3:30 PM - 3:45 PM
[AAS09-06] Interfaces between the past meteorological data such as reanalysis data and social applications
Keywords:reanalysis, social applications, machine learning, earth digital twin
Although the value of reanalysis data has been highly recognized by the scientists, it is not widely used in the community of industry and society. However, we are now at the stage of stepping into a new phase of reanalysis application under the recent rapid development of machine learning and artificial intelligence, as well as the recent availability of high-resolution products of the regional reanalysis.
The ClimCORE project, which has been promoted at the RCAST, University of Tokyo, under JST's COI-NEXT, is currently producing a 5 km mesh regional reanalysis based on the data assimilation system of the JMA's mesoscale NWP model “MSM”. In addition, since MSM uses analytical rainfall data for data assimilation to improve the accuracy of precipitation forecasting, reprocessing of Radar/Raingauge Analyzed Precipitation data has been conducted retrospectively. Since some parts of these products have been produced, utilization research has been developed within the project. In this paper, we try to classify the utilization of reanalysis data based on the experience of the utilization research within the project.
First, there are some social applications which can directly utilize the past data. For example, the potential of renewable energy, various agricultural production potentials, and the frequency of extreme events that cause disasters are examples of direct usages where statistics based on historical data itself are important. For example, historical data is already available in the form of the Agro-Meteorological Grid Square Data from NARO, and reanalysis data has been already used in the renewable energy and insurance industries. One of the challenges in applying these data to society is the granularity of the information. For example, while AMeDAS has been in operation for about 50 years, the horizontal resolution is on the order of 20 km, and there is no observation data for offshore areas except for islands. The value of the regional reanalysis data is already recognized in these applications.
For wind fields, which can be utilized for potential capacity of wind power generation, the reanalysis data might be the only solution for the past periods. The horizontal resolution of the reanalysis is important because wind is highly affected by topography. Reanalysis data is also considered to be highly important for investigating extreme events causing disasters in the past.
On the other hand, in many fields, the better future prediction, such as weather forecasting for tomorrow and climate prediction for the next few decades, is highly required. In responding to these requirements, combining reanalysis data with numerical weather predictions and climate predictions is a key issue. In the case of numerical weather prediction, it is possible to improve quality of prediction by learning from long term homogeneous reanalysis data in the past and applying it for daily forecast. By using techniques such as super-resolution based on deep learning, we will be able to obtain downscaled mesh data with bias correction and translation, which corresponds to the extension of conventional guidance methods.
AI weather forecasting is a technology that has made rapid progress in the past few years by utilizing reanalysis data. AI weather model can predict the weather itself through machine learning of reanalysis data. Some results have been reported that surpass traditional numerical weather forecasting based on physical laws. It should be noted that AI weather forecasting requires reanalysis data (using physical laws).
In climate prediction, downscaling by super-resolution techniques is possible in the same way as above, and regional reanalysis data is important for the teacher data. Biases in regional climate models should be reduced as much as possible for adaptation measures. It is possible to evaluate biases comparing the reanalysis and the simulation and these biases will be reflected in the reduction of biases in the climate prediction data.
Finally, we would like to mention the superiority of reanalysis data over observation-based analysis data. Observation based data such as the shortwave radiation estimated by the geostationary satellite is more accurate than reanalysis data. But reanalysis data can be extended to the spatial dimension by downscaling and to the temporal dimension by prediction to the future. Such extensibility is owing to 4-D structure of reanalysis data, which we may call the Earth Digital Twin, data infrastructure for various social applications.
This research was supported by the JST Co-Creation Field Formation Support Program JPMJPF2013.
The ClimCORE project, which has been promoted at the RCAST, University of Tokyo, under JST's COI-NEXT, is currently producing a 5 km mesh regional reanalysis based on the data assimilation system of the JMA's mesoscale NWP model “MSM”. In addition, since MSM uses analytical rainfall data for data assimilation to improve the accuracy of precipitation forecasting, reprocessing of Radar/Raingauge Analyzed Precipitation data has been conducted retrospectively. Since some parts of these products have been produced, utilization research has been developed within the project. In this paper, we try to classify the utilization of reanalysis data based on the experience of the utilization research within the project.
First, there are some social applications which can directly utilize the past data. For example, the potential of renewable energy, various agricultural production potentials, and the frequency of extreme events that cause disasters are examples of direct usages where statistics based on historical data itself are important. For example, historical data is already available in the form of the Agro-Meteorological Grid Square Data from NARO, and reanalysis data has been already used in the renewable energy and insurance industries. One of the challenges in applying these data to society is the granularity of the information. For example, while AMeDAS has been in operation for about 50 years, the horizontal resolution is on the order of 20 km, and there is no observation data for offshore areas except for islands. The value of the regional reanalysis data is already recognized in these applications.
For wind fields, which can be utilized for potential capacity of wind power generation, the reanalysis data might be the only solution for the past periods. The horizontal resolution of the reanalysis is important because wind is highly affected by topography. Reanalysis data is also considered to be highly important for investigating extreme events causing disasters in the past.
On the other hand, in many fields, the better future prediction, such as weather forecasting for tomorrow and climate prediction for the next few decades, is highly required. In responding to these requirements, combining reanalysis data with numerical weather predictions and climate predictions is a key issue. In the case of numerical weather prediction, it is possible to improve quality of prediction by learning from long term homogeneous reanalysis data in the past and applying it for daily forecast. By using techniques such as super-resolution based on deep learning, we will be able to obtain downscaled mesh data with bias correction and translation, which corresponds to the extension of conventional guidance methods.
AI weather forecasting is a technology that has made rapid progress in the past few years by utilizing reanalysis data. AI weather model can predict the weather itself through machine learning of reanalysis data. Some results have been reported that surpass traditional numerical weather forecasting based on physical laws. It should be noted that AI weather forecasting requires reanalysis data (using physical laws).
In climate prediction, downscaling by super-resolution techniques is possible in the same way as above, and regional reanalysis data is important for the teacher data. Biases in regional climate models should be reduced as much as possible for adaptation measures. It is possible to evaluate biases comparing the reanalysis and the simulation and these biases will be reflected in the reduction of biases in the climate prediction data.
Finally, we would like to mention the superiority of reanalysis data over observation-based analysis data. Observation based data such as the shortwave radiation estimated by the geostationary satellite is more accurate than reanalysis data. But reanalysis data can be extended to the spatial dimension by downscaling and to the temporal dimension by prediction to the future. Such extensibility is owing to 4-D structure of reanalysis data, which we may call the Earth Digital Twin, data infrastructure for various social applications.
This research was supported by the JST Co-Creation Field Formation Support Program JPMJPF2013.