Machine Intelligence-Based Reference Evapotranspiration Modelling: An application of Neural Networks

Abstract

After inventing Artificial Neural Networks, a deep learning algorithm, simulation of hydrology and water resource-related problems become more efficient. The investigation aimed to discover an efficient Artificial Neural Networks (ANN) model for obtaining weekly reference evapotranspiration (ET0) in the Tirupati region. Air temperature (T), Sunshine hours (S), Wind speed (W) and Relative Humidity (RH) are among the climate variables commonly utilized to evaluate the ET0. Multiple and partial correlation analyses were performed between the ET0 calculated by the Penman-Monteith (PM) method (PMET0) and these variables by deleting one variable each time to determine the most impacting variable, RH, W, S, and T were found to be impacting variables in the order of lowest to highest. As a result, the most desirable ANN model (ANN ET0) was created using all the variables as inputs and eliminating one of the least influential variables each time to assess ET0. The ANN models are developed and validated using climatic data from 1992 to 2001. The model's ability was evaluated using numerical indicators and scatter & comparison plots by matching the PM ET0 to the ANN ET0. The numerical indexes are employed to validate the usefulness of the generated models. The ANN (1-5-1) considering one input variable (T), ANN (2-5-1) considering two input variables (T & S), ANN (3-4-1) considering three input variables (T, S, & W), and ANN (4-3- 1) considering four input variables (T, S, W, & RH), were found to have 83.53%, 89.85%, 94.21%, and 99.30% efficiency during the validation, respectively. Therefore, the ANN models may accurately predict the weekly ET0 in the research area and elsewhere in climatological situations similar to the study area.