A support vector machine-based method for improving real-time hourly precipitation forecast in Japan

Abstract

• Improved real-time hourly precipitation forecast accuracy using machine learning. • Investigated the use of CDF-transform for short-range forecast bias correction. • Provided forcing data for the next step to enhance flood early warning accuracy. Real-time precipitation forecast facilitates water management and water-associated disaster early warning. However, numerical weather prediction (NWP) models provide precipitation forecasts with bias. This study proposed to combine support vector machine (SVM) regression with quantile-based bias correction method to improve real-time 39-hour precipitation forecasts in Japan. Five methods were compared and evaluated against observations, which include SVM regression, quantile mapping (QM), cumulative distribution function transform (CDFt), and the combination of SVM and QM (or CDFt). Results indicated that the combination of SVM and CDFt (i.e., SVM-CDFt) generally provided the highest accuracy with good computational efficiency. SVM alone improved the spatial representation of hourly precipitation with a correlation coefficient increased from 0.387 to 0.490 in January and from 0.235 to 0.296 in July in the cross-validation experiment. However, SVM underestimated the variability of hourly precipitation and heavy precipitation events. QM and CDFt perform well in correcting the bias in modeled precipitation, while they have limited capability in correcting the rainband location. Combining SVM and quantile-based method took advantage of both approaches, providing a more consistent variability with observations and better predicted extreme precipitation events, although overestimation of rainfall area was witnessed. The simple concept, high computational efficiency, as well as evident improvement in forecast accuracy make the combined cases, especially the SVM-CDFt method, beneficial for real-time precipitation forecast and flood early warning.