Input file names of hourly data required for RNN--WL model training and verification

Regional_Neural_Network
Page Navigation
10

Fletcher and Goss's (1993) approximation was applied to approximately estimate

the number of neurons in the hidden layer. The network was trained using the data

set and then validated with another data set. Through sensitivity study, the optimal

network size was selected as that size which resulted in the minimum error and

maximum correlation in the validation data set.

In order to account for the phase difference of water levels between inputs and

outputs, the last 4 hourly data points from the input time series of water levels at

the NOAA stations at each time step were used to predict water level at an inlet at the

given time step. This is similar to the autoregressive and moving average variables in

stochastic modeling. The neural network relationship between water level at inlet,

h(t)inlet , and at NOAA permanent station, h(t)NOAA , in the RNN--WL neural net-

work model can be illustrated by the following equation

RNN--WL[h(t),h(t 1),h(t 2),h(t 3)]NOAA Station

h(t)inlet

(a)

After a series of sensitivity tests, a network with 25 neurons in the hidden layer was

adopted. A schematic diagram of the RNN--WL model is given in Fig. 5(a). The

model is generalized for convenient user input. A simple graphic-user-interface was

developed for users to easily load data files in Windows environment (Fig. 5(b)).

Data sets required for model training and verification in Table 1. Default model

parameters are given in Table 2.

Table 1

Input file names of hourly data required for RNN--WL model training and verification

Remote NOAA station

Local inlet station

Dataset--1--NOAA

Dataset--1--Local

1. Model training

Datatset--2--NOAA

Dataset--2--Local

2. Model verification