NARX Recurrent Neural-Network Model for Long-Term Forecasting of Daily flow

Location

Eccles Conference Center

Event Website

http://water.usu.edu

Start Date

4-1-2014 7:05 PM

End Date

4-1-2014 7:10 PM

Description

Forecasted river flow can be utilized to optimally allocate water to agriculture, municipal, and industrial demands for the coming year. This research forecasts long-term daily inflow into Dez Reservoir. To develop a NARX Recurrent Neural Network (NARX-RNN) model, we evaluate different input delays, number of neurons in the hidden layer, types of transfer functions, and different numbers of output delays. The best NARX-RNN model, based upon root mean square error (RMSE) and mean bias error (MBE), uses a log-sigmoid transfer function, 4 neurons in the hidden layer and 10 outputs delay. We also use an Auto Regressive Integrated Moving Average (ARIMA) model which uses Fourier series parameters to eliminate seasonal trend. Training and testing both models employs daily discharge data from 1975 to 2011. Comparison of observed and forecasted reservoir inflow results shows that NARX-RNN operates more successfully than ARIMA model in 365-day daily inflow forecasting, especially for peak flows.

This document is currently not available here.

Share

COinS
 
Apr 1st, 7:05 PM Apr 1st, 7:10 PM

NARX Recurrent Neural-Network Model for Long-Term Forecasting of Daily flow

Eccles Conference Center

Forecasted river flow can be utilized to optimally allocate water to agriculture, municipal, and industrial demands for the coming year. This research forecasts long-term daily inflow into Dez Reservoir. To develop a NARX Recurrent Neural Network (NARX-RNN) model, we evaluate different input delays, number of neurons in the hidden layer, types of transfer functions, and different numbers of output delays. The best NARX-RNN model, based upon root mean square error (RMSE) and mean bias error (MBE), uses a log-sigmoid transfer function, 4 neurons in the hidden layer and 10 outputs delay. We also use an Auto Regressive Integrated Moving Average (ARIMA) model which uses Fourier series parameters to eliminate seasonal trend. Training and testing both models employs daily discharge data from 1975 to 2011. Comparison of observed and forecasted reservoir inflow results shows that NARX-RNN operates more successfully than ARIMA model in 365-day daily inflow forecasting, especially for peak flows.

https://digitalcommons.usu.edu/runoff/2014/2014Posters/1