Comparison of Dynamic Artificial Neural Network and Multivariate Linear Regression Models for Inflow Forecasting Using Remote Sensing Data

This study aims to compare the ability of dynamic artificial neural network (DANN) and multivariate linear regression (LR) in forecasting monthly inflow to Shahcheraghi reservoir in Semnan province, Iran. The input data consisted monthly flow discharge, precipitation, mean temperature and snow cover area. Snow cover area was estimated using NOAA-AVHRR images, based on thresholds in histograms of different phenomena in visible and thermal channels. Dynamic artificial neural networks were determined with one hidden layer, Levenberg-Marquardt as training function, and sigmoid as transfer function Moreover, five DANN and five LR models were run with different input data and the results were compared. Root mean square (RMSE), mean bias error (MBE), mean absolute relative error (MARE), maximum relative error (REmax) and R2 (coefficient of determination) are the criteria that were used for models evaluation. The best result is gained with three inputs (inflow discharge, precipitation and snow cover area) by DANN. Regarding linear regression as a classic model in inflow forecasting, the improvement of the results by using DANN was obvious. The REmax of the selected DANN model was almost 85% less than REmax of the selected LR.