基于AutoML的长江下游周旬尺度枯水位及潮位预报

Forecast of low water level and tide level in lower reaches of the Yangtze River on week and ten-day scales based on AutoML

基于微软在2018年发布的自动化机器学习建模平台Azure AutoML,探索长江下游水位及潮位预报模型构建与应用。以大通枯水期周旬尺度最低水位、南京未来14与20次低潮中最低潮位的预报为例,开展研究并从模型构建与评估、预报精度、输入因子重要性等角度进行分析。研究结论表明:微软Azure AutoML平台可便捷地进行自动化机器学习模型的构建;两站点的预报模型在2014—2020年模型构建过程、2021年模型精度分析过程中均取得较高的精度指标,但在最低水位(潮位)波动变幅较大的阶段,预报模型的性能有待进一步提高;大通站不同预见期预报模型的重要输入因子较为一致,排名前三的重要变量依次为八里江、大通、安庆水位;南京潮位预报规律较为复杂,应尽量纳入更长时段的前期潮位信息。

Azure AutoML is an automated machine learning modeling platform released by Microsoft in 2018.Based on this platform,we explore the building and application of water level and tide level forecasting models in the lower reaches of the Yangtze River.We carry out research by taking the prediction of the lowest water level in week and ten-day scales in the dry season of Datong and the lowest tide level in the future 14 and 20 low tides in Nanjing as an example.In addition,we analyze from the perspectives of model building and evaluation,prediction accuracy,and the importance of input factors.The results show that the Microsoft Azure AutoML platform can facilitate the building of automated machine learning models.The forecast model of the two stations can achieve a high overall accuracy index during the model building from 2014 to 2020 and model accuracy analysis in 2021.However,under the greatly fluctuating lowest water level(tide level),the performance of the forecast model needs to be further improved.The important input factors of the forecast model for different forecast periods of Datong station are relatively consistent,and the top three important variables are the water level of Balijiang,Datong,and Anqing in turn.The forecast law of tide level in Nanjing is relatively complex and should be included in the early tide level information for a longer period as far as possible.