一种机器学习方法在湖北定时气温预报中的应用试验
Using a Machine Learning Method for Temperature Forecast in Hubei Province
中文摘要：
利用2015—2017年湖北89个气象站地面观测温度、欧洲中心再分析资料和0～12 h预报资料回归模式输出要素与地面气温之间的关系，建立了LightGBM模型，并在2018年数据集上进行测试。结果表明，定时气温平均绝对误差由模式本身的1.8 ℃下降到1.1 ℃，2 ℃以内预报准确率由65.9%上升至86.6%，决定系数（拟合优度）高达0.97。该模型已经在武汉中心气象台业务化，初步选取定时气温中的极值进行2018年2—6月预报评分，24 h高、低温预报准确率分别为76.9%和91.4%，在客观产品中排名前列，较数值预报模式产品提升明显，低温预报准确率超过预报员水平。LightGBM作为一个年轻的机器学习框架，在气象要素预报方面具备良好的应用前景。
中文关键词：
机器学习，决策树，LightGBM，气温，预报试验
KeyWords：
machine learning, decision tree, LightGBM, temperature, forecast experiment
Abstract：
A prediction model using LightGBM was established according to a regression analysis using ground observations of temperature, ERA-Interim data and 0-12 h forecast data. Tests on the data set from January to April in 2018 showed that the mean absolute error of the temperature decreased from 1.8℃ to 1.1℃ with the EC model, the prediction accuracy increased from 65.9% to 86.6% and the coefficient of determination was 0.97. This model has been operationalized by the Wuhan Central Meteorological Observatory. The forecast score from February to June in 2018 showed that the accuracy of the highest and lowest temperature forecasts were 76.1% and 91.5%, respectively, which was significantly higher than that of the numerical weather prediction, and the minimum temperatures were better than current forecasters. As a young machine learning frame, LightGBM has strong prospects for application in meteorological forecasting.