气象要素缺失条件下不同机器学习模型计算参考作物蒸散量比较

Comparison of Reference Crop Evapotranspiration Calculated by Different Machine Learning Models in Absence of Meteorological Elements

为了实现气象要素缺失条件下对参考作物蒸散量(ET_(0))的预测,以山西果树研究所Adcon-Ws无线自动气象站2020-2021年每日最高气温(T_(max))、最低气温(T_(min))、2 m高风速(u_(2))、相对湿度(RH)和日照时数(n)数据为例,构建了9种气象要素缺失组合下的决策树(CART)、随机森林(RF)、梯度提升决策树(GBDT)、极端梯度提升(XGBoost)、支持向量机(SVR)、BP神经网络(BPNN)和深度学习(DL)7种ET_(0)机器学习模型,以PM公式计算值作为标准值,并与经验法Hargreaves-Samani、Irmak-Allen、Makkink和Priestley-Taylor进行对比。结果表明,在所有气象要素组合中,深度学习和BP神经网络均能取得较高的模拟精度并且有较好的泛化能力,其他模型在不同气象要素缺失组合中模拟精度和泛化能力有不同的排名,但整体效果较好的是支持向量机。不同气象要素对模型模拟ET_(0)的影响程度不同,影响由大到小排序依次为n、T_(max)、T_(min)、RH、u_(2)。与4种经验法相比,机器学习模型模拟精度均大于输入相同组合的经验法。

In order to realize the prediction of reference crop evapotranspiration(ET_(0))under the absence of meteorological elements,based on the data of the daily maximum temperature(T_(max)),minimum temperature(T_(min)),2 m high wind speed(u_(2)),relative humidity(RH)and sunshine hours(n)from the Adcon-Ws wireless automatic weather station of Shanxi Fruit Tree Research Institute in 2020-2021,seven machine learning models,such as decision tree(CART),random forest(RF),gradient boosting decision tree(GBDT),extreme gradient boosting(XGBoost),support vector machine(SVR),BP neural network(BPNN)and deep learning(DL)were used for calculating ET_(0)under the absence of combination of 9 meteorological elements.The PM formula calculated value was used as the standard value and the predicted results were compared with that of empirical methods,such as Hargreaves-Samani,Irmak-Allen,Makkink and Priestley-Taylor.The results showed that deep learning and BP neural network could achieve higher simulation accuracy and better generalization ability in all meteorological element combinations.Other models had different rankings in simulation accuracy and generalization ability under different meteorological absence combinations,but the overall effect of support vector machine was better.The influence of different meteorological elements on the model simulation ET0 was different,and the order from large to small was n,T_(max),T_(min),RH,u_(2).Compared with the four empirical methods,the accuracy of the machine learning model was better for the same input combination.