Ensemble simulation is usually done with multiple runs of a deterministic model. This poster present an analog-based forecast system that only runs the deterministic model once, and then generate probability information.
Analog Ensemble technique is applied to a variety of problems including the short-term weather forcast, the photovoltaic energy production forecast, and forecast downscaling.
Analog Ensemble is a scalable and efficient algorithm with an HPC infrastructure.
Numeric weather prediction is undergoing a revolution resulting from the continuous advances in scientific knowledge and technologies. With dozens of weather models emerging that all generate different predictions from each other, forecasts have been gradually shifting from a deterministic form to a probabilistic form which shows the increasing concerns of, not just the absolute prediction values, but the confidence of predictions and the uncertainty of models.
As a computational problem, generating uncertainty information can be an expensive task. Conventionally, prediction models are initiated with slightly perturbed parameters and then the diversion of model results can be a measure of model uncertainty. However, the multi-simulation approach drastically increases the computational requirement so that it can potentially exceed the ability of the state-of-art high-performance computing platforms. Meanwhile, if spatial and temporal resolutions are of concern, this approach is far from being efficient and viable.
The Parallel Ensemble Forecast system is designed to generate probabilistic weather forecasts by using the revolutionary numerical weather prediction technique, Analog Ensemble. It is a data-driven method that derives probability information of a deterministic prediction model using past forecasts and observations without multiple simulation runs. Integrated with high-performance platforms, the system distributes computational tasks among nodes and therefore further boosts the data simulation process.
If it is not displaying normally, you can access the PDF here.
Click here to see the abstract from 2018 AGU database for more information and citation. Thank you.