Climate feature factor extraction method based on combination of supervised learning and unsupervised learning

Abstract

The invention provides a climate feature factor extraction method based on combination of supervised learning and unsupervised learning. The method comprises the steps of obtaining historical data of factors; wherein the historical data of the factors comprises historical data of physical quantity field factors and historical data of forecast objects; standardizing the historical data of the factors to obtain standardized data; analyzing the standardized data by adopting a supervised learning regression class, and extracting a mean square error field of the standardized data; analyzing the standardized data by adopting a correlation coefficient class, and extracting a correlation coefficient field of the standardized data; acquiring a forecast factor set of a mean square error field; acquiring a forecast factor set of the correlation coefficient field; and merging the forecast factor set of the mean square error field and the forecast factor set of the correlation coefficient field to obtain a multi-factor sequence. The invention has the advantages that the situation that information of a low-influence area is covered by an overall high-correlation and high-influence area when factors are extracted in a large range is avoided; and excessive discarding of low-signal information inconsistent with the overall field element modality is avoided.

Description

technical field [0001] The invention is mainly used in the field of meteorological forecasting and forecasting, especially relates to a machine learning feature engineering method for factor extraction in the process of statistical modeling of climate forecasting, and has certain reference application value for the interpretation and application of medium and short-term numerical model products. Background technique [0002] Meteorological predictions are often carried out using machine learning-related methods. However, in terms of machine learning statistical modeling, the impact of feature engineering and feature selection on the analysis results is often more important than the selection of machine learning models. [0003] In monthly and seasonal short-term climate prediction, traditional methods are often based on linear methods to extract and find features, and the use of nonlinear information of the extremely complex global climate system is limited, so it is necessa...

AI Technical Summary

Problems solved by technology

[0006] 1. The climate system is very complex. Traditional factor extraction methods often use a linear method, but there is no good extraction scheme for high-impact areas that may be non-linear impact statistics

[0007] 2. When extracting elements of the global field, areas with high overall correlation and high impact often cover up information in areas with low impact, and information in low-influence areas that is not extracted and participates in modeling will affect the stability and prediction effect of the final modeling

Images