A Research for 2-m Temperature Prediction Based on Machine Learning

Abstract

In this paper, we proposed two improved methods to solve the problems of current machine learning algorithm in the study of weather forecast revision. So this study can be divided into two parts. First of all, contrary to the previous machine learning algorithm with only one machine learning model applied for a vast area, a clustering algorithm is proposed to divide the vast area into multiple areas with different models for forecast correction. Secondly, Considering ECMWF the forecast of ECMWF with multiple initial times, a double-model correction model based on XGBoost machine learning algorithm is proposed. To test the new approach for the daily max and min temperature forecasts, the 2-m surface air temperature in the China area from the ECMWF (European Centre for Medium-Range Weather Forecasts) and the observatory meteorological observation data from china national stations are used. The data ranged from July 2017 to September 2019. We used meal weight interpolation method and sliding time window method to construct sample data. The new approach is compared with the multiple linear regression, random forest and SVM algorithms which are often used in weather forecast correction, it shows better numerical performance. The root mean square error (RMSE) and forecast accuracy are used to evaluate the model, and the forecast is better after the model is revised. The applicability of the model is verified by continuous test data (September 2019), and the experimental results show that the model has good practical value.