The Correction Method of Visibility Forecasts Based on the Machine Learning

Abstract

The visibility predicted by the GRAPES-CUACE model and the objective forecast of visibility are corrected for this study. An improved correction method using the Machine learning for visibility predictions is established. Temporal and spatial variation characteristics of visibility before and after the correction are analyzed focusing on the applicability and improvement effect of the Machine learning. The error statistics and TS cores test shows that regardless of the CUACE model or the objective method, the visibility is generally lower than the observation, and the correlation coefficient between the CUACE model prediction and the observation is low. The visibility predictions effect of North China is better other regions, the corrected value is closer to the measured value, and the correlation coefficient is significantly improved. The correction method can improve the forecasting effect of visibility, mainly because the measured values and meteorological conditions of visibility are introduced, and the differences in meteorological conditions in different regions are considered, and the visibility of each model prediction is dynamically corrected in real time. However, there are still regional differences in the correction effects of different regions. In the future research, better design schemes will be adopted, and meteorological data assimilation methods will be used to reduce weather forecast errors to improve the correction effect.