify discharge change characteristics; extract limited feature samples at regular intervals in a whole discharge cycle to avoid sensitive outliers; Step 4: Take the extracted limited feature samples as the data set, and give the loss function and learning rate of GBDT in real time. Use GBDT to solve the optimal SOC value of ZBFB and develop a SOC solution paradigm that can realize real-time verification. Step 5: Apply limit cycle theory to derive the judgment and verification criteria for using GBDT to calculate the high-precision SOC of ZBFB. By reducing the number of GBDT applications, its calculation efficiency in a complete charge and discharge cycle can be improved to improve the real-time performance of the proposed GBDT algorithm.

2. The method for verifying the voltage inflection point SOC of a flow battery based on GBDT machine learning prediction according to claim 1, characterized in that: In step 2, the sampling period is 1 minute, and the measured data sampled include the voltage, current, circulation pump speed, charge and discharge time, and measured SOC of the zinc-bromine flow battery experimental platform.

3. The method for verifying the voltage inflection point SOC of a flow battery based on GBDT machine learning prediction according to claim 1, characterized in that: In step 3, {(t1, SOC1), (t2, SOC2), … (t k ,SOC k ),…,(t N ,SOC N )} A total of N representative small sample feature data.

4. The method for verifying the voltage inflection point SOC of a flow battery based on GBDT machine learning prediction according to claim 1, characterized in that: In step 4, the SOC of ZBFB in the time domain after the voltage inflection point is evaluated in real time, and the implementation process of the constructed GBDT model is as follows: (1) Input feature sample data D = {(x1, y1), (x2, y2), … (x k ,y k ),…,(x N ,y N )}, input loss function L(y,f(x)); where N is the number of samples, (x k ,y k ) represents the kth feature sample data, 1≤k≤N, χ represents the input space; 1≤k≤N, Y represents the output space; y and f(x) represent the output data and the learner function respectively; (2) Initialize weak learner: And initialize the number of iterations m = 1; where f0(x) is the initialization weak learner function, x is the input feature, and c is the constant taken when the loss function is minimized; (3) For any sample data (x k ,y k ), the residual is calculated as: Among them, r m,k is the pseudo residual of the k-th feature sample data in the m-th iteration, that is, is the negative gradient of the loss function; (4) The calculated residual value is used as the true value of the next training, and the fitting data {(x k ,r m,k )} k＝1,2,...,N As the sample data for the next training, a new regression tree T is obtained m , and its corresponding leaf node area is R m,k ,k=1,2,…,J m , where J m is the number of leaf nodes of the mth regression tree; (5) Apply regression tree T m Calculate the best fit value for each leaf node of: Among them, c m,k is the best fitting value of the leaf area, f m-1 (x k ) represents the sample data x given by the learner in the m-1th iteration k The predicted value of (6) Refresh the strong learner: Among them, f m (x) is the learner of the mth iteration, I is an indicator whose value is 1, which is used to fit the classifier in the node area; (7) After completing M rounds of iterations, the final boosted regression tree is obtained: Among them, M is the number of iterations, that is, the number of trees; (8) Obtain and analyze the optimal decision tree (9) Verification by comparison with basic samples (10) If If the predicted value is reasonable, output Otherwise, m=m+1, and the above calculation process is repeated starting from (3).

5. The method for verifying the voltage inflection point SOC of a flow battery based on GBDT machine learning prediction according to claim 1, characterized in that: In step 4, the optimal decision tree of ZBFB at the voltage inflection point is as follows:

6. The method for verifying the voltage inflection point SOC of a flow battery based on GBDT machine learning prediction according to claim 1, characterized in that: In step five, within a complete charge and discharge cycle, the necessity of GBDT for SOC verification in different discharge periods is determined by analyzing the discharge voltage limit cycle and the discharge current limit cycle, and the calculation efficiency of GBDT within a complete charge and discharge cycle is improved by reducing the number of applications of GBDT.

Images