Output characteristic prediction device, output characteristic prediction system, and output characteristic prediction method

Abstract

The present invention provides an output characteristic prediction device, an output characteristic prediction system, and an output characteristic prediction method that can improve the accuracy of output characteristic prediction. [Solution] The prediction unit repeatedly executes the prediction process. The identification unit identifies source vehicles similar to the target vehicle. The model improvement unit improves the regression model of the target vehicle by performing transfer learning using the regression models of similar source vehicles. The calculation unit calculates the prediction accuracy index of the improved regression model by interpolation. If the prediction accuracy in the current prediction process is better than in the previous prediction process, the setting unit updates the source vehicles with high similarity to the target vehicle to the source vehicles identified as similar to the target vehicle in the current prediction process. The extrapolation unit extrapolates the battery output characteristics of the target vehicle using the regression model of the target vehicle, which has been improved by performing transfer learning using the regression models of source vehicles with high similarity to the target vehicle.

Description

Technical Field

[0001] The present invention relates to an output characteristic prediction device, an output characteristic prediction system, and an output characteristic prediction method.

Background Art

[0002] There is a need to predict the output characteristics of a battery provided in a vehicle. The output voltage prediction system described in Patent Document 1 includes a plurality of vehicles provided with fuel cells and an information processing device that predicts the output voltage of the fuel cells. The information processing device generates a prediction model for predicting the output voltage of the fuel cells by performing machine learning using time-series data of a plurality of vehicles. This prediction model is a common prediction model for a plurality of vehicles. The information processing device predicts the output voltage of the fuel cells using the generated prediction model.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] It is required to improve the prediction accuracy of the output characteristics of the battery of each vehicle as compared with the prior art.

Means for Solving the Problems

[0005] An output characteristic prediction device for solving the above problems is an output characteristic prediction device that predicts the output characteristics of a battery provided in a vehicle, including an acquisition unit that individually acquires vehicle data including time-series data from a plurality of the vehicles, a model generation unit that generates a regression model for each of the vehicles to output a target variable including the output characteristics when an explanatory variable is input by performing machine learning using the time-series data for a learning target period for each of the vehicles, and a prediction unit that executes a prediction process for predicting the output characteristics of the battery for each of the vehicles. Each time the acquisition unit newly acquires the vehicle data for a predetermined period from the plurality of vehicles, the prediction unit executes the prediction process. When one of the plurality of vehicles is set as a target vehicle and the vehicles other than the target vehicle are set as source vehicles, the prediction unit includes a specifying unit that specifies one or more source vehicles similar to the target vehicle based on the vehicle data for the predetermined period of the plurality of vehicles, a model improvement unit that improves the regression model of the target vehicle by performing transfer learning using the regression coefficients of the regression models of one or more source vehicles similar to the target vehicle, a calculation unit that calculates a prediction accuracy index of the improved regression model of the target vehicle by interpolating the output characteristics of the battery of the target vehicle in the predetermined period by using the improved regression model of the target vehicle, a setting unit that sets one or more source vehicles having a high similarity to the target vehicle based on the prediction accuracy index, and an extrapolation unit that extrapolates the output characteristics of the battery of the target vehicle by using the regression model of the target vehicle improved by performing transfer learning using the regression coefficients of the regression models of one or more source vehicles having a high similarity to the target vehicle. When the prediction accuracy in the current prediction process is higher than the prediction accuracy in the previous prediction process, the setting unit updates the source vehicle having a high similarity to the target vehicle from the source vehicle set in the previous prediction process to the source vehicle specified as being similar to the target vehicle in the current prediction process.

[0006] According to the above configuration, each time the acquisition unit acquires new vehicle data for a predetermined period from multiple vehicles, the prediction unit performs prediction processing. In the prediction processing, the identification unit identifies source vehicles similar to the target vehicle based on the vehicle data for a predetermined period from multiple vehicles. Therefore, it can handle changes in the combination of similar vehicles as time passes.

[0007] Furthermore, according to the above configuration, the setting unit sets source vehicles with a high similarity to the target vehicle based on the comparison result between the prediction accuracy in the current prediction process and the prediction accuracy in the previous prediction process. Therefore, source vehicles that improve prediction accuracy can be set as source vehicles with a high similarity to the target vehicle.

[0008] The extrapolation unit extrapolates the battery output characteristics of the target vehicle using a regression model of the target vehicle that has been improved by transfer learning using the regression coefficients of the regression model of the source vehicle, which has a high similarity to the target vehicle. Therefore, the prediction accuracy of the battery output characteristics is improved.

[0009] With respect to the output characteristic prediction device, the vehicle data may include, as time-series data, time-series data relating to the power generation control of the battery and time-series data relating to the usage status of the vehicle.

[0010] According to the above configuration, regression models can be generated effectively. For the output characteristic prediction device, the value at each point in time in the time series data may be the average, median, or cumulative value of the values ​​over a predetermined calculation period.

[0011] According to the above configuration, the amount of time series data can be reduced, thereby reducing the processing load involved in generating the regression model. Furthermore, if the value at each point in the time series data is the median of the values ​​over a predetermined calculation period, noise can be effectively reduced.

[0012] An output characteristic prediction system that solves the above problems comprises a plurality of vehicles equipped with batteries, and an output characteristic prediction device that predicts the output characteristics of the batteries, wherein the output characteristic prediction device is the output characteristic prediction device described above.

[0013] Regarding the output characteristic prediction system, the vehicle comprises a fuel cell as the battery, an energy storage device capable of charging the output power of the fuel cell, a charge rate detection unit for detecting the charge rate of the energy storage device, and an FC control device for controlling the power generation of the fuel cell. The FC control device is capable of performing switching control to switch the output power of the fuel cell in stages based on the charge rate of the energy storage device detected by the charge rate detection unit, and the vehicle data may include time-series data of the output characteristics of the fuel cell during the execution of the switching control as time-series data.

[0014] With the above configuration, the output characteristics of the fuel cell remain stable while the switching control is in operation. By generating a regression model using time-series data of the fuel cell's output characteristics during the switching control, the prediction accuracy of the fuel cell's output characteristics can be improved.

[0015] The output characteristic prediction method for solving the above problem is an output characteristic prediction method for predicting the output characteristics of a battery installed in a vehicle using an output characteristic prediction device, wherein the output characteristic prediction device comprises: an acquisition unit that individually acquires vehicle data including time series data from a plurality of vehicles; a model generation unit that generates a regression model for each vehicle that outputs a target variable including the output characteristics when an explanatory variable is input, by machine learning using the time series data for a learning target period for each vehicle; and a prediction unit that performs a prediction process to predict the output characteristics of the battery for each vehicle, wherein the prediction process is performed each time the acquisition unit newly acquires vehicle data for a predetermined period from the plurality of vehicles, and when one of the plurality of vehicles is designated as a target vehicle and the other vehicles are designated as source vehicles, the prediction process includes the steps of: an identification unit of the prediction unit identifying one or more source vehicles similar to the target vehicle based on the vehicle data for the predetermined period from the plurality of vehicles; and a model improvement unit of the prediction unit using the regression coefficients of the regression model of one or more source vehicles similar to the target vehicle for transfer learning The process includes the steps of: improving the regression model of the target vehicle by learning; calculating a prediction accuracy index for the improved regression model of the target vehicle by interpolating the output characteristics of the battery of the target vehicle over a predetermined period using the improved regression model of the target vehicle; setting one or more source vehicles with high similarity to the target vehicle based on the prediction accuracy index; and extrapolating the output characteristics of the battery of the target vehicle by the regression model of the target vehicle, which has been improved by transfer learning using the regression coefficients of the regression models of one or more source vehicles with high similarity to the target vehicle, wherein in the step of setting source vehicles with high similarity to the target vehicle, if the prediction accuracy in the current prediction process is higher than the prediction accuracy in the previous prediction process, the setting unit selects the source vehicles with high similarity to the target vehicle from the source vehicles set in the previous prediction process.In this prediction process, the source vehicle identified as similar to the target vehicle is updated.

[0016] According to the above method, each time the acquisition unit acquires new vehicle data for a predetermined period from multiple vehicles, the prediction unit performs prediction processing. The prediction processing includes a step in which the identification unit identifies source vehicles similar to the target vehicle based on the vehicle data for a predetermined period from multiple vehicles. Therefore, it can handle changes in the combination of similar vehicles over time.

[0017] Furthermore, according to the above method, the prediction process includes a step in which the setting unit sets source vehicles with high similarity to the target vehicle based on a comparison of the prediction accuracy in the current prediction process with the prediction accuracy in the previous prediction process. Therefore, source vehicles that improve prediction accuracy can be set as source vehicles with high similarity to the target vehicle.

[0018] The prediction process includes a step in which the extrapolation unit extrapolates the battery output characteristics of the target vehicle using an improved regression model of the target vehicle, which is achieved by transfer learning using the regression coefficients of the regression model of the source vehicle that has a high similarity to the target vehicle. Therefore, the prediction accuracy of the battery output characteristics is improved. [Effects of the Invention]

[0019] According to the present invention, the prediction accuracy of output characteristics can be improved. [Brief explanation of the drawing]

[0020] [Figure 1] Figure 1 is a block diagram showing the configuration of the output characteristic prediction system. [Figure 2] Figure 2 is a block diagram showing the vehicle configuration. [Figure 3] Figure 3 is a block diagram showing the functional elements of the output characteristic prediction device. [Figure 4] Figure 4 shows an example of vehicle data. [Figure 5] Figure 5 shows the process executed by the model generation unit. [Figure 6] Figure 6 is a diagram illustrating the learning period and the specified period. [Figure 7] Figure 7 is a flowchart showing the prediction process performed by the prediction unit. [Figure 8] Figure 8 is a flowchart showing the processes executed by a specific unit. [Figure 9] Figure 9 shows an example of hierarchical clustering results. [Figure 10] Figure 10 shows an example of grouping results based on each variable. [Figure 11] Figure 11 shows the process executed by the model improvement unit. [Modes for carrying out the invention]

[0021] Hereinafter, one embodiment of the output characteristic prediction device 11, the output characteristic prediction method, and the output characteristic prediction system 100 will be described with reference to Figures 1 to 11. In the following description, the output characteristic prediction device 11 will simply be referred to as "prediction device 11".

[0022] As shown in Figure 1, the output characteristic prediction system 100 comprises multiple vehicles 10 and a prediction device 11. In this embodiment, the output characteristic prediction system 100 comprises five vehicles 10. When distinguishing between the five vehicles 10, they are designated as the first vehicle 10a, the second vehicle 10b, the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e. In this embodiment, each vehicle 10 is a forklift.

[0023] Each vehicle 10 and the prediction device 11 are connected via a network NW to enable the transmission and reception of information. The network NW includes, for example, the Internet, WAN (Wide Area Network), LAN (Local Area Network), provider terminals, wireless communication networks, wireless base stations, dedicated lines, etc.

[0024] <Vehicle 10> As shown in Figure 2, each vehicle 10 is equipped with a fuel cell system 20. The fuel cell system 20 includes a hydrogen tank 21, an air compressor 22, a fuel cell 23 acting as a battery, and an energy storage device 24.

[0025] The hydrogen tank 21 stores hydrogen gas. The air compressor 22 compresses air. The fuel cell 23 is a stack of multiple battery cells. The battery cells are, for example, solid molecular fuel cells. The fuel cell 23 generates electricity through an electrochemical reaction between hydrogen supplied from the hydrogen tank 21 and oxygen in the compressed air supplied from the air compressor 22. The fuel cell 23 emits anode-off gas and cathode-off gas. The anode-off gas contains unreacted hydrogen in the fuel cell 23 and generated water produced when hydrogen and oxygen react in the fuel cell 23. The cathode-off gas contains air containing unreacted oxygen in the fuel cell 23 and generated water produced when hydrogen and oxygen react in the fuel cell 23.

[0026] The energy storage device 24 is a secondary battery or a capacitor. The charging and discharging of the energy storage device 24 is performed according to the relationship between the output power of the fuel cell 23 and the power required by the load 41, which will be described later. If the output power of the fuel cell 23 exceeds the power required by the load 41, the surplus power is charged into the energy storage device 24. If the output power of the fuel cell 23 is less than the power required by the load 41, the deficit is discharged from the energy storage device 24.

[0027] The fuel cell system 20 includes a first pressure detection unit 31, a second pressure detection unit 32, a flow rate detection unit 33, a current detection unit 34, a voltage detection unit 35, a temperature detection unit 36, and a charge level detection unit 37. The flow rate detection unit 33 and the temperature detection unit 36 ​​may be composed of a single detection unit.

[0028] The first pressure detection unit 31 detects the pressure of the hydrogen supplied to the fuel cell 23. The second pressure detection unit 32 detects the pressure of the anode off-gas discharged from the fuel cell 23. The flow rate detection unit 33 detects the amount of air supplied to the fuel cell 23 (hereinafter referred to as "supplied air amount"). The current detection unit 34 detects the output current of the fuel cell 23. The voltage detection unit 35 detects the output voltage of the fuel cell 23. The temperature detection unit 36 ​​detects the ambient temperature around the vehicle 10. The charge level detection unit 37 detects the charge level of the energy storage device 24. These various detection units perform detections at predetermined detection intervals. The detection interval may be the same for all detection units, or it may differ for each detection unit.

[0029] The fuel cell system 20 includes an FC control device 25. The FC control device 25 comprises a processor 25a and a memory unit 25b. Examples of the processor 25a include a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and a DSP (Digital Signal Processor). The memory unit 25b includes RAM (Random access memory), ROM (Read Only Memory), and rewritable non-volatile memory. Examples of non-volatile memory include EEPROM (Electrically Erasable Programmable Read-Only Memory) and flash memory. The memory unit 25b stores program code or instructions configured to cause the processor to execute processing. The memory unit 25b, i.e., the computer-readable medium, includes any available medium that can be accessed by a general-purpose or dedicated computer. The FC control device 25 may be composed of hardware circuits such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). The FC control device 25, which is a processing circuit, may include one or more processors that operate according to a computer program, one or more hardware circuits such as ASICs or FPGAs, or a combination thereof.

[0030] The FC control device 25 acquires the detection results from various detection units. In this embodiment, the FC control device 25 acquires the detection results from the first pressure detection unit 31, the second pressure detection unit 32, the flow rate detection unit 33, the current detection unit 34, the voltage detection unit 35, the temperature detection unit 36, and the charge level detection unit 37. The FC control device 25 detects the difference between the pressure of the hydrogen supplied to the fuel cell 23 and the pressure of the anode off gas discharged from the fuel cell 23 (hereinafter referred to as the "hydrogen pressure difference"). The FC control device 25 detects the output power of the fuel cell 23 from the output current and output voltage of the fuel cell 23.

[0031] The FC control device 25 controls the power generation of the fuel cell 23. The output power of the fuel cell 23 changes depending on the amount of hydrogen supplied to the fuel cell 23 and the amount of oxygen supplied to the fuel cell 23. The FC control device 25 controls the power generation of the fuel cell 23 by controlling the amount of hydrogen supplied to the fuel cell 23 and the amount of oxygen supplied to the fuel cell 23.

[0032] The FC control device 25 of this embodiment can perform switching control to switch the output power of the fuel cell 23 in stages based on the charge rate of the energy storage device 24 detected by the charge rate detection unit 37. For example, the FC control device 25 switches the output power of the fuel cell 23 in four stages: 0 [kW], 3 [kW], 5 [kW], and 8 [kW].

[0033] When the charge level of the energy storage device 24 falls below a predetermined first threshold, the FC control device 25 switches the output power of the fuel cell 23 from 0 kW to 3 kW. When the charge level of the energy storage device 24 falls below a predetermined second threshold which is lower than the first threshold, the FC control device 25 switches the output power of the fuel cell 23 from 3 kW to 5 kW. When the charge level of the energy storage device 24 falls below a predetermined third threshold which is lower than the second threshold, the FC control device 25 switches the output power of the fuel cell 23 from 5 kW to 8 kW.

[0034] When the charge level of the energy storage device 24 reaches a predetermined fourth threshold or higher, the FC control device 25 switches the output power of the fuel cell 23 from 8 kW to 5 kW. When the charge level of the energy storage device 24 reaches a predetermined fifth threshold or higher, which is greater than the fourth threshold, the FC control device 25 switches the output power of the fuel cell 23 from 5 kW to 3 kW. When the charge level of the energy storage device 24 reaches a predetermined sixth threshold or higher, which is greater than the fifth threshold, the FC control device 25 switches the output power of the fuel cell 23 from 3 kW to 0 kW.

[0035] The vehicle 10 includes a load 41, a key switch 42, an operating unit 43, and a vehicle control device 44. The load 41 is driven by electricity supplied from the fuel cell 23 or the energy storage device 24. In this embodiment, the load 41 includes a cargo handling motor that operates the vehicle 10 for loading and unloading, and a driving motor that operates the vehicle 10 for driving.

[0036] The key switch 42 can be switched on or off by the user. When the key switch 42 is switched from off to on, the operation of the vehicle 10 begins. When the key switch 42 is switched from on to off, the operation of the vehicle 10 stops.

[0037] The control unit 43 is operated by the user of the vehicle 10. The control unit 43 includes a cargo handling lever to operate the vehicle 10 in cargo handling operations and an accelerator to operate the vehicle 10 in driving operations. The vehicle control device 44 comprises a processor 44a and a storage unit 44b. Examples of the processor 44a include a CPU, GPU, and DSP. The storage unit 44b includes RAM, ROM, and rewritable non-volatile memory. Examples of non-volatile memory include EEPROM and flash memory. The storage unit 44b stores program code or instructions configured to cause the processor to perform processing. The storage unit 44b, i.e., the computer-readable medium, includes any available medium accessible by a general-purpose or dedicated computer. The vehicle control device 44 may be composed of hardware circuits such as ASICs and FPGAs. The vehicle control device 44, which is a processing circuit, may include one or more processors that operate according to a computer program, one or more hardware circuits such as ASICs and FPGAs, or a combination thereof.

[0038] The vehicle control device 44 can detect the switching of the key switch 42. The vehicle control device 44 can also detect the idling state of the vehicle 10. The idling state refers to a state in which the operation of the vehicle 10 has started, and the vehicle 10 is not performing any driving or loading / unloading operations. The vehicle control device 44 detects that the vehicle 10 is in the idling state when the key switch 42 is in the ON state and the operation unit 43 is not being operated.

[0039] As shown in Figure 1, each vehicle 10 has a communication unit 45. The communication unit 45 sends and receives various information to and from the prediction device 11 via a network NW. The communication unit 45 transmits various data to the prediction device 11. In this embodiment, the communication unit 45 transmits to the prediction device 11 data on the output voltage of the fuel cell 23, data on the hydrogen pressure difference, data on the amount of supplied air, data on the switching of the output power of the fuel cell 23, data on the switching of the key switch 42, data on the ambient temperature around the vehicle 10, data on the idling of the vehicle 10, and data on the output power of the fuel cell 23.

[0040] The output voltage data of the fuel cell 23 includes the detected value and detection time of the output voltage of the fuel cell 23. The hydrogen pressure difference data includes the detected value and detection time of the hydrogen pressure difference. The supply air amount data includes the detected value and detection time of the supply air amount. The data related to the switching of the output power of the fuel cell 23 includes the output power switching time. The data related to the switching of the key switch 42 includes the switching time of the key switch 42. The ambient temperature data around the vehicle 10 includes the detected value and detection time of the ambient temperature. The data related to the idling of the vehicle 10 includes the start time and end time of the idling state. The output power data of the fuel cell 23 includes the detected value and detection time of the output power of the fuel cell 23.

[0041] <Prediction device 11> The prediction device 11 predicts the output characteristics of the battery for each vehicle 10. The prediction device 11 predicts the output characteristics of the battery for a period beyond the prediction time. In this embodiment, the prediction device 11 predicts the output voltage of the fuel cell 23 as the output characteristics of the battery. The prediction device 11 includes a control unit 12 and a communication unit 13.

[0042] The communication unit 13 transmits and receives various information to and from each vehicle 10 via the network NW. The communication unit 13 receives various data from each vehicle 10. In this embodiment, the communication unit 13 receives from each vehicle 10 the output voltage data of the fuel cell 23, the hydrogen pressure difference data, the amount of supplied air data, data related to the switching of the output power of the fuel cell 23, data related to the switching of the key switch 42, the ambient temperature around the vehicle 10, data related to the idling of the vehicle 10, and the output power data of the fuel cell 23.

[0043] The control unit 12 comprises a processor 12a and a storage unit 12b. Examples of the processor 12a include a CPU, GPU, and DSP. The storage unit 12b includes RAM, ROM, and rewritable non-volatile memory. Examples of non-volatile memory include EEPROM and flash memory. The storage unit 12b stores program code or instructions configured to cause the processor to perform processing. The storage unit 12b, i.e., the computer-readable medium, includes any available medium accessible by a general-purpose or dedicated computer. The control unit 12 may be composed of hardware circuits such as ASICs or FPGAs. The control unit 12, which is a processing circuit, may include one or more processors operating according to a computer program, one or more hardware circuits such as ASICs or FPGAs, or a combination thereof.

[0044] As shown in Figure 3, the control unit 12 includes an acquisition unit 14, a model generation unit 15, and a prediction unit 16. The acquisition unit 14, the model generation unit 15, and the prediction unit 16 are functional elements that operate when the processor 12a executes a program stored in the storage unit 12b.

[0045] <Acquisition part 14> The acquisition unit 14 individually acquires vehicle data D, including time-series data, from multiple vehicles 10. The acquisition unit 14 acquires first vehicle data Da from the first vehicle 10a. The acquisition unit 14 acquires second vehicle data Db from the second vehicle 10b. The acquisition unit 14 acquires third vehicle data Dc from the third vehicle 10c. The acquisition unit 14 acquires fourth vehicle data Dd from the fourth vehicle 10d. The acquisition unit 14 acquires fifth vehicle data De from the fifth vehicle 10e.

[0046] As shown in Figure 4, the vehicle data D of this embodiment includes time-series data of eight variables. Specifically, the vehicle data D includes time-series data of the output voltage Y1 of the fuel cell 23, time-series data of the hydrogen pressure difference Y2, time-series data of the supplied air amount Y3, time-series data of the number of times the output power of the fuel cell 23 is switched X1, time-series data of the number of times the vehicle 10 is started and stopped X2, time-series data of the ambient temperature X3 around the vehicle 10, time-series data of the number of times the vehicle 10 is idled X4, and time-series data of the output power X5 of the fuel cell 23.

[0047] The time-series data of the output voltage Y1 of the fuel cell 23, the time-series data of the hydrogen pressure difference Y2, and the time-series data of the supplied air amount Y3 are time-series data related to the power generation control of the fuel cell 23. The time-series data of the number of times the output power of the fuel cell 23 is switched X1, the time-series data of the number of times the vehicle 10 is started and stopped X2, the time-series data of the ambient temperature X3 around the vehicle 10, the time-series data of the number of times the vehicle 10 is idled X4, and the time-series data of the output power X5 of the fuel cell 23 are time-series data related to the usage status of the vehicle 10.

[0048] In this embodiment, the acquisition unit 14 calculates the median value of the output voltage of the fuel cell 23 over a predetermined calculation period from the output voltage data of the fuel cell 23. The predetermined calculation period is, for example, one week. In this embodiment, the median value of the output voltage of the fuel cell 23 over the predetermined calculation period is defined as "output voltage Y1 of the fuel cell 23". Therefore, the value at each point in time in the time-series data of the output voltage Y1 of the fuel cell 23 is the median value of the output voltage of the fuel cell 23 over the predetermined calculation period.

[0049] The acquisition unit 14 calculates the median value of the hydrogen pressure difference over a predetermined calculation period from the hydrogen pressure difference data. In this embodiment, the median value of the hydrogen pressure difference over a predetermined calculation period is defined as "hydrogen pressure difference Y2". Therefore, the value at each point in the time-series data of hydrogen pressure difference Y2 is the median value of the hydrogen pressure difference over the predetermined calculation period.

[0050] The acquisition unit 14 calculates the median value of the supplied air quantity over a predetermined calculation period from the supplied air quantity data. In this embodiment, the median value of the supplied air quantity over a predetermined calculation period is defined as "supplied air quantity Y3". Therefore, the value at each point in time in the time-series data of supplied air quantity Y3 is the median value of the supplied air quantity over the predetermined calculation period.

[0051] The acquisition unit 14 calculates the cumulative number of times the fuel cell 23's output power is switched during a predetermined calculation period from data relating to the switching of the fuel cell 23's output power. In this embodiment, the cumulative number of times the fuel cell 23's output power is switched during a predetermined calculation period is defined as "Fuel cell 23 output power switching count X1". Therefore, the value at each point in time in the time-series data of fuel cell 23 output power switching count X1 is the cumulative number of times the fuel cell 23's output power is switched during a predetermined calculation period.

[0052] The acquisition unit 14 calculates the cumulative number of times the vehicle 10 has been started and stopped during a predetermined calculation period from the data related to the switching of the key switch 42. In this embodiment, the cumulative number of times the vehicle 10 has been started and stopped during a predetermined calculation period is defined as "Number of times the vehicle 10 has been started and stopped X2". Therefore, the value at each point in time in the time-series data of Number of times the vehicle 10 has been started and stopped X2 is the cumulative number of times the vehicle 10 has been started and stopped during a predetermined calculation period.

[0053] The acquisition unit 14 calculates the average value of the ambient temperature around the vehicle 10 over a predetermined calculation period from the ambient temperature data around the vehicle 10. In this embodiment, the average value of the ambient temperature around the vehicle 10 over a predetermined calculation period is defined as "Ambient Temperature X3 around Vehicle 10". Therefore, the value at each point in time in the time-series data of Ambient Temperature X3 around Vehicle 10 is the average value of the ambient temperature around the vehicle 10 over a predetermined calculation period.

[0054] The acquisition unit 14 calculates the cumulative number of idling cycles of the vehicle 10 over a predetermined calculation period from data related to the vehicle 10's idling. In this embodiment, the cumulative number of idling cycles of the vehicle 10 over a predetermined calculation period is defined as "Vehicle 10 Idling Cycle X4". Therefore, the value at each point in time in the time-series data of Idling Cycle X4 is the cumulative number of idling cycles over a predetermined calculation period.

[0055] The acquisition unit 14 calculates the average value of the fuel cell 23's output power over a predetermined calculation period from the data of the fuel cell 23's output power. In this embodiment, the average value of the fuel cell 23's output power over a predetermined calculation period is defined as "Fuel cell 23 output power X5". Therefore, the value at each point in time in the time-series data of fuel cell 23 output power X5 is the average value of the fuel cell 23's output power over a predetermined calculation period.

[0056] As described above, the FC control device 25 of this embodiment is capable of performing switching control to switch the output power of the fuel cell 23 in stages based on the charge rate of the energy storage device 24. The vehicle data D of this embodiment includes time-series data of the output voltage Y1 of the fuel cell 23 during the execution of the switching control, and time-series data of the output power X5 of the fuel cell 23 during the execution of the switching control.

[0057] The acquisition unit 14 acquires vehicle data D from each vehicle 10 for a predetermined learning period α. The predetermined learning period α is, for example, several years. The acquisition unit 14 continues to acquire vehicle data D from each vehicle 10 even after acquiring vehicle data D for the learning period α.

[0058] <Model generation unit 15> The model generation unit 15 generates a regression model M1 for each vehicle 10 by performing machine learning using time series data for a learning period α minutes for each vehicle 10. In this embodiment, the model generation unit 15 generates the regression model M1 by performing machine learning using time series data of eight variables for a learning period α minutes.

[0059] As shown in Figure 5, the model generation unit 15 generates a first regression model M1a for the first vehicle 10a by performing machine learning using time series data of the first vehicle data Da for a learning period of α minutes. The model generation unit 15 generates a second regression model M1b for the second vehicle 10b by performing machine learning using time series data of the second vehicle data Db for a learning period of α minutes. The model generation unit 15 generates a third regression model M1c for the third vehicle 10c by performing machine learning using time series data of the third vehicle data Dc for a learning period of α minutes. The model generation unit 15 generates a fourth regression model M1d for the fourth vehicle 10d by performing machine learning using time series data of the fourth vehicle data Dd for a learning period of α minutes. The model generation unit 15 generates a fifth regression model M1e for the fifth vehicle 10e by performing machine learning using time series data of the fifth vehicle data De for a learning period of α minutes.

[0060] The regression model M1 outputs a dependent variable when an explanatory variable is input. The dependent variable includes the output voltage of the fuel cell 23 as an output characteristic of the battery. In other words, the regression model M1 is a predictive model that predicts the output voltage of the fuel cell 23 as an output characteristic of the battery.

[0061] In this embodiment, there are three objective variables. Specifically, the objective variables are the output voltage Y1 of the fuel cell 23, the hydrogen pressure difference Y2, and the amount of supplied air Y3. In this embodiment, there are eight explanatory variables. Specifically, the explanatory variables are the output voltage Y1 of the fuel cell 23, the hydrogen pressure difference Y2, the amount of supplied air Y3, the number of times the output power of the fuel cell 23 is switched X1, the number of times the vehicle 10 is started and stopped X2, the ambient temperature around the vehicle 10 X3, the number of times the vehicle 10 idles X4, and the output power of the fuel cell 23 X5.

[0062] The output voltage Y1, hydrogen pressure difference Y2, and supplied air amount Y3 of the fuel cell 23 are endogenous variables. The number of times the output power of the fuel cell 23 is switched X1, the number of times the vehicle 10 is started and stopped X2, the ambient temperature around the vehicle 10 X3, the number of times the vehicle 10 idles X4, and the output power of the fuel cell 23 X5 are exogenous variables.

[0063] The regression model M1 of this embodiment is represented by a vector auto-regression (VAR) model shown in Equation (1).

[0064]

Number

[0065] Y t-1 vector = (Y1 t-1 , Y2 t-1 , Y3 t-1 ). y1 t-1 is the output voltage of the fuel cell 23 at time point t-1, which is one time point before time point t. y2 t-1 is the hydrogen pressure difference at time point t-1. y3 t-1 is the supplied air volume at time point t-1. W 内 is the regression coefficient of the Y t-1 vector. In this embodiment, W 内 is a 3×3 matrix.

[0066] X t vector = (X1 t , X2 t , X3 t , X4 t , X5 t ). x1 t is the number of switching times of the output power of the fuel cell 23 at time point t. x2 t is the number of start times and stop times of the operation of the vehicle 10 at time point t. x3 t is the outside air temperature around the vehicle 10 at time point t. x4 t is the number of idling times of the vehicle 10 at time point t. x5 tThis is the output power of the fuel cell 23 at time t. 外 X t These are the regression coefficients of the vector. In this embodiment, W 外 This is a 3x5 matrix.

[0067] The vector C is (C1, C2, C3). C1, C2, and C3 are all constants. ε t Vector = (ε1 t ,ε2 t ,ε3 t ) is. ε1 t ,ε2 t ,ε3 t These represent noise at time t.

[0068] Y t-1 Vector regression coefficient W 内、 and X t Vector regression coefficient W 外 Let W be the regression coefficient consisting of the following. In this embodiment, the regression coefficient W is a 3 × 8 matrix. The regression coefficient W is estimated, for example, by the ridge regression shown in equation (2).

[0069]

number

[0070] The model generation unit 15 generates a regression model M1 by estimating the regression coefficient W. The model generation unit 15 stores the generated regression model M1 for each vehicle 10 in the storage unit 12b.

[0071] <Prediction section 16> The prediction unit 16 performs a prediction process to predict the output voltage of the fuel cell 23 for each vehicle 10. The prediction unit 16 repeatedly performs the prediction process. The prediction unit 16 performs the prediction process for the first time when the model generation unit 15 generates the regression model M1. In addition, the prediction unit 16 performs the prediction process each time the acquisition unit 14 acquires new vehicle data D for a predetermined period β from multiple vehicles 10. The predetermined period β is, for example, several months.

[0072] As shown in Figure 6, for example, the learning period α is the period from time t1 to time ta. In this case, when the acquisition unit 14 acquires vehicle data D from time ta to time tb after a predetermined period β has elapsed, the prediction unit 16 executes prediction processing. When the acquisition unit 14 acquires vehicle data D from time tb to time tc after a predetermined period β has elapsed, the prediction unit 16 executes prediction processing.

[0073] The prediction process performed by the prediction unit 16 is described in detail below. In the following description, one of the multiple vehicles 10 is designated as the target vehicle T, and the other vehicles 10 are designated as source vehicles S. Note that the target vehicle T does not refer to only one specific vehicle 10 (for example, only the first vehicle 10a). Each process in the prediction process is executed by replacing the target vehicle T one by one among the multiple vehicles 10.

[0074] As shown in Figure 3, the prediction unit 16 includes a specific unit 16a, a model improvement unit 16b, a calculation unit 16c, a setting unit 16d, and an extrapolation unit 16e. As shown in Figure 7, the identification unit 16a identifies one or more source vehicles S similar to the target vehicle T based on vehicle data D of multiple vehicles 10 over a predetermined period β (step S10). In this embodiment, the identification unit 16a identifies source vehicles S similar to the target vehicle T based on time-series data of eight variables over a predetermined period β of multiple vehicles 10. In the first prediction process, the identification unit 16a identifies source vehicles S similar to the target vehicle T based on time-series data of eight variables over a learning period α of multiple vehicles 10.

[0075] As shown in Figure 8, the identification unit 16a outputs a similarity matrix (step S11) by calculating the similarity of time series data between two vehicles 10 for all combinations of vehicles 10 for each variable. In this embodiment, the identification unit 16a calculates the similarity of time series data between two vehicles 10 using Dynamic Time Warping (DTW). In this embodiment, there are 5 vehicles 10. Also, there are 8 variables in this embodiment. Therefore, in this embodiment, the identification unit 16a outputs 8 5x5 similarity matrices.

[0076] The identification unit 16a performs hierarchical clustering for each variable (step S12). The memory unit 12b of the prediction device 11 stores a trained model M2 that has been trained to output clustering results when a similarity matrix is ​​input (see Figure 1). When the identification unit 16a inputs the similarity matrix to the trained model M2, the trained model M2 outputs clustering results for multiple vehicles 10.

[0077] Figure 9 shows an example of clustering results for five vehicles 10. Two vehicles, the first vehicle 10a and the second vehicle 10b, and three vehicles, the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e, are branched at the highest level.

[0078] As shown in Figure 8, the identification unit 16a groups the multiple vehicles 10 into a predetermined number of groups based on the clustering results for each variable (step S13). In this embodiment, the predetermined number of groups is set to 2 for each variable. Therefore, the identification unit 16a groups the 5 vehicles 10 into 2 groups. Specifically, the identification unit 16a groups the 5 vehicles 10 into group 1 or group 2 based on the clustering results of the highest level.

[0079] In the example shown in Figure 9, the specific unit 16a groups the first vehicle 10a and the second vehicle 10b into group 1. The specific unit 16a groups the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e into group 2.

[0080] As shown in Figure 8, the identification unit 16a identifies source vehicles S similar to the target vehicle T based on the grouping results based on each explanatory variable (step S14). Specifically, the identification unit 16a identifies source vehicles S whose combination of grouping results based on each explanatory variable matches that of the target vehicle T as source vehicles S similar to the target vehicle T.

[0081] In the example shown in Figure 10, the grouping results based on the variables Y1, Y2, Y3, X1, X2, X3, X4, and X5 for the first vehicle 10a and the second vehicle 10b are groups 1, 1, 1, 1, 1, 1, 1, and 1, respectively. That is, the combinations of grouping results based on each explanatory variable are the same for the first vehicle 10a and the second vehicle 10b. Therefore, the first vehicle 10a and the second vehicle 10b are similar.

[0082] In the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e, the grouping results based on the variables Y1, Y2, Y3, X1, X2, X3, X4, and X5 are groups 2, 2, 2, 2, 2, 2, and 2, respectively. That is, the combinations of grouping results based on each variable are the same in the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e. Therefore, the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e are similar.

[0083] For example, if the target vehicle T is the first vehicle 10a, and the source vehicles S are the second vehicle 10b, the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e, the identification unit 16a identifies the second vehicle 10b as a source vehicle S similar to the target vehicle T.

[0084] For example, if the target vehicle T is the third vehicle 10c and the source vehicles S are the first vehicle 10a, the second vehicle 10b, the fourth vehicle 10d, and the fifth vehicle 10e, the identification unit 16a identifies the fourth vehicle 10d and the fifth vehicle 10e as source vehicles S similar to the target vehicle T.

[0085] As shown in Figures 7 and 11, the model improvement unit 16b improves the regression model M1 of the target vehicle T by performing transfer learning using the regression coefficients W of the regression model M1 of one or more source vehicles S that are similar to the target vehicle T (step S20).

[0086] The regression coefficients W of the regression model M1 of the target vehicle T estimated by the model generation unit 15 are used to obtain the regression coefficients W of the target vehicle T before transfer learning. T The regression coefficients W of the regression model M1 of the source vehicle Sj estimated by the model generation unit 15 are set to the regression coefficients W of the source vehicle Sj before transfer learning. Sj The model improvement unit 16b calculates the regression coefficients W of the target vehicle T before transfer learning. T And the regression coefficients W before transfer learning for source vehicle Sj which is similar to target vehicle T. Sj By using transfer learning, the regression coefficients W of the target vehicle T after transfer learning can be obtained. T fix We estimate this.

[0087] In this embodiment, the regression coefficient W after transfer learning of the target vehicle T T fix This is estimated by the equation shown in equation (3).

[0088]

number

[0089] ρ(T,Sj) is a coefficient determined based on whether the target vehicle T and the source vehicle Sj are similar or not. If the target vehicle T and the source vehicle Sj are similar, ρ(T,Sj) = 1. If the target vehicle T and the source vehicle Sj are not similar, ρ(T,Sj) = 0. In other words, if the target vehicle T and the source vehicle Sj are similar, the transfer learning term ≠ 0, so transfer learning is performed. If the target vehicle T and the source vehicle Sj are not similar, the transfer learning term = 0, so transfer learning is not performed.

[0090] The transfer learning term determines the regression coefficient W of the target vehicle T before transfer learning. T And the regression coefficients W before transfer learning for source vehicle Sj which is similar to target vehicle T. Sj The regression coefficients W after transfer learning of the target vehicle T are set to average out. T fix It is estimated that...

[0091] For example, the target vehicle T is the first vehicle 10a, and the source vehicles S1, S2, S3, and S4 are the second vehicle 10b, the third vehicle 10c, the fourth vehicle 10d, and the fifth vehicle 10e, respectively. In this case, the second vehicle 10b is similar to the first vehicle 10a, so ρ(T,S1)=1. The third vehicle 10c is not similar to the first vehicle 10a, so ρ(T,S2)=0. The fourth vehicle 10d is not similar to the first vehicle 10a, so ρ(T,S3)=0. The fifth vehicle 10e is not similar to the first vehicle 10a, so ρ(T,S4)=0. Therefore, the W of the first vehicle 10a after transfer learning is T fix This is W before transfer learning of the second vehicle 10b. S1 It is estimated by transfer learning using [this method].

[0092] For example, the target vehicle T is the third vehicle 10c, and the source vehicles S1, S2, S3, and S4 are the first vehicle 10a, the second vehicle 10b, the fourth vehicle 10d, and the fifth vehicle 10e, respectively. In this case, the first vehicle 10a is similar to the third vehicle 10c, so ρ(T,S1)=0. The second vehicle 10b is not similar to the third vehicle 10c, so ρ(T,S2)=0. The fourth vehicle 10d is similar to the third vehicle 10c, so ρ(T,S3)=1. The fifth vehicle 10e is similar to the third vehicle 10c, so ρ(T,S4)=1. Therefore, the W of the third vehicle 10c after transfer learning is T fix This is W before transfer learning of the 4th vehicle 10d. S3 And the W of the 5th vehicle 10e before transfer learning S4 It is estimated by using transfer learning.

[0093] The model improvement unit 16b uses the regression coefficients W in the regression model M1 of the target vehicle T, and the regression coefficients W before transfer learning. T Therefore, the regression coefficient W after transfer learning T fix By updating, the regression model M1 of the target vehicle T is improved. The model improvement unit 16b stores the improved regression model M1 of each vehicle 10 in the storage unit 12b.

[0094] The calculation unit 16c calculates a prediction accuracy index for the improved regression model M1 of the target vehicle T by interpolating the output voltage of the fuel cell 23 of the target vehicle T during a predetermined period β using the improved regression model M1 of the target vehicle T (step S30). In the first prediction process, the calculation unit 16c calculates a prediction accuracy index for the improved regression model M1 of the target vehicle T by interpolating the output voltage of the fuel cell 23 of the target vehicle T during the learning period α using the improved regression model M1 of the target vehicle T.

[0095] More specifically, the calculation unit 16c interpolates the output voltage of the fuel cell 23 of the target vehicle T over a predetermined period β using the improved regression model M1 of the target vehicle T. The calculation unit 16c obtains the measured value of the output voltage of the fuel cell 23 of the target vehicle T over a predetermined period β from the time-series data of the target vehicle T over a predetermined period β. The calculation unit 16c calculates a prediction accuracy index for the improved regression model M1 of the target vehicle T from the obtained measured value of the output voltage of the fuel cell 23 of the target vehicle T over a predetermined period β and the interpolated predicted value of the output voltage of the fuel cell 23 of the target vehicle T over a predetermined period β. As a prediction accuracy index, the calculation unit 16c calculates, for example, the mean squared error (MSE) or the root mean squared error (RMSE). In this case, the smaller the value of the prediction accuracy index, the better the prediction accuracy.

[0096] The setting unit 16d sets a source vehicle S that has a high similarity to the target vehicle T based on the prediction accuracy index calculated by the calculation unit 16c. The setting unit 16d determines whether the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process (step S40). If the value of the prediction accuracy index in the current prediction process is less than the value of the prediction accuracy index in the previous prediction process, the setting unit 16d determines that the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process. If the value of the prediction accuracy index in the current prediction process is the same as the value of the prediction accuracy index in the previous prediction process, the setting unit 16d determines that the prediction accuracy in the current prediction process is approximately the same as the prediction accuracy in the previous prediction process. If the value of the prediction accuracy index in the current prediction process is greater than the value of the prediction accuracy index in the previous prediction process, the setting unit 16d determines that the prediction accuracy in the current prediction process is worse than the prediction accuracy in the previous prediction process.

[0097] If the prediction accuracy in this prediction process is better than that in the previous prediction process, then one or more source vehicles S identified as similar to target vehicle T in this prediction process have a higher degree of similarity to target vehicle T than one or more source vehicles S identified as similar to target vehicle T in the previous prediction process. If the prediction accuracy in this prediction process is the same as that in the previous prediction process, then one or more source vehicles S identified as similar to target vehicle T in this prediction process and one or more source vehicles S identified as similar to target vehicle T in the previous prediction process have a similar degree of similarity to target vehicle T. If the prediction accuracy in this prediction process is worse than that in the previous prediction process, then one or more source vehicles S identified as similar to target vehicle T in this prediction process have a lower degree of similarity to target vehicle T than one or more source vehicles S identified as similar to target vehicle T in the previous prediction process.

[0098] Therefore, the setting unit 16d sets a source vehicle Sj that has a high similarity to the target vehicle T, based on whether the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process.

[0099] If the setting unit 16d determines that the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process (YES in step S40), it executes the process in step S50. In step S50, the setting unit 16d updates the source vehicle S with a high similarity to the target vehicle T from the source vehicle S set in the previous prediction process to the source vehicle S identified as similar to the target vehicle T in the current prediction process.

[0100] If the setting unit 16d determines that the prediction accuracy in the current prediction process is the same as, or worse than, the prediction accuracy in the previous prediction process (NO in step S40), it proceeds to the process in step S60. In other words, the setting unit 16d does not update the source vehicle S with high similarity to the target vehicle T from the source vehicle S set in the previous prediction process to the source vehicle S identified as similar to the target vehicle T in the current prediction process. Therefore, the source vehicle S with high similarity to the target vehicle T remains unchanged from the source vehicle S set in the previous prediction process.

[0101] In the first prediction process, since there is no prediction accuracy metric from the previous prediction process, the setting unit 16d cannot compare the prediction accuracy from the previous prediction process with the prediction accuracy from the current prediction process. In this case, the setting unit 16d sets the source vehicle S, which has a high similarity to the target vehicle T, as the source vehicle S identified as similar to the target vehicle T in the first prediction process.

[0102] In step S60, the extrapolation unit 16e extrapolates the output voltage of the fuel cell 23. The extrapolation unit 16e predicts the output voltage of the fuel cell 23 of the target vehicle T using an improved regression model M1 of the target vehicle T, which is obtained by transfer learning using the regression coefficients W of the regression model M1 of one or more source vehicles S that have a high similarity to the target vehicle T.

[0103] The extrapolation unit 16e inputs explanatory variables into the improved regression model M1 of the target vehicle T, causing the improved regression model M1 of the target vehicle T to output the dependent variable. As described above, the dependent variable output by the regression model M1 includes the output voltage of the fuel cell 23. Thus, the extrapolation unit 16e extrapolates the output voltage of the fuel cell 23 of the target vehicle T.

[0104] For example, suppose the target vehicle T is the first vehicle 10a, and the source vehicle S with a high similarity to the target vehicle T is the second vehicle 10b. In this case, the extrapolation unit 16e predicts the output voltage of the fuel cell 23 of the first vehicle 10a using the first regression model M1a, which has been improved by transfer learning using the regression coefficient W of the second regression model M1b.

[0105] [Operation and Effects of This Embodiment] The operation and effects of this embodiment will now be explained. (1) The prediction device 11 predicts the output voltage of the fuel cell 23 installed in the vehicle 10. The prediction device 11 comprises an acquisition unit 14, a model generation unit 15, and a prediction unit 16. The acquisition unit 14 individually acquires vehicle data D, including time series data, from multiple vehicles 10. The model generation unit 15 generates a regression model M1 for each vehicle 10 by machine learning using time series data for a learning period α minutes for each vehicle 10, which outputs a target variable including the output voltage of the fuel cell 23 when an explanatory variable is input.

[0106] The prediction unit 16 performs a prediction process to predict the output voltage of the fuel cell 23 for each vehicle 10. The prediction unit 16 performs the prediction process each time the acquisition unit 14 acquires new vehicle data D for a predetermined period β from multiple vehicles 10. The prediction unit 16 includes a specification unit 16a, a model improvement unit 16b, a calculation unit 16c, a setting unit 16d, and an extrapolation unit 16e.

[0107] The identification unit 16a identifies one or more source vehicles S similar to the target vehicle T based on vehicle data D of multiple vehicles 10 for a predetermined period β. The model improvement unit 16b improves the regression model M1 of the target vehicle T by performing transfer learning using the regression coefficients W of the regression model M1 of one or more source vehicles S similar to the target vehicle T. The calculation unit 16c calculates a prediction accuracy index of the improved regression model M1 of the target vehicle T by interpolating the output voltage of the fuel cell 23 of the target vehicle T for a predetermined period β using the improved regression model M1 of the target vehicle T.

[0108] The setting unit 16d sets one or more source vehicles S that have a high similarity to the target vehicle T based on the prediction accuracy index. If the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process, the setting unit 16d updates the source vehicles S that have a high similarity to the target vehicle T from the source vehicles S set in the previous prediction process to the source vehicles S that have been identified as similar to the target vehicle T in the current prediction process.

[0109] The extrapolation unit 16e extrapolates the output voltage of the fuel cell 23 of the target vehicle T using a regression model M1 of the target vehicle T that has been improved by transfer learning using the regression coefficients W of the regression model M1 of one or more source vehicles S that have a high similarity to the target vehicle T.

[0110] In this configuration, each time the acquisition unit 14 acquires new vehicle data D for a predetermined period β from multiple vehicles 10, the prediction unit 16 performs a prediction process. In the prediction process, the identification unit 16a identifies source vehicles S similar to the target vehicle T based on the vehicle data D for a predetermined period β from multiple vehicles 10. Therefore, it can handle changes in the combination of similar vehicles 10 as time passes.

[0111] Furthermore, with this configuration, the setting unit 16d determines a source vehicle S with a high similarity to the target vehicle T based on a comparison of the prediction accuracy in the current prediction process and the prediction accuracy in the previous prediction process. Therefore, a source vehicle S that improves prediction accuracy can be selected as a source vehicle S with a high similarity to the target vehicle T.

[0112] The extrapolation unit 16e extrapolates the output voltage of the fuel cell 23 of the target vehicle T using an improved regression model M1 of the target vehicle T, which is obtained by transfer learning using the regression coefficients W of the regression model M1 of the source vehicle S that has a high similarity to the target vehicle T. Therefore, the prediction accuracy of the output voltage of the fuel cell 23 is improved.

[0113] (2) Vehicle data D includes, as time-series data, time-series data relating to the power generation control of the fuel cell 23 and time-series data relating to the usage status of the vehicle 10. This configuration allows for the efficient generation of the regression model M1 and the identification of source vehicles S similar to the target vehicle T.

[0114] (3) The value at each point in time series data is the mean, median, or cumulative value of the values ​​over a predetermined calculation period. This configuration reduces the amount of time-series data, thereby reducing the processing load involved in generating the regression model M1 and identifying source vehicles S similar to the target vehicle T. Furthermore, if the value at each point in the time-series data is the median of the values ​​over a predetermined calculation period, noise can be effectively reduced.

[0115] (4) The vehicle 10 includes a fuel cell 23 as a battery, an energy storage device 24, a charge level detection unit 37, and an FC control device 25. The energy storage device 24 can charge the output power of the fuel cell 23. The charge level detection unit 37 detects the charge level of the energy storage device 24. The FC control device 25 controls the power generation of the fuel cell 23. The FC control device 25 can perform switching control to switch the output power of the fuel cell 23 in stages based on the charge level of the energy storage device 24 detected by the charge level detection unit 37. The time series data includes time series data of the output voltage Y1 of the fuel cell 23 during the execution of the switching control and time series data of the output power X5 of the fuel cell 23 during the execution of the switching control.

[0116] With this configuration, the output voltage and output power of the fuel cell 23 remain stable during switching control. By generating a regression model M1 using time-series data of the output voltage Y1 and output power X5 of the fuel cell 23 during switching control, the prediction accuracy of the output voltage of the fuel cell 23 can be improved.

[0117] (5) When calculating the similarity of time series data between two vehicles 10, the similarity may be calculated as low even if the time series data of the two vehicles 10 are similar if any of the following three patterns apply: The first pattern is when the time series data contains noise. The second pattern is when the time series data of one vehicle 10 is temporally out of sync with the time series data of the other vehicle 10. The third pattern is when the time series data of one vehicle 10 and the time series data of the other vehicle 10 are partially similar.

[0118] The identification unit 16a of this embodiment calculates the similarity of time series data between two vehicles 10 using a dynamic time stretching method. In this case, even if any of the three patterns described above apply, the similarity of time series data between the two vehicles 10 is low and difficult to calculate. Therefore, the accuracy of identifying a source vehicle S similar to the target vehicle T can be improved.

[0119] (6) The prediction device 11 of this embodiment can accurately predict the output voltage of the fuel cell 23. This makes it possible to accurately predict the lifespan of the fuel cell 23. The lifespan of the fuel cell 23 is the timing at which the output voltage of the fuel cell 23 falls below a predetermined output voltage threshold.

[0120] [Example of changes] Furthermore, each of the above embodiments can be implemented with the following modifications. The above embodiments and the following modifications can be combined with each other to the extent that they do not contradict each other technically.

[0121] ○ The number of vehicles 10 is not limited to 5 and may be changed as appropriate. The number of vehicles 10 may be, for example, several tens to several thousand. ○ Vehicle 10 is not limited to a forklift. Vehicle 10 may be an industrial vehicle other than a forklift, or a passenger car.

[0122] ○ The battery is not limited to the fuel cell 23. The battery may be, for example, a lithium-ion battery (LIB) or a lithium-air battery (LAB).

[0123] ○ The output characteristics of the fuel cell 23 predicted by the output characteristics prediction device 11 may also be the output power of the fuel cell 23. ○ If the dependent variable includes the output characteristics of the battery, the number and types of dependent variables may be changed as appropriate.

[0124] ○ The number and types of explanatory variables may be changed as appropriate. ○ The regression model M1 is not limited to a VAR model. The regression model M1 may be, for example, an autoregressive (AR) model.

[0125] ○ The method for estimating the regression coefficient W is not limited to ridge regression and may be modified as appropriate. ○ The method for identifying a source vehicle S similar to the target vehicle T is not limited to the method of the above embodiment and may be modified as appropriate.

[0126] ○ The vehicle data D used to identify a source vehicle S similar to the target vehicle T does not necessarily have to be time-series data. ○ The number of variables used to identify a source vehicle S similar to the target vehicle T is not limited to eight and may be changed as appropriate.

[0127] ○ The method for calculating the similarity of vehicle data D between two vehicles 10 is not limited to the dynamic time stretching method and may be modified as appropriate. ○ The number of groups into which the vehicles 10 are grouped is not limited to two; it may be three or more.

[0128] ○ The predetermined number of groups when grouping the vehicles 10 may differ for each variable. ○ Clustering methods are not limited to hierarchical clustering.

[0129] [Note] The technical concepts that can be understood from the above embodiments and modified examples are described below. <Note 1> An output characteristic prediction device for predicting the output characteristics of a battery installed in a vehicle, comprising: an acquisition unit that individually acquires vehicle data including time-series data from a plurality of vehicles; a model generation unit that generates a regression model for each vehicle that outputs a target variable including the output characteristics when an explanatory variable is input, by machine learning using the time-series data for a learning period for each vehicle; and a prediction unit that performs a prediction process to predict the output characteristics of the battery for each vehicle, wherein the prediction unit performs the prediction process each time the acquisition unit newly acquires vehicle data for a predetermined period from the plurality of vehicles, and when one of the plurality of vehicles is designated as a target vehicle and the other vehicles are designated as source vehicles, the prediction unit includes an identification unit that identifies one or more source vehicles similar to the target vehicle based on the vehicle data for the predetermined period from the plurality of vehicles, and performs transfer learning using the regression coefficients of the regression model of one or more source vehicles similar to the target vehicle, thereby predicting the regression of the target vehicle An output characteristic prediction device comprising: a model improvement unit that improves the model; a calculation unit that calculates a prediction accuracy index for the improved regression model of the target vehicle by interpolating the output characteristics of the battery of the target vehicle over a predetermined period using the improved regression model of the target vehicle; a setting unit that sets one or more source vehicles with high similarity to the target vehicle based on the prediction accuracy index; and an extrapolation unit that extrapolates the output characteristics of the battery of the target vehicle using the regression model of the target vehicle which has been improved by transfer learning using the regression coefficients of the regression model of one or more source vehicles with high similarity to the target vehicle, wherein the setting unit updates the source vehicles with high similarity to the target vehicle from the source vehicles set in the previous prediction process to the source vehicles identified as similar to the target vehicle in the current prediction process if the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process.

[0130] <Note 2> The output characteristic prediction device according to Appendix 1, wherein the vehicle data includes, as time-series data, time-series data relating to the power generation control of the battery and time-series data relating to the usage status of the vehicle.

[0131] <Note 3> The output characteristic prediction device described in Appendix 1 or Appendix 2, wherein the value at each point in time in the aforementioned time series data is the average, median, or cumulative value of the values ​​over a predetermined calculation period.

[0132] <Note 4> An output characteristic prediction system comprising a plurality of vehicles equipped with batteries, and an output characteristic prediction device for predicting the output characteristics of the batteries, wherein the output characteristic prediction device is an output characteristic prediction device described in any one of the appendices 1 to 3.

[0133] <Note 5> The vehicle comprises a fuel cell as a battery, an energy storage device capable of charging the output power of the fuel cell, a charge rate detection unit for detecting the charge rate of the energy storage device, and an FC control device for controlling the power generation of the fuel cell, wherein the FC control device is capable of performing switching control to switch the output power of the fuel cell in stages based on the charge rate of the energy storage device detected by the charge rate detection unit, and the vehicle data includes, as time series data, time series data of the output characteristics of the fuel cell during the execution of the switching control, as described in Appendix 4 of the output characteristics prediction system. [Explanation of symbols]

[0134] 10...Vehicle, 11...Output characteristic prediction device, 14...Acquisition unit, 15...Model generation unit, 16...Prediction unit, 16a...Specification unit, 16b...Model improvement unit, 16c...Calculation unit, 16d...Setting unit, 16e...Extrapolation unit, 23...Fuel cell as a battery, 24...Energy storage device, 25...FC control device, 37...Charge rate detection unit, 100...Output characteristic prediction system, D...Vehicle data, M1...Regression model.

Claims

1. An output characteristic prediction device for predicting the output characteristics of a battery installed in a vehicle, An acquisition unit that individually acquires vehicle data, including time-series data, from multiple of the aforementioned vehicles, A model generation unit generates a regression model for each vehicle that outputs a target variable including the output characteristics when an explanatory variable is input, by performing machine learning using the time series data for the learning period for each vehicle. A prediction unit that performs a prediction process to predict the output characteristics of the battery for each vehicle, Equipped with, The prediction unit executes the prediction process each time the acquisition unit newly acquires vehicle data for a predetermined period from the multiple vehicles. When one of the aforementioned multiple vehicles is designated as the target vehicle and the other vehicles are designated as source vehicles, The prediction unit, An identification unit identifies one or more source vehicles similar to the target vehicle based on the vehicle data of the plurality of vehicles for the predetermined period, A model improvement unit improves the regression model of the target vehicle by performing transfer learning using the regression coefficients of the regression model of one or more source vehicles similar to the target vehicle. A calculation unit calculates a predictive accuracy index for the improved regression model of the target vehicle by interpolating the output characteristics of the battery of the target vehicle during the predetermined period using the improved regression model of the target vehicle, A setting unit sets one or more source vehicles that have a high similarity to the target vehicle based on the prediction accuracy index, An extrapolation unit extrapolates the output characteristics of the battery of the target vehicle using the regression model of the target vehicle, which has been improved by transfer learning using the regression coefficients of the regression model of one or more source vehicles that have a high similarity to the target vehicle, It has, The setting unit is an output characteristic prediction device that, when the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process, updates the source vehicle with a high similarity to the target vehicle from the source vehicle set in the previous prediction process to the source vehicle identified as similar to the target vehicle in the current prediction process.

2. The output characteristic prediction device according to claim 1, wherein the vehicle data includes, as time-series data, time-series data relating to the power generation control of the battery and time-series data relating to the usage status of the vehicle.

3. The output characteristic prediction device according to claim 1, wherein the value at each point in time in the aforementioned time series data is the average, median, or cumulative value of the values ​​over a predetermined calculation period.

4. Multiple vehicles equipped with batteries, An output characteristic prediction device for predicting the output characteristics of the aforementioned battery, An output characteristic prediction system comprising, The output characteristic prediction device is the output characteristic prediction device described in claim 1, which is the output characteristic prediction system.

5. The aforementioned vehicle is The aforementioned fuel cell as a battery, A power storage device capable of charging the output power of the fuel cell, A charge rate detection unit for detecting the charge rate of the energy storage device, An FC control device for controlling the power generation of the fuel cell, Equipped with, The FC control device is capable of performing switching control to switch the output power of the fuel cell in stages based on the charge rate of the energy storage device detected by the charge rate detection unit. The output characteristic prediction system according to claim 4, wherein the vehicle data includes, as time-series data, time-series data of the output characteristics of the fuel cell during the execution of the switching control.

6. An output characteristic prediction method for predicting the output characteristics of a battery installed in a vehicle using an output characteristic prediction device, The output characteristic prediction device is An acquisition unit that individually acquires vehicle data, including time-series data, from multiple of the aforementioned vehicles, A model generation unit generates a regression model for each vehicle that outputs a target variable including the output characteristics when an explanatory variable is input, by performing machine learning using the time series data for the learning period for each vehicle. A prediction unit that performs a prediction process to predict the output characteristics of the battery for each vehicle, Equipped with, The prediction process is performed each time the acquisition unit acquires new vehicle data for a predetermined period from the multiple vehicles. When one of the aforementioned multiple vehicles is designated as the target vehicle and the other vehicles are designated as source vehicles, The aforementioned prediction process is, The identification unit of the prediction unit identifies one or more source vehicles similar to the target vehicle based on the vehicle data of the plurality of vehicles for the predetermined period, The model improvement unit of the prediction unit improves the regression model of the target vehicle by performing transfer learning using the regression coefficients of the regression model of one or more source vehicles similar to the target vehicle. The calculation unit of the prediction unit calculates a prediction accuracy index of the improved regression model of the target vehicle by interpolating the output characteristics of the battery of the target vehicle during the predetermined period using the improved regression model of the target vehicle. The setting unit of the prediction unit sets one or more source vehicles that have a high similarity to the target vehicle based on the prediction accuracy index, The extrapolation unit of the prediction unit extrapolates the output characteristics of the battery of the target vehicle using the regression model of the target vehicle, which has been improved by transfer learning using the regression coefficients of the regression model of one or more source vehicles that have a high similarity to the target vehicle. Includes, Output characteristic prediction method in which, in the step of setting a source vehicle that has a high similarity to the target vehicle, the setting unit updates the source vehicle that has a high similarity to the target vehicle from the source vehicle set in the previous prediction process to the source vehicle that has been identified as similar to the target vehicle in the current prediction process if the prediction accuracy in the current prediction process is better than the prediction accuracy in the previous prediction process.

Images