The particle filter-based back propagation neural network for evapotranspiration estimation

Abstract

ABSTRACT Back propagation (BP) neural networks are one of the most commonly used artificial neural networks, widely applied in many theoretical and practical areas. The performance of BP neural networks is highly impressed by their training algorithm. Traditional BP learning algorithms have some problems and difficulties such as poor rate of convergence and easily getting stuck in local minimum. This paper proposes a hybrid particle filter-based BP (PF-BP) neural network, and suggests which PF is used to estimate (training) connection weights. The PF-BP method is employed for Evapotranspiration estimation of Tabriz region (semi-arid region) in Iran by dataset collected from 1992 to 2011. The efficiency and effectiveness of the PF-BP model is compared with standard BP, genetic algorithm-BP, imperialist competition algorithm-BP and empirical models, namely, Penman-FAO, and linear regression methods. Results reveal that the network that is used by this method has the best performance and generalization ability.