基于机器学习的模式温度预报订正方法研究

机器学习在气象数值模式的后处理中表现优越，但其稳定性和适用性有待深入探究。本文选取了ECWMF模式包括2米温度、风、降水等多气象要素预报产品和安徽省80个国家气象站观测2米温度实况资料，分析了EC模式在安徽省站点温度预报误差，利用决策树、随机森林、LightGBM三种机器学习算法订正EC 模式0-72小时温度站点预报，并将其与传统MOS订正方法和SPCC主观预报产品进行了对比。结果表明：EC模式高温预报误差明显高于低温预报，在安徽皖南山区和大别山区存在较大误差；机器学习算法中最高温度预报随机森林表现最优，最低温度预报LightGBM最优，比EC模式平均绝对误差MAE分别降低了0.55℃、0.2℃，均方根误差RMSE分别降低0.6℃、0.31℃，预报准确率提高了18.16%和5.19%；高山站独立建模并融合周围站的信息能有效降低模型误差；相比SPCC主观预报产品，机器学习预报模型在高温和寒潮过程中互有优劣，但在天气转折初期落后；机器学习可以作为常规预报模式的补充，能显著优化或改善传统预报中温度预测精度，特别是对于数据缺乏的高山站点。

Research on Modified Method of Model Temperature Forecast Based on Machine Learning

Machine learning performs well in the post-processing of meteorological numerical models, but its stability and applicability need further investigation. In this paper, the ECWMF model, including the forecast products of multi-meteorological factors such as 2-meter temperature, wind and precipitation, and the actual data of 2-meter temperature observed by 80 national meteorological stations in Anhui Province are used to analyze the temperature forecast error of EC model in Anhui Province. Three machine learning algorithms, decision tree, random forest and LightGBM, were used to revise EC model 0-72 hours temperature site forecast and compared with traditional MOS correction method and SPCC subjective forecast product. The results show that the high temperature forecast error of EC model is significantly higher than that of low temperature forecast, and there is a large error in southern Anhui mountainous area and Dabie mountainous area. In the machine learning algorithm, the highest temperature forecast random forest has the best performance, the lowest temperature forecast LightGBM has the best performance. Compared with the EC model, the average absolute error (MAE) was reduced by 0.55℃ and 0.2℃, and the root means square error (RMSE) was reduced by 0.6℃ and 0.31℃, respectively. The prediction accuracy was increased by 18.16% and 5.19%. The independent modeling of alpine station and the fusion of information from surrounding stations can effectively reduce the model error. Compared with SPCC subjective prediction products, the machine learning prediction model has advantages and disadvantages in the process of high temperature and cold waves but lags in the initial stages of the weather transition. Machine learning can be used as an adjunct to conventional forecasting models and can significantly optimize or improve the accuracy of temperature prediction in traditional forecasts, particularly for high-mountain sites where data is scarce.