Meteorology
Scenarios
High-performance computers have long been essential to the meteorology industry. In weather forecast, many factors need to be identified by numerical predictions, which lead to massive computing requirements. Three methods, including statistical forecast, synoptic forecast, and numerical weather prediction (NWP) have been used for weather forecast. Among them, the statistical forecast is based on the meteorological observation data accumulated over the years. It is generally used for mid- and long-term (more than 10 days) forecast, but the accuracy is low. In the mid- (3–10 days) and short-term (12 hours – 3 days) and ultra-short-term (12 hours) weather forecast, the synoptic forecast and NWP are mainly used. In particular, NWP becomes the most important method for more accurate weather forecast. Table 1 describes the weather forecast models and application.
Mesoscale Weather Forecast Model |
Description |
Application |
|---|---|---|
MM5 |
The MM5 is the fifth-generation mesoscale weather forecast model jointly developed by National Atmospheric Research Center (NCAR) and Pennsylvania State University. |
The MM5 is a mesoscale and small-scale non-static dynamic weather model used for weather forecast simulation. It is the most widely used mesoscale forecast model in the meteorological field. The latest version of MM5 is MM5V3. The R&D of MM5V3 has been stopped, and the focus has switched to the Weather Research Forecasting (WRF) model. |
WRF |
The WRF model is a next-generation mesoscale model developed based on the MM5 model. It is developed by many scientists of research departments and universities in the United States. |
The WRF model combines portability with easy maintenance and use. It also features high scalability and efficiency, making it a widely used model. The public WRF model is maintained and supported by the NCAR and released for free. |
GRAPES |
Global/Regional Assimilation and Prediction System (GRAPES) is a next-generation numerical forecast system developed by China Meteorological Administration (CMA). It is an advanced system adaptive to both mesoscale and large-scale weather forecast. |
The long-term goal of GRAPES is to contribute to the R&D of China's numerical forecast system in the 21st century, improve the accuracy of numerical forecast, and provide scientific theoretical and technical support for the weather forecast services of CMA. The GRAPES model will play a significant role in supercharging the fast-growing meteorological services. |
AREMS |
To provide meteorological services of higher accuracy, Wuhan Institute of Heavy Rain, CMA and the Institute of Atmospheric Physics, Chinese Academy of Sciences (CAS) have developed the AREMS mesoscale rainstorm numerical forecast model based on the REM model (also called the ETA model). |
The AREMS model has demonstrated high accuracy in the rainstorm forecast tests in Jianghuai River Basin. It accurately predicts the position and motion trend of rain belt, as well as the position and intensity of rainstorm center in the region. In particular, the TS score of the AREMS model in the rainstorm forecast within 12 to 36 hours is significantly higher than that of other models. |
Challenges
- Massive computing
Mesoscale weather forecast models (MM5, WRF, GRAPES, and AREMS) require massive computing. The calculation must be completed within a specified period (usually no more than 2 hours) to provide real-time weather forecast. As weather forecast is required to deliver higher precision in a wider area, the computing volume of the model grows exponentially. 16 times higher computing volume is needed to double the prediction accuracy. As a result, it is impossible for a single CPU or an ordinary computer to meet the high computing requirements of numerical weather forecast, making parallel computing a must-have. To this end, model prediction software is parallelized through message transfer or shared memory, and high-performance parallel computers are required. Currently, most mesoscale weather forecast models have been parallelized. For example, MM5 and WRF support parallel MPI message transfer and OpenMP shared storage memory, as well as the MPI+OpenMP hybrid running mode. The GRAPES supports only the MPI parallel mode, and the AREMS also supports MPI parallel computing.
- Intensive communication
The weather forecast models use software and the finite difference lattice mode for parallel computing. Therefore, when the mesoscale weather forecast model is run, the volume of communication between processors is high, and the model requires high communication performance. For example, the communication of MM5 and WRF includes not only the inter-domain communication between the parent and the nested domain, but also the communication between each domain. This makes the high-performance communication network a must-have for high-performance computers.
- Requirements for high real-time performance and scheduled running
The computing cluster for the meteorology industry needs to deliver good real-time performance to support accurate weather forecast. In addition, the forecast system needs to run automatically at fixed time without manual intervention. Generally, the scheduled jobs are completed in two to four fixed time segments every day. Each time segment lasts for two hours. These jobs run at the same time every day, and you must ensure that they can be completed on time.
- High computing requirements in main mode
The software processing procedure can be divided into three phases: pre-processing, main mode, and post-processing. The pre-processing phase includes document download and data assimilation. The post-processing phase covers the graphical processing of generated products. These two phases do not have high requirements on computers. The main mode is the most important part of the entire system and demands for the most computing requirements and high performance of computers.