Applying machine learning for precipitation forecasting in uneven rainfall regions of Taiwan

Study region: Taiwan, with its steep topography, high spatiotemporal rainfall variability, and frequent extreme weather events, which make it difficult to predict the weather accurately. Predictability of rainfall in the short term is essential in flood prevention, hydrological modeling, and water resource management in this area. Study focus: This paper introduces a superior machine learning model in precipitation prediction with surface observation data. Even though the conventional method of Numerical Weather Prediction (NWP) models is still in use, it is computationally demanding and time-consuming. The machine learning models offer a more data-driven option, but heavy rainfall is hard to predict because precipitation data is non-Gaussian and highly skewed. In order to overcome this shortcoming, the Zero-Order Hold (ZOH) method is used to preprocess temporal data, and a regression-based ZD-XGBoost model is suggested. The objective function of the model is meant to fit the uneven distribution of rainfalls and enhance extreme precipitation prediction. New hydrological insights for the region: The suggested ZD-XGBoost model maintains the accuracy of the light rainfall and minimizes the errors of heavy rainfall prediction by about 20% and 40% respectively, on hourly and daily data. These findings indicate that the distribution-conscious machine learning with surface observations can be effectively used to reduce the problem of data imbalance and improve the reliability of rainfall prediction in topographically complex and rainfall-heterogeneous areas like Taiwan.