Study of the Impact of the Topology of Artificial Neural Networks for the Prediction of Meteorological Data

Abstract

In this chapter, we present an analysis of the capacity of six topologies of neural networks—multilayer perceptron (MLP), complete (CP), random (RD), scale-free (SF), small world (SW), and hybrid (HY)—to perform next-step predictions of temperature and solar radiation. For this purpose, 100 networks of each complex topology, including an MLP network and a complete network, were created, and each network contained twenty-six neurons (five entries, twenty processing units, and one output). The networks were trained for 2000 epochs and used for the next-step prediction of the selected climate variables. The MLP and Complete networks were trained 100 times, and the averages were taken for comparison. The parameters of comparison were the root mean square error (RMSE), the mean time to learn (ETL), and the mean time to predict the next step after training (ETP), using 1000 h of input data obtained from INMET (National Institute of Meteorology), a station on the coast of Bahia. The set of 1000 h of data was divided into two parts: 700 h for network training and 300 h to verify the effectiveness of learning.