Physicochemical characterization of crude oil fractions by artificial neural networks

This paper presents a novel approach to the problem of characterization of petroleum fractions. An artificial neural network consisting of a three-lay er perceptron is used to predict volume and weight yields, viscosity, speci fic gravity and sulphur content. The network was trained using assay data r elative to crude oils from central Libya and south-west Iran. After trainin g, the predictive capabilities of the perceptron were tested on systems not included in the learning set. In addition, a comparison was made with the estimates provided by a widespread crude-oil evaluation procedure. The results obtained indicate that accuracies can be achieved that are even better than those derived from current estimation methods.