Application of artificial neural networks in the prediction of sugarcane juice Pol/Aplicacao de redes neurais artificiais na predicao do Pol do caldo da cana-de-acucar

Innovative techniques that seek to minimize the costs of production and the laboriousness of certain operations are one of the great challenges in the sugar-energy sector nowadays. Thus, the objective of the present study was to estimate the Pol values of sugarcane juice as a function of [degrees]Brix and wet cake weight (WCW) using artificial neural network (ANN) modeling. A database was organized consisting of 204 technological analyses from a field experiment with 15 treatments and 2 years of evaluation. 75% of the data were used for the calibration of the model and 25% for its validation. Multilayer Perceptron ANNs were used for calibration and validation of the data. Before calibration, the variables were normalized. The training algorithm used was backpropagation and the activation function was the sigmoid. The ANNs were established with two hidden layers and the number of neurons ranging from 4 to 20 in each. The 15 ANNs with the lowest root mean square errors were randomly presented by the software, among which 6 were chosen to verify the accuracy. The ANNs had a high accuracy in the estimation of sugarcane juice Pol, both in the calibration phase ([R.sup.2] = 0.948, RMSE = 0.36%) and in the validation ([R.sup.2] = 0.878, RMSE = 0.41%), and can replace the standard method of analysis. Simpler networks can be trained to have the same accuracy as more complex networks. Key words: trs, Brix, sucrose, technological quality Tecnicas inovadoras que busquem minimizar os custos de producao e a onerosidade de determinadas operacoes sao um dos grandes desafios atualmente no setor sucroenergetico. Nesse sentido, objetivou-se estimar os valores do Pol do caldo da cana-de-acucar, em funcao do [degrees]Brix e do peso do bolo umido (PBU), utilizando modelagem por redes neurais artificiais (RNAs). Foi organizado um banco de dados constituido de 204 analises tecnologicas provenientes de um experimento de campo com 15 tratamentos e 2 anos de avaliacao. Foram utilizados 75% dos dados para a calibracao do modelo e 25% para a validacao. Foram utilizadas RNAs do tipo Multilayer Perceptron para calibracao e validacao dos dados. Antes da calibracao, as variaveis foram normalizadas. O algoritmo de treinamento utilizado foi o backpropagation e a funcao de ativacao foi a sigmoide. As RNAs foram estabelecidas com duas camadas ocultas e o numero de neuronios variando de 4 a 20 em cada. As 15 RNAs com menor raiz do erro quadratico medio foram apresentadas aleatoriamente pelo so