Numerical mode correction method based on sequential regression learning

Abstract

The invention belongs to the technical field of machine learning and meteorological prediction, and particularly relates to a numerical mode rainfall correction method based on sequential regression learning. The method comprises the following steps: performing feature selection on meteorological features of a numerical mode, and selecting effective features by utilizing correlation between the meteorological features and a ground observation rainfall value; carrying out region segmentation on global data according to longitude and latitude of a ground observation station to generate a plurality of smaller space ranges, and then carrying out regularization processing on features in the space ranges; putting the regularized features into an auto-encoder for training to obtain noise-removedmixed features; and finally, obtaining rainfall confidence and sequential regression distribution through the rainfall probability prediction network and the sequential regression distribution prediction network, and fusing the rainfall confidence and the sequential regression distribution to generate a finally corrected rainfall value. According to the method, the characteristics of ordered discrete continuous values, namely rainfall values, can be well extracted, so that the model can well learn rainfall prediction errors in a numerical mode, and a better correction effect is achieved.

Description

technical field [0001] The invention belongs to the technical field of machine learning and meteorological forecasting, and in particular relates to a numerical model correction method based on ordinal regression learning. Background technique [0002] Numerical model forecasting and correction is one of the mainstream methods in weather forecasting. Using the knowledge of atmospheric dynamics and meteorology to generate forecasts of weather indicators, and then relying on expert experience to manually generate local forecast correction results as the final weather forecast results. Most of the current methods in this field mainly rely on the subjective experience of the forecaster. [0003] The earliest numerical weather prediction (NWP) dates back to the work of Charney et al. [1] in 1950, which was based on atmospheric dynamics to generate forecasts of air temperature and precipitation. [2][3] are two typical single-model methods, they use the post-processing method to ...

AI Technical Summary

Problems solved by technology

However, multi-model methods often require a large amount of data, and the accuracy of the produced results is limited, which cannot produce more refined prediction results.

In addition, the above methods all rely on the subjective knowledge of the forecaster, a large number of manual features and linear models to correct the error

Images