Neural Network-Based Prediction Model for Evaporation Using Weather Data

Abstract

The information on evaporation has become important during the last decades because of monitoring and management of water resources and crop water requirement. The principal weather variables affecting evaporation are radiation, air temperature, humidity, rainfall and wind speed. Models or procedures have been developed to assess the evaporation from weather variables; it is mainly based on regression models (linear and nonlinear) in studying relationship of evaporation with weather variables (as such or in some transformed forms). Nowadays, artificial neural networks (ANNs) techniques have become very popular because of their wide range of applicability and the ease with which they can treat complex problems even if the data are vague and noisy. From modeling perception, neural networks are interesting because of their potential use in prediction. This approach has been illustrated for prediction of evaporation for four locations, viz. Maruteru (16.6269° N, 81.7389° E), Palampur (32.1109° N, 76.5363° E), Patancheru (17.5287° N, 78.2667° E) and Raipur (21.2514° N, 81.6296° E) based on fifteen years’ historical data. Consider the different lag period (4–6 weeks) on maximum and minimum temperature, relative humidity, basic sunshine hours and rainfall as the input variables and evaporation as the output variable. ANN models with back propagation as the learning algorithm with varying architecture of hidden layers (one and two) and varying neurons in hidden layers were developed. The developed models were validated for subsequent weeks (52 standard meteorological weeks) for different locations. On the basis of mean squared errors, ANN-based models with 2 layers each having 5 & 6 neurons in respective layers showed the most promising results for Maruteru, while in Palampur, Patancheru and Raipur, the 2L-5-6N, 2L-6-5N and 2L-5-5N architectures gave the best results.