Redes neurais artificiais para estimar a precipitação na irrigação por aspersão

Abstract

This work presents a theoretical-conceptual approach of the sprinkler irrigation model, reporting its methods, efficiency and management. It presents the concepts, learning techniques, topologies and algorithms for the training of artificial neural networks. The objective of the research was to build a model based on artificial neural networks capable of estimating the precipitation of a sprinkler. The backpropagation training algorithm was used with and without the term momentum and several learning rates in order to lead to the choice of the best network using SNNS software. The network (6X459X1) with 6 neurons in the input layer, 459 in the hidden layer and 1 in the output layer was the one that presented the lowest error (MSE) obtained with a learning rate of 0.7 and with the use of the term moment (μ = 0.3). The simulations presented good results in the statistical analysis, with coefficient of determination (R2) equal to 0.93 for linear adjustment and normal distribution for Lilliefors. Anova's analysis reflected valid homogeneity and accepted the hypothesis for the Cochran C and Bartlett test. The Pearson coefficient resulted in a very strong correlation between the simulated and observed results, and the comparison was around 0.0 to 0.29 mm of precipitation. The tests were promising in terms of estimates for sprinkler irrigation.