KR102689216B1 - A system for operating and managing battery packs by KNN machine learning algorithm - Google Patents

Abstract

The embodiment relates to a battery pack operation management system using the KNN machine learning algorithm.
Specifically, this configuration first includes a battery pack in which multiple battery modules containing multiple battery cells connected in parallel are connected in series, and a sensor is provided on one side of the battery pack to measure voltage, current, and temperature in units of battery modules and battery cells. A microcomputer that receives measurement data measured by the unit and the sensor unit and performs KNN machine learning to control charging and discharging of the battery pack and voltage and current of the battery pack, battery module, and battery cell, and the battery It includes a communication unit for the pack, sensor unit, and microcomputer to communicate with each other, and a cell BMS for managing the battery cell is mounted on one side of the battery cell, and a module BMS for managing the battery module is installed on one side of the battery module. It is characterized in that it is installed.
And, additionally, for batteries, operating cycle conditions play an important role in identifying parameters in determining performance. And, in this case, the number of charge and discharge is considered a common parameter, and the C-rate of 0.5C at an ambient temperature of 25 degrees becomes the most important variable. Therefore, we analyze the SOC status considering the aging change of the battery, and in this case, we use the kNN regression algorithm to analyze the degradation parameters of the charging voltage (V) and charging current (I) with respect to the charging time (T). The battery's state of charge value is estimated by learning using the Euclidean distance calculation method. In addition, charge/discharge voltage and current in response to temperature changes are major factors that determine battery SOC characteristics, and the state of charge value of these batteries is estimated differently depending on temperature changes.
Therefore, through this, the battery pack can be efficiently managed by predicting changes over time in the battery.

Description

Battery pack operation management system using KNN machine learning algorithm {A system for operating and managing battery packs by KNN machine learning algorithm}

The content disclosed in this specification relates to an AI (Artificial Intelligence) battery pack operation management system, and more specifically, monitors the charging and discharging status of the battery pack using the KNN (K Nearest Neighbors) machine learning algorithm, This is about a battery pack operation management system using the KNN machine learning algorithm that can control to maintain the optimal state and predict aging of the battery pack.

Recently, the use of batteries has been increasing in various fields such as ESS (Energy Storage System), electric vehicles, and electric motorcycles. However, as the charging and discharging of the battery is repeated more than a certain number of times, the performance of the battery deteriorates and the lifespan of the battery is drastically shortened. Specifically, when the number of charging and discharging times of the battery increases to 100, 200, and 300, the total chargeable capacity of the battery may decrease, causing rapid discharging of the battery.

This process of deteriorating battery performance due to continuous charging/discharging of the battery is called secular change. If this aging change in the battery accelerates, a difference may occur in the voltage balance of each battery cell that makes up the battery pack, and not only does the battery's chargeable capacity (State of Health, SoH) decrease, but also overcharging and overcharging of each battery cell may occur. There is a risk that overdischarge may occur. At this time, in the battery, SoH refers to the chargeable capacity of the battery, and SoC (State of Charge) refers to the amount of charging energy currently stored in the battery and available for use.

In addition, electric motorcycles equipped with batteries whose performance has deteriorated due to changes over time have lower performance compared to existing engine motorcycles, so their ability to replace existing motorcycles is limited.

Therefore, there is an urgent need to introduce technology to control and manage the battery pack by predicting changes over time in the battery so that the life of the battery pack can be extended beyond a certain period of time to battery packs for electric motorcycles.

KR 1020210042402 A KR 100814811 B1

The disclosed content uses a KNN machine learning algorithm that can check the charging and discharging status of the battery pack, predict aging changes in the battery pack, and control the battery pack to maintain an optimal state using a machine learning algorithm. The goal is to provide a battery pack operation management system.

The battery pack operation management system using the KNN machine learning algorithm according to the embodiment is,

First, a battery pack with a plurality of battery modules containing a plurality of battery cells connected in parallel connected in series, a sensor unit provided on one side of the battery pack and measuring voltage, current, and temperature in units of battery modules and battery cells, and the sensor unit. Receives data measured by and performs KNN machine learning to control the charging and discharging of the battery pack and the voltage and current of the battery pack, battery module, and battery cell, the sensor unit, and the microcomputer communicate with each other. A cell BMS for managing the battery cell is mounted on one side of the battery cell, and a module BMS for managing the battery module is mounted on one side of the battery module.

Additionally, operating cycle conditions play an important role in identifying parameters in determining performance in batteries. And, in this case, the number of charge and discharge is considered a common parameter, and the C-rate of 0.5C becomes the most important variable at an ambient temperature of 25ㅀC. So, here we analyze the SOC status considering the aging changes of the battery, and in this case, we use the kNN regression algorithm to determine the degradation parameters of the charging voltage (V) and charging current (I) with respect to the charging time (T). Based on this, the charge state value of the battery is learned and estimated using the Euclidean distance calculation method. In addition, charge/discharge voltage and current in response to temperature changes are major factors that determine battery SOC characteristics, and the state of charge value of these batteries is estimated differently depending on temperature changes.

According to embodiments, the battery pack operation management system using the KNN machine learning algorithm can predict aging changes in the battery and efficiently manage the battery pack, so the battery pack can be applied to various fields such as electric vehicles and electric motorcycles. There is an effect.

Figure 1 is a configuration diagram of a battery pack operation management system using the KNN machine learning algorithm according to an embodiment.
FIG. 2 is a configuration diagram of the battery pack shown in FIG. 1.
Figure 3 is a diagram for explaining the charging and discharging control process of the battery pack.
Figure 4 is a diagram to explain the control method of the KNN machine learning algorithm.
Figure 5 is a diagram for explaining the cell BMS attached to each cell.
Figure 6 is a diagram for explaining the module BMS attached to each module.
7 to 9 are diagrams for explaining a learning model according to an embodiment.
Figure 10 is a diagram showing the experimental results of the learning model.

FIG. 1 is a configuration diagram of a battery pack operation management system 10 using the KNN machine learning algorithm 48 according to an embodiment, and FIG. 2 is a configuration diagram of the battery pack 20 shown in FIG. 1.

Referring to FIG. 1, the battery pack operation management system 10 using the KNN machine learning algorithm according to one embodiment includes a battery pack 20, a sensor unit 30, a communication unit 35, and a microcomputer 40. It is composed.

Referring to FIG. 2, the battery pack 20 includes a plurality of battery modules 21 connected in series, and the battery module 21 includes a plurality of battery cells 23 connected in parallel.

Specifically, the battery pack 20 includes 20 battery modules 21 connected in series, and the battery module 21 includes 4 battery cells 23 connected in parallel. In other words, the battery pack 20 includes a total of 80 battery cells 23, 20 series x 4 parallel (20 series x 4 parallel).

And, for each battery cell (23), #(1, 1), #(1, 2), #(1, 3), #(1, 4),… A cell number is given, such as ,#(20, 4). Additionally, each battery module (21) is #1, #2, #3,… , a module number is assigned, such as #20.

As the number of battery cells connected in series and parallel increases, the capacity of the battery pack 20 including the battery cells 23 further increases.

Additionally, a cell BMS 24 (Battery Management System) is installed on one side of the battery cell 23 to manage the battery cell 23. Additionally, a module BMS 22 (Battery Management System) for managing the battery module 21 is installed on one side of the battery module 21. Additionally, the sensor unit 30 is provided on one side of the battery pack 20 and measures voltage, current, and temperature for each battery module 21 and battery cell 23.

Additionally, the communication unit 35 serves to enable the battery pack 20, sensor unit 30, and microcomputer 40 to communicate with each other. At this time, the communication unit 35 is preferably implemented in the form of a photo coupler in which the light emitting unit and the light receiving unit are insulated from each other and can couple electrical signals into light.

Then, the communication unit 35 transmits the measurement data of the battery module 21 or the battery cell 23 measured by the sensor unit 30 to the microcomputer 40 and sends a command signal generated by the microcomputer 40. It is transmitted to the battery pack (20).

And, based on the measurement data measured by the sensor unit 30, the microcomputer 40 performs the charging and discharging of the battery pack 20, the battery pack 20, the battery module 21, and the battery cell 23. Controls voltage and current. At this time, it is desirable that the microcomputer 40 programming be implemented using artificial intelligence (AI) techniques.

Then, the microcomputer 40 identifies the battery cells 23 whose actual voltage does not fall within the reference range between the preset lower and upper limit voltages and determines them to be defective. At this time, the upper limit voltage refers to the voltage that serves as the standard for full charging of the battery cell 23, and the lower limit voltage refers to the voltage that serves as the standard for full discharging of the battery cell 23.

Specifically, when the actual voltage of the battery cell 23 exceeds the upper limit voltage, the microcomputer 40 determines the battery cell 23 to be in an overvoltage state and discharges the battery cell 23, thereby discharging the battery cell 23. ) is controlled so that the actual voltage converges to the upper limit voltage within the reference section.

In addition, when the actual voltage of the battery cell 23 is less than the lower limit voltage, the microcomputer 40 determines the battery cell 23 to be in a low voltage state and charges the battery cell 23, thereby Control the actual voltage to converge to the lower limit voltage within the reference section.

For example, if the upper and lower limit voltages are 4.0V and 2.8V, respectively, and the actual voltages of battery cell 23 #1 and battery cell 23 #2 are 4.2V and 2.6V, respectively, the microcomputer 40 can determine battery cell 23 #1 to be in an overvoltage state, discharge battery cell 23 #1, and control the actual voltage of battery cell 23 #1 to converge to 4.0V within the reference section. there is.

Then, the microcomputer 40 determines that the battery cell 23 #2 is in a low voltage state and charges the battery cell 23 #2, so that the actual voltage of the battery cell 23 #2 is 2.8V within the reference section. It can be controlled to converge to .

Figure 3 is a diagram for explaining the charging and discharging control process of the battery pack 20.

Referring to FIG. 3, the battery pack operation management system 10 using the KNN machine learning algorithm according to one embodiment includes a charger connector 51, a voltage sensor 52, a first switching element 53, and a current sensor 54. ) and a second switching element 55.

First, the charger connector 51 is connected to the charger 50 and supplies power of the reference voltage and reference current previously set in the charger 50 to the battery pack 20.

Here, the reference voltage is a voltage that serves as a charging standard for the battery pack 20, and the reference current is a current that serves as a charging standard for the battery pack 20.

For example, the charger connector 51 can supply power with a reference voltage of 84V and a reference current of 32A to the battery pack 20 through the charger 50.

Additionally, the voltage sensor 52 is provided on one side of the battery pack 20 and measures the actual voltage of the battery pack 20. The actual voltage data of the battery pack 20 measured by the voltage sensor 52 is transmitted to the microcomputer 40 through the communication unit 35.

In addition, the first switching element 53 is provided between one end of the charger connector 51 and the voltage sensor 52, and the on-time time is determined by the actual voltage of the battery pack 20 measured by the voltage sensor 52. It is configured to be variable.

Additionally, the current sensor 54 is provided on the other side of the battery pack 20 and measures the actual current of the battery pack 20 being charged. The actual current data of the battery pack 20 measured by the current sensor 54 is transmitted to the microcomputer 40 through the communication unit 35.

In addition, the second switching element 55 is provided between the other end of the charger connector 51 and the current sensor 54, and the on-time time is determined by the actual current of the battery pack 20 measured by the current sensor 54. It is configured to be variable.

Power supplied from the charger connector 51 passes through the first switching element 53 and the voltage sensor 52 and is supplied to the battery pack 20. Then, the power passes through the battery pack 20, current sensor 54, and second switching element 55, and is supplied to the charger connector 51.

Then, the microcomputer 40 calculates the actual voltage difference by subtracting the actual voltage measured by the voltage sensor 52 from the preset reference voltage, and then switches the first switching element 53 in proportion to the actual voltage difference. Control the on time to be variable.

For example, if the reference voltage and reference current are 84V and 32A, respectively, and the actual voltage and actual current of the battery pack 20 are 60V and 3A, respectively, the microcomputer 40 generates an actual voltage difference of 24V (=84V-60V). In proportion to , the first switching element 53 can be controlled to remain on for 3 hours.

Meanwhile, when the charger connector 51 supplies power with a reference current of 32A to the battery pack 20 for 3 hours, and the actual voltage of the battery pack 20 increases from 60V to 76V, the microcomputer 40 increases the actual voltage. In proportion to the difference of 8V (=84V-76V), the first switching element 53 can be controlled to remain on for 1 hour.

Through this, the charger connector 51 can supply power with a reference current of 32A to the battery pack 20 for 1 hour.

Meanwhile, if the charger connector 51 initially supplies excessive current to the battery pack 20, multiple batteries included inside the battery pack 20 may be damaged by overcharging of the battery pack 20.

Accordingly, the microcomputer 40 calculates the actual current difference by subtracting the actual current measured by the current sensor 54 from the reference current, and then controls the on time of the second switching element 55 to vary in proportion to the actual current difference. can do. Through this process, the microcomputer 40 can control the amount of current supplied from the charger connector 51 to the battery pack 20.

Meanwhile, the user can check the data processed by the microcomputer 40 through the dashboard 60 or mobile app 61. Based on the data, the user can repair or replace the battery pack 20 whose performance has deteriorated, or use the data as driving information.

Next, the control method of the KNN machine learning algorithm 48 will be described in more detail.

Figure 4 is a diagram for explaining the control method of the KNN machine learning algorithm 48.

Referring to FIG. 4, the microcomputer 40 includes a first amplifier 41, a first comparator 42, a PI voltage controller 43, a current converter 44, a second amplifier 45, and a second comparator 46. ), a learning model (47), and a KNN machine learning algorithm (48).

First, the first amplifier 41 increases the actual voltage measured by the voltage sensor 52 by a certain multiple.

And, the first comparator 42 outputs a voltage difference obtained by subtracting the actual voltage of the battery pack 20 output from the first amplifier 41 from the reference voltage. At this time, the reference voltage and reference current are included in a command signal called a command and are input to the first comparator 42.

Additionally, the PI voltage controller 43 receives the voltage difference calculated by the first comparator 42 and outputs a voltage compensation value obtained by compensating the voltage difference to the actual voltage of the battery pack 20.

Then, the current converter 44 outputs a current predicted value obtained by converting the voltage compensation value. At this time, the current predicted value means the current scheduled to be charged with the battery pack 20.

And, the second amplifier 45 increases the actual current measured by the current sensor 54 by a certain multiple.

Then, the second comparator 46 outputs an actual current difference obtained by subtracting the actual current of the battery pack 20 output from the second amplifier 45 from the current predicted value.

And, in the learning model 47, a reference current difference that serves as a comparison standard for the actual current difference is set in advance.

Then, the KNN machine learning algorithm 48 compares the reference current difference of the learning model 47 and the actual current difference of the second comparator 46 and outputs state data of the battery pack 20. At this time, the state data of the battery pack 20 is set to one of overcharge, uncharge, and overdischarge states, and includes voltage and current values that must be supplied to the battery pack 20.

For reference, the learning model 47 according to this embodiment will be described later with reference to FIGS. 7 and 8 below.

In addition, the KNN machine learning algorithm 48 outputs the on-time of the first switching element 53 and the second switching element 55, respectively, based on the status data of the battery pack 20, and communicates the communication unit 35. It is transmitted to the first switching element 53 and the second switching element 55 through .

FIG. 5 is a diagram for explaining the cell BMS 24 included in the battery pack 20.

Referring to FIG. 5, the battery pack operation management system 10 using the KNN machine learning algorithm according to one embodiment further includes an overcharge control sensor 110 and a cell power switching element 111.

The overcharge control sensor 110 included in the cell BMS 24 receives the voltage value of each cell as an input and controls the on time of the cell power switching element 111. If the cell power switching element 111 is turned on, the cell voltage is discharged through this to prevent overcharging.

For example, when the overcharge reference voltage of the overcharge control sensor 110 is 4.2V and the voltage of the battery cell 23 is 4.3V, the input value of the overcharge control sensor 110 is greater than the overcharge reference voltage, so the cell power By turning the switching element 111 on, the voltage of the battery cell 23, which has a voltage higher than the overcharge reference voltage, is discharged to converge to the overcharge reference voltage.

FIG. 6 is a diagram for explaining the module BMS 22 included in the battery pack 20.

Referring to FIG. 6, the battery pack operation management system 10 using the KNN machine learning algorithm according to one embodiment further includes an overcharge control sensor 112 and a cell power switching device 113.

The overcharge control sensor 110 included in the module BMS 24 receives the voltage value of each cell as an input, controls the on time of the cell power switching element 111, and determines whether the battery cell 23 is overcharged by the microcomputer 40. ) to send. If the cell power switching element 111 is turned on, the cell voltage is discharged through this to prevent overcharging.

For example, when the overcharge reference voltage of the overcharge control sensor 110 is 4.2V and the voltage of the battery cell 23 is 4.3V, the input value of the overcharge control sensor 110 is greater than the overcharge reference voltage, so the cell power By turning the switching element 111 on, the voltage of the battery cell 23, which has a voltage higher than the overcharge reference voltage, is discharged to converge to the overcharge reference voltage.

And, from the moment the battery cell 23, which was below 4.2V during charging, becomes higher than 4.2V, the overcharge signal transmitted from the overcharge control sensor 110 to the microcomputer 40 changes from normal to overcharge. As a result, the microcomputer 40 can monitor which battery cell 23 is overcharged.

The battery pack operation management system using the KNN machine learning algorithm according to one embodiment has the effect of checking the charging and discharging status of the battery pack using the machine learning algorithm and controlling the battery pack to maintain the optimal state. There is.

The battery pack operation management system using the KNN machine learning algorithm according to one embodiment predicts spasm changes in the battery and estimates the SOC state by considering the spasm changes in the battery, so that the battery pack can be efficiently managed. The pack can be applied to various fields such as electric cars and electric motorcycles.

7 to 9 are diagrams for explaining the learning model 47 according to an embodiment.

Specifically, FIGS. 7 and 8 are diagrams for explaining the learning model 47 according to an embodiment, and FIG. 9 is a diagram for explaining the kNN algorithm used in this case.

As shown in FIGS. 7 to 9, the learning model 47 according to one embodiment first determines that the usage rate of the battery gradually decreases due to many factors and interconnected performance degradation mechanisms, so the operating cycle condition is used to identify parameters. plays an important role.

And, in this case, the number of charge and discharge is considered a common parameter, and the C-rate of 0.5C becomes the most important variable at an ambient temperature of 25ㅀC.

So, in one embodiment, the SOC state is analyzed considering the aging change of the battery, and in this case, the kNN regression algorithm is used to mediate the deterioration of the charging voltage (V) and charging current (I) with respect to the charging time (T). Based on the variables, the battery's state of charge value is estimated by learning using the Euclidean distance calculation method.

In addition, charge/discharge voltage and current in response to temperature changes are major factors that determine battery SOC characteristics, and the state of charge value of these batteries is estimated differently depending on temperature changes.

1. Specifically, first, the learning model 47 must define the SOC state for this purpose, and considering the previous battery aging change, it analyzes and determines the SOC state appropriate for each characteristic aging change.

In particular, in this case, the SOC status is determined differently for a number of different devices, that is, depending on the situation or condition of use, thereby providing accurate information in various situations.

So, for example, this SOC status can be compared with the SOC status of each device in use to determine whether the device is normal, so that the device can be maintained in an appropriate state and used smoothly for a relatively long period of time.

Additionally, these characteristics indicate specific aspects (e.g., voltage variability, etc.) that degrade performance in various charge/discharge operations.

And, for example, by extracting the corresponding reference current value for each SOC state for each characteristic secular change and setting it in the device to be used, comparing the reference current difference and actual current difference to determine whether the device is normal (Figure 4).

For example, as described above, if the difference between the reference current difference and the actual current difference is greater than the set difference, measures such as appropriately changing the usage situation may be taken to continue using the device.

Additionally, in this case, in addition to the reference current, voltage and temperature may be used as information indicating the SOC state.

For this purpose, specifically, the charging state of the battery is defined first, based on the degradation parameters of charging voltage (V) and charging current (I) with respect to charging time (T), and in addition, it is variable depending on temperature. It is assumed that

In other words, the charging state value of the battery is estimated by learning using the Euclidean distance calculation method based on the degradation parameters of charging voltage (V) and charging current (I) with respect to charging time (T). In addition, charge/discharge voltage and current in response to temperature changes are major factors that determine battery SOC characteristics, and the state-of-charge value of these batteries is estimated differently depending on temperature changes.

So, in these cases, this state of charge is generally expressed in terms of temperature and charge/discharge rate, and looking a little further, it is expressed as battery capacity relative to nominal or available capacity relative to temperature and charge/discharge rate, etc.

In other words, the state of charge (SOC) of the battery is calculated by expressing this capacity as Equation 1 below.

[Equation 1]

(Where, Q: Total battery current capacity [Ah], Q _c : Battery current remaining amount over time [Ah], Q _e : Battery charging current amount [Ah].)

Additionally, this information may be set differently for a number of different devices, that is, for each surrounding situation and condition in which the device is used. For example, charging voltage, etc. may be set differently.

2. Then, the SOC calculated in this way is sent to the kNN classifier for SOC prediction.

That is, SOC information corresponding to each charging voltage, for example, for each different device, that is, SOC information showing characteristic changes over time, is classified and registered.

3. Additionally, depth of discharge (DOD) is also an important parameter for estimating battery life. For example, high SOC and low DOD quickly reduce battery life. Therefore, the depth of discharge, that is, the amount of discharge of the battery, is also calculated using Equation 2 below.

[Equation 2]

For reference, depth of discharge is the opposite of SOC. SOC, or remaining capacity, is expressed using a percentage value of 100% when the battery capacity is full and 0% when the battery capacity is exhausted, whereas depth of discharge is expressed as a percentage. There are two ways to express it: ampere-hours (Ah) and percentage figures.

For example, if the discharge depth of a battery with a capacity of 50Ah is expressed in ampere hours, a discharge depth of 0Ah means that the battery is not discharged (the remaining capacity is 100%), and a discharge depth of 50Ah means that the battery is completely discharged. This refers to the state (state where the remaining capacity is 0%).

Also, when expressing the discharge depth of a battery as a percentage, if the discharge depth is 0%, the remaining capacity is 100%, and if the discharge depth is 100%, the remaining capacity is 0%.

In general, even after the battery has been used to its indicated capacity, it can be used slightly more, so the depth of discharge can exceed 100%. Since such a situation cannot be expressed with remaining capacity, it is also expressed using discharge depth.

In addition, as above, the discharge depth information corresponding to each charging voltage, for example, for each different device, that is, the discharge depth information showing characteristic changes over time, is classified and registered.

4. Meanwhile, a large amount of information set is derived through experiments to confirm the accuracy of the information under the set charging voltage conditions of the battery cell using the calculated SOC and DOD conditions.

For example, after charging to 4.0V and resting for a set time of 1 hour, the battery is discharged to a set critical discharge voltage, for example, 2.7V, under set discharge current conditions.

Therefore, this voltage, current, and time information is collected and stored as SOC learning information, that is, as an information set.

5. So, machine learn the SOC learning information using the kNN algorithm according to the example below.

And also, in this case, this configuration is created through a neural network circuit for SOC estimation for secular changes.

Specifically, according to one embodiment, this configuration consists of charging voltage (V) and charging current (I) for charging time (T) in the input layer, 4 nodes in the hidden layer, and 2 output nodes in the output layer. Predict SOC results for battery aging.

In this case, the weights of the kNN classifier model are adjusted so that the calculated SOC is compared to the actual SOC to minimize the error (weights are explained in more detail below).

Here, the kNN algorithm is as follows according to one embodiment (see FIG. 9).

1) First, voltage, current, time, and temperature information are input as SOC learning information.

2) Next, obtain the k nearest neighbor points that were close to the target value from the SOC learning information.

3) So, based on k neighbors, the Euclidean distance of each neighbor is calculated using Equations 3 to 5 below.

[Equation 3]

(where x _t : predicted value for input, f(x _t) : predicted SOC value in training data set, w _t : weight, k : number of predicted weights)

[Equation 4]

(Where, d _p,i : Weighted distance between test point and adjacent point, t : Decrease rate parameter in w _t .)

[Equation 5]

(Here, N : Euclidean space number of 2-point distance difference, x _t, x _j : t-th and j-th SOC point data values, w _t : t-th weighting)

4) Then, after calculating the distance through this, the nearest neighbor is selected.

5) So, assign the kth layer based on adjacent points.

6) Next, calculate the number of data points in each category.

7) So, assign the data points to the nearest and smallest kth layer.

Through this, the SOC status and discharge depth information appropriate for each year's changes are obtained, and independent and dependent variables for diagnosing the device are set respectively.

For example, the independent variable is set to SOC status or, in addition, discharge depth information, and the dependent variable is set to charging voltage, etc.

Additionally, this information may be determined differently for each device.

So, through this, we can create learning and training information with this information and obtain a format for diagnosis.

On the one hand, this charging voltage becomes a reference voltage preset in the device as described above, and this reference voltage is used as a reference voltage for charging the battery pack.

For example, it may be used with a reference voltage of 84V, that is, the charger connector is set to a reference voltage of 84V through the charger and supplies this to the battery pack.

8) Additionally, in order to obtain the above-described weight, that is, a weight determined differently depending on the device for accuracy of SOC information, the error rate is calculated through the average absolute error.

9) Then, the weights are optimized to minimize the error calculated in this way.

Meanwhile, in this case, this format identifies and diagnoses the SOC status through the following operations, so that it can be used smoothly and continuously by changing the usage conditions of the device.

Specifically, this operation is as follows (see Figure 4).

First, as described above, measure the actual voltage on the device in use and increase it by a certain multiple.

Then, the voltage difference obtained by subtracting the actual voltage of this device from the reference voltage is output.

Additionally, the PI voltage controller receives this voltage difference and outputs a voltage compensation value by compensating this voltage difference to the actual voltage. For reference, this operation is performed in the same way for current.

In addition, in the learning model, a reference voltage difference that serves as a comparison standard for the actual voltage difference is set in advance.

In addition, the KNN machine learning algorithm obtains status data of the battery pack by comparing the reference voltage difference of the learning model and the actual voltage difference. At this time, the status data of the battery pack is identified as one of overcharge, uncharge, and overdischarge.

So, control is performed in an appropriate state, and the specific operation is as follows.

For example, the cell BMS described above receives the voltage value corresponding to the difference between each cell as input and controls the on time of the cell power switching device.

In other words, if the cell power switching element is turned on, the cell voltage is discharged through this, thereby preventing overcharging.

For example, when the overcharge reference voltage is 4.2V and the voltage of the battery cell is 4.3V, the input value is greater than the overcharge reference voltage, so the cell power switching element is turned on, so the battery cell with a voltage higher than the overcharge reference voltage. Let the voltage converge to the overcharge reference voltage.

For reference, this format additionally performs the following operations.

For example, when charging various devices, the required usage amount is created for each charging voltage.

In this case, depending on the charging voltage in the charging operation or discharging operation, it is divided into a charging section that rises to a certain amount at the start of the work process in the process of charging the device, a holding section that maintains this, and a discharge section that falls. .

Also, since the required usage varies depending on each section, the information set is divided by section and a format is created. Also, depending on various situations, data may be stored without performing actual work, so different formats are created for various situations.

Therefore, formats are classified differently depending on the charging situation and device, and are also created by dividing them into sections. Additionally, a new format can be created for each unit of each facility, or a format can be created as a bundle of several facilities by setting standards. This allows the appropriate method to be determined depending on the characteristics of the data.

Next, if a large number of data is not collected due to errors in devices and facilities in the information collected in real time, or if outliers such as various conditions occur, the corresponding data file must be removed. After creating a basic data set, add additional properties such as required cumulative usage.

Additionally, if some information is not collected due to information interruption, the information is removed.

Next, determine valid properties for each format, obtain normal values, and determine independent and dependent variables.

FIG. 10 is a diagram showing experimental results of the learning model 47 according to one embodiment.

As shown in FIG. 10, the experimental results of the learning model 47 according to one embodiment are first given as inputs of voltage, current, temperature, and time. And, the target output is SOC%. Initially, the information is split into training and testing. In this case, 70% information is learned with the kNN method. Additionally, during training, focus on the kth layer where accurate results can be obtained. Here, we show a comparison between the kNN results and the actual during charging, using (a) 10 cycles, (b) 50 cycles, (c) 100 cycles, and (d) 150 cycles.

Therefore, as shown in Table 1 below, the predicted values and experimental results show excellent performance with average absolute errors of 0.08%, 0.58%, 0.35%, and 0.45% in charge and discharge tests of 10, 50, 100, and 150 times, respectively. shows.

Number of charges mean absolute error 10 cycles 0.08% 50 cycles 0.58% 100 cycles 0.35% 150 cycles 0.45%

10: Battery pack operation management system
20: Battery pack
21: Battery module
22: Module BMS
23: Battery cell
24: Cell BMS
30: sensor unit
35: Department of Communications
40: Microcomputer
41: first amplifier
42: first comparator
43: PI voltage controller
44: current transducer
45: second amplifier
46: second comparator
47: Learning model
48: KNN machine learning algorithm
50: Charger
51: charger connector
52: voltage sensor
53: first switching element
54: current sensor
55: second switching element
60: Dashboard
61: Mobile app
70: gate driver

Claims (5)

a) a battery pack (20) in which a plurality of battery modules (21) including a plurality of battery cells (23) connected in parallel are connected in series;
A sensor unit 30 provided on one side of the battery pack 20 to measure voltage, current, and temperature for each battery module 21 and battery cell 23; and
Based on the measurement data measured by the sensor unit 30, the charging and discharging of the battery pack 20, the voltage and current of the battery pack 20, the battery module 21, and the battery cell 23 are measured. Controlling microcomputer (40);
It includes a communication unit 35 for the battery pack 20, the sensor unit 30, and the microcomputer 40 to communicate with each other,
On one side of the battery cell 23,
A cell BMS (24) is installed to manage the battery cell (23),
On one side of the battery module 21,
A module BMS (22) is installed to manage the battery module (21),

b) a charger connector 51 that is connected to a charger and supplies power of a preset reference voltage and reference current to the battery pack 20;
A voltage sensor 52 provided on one side of the battery pack 20 to measure the actual voltage of the battery pack 20;
A first switching device that is provided between one end of the charger connector 51 and the voltage sensor 52 and whose on-time is variable depending on the actual voltage of the battery pack 20 measured by the voltage sensor 52. Small (53);
A current sensor 54 provided on the other side of the battery pack 20 to measure the actual current of the battery pack 20; and
A second switching element provided between the other end of the charger connector 51 and the current sensor 54 and whose on-time is variable depending on the actual current of the battery pack 20 measured by the current sensor 54. (55); Including,


c) a first amplifier 41 that increases the actual voltage measured by the voltage sensor 52 by a certain multiple;
a first comparator 42 that outputs a voltage difference obtained by subtracting the actual voltage of the battery pack 20 output from the first amplifier 41 from the reference voltage;
a PI voltage controller 43 that receives the voltage difference calculated by the first comparator 42 and outputs a voltage compensation value that compensates the voltage difference to the actual voltage of the battery pack 20;
a current converter 44 that outputs a current predicted value converted from the voltage compensation value;
a second amplifier 45 that increases the actual current measured by the current sensor 54 by the predetermined multiple;
a second comparator 46 that outputs an actual current difference obtained by subtracting the actual current of the battery pack 20 output from the second amplifier 45 from the current predicted value;
A learning model 47 in which a reference current difference that serves as a comparison standard for the actual current difference is set in advance; and
A KNN machine learning algorithm 48 that compares the reference current difference of the learning model 47 and the actual current difference of the second comparator 46 and outputs state data of the battery pack 20; Includes,

The learning model 47 is,
1. First, calculate the state of charge (SOC) of the battery by expressing it in Equation 1 below,
[Equation 1]

(Where, Q: Total battery current capacity [Ah], Q _c : Battery current remaining amount over time [Ah], Q _e : Battery charging current amount [Ah].)
2. The calculated SOC is sent to the kNN classifier for SOC prediction,
3. Calculate the battery discharge amount using Equation 2 below,
[Equation 2]

4. Using the calculated SOC and DOD (Depth of Discharge) conditions, charge up to the set critical charge voltage under the set charging voltage conditions of the battery cell, rest for a set time, and then discharge up to the set critical discharge voltage under set discharge current conditions. Collect voltage, current, and time information as SOC learning information,
5. The neural network circuit uses the kNN (K Nearest Neighbors) algorithm to machine learn SOC learning information and estimate SOC for secular changes. The input layer includes charging voltage (V) and charging current (I) for charging time (T). ) and 4 nodes in the hidden layer and 2 output nodes in the output layer to predict the SOC result for the battery's aging changes,
The calculated SOC adjusts the weights of the kNN classifier model to minimize the error compared to the actual SOC, where:

The kNN algorithm is,
1) Voltage, current, time, and temperature information are input as SOC learning information,
2) Find the k nearest neighbor points that were close to the target value from the SOC learning information,
3) Based on k neighbors, calculate the Euclidean distance of each neighbor using Equations 3 to 5 below,
[Equation 3]

(where x _t : predicted value for input, f(x _t) : predicted SOC value in training data set, w _t : weight, k : number of predicted weights)
[Equation 4]

(Where, d _p,i : Weighted distance between test point and adjacent point, t : Decrease rate parameter in w _t .)
[Equation 5]

(Here, N : Euclidean space number of 2-point distance difference, x _t, x _j : t-th and j-th SOC point data values, w _t : t-th weighting)
4) After calculating the distance, select the nearest neighbor,
5) Assign the kth layer based on adjacent points,
6) Calculate the number of data points in each category,
7) Assign the data points to the nearest and smallest kth layer,
8) Calculate the error rate through the average absolute error,
9) Optimizing the weights so that the calculated error is minimized; A battery pack operation management system using the KNN machine learning algorithm characterized by . According to clause 1,
The current predicted value is
This is the current scheduled to be charged with the battery pack 20,
The status data of the battery pack 20 is
It is set to one of overcharge, uncharge, and overdischarge states, and includes voltage and current values to be supplied to the battery pack 20,
The KNN machine learning algorithm is
Based on the status data of the battery pack 20, the on-time of the first switching element 53 and the second switching element 55 are output, respectively, and the first switching element ( 53) and a second switching element 55, respectively. A battery pack operation management system using the KNN machine learning algorithm. According to clause 1,
The microcomputer 40 is
When the actual voltage of the battery cell 23 measured by the sensor unit 30 exceeds the preset upper limit voltage, the battery cell 23 is determined to be in an overvoltage state, and the battery cell 23 is discharged. Controlling the actual voltage of the battery cell 23 to converge to the upper limit voltage within the reference section,
When the actual voltage of the battery cell 23 measured by the sensor unit 30 is less than a preset lower limit voltage, the battery cell 23 is determined to be in a low voltage state and the battery cell 23 is charged, A battery pack operation management system using the KNN machine learning algorithm, characterized in that the actual voltage of the battery cell 23 is controlled to converge to the lower limit voltage within the reference section. According to clause 3,
The upper limit voltage is
This is the voltage that serves as the standard for charging battery cells,
The lower limit voltage is
This is the voltage that serves as the standard for discharging battery cells,
The standard section is
A battery pack operation management system using the KNN machine learning algorithm, characterized in that the voltage of the battery cell is above the lower limit voltage and below the upper limit voltage. According to clause 3,
The microcomputer 40 is
After calculating the actual voltage difference by subtracting the actual voltage measured by the voltage sensor 52 from the reference voltage, the on time of the first switching element 53 is varied in proportion to the actual voltage difference. Control as much as possible,
After calculating the actual current difference by subtracting the actual current measured by the voltage sensor 52 from the reference current, the on time of the second switching element 55 is controlled to vary in proportion to the actual current difference. ,
The reference voltage is
This is the voltage that serves as a charging standard for the battery pack 20,
The reference current is
A battery pack operation management system using the KNN machine learning algorithm, characterized in that the current is the charging standard for the battery pack (20).