Water volume prediction device, water volume prediction method, and program

Abstract

To provide a water volume prediction device, a water volume prediction method, and a program that can predict water volume more accurately without requiring specialized knowledge. [Solution] The water volume prediction device of the embodiment includes a reception unit that receives input of information including inflow data to a target location, an estimation unit that uses a predetermined large-scale language model to estimate parameters of a predetermined physical model that takes the inflow data as input and outflow data as output, and a prediction unit that inputs the information including the inflow data and the parameters into the physical model to predict the outflow data at the target location, wherein the predetermined physical model includes a tank model.

Description

Technical Field

[0001] The present invention relates to a water volume prediction device, a water volume prediction method, and a program.

Background Art

[0002] In recent years, for the analysis and prediction of natural disasters, the utilization of deep learning using a large amount of data accumulated in the past has been promoted. In this regard, a technique for predicting the river water level caused by meteorological disasters using a prediction model of an ANN (Artificial Neural Network) using deep learning is known (see, for example, Non-Patent Document 1). Further, in recent years, a technique for evaluating flood risks and the like by image analysis via generative AI (Generative Artificial Intelligence) is known (see, for example, Non-Patent Document 2).

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the flow rate data used to train flood prediction models includes human judgments (judgments based on experience and intuition), such as dam discharge adjustments, sluice gate operations, and the operation of drainage pumps. Deep learning can only make predictions based on patterns contained in the data, so it is difficult to create patterns in data where variability occurs due to human judgments. Furthermore, methods and technologies are needed to obtain the optimal answers to the queries made by the generative AI, and specialized knowledge is required to judge the certainty of those answers. Therefore, in the past, there was a possibility that accurate water volume predictions could not be made.

[0005] The aspects of the present invention have been made in consideration of these circumstances, and aim to provide a water volume prediction device, a water volume prediction method, and a program that can predict water volume more accurately without requiring specialized knowledge. [Means for solving the problem]

[0006] The water volume prediction device, water volume prediction method, and program according to this invention employ the following configuration. A water volume prediction device according to a first aspect of the present invention comprises a reception unit that receives input of information including inflow data to a target location, an estimation unit that estimates parameters of a predetermined physical model that takes the inflow data as input and outflow data as output using a predetermined large-scale language model, and a prediction unit that inputs the information including the inflow data and the parameters to the physical model to predict the outflow data at the target location, wherein the predetermined physical model includes a tank model.

[0007] A water volume prediction device according to a second aspect of the present invention further comprises a determination unit that determines whether the error between the discharge data predicted by the prediction unit and predetermined discharge data is within a predetermined range, the estimation unit updates the parameters when the determination unit determines that the error is not within the predetermined range, and the prediction unit predicts the discharge data again using the updated parameters.

[0008] A third aspect of the present invention is a water volume prediction device in which the prediction unit further predicts at least one of the flood situation and the inundation situation based on the runoff data.

[0009] A fourth aspect of the present invention, a water volume prediction device, further comprises a providing unit that generates and provides to the user information indicating the results predicted by the prediction unit.

[0010] A fifth aspect of the present invention is a water volume prediction device in which the estimation unit further refers to a dataset in which the previously stored inflow data, outflow data, and parameters are associated, and extracts a dataset in which the similarity to the inflow data received by the reception unit is equal to or greater than a threshold, and estimates the parameters.

[0011] A sixth aspect of the present invention is a water volume prediction method in which a computer receives input of information including inflow data to a target location, estimates parameters of a predetermined physical model that takes the inflow data as input and outflow data as output using a predetermined large-scale language model, inputs the information including the inflow data and the parameters into the physical model to predict the outflow data at the target location, and the predetermined physical model includes a tank model.

[0012] A program according to a seventh aspect of the present invention is a program which causes a computer to accept input of information including inflow data to a target location, estimates parameters of a predetermined physical model that takes the inflow data as input and outflow data as output using a predetermined large-scale language model, inputs the information including the inflow data and the parameters into the physical model to predict the outflow data at the target location, and the predetermined physical model includes a tank model. [Effects of the Invention]

[0013] According to the embodiments of the present invention described above, water volume can be predicted more accurately without requiring specialized knowledge.

Brief Description of the Drawings

[0014] [Figure 1] It is a diagram showing an example of the schematic configuration of the water volume prediction system 1 including the water volume prediction device 100 of the embodiment. [Figure 2] It is a diagram showing an example of the functional configuration of the terminal device 200. [Figure 3] It is a diagram for explaining an example of water volume prediction in the embodiment. [Figure 4] It is a diagram for explaining an example of the content of the dataset 196. [Figure 5] It is an example of the data input to the generated AI model 192 in the first embodiment. [Figure 6] It is a diagram showing an example of other input data in the first embodiment. [Figure 7] It is a diagram showing an example of the estimation result of the proportionality constant (a) obtained using the generated AI model 192 of the first embodiment. [Figure 8] It is an example of the data input to the generated AI model 192 in the second embodiment. [Figure 9] It is a diagram showing an example of the inflow rate (x) input as inquiry information. [Figure 10] It is a diagram showing an example of the response data in the second embodiment. [Figure 11] It is a flowchart showing an example of the process executed by the water volume prediction device 100.

Modes for Carrying Out the Invention

[0015] Hereinafter, with reference to the drawings, embodiments of the water volume prediction device, water volume prediction method, and program of the present invention will be described. In the following, for example, a water volume prediction device that predicts outflow data for inflow data at a target location or predicts the impact (e.g., flood situation, inundation situation, etc.) in the vicinity (peripheral area) of the target location based on the outflow data will be described. The target location is, for example, a river, a dam, a water storage tank, etc., but is not limited thereto. Further, the inflow data is, for example, the inflow volume of water such as precipitation, and the outflow data is information related to the water volume such as the outflow volume of water and the water level, but is not limited thereto. Further, the water volume may include not only the amount of water but also the amount of sediment and the like.

[0016] <Water volume prediction system> FIG. 1 is a diagram showing an example of the schematic configuration of a water volume prediction system 1 including a water volume prediction device 100 according to an embodiment. The water volume prediction system 1 includes, for example, a water volume prediction device 100 and at least one or more terminal devices 200-1 to 200-n. Hereinafter, unless otherwise distinguished and described for each of the terminal devices 200-1 to 200-n, they will be simply referred to as "terminal device 200" for description. The water volume prediction device 100 and the terminal device 200 communicate via, for example, a network NW. The network NW includes, for example, a cellular network, a Wi-Fi network, Bluetooth (registered trademark), the Internet, a WAN (Wide Area Network), a LAN (Local Area Network), a public line, a provider device, a dedicated line, a wireless base station, and the like.

[0017] The water volume prediction device 100 predicts the water volume at the target location based on input data from a user who uses the terminal device 200. Further, the water volume prediction device 100 generates information (e.g., an image, etc.) for providing the prediction result to the user, displays it on the display unit of the water volume prediction device 100, or transmits it to the terminal device 200 to provide it to the user. The water volume prediction device 100 may be, for example, a server device or a PC (Personal Computer), or may be a cloud server configured by cloud computing including one or more processing devices.

[0018] The terminal device 200 is a terminal used by the user. For example, the terminal device 200 receives input data including target locations for water volume prediction and inflow data, transmits the received input data to the water volume prediction device 100, and receives and provides to the user the prediction results such as outflow data obtained from the water volume prediction device 100. The terminal device 200 may be, for example, a PC or tablet terminal, or a communication terminal such as a smartphone.

[0019] Next, we will specifically explain the functional configuration of the water volume prediction device 100 and the terminal device 200. <Water volume prediction device 100> The water volume prediction device 100 includes, for example, a communication unit 110, a reception unit 120, an output unit 130, an estimation unit 140, a prediction unit 150, a determination unit 160, a supply unit 170, a control unit 180, and a storage unit 190. Some or all of the estimation unit 140, the prediction unit 150, the determination unit 160, the supply unit 170, and the control unit 180 are realized by a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or GPU (Graphics Processing Unit), or by the cooperation of software and hardware. The program may be stored in advance on the HDD (Hard Disk Drive) or flash memory of the water volume prediction device 100, or it may be stored on a removable storage medium such as a DVD, CD-ROM, or memory card, and installed on the HDD or flash memory of the water volume prediction device 100 when the storage medium (non-transient storage medium) is inserted into the drive device.

[0020] The memory unit 190 is implemented by, for example, an HDD, flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), ROM (Read Only Memory), or RAM (Random Access Memory). The memory unit 190 stores, for example, a generated AI model 192, a physical model 194, a dataset 196, a program, and various other information.

[0021] The generative AI model 192 is a model that autonomously generates and outputs parameters (e.g., setting conditions) to be input to the physical model 194 in response to predetermined input information (e.g., input information corresponding to the type of object). The generative AI model may include, for example, large language models (LLMs) specialized in natural language processing (NLP), or it may include conversational AI models specialized in enabling natural conversation (dialogue) with a person (user) by applying large language models. In a conversational AI model, for example, various questions for collecting elements (information) necessary for estimating parameters, etc., based on dialogue with the user are stored in advance, and the next question is selected according to the user's answer to the question, and the necessary elements are collected by repeating the dialogue including questions and answers.

[0022] Furthermore, the generation AI model 192 has an LLM and a diffusion model that performs image generation, and is a multimodal model that can handle text data, image data, etc. as input data. In this embodiment, the generation AI model 192 is used to operate a physical model (hydraulic model) 194 that provides physical information such as leak data without human intervention (for example, by adjusting parameters described later). The generation AI model 192 may use, for example, ChatGPT®, or other models or systems (for example, Gemini®, OpenAI®, BingAI). The generation AI model 192 may be stored in the storage unit 190, or it may be provided on an external device (for example, a generation AI server, etc.) via a network NW. Figure 1 shows an example in which the generation AI model 192 is stored in the storage unit 190.

[0023] Physical model 194 is a model that predicts outflow data, etc., based on inputs such as inflow data and parameters. Physical model 194 is, for example, a tank model. A tank model replaces the target location with a container (tank) that has an outflow hole according to the characteristics of that location (shape, size, geology, etc.), and derives the amount of water outflow in relation to the inflow data and parameters. In a tank model, for example, the relationship between rainfall and flow rate in the target area is determined using parameters, so setting the parameters is extremely important.

[0024] The physical model 194 may be used individually with the generated AI model 192, or used in combination. Furthermore, the physical model 194 may be stored in the memory unit 190, or it may be provided on an external device via a network NW. Figure 1 shows an example where the physical model 194 is stored in the memory unit 190. The generated AI model 192 and the physical model 194 described above may be updated as needed.

[0025] Dataset 196 is a collection of data consisting of, for example, precipitation data over time, water level data, and text data with precipitation and water level values ​​in tabular format using Markdown. Dataset 196 may be stored for each target location and may contain information about parameters used in the physical model 194. Dataset 196 is used, for example, when estimating parameters to be input into the physical model 194. Dataset 196 includes, for example, actual measurement data such as past flood data. Using this data makes it possible to achieve highly accurate flood predictions.

[0026] The communication unit 110 communicates with terminal devices 200 and other external devices, for example, via a network NW. For example, the communication unit 110 transmits information processed by each component within the water volume prediction device 100 to terminal devices 200 and other external devices. The communication unit 110 also outputs information received from terminal devices 200 and other external devices to predetermined components within the water volume prediction device 100. The above-described transmission and reception processing may be controlled by the control unit 180 or by other components.

[0027] The reception unit 120 receives input of various types of information from the administrator of the water volume prediction device 100 (the administrator of the water volume prediction system 1) and users of the terminal device 200. This information includes, for example, various input data (such as the target location and inflow data) for obtaining prediction results such as the outflow volume at the target location. The administrator mentioned above is also an example of a "user" of this system.

[0028] The output unit 130 includes, for example, a display unit and a speaker, and outputs predetermined information to each. The display unit is, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) display. The display unit displays a screen showing various information (text and images) in the embodiment. The speaker outputs predetermined sound. The receiving unit 120 may be configured integrally with the display unit as a touch panel. For example, the output unit 130 outputs, for example, information received by the receiving unit 120 or information generated by the providing unit 170.

[0029] The estimation unit 140 uses the generation AI model 192 to estimate parameters for input to the physical model 194, which outputs (predicts) outflow data in relation to inflow data. Details of the functions of the estimation unit 140 will be described later.

[0030] The prediction unit 150 inputs the input data, including inflow data, and the parameters estimated by the estimation unit 140, into the physical model 194 to predict the outflow data at the target location. Based on the predicted outflow data, the prediction unit 150 also predicts at least one of the flood conditions and inundation conditions as the impact on the area near the target location (surrounding area). Details of the functions of the prediction unit 150 will be described later.

[0031] The determination unit 160 determines whether the error between the result predicted by the prediction unit 150 (for example, predicted outflow volume) and a predetermined value (comparative outflow volume) is within a predetermined range. The comparative outflow volume is a reference value that can be guaranteed by, for example, physical laws, and may be a value set based on past measured data at the same location, or a predetermined standard value. The comparative outflow volume may be stored in the storage unit 190 in advance, or it may be obtained from an external device via a network NW. If the determination unit 160 determines that the error is not within the predetermined range (exceeds the predetermined range), it requests the estimation unit 140 to update the parameters. The estimation unit 140 receives the update request from the determination unit 160 and updates the parameters. At this time, the estimation unit 140 may adjust the amount of change in the parameters based on the magnitude of the error and the relationship between the predicted outflow volume and the comparative outflow volume, or it may adjust based on a predetermined fixed change amount. Then, the prediction unit 150 predicts the outflow data at the target location again using the updated parameters. The above process is repeated until the error falls within a predetermined range, thereby obtaining more accurate prediction results.

[0032] The provisioning unit 170 generates various screens for users to input data such as target areas and inflow data, and transmits the generated screens to the terminal device 200 or displays them on the display unit of the output unit 130, etc., to provide them to users (including system administrators). The provisioning unit 170 also generates information such as images and audio (provided information) that shows the results predicted by the prediction unit 150, and outputs the generated provided information to the output unit 130 or transmits it to the terminal device 200 to provide it to users.

[0033] The control unit 180 controls the entire functional configuration of the water volume prediction device 100. For example, the control unit 180 controls the transmission and reception of information (data) in the communication unit 110, causes the estimation unit 140 to estimate parameters etc. based on the information received by the reception unit 120, and causes the prediction unit 150 to perform prediction processing. The control unit 180 also causes the determination unit 160 to perform determination processing, causes the provision unit 170 to generate provision information, and provides it to the user.

[0034] <Terminal device 200> Figure 2 shows an example of the functional configuration of the terminal device 200. The terminal device 200 includes, for example, a terminal-side communication unit 210, an input unit 220, an output unit 230, a control unit 240, an application execution unit 250, and a terminal-side storage unit 260. Some or all of the control unit 240 and the application execution unit 250 are realized, for example, by a hardware processor such as a CPU executing a program (software). Some or all of these components may also be realized by hardware (including circuitry) such as an LSI, ASIC, FPGA, or GPU, or by the cooperation of software and hardware. The program may be stored in advance in a storage device such as an HDD or flash memory (a storage device with a non-transient storage medium), or it may be stored in a removable storage medium such as a DVD, CD-ROM, or memory card (a non-transient storage medium), and installed in the storage device when the storage medium is inserted into a drive device or a card slot of the terminal device 200.

[0035] The terminal-side storage unit 260 may be implemented using the various storage devices described above, or by EEPROM, ROM, RAM, etc. For example, the terminal-side storage unit 260 stores applications 262, programs, and various other information.

[0036] The terminal-side communication unit 210 communicates with the water volume prediction device 100 and other external devices, for example, via a network NW. For example, the terminal-side communication unit 210 transmits information processed by each component within the terminal device 200 to the water volume prediction device 100 and other external devices. The terminal-side communication unit 210 also outputs information received from the water volume prediction device 100 and other external devices to a predetermined component within the water volume prediction device 100. The above-described transmission and reception processing may be controlled by the control unit 240 or by other components.

[0037] The input unit 220 accepts user input, for example, through the operation of various keys or buttons. The input unit 220 may also be equipped with a microphone to accept user voice input.

[0038] The output unit 230 includes, for example, a display unit 232 and a speaker 234. The output unit 230 outputs predetermined information to the display unit 232 and the speaker 234. The display unit 232 is, for example, an LCD or an organic EL display. The input unit 220 may be configured integrally with the display unit 232 as a touch panel. The display unit 232 displays various information according to the embodiment. The speaker 234 outputs predetermined sound. For example, the output unit 230 outputs a screen or sound corresponding to the information provided by the water volume prediction device 100, or outputs information input by the input unit 220.

[0039] The control unit 240 controls the overall functions of the terminal device 200. For example, the control unit 240 controls the transmission and reception of information (data) in the terminal-side communication unit 210, controls the input and output of data by the input unit 220 and output unit 230, and controls the execution of the application 262 by the application execution unit 250.

[0040] The application execution unit 250 is realized by executing an application 262 stored in the terminal-side storage unit 260. The application 262 is, for example, downloaded from an external device via a network NW and installed on the terminal device 200. Alternatively, the application 262 may be an application program (software) provided by the water volume prediction device 100, and may, for example, perform processing to access a predetermined address (URL; Uniform Resource Locator) via a browser. The application 262 transmits input data such as target locations and inflow data entered by the user from the terminal device 200 to the water volume prediction device 100 to execute water volume prediction processing and obtain the execution result. When the application 262 transmits information entered by the input unit 220 to the water volume prediction device 100, it may perform processing to include identification information that identifies the terminal device 200 and identification information that identifies the user in the information to be transmitted.

[0041] <Regarding water volume forecasts> Next, the water volume prediction in the embodiment will be explained using a diagram. Figure 3 is a diagram illustrating an example of water volume prediction in the embodiment. In the water volume prediction system 1 of this embodiment, a system (CoT simulator) is constructed that links a generating AI model 192 and a simulator using a physical model 194 via CoT (Chain of Thought) prompting, which can generate the thought process performed by a human. For example, in the embodiment, the target is flood prediction due to climate change, and a CoT simulator is constructed that accepts input of inflow data such as precipitation at the target site and predicts the outflow of rivers, etc. As an example of the physical model 194 (computer that outputs physical information) used in this simulator, a tank model will be used in the explanation below. In the explanation below, a simple single-stage tank model will be used, but a multi-stage model may be used depending on the characteristics of the target site. For example, if the water depth (or flow rate in the tank) of a single-stage tank model is η, the outflow (Q) is given by the following equation (1). Q = (a) × η …(1)

[0042] Here, (a) in equation (1) is a proportionality constant related to the outflow rate and is an example of a parameter (internal parameter). The estimation unit 140 estimates (automatically adjusts) the parameters to be input to this tank model using the generating AI model 192. The prediction unit 150 predicts outflow data, etc., using the parameters estimated by the estimation unit 140. The provision unit 170 generates and provides to the user information including the results predicted by the prediction unit 150.

[0043] To explain in more detail, for example, as shown in Figure 3, the user inputs input data (prompts) including inflow data such as text data and image data via a terminal device 200, etc., and queries the water volume prediction device 100 for the prediction results of outflow data at the target location. The input data (prompts) include information such as, for example, "If there is a rainfall of 100 [mm] per hour in the river at point A, what will the flood situation of the river be like 2 hours later?" The reception unit 120 or the control unit 180 performs known language analysis on the above sentence and obtains the content of the input data from the analysis results, such as the target location "Point A", inflow data "100 [mm] of rainfall per hour", and query (prediction target) "flood situation of the river 2 hours later".

[0044] On the water volume prediction device 100 side, in addition to inputting the aforementioned target location (location A) and inflow data (rainfall of 100 mm per hour) into the tank model, which is the physical model 194, the device also refers to the dataset 196 to search for data similar to the content of the input data, and estimates the parameters to be input into the tank model (for constructing the tank model) from the search results.

[0045] Figure 4 is a diagram illustrating an example of the contents of dataset 196. The dataset 196 shown in Figure 4 associates precipitation data, water level data, situational descriptive text, parameters, etc., with changes over time. More specifically, in the example in Figure 4, to make the prediction target easier to understand, the precipitation data is represented as parametric time-series data that can be expressed using a Gaussian distribution. Furthermore, the precipitation data stores various distributions using three parameters: mean, variance, and total precipitation. In the example in Figure 4, the precipitation data stores a total of 576 datasets including multiple locations, and the water level data, which is the output result when these datasets are input into the tank model, is stored in a time-series data format with a time delay. In this embodiment, a single-stage tank model is used, and three types of internal parameters are set within that tank model, storing 576 × 3 sets of water level data. Internal parameters are, for example, parameters used to predict (calculate) runoff data using the physical model 194, and are the proportionality constant (a) mentioned above. The types and amount of data in the dataset 196, the types and number of parameters, etc., in the embodiment are not limited to the example in Figure 4, but can be appropriately adjusted according to the types and number of parameters of the physical model 194, the number of target locations to be stored, etc. In addition, the dataset 196 may also be associated with text data that describes the situation when the water level shown in the water level data is reached (text data showing the relationship between precipitation and water level), parameters other than internal parameters, image data such as graphs, etc.

[0046] For example, the estimation unit 140, via the generation AI model 192 (LLM) described above, refers to the dataset 196 based on the target location (location A) and inflow data, and extracts data that is close to the target location entered by the user (including the same location) and has a high similarity to the inflow data (for example, precipitation data with a similarity above a threshold). If the input data includes information such as a description of the situation from the user, that information may be used to extract data with a high similarity to the situation from the dataset 196.

[0047] Furthermore, if the incoming data includes image data, the estimation unit 140 may perform image analysis via the generation AI model 192 (or other image analysis models). Image analysis may use known standard image analysis models, or models such as CLIP (Contrastive Language-Image Pre-training) may be used. For example, a standard image analysis model estimates text data corresponding to image features using a pre-trained model in which an image feature extractor and a linear classifier are trained as a pair. On the other hand, CLIP trains an image encoder and a text encoder as a pair to estimate the correct combination for a series of images and texts. CLIP is a model that has been pre-trained to ensure that the distance between text and image pairs that have a correspondence relationship is small after mapping. Therefore, by using CLIP, text data such as precipitation data and water level data corresponding to images can be extracted more appropriately than with a standard image analysis model.

[0048] The generative AI model 192 then acquires water level data and internal parameters associated with highly similar precipitation data. Note that, as shown in Figure 4, when the generative AI model 192 acquires three types of internal parameters from the dataset 196, it may select one of the three internal parameters, estimate a single parameter by combining multiple parameters, or estimate a parameter by associating it with other information (input data). Furthermore, the generative AI model 192 may select (or estimate) parameters for the user through interaction with the user.

[0049] Next, the prediction unit 150 inputs inflow data (rainfall of 100 mm per hour) and estimated parameters into a tank model (physical model 194), and operates the tank model to predict the flood situation of the river two hours later corresponding to the inquiry. At this time, the prediction unit 150 constructs a tank model corresponding to the estimated parameters, inputs inflow data into the constructed tank model to calculate the water level (outflow) for a predetermined time (for example, one hour), and predicts the water level two hours later based on the calculated water level. Furthermore, the prediction unit 150 predicts the flood situation of the target river (near point A, which is the target location) from the water level two hours later. The flood situation may be, for example, the water level or flow rate at the target location, or the scale of the flood damage caused by the occurrence of a flood. Alternatively, the flood situation may be the flood risk level obtained by comparing the predicted water level with the height information of surrounding levees, etc., obtained from map information, or by comparing the predicted water level or flow rate with the water level or flow rate values ​​from past disasters. Furthermore, flood conditions may include information regarding the status (scale) of flood warnings and evacuation warnings.

[0050] In this embodiment, the prediction unit 150 may predict inundation conditions in addition to (or instead of) flood conditions. Inundation conditions include, for example, the inundation depth (height from the ground to the water surface in the inundated area) and the estimated inundation area based on that, if there are buildings such as houses near the target river. The topography and buildings near the target river in the inundation conditions may be obtained by referring to map information, for example, or they may be standard topography and buildings assumed for the simulation. The map information mentioned above may be stored in the storage unit 190 or obtained from an external device via a network NW. Furthermore, the prediction results for flood conditions and inundation conditions may include mapped image data and graphed image data (e.g., flood waveforms, inundation distribution, etc.) of each condition.

[0051] Furthermore, in this embodiment, as shown in Figure 4, the determination unit 160 determines whether the error between the predicted discharge amount (river discharge amount (water level)), which is the prediction result, and a predetermined comparative discharge amount (water level) is within a predetermined range. If the error exceeds the predetermined range, it outputs a request to the generating AI model 192 to update the parameters. Here, the predetermined comparative discharge amount may be, for example, water level data obtained by searching the dataset 196, or it may be a value based on that water level data. Since the dataset 196 includes past measured data, it is possible to determine whether the prediction result using the physical model 194 is accurate data (whether it is an appropriate value that can be guaranteed by physical laws, etc.) by comparing it with that data.

[0052] The estimation unit 140 updates the parameters input to the physical model 194 to different parameters via the generation AI model 192. The updated parameters may be parameters stored in the dataset 196 (for example, the remaining unused parameters from the three types of parameters (internal parameters) mentioned above), or parameters adjusted based on the nature of the error. The prediction unit 150 inputs the updated parameters into the tank model and performs the prediction again. If the determination unit 160 determines that the error is within a predetermined range, the provision unit 170 generates provision information (text data, image data, etc.) using the prediction result information and outputs the generated provision information to the user (terminal device 200) as an answer to the inquiry. The generation AI model 192 may also be used to generate the provision information.

[0053] Thus, in this embodiment, the internal parameters of the tank model necessary for predicting leakage data can be automatically adjusted by the generating AI model 192, so that users can obtain appropriate answers that can be guaranteed by physical laws without having specialized knowledge. Furthermore, according to this embodiment, by retraining the generating AI with human experience and intuition regarding natural disasters, as well as human wisdom such as mathematics and physical laws, and adjusting the internal parameters, it is possible to construct a generating AI that specializes in comprehensive knowledge of natural disasters, for example.

[0054] <Examples> The following describes several embodiments using the water volume prediction system 1 in the embodiment. In addition, two embodiments relating to the coordination between the generated AI model 192 (LLM) and the physical model (computer) 184 will be shown below.

[0055] [First Embodiment] The first embodiment is an example of estimating the proportionality constant (a) for the outflow rate, which is the output result of the tank model, by providing tabular data of the temporal change in precipitation to the generating AI model 192. Figure 5 shows an example of the data input to the generating AI model 192 in the first embodiment. In Figure 5, [Table 1] shows the relationship between hourly inflow (x) and outflow (y) using the tank model from 1:00 to 12:00. The proportionality constant (a), which is a parameter of this tank model, is assumed to be unknown.

[0056] Here, let's assume that the prediction unit 150 was made to predict the flow rate using a tank model (initial tank) with a proportionality constant (a) of "0.2", and the result was as shown in [Table 2]. In this case, the peak outflow is slightly larger compared to the result in [Table 1]. Therefore, the estimation unit 140 inputs the contents shown in [Table 1] and [Table 2] (data on the time change of precipitation) into the generated AI model 192, and also inputs inquiry information such as, "What should the value of the proportionality constant (a) be in order to bring the result shown in [Table 2] closer to the result shown in [Table 1]?" The estimation unit 140 may also accept image data such as graphs as input instead of table data as shown in Figure 5. Figure 6 is a diagram showing an example of other input data in the first embodiment. In the image data IMG1 shown in Figure 6, the horizontal axis shows time, and the vertical axis shows the inflow (x), the outflow (y) shown in Table 5, and the outflow (y') shown in Table 6. Furthermore, even when images such as graphs are input, the image analysis processing of the generating AI model 192 can appropriately extract the data. This input data may be generated by the estimation unit 140 or received as a prompt by the reception unit 120.

[0057] The generating AI model 192 estimates the proportionality constant (a) using, for example, a dataset or by acquiring prediction results from the prediction unit 150, and outputs the estimation result. The generating AI model 192 may also estimate the proportionality constant (a) in cooperation with other models or search systems. Figure 7 shows an example of the estimation result of the proportionality constant (a) obtained using the generating AI model 192 of the first embodiment. The generating AI model 192 outputs text data TXT1 containing explanatory information as shown in Figure 7 for the above input. Note that the content shown in Figure 7 (interactive data) is merely an example and is not limited to this; at least a part of it may be omitted or changed. According to the example in Figure 7, the estimated value of the proportionality constant (a) is output as 0.1 to 0.15.

[0058] According to the first embodiment, by using the generating AI model 192 to present a new outflow amount based on the tank model constructed from input information of inflow (precipitation) and outflow, the proportionality constant of the tank model can be appropriately estimated. In this way, in the first embodiment, the parameters to be input to the physical model 194 can be estimated using the generating AI model 192, and water volume prediction can be performed more accurately using these parameters.

[0059] [Second Example] The second embodiment is an example in which the generating AI model 192 can make appropriate predictions that satisfy physical laws even when given input data related to a simple flood event. Figure 8 shows an example of the data input to the generating AI model 192 in the second embodiment. In the example in Figure 8, the time-series relationship (tank model) between the amount of water flowing into the tank (x) and the amount of water flowing out (y) proportional to the volume of water in the tank is shown when water flows into the tank every two hours from 0:00 to 24:00. Figure 9 shows an example of the amount of water flowing in (x) that is input as query information. In the second embodiment, the user inputs input data (prompt) from the terminal device 200, which includes the tank model with the relationship shown in Figure 8 (an example of inflow data) and query information asking what the amount of water flowing out will be when an inflow amount (x) like the one shown in Figure 9 is received.

[0060] The estimation unit 140 provides the generating AI model 192 with information (table data) showing the time-series relationship between inflow and outflow volumes as shown in Figure 8, which is used as input to the tank model, and causes the generating AI model 192 to construct the tank model. In this case, for example, the parameters can be extracted from the dataset 196 with a similarity to the data shown in Figure 8 that is above a threshold, and the parameters associated with the extracted data can be used. Furthermore, in response to queries for time-series data of inflow volumes as shown in Figure 9, the estimation unit 140 works in conjunction with the tank model to cause the prediction unit 150 to predict the outflow volume.

[0061] Figure 10 shows an example of the prediction results in the second embodiment. The example in Figure 10 includes text data TXT2 and image data IMG2. The text data TXT2 is data (interactive data) that includes explanatory information output by the generation AI model 192, but this is merely an example and is not limited to this, and at least a part of it may be omitted or changed. Similarly, the image data IMG2 is data predicted by the prediction unit 150, but this is also just an example and is not limited to this. In Figure 10, x represents the inflow amount corresponding to the precipitation data shown in Figure 9, y represents the outflow amount obtained from the tank model input in Figure 8, and y' represents the predicted value of the outflow amount when the tank model constructed by the generation AI model 192 of this embodiment is used for x. Comparing the values ​​of y and y' shown in Figure 10, the results are generally reasonable, although the timing of the outflow peak is slightly earlier.

[0062] According to the second embodiment, for example, even if the generating AI model 192 constructs a tank model from time-series data input information regarding inflow (precipitation) and outflow, it can predict an appropriate outflow amount for new precipitation input data. Thus, this embodiment can be applied to various water volume predictions.

[0063] Furthermore, by inputting information about the target location into the first and second embodiments described above, more accurate estimations and predictions suitable for that location (region) can be made. At least a portion of the contents shown in Figures 7 and 10 above may be generated as provided information by the providing unit 170 and provided to the user via the terminal device 200.

[0064] Thus, according to this embodiment, since the generating AI model 192 uses a collaborative method in which it manipulates the physical model 194, users can make more accurate water volume predictions without having specialized knowledge, and furthermore, various predictions (calculations) can be made in real time.

[0065] <Processing flowchart> Next, an example of the process performed by the water volume prediction device 100 will be explained using a flowchart. Figure 11 is a flowchart of an example of the process performed by the water volume prediction device 100. In the example in Figure 11, the reception unit 120 receives input data including inflow data for the target location (step S100). Next, the estimation unit 140 estimates the parameters to be input to the physical model 194 (for example, to construct a tank model) using the generated AI model 192 (step S110). Next, the prediction unit 150 uses the input data and parameters to predict the outflow data using the physical model 194 (step S120).

[0066] Next, the determination unit 160 performs error determination processing based on the prediction result (step S130) and determines whether the error between the leaked data obtained as the prediction result and the predetermined leaked data is within a predetermined range (step S140). If it is determined that the error is not within the predetermined range (exceeds the predetermined range), the estimation unit 140 updates the parameters using the generated AI model 192 (step S150) and returns to the process in step S120. As a result, the prediction unit 150 predicts the leaked data again using the updated parameters. Also, if it is determined in the process of step S140 that the error is within the predetermined range, the provision unit 170 generates provision information to be provided to the user (step S160) and outputs the generated provision information to the terminal device 200, etc., and provides it to the user (step S170). With this, the processing of this flowchart is completed.

[0067] <Variation> In this embodiment, a tank model was used as an example of physical model 194, but other physical models may be used in addition to (or instead of) the tank model. In this case, for example, a model based on the storage function method may be used as another physical model. A model based on the storage function method separates rainwater into components that run off as floods in a short time and components that run off over a long time due to water retention, etc., and represents the storage (delay) of the flood-running component at the target location (watershed). This model also requires parameters such as constants that are set for each target location (watershed) when calculating the amount of runoff. Therefore, in this embodiment, even when applying this model, the parameters can be estimated using the generated AI model 192, and the runoff data can be predicted more accurately using the estimated parameters.

[0068] Furthermore, in the embodiment, the generative AI model 192 may use RAG (Retrieval Augmented Generation), which combines LLM with search technology. The generative AI model 192 using LLM can perform dialogue with pre-formatted prompt input by providing some auxiliary information to the pre-trained model to make inferences about unknown situations without prior retraining (Zero-shot inference), or by making inferences based on pre-given tasks (Few-shot inference). RAG utilizes this Few-shot inference function and embeds available data into a template, enabling inference using a proprietary database without further training of the LLM itself. Therefore, for example, when using data of past similar phenomena in dataset 196, by adding past data with RAG, it is possible to estimate parameters with high accuracy without retraining the LLM, and furthermore, water volume prediction with higher accuracy can be achieved using these parameters.

[0069] Furthermore, in this embodiment, at least some of the functions of the water volume prediction device 100 may be provided in the terminal device 200. Also, the water volume prediction device 100 may have separate devices within the system that handle the functions of the estimation unit 140, which performs processing using the generation AI model 192, and the prediction unit 150, which performs predictions using the physical model 194.

[0070] As described above, the water volume prediction device 100 of the embodiment includes a reception unit that receives input of information including inflow data to a target location, an estimation unit that uses a predetermined large-scale language model to estimate parameters of a predetermined physical model that takes the inflow data as input and outflow data as output, and a prediction unit that inputs the information including the inflow data and the parameters into the physical model to predict the outflow data at the target location. The predetermined physical model includes a tank model, which enables more accurate water volume prediction without requiring specialized knowledge.

[0071] For example, in one embodiment, focusing on flood prediction due to climate change, the system automatically adjusts parameter values ​​via a generative AI model (LLM) to construct a simulator (a tank model, a type of hydrological model) that calculates river runoff based on input data such as precipitation at a specific location. Then, the runoff is predicted using the tank model adjusted with optimal parameter values. By constructing this system, by prompting input of information such as precipitation at the prediction location, it becomes possible to predict the impact on the surrounding area (e.g., flood conditions, inundation conditions, etc.) based on the runoff, such as appropriate water level rise and flood risk. Therefore, according to this embodiment, detailed risk predictions for natural disasters can be made based on comprehensive human knowledge without requiring specialized knowledge.

[0072] Furthermore, according to the embodiment, when dealing with phenomena that follow physical laws when prompting input, a physical model is combined with the generating AI, and a CoT simulator that learns human thought processes is introduced, allowing even inexperienced prompt engineers or those with insufficient expertise to obtain plausible answers.

[0073] Although embodiments for carrying out the present invention have been described above using examples, the present invention is not limited in any way to these embodiments, and various modifications and substitutions can be made without departing from the spirit of the present invention. [Explanation of Symbols]

[0074] 1...Water volume prediction system, 100...Water volume prediction device, 110...Communication unit, 120...Reception unit, 130, 230...Output unit, 140...Estimation unit, 150...Prediction unit, 160...Determination unit, 170...Provision unit, 180, 240...Control unit, 190...Storage unit, 200...Terminal device, 210...Terminal-side communication unit, 220...Input unit, 232...Display unit, 234...Speaker, 250...Application execution unit, 260...Terminal-side storage unit, 262...Application

Claims

1. A reception area that accepts input of information including data on inflow to the target location, An estimation unit that uses a predetermined large-scale language model to estimate the parameters of a predetermined physical model that takes the inflow data as input and the outflow data as output, The system includes a prediction unit that inputs information including the inflow data and the parameters into the physical model to predict the outflow data at the target location, The predetermined physical model includes a tank model, Water volume prediction device.

2. The system further includes a determination unit that determines whether the error between the leak data predicted by the prediction unit and predetermined leak data is within a predetermined range. The estimation unit updates the parameters when the determination unit determines that the error is not within a predetermined range. The prediction unit then uses the updated parameters to predict the leaked data again. The water volume prediction device according to claim 1.

3. The prediction unit predicts at least one of the flood situation and the inundation situation based on the runoff data. The water volume prediction device according to claim 1.

4. The system further includes a providing unit that generates and provides to the user information indicating the results predicted by the prediction unit. The water volume prediction device according to claim 1.

5. The estimation unit refers to a dataset in which inflow data, outflow data, and parameters are associated with data stored in advance, extracts datasets whose similarity to the inflow data received by the reception unit is above a threshold, and estimates the parameters. The water volume prediction device according to claim 1.

6. Computers We accept input of information including data on inflow to the target location. Using a predetermined large-scale language model, the parameters of a predetermined physical model that takes the inflow data as input and the outflow data as output are estimated. The information including the inflow data and the parameters are input into the physical model to predict the outflow data at the target location. The predetermined physical model includes a tank model, Methods for predicting water volume.

7. On the computer, The system accepts input of information including data on inflow to the target location. Using a predetermined large-scale language model, the parameters of a predetermined physical model that takes the inflow data as input and the outflow data as output are estimated. The information including the inflow data and the parameters are input into the physical model to predict the outflow data at the target location. The predetermined physical model includes a tank model, program.

Images