Investigating the performance of neural network backpropagation algorithms for TEC estimations using South African GPS data

摘要

In this work, results obtained by investigating the application ofdifferent neural network backpropagation training algorithms are presented.This was done to assess the performance accuracy of each training algorithmin total electron content (TEC) estimations using identical datasets inmodels development and verification processes. Investigatedtraining algorithms are standard backpropagation (SBP), backpropagation withweight delay (BPWD), backpropagation with momentum (BPM) term,backpropagation with chunkwise weight update (BPC) and backpropagation forbatch (BPB) training. These five algorithms are inbuilt functions within theStuttgart Neural Network Simulator (SNNS) and the main objective was to findout the training algorithm that generates the minimum error between the TEC derived from Global Positioning System (GPS) observations and the modelled TECdata. Another investigated algorithm is the MatLab basedLevenberg-Marquardt backpropagation (L-MBP), which achieves convergence afterthe least number of iterations during training. In this paper, neural network(NN) models were developed using hourly TEC data (for 8 years: 2000–2007)derived from GPS observations over a receiverstation located at Sutherland (SUTH) (32.38° S, 20.81° E), SouthAfrica. Verification of the NN models for all algorithms considered wasperformed on both "seen" and "unseen" data. Hourly TEC values over SUTHfor 2003 formed the "seen" dataset. The "unseen" dataset consisted ofhourly TEC data for 2002 and 2008 over Cape Town (CPTN) (33.95° S,18.47° E) and SUTH, respectively. The models' verification showed that allalgorithms investigated provide comparable results statistically, but differsignificantly in terms of time required to achieve convergence duringinput-output data training/learning. This paper therefore provides a guide toneural network users for choosing appropriate algorithms based on theavailability of computation capabilities used for research.