Linear and non-linear ensemble concepts for pan evaporation modeling


Abdullahi J., Elkiran G., Aslanova F., Orhon D.

Desalination and Water Treatment, vol.293, pp.67-81, 2023 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 293
  • Publication Date: 2023
  • Doi Number: 10.5004/dwt.2023.29531
  • Journal Name: Desalination and Water Treatment
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Biotechnology Research Abstracts, CAB Abstracts, Environment Index, Geobase, Pollution Abstracts, Veterinary Science Database, Civil Engineering Abstracts
  • Page Numbers: pp.67-81
  • Keywords: Artificial neural network, Ensemble modeling, Erbil, Pan evaporation, station
  • Istanbul Technical University Affiliated: Yes

Abstract

Modeling of pan evaporation (Ep) is of paramount importance in the evaluation of drinking water supplies, planning of regional water resources and reservoir management. The main aim of this study is to investigate the accuracy of linear and non-linear ensemble approaches for monthly Ep modeling in Erbil and Salahaddin meteorological stations of Iraq. For this purpose, sensitivity analysis was performed to determine the dominant input parameters. The results showed that Tmean, Tmax and Tmin are the most effective parameters. Thereafter, two scenarios were involved for the Ep modeling. In scenario 1, the ability of artificial neural network, least-squares support-vector machine and multiple linear regression models was examined for the estimation of Ep. The results demonstrated that different input combinations led to different performance, model 3 (which has Tmean, Tmax, Tmin, RH) for Erbil station and model 2 (which has Tmean, Tmax, Tmin) for Salahaddin station provided the best performance among several models developed. In scenario 2, linear and non-linear ensemble approaches were employed as simple linear average, weighted linear average and non-linear ensemble (NLE) models to improve predictions of the single models. The results reported that ensemble modeling could improve performance of single models and NLE model provided the best results due to its non-linear nature. The general results demonstrated that the proposed ensemble models could improve predictions of single models up to 5% and 16% for Erbil and Salahaddin stations, respectively.