Forecasting Evaporative Loss by Least-Square Support-Vector Regression and Evaluation with Genetic Programming, Gaussian Process, and Minimax Probability Machine Regression: Case Study of Brisbane City: Fid-Bau Portal

Forecasting Evaporative Loss by Least-Square Support-Vector Regression and Evaluation with Genetic Programming, Gaussian Process, and Minimax Probability Machine Regression: Case Study of Brisbane City

Deo, Ravinesh C. / Samui, Pijush

Abstract

Daily evaporative loss ( $E_{p}$ ) forecasting models are decisive tools with potential applications in hydrology, the design of water systems, urban water assessments, and irrigation management. This paper performs a case study for forecasting daily $E_{p}$ for Brisbane city using least-square support-vector regression (LSSVR). A limited set of predictor data with solar radiation and exposure, maximum/minimum temperatures, wind speed, and precipitation (March 1, 2014 to March 31, 2015) is adopted to develop the predictive model. The results are evaluated with Gaussian process regression (GPR), minimax probability machine regression (MPMR), and genetic programming (GP) models. In the testing phase, a correlation coefficient of 0.895 is attained between the observed and forecasted $E_{p}$ by LSSVR that contrasted 0.875 (GPR), 0.864 (MPMR), and 0.628 (GP). A sensitivity test of predictor variables shows that approximately 28.5% of features are extracted from solar radiation data with 18.1% (wind speed), 16.6% (precipitation), and 10–15% (minimum and maximum temperature). The root-mean square error for LSSVR is lower than the GPR, MPMR, and GP models by 16.2, 11.4, and 79.4%, and the cumulative frequency of forecasting error attained for LSSVR is the highest within the smallest error band. The results confirm the better utility of LSSVR in relation to GP, GPR, and MPMR models for forecasting daily evaporative loss.