A Comparative Analysis of Machine Learning Approaches to Gap Filling Meteorological Datasets (Results Only)

This dataset contains the results from our evaluation of methodologies for filling gaps in meterological data. Variables were chosen to represent a standard set of measurements for purposes typically performed using Weather Stations installed in urban and rural areas. There are 4 dimensions by which we measure and validate each of the gap filling models across a large set of experimental configurations: meteorological variables TargetVar (dewpoint, humidity, leaf wetness, temp); feature_set (AWS, AWS-ERA5, ERA5, ERA5_Debias, Spatial); 3 types of machine learning algorithms ML (linear regression, random forests, LightGBM) combined with 2 non-ML algorithms; and gap_length: 1, 4, 36 and 288. A total of 1,720 experiments were conducted: 1,440 machine learning experiments; 160 using a spatial algorithm and 120 experiments using ERA5. For the 3 machine learning experiments, the average result was selected for each of 10 sites for 4 gap sizes (40 results) with the 3 ML models using 3 different feature sets (120 results). Data is provided in both CSV format and as a MySQL dump. Resultsets are accompanied with the SQL expression used to generate the result.本数据集收录了我们针对气象数据缺失补全方法开展的评估成果。本次选取的变量可代表城乡布设气象站常规开展的标准监测项目集合。本次评估共从4个维度对多组实验配置下的各类补全模型进行测试与验证：气象变量（TargetVar），涵盖露点、湿度、叶面湿度、气温四类指标；特征集（feature_set），包含AWS、AWS-ERA5、ERA5、ERA5_Debias、Spatial共5种类型；机器学习算法（ML）类型，涵盖线性回归、随机森林、LightGBM共3类机器学习算法，外加2种非机器学习算法；以及缺失时长（gap_length），分别为1、4、36与288个时间单位。本次研究共开展1720组对照实验：其中机器学习实验1440组，空间算法实验160组，基于ERA5的实验120组。针对机器学习实验，我们为10个监测站点的4种缺失时长各选取平均实验结果（共40组结果），搭配3种不同特征集与3类机器学习模型进行测试，最终形成120组实验结果。数据集同时提供CSV格式文件与MySQL转储文件。各结果集均附带生成该结果所用的SQL查询语句。