Monthly ETo Forecasting Using PO-LSSVM Based on Wavelet Packet Decomposition and Analysis on the Prediction Uncertainty

Abstract

The least square support vector machine (LSSVM) is a widely-used machine learning model for classification and regression. However, its performance is affected by two control parameters γ and σ. As a novel artificial intelligence technology, optimization algorithms have been broadly used to train machine learning model because of their simplicity and efficiency. Inthiswork, a recently proposed novel optimization algorithm named political optimizer (PO) is adopted to select the appropriate values of two parameters γ and σ in LSSVM for elimination of the subjectivity in artificial parameter adjustment and enhancement in its performance. With the help of wavelet packet decomposition, the combined model PO-LSSVM is evaluated for a case study in forecasting monthly ETo, namely the reference crop evapotranspiration which is of great importance for agricultural water resources management. And the comparative study of PO-LSSVM with control models PSO-SVR (particle swarm optimization algorithm-support vector regression), PO-SVR and PSO-LSSVM is conducted with Friedman test and Wilcoxon signed rank test based on the evaluation results. The results demonstrate that PO-LSSVM ranks first for providing the highest accuracy and stability and PO is statistically better than PSO in optimizing the parameters. Therefore, the PO algorithm is feasible and efficient in the optimization of super parameters for LSSVM. The established PO-LSSVM is of highly accuracy and efficiency in forecasting ETo based on our case study which can be a reference for practical production. In addition, the analysis on the prediction uncertainty is carried out to indicate scientific guidance for water resources allocation from a unique perspective.