Machine Learning Methods Based on Limited Meteorological Data to Simulate Potential Evapotranspiration: A Case Study of Source Region of Yellow River Basin

Abstract

Simulation of potential evapotranspiration (PET) is an important part of drought warning and water resource planning. However, the commonly used empirical models need to input a large number of meteorological elements. Therefore, to improve the efficiency and accuracy of PET simulation in areas lacking meteorological data, this study evaluated the performance of Extreme learning machine (ELM), multi-layer perceptron (MLP), and Random Forest (RF) three machine learning methods to simulate daily PET using limited meteorological data in the source region of the Yellow River (SYRB). Two bionic optimization algorithms, Grey Wolf Optimizer (GWO) and Sparrow Search Algorithm (SSA), were used to optimise the hyperparameters of the model to improve the accuracy of the model. In addition, the effect of months on daily PET simulations was evaluated. The results showed that the daily maximum temperature (T _max) was the most important factor affecting the PET simulation, and the daily average relative humidity (RH) and wind speed (U ₁₀) were the secondary factors. It is recommended to use T _max, RH, U ₁₀, and sunshine duration as the optimum combination of input (R ² > 0.95). In the case of limited meteorological data, the input combination of T _max, RH, U ₁₀, or T _max, RH (R ² > 0.75) was considered. Considering the accuracy and the time and space overhead of the model, the ELM-GWO model is recommended. When month information was used as an input factor, model performance improved in all scenarios, and June to July was the most accurate month for the model to simulate daily PET. This research resultwill allow researchers to choose the appropriate meteorological factor when simulating the PET to provide the reference.