Estimation device, estimation method, and program

Abstract

This estimation device 1 is provided with: a conversion unit 22 that downscales a state variable used for prediction in each model, and a prediction result of a future time predicted by each model to a prescribed resolution; a learning unit 23 that performs learning of a prediction model for predicting a predicted value of future from the state variable being downscaled, using the prediction result being downscaled as supervisory data; and an estimation unit 24 that enters a new state variable, which have been downscaled to the prescribed resolution, into the prediction model to estimate a predicted value.

Description

Estimation device, estimation method, and program

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

[0002] In weather forecasting, numerical prediction models with different spatial scales are constructed according to the phenomena to be predicted (Non-Patent Documents 1 and 2). Here, spatial scale refers to the size or resolution of the area to be predicted.

[0003] High-resolution models require a large amount of calculation, so by using a model with an appropriate resolution for the phenomenon you want to predict, you can optimize calculation costs and obtain the required accuracy. In addition, by making optimal use of computational resources, you can achieve both improved forecast accuracy and reduced calculation volume.

[0004] On the other hand, when using models with different resolutions, integrating these models becomes more complex, and a method for appropriately combining data or forecast results with different resolutions is required.

[0005] There is a technique for downscaling state variables and the like input to a prediction model using long-term modification technology (Non-Patent Document 3).

[0006] "Japan Meteorological Agency, Volume 51 (2018), Fundamentals of the 10th Generation Numerical Analysis and Forecasting System and Numerical Prediction, Chapter 4: Numerical Prediction Models," [online], Japan Meteorological Agency, [Retrieved March 25, 2024], Internet <URL: https: / / www.jma.go.jp / jma / kishou / books / nwptext / 51 / 2_chapter4.pdf> "Types of Numerical Weather Prediction Models," [online], Japan Meteorological Agency, [Retrieved March 25, 2024], Internet <URL: https: / / www.jma.go.jp / jma / kishou / know / whitep / 1-3-4.html> Sekiyama, TT, "Statistical Downscaling of Temperature Distributions from the Synoptic Scale to the Mesoscale Using Deep Convolutional Neural Networks," July 22, 2020

[0007] However, none of the non-patent documents discloses or suggests a technique for appropriately combining models with different resolutions.

[0008] The present disclosure has been made in consideration of the above circumstances, and an object of the present disclosure is to provide a technology for appropriately combining models with different resolutions.

[0009] An estimation device according to one aspect of the present disclosure includes a conversion unit that downscales the state variables used for prediction in each model and the prediction results for future times predicted in each model to a predetermined resolution, a learning unit that uses the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables, and an estimation unit that inputs the new state variables downscaled to the predetermined resolution into the prediction model and estimates a predicted value.

[0010] In an estimation method according to one aspect of the present disclosure, a computer downscales the state variables used for prediction in each model and the prediction results for future times predicted in each model to a predetermined resolution, uses the downscaled prediction results as training data, learns a prediction model that predicts future predicted values ​​from the downscaled state variables, and inputs the new state variables downscaled to the predetermined resolution into the prediction model to estimate predicted values.

[0011] A program according to one aspect of the present disclosure causes a computer to execute a procedure of downscaling the state variables used for prediction in each model and the prediction results for future times predicted in each model to a predetermined resolution, using the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables, and inputting the new state variables downscaled to the predetermined resolution into the prediction model to estimate the predicted values.

[0012] According to the present disclosure, it is possible to provide a technique for appropriately combining models with different resolutions.

[0013] 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.

[0014] 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.

[0015] (Estimation Device) As shown in FIG. 1 , an estimation device 1 according to the present disclosure integrates state variables used in predictions in each model and prediction results for future times predicted by each model to learn a prediction model that estimates weather phenomena using a surrogate model. The estimation device 1 downscales the state variables used in predictions in each model and prediction results for future times predicted by each model to a predetermined resolution. This allows the estimation device 1 to appropriately combine models with different resolutions.

[0016] 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.

[0017] The prediction model data 11 is data learned by the learning unit 23 and identifies a model that predicts future weather phenomena from existing state variables. The prediction model data 11 may be, for example, a set of variables of a trained deep neural network. The prediction model data 11 is generated by integrating multiple prediction models. In the present disclosure, a state variable is a value that identifies a weather phenomenon, such as precipitation, temperature, humidity, or UV radiation. The prediction result may be a value of a future state variable, or a weather phenomenon identified from one or more future state variables.

[0018] The estimated data 12 is data of the prediction results of future weather phenomena estimated from existing state variables using the prediction model data 11. For state variables at time Tx, the estimated data 12 is the prediction result for time Ty, which is later than time Tx.

[0019] The input unit 21 inputs data required for each process to the learning unit 23 and the estimation unit 24 .

[0020] The input unit 21 inputs the state variables used for prediction in each model and the prediction results for future times predicted in each model to the conversion unit 22. In the example shown in Fig. 1, a case will be described in which the input unit 21 inputs the state variables and prediction results for three prediction models to the conversion unit 22, but this is not limiting. The input unit 21 may input the state variables and prediction results for two or more prediction models to the conversion unit 22.

[0021] The input unit 21 inputs, for one prediction model, the state variables at time Tx and the prediction result at time Ty to the conversion unit 22. Here, time Ty is in the future than time Tx.

[0022] For each model, the state variables used for prediction and the prediction results for future times predicted by each model may have different resolutions. Here, the resolution corresponds to the size of a region when a state variable or prediction result is given for a region. In this disclosure, reducing the region is referred to as downscaling or conversion using super-resolution technology.

[0023] In the present disclosure, the size of the area in which the state variables and prediction results are given in the first prediction model, the size of the area in which the state variables and prediction results are given in the second prediction model, and the size of the area in which the state variables and prediction results are given in the third prediction model may be different from each other.

[0024] The converter 22 downscales the state variables used in the predictions in each model and the prediction results for future times predicted by each model to a predetermined resolution. Here, the predetermined resolution is the highest resolution among the resolutions of the models. The converter 22 converts the state variables and prediction results of each model into state variables and prediction results of the predetermined resolution, for example, using the super-resolution technology described in Non-Patent Document 3.

[0025] The conversion unit 22 inputs the prediction results of each model at time Ty after downscaling as training data to the learning unit 23. The conversion unit 22 inputs the state variables of each model at time Tx after downscaling as training data to the learning unit 23.

[0026] The learning unit 23 uses the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables.

[0027] The learning unit 23 adjusts the variables of the deep neural network so that the difference between the prediction result at time Ty predicted from the state variables downscaled by the conversion unit 22 and the prediction result at time Ty downscaled by the conversion unit 22 is minimized.

[0028] 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 prediction result and the prediction result 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.

[0029] Once the prediction model has been learned by the learning unit 23, future prediction results are estimated using this prediction model.

[0030] The estimation unit 24 obtains new state variables at time Tx that have been downscaled by the conversion unit 22. Here, the conversion unit 22 downscales the state variables and prediction results of each model so that they have the same resolution as the state variables and prediction results input to the learning unit 23.

[0031] The estimation unit 24 inputs the state variables at time Tx downscaled by the conversion unit 22 into a prediction model to estimate future prediction results. The new state variables at time Tx are data used for estimation. The estimation unit 24 outputs estimation data 12 that specifies the estimated future prediction results.

[0032] 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.

[0033] The estimation device 1 performs the process of step S11 for each model integrated in the learning. In step S11, the estimation device 1 downscales the state variables and prediction results for the model to be processed to a predetermined resolution.

[0034] When the processing of step S11 for each model to be integrated is completed, the estimation device 1 proceeds to step S12. In step S12, the estimation device 1 inputs the state variables of each downscaled model as training data to the learning unit 23. In step S13, the estimation device 1 inputs the prediction results of each downscaled model as training data to the learning unit 23.

[0035] The processes of steps S14 and S15 are repeated until the convergence condition is satisfied. In step S14, the estimation device 1 calculates the difference between the prediction result obtained by inputting the downscaled state variables into the current deep neural network and the downscaled prediction result input as training 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.

[0036] 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.

[0037] 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.

[0038] The estimation device 1 performs the process of step S21 for each model used for estimation. In step S21, the estimation device 1 downscales the state variables and prediction results for the model to be processed to a predetermined resolution.

[0039] When the process of step S21 for each model used for estimation is completed, the estimation device 1 proceeds to step S22. In step S22, the estimation device 1 inputs the state variables of each downscaled model into the prediction model output by the process of FIG.

[0040] In step S23, the estimation device 1 outputs the estimation data 12 obtained by inputting the state variables of each downscaled model used for estimation into the prediction model.

[0041] In the estimation device 1 according to the present disclosure, the conversion unit 22 standardizes the state variables and prediction results of each model to a predetermined resolution, thereby realizing faster and more stable learning and inference processes of the surrogate model. Specifically, since the state variables and prediction results of each model have a common resolution, calculation time can be reduced by using a graphics processing unit (GPU) or the like with high computing performance.

[0042] Here, since the state variables and prediction results of each model are downscaled, it is conceivable that the amount of calculation in the learning process and inference process will increase as the number of regions increases. However, it is expected that the amount of calculation in the learning unit 23 or the estimation unit 24 can be reduced by, for example, upscaling the resolution of regions that do not affect the prediction accuracy in the learning unit 23 or the estimation unit 24.

[0043] In this way, the estimation device 1 according to the present disclosure can appropriately combine models with different resolutions.

[0044] 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.

[0045] The estimation device 1 may be implemented by one computer or by multiple computers, or may be a virtual machine implemented on a computer.

[0046] 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.

[0047] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the present disclosure.

[0048] 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 downscales, to a predetermined resolution, state variables used for prediction in each model and the prediction results for future times predicted in each model; a learning unit that uses the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables; and an estimation unit that inputs the new state variables downscaled to the predetermined resolution into the prediction model and estimates predicted values.

2. The estimation device according to claim 1, wherein the predetermined resolution is the highest resolution among the resolutions of the models.

3. An estimation method in which a computer downscales the state variables used for prediction in each model and the prediction results for future times predicted in each model to a predetermined resolution, uses the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables, and inputs the new state variables downscaled to the predetermined resolution into the prediction model to estimate predicted values.

4. A program that causes a computer to execute the following steps: downscale the state variables used for prediction in each model and the prediction results for future times predicted in each model to a specified resolution; use the downscaled prediction results as training data to learn a prediction model that predicts future predicted values ​​from the downscaled state variables; and input the new state variables downscaled to the specified resolution into the prediction model to estimate the predicted values.

Images