A machine learning bias correction method for precipitation
corresponding to weather conditions using simple input data
- Takao Yoshikane,
- Kei Yoshimura
Abstract
Various bias correction methods have recently been proposed using
machine learning techniques. Generally, machine learning methods are
fairly complicated, and it is extremely difficult to explain how machine
learning corrects model biases. Accordingly, researchers perpetually
seek to apply machine learning methods to diverse cases and to determine
whether these methods are reliable. Here, we developed a machine
learning method using simple input data by assuming a relation between
observed and simulated precipitation corresponding to weather
conditions. This simple method can find the optimal relation without
employing dimension reduction and can facilitate the comprehension of
precipitation characteristics. According to a validation experiment,
this simple method can correct the precipitation frequency corresponding
to the orography and estimate the local precipitation distribution
characteristics, resulting in values similar to the observed data even
when data are forecasted more than 24 hours from the initial time.