Prediction device, prediction method, learning device, learning method, and program
The integration of time-series and periodic data through an LSTM model addresses the accuracy decline in DNN-based water level prediction, ensuring precise forecasts over longer durations.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- NAT AGRI & FOOD RES ORG
- Filing Date
- 2022-10-25
- Publication Date
- 2026-07-22
AI Technical Summary
Conventional deep neural network (DNN) based water level prediction techniques suffer from decreased accuracy as the lead time for predicting water levels in regulating ponds and rivers becomes longer.
A prediction device and method that utilizes a long short-term memory (LSTM) architecture trained with both time-series and periodic data extracted from water level measurements, using Fourier transforms to enhance prediction accuracy for longer lead times.
Enables accurate prediction of water levels in retention ponds and rivers even with extended lead times by incorporating periodic components into the learning process.
Smart Images
Figure 0007893472000003 
Figure 0007893472000004 
Figure 0007893472000005
Abstract
Description
Technical Field
[0004] ,
[0006] , , , , , ,
[0005] , , ,
[0003] , , , ,
[0001] The present invention relates to a prediction device, a prediction method, a learning device, a learning method, and a program.
Background Art
[0002] Conventionally, a technique for predicting in real time the water levels in the regulating ponds of drainage airports or rivers where drainage management is performed has been known. For example, Non-Patent Document 1 discloses a technique for predicting the water level of a river using a DNN (deep neural network).
Prior Art Documents
Non-Patent Documents
[0003]
Non-Patent Document 1
Summary of the Invention
Problems to be Solved by the Invention
[0004] Generally, while DNN requires a large amount of teacher data and sufficient learning time, once learning is completed, rapid prediction becomes possible. However, in the conventional water level prediction technique using DNN, the prediction accuracy tends to decrease as the lead time (prediction time) becomes longer for water level prediction up to several hours ahead.
[0005] The present invention has been made in consideration of such circumstances, and one of its objectives is to provide a prediction device, a prediction method, a learning device, a learning method, and a program that can predict the water levels in regulating ponds and rivers with high accuracy even when the lead time becomes long. [[ID=4I]]
Means for Solving the Problems
[0006] A prediction device according to one aspect of the present invention includes an acquisition unit that acquires time-series data of water level measured by a water level gauge, and a prediction unit that predicts future values of water level by inputting input data based on the time-series data into a trained model that has been trained to output future values of water level when input data based on time-series data of water level is input, wherein the input data includes at least periodic data obtained by performing a process to extract periodic components of the change in water level from the time-series data.
[0007] The process for extracting the periodic component of the water level change may be a process that extracts peak values by applying a Fourier transform to the time series data, and then obtains the periodic data by applying an inverse Fourier transform to the extracted peak values.
[0008] The process for extracting the periodic component of the water level change may be a process of obtaining the periodic data by applying a Fourier transform to the time series data to extract peak values, applying an inverse Fourier transform to the extracted peak values, and extending the data obtained from the inverse Fourier transform over a future interval according to the periodic component.
[0009] The input data may include the time-series data and the periodic data.
[0010] A learning device according to one aspect of the present invention is a learning device that learns to output future values of water level when input data based on time-series data of water level measured by a water level gauge is input, The input data includes at least periodic data obtained by performing a process to extract the periodic component of the water level change from the time series data. [Effects of the Invention]
[0011] According to the present invention, even with a long lead time, the water level of retention ponds and rivers can be predicted with high accuracy. [Brief explanation of the drawing]
[0012] [Figure 1] This figure shows an example of the operating environment and configuration of the prediction device 100. [Figure 2] This figure shows an overview of the water level prediction performed by the prediction device 100. [Figure 3] This diagram illustrates the general method for acquiring periodic data 134. [Figure 4] This figure shows an example of the results obtained by applying a Fourier transform to time series data 132. [Figure 5] This figure shows an example of periodic data 134 obtained by the inverse Fourier transform. [Figure 6] This flowchart shows an example of the processing flow performed by the prediction device 100. [Figure 7] This figure shows an example of the configuration of the learning device 200. [Figure 8] This figure shows an example of periodic data 134 related to a modified example. [Modes for carrying out the invention]
[0013] [overview] The following describes a prediction device 100 according to an embodiment of the present invention, with reference to the drawings. Figure 1 is a diagram showing an example of the operating environment and configuration of the prediction device 100. The prediction device 100 operates in cooperation with, for example, a water level gauge 10 and a transmitter 20.
[0014] The water level gauge 10 is a sensor that is installed, for example, on a pipe P laid in a regulating reservoir of a drainage pump station, and measures the height of the water surface (i.e., water level) from a certain reference surface in the regulating reservoir over time. In Figure 1, the water level gauge 10 is shown as a contact-type water level gauge such as a pressure-type water level gauge or a capacitive-type water level gauge, but the present invention is not limited to such a configuration, and the water level gauge 10 may be a non-contact-type water level gauge such as a radio wave-type water level gauge or an ultrasonic-type water level gauge. Furthermore, in this embodiment, the water level gauge 10 is described as measuring the water level of a regulating reservoir, but the present invention is not limited to such a configuration, and the water level gauge 10 may, for example, measure the water level of a river.
[0015] The transmitter 20 is connected to the water level gauge 10 by a cable (not shown) passing through the pipe P, for example. The transmitter 20 transmits the time-series data of the water level in the regulating pond measured by the water level gauge 10 to the prediction device 100 via the network NW. The network NW is an arbitrary network such as a LAN, a WAN, or an Internet line, and may be wired or wireless. Also, the water level gauge 10 itself may include the transmitter 20, and the water level gauge 10 and the transmitter 20 may be integrated.
[0016] The prediction device 100 is a server device such as a web server that predicts the future value of the water level in the regulating pond of the drainage airport. FIG. 2 is a diagram showing an overview of the water level prediction executed by the prediction device 100. As shown in FIG. 2, the prediction device 100 inputs the time-series data 132 of the water level in the regulating pond measured by the water level gauge 10 and the periodic data 134 obtained by executing a process of extracting the periodic component of the change in the water level from the time-series data 132 (details will be described later) into the learned model 136, thereby obtaining the future value of the water level.
[0017] The time-series data 132 represents, for example, the water level values measured by the water level gauge 10 at one-hour intervals. Among them, for example, the water level data at one-hour intervals obtained up to 1 to 6 hours before is input as input data into the learned model 136. Further, the periodic data 134 is time-series data of the water level obtained by executing a process of extracting the periodic component from the time-series data 132 collected over a predetermined past period (for example, several years). The output data is, for example, the future value of the water level at one-hour intervals up to 1 to 6 hours later. The settings such as the water level values 1 to 6 hours before to be input and the water level values 1 to 6 hours later to be predicted are merely examples, and the input and output of the water level values may be performed at shorter intervals (for example, 30-minute intervals) or longer intervals (for example, 1.5-hour intervals).
[0018] The learned model 136 is a long short-term memory (LSTM) architecture that can handle abrupt changes while memorizing the trends of long-term continuous data. As a water level prediction model, it is generated using the time series data 132 and periodic data 134 corresponding to the above-described inputs and outputs as teacher data. That is, in the present embodiment, the learned model 136 is learned by taking into account not only the time series data 132 of the water level but also the periodic data 134 of the water level. As a result, compared with the conventional water level prediction technology using a DNN, even if the lead time is long, the water levels of the regulating pond and the river can be predicted with high accuracy. Hereinafter, the functions of the prediction device 100 will be described in more detail.
[0019] [Functional Configuration] The prediction device 100 includes, for example, a data acquisition unit 110, a water level prediction unit 120, and a storage unit 130. The data acquisition unit 110 and the water level prediction unit 120 are each 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 in a storage device such as an HDD (Hard Disk Drive) 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 or CD-ROM (a non-transient storage medium) and installed when the storage medium is mounted in a drive device. The memory unit 130 is implemented by, for example, ROM, flash memory, SD card, RAM, HDD, registers, etc. The memory unit 130 stores, for example, time-series data 132, periodic data 134, and a trained model 136.
[0020] The data acquisition unit 110 receives time-series data of the water level of the regulating reservoir measured by the water level gauge 10 from the transmitter 20 via the network NW and stores it in the storage unit 130 as time-series data 132. The data acquisition unit 110 further extracts the periodic component of the water level change from the latest stored time-series data 132, more specifically by performing a Fourier transform (and inverse Fourier transform) to obtain periodic data 134.
[0021] Figure 3 is a diagram illustrating the overview of the method for acquiring periodic data 134. In the graph of time-series data 132 shown in the upper left of Figure 3, the horizontal axis represents time and the vertical axis represents water level. The data acquisition unit 110 first performs a Fourier transform on the time-series data 132 according to the following equation (1).
[0022]
number
[0023] In equation (1) above, f represents a function of water level over time in the time series data 132, N represents the number of data points in the time series data 132 to be subjected to the Fourier transform, ω represents the frequency, and t represents time. Equation (1) represents a calculation formula using the discrete Fourier transform, but the fast Fourier transform is used in actual calculations.
[0024] Figure 4 shows an example of the results obtained by applying a Fourier transform to time-series data 132. In the graph shown in Figure 4, the horizontal axis represents frequency and the vertical axis represents spectral density. The data acquisition unit 110 extracts frequencies with a spectral density greater than or equal to a threshold Th as peak values, while removing frequencies with a spectral density less than the threshold Th as noise. Alternatively, the data acquisition unit 110 may pre-define a range of specific frequencies and extract the spectral density within that range as peak values. For example, in Figure 4, the data acquisition unit 110 extracts frequency ω1 corresponding to the diurnal period and frequency ω2 corresponding to the semi-diurnal period. However, based on prior verification, a first range R1 including frequency ω1 and a second range R2 including frequency ω2 may be pre-defined, and frequencies within these ranges may be extracted as peak values while other frequencies are removed as noise.
[0025] The data acquisition unit 110 extracts peak values of spectral density, and then performs an inverse Fourier transform on the extracted peak values according to the following equation (2).
[0026]
number
[0027] In equation (2) above, F represents a function indicating the spectral density obtained by the Fourier transform in equation (1), N represents the number of data points in the time series data 132 to which the inverse Fourier transform is applied, ω represents the frequency, and t represents time. The data acquisition unit 110 acquires periodic data 134 as a result of applying the inverse Fourier transform.
[0028] Figure 5 shows an example of periodic data 134 obtained by the inverse Fourier transform. Of the periodic data 134 shown in Figure 5, the upper graph represents the water level on a daily cycle corresponding to frequency ω1, and the lower graph represents the water level on a semi-daily cycle corresponding to frequency ω2. The data acquisition unit 110 obtains the periodic data 134 to be input to the trained model 136 in this manner.
[0029] When the water level prediction unit 120 acquires time-series data 132 and periodicity data 134, it inputs this data into a trained model 136 to obtain future water level values. As described above, in this embodiment, the water level prediction unit 120 acquires future water level values at one-hour intervals from 1 to 6 hours later, for example. In this way, by predicting future water level values using a trained model 136 that has been trained to take into account not only the time-series water level data 132 but also the periodicity data 134 of the water level, it is possible to predict the water level of the retention pond or river with high accuracy even if the lead time is long.
[0030] In the above description, the data acquisition unit 110 derives periodic data 134 for input to the trained model 136 based on the latest time-series data 132 stored in the storage unit 130, but the present invention is not limited to such a configuration. For example, the data acquisition unit 110 may derive periodic data 134 periodically (for example, daily), and when predicting water levels, it may input the periodic data 134 derived at the previous timing (for example, the previous day) to the trained model 136.
[0031] Next, with reference to Figure 6, the processing flow performed by the prediction device 100 will be described. Figure 6 is a flowchart showing an example of the processing flow performed by the prediction device 100. The processing described in this flowchart may be performed in a predetermined control cycle (for example, every hour), or it may be performed in response to a command from a terminal device held by a user of the prediction device 100.
[0032] First, the data acquisition unit 110 receives time-series data 132 of the water level of the regulating reservoir measured by the water level gauge 10 from the transmitter 20 via the network NW (step S100). Next, the data acquisition unit 110 extracts the periodic component of the water level change from the time-series data 132 of the water level, more specifically, by performing a Fourier transform (and inverse Fourier transform) to obtain periodic data 134 of the water level (step S102). Next, the water level prediction unit 120 inputs the time-series data 132 and the periodic data 134 into the trained model 136 (step S104). Next, the water level prediction unit 120 obtains the future values of the water level output by the trained model 136 and predicts the water level of the regulating reservoir (step S106). This completes the processing of this flowchart.
[0033] [Learning device] Next, with reference to Figure 7, a learning device 200 that generates the trained model 136 of this embodiment will be described. Figure 7 is a diagram showing an example of the configuration of the learning device 200. As shown in Figure 7, the learning device 200 includes, for example, a learning unit 210 and a storage unit 220. The storage unit 220 stores, for example, training data 222 and an LSTM model 224. The learning unit 210 is realized, for example, by a hardware processor such as a CPU executing a program (software). Part or all of the learning unit 210 may 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 equipped with a non-transient storage medium), or it may be stored in a removable storage medium such as a DVD or CD-ROM (a non-transient storage medium) and installed when the storage medium is mounted on a drive device. The memory unit 220 can be implemented by, for example, ROM, flash memory, SD card, RAM, HDD, registers, etc.
[0034] The training data 222 includes time-series water level data 132 and periodicity data 134 over a predetermined period (e.g., several years) derived from the time-series data 132. The learning unit 210 learns the LSTM model 224 so that when it receives, for example, hourly water level data acquired from 1 to 6 hours prior and the periodicity data 134 over the predetermined period from the time-series data 132, it outputs future values of the water level at hourly intervals from 1 to 6 hours later. More specifically, the learning unit 210 learns the LSTM model 224 by adjusting its parameters, for example using stochastic gradient descent, to optimize the sum of squared errors between predicted and observed water levels as the loss function. This generates the trained model 136.
[0035] In this embodiment, the prediction device 100 and the learning device 200 are configured as separate devices. However, the present invention is not limited to such a configuration, and the prediction device 100 and the learning device 200 may be configured as a single information processing device. In that case, the information processing device will include all the functional parts of the prediction device 100 and the learning device 200 described above.
[0036] As described above, according to this embodiment, time-series data of water level measured by a water level gauge is acquired, and input data based on the time-series data is input to a trained model that has been trained to output future values of water level when input data based on time-series data of water level is input, thereby predicting future values of water level. The input data includes at least periodic data obtained by performing a process to extract periodic components from the time-series data. As a result, even if the lead time is long, the water level of the retention pond or river can be predicted with high accuracy.
[0037] [Differentiation] In the above embodiment, the data acquisition unit 110 obtains periodic data 134 by applying a Fourier transform (and inverse Fourier transform) to the time-series water level data 132 obtained up to the present time, and predicts future water level values by inputting the time-series data 132 and the periodic data 134 into a trained model 136. However, the present invention is not limited to such a configuration, and the periodic data 134 may include periodic data in future intervals.
[0038] Figure 8 shows an example of periodic data 134 related to the modified example. First, the data acquisition unit 110 applies a Fourier transform to the time-series data 132 up to the present time using the same method as described above to extract each peak value of the spectral density, and then applies an inverse Fourier transform to the extracted peak values to obtain simple harmonic motion of the water level with respect to time. Next, the data acquisition unit 110 generates simple harmonic motion in the future interval by shifting (extending) the waveform of the simple harmonic motion obtained by the inverse Fourier transform in the time direction. The data acquisition unit 110 can generate periodic data 134 related to the modified example by combining the simple harmonic motion from the past to the present time and the simple harmonic motion in the future interval.
[0039] If, as a result of extracting multiple peak values from the spectral diagram, the period is a group of multiple frequencies, the data acquisition unit 110 can extend each simple harmonic motion generated from the peak values corresponding to the multiple frequencies to a future interval and simply superimpose them linearly to generate periodicity data 134. By using the periodicity data 134 thus generated as training data 222 to generate a trained model 136, the water level prediction unit 120 can predict future water levels by considering the future periodicity of the water level.
[0040] 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]
[0041] 10 Water level gauge 20 Transmitters 100 Prediction Devices 110 Data Acquisition Unit 120 Water Level Prediction Section 130, 220 storage section 132 Time Series Data 134 Periodic Data 136 pre-trained models 200 Learning Devices 210 Learning Department 222 Teacher data 224 LSTM model P Pipe NW Network
Claims
1. An acquisition unit that acquires time-series data of water level measured by a water level gauge, The system includes a prediction unit that predicts future values of the water level by inputting input data based on the aforementioned time series data into a trained model that has been trained to output future values of the water level when input data based on time series data of the water level is input, The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. Prediction device.
2. An acquisition unit that acquires time-series data of the water level measured by a water level gauge, The system includes a prediction unit that predicts future values of the water level by inputting input data based on the aforementioned time series data into a trained model that has been trained to output future values of the water level when input data based on time series data of the water level is input, The input data includes at least periodic data obtained by performing a process to extract the periodic component of the water level change from the time series data, The process for extracting the periodic component of the water level change involves applying a Fourier transform to the time series data to extract peak values, and then applying an inverse Fourier transform to the extracted peak values to obtain the periodic data. Prediction device.
3. An acquisition unit that acquires time-series data of the water level measured by a water level gauge, The system includes a prediction unit that predicts future values of the water level by inputting input data based on the aforementioned time series data into a trained model that has been trained to output future values of the water level when input data based on time series data of the water level is input, The input data includes at least periodic data obtained by performing a process to extract the periodic component of the water level change from the time series data, The process for extracting the periodic component of the water level change involves applying a Fourier transform to the time series data to extract peak values, applying an inverse Fourier transform to the extracted peak values, and extending the data obtained from the inverse Fourier transform over a future interval according to the periodic component to obtain the periodic data. Prediction device.
4. An acquisition unit that acquires time-series data of the water level measured by a water level gauge, The system includes a prediction unit that predicts future values of the water level by inputting input data based on the aforementioned time series data into a trained model that has been trained to output future values of the water level when input data based on time series data of the water level is input, The input data includes at least periodic data obtained by performing a process to extract the periodic component of the water level change from the time series data, The input data includes the time-series data and the periodic data. Prediction device.
5. Computers Time-series data of water levels measured by water level gauges is acquired. The future value of the water level is predicted by inputting the input data based on the aforementioned time series data into a trained model that has been trained to output the future value of the water level when input data based on the time series data of the water level is input. The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. Prediction method.
6. On the computer, Time-series data of water levels measured by a water level gauge is acquired. By inputting the input data based on the aforementioned time series data into a trained model that has been trained to output future values of the water level when input data based on the water level time series data is input, the future values of the water level are predicted. The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. program.
7. A learning device that, upon receiving input data based on time-series data of water level measured by a water level gauge, learns to output future values of the water level, The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. Learning device.
8. Computers A learning method that, upon receiving input data based on time-series data of water level measured by a water level gauge, learns to output future values of the water level, The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. Learning methods.
9. On the computer, A program that, upon receiving input data based on time-series data of water levels measured by a water level gauge, is trained to output future values of the water level, The input data includes at least periodic data obtained by performing a process to extract periodic components of the water level change from the time series data that are unique to the time series data. program.