基于极限学习机的参考作物蒸散量预测模型

为实现气象资料缺乏情况下参考作物蒸散量（reference crop evapotranspiration, ET0）高精度预测，以气象因子的不同组合为输入参数，利用FAO-56 Penman-Monteith公式计算的ET0作为预测标准值建立基于极限学习机（extreme learning machine, ELM）的ET0预测模型。选取川中丘陵区7个气象站点1963－2012年逐日气象资料进行模型训练与测试，并将模拟结果同Hargreaves、Priestley-Taylor、Makkink及Irmark-Allen等4种常用模型进行对比。结果表明：ELM模型能很好地反映气象因子同ET0间复杂的非线性关系，且模拟精度较高；基于最高和最低温度的ELM模型模拟精度（均方根误差和模型效率系数分别为0.504 mm/d和0.827）高于Hargreaves模型（均方根误差和模型有效系数分别为0.692 mm/d和0.741）；基于最高、最低温度和辐射的ELM模型模拟精度（均方根误差和模型有效系数分别为0.291 mm/d和0.938）明显高于Priestley-Taylor（均方根误差和模型有效系数分别为0.467 mm/d和0.823）、Makkink（均方根误差和模型有效系数分别为0.540 mm/d和0.800）和Irmark-Allen模型（均方根误差和模型有效系数分别为0.880 mm/d和0.623）。因此基于最高、最低温度和辐射的ELM模型可以作为气象资料缺乏情况下川中丘陵区ET0计算的推荐模型。该研究可为川中丘陵区气象资料缺乏情境下ET0精确计算提供科学依据。

Prediction model of reference crop evapotranspiration based on extreme learning machine

Abstract: Reference crop evapotranspiration (ET0) is an essential parameter of water resource planning and management. Accurate estimation of ET0 becomes vital in planning and optimizing irrigation schedules and irrigation systems management. Numerous methods have been proposed for estimating ET0, among which Penman-Monteith (P-M) model recommended by Food and Agriculture Organization of the United Nations (FAO) in 1998 is the best one. FAO accepted the P-M model as the standard and sole equation for ET0 estimation since it provided the most accurate results across the world wherever in an arid or humid environment. But the main problems for computing ET0 by the P-M model are its complicated nonlinear process and requirements of many climatic variables. Thus, it is urgent to develop a simpler and more appropriate model in areas with limited data especially in developing countries like China. In the current study, the applicability of extreme learning machine (ELM) in ET0 modeling based on limited data was assessed in the humid environment in hilly area of central Sichuan, China. In addition, four climate-based models (Hargreaves, Priestley-Taylor, Makkink and Irmark-Allen) and the ELM model were tested against the P-M model to study their performance by using three commonly used criteria: root mean square error (RMSE), coefficient of determination (R2) and efficiency coefficient (Ens). From the statistical results, the ELM model performed well in expressing the nonlinear relationship between ET0 and meteorological factors; when based on temperature data, the ELM model performed better than Hargreaves model which is an empirical temperature-based model. When radiation and temperature data were introduced in the ELM model, the error decreased significantly, and it was much more accurate than the Priestley-Taylor, Makkink and Irmark-Allen model. It was found that the ELM model, which required maximum air temperature, minimum air temperature and sunshine duration input variables, had the best accuracy and was the optimal approach to estimate ET0 when the complete weather data required by the P-M model were not available. The further assessment of ELM was conducted and it was confirmed that the model could provide a powerful tool in estimating ET0 in the humid environment like hilly area of central Sichuan when lack of meteorological data. The research could provide a reference to accurate ET0 estimation in hilly area of central Sichuan.