Data assimilation method

Abstract

A sensitivity calculation unit 11 performs sensitivity analysis on a change in a parameter Q for which prediction accuracy is to be improved, and identifies a parameter P that affects the parameter Q. An effect amount calculation unit 12 performs, until a time partway through a time at which prediction is to be performed for data assimilation experiments under individual conditions, a data assimilation experiment under a condition, calculates the prediction accuracy for the parameter P, and, on the basis of the prediction accuracy for the parameter P, determines whether to continue or end the data assimilation experiment under said condition.

Description

Data assimilation methods

[0001] The present disclosure relates to data assimilation methods.

[0002] When predicting phenomena through simulations, a widely used method is to incorporate observational data into the initial and predicted values ​​to obtain more realistic results and improve the accuracy of the predictions. This process is called data assimilation and is used in various fields such as meteorology and engineering.

[0003] Non-Patent Document 1 improved forecast accuracy by using a supercomputer and performing high-frequency data assimilation every 30 seconds using a large-scale data assimilation system built on the supercomputer. Non-Patent Document 1 also evaluated the effect on forecast accuracy of a linear precipitation band that occurred in Kyushu in 2020 if there was a precipitation radar covering the entire Kyushu region using OSSE (Observation System Simulation Experiment), and showed that the accuracy of precipitation forecasts improved in Kumamoto, which was particularly hard hit.

[0004] On the other hand, since obtaining observational data is costly, there are studies that attempt to use some method to identify observation locations and items that are effective in improving prediction accuracy.

[0005] Non-patent document 2 examines the impact of radiosonde observations at 12 locations in the Northern Hemisphere on global weather forecasts using a weather simulation that implements a method called Forecast Sensitivity to Observation (FSO), and shows that radiosonde observations in the Arctic are the most effective for forecasting weather in North America 4-7 days in advance.

[0006] Yasumitsu Maejima, et al.,“Observing System Simulation Experiments of a Rich Phased Array Weather Radar Network Covering Kyushu for the July 2020 Heavy Rainfall Event,”SOLA, 2022, Vol. 18, pp. 25-32Akira Yamazaki, et al.,“EFSO at Different Geographical Locations Verified with Observing System Experiments,”Weather and Forecasting, Volume 36, Issue 4, 2021, pp. 1219-1236Norihiro Oyama, et al., “Deep generative model super-resolves spatially correlated multiregional climate data,” Scientific Reports, 2023

[0007] In data assimilation experiments, multiple conditions must be comprehensively verified to find optimal parameters. Weather forecasts, in particular, require high-resolution simulations under each condition, which requires a large amount of computer resources.

[0008] To speed up data assimilation experiments, there are methods to reduce the amount of calculation by narrowing down the verification conditions, and methods to reduce the amount of calculation in simulations by utilizing surrogate models using deep learning.

[0009] Non-Patent Document 3 proposes a new approach to achieve highly accurate downscaling in meteorological simulations, and by utilizing machine learning, achieves high resolution while reducing the amount of calculation.

[0010] However, when you want to improve the forecast accuracy of a specific parameter at a specific time, these measures are insufficient because you cannot obtain results unless you run a simulation up to the forecast time. For example, to predict precipitation 24 hours in the future, you need to run a simulation for 24 hours.

[0011] The present disclosure has been made in view of the above, and aims to efficiently search for optimal conditions in data assimilation experiments.

[0012] A data assimilation method of one embodiment of the present disclosure performs a sensitivity analysis on changes in a first parameter for which it is desired to improve the prediction accuracy, identifies a second parameter that affects the first parameter, performs a data assimilation experiment for each condition up to a time partway through the time to be predicted, calculates the prediction accuracy of the second parameter, and determines whether to continue or terminate the data assimilation experiment for that condition based on the prediction accuracy.

[0013] According to the present disclosure, optimal conditions can be efficiently searched for in data assimilation experiments.

[0014] FIG. 1 is a diagram showing an example of the configuration of an information processing device according to this embodiment. FIG. 2 is a flowchart showing an example of the processing flow of the information processing device. FIG. 3 is a diagram showing a typhoon reproduced using WRF. FIG. 4 is a diagram showing initial values ​​assigned during sensitivity analysis. FIG. 5 is a diagram showing an example of sensitivity calculated retroactively. FIG. 6 is a diagram showing an example of highly sensitive parameters extracted from FIG. 5. FIG. 7 is a diagram showing an example of the hardware configuration of the information processing device.

[0015] [Device Configuration] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings.

[0016] An example of the configuration of an information processing device 10 according to this embodiment will be described with reference to FIG. 1. The information processing device 10 shown in the figure calculates the accuracy of an assimilation experiment under each condition at regular intervals and determines whether to terminate the assimilation experiment if the condition does not reach the predetermined accuracy. In calculating the accuracy, a parameter sensitive to a specific parameter at a specific time for which prediction accuracy is desired to be improved is identified through sensitivity analysis, and the prediction accuracy of that parameter is referenced. For example, when predicting parameter Q 24 hours after time T (T+24), a parameter P sensitive to parameter Q at time T+24 at time T+12 is identified, and the prediction accuracy of parameter P at time T+12 is referenced.

[0017] The information processing device 10 in FIG. 1 includes a sensitivity calculation unit 11 and an effect amount calculation unit 12 .

[0018] The sensitivity calculation unit 11 performs sensitivity analysis on changes in a parameter Q for which prediction accuracy is desired to be improved, and identifies a parameter P that has an effect on the parameter Q. For example, the sensitivity calculation unit 11 uses the parameter Q as an input and performs inverse calculation of a simulation going back in time to identify a parameter P that has a large effect on changes in the parameter Q at a specific time in the past. From the perspective of reducing the amount of calculation, the specific time in the past may be any time before the specific time for which prediction accuracy is desired to be improved.

[0019] The effect size calculation unit 12 performs a data assimilation experiment (simulation) for each condition of the data assimilation experiment up to a point in time during the time to be predicted (time within the simulation), calculates the prediction accuracy of parameter P, which is sensitive to changes in parameter Q, and determines to continue the data assimilation experiment if the prediction accuracy achieves the desired accuracy, or to terminate the data assimilation experiment if the desired accuracy is not achieved. The prediction accuracy of parameter P is obtained by calculating the root mean square error (RMSE) between the predicted value of parameter P and the true value or analytical value, or by calculating it from the rate of error growth of parameter P in the simulation. The desired accuracy may be determined in advance as a specific numerical value, or may be determined according to a value such as 10% of the value of parameter P.

[0020] [Processing Flow] An example of the processing flow of the information processing apparatus of this embodiment will be described with reference to the flowchart of FIG.

[0021] In step S11, the sensitivity calculation unit 11 receives information on a parameter Q for which the prediction accuracy is desired to be improved.

[0022] In step S12 , the sensitivity calculation unit 11 performs sensitivity analysis to obtain a parameter P that has sensitivity to the parameter Q at a specific time, and outputs information about the parameter P to the effect amount calculation unit 12 .

[0023] In step S13, the effect amount calculation unit 12 performs a data assimilation experiment for each condition of the data assimilation experiment up to a specific time, and calculates the accuracy of the parameter P having sensitivity at the specific time.

[0024] In step S14, the effect amount calculation unit 12 determines whether or not to continue the data assimilation experiment for each condition, based on the accuracy calculated in step S13.

[0025] Data assimilation experiments are conducted only under promising conditions by stopping data assimilation experiments under conditions that are judged to be terminated and continuing data assimilation experiments under conditions that are judged to be continued.

[0026] For the ongoing data assimilation experiment, the processes of steps S13 and S14 may be repeated until the next intermediate time.

[0027] [Example] Next, an example will be described in which the method of the present disclosure is applied to a data assimilation experiment to predict precipitation for Typhoon No. 11 in 2022. The parameter Q, for which prediction accuracy is desired to be improved, is set to "precipitation at the typhoon center six hours later." The Weather Research and Forecasting Model (WRF), which can reproduce typhoons, is used as the numerical model. Figure 3 shows a diagram of Typhoon No. 11 in 2022 reproduced using WRF. To predict precipitation (parameter Q) at 12:00 on August 31, 2022, observation data is assimilated at 6:00 on August 31, 2022.

[0028] First, parameter P, which affects parameter Q, is identified. The sensitivity calculation unit 11 performs a sensitivity analysis on the precipitation amount near the typhoon center going back three hours. The sensitivity of each parameter at 9:00 a.m. to parameter Q is examined. Specifically, a sensitivity of 1 is assigned only to parameter Q (Figure 4), and calculations are performed going back in time to calculate how the sensitivity of 1 spreads over three hours. Figure 5 shows the sensitivity when going back three hours. Calculating going back in time means back-analyzing (solving backward) the equations in the simulation model. For the back-analysis, an adjoint model provided in WRF is used, and calculations are performed within the range in which linearity can be assumed.

[0029] The sensitivity calculation unit 11 extracts a parameter P with high sensitivity. For example, the sensitivity calculation unit 11 extracts only points with a sensitivity of 0.3 or higher. As the parameter P, locations and items (temperature, air pressure, water vapor content, etc.) with high sensitivity at 9:00 (3 hours before) are extracted. High sensitivity includes both positive sensitivity, where the output value increases as the value increases, and negative sensitivity, where the output value increases as the value decreases. In this example, it is assumed that "humidity at multiple points northeast of the typhoon" is extracted as the parameter P, as shown in Figure 6.

[0030] Next, a data assimilation experiment under each condition is performed until partway through the simulation time, the prediction accuracy of the parameter P is calculated, and the data assimilation experiment to continue is extracted. The effect amount calculation unit 12 obtains the prediction results up to 9:00 in the data assimilation experiment under each condition, and calculates the accuracy of the parameter P. In this embodiment, the data assimilation experiment is not performed until 12:00, and the results up to 9:00 are output. The accuracy of the parameter P is found by calculating the RSME between the parameter P and the true value or analytical value.

[0031] The effect amount calculation unit 12 determines whether to continue the data assimilation experiment for each condition based on the accuracy of the parameter P. For example, the effect amount calculation unit 12 determines to continue the data assimilation experiment for a condition where the RMSE is less than 5%, and determines to terminate the data assimilation experiment for a condition where the RMSE is 5% or more.

[0032] As described above, according to this embodiment, the sensitivity calculation unit 11 performs sensitivity analysis on changes in parameter Q, for which prediction accuracy is desired to be improved, and identifies parameter P that affects parameter Q. The effect amount calculation unit 12 performs data assimilation experiments for each condition up to a time partway through the time to be predicted, calculates the prediction accuracy of parameter P, and determines whether to continue or terminate the data assimilation experiment for that condition based on the prediction accuracy of parameter P. In this way, by calculating the expected improvement in prediction accuracy before running the simulation to the end and terminating those that are unlikely to achieve the desired accuracy, the overall amount of calculation can be reduced. As a result, it becomes possible to verify more conditions and search for a more optimal assimilation method.

[0033] The information processing device 10 described above can be, for example, a general-purpose computer system including a central processing unit (CPU) 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 7. In this computer system, the information processing device 10 is realized by the CPU 901 executing a predetermined program loaded onto the memory 902. This program can be recorded on a computer-readable non-transitory recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or can be distributed via a network.

[0034] 10 Information processing device 11 Sensitivity calculation unit 12 Effect amount calculation unit

Claims

1. A data assimilation method in which a computer performs sensitivity analysis on changes in a first parameter for which the prediction accuracy is desired to be improved, identifies a second parameter that affects the first parameter, performs data assimilation experiments for each condition up to a time partway through the time to be predicted, calculates the prediction accuracy of the second parameter, and determines whether to continue or terminate the data assimilation experiment for that condition based on the prediction accuracy.

2. A data assimilation method according to claim 1, comprising: inputting a first parameter for which prediction accuracy is desired to be improved; performing a reverse simulation going back in time; and identifying a second parameter that has a large influence on changes in the first parameter at past times.

Images