Application of Multi-Layered Perceptron Model to the Estimation of Chlorinity Variations in a Tidal River

Abstract

Applicability of a neural network model to the estimation of chlorinity variations in a tidal river has been studied in relation to optimal inputs for the model. Feedforward multi-layered perceptron model was adopted along with the generalized delta learning rule for determining weighs and thresholds of the network and the genetic algorithm for learning steepness parameters of sigmoid functions. The chlorinity data observed during non-irrigation season at the estuary in the Chikugo River, and at Morotomi located at about 7.3 km upstream from the river mouth were analyzed by the model.
To estimate a chlorinity at the present step, observed fresh water discharges and observed or calculated water-stages at the past several time steps were allocated input-layer of three?layered perceptron model. The applicability of the model was estimated using square errors obtained from the difference between calculated and observed chlorinities, and the optimal inputs to represent chlorinity variations were determined at the least square error.
The learning method using both the generalized delta learning rule and the genetic algorithm had good performance to represent chlorinity variations. The model needed longer history of fresh water discharge and water-stage as input at the estuary, because of strong mixture of fresh and salt water compared with the upstream region.