Study on Machine Learning Correction of Precipitation Forecast and Its Validation on Two Types of Precipitation
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 prediction and objective corrections for precipitation become considerably more difficult than for other meteorological variables like wind and temperature. Utilizing station observations and numerical model forecasts, this paper explores the application effect of machine learning algorithm (LightGBM) in correcting 3 h accumulated precipitation forecasts for the Beijing-Tianjin-Hebei Region through strategies such as constructing and sampling precipitation datasets, inputting relevant physical features, and training on residuals. The results demonstrate that, to address the long-tailed distribution challenge of precipitation samples, when constructing the dataset it is crucial to comprehensively consider maintaining the true distribution of precipitation while moderately increasing the proportion of samples with stronger precipitation intensities. This is a key step in enhancing the correction effectiveness of heavy precipitation. Statistical tests on the independent test set show that the LightGBM correction scheme achieves significant improvements in skill scores for precipitation ranging from 0.1 mm to 20 mm compared to the raw model forecasts, and the increase rate ascends upward as the threshold rises. Statistical tests and individual case evaluations of precipitation by type show that the LightGBM correction scheme presents comprehensive adjustments in rainfall intensity and fall area in different types of precipitation forecasts. Among them, the number of forced precipitation samples by severe weather systems is relatively superior, and the correction effects on each precipitation threshold are more remarkable. Additionally, the evaluation of classified heavy precipitation indicates that it is more challenging to achieve improvements in the correction of convective heavy precipitation forced by weak weather systems, in particular in the cases with precipitation ≥15 mm. The adjustment of the fall area of precipitation is more challenging. This suggests that the unbalanced distribution of precipitation samples remains a challenge for machine learning correction. However, machine learning has shown particular promise for correcting larger magnitudes of heavy precipitation events with lower forecast accuracy from the model, that is, the lower the model’s forecast ability, the greater the room and extent for the correction improvement of machine learning. The analysis of feature importance shows that the input of physical features such as dynamics, thermodynamics and water vapor, which are closely related to precipitation, has a positive contribution to enhancing the correction score of LightGBM.