A Support Vector Machine-based Method for Improving Real-time Hourly Precipitation Forecast in Japan
Abstract
Japan is a country greatly influenced by extreme precipitation. An accurate and effective real-time precipitation forecast is crucial for flood early warning and water management. This study proposed to combine support vector machine (SVM) regression with quantile-based bias correction method to improve real-time 39-hour precipitation forecasts simulated from a non-hydrostatic numerical weather prediction model in Japan. Five methods, which include SVM regression, quantile mapping (QM), cumulative distribution function transform (CDFt), and the combination of SVM and QM (or CDFt), were compared and evaluated against observations. Results suggested the combination of SVM and CDFt (i.e., SVM-CDFt) provided the highest accuracy with satisfactory computational efficiency. SVM exhibited the capability of improving the spatial representation of precipitation with the use of neighborhood grids. The correlation coefficient between forecasted and observed hourly precipitation increased from 0.39 to 0.49 in January and from 0.24 to 0.30 in July in the cross-validation experiment. However, SVM underestimated the variability of hourly precipitation, especially for heavy precipitation events. Quantile-based methods corrected forecast bias, whereas showed limited ability to correct rainband location estimation. Combining SVM and quantile-based method took advantage of both approaches and yielded precipitation forecasts with higher accuracy, although an overestimation of rainfall area during extreme rainfall events was witnessed. Overall, the simple concept, high computational efficiency, as well as evident improvement in forecast accuracy make the combined cases, especially the SVM-CDFt method, promising to be used in an operating system for real-time flood early warning.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H22J..04Y