Benchmarking Quantitative Precipitation Forecast Using a Composite of Numerical Modeling and Deep Neural Networks
Abstract
Quantitative precipitation forecast, which informs the expected amount of precipitation accumulated over a specified region over a specified period, remains a challenging task for current numerical weather/climate models. The difficulty lies in inferring the unobserved status and unresolved microphysics of cloud hydrometeors with different phases, and estimating the falling of the resulting precipitates. Since a large number of assumptions and empirical parameter values are involved in parameterizing the cloud and precipitation process, it is difficult to carry out deductive diagnosis to improve the relevant parameterization schemes. Recently, machine learning has attracted attention of the earth system modeling community, since it offers an alternative for parameterizing the unresolved processes by learning from high quality examples, such as observations or high resolution numerical simulations. We offer a novel dataset that enables learning quantitative precipitation forecast by matching long-term (1980-2019), quality-controlled, hourly, gauge-point scale precipitation observations with their surrounding atmospheric moisture and mass field analysis. Based on this dataset, we conduct a series of experiments that testify the impacts of predictor selections (from synoptic features to smaller spatial-scale features), neural network architectures (convolution, recurrent, transformer, etc.), and hyperparameter choices. We show that hourly precipitation predictions from the optimal deep neural network models give r ≈ 0.8 at 2 ×2.5 spatial scale, while the baseline numerical model achieved r ≈ 0.5. The best performance at hourly, gauge-point scale reaches the order of r ≈ 0.6 for some gauges. However, there is high skill variance in estimating precipitation at such a stringent spatiotemporal resolution. To further test the proposed model in practical forecasts, dynamical retrospective forecast experiments for two atmospheric river land-falling events are carried out using the Weather Research and Forecasting (WRF) model. The WRF dynamical simulations are used to force the trained neural network model for alternative precipitation process predictions. Simulation results verified the consistency and robustness of the proposed approach.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMGC43D1349P
- Keywords:
-
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- COMPUTATIONAL GEOPHYSICS;
- 1626 Global climate models;
- GLOBAL CHANGE;
- 1942 Machine learning;
- INFORMATICS;
- 4313 Extreme events;
- NATURAL HAZARDS