ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
Study on Machine Learning Correction of Precipitation Forecast and the Improvements on Two Types of Heavy Rain Forecasts
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Heavy precipitation is one of the most widespread hazardous weather affecting the socioeconomic stability and people"s livelihoods in China. Accurately forecasting such events poses significant scientific challenges. The multi-scale nonlinear characteristics of moist physical processes make numerical weather forecasting and objective corrections for precipitation considerably more difficult than for other meteorological variables like wind and temperature. This study investigates correcting 3-hour precipitation forecasts in the Beijing-Tianjin-Hebei region. Utilizing station observations and numerical model predictions, we explore the application of machine learning in correcting heavy precipitation through strategies such as constructing and sampling precipitation datasets, inputting relevant physical features, and training on residuals. The results demonstrate that addressing the long-tailed distribution challenge of precipitation samples—while maintaining the true distribution of precipitation and moderately increasing the proportion of heavy precipitation samples—is crucial for enhancing correction effectiveness. Statistical tests on the test set show that the machine learning correction scheme achieves significant improvements in skill scores for precipitation ranging from 0.1 to 20 mm compared to the raw model forecasts, with the enhancement magnitude increasing as the threshold rises. Additionally, the evaluation of categorized heavy precipitation indicates that it is more challenging to achieve improvements in convective heavy precipitation forced by weak weather systems. However, machine learning demonstrates particular promise for correcting heavy precipitation events with lower forecast accuracy from the model. Specifically, the lower the model"s predictive capability, the greater the space and extent for improvement through machine learning corrections. Feature importance analysis reveals that incorporating physically factors—particularly dynamic-thermodynamic, moisture parameters and the duration of convection—positively enhances model forecast scores while simultaneously mitigating machine learning models" tendency toward overprediction of light to moderate precipitation events.
Keywords:
Project Supported:
Clc Number:
Mobile website
