Estimation device and estimation system
Patent Information
- Application Number
- PCT/JP2025/045538
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-02-26
- Filing Date
- 2025-12-25
- Publication Date
- 2026-09-03
Smart Images
Figure JP2025045538_03092026_PF_FP_ABST
Abstract
Description
Estimation device and estimation system
[0001] The present invention relates to an estimation device and an estimation system.
[0002] Batteries are used, for example, to supply power to drive sources such as vehicle motors. When controlling battery charging or vehicle operation, it is necessary to estimate the battery's State of Charge (SOC). One method for estimating SOC is to use the battery's SOC-OCV characteristics. In this method, the open-circuit voltage (OCV) is estimated from the measured values of the current flowing through the battery and the terminal voltage, and the SOC is estimated by referring to the SOC-OCV characteristics.
[0003] In recent years, LFP (lithium iron phosphate) batteries, which are low-cost and have a long lifespan, have attracted attention as vehicle batteries. LFP batteries have a relatively flat SOC-OCV characteristic. Therefore, in the case of LFP batteries, estimation methods using SOC-OCV characteristics may result in large estimation errors. Patent document 1 proposes a method for estimating SOC for LFP batteries based on the internal resistance of the battery.
[0004] Patent No. 6733485
[0005] However, the internal resistance of a battery depends not only on the State of Core (SOC) but also on factors such as the magnitude and direction of the current. Therefore, when estimating SOC based on internal resistance, it is necessary to perform numerous experiments and simulations to create maps and conversion logic. Furthermore, when estimating SOC, it is expected that the estimation accuracy will improve by using data from multiple points. However, if the estimation device performs calculations using maps and conversion logic for each of the multiple data points, the computational load will increase. It is desirable to improve the accuracy of SOC estimation while reducing the computational load on the estimation device.
[0006] One aspect of the present invention is an estimation device for estimating the charge level of a battery, comprising: a first estimation unit that estimates the overvoltage response of the battery to a predetermined input pattern of the current based on measured values of the current flowing through the battery and the terminal voltage; and a second estimation unit that estimates the charge level by inputting at least the measured values of the current and the time-series data of the overvoltage response estimated by the first estimation unit into an estimation model created by machine learning, wherein the estimation model is created by learning the correspondence between input data and output data, the charge level, and the input signal includes at least the current flowing through the battery and the time-series data of the overvoltage response.
[0007] According to the present invention, it is possible to improve the estimation accuracy of the SOC while reducing the computational load on the estimation device.
[0008] This is a block diagram showing the configuration of the estimation device according to the embodiment. This is a diagram showing an equivalent circuit representing a battery system. This is a graph showing an example of the overvoltage response to an impulse input of current. This is a diagram showing an example of machine learning of the estimation model EM. (a) is a graph showing time-series data of current used for SOC estimation, and (b) is a graph showing time-series data of terminal voltage. This is a graph showing the estimation result of SOC. This is a graph showing the true value of SOC. This is a flowchart showing the processing of the estimation device in this embodiment. This is a diagram showing the configuration of the estimation system including the estimation device. This is a block diagram showing the configuration of the estimation device according to Modification 1.
[0009] Hereinafter, an estimation device according to an embodiment of the present invention will be described with reference to the drawings. Figure 1 is a block diagram showing the configuration of the estimation device 1 according to an embodiment. As shown in Figure 1, the estimation device 1 estimates the state of charge (SOC) of a battery 5 installed in a vehicle such as an electric vehicle or a hybrid electric vehicle. The battery 5 is a rechargeable secondary battery. The type of battery 5 is not limited, but as an example, an LFP (lithium iron phosphate) battery can be used. An LFP battery uses lithium (Li), iron (Fe), and phosphorus (P) as the positive electrode material. Compared to secondary batteries that use rare metals such as nickel and cobalt, LFP batteries are attracting attention because they are inexpensive, have high thermal stability, and have a long lifespan. The battery 5 supplies power to the electric motor, which is the drive source of the vehicle, by discharging. Also, when the vehicle is braked, it is charged by regenerative energy from the electric motor. The battery 5 can also be charged by external charging equipment such as a rapid charger or a household outlet.
[0010] A voltage sensor 6, a current sensor 7, and a temperature sensor 8 are connected to the battery 5. The voltage sensor 6 measures the terminal voltage v(k) of the battery 5. The current sensor 7 measures the current i(k) flowing through the battery 5. The temperature sensor 8 measures the surface temperature T(k) of the battery 5. The voltage sensor 6, current sensor 7, and temperature sensor 8 perform measurements at a predetermined sampling period, for example, from when the vehicle's ignition is turned ON to when it is turned OFF. The estimation device 1 is connected to the voltage sensor 6, current sensor 7, and temperature sensor 8 by wire or wireless connection and can acquire the measured values of each. The estimation device 1 uses these measured values to estimate the State of Coherence (SOC).
[0011] The estimation device 1 can be implemented, for example, in an information processing device such as an ECU (Electronic Control Unit) installed in a vehicle. Although not shown in the diagram, the ECU consists of a processor such as a CPU (Central Processing Unit) and memory such as ROM (Read-only Memory) and RAM (Random Access Memory). Various programs executed by the estimation device 1 are stored in the memory, and the functional configuration of the estimation device 1 is realized when the processor executes the programs. Various data necessary for the processing of the estimation device 1 are also stored in the memory, and processing results are temporarily stored there as well. Although not explained in detail, in addition to the SOC estimation process, the ECU may also perform charge / discharge control of the battery 5 and vehicle operation control based on the SOC. Alternatively, the ECU may output the estimated SOC to an external control device.
[0012] As shown in Figure 1, the estimation device 1 includes an effective resistance estimation unit 11 (first estimation unit), a low-pass filter 12, and an SOC estimation unit 14 (second estimation unit) as a functional configuration for estimating SOC. The effective resistance estimation unit 11 estimates the effective resistance R(k) based on the terminal voltage v(k) and current i(k) of the battery 5 obtained from the voltage sensor 6 and the current sensor 7. The effective resistance R(k) represents the overvoltage response of the battery 5 to a predetermined current input pattern and is one of the parameters that represent the characteristics of the internal impedance of the battery 5. "Overvoltage response of the battery 5 to a predetermined current input pattern" means, for example, the impulse response to an impulse current input, or the step response to a step input.
[0013] The low-pass filter 12 performs low-pass filtering on the measured value of the battery current i(k) obtained from the current sensor 7. The low-pass filter 12 can be, for example, a moving average filter or a time-lag filter. In this embodiment, an example is described in which the low-pass filter 12 is a moving average filter and outputs a moving average value ibar(k) from the measured value of the current i(k). Here, "bar" means the average value.
[0014] The SOC estimation unit 14 inputs the current flowing through the battery 5, the time-series data of the effective resistance (overvoltage response), and the temperature of the battery 5 into the estimation model EM to determine the SOCx of the battery 5. c The SOC estimation unit 14 specifically inputs the moving average value ibar(k) of the current obtained by low-pass filtering the measured value i(k) of the current, the parameter bhat(k) of the effective resistance R(k) estimated by the effective resistance estimation unit 11, and the measured value of the battery temperature T(k) obtained from the temperature sensor 8 into the estimation model EM. Here, "hat" means the estimated value.
[0015] Furthermore, instead of using the temperature measurement from the temperature sensor 8, the temperature of the battery 5 may be estimated using a thermal equivalent circuit model that takes into account the heat generated due to the internal resistance of the battery 5. In this case, since the internal temperature of the battery can be estimated, it is expected that the accuracy of the SOC calculation will be further improved.
[0016] Figure 2 shows an equivalent circuit representing the battery system. In Figure 2, i(k) is the current flowing through battery 5, with the discharge direction being positive. Also, v(k) is the terminal voltage, and v oc (k) is the open-circuit voltage. Here, k represents a discrete time, for example, i(k) is the time kT s This refers to the value of the current i at T. ｓ is the sampling period. As shown in Figure 2, the battery system can be represented as an equivalent circuit consisting of electromotive force and internal impedance. The voltage at the electromotive force is the open-circuit voltage v. oc (k) is such that the voltage across the internal impedance is the overvoltage η(k).
[0017] Generally, when battery 5 is discharged, the terminal voltage v(k) is equal to the open-circuit voltage v oc (k) becomes lower, and during charging, the terminal voltage v(k) is equal to the open-circuit voltage v oc It becomes higher than (k). This is due to the presence of the battery's internal impedance, and such a terminal voltage v(k) and open-circuit voltage v ocThe difference in (k) is called overpotential. That is, the overpotential η(k) is represented by the following formula (1). Where v(k), v oc (k), the directions of η(k) are as shown in Figure 2. In formula (1), the current i(k) and the open circuit voltage v oc (k) correspond to the SOC x c (k) of the battery 5. It is known that they are associated via (k). SOC x c (k) is obtained by the following formula (2) using the current i(k). Where FCC is the full charge capacity. In addition, it is assumed that the current i(k) is zero-order held between sample points.
[0018] The estimation device 1 estimates the SOC x c (k) defined by this formula (2) using the current i(k) measurable from the battery 5 and the terminal voltage v(k). For example, as a method for calculating SOC based on formula (2), the current integration method is known. However, the current integration method has the problem that it cannot cope with errors in the initial integration value and current measurement errors. As another method, an estimation method using the correspondence between open circuit voltage and SOC (SOC-OCV characteristics) is also known. However, the aforementioned LFP battery has a relatively flat SOC-OCV characteristic, the region where the open circuit voltage changes depending on SOC is small, and the change width is not sufficient. Therefore, when an LFP battery is used as the battery 5, the estimation method using SOC-OCV characteristics may not ensure the SOC estimation accuracy.
[0019] Therefore, the estimation device 1 of the present embodiment estimates the SOC by using the correspondence between the internal impedance of the battery 5 and the SOC instead of the SOC-OCV characteristic. Specifically, the effective resistance estimation unit 11 of the estimation device 1 models the internal impedance of the battery 5 and calculates the effective resistance R(k) representing the characteristics of the internal impedance. The effective resistance estimation unit 11 uses a μ Markov model M as a model of the internal impedance of the battery 5.
[0020] Figure 3 is a graph showing an example of the overvoltage response to an impulse current input. Figure 3(a) is the time-series data of the current i(k) measured at a predetermined sampling period (e.g., 0.1 seconds). Figure 3(b) is the time-series data of the overvoltage η(k) corresponding to Figure 3(a), and Figure 3(c) is the time-series data of the effective resistance R(k) shown in Figure 3(b). As shown in Figures 3(a) and 3(b), when an impulse current is input, the overvoltage of the battery 5 increases rapidly and then decreases gradually. As shown in Figure 3(c), the effective resistance R(k) can be expressed as the integral value of the overvoltage η(k) from the start of current input to a predetermined time (x seconds) (effective resistance for x seconds). The effective resistance R(k) reflects the change in the overvoltage η(k) shown in Figure 3(b). It fluctuates significantly in the range SR from the start of current input to a predetermined time (x seconds), but thereafter the fluctuations subside and it maintains a nearly constant value.
[0021] As described above, the overvoltage η(k) is the voltage at the internal impedance of the battery 5. Therefore, the effective resistance R(k), which indicates the overvoltage response of the battery 5, can be used as a parameter indicating the change in the internal impedance of the battery 5. Here, using the μMarkov model M, for example, it is possible to estimate the effective resistance R(k) every 0.1 seconds over a period of 0.1 seconds to 5 seconds and obtain time-series data consisting of 50 estimated values. In the time-series data, SOCx c If the effective resistance R(k) varies significantly depending on (k) (see range SR in Figure 3(c)), the machine learning device 200 (see Figure 4) described later will capture the characteristics of that variation and determine the SOCx c (k) can be estimated with high accuracy.
[0022] This section describes a method for estimating the effective resistance R(k) using the μMarkov model M. The value of the step response of the internal impedance G(q) in the battery system shown in Figure 2 at time K is the K-step effective resistance R K It can be defined as follows: Effective resistance R K This can be calculated using the impulse response g(k) of the internal impedance G(q) by the following equation (3). That is, the effective resistance R K can be obtained by calculating the impulse response g(k) of the internal impedance G(q) and calculating the sum up to the K-th term. The μ-Markov model M represents the internal impedance G(q), and the effective resistance R K has been proposed as a model that can efficiently calculate . The μ-Markov model M is a combination of an FIR (Finite Impulse Response) model and an ARX (Auto-Regressive with eXogenous) model. The μ-Markov model M is represented by the following equation (4). The first term on the right-hand side of equation (4) represents the FIR model, and the second term and the third term on the right-hand side represent the ARX model. In this μ-Markov model M, the effective resistance R in the FIR model K can be interpreted as replacing terms with k>K that are unnecessary for calculation with a rational transfer function. That is, the μ-Markov model M can reduce the number of parameters unnecessary for estimating effective resistance in the FIR model, and efficiently represent a linear system.
[0023] Here, if the order p is selected to be equal to the order of the true system, there exists a set of parameters for which equation (4) is equivalent to the true system. On the other hand, if the order p is different from the true order, such a set of parameters does not exist. However, the estimated value b of the impulse response l has been shown to provide an effective estimated value without depending on the order p of the system, and this is an excellent property of the μ-Markov model M.
[0024] The effective resistance estimation unit 11 uses the μ-Markov model M to obtain the effective resistance R K for estimation, and uses the recursive least squares method. The recursive least squares method is a recursive version of the least squares method for linear regression models, as shown in the following equations (5) to (10).
[0025] In an actual battery 5, dynamic characteristics change moment by moment depending on the SOC and temperature. Therefore, the forgetting factor λ is used in equation (7). The forgetting factor is a parameter set by the user within the range of 0 < λ ≤ 1. Reducing the value of the forgetting factor λ improves the followability to parameter fluctuations. On the other hand, since less data is effectively used, sensitivity to noise increases and estimation errors become larger. It is desirable to determine the value of the forgetting factor λ in consideration of the trade-off between followability to parameter fluctuations and sensitivity to noise.
[0026] The effective resistance estimation unit 11 uses the current i(k) measured by the current sensor 7 as the input u(k) and the terminal voltage v(k) measured by the voltage sensor 6 as the output y(k), and calculates a regression vector φ (see equation (7)) that constitutes the μ-Markov model M. The effective resistance estimation unit 11 calculates a parameter θhat (see equation (6)) from the calculated regression vector φ. For SOC estimation, among the calculated parameters θhat, the parameter bhat (b0...b K ) is used.
[0027] Here, for SOC estimation, all the calculated parameters b0...b K , it is also possible to use only parameters in a range where the change in effective resistance R(k) (overvoltage response) is large (for example, the range SR shown in (c) of FIG. 3). As shown in (c) of FIG. 3, the effective resistance R(k) of the battery fluctuates greatly immediately after the start of current input, and the fluctuation gradually converges to a substantially constant value. In machine learning, since SOCx c (k) is estimated by capturing the characteristics of large fluctuations of the effective resistance R(k), it is assumed that parameters in a range with poor fluctuations contribute little to the SOC estimation accuracy.
[0028] As described above, the μ-Markov model M itself is a system that reduces unnecessary data points, and further, by excluding parameters with low contribution to estimation accuracy, the computational load on the estimation device 1 can be reduced.
[0029] The effective resistance estimation unit 11, for example, obtains parameters b0...b KFrom among these parameters, parameters whose fluctuations exceed a predetermined amount may be extracted and input to the SOC estimation unit 14 as time-series data of the effective resistance R(k). The effective resistance estimation unit 11 may, for example, calculate the amount of fluctuation of the parameter at each time point from the previous time point, and determine the range in which the fluctuation amount exceeds a predetermined amount as the range from which to extract parameters. This predetermined amount can be set to, for example, a value of at least four times the voltage detection resolution to suppress the effect of quantization errors. Alternatively, if the fluctuation pattern of the effective resistance is predictable, the effective resistance estimation unit 11 may, for example, extract parameters included in a predetermined period from the start of current input.
[0030] Here, effective resistance is a parameter that indicates the characteristics of the internal impedance and depends on the State of Control (SOC). However, effective resistance also depends on the magnitude and direction of the current flowing through the battery 5, and the temperature of the battery 5, in addition to the SOC. Therefore, in order to accurately estimate the SOC based on effective resistance, it is desirable to consider the current and temperature of the battery 5 when making the estimation. For this reason, the SOC estimation unit 14 performs the estimation using at least the data of the effective resistance R(k) and current i(k) of the battery 5. More preferably, the SOC estimation unit 14 performs the estimation using data of three points: effective resistance R(k), current i(k), and temperature T(k).
[0031] In the conventional approach to estimating SOC by considering three points—effective resistance R(k), current i(k), and temperature T(k)—it is necessary to perform numerous experiments and simulations to create maps and conversion logic. Furthermore, complex programs must be created to execute the conversion logic by referring to the maps. In addition, the estimation method using the μMarkov model M described above allows for the estimation of effective resistance R(k) every 0.1 seconds, for example, and the time-series data of effective resistance R(k) consists of numerous estimated values. Performing complex calculations on numerous estimated values leads to an increased computational load.
[0032] Therefore, in this embodiment, the SOC estimation unit 14 receives an input signal (input data) and an output signal (output data) called SOCx cThe SOC is estimated using an estimation model EM that has learned the correspondence with (k). The input signal includes at least the current i(k) flowing through the battery 5 and the effective resistance R(k), which is time-series data of the overvoltage response. The input signal may more preferably further include the battery temperature T(k). Figure 4 shows an example of machine learning of the estimation model EM. As shown in Figure 4, the estimation model EM can be pre-created, for example, in a machine learning device 200 and then implemented in the estimation device 1. The machine learning device 200 may be, for example, a device installed outside the vehicle. In this case, the estimation model EM created by the machine learning device 200 may be implemented in the estimation device 1 before it is installed in the vehicle. Alternatively, the estimation model EM may be delivered to the estimation device 1 after it has been installed in the vehicle via the cloud or the like. The machine learning device 200 may also be installed in the vehicle. The machine learning device 200 may be implemented in an information processing device separate from the information processing device that implements the estimation device 1, or it may be implemented in the same information processing device. The machine learning device 200 may create or update an estimation model EM using data obtained by operating the battery 5 while the vehicle is running.
[0033] The SOC estimation model EM can be expressed, for example, by the following equation (11). bhat is an estimated value θhat, which is a parameter of the effective resistance R(k) shown in equation (6), where b0...b K This is a vector containing only the estimated values, and is represented by the following equation (12). ibar is the moving average value of the current i(k), and is expressed by the following equation (13). The moving average value ibar can be obtained by applying a low-pass filter similar to the low-pass filter 12 shown in Figure 1 to the measured value of current i(k). T(k) is the temperature of the battery 5.
[0034] In this embodiment, the nonlinear function f in equation (11) is constructed using a machine learning model. The machine learning model is not limited to a specific model, but for example, a neural network (NN), a random forest, etc., can be used. As shown in Figure 4, for the nonlinear function f, a dataset DS is required, which consists of input data parameters bhat(k), moving average value of current ibar(k), temperature T(k), and output data, the true value of SOC. By learning the correspondence between the input data and the output data, the machine learning model can create an estimation model EM that estimates SOC based on the input data. The machine learning model can efficiently learn the nonlinear function f by, for example, using backpropagation on the dataset DS.
[0035] In equation (11), only a finite-length moving average is used for the current i(k). Also, the measured value of the terminal voltage v(k) of battery 5 is not used directly. Therefore, the estimation method of SOC corresponding to the current integration method in equation (2), and the open-circuit voltage v from the terminal voltage v(k) are not used. oc It is considered impossible to create a model that corresponds to the conventional estimation method of calculating (k) to estimate the SOC. Therefore, the SOC estimation using the machine learning model in this embodiment estimates the SOC using a mechanism different from the conventional estimation method.
[0036] The training dataset DS can be created, for example, by the following method: - Create a charge / discharge measurement device equipped with a battery of the same type as the battery 5 used in the vehicle (e.g., an LFP battery) that simulates the discharge and charge that occur when the vehicle is running. - For discharge, connect multiple resistors of different sizes in parallel to the battery 5 and control the load by switching these connections with a MOSFET (Metal-oxide-semiconductor Field-effect Transistor), etc. - For charging, control the device to switch between CC (Constant Current) charging and CV (Constant Voltage) charging. - Operate the charge / discharge measurement device with a program that performs random charging and discharging. - Measure the current i(k), terminal voltage v(k), and temperature T(k) of the battery 5 with a voltage sensor 6, a current sensor 7, and a temperature sensor 8. - Input the measured values of terminal voltage v(k) and current i(k) into a μMarkov model M to calculate the parameter θhat(k) of the effective resistance R(k). - Calculate the moving average value ibar(k) by applying a low-pass filter to the current i(k). - Calculate the true SOC value using the current integration method of equation (2) or the measured values of terminal voltage v(k) and current i(k). - Create a dataset DS with the parameter bhat(k) of the effective resistance R(k), the moving average value ibar(k) of the current, and the temperature T(k) as input data, and the true SOC value as output data.
[0037] [Experimental Example] The usefulness of SOC estimation using a machine learning-based estimation model EM will be explained based on an experimental example. Figure 5(a) is a graph showing the data for current i(k) used in SOC estimation, and Figure 5(b) is a graph showing the data for terminal voltage v(k). Although not shown in the illustration, data for the temperature T(k) of battery 5 is also prepared in the experiment. The estimation model EM used in the experiment was created by inputting the aforementioned training dataset DS (see Figure 4) into an NN model. For training, the regression learner app included in the Statistics and Machine Learning Toolbox of MATLAB® was used, with the number of fully connected layers set to 1, the layer size to 100, and the iteration limit to 500.
[0038] The current i(k) and terminal voltage v(k) data shown in Figure 5 were input into the μMarkov model M to calculate the parameter θhat(k) of the effective resistance R(k). Furthermore, a low-pass filter was applied to the current i(k) data to calculate the moving average value ibar(k). The data prepared for input into the estimation model EM included the parameter bhat(k) of the effective resistance R(k) over a 5-second period, the effective resistance value over 5 seconds, the moving average value of the current ibar(k), and the temperature T(k).
[0039] Using the estimation model EM, the State of Cost (SOC) was estimated under the following conditions 1 to 4: • Example 1: The SOC was estimated by inputting three points into the estimation model EM: the parameter bhat(k) of the effective resistance R(k) over 5 seconds, the moving average value of the current ibar(k), and the temperature T(k). • Example 2: The SOC was estimated by inputting three points into the estimation model EM: the effective resistance value, the moving average value of the current ibar(k), and the temperature T(k). • Example 3: The SOC was estimated by inputting two points into the estimation model EM: the parameter bhat(k) of the effective resistance R(k) over 5 seconds, and the moving average value of the current ibar(k). • Example 4: The SOC was estimated by inputting only the parameter bhat(k) of the effective resistance R(k) over 5 seconds into the estimation model EM.
[0040] Figure 6(a) is a graph showing the true SOC value. Figure 6(b) is a graph showing the SOC estimation result for the first example. Figure 6(c) is a graph showing the SOC estimation result for the second example. Figure 7(a) is a graph showing the SOC estimation result for the third example. Figure 7(b) is a graph showing the SOC estimation result for the fourth example.
[0041] The first example is an experimental example in accordance with the embodiment, in which the SOC is estimated using multiple parameter bhat(k) of the effective resistance R(k) obtained from the μMarkov model. On the other hand, the second example estimates the SOC using a single 5-second effective resistance value. Both the first example (see Figure 6(b)) and the second example (see Figure 6(c)) track the true SOC value (see Figure 6(a)), but the first example has fewer outliers than the second example, indicating that it tracks the true SOC more closely. From this, the usefulness of SOC estimation using multiple data points can be confirmed.
[0042] In the third and fourth examples, similar to the first example, the multi-point parameter bhat(k) of the effective resistance R(k) obtained from the μMarkov model is used, but in the third example, the temperature T(k) is not input into the estimation model EM. In the fourth example, the moving average value of the current ibar(k) and the temperature T(k) are not input into the estimation model EM. Both the third example (see Figure 7(a)) and the fourth example (see Figure 7(b)) track the true SOC value (see Figure 6(a)), but the third example has fewer outliers than the fourth example. Furthermore, comparing the third and fourth examples with the first example (see Figure 6(b)), it can be seen that the first example has fewer outliers and tracks the true SOC value more closely. The RMSE (root mean square error) for the first and third examples were 2.88 and 3.00, respectively.
[0043] As mentioned above, the estimation model EM uses the input signal and SOCx ｃ It is created by learning the correspondence with (k). From the experimental results of the first, third, and fourth examples, it can be seen that by increasing the number of input signal points to the estimation model EM, the estimation model EM can capture the characteristics of a variety of fluctuations, and the estimation accuracy of SOC can be further improved.
[0044] [Estimation Process for Effective Resistance] Figure 8 is a flowchart showing the processing of the estimation device 1 in this embodiment. As shown in Figure 8, the estimation device 1 acquires measured values of terminal voltage v(k), current i(k), and temperature T(k) from the voltage sensor 6, current sensor 7, and temperature sensor 8 at a predetermined sampling period (step S01). The effective resistance estimation unit 11 estimates the effective resistance R(k) using the μMarkov model M with the measured values of terminal voltage v(k) and current i(k) (step S02). The low-pass filter 12 calculates the moving average value ibar(k) by applying a low-pass filter to the measured value of current i(k) (step S03). The SOC estimation unit 14 calculates the SOCx using the estimation model EM. c (k) is estimated (step S04). The SOC estimation unit 14 inputs at least two data points, the effective resistance parameter bhat(k) and the moving average value of the current ibar(k), into the estimation model to determine SOCxc (k) can be estimated. The SOC estimation unit 14 inputs the temperature T(k) in addition to the data from the two points mentioned above into the estimation model to estimate SOCx c (k) can be estimated.
[0045] [Estimation System] Figure 9 shows the configuration of the estimation system 100, which includes the estimation device 1. As shown in Figure 9, the estimation system 100 comprises one or more vehicles VH equipped with the estimation device 1, and a machine learning device 2 that performs machine learning of the estimation model EM based on data collected from the estimation device 1. The machine learning device 2 can perform machine learning of the estimation model EM in the same manner as the machine learning device 200 shown in Figure 4. The estimation device 1 of each vehicle VH has the configuration shown in Figure 1 and has an estimation model EM that has been created in advance by machine learning implemented on it. Each estimation device 1 performs SOC estimation using the estimation model EM while the vehicle VH is running.
[0046] The machine learning device 2 can be composed of an information processing device equipped with a processor such as a CPU and memory such as ROM and RAM. The memory stores various programs to be executed by the machine learning device 2. The processor executes these programs, thereby realizing the functional configuration of the machine learning device 2. Furthermore, the memory also stores an estimation model EM similar to the estimation model EM provided in each estimation device 1.
[0047] The machine learning device 2 is connected to each estimation device 1 via a network. The machine learning device 2 comprises a data acquisition unit 21, a storage unit 22, and a machine learning unit 23. The storage unit 22 consists of memory. For example, the storage unit 22 stores data necessary for processing each functional configuration of the machine learning device 2. The storage unit 22 also temporarily stores the processing results of each functional configuration.
[0048] The data acquisition unit 21 receives input data from the estimation device 1 installed in the vehicle VH to the estimation model EM (effective resistance parameter bhat(k), moving average value of current ibar(k), temperature T(k)) and output data, the SOC estimation value x c(k) is collected. The data collection unit 21 stores the collected data in the storage unit 22 as a training dataset DS. The data collection unit 21 can, for example, randomly select vehicles VH based on predetermined conditions and collect data. The predetermined conditions can be set based on, for example, the type of vehicle VH, geographical information of the area in which the vehicle VH travels, and environmental information such as temperature or weather. The data collection unit 21 can attach tags to the dataset DS stored in the storage unit 22 to classify the type of vehicle VH, geographical information, environmental information, etc.
[0049] The machine learning unit 23 performs machine learning on the estimation model EM using the dataset DS stored in the memory unit 22. The dataset DS stored in the memory unit 22 is collected from actual running vehicles VH and reflects the type of vehicle VH, geographical information, environmental information, etc. By performing machine learning on the estimation model EM using this dataset DS, the estimation model EM can be updated to reflect the actual running conditions of the vehicle VH.
[0050] The machine learning unit 23 transmits update data UD of the estimated model EM to the estimation device 1 of the vehicle VH. The estimation device 1 can update the estimated model EM by installing the received update data UD into the estimated model EM. The machine learning unit 23 may transmit the same update data UD to all estimation devices 1, or it may create multiple different update data UDs and select the update data UD to transmit according to the vehicle VH. For example, the machine learning unit 23 can search for tags to extract a machine learning dataset MDS that meets specific conditions. The machine learning unit 23 can update the estimated model EM using only the extracted machine learning dataset MDS. This makes it possible to optimize the estimated model EM according to the type of vehicle VH, geographical conditions, environmental conditions, etc. Then, the update data UD of the estimated model EM is transmitted to the estimation device 1 of the vehicle VH that matches the type of vehicle VH, geographical conditions, environmental conditions, etc. This allows the estimation device 1 to use the optimized estimated model EM, and an improvement in the estimation accuracy of SOC is expected.
[0051] As described above, the estimation device 1 of this embodiment has the following configuration: (1) The estimation device 1 is the SOCx of the battery 5 c The (k) (charge rate) is estimated. The estimation device 1 comprises an effective resistance estimation unit 11 (first estimation unit) and an SOC estimation unit 14 (second estimation unit). The effective resistance estimation unit 11 estimates the effective resistance R(k), which is the overvoltage response of the battery 5 to a predetermined input pattern of current, based on measured values of the current i(k) and terminal voltage v(k) flowing through the battery 5. The SOC estimation unit 14 inputs at least the current i(k) flowing through the battery 5 and the parameter bhat(k), which is time-series data of the effective resistance estimated by the effective resistance estimation unit 11, into an estimation model EM created by machine learning, and estimates the SOCx of the battery 5. c (k) is estimated. The estimation model EM can be created by learning the correspondence between the input data and the output data, which is the SOC. The input data can be created by learning the correspondence between at least the current i(k) flowing through the battery 5, the parameter bhat(k) of the effective resistance R(k), and the SOC of the battery 5. Note that "current i(k) flowing through the battery 5" may be the measured value of current i(k) itself, or it may be the moving average value ibar(k) obtained by processing the measured value of current i(k) with a low-pass filter.
[0052] The estimation device 1 of this embodiment can improve estimation accuracy while reducing the computational load when estimating the SOC. LFP batteries are attracting attention as batteries 5 that supply power to the drive source of a vehicle VH. However, LFP batteries have a relatively flat SOC-OCV characteristic, and estimation errors are likely to occur in SOC estimation methods based on the SOC-OCV characteristic. The estimation device 1 of this embodiment focuses on the correspondence between the SOC and the internal impedance of the battery 5, and estimates the SOC based on the effective resistance, which is one of the parameters that indicate the characteristics of the internal impedance.
[0053] Here, the effective resistance also fluctuates due to various factors other than SOC. In a typical approach, if these factors are also considered, it is necessary to perform numerous experiments and simulations to generate maps and transformation logic. Furthermore, it is expected that the estimation accuracy will improve if the estimation device 1 performs estimation using a large number of data points. However, if the estimation device 1 performs calculation processing using the aforementioned maps and logic for each of the numerous data points, the computational load will increase.
[0054] In the estimation device 1 of this embodiment, the SOC estimation unit 14 estimates the SOC using an estimation model EM created using the training dataset DS. By utilizing machine learning, the effort required for experiments and simulations to generate maps and transformation logic can be reduced. Furthermore, the SOC estimation unit 14 does not need to perform complex calculations using maps or transformation logic when estimating the SOC. Therefore, the estimation device 1 can improve the accuracy of SOC estimation by using a large amount of data while reducing the computational load.
[0055] (2) In the estimation device 1 of (1), the input data to the estimation model EM may further include the temperature T(k) of the battery 5. The SOC estimation unit 14 can estimate the SOC of the battery 5 by inputting the current i(k) flowing through the battery 5, the parameter bhat(k) of the effective resistance R(k) estimated by the effective resistance estimation unit 11, and the temperature T(k) of the battery 5 to the estimation model EM. The "temperature T(k) of the battery 5" may be a measurement value obtained from a temperature sensor 8 that measures the surface temperature of the battery 5, or it may be the internal temperature of the battery 5 estimated by a thermal equivalent circuit model or the like.
[0056] One factor that affects the effective resistance R(k) is the temperature T(k) of the battery 5. In this embodiment, three points are used as input data for the estimation model EM: the effective resistance R(k), current i(k), and temperature T(k) of the battery 5. This allows for the calculation of SOCx cThe estimation accuracy of (k) can be improved. In conventional estimation methods, generating maps and transformation logic while considering the three points is laborious and the computational processing is complex. In this embodiment, machine learning can be used to create an estimation model EM that considers the three points while reducing the labor involved. Furthermore, the estimation accuracy of SOC can be improved while reducing the computational load on the estimation device 1.
[0057] (3) In the estimation device 1 of (1) or (2) above, the effective resistance estimation unit 11 estimates the effective resistance R(k) of the battery 5 by inputting the measured values of the current i(k) and the terminal voltage v(k) into the μMarkov model M.
[0058] The μMarkov model M combines the FIR model and the ARX model, and can reduce the number of parameters unnecessary for estimating the effective resistance R(k) in the FIR model. As a result, the estimation device 1 can improve the accuracy of the effective resistance R(k) estimation while reducing the computational load.
[0059] (4) In any of the estimation devices 1 described in (1) to (3), the SOC estimation unit 14 extracts a range SR in which the variation of the effective resistance R(k) exceeds a predetermined amount from the time-series data of the effective resistance R(k) estimated by the effective resistance estimation unit 11, and inputs it into the estimation model EM.
[0060] By extracting only the range expected to contribute to the SOC estimation accuracy from the time-series data of effective resistance R(k), the computational load of the estimation device 1 can be further reduced while maintaining the SOC estimation accuracy. Specifically, for example, the effective resistance estimation unit 11 can calculate the amount of change from the previous time point for the parameter bhat(k) at each time point, extract the parameter bhat(k) in the range where the amount of change is greater than or equal to a predetermined amount, and input it to the SOC estimation unit 14. Alternatively, the effective resistance estimation unit 11 can extract the parameter bhat(k) included in a predetermined period from the start of current input and input it to the SOC estimation unit 14.
[0061] (5) In any of the estimation devices 1 described in (1) to (4), the SOC estimation unit 14 inputs the measured value of the current i(k) that has been low-pass filtered by the low-pass filter 12 to the estimation model EM. The low-pass filter 12 can be, for example, a moving average filter. In this case, the SOC estimation unit 14 inputs the moving average value ibar(k) of the current to the estimation model EM.
[0062] The internal impedance of battery 5 changes depending on the magnitude of the current i(k) input to battery 5. It is assumed that there will be a delay between the timing of the change in the measured value of current i(k) and the timing of the change in the internal impedance of battery 5. Therefore, in this embodiment, a low-pass filter is applied to the measured value of current i(k) to reflect the delay in the change in internal impedance. As a result, the current i(k) input to the estimation model EM corresponds to the effective resistance R(k) which represents the change in overvoltage η(k), thereby improving the estimation accuracy of SOC.
[0063] The estimation system 100 according to this embodiment comprises, for example, the following configuration: (7) The estimation system 100 comprises one or more vehicles VH and a machine learning device 2. The vehicle VH is provided with an estimation device 1 as described in any of (1) to (5) above and (6) below. Power is supplied to the drive source of the vehicle VH by a battery 5. The machine learning device 2 is connected to the estimation device 1 of each vehicle VH via a network. The estimation device 1 receives input data to the estimation model EM (effective resistance parameter bhat(k), moving average value of current ibar(k), temperature T(k)) and output data to the estimation model EM (SOC estimate x c (k)) is sent to the machine learning device 2. The machine learning device 2 updates the estimated model EM using the input and output data received from the estimation device 1. The machine learning device 2 sends the updated data UD (data related to the updated estimated model EM) to the estimation device 1.
[0064] Thus, in the estimation system 100, the machine learning device 2 can collect data from the actual running vehicle VH and update the estimation model EM in real time. Furthermore, the machine learning device 2 can optimize the estimation model EM according to the type of vehicle VH, geographical conditions, environmental conditions, etc., for each estimation device 1, thereby further improving the estimation accuracy of the SOC in each estimation device 1. In addition, by having the external machine learning device 2 update the estimation model EM, the computational load on the estimation device 1 for each vehicle VH can be reduced.
[0065] [Modification 1] Figure 10 is a block diagram showing the configuration of the estimation device 1A according to Modification 1. Although not shown in Figure 10, the estimation device 1A is connected to a voltage sensor 6, a current sensor 7, and a temperature sensor 8 (see Figure 1), similar to the embodiment, and receives measured values of terminal voltage v(k), current i(k), and temperature T(k) from each. As shown in Figure 10, the estimation device 1A includes high-pass filters 15 and 16. The high-pass filters 15 and 16 perform similar high-pass filtering. The high-pass filter 15 performs high-pass filtering on the measured value of terminal voltage v(k) obtained from the voltage sensor 6 and inputs it to the effective resistance estimation unit 11. The high-pass filter 16 performs high-pass filtering on the measured value of current i(k) obtained from the current sensor 7 and inputs it to the effective resistance estimation unit 11.
[0066] As shown in the embodiment (see Figure 1), it is possible to estimate the effective resistance R(k) by directly inputting the measured value of the terminal voltage v(k) into the μMarkov model M. However, as shown in Figure 2, the terminal voltage v(k) includes an overvoltage η(k) and an open-circuit voltage V oc The component (k) is included. Since the effective resistance R(k) exhibits an overvoltage response, the estimation accuracy can be further improved by inputting a component close to the overvoltage η(k) into the μMarkov model M.
[0067] In Modification 1, a high-pass filter is applied to the measured value of the terminal voltage v(k), thereby reducing the open-circuit voltage V ocThe low-frequency component corresponding to (k) is removed, and a component close to the overvoltage η(k) can be extracted. However, the high-pass filter also removes the low-frequency component of the overvoltage η(k) itself. Therefore, in order to maintain consistency of the elements input to the μMarkov model M, in the modified example 1, the same high-pass filter is also applied to the measured value of current i(k). This improves the estimation accuracy of the effective resistance R(k).
[0068] As described above, the estimation device 1A according to the modified example 1 has the following configuration. (6) In any of (1) to (5) above, the effective resistance estimation unit 11 estimates the effective resistance R(k) based on the measured values of the current i(k) and terminal voltage v(k) that have been high-pass filtered by the high-pass filters 15 and 16.
[0069] By applying a high-pass filter to the measured terminal voltage v(k), the open-circuit voltage V oc By removing the low-frequency component corresponding to (k), a component close to the overvoltage η(k) can be extracted. By inputting a component close to the overvoltage η(k) into the μMarkov model M, the estimation accuracy of the effective resistance R(k), which is the overvoltage response, can be further improved. In addition, by applying a high-pass filter to the measured value of current i(k), the consistency of the elements input to the μMarkov model M can be maintained.
[0070] The estimation device 1A according to Modification 1 is also applicable to the estimation system 100 shown in Figure 9. That is, at least one vehicle VH constituting the estimation system 100 may be equipped with the estimation device 1A according to Modification 1.
[0071] 1, 1A: Estimation device 2, 200: Machine learning device 5: Battery 6: Voltage sensor 7: Current sensor 8: Temperature sensor 11: Effective resistance estimation unit (first estimation unit) 12: Low-pass filter 14: SOC estimation unit (second estimation unit) 15, 16: High-pass filter 100: Estimation system M: μMarkov model EM: Estimation model VH: Vehicle DS: Dataset
Claims
1. An estimation device for estimating the charge level of a battery, comprising: a first estimation unit that estimates the overvoltage response of the battery to a predetermined input pattern of the current based on measured values of the current flowing through the battery and the terminal voltage; and a second estimation unit that estimates the charge level by inputting at least the measured values of the current and the time-series data of the overvoltage response estimated by the first estimation unit into an estimation model created by machine learning, wherein the estimation model is created by learning the correspondence between input data and output data, the charge level, and the input data includes at least the current flowing through the battery and the time-series data of the overvoltage response.
2. The estimation device according to claim 1, wherein the input data further includes the temperature of the battery, and the second estimation unit inputs the battery temperature in addition to the measured value of the current and the time-series data of the overvoltage response to the estimation model to estimate the charge level of the battery.
3. The estimation device according to claim 1, wherein the first estimation unit estimates the overvoltage response of the battery by inputting the measured values of the current and terminal voltage into a μMarkov model.
4. The estimation device according to claim 1, wherein the second estimation unit extracts a range from the time-series data of the overvoltage response in which the variation of the overvoltage response is greater than or equal to a predetermined amount, and inputs it into the estimation model.
5. The estimation device according to claim 1, wherein the second estimation unit inputs the measured value of the current, which has been processed with a low-pass filter, to the estimation model.
6. The estimation device according to claim 1, wherein the first estimation unit estimates the overvoltage response based on high-pass filtered current and terminal voltage measurements.
7. An estimation system comprising: one or more vehicles, each equipped with an estimation device as described in any one of claims 1 to 6, the vehicle being powered by a battery to a drive source; and a machine learning device connected to the estimation device via a network, wherein the estimation device transmits input data to the estimation model and output data of the estimation model to the machine learning device; and the machine learning device updates the estimation model using the input data and output data received from the estimation device, and transmits data relating to the updated estimation model to the estimation device.