Modeling and simulation of OBPN for forecasting of long-range monsoon AWR over a smaller homogeneous region through astrological parameters and its verification for 2019

Abstract

In this research work the astrological parameters has been identified, those are physically linked with long range area weighted rainfall (LR-AWR) of 132 rain-gauge stations over Mahanadi river basin Chhattisgarh, India which is geographically located at 80° 28′ to 86° 43′ E and 19° 8′ to 23° 32′ N. To forecast of LR-AWR over this smaller region for 2018 and 2019, an optimum back-propagation neural network (OBPN) is implemented through various experiments wherein its architectural parameters such as number of input vector (n), number of output neuron (m), hidden layer neurons (p), learning rate (α), momentum factor (µ), and optimum learning conditions such as local minima ($l_{\rm{min} }$), global minima ($g_{\rm{min} }$), global training point (GTP), and optimum learning positions (OLP) are optimized. In this case, the OBPN is found as n = 4, m = 1, p = 2, α = 0.61, and µ = 0.89. During its training with various experimental results and their analysis, $l_{\rm{min} }$ is obtained at e = 4864 epochs, GTP is finalized for e = 300,000 epochs, $g_{\rm{min} }$ is obtained at MSE = 1.525511836204590E−04 at 300,000 epochs, and OLP is obtained as 1.264603588205336 (MAD % of LPA). As far as the performance is concerned, the OBPN has shown excellent performance during training process except for the years 2003, 2009, 2013, and for other years the deviation between actual and predicted % of LPA is found to be just 5–8% or say below 10%, while for years 2007, 2008, and 2006 deviation is almost 0%. During the testing period except for year 2015 it performed excellently. For 2014, 2016 it is 8.5% and 6.4%, respectively; however for the year 2017 it is only 0.7%. OBPN successfully forecasted LR-AWR for the years 2018 and 2019 with 9.6% and 7.2% deviations, respectively, as presented through this paper.