Estimation device, estimation method, and program

Abstract

An estimation device 1 is provided with: a conversion unit 22 that converts the confidence level of each pixel of spatial observation data into a weighting factor; a training unit 23 that trains, using physical quantities relating to meteorological phenomena at times later than the time when the observation data was obtained as training data, a prediction model that predicts a future physical quantity from spatiotemporal information of observation data, the observation value of each pixel of the observation data, and the weighting factor for each pixel; and an estimation unit 24 that enters spatiotemporal information of new observation data, the observation value of each pixel of the new observation data, and the weighting factor for each pixel of the new observation data into the prediction model to estimate a physical quantity. The weighting factor for each pixel of the new observation data is obtained by the conversion unit 22 converting the confidence level of the pixel of the new observation data.

Description

Estimation device, estimation method, and program

[0001] The present disclosure relates to a learning device, a learning method, and a program.

[0002] A surrogate model is known (see Non-Patent Document 1). The surrogate model estimates the probability distribution of future precipitation using satellite images and radar images as input.

[0003] Ensemble forecasting is a type of numerical forecasting based on conventional physical models. Ensemble forecasting involves making predictions based on perturbed initial values ​​to account for the uncertainty of the initial values, and predicting one potential future outcome from among the possible outcomes.

[0004] On the other hand, Non-Patent Document 1 significantly reduces the amount of calculation by directly estimating the probability distribution, specifically by taking into account the uncertainty due to the initial value.

[0005] Casper Kaae Sonderby, et al., “MetNet: A Neural Weather Model for Precipitation Forecasting”, https: / / arxiv.org / abs / 2003.12140

[0006] However, there is a problem in that the uncertainty in the observation data is not taken into account in the training and inference of the surrogate model.

[0007] The present disclosure has been made in consideration of the above circumstances, and an object of the present disclosure is to provide a technology that can predict future weather phenomena while taking into account uncertainty in observation data.

[0008] An estimation device according to one aspect of the present disclosure includes a conversion unit that converts the reliability of each pixel of spatial observation data into a weighting coefficient; a learning unit that uses physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as teacher data, and learns a prediction model that predicts future physical quantities from spatiotemporal information of the observation data, the observation values ​​of each pixel of the observation data, and the weighting coefficients of each pixel; and an estimation unit that inputs the spatiotemporal information of new observation data, the observation values ​​of each pixel of the new observation data, and the weighting coefficients of each pixel of the new observation data into the prediction model to estimate the physical quantities, and the weighting coefficients of each pixel of the new observation data are converted by the conversion unit from the reliability of each pixel of the new observation data.

[0009] In one aspect of the present disclosure, an estimation method includes a computer: converting the reliability of each pixel of spatial observation data into a weighting coefficient; using physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as training data; learning a prediction model that predicts future physical quantities from spatiotemporal information of the observation data, the observation values ​​of each pixel of the observation data, and the weighting coefficients for each pixel; converting the reliability of each pixel of new observation data into a weighting coefficient; and inputting the spatiotemporal information of the new observation data, the observation values ​​of each pixel of the new observation data, and the weighting coefficients for each pixel of the new observation data into the prediction model to estimate the physical quantities.

[0010] A program according to one aspect of the present disclosure causes a computer to execute the following steps: converting the reliability of each pixel of spatial observation data into a weighting coefficient; learning a prediction model that predicts future physical quantities from spatiotemporal information of the observation data, the observation values ​​of each pixel of the observation data, and the weighting coefficients of each pixel, using physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as training data; converting the reliability of each pixel of new observation data into a weighting coefficient; and inputting the spatiotemporal information of the new observation data, the observation values ​​of each pixel of the new observation data, and the weighting coefficients of each pixel of the new observation data into the prediction model to estimate the physical quantities.

[0011] According to the present disclosure, it is possible to provide a technology that can predict future weather phenomena by taking into account uncertainties in observation data.

[0012] Fig. 1 is a diagram illustrating functional blocks of an estimation device according to the present disclosure. Fig. 2 is a flowchart illustrating an example of a learning process during learning. Fig. 3 is a flowchart illustrating an example of an estimation process during estimation. Fig. 4 is a diagram illustrating the hardware configuration of a computer used in the estimation device.

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the description of the drawings, the same parts are designated by the same reference numerals and the description thereof will be omitted.

[0014] (Estimation Device) As shown in Fig. 1, an estimation device 1 according to the present disclosure learns a prediction model that estimates meteorological phenomena using a surrogate model, taking into account the reliability of observation data. The estimation device 1 calculates weighting coefficients for observation locations based on the reliability of the observation data, inputs the weighting coefficients into a deep neural network, and learns the prediction model.

[0015] 1 , the estimation device 1 includes prediction model data 11 and estimation data 12, and functions of an input unit 21, a conversion unit 22, a learning unit 23, and an estimation unit 24. Each piece of data is stored in a storage device such as a memory 902 or a storage 903. Each function is implemented in a CPU 901.

[0016] The prediction model data 11 is data that is learned by the learning unit 23 and identifies a model that predicts physical quantities of future meteorological phenomena from existing observation data. The prediction model data 11 may be, for example, a set of variables of a trained deep neural network. The prediction model data 11 may estimate physical quantities for each region and estimate a prediction probability distribution. In the present disclosure, physical quantities are values ​​that identify meteorological phenomena, such as precipitation, temperature, humidity, and ultraviolet radiation.

[0017] The estimated data 12 is data on physical quantities of future meteorological phenomena estimated from existing observation data using the prediction model data 11. For observation data at time Tx, the estimated data 12 is a predicted probability distribution for time Ty, which is later than time Tx.

[0018] The input unit 21 inputs data required for each process to the learning unit 23 and the estimation unit 24 .

[0019] The input unit 21 inputs the physical quantity at time Ty as teacher data to the learning unit 23. The input unit 21 inputs (i) spatiotemporal information of the observation data, (ii) the observation data at time Tx, and (iii) the reliability of the observation data at time Tx, which are referred to by the learning unit 23 or the estimation unit 24, to the conversion unit 22.

[0020] In the present disclosure, observation data is data that associates an observed space with a physical quantity in that space. In the present disclosure, one unit of the observed space is referred to as a pixel. The spatiotemporal information identifies the space and time at which the observation data was acquired. The time Ty is in the future than the time Tx. The observation data includes the physical quantity of each pixel. The reliability of the observation data includes the reliability of each pixel of the observation data.

[0021] In the present disclosure, reliability is assigned to each pixel constituting the observation data. The reliability is determined based on the reliability or error of a physical quantity in the observation data. The reliability of each pixel of the observation data is determined, for example, based on the observation method of the observation data. Alternatively, the reliability of each pixel of the observation data is determined, for example, based on the meteorological phenomenon observed by the observation data.

[0022] The reliability of each pixel is, for example, QA-flag in the case of JAXA's (Japan Aerospace Exploration Agency) climate change observation satellite "Shikisai" (GCOM-C: Global Change Observation Mission - C) (https: / / shikisai.jaxa.jp / faq / faq0031_j.html).

[0023] Alternatively, observation data obtained by an X-band multi-parameter radar may be unavailable at locations far from the radar's position or farther away than the area of ​​heavy rainfall in the direction of the radar's illumination. Taking these circumstances into consideration, the reliability of each pixel is set based on the likelihood or likelihood of errors occurring (https: / / www.jstage.jst.go.jp / article / jjshwr / 22 / 5 / 22_5_372 / _pdf / -char / ja).

[0024] The conversion unit 22 converts the reliability of each pixel of the spatial observation data into a weighting coefficient. During learning, the conversion unit 22 converts the reliability of each pixel of the observation data into a weighting coefficient for the observation point of each pixel, and inputs the converted coefficient to the learning unit 23. During estimation, the conversion unit 22 converts the reliability of each pixel of the observation data into a weighting coefficient for the observation point of each pixel, and inputs the converted coefficient to the estimation unit 24.

[0025] The converter 22 normalizes the reliability of each observation value and converts it into a weighting coefficient that has a positive correlation with the normalized reliability. The converter 22 calculates a normalized weighting coefficient for the reliability of each pixel of the observation data so that a weighting coefficient of 0 is assigned to a completely uncertain observation value and a weighting coefficient of 1 is assigned to a completely certain observation value.

[0026] Furthermore, when acquiring multiple pieces of observation data observed using multiple observation methods, the conversion unit 22 converts coordinates associated with each pixel of each piece of observation data into common coordinates and inputs them to the learning unit. The coordinates that identify each pixel in the observation data differ depending on the observation method. In order to absorb the difference in coordinates due to the observation method, the conversion unit 22 converts the coordinates of each pixel in each piece of observation data into common coordinates. The conversion unit 22 calculates an observation value and a weighting coefficient for each pixel in the common coordinates for each piece of observation data.

[0027] The conversion unit 22 inputs the weighted observation data to the learning unit 23 or the estimation unit 24. The weighted observation data is data that associates the spatiotemporal information of the observation data, the observation value of each pixel of the observation data, and the weighting coefficient of each pixel.

[0028] The learning unit 23 uses the physical quantities related to meteorological phenomena at a time Ty in the future from the time Tx at which the observation data was acquired as training data, and learns a prediction model that predicts future physical quantities from the spatiotemporal information of the observation data, the observation values ​​of each pixel of the observation data, and the weighting coefficients of each pixel.

[0029] The learning unit 23 adjusts the variables of the deep neural network so that the difference between the physical quantity at time Ty predicted from the weighted observation data input from the conversion unit 22 and the physical quantity at time Ty input from the input unit 21 as teacher data is minimized.

[0030] The learning unit 23, for example, repeats the adjustment of the variables of the deep neural network until a predetermined condition is satisfied, and outputs the prediction model data 11. The predetermined condition may be, for example, that the difference between the predicted physical quantity and the physical quantity of the teacher data is equal to or less than a predetermined value, that the adjustment of the variables is repeated a predetermined number of times, that the adjustment of the variables is repeated for a predetermined time, etc.

[0031] Once the prediction model has been learned by the learning unit 23, future physical quantities are estimated using this prediction model. When new observation data is acquired, the estimation unit 24 inputs weighted observation data for the new observation data to the estimation unit 24. The weighted observation data for the new observation data is data including spatiotemporal information of the new observation data, the observation value of each pixel of the new observation data, and a weighting coefficient for each pixel of the new observation data.

[0032] The estimation unit 24 inputs the spatiotemporal information of the new observation data, the observation value of each pixel of the new observation data, and the weighting coefficient of each pixel of the new observation data into a prediction model to estimate future physical quantities. The new observation data is the observation data used for estimation. The weighting coefficient of each pixel of the new observation data is converted by the conversion unit 22 from the reliability of each pixel of the new observation data. The estimation unit 24 outputs estimation data 12 that specifies the estimated future physical quantities.

[0033] The processing of the estimation device 1 during learning in the estimation method according to the present disclosure will be described with reference to Fig. 2. The content and procedure of the processing shown in Fig. 2 are an example and are not limited to this.

[0034] The estimation device 1 performs the processes of steps S11 and S12 for each piece of observation data used for learning. In step S11, the estimation device 1 converts the reliability of each pixel of the observation data to be processed into a weighting coefficient. In step S12, the estimation device 1 converts the coordinates of each pixel into common coordinates.

[0035] When the processes of steps S11 and S12 for each observation data used in learning are completed, the estimation device 1 proceeds to step S13. In step S13, the estimation device 1 inputs the weighted observation data and the teacher data to the learning unit 23.

[0036] The processes of steps S14 and S15 are repeated until the convergence condition is satisfied. In step S14, the estimation device 1 inputs the weighted observation data to the current deep neural network and calculates the difference between the predicted physical quantity and the teacher data. In step S15, the estimation device 1 updates the variables of the deep neural network so as to reduce the difference calculated in step S14.

[0037] If the convergence condition is satisfied, the estimation device 1 proceeds to step S16. In step S16, the estimation device 1 outputs the prediction model obtained as a result of the processes in steps S14 and S15.

[0038] The processing of the estimation device 1 during estimation in the estimation method according to the present disclosure will be described with reference to Fig. 3. The content and procedure of the processing shown in Fig. 3 are an example and are not limited to this.

[0039] The estimation device 1 performs the processes of steps S21 and S22 for each piece of observation data used for estimation. In step S21, the estimation device 1 converts the reliability of each pixel of the observation data to be processed into a weighting coefficient. In step S22, the estimation device 1 converts the coordinates of each pixel into common coordinates.

[0040] When the processes of steps S21 and S22 are completed for each observation data used for estimation, the estimation device 1 proceeds to step S23. In step S23, the estimation device 1 inputs the weighted observation data into the prediction model output by the process of FIG.

[0041] In step S24, the estimation device 1 inputs the weighted observation data used for estimation into a prediction model and outputs the estimated data 12 obtained.

[0042] The estimation device 1 according to the present disclosure performs learning in consideration of the influence of the reliability of observation data. The estimation device 1 can more accurately reflect inaccuracies in the real world in the model, thereby improving the reliability of the model.

[0043] The estimation device 1 according to the present disclosure can efficiently train a model by using the reliability of observation data, which allows the estimation device 1 to reduce the computational cost of training, and is therefore expected to enable real-time prediction or optimization.

[0044] In this way, the estimation device 1 according to the present disclosure can predict future meteorological phenomena taking into account the uncertainty in the observation data.

[0045] The estimation device 1 of the present embodiment described above uses, for example, a general-purpose computer system including a CPU (Central Processing Unit, processor) 901, a memory 902, a storage 903 (HDD: Hard Disk Drive, SSD: Solid State Drive), a communication device 904, an input device 905, and an output device 906. In this computer system, the CPU 901 executes a program loaded on the memory 902, thereby realizing each function of the estimation device 1.

[0046] The estimation device 1 may be implemented by one computer or by multiple computers. The estimation device 1 may also be a virtual machine implemented on a computer.

[0047] The program of the estimation device 1 can be stored in a computer-readable recording medium such as a HDD, an SSD, a Universal Serial Bus (USB) memory, a Compact Disc (CD), or a Digital Versatile Disc (DVD), or can be distributed via a network. The computer-readable recording medium is, for example, a non-transitory recording medium.

[0048] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the present disclosure.

[0049] REFERENCE SIGNS LIST 1 Estimation device 11 Prediction model data 12 Estimation data 21 Input unit 22 Conversion unit 23 Learning unit 24 Estimation unit 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device

Claims

1. An estimation device comprising: a conversion unit that converts the reliability of each pixel of spatial observation data into a weighting coefficient; a learning unit that uses physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as teacher data, and learns a prediction model that predicts future physical quantities from spatiotemporal information of the observation data, the observed values ​​of each pixel of the observation data, and the weighting coefficients of each pixel; and an estimation unit that inputs the spatiotemporal information of new observation data, the observed values ​​of each pixel of the new observation data, and the weighting coefficients of each pixel of the new observation data into the prediction model to estimate the physical quantities, wherein the weighting coefficients of each pixel of the new observation data are converted by the conversion unit from the reliability of each pixel of the new observation data.

2. The estimation device according to claim 1, wherein the conversion unit normalizes the reliability of each observed value and converts it into a weighting coefficient that has a positive correlation with the normalized reliability.

3. The estimation device according to claim 1, wherein when the conversion unit acquires multiple pieces of observation data observed using multiple observation methods, the conversion unit converts coordinates associated with each pixel of each piece of observation data into common coordinates and inputs them to the learning unit.

4. The estimation device according to claim 1, wherein the reliability of each pixel of the observation data is determined from the observation method of the observation data.

5. The estimation device according to claim 1, wherein the reliability of each pixel of the observation data is determined from the meteorological phenomenon observed by the observation data.

6. An estimation method in which a computer converts the reliability of each pixel of spatial observation data into a weighting coefficient, learns a prediction model that predicts future physical quantities from the spatiotemporal information of the observation data, the observed value of each pixel of the observation data, and the weighting coefficient of each pixel, using physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as training data, converts the reliability of each pixel of new observation data into a weighting coefficient, and inputs the spatiotemporal information of the new observation data, the observed value of each pixel of the new observation data, and the weighting coefficient of each pixel of the new observation data into the prediction model to estimate the physical quantities.

7. A program for causing a computer to execute the steps of: converting the reliability of each pixel of spatial observation data into a weighting coefficient; using physical quantities related to meteorological phenomena at times in the future from the time the observation data was acquired as training data, learning a prediction model that predicts future physical quantities from the spatiotemporal information of the observation data, the observation values ​​of each pixel of the observation data, and the weighting coefficients of each pixel; converting the reliability of each pixel of new observation data into a weighting coefficient; and inputting the spatiotemporal information of the new observation data, the observation values ​​of each pixel of the new observation data, and the weighting coefficients of each pixel of the new observation data into the prediction model to estimate the physical quantities.

Images